diff options
Diffstat (limited to 'drivers/webp/dsp')
58 files changed, 20965 insertions, 2234 deletions
diff --git a/drivers/webp/dsp/alpha_processing.c b/drivers/webp/dsp/alpha_processing.c new file mode 100644 index 0000000000..1716cace8d --- /dev/null +++ b/drivers/webp/dsp/alpha_processing.c @@ -0,0 +1,383 @@ +// Copyright 2013 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Utilities for processing transparent channel. +// +// Author: Skal (pascal.massimino@gmail.com) + +#include <assert.h> +#include "./dsp.h" + +// Tables can be faster on some platform but incur some extra binary size (~2k). +// #define USE_TABLES_FOR_ALPHA_MULT + +// ----------------------------------------------------------------------------- + +#define MFIX 24 // 24bit fixed-point arithmetic +#define HALF ((1u << MFIX) >> 1) +#define KINV_255 ((1u << MFIX) / 255u) + +static uint32_t Mult(uint8_t x, uint32_t mult) { + const uint32_t v = (x * mult + HALF) >> MFIX; + assert(v <= 255); // <- 24bit precision is enough to ensure that. + return v; +} + +#ifdef USE_TABLES_FOR_ALPHA_MULT + +static const uint32_t kMultTables[2][256] = { + { // (255u << MFIX) / alpha + 0x00000000, 0xff000000, 0x7f800000, 0x55000000, 0x3fc00000, 0x33000000, + 0x2a800000, 0x246db6db, 0x1fe00000, 0x1c555555, 0x19800000, 0x172e8ba2, + 0x15400000, 0x139d89d8, 0x1236db6d, 0x11000000, 0x0ff00000, 0x0f000000, + 0x0e2aaaaa, 0x0d6bca1a, 0x0cc00000, 0x0c249249, 0x0b9745d1, 0x0b1642c8, + 0x0aa00000, 0x0a333333, 0x09cec4ec, 0x0971c71c, 0x091b6db6, 0x08cb08d3, + 0x08800000, 0x0839ce73, 0x07f80000, 0x07ba2e8b, 0x07800000, 0x07492492, + 0x07155555, 0x06e45306, 0x06b5e50d, 0x0689d89d, 0x06600000, 0x063831f3, + 0x06124924, 0x05ee23b8, 0x05cba2e8, 0x05aaaaaa, 0x058b2164, 0x056cefa8, + 0x05500000, 0x05343eb1, 0x05199999, 0x05000000, 0x04e76276, 0x04cfb2b7, + 0x04b8e38e, 0x04a2e8ba, 0x048db6db, 0x0479435e, 0x04658469, 0x045270d0, + 0x04400000, 0x042e29f7, 0x041ce739, 0x040c30c3, 0x03fc0000, 0x03ec4ec4, + 0x03dd1745, 0x03ce540f, 0x03c00000, 0x03b21642, 0x03a49249, 0x03976fc6, + 0x038aaaaa, 0x037e3f1f, 0x03722983, 0x03666666, 0x035af286, 0x034fcace, + 0x0344ec4e, 0x033a5440, 0x03300000, 0x0325ed09, 0x031c18f9, 0x0312818a, + 0x03092492, 0x03000000, 0x02f711dc, 0x02ee5846, 0x02e5d174, 0x02dd7baf, + 0x02d55555, 0x02cd5cd5, 0x02c590b2, 0x02bdef7b, 0x02b677d4, 0x02af286b, + 0x02a80000, 0x02a0fd5c, 0x029a1f58, 0x029364d9, 0x028ccccc, 0x0286562d, + 0x02800000, 0x0279c952, 0x0273b13b, 0x026db6db, 0x0267d95b, 0x026217ec, + 0x025c71c7, 0x0256e62a, 0x0251745d, 0x024c1bac, 0x0246db6d, 0x0241b2f9, + 0x023ca1af, 0x0237a6f4, 0x0232c234, 0x022df2df, 0x02293868, 0x02249249, + 0x02200000, 0x021b810e, 0x021714fb, 0x0212bb51, 0x020e739c, 0x020a3d70, + 0x02061861, 0x02020408, 0x01fe0000, 0x01fa0be8, 0x01f62762, 0x01f25213, + 0x01ee8ba2, 0x01ead3ba, 0x01e72a07, 0x01e38e38, 0x01e00000, 0x01dc7f10, + 0x01d90b21, 0x01d5a3e9, 0x01d24924, 0x01cefa8d, 0x01cbb7e3, 0x01c880e5, + 0x01c55555, 0x01c234f7, 0x01bf1f8f, 0x01bc14e5, 0x01b914c1, 0x01b61eed, + 0x01b33333, 0x01b05160, 0x01ad7943, 0x01aaaaaa, 0x01a7e567, 0x01a5294a, + 0x01a27627, 0x019fcbd2, 0x019d2a20, 0x019a90e7, 0x01980000, 0x01957741, + 0x0192f684, 0x01907da4, 0x018e0c7c, 0x018ba2e8, 0x018940c5, 0x0186e5f0, + 0x01849249, 0x018245ae, 0x01800000, 0x017dc11f, 0x017b88ee, 0x0179574e, + 0x01772c23, 0x01750750, 0x0172e8ba, 0x0170d045, 0x016ebdd7, 0x016cb157, + 0x016aaaaa, 0x0168a9b9, 0x0166ae6a, 0x0164b8a7, 0x0162c859, 0x0160dd67, + 0x015ef7bd, 0x015d1745, 0x015b3bea, 0x01596596, 0x01579435, 0x0155c7b4, + 0x01540000, 0x01523d03, 0x01507eae, 0x014ec4ec, 0x014d0fac, 0x014b5edc, + 0x0149b26c, 0x01480a4a, 0x01466666, 0x0144c6af, 0x01432b16, 0x0141938b, + 0x01400000, 0x013e7063, 0x013ce4a9, 0x013b5cc0, 0x0139d89d, 0x01385830, + 0x0136db6d, 0x01356246, 0x0133ecad, 0x01327a97, 0x01310bf6, 0x012fa0be, + 0x012e38e3, 0x012cd459, 0x012b7315, 0x012a150a, 0x0128ba2e, 0x01276276, + 0x01260dd6, 0x0124bc44, 0x01236db6, 0x01222222, 0x0120d97c, 0x011f93bc, + 0x011e50d7, 0x011d10c4, 0x011bd37a, 0x011a98ef, 0x0119611a, 0x01182bf2, + 0x0116f96f, 0x0115c988, 0x01149c34, 0x0113716a, 0x01124924, 0x01112358, + 0x01100000, 0x010edf12, 0x010dc087, 0x010ca458, 0x010b8a7d, 0x010a72f0, + 0x01095da8, 0x01084a9f, 0x010739ce, 0x01062b2e, 0x01051eb8, 0x01041465, + 0x01030c30, 0x01020612, 0x01010204, 0x01000000 }, + { // alpha * KINV_255 + 0x00000000, 0x00010101, 0x00020202, 0x00030303, 0x00040404, 0x00050505, + 0x00060606, 0x00070707, 0x00080808, 0x00090909, 0x000a0a0a, 0x000b0b0b, + 0x000c0c0c, 0x000d0d0d, 0x000e0e0e, 0x000f0f0f, 0x00101010, 0x00111111, + 0x00121212, 0x00131313, 0x00141414, 0x00151515, 0x00161616, 0x00171717, + 0x00181818, 0x00191919, 0x001a1a1a, 0x001b1b1b, 0x001c1c1c, 0x001d1d1d, + 0x001e1e1e, 0x001f1f1f, 0x00202020, 0x00212121, 0x00222222, 0x00232323, + 0x00242424, 0x00252525, 0x00262626, 0x00272727, 0x00282828, 0x00292929, + 0x002a2a2a, 0x002b2b2b, 0x002c2c2c, 0x002d2d2d, 0x002e2e2e, 0x002f2f2f, + 0x00303030, 0x00313131, 0x00323232, 0x00333333, 0x00343434, 0x00353535, + 0x00363636, 0x00373737, 0x00383838, 0x00393939, 0x003a3a3a, 0x003b3b3b, + 0x003c3c3c, 0x003d3d3d, 0x003e3e3e, 0x003f3f3f, 0x00404040, 0x00414141, + 0x00424242, 0x00434343, 0x00444444, 0x00454545, 0x00464646, 0x00474747, + 0x00484848, 0x00494949, 0x004a4a4a, 0x004b4b4b, 0x004c4c4c, 0x004d4d4d, + 0x004e4e4e, 0x004f4f4f, 0x00505050, 0x00515151, 0x00525252, 0x00535353, + 0x00545454, 0x00555555, 0x00565656, 0x00575757, 0x00585858, 0x00595959, + 0x005a5a5a, 0x005b5b5b, 0x005c5c5c, 0x005d5d5d, 0x005e5e5e, 0x005f5f5f, + 0x00606060, 0x00616161, 0x00626262, 0x00636363, 0x00646464, 0x00656565, + 0x00666666, 0x00676767, 0x00686868, 0x00696969, 0x006a6a6a, 0x006b6b6b, + 0x006c6c6c, 0x006d6d6d, 0x006e6e6e, 0x006f6f6f, 0x00707070, 0x00717171, + 0x00727272, 0x00737373, 0x00747474, 0x00757575, 0x00767676, 0x00777777, + 0x00787878, 0x00797979, 0x007a7a7a, 0x007b7b7b, 0x007c7c7c, 0x007d7d7d, + 0x007e7e7e, 0x007f7f7f, 0x00808080, 0x00818181, 0x00828282, 0x00838383, + 0x00848484, 0x00858585, 0x00868686, 0x00878787, 0x00888888, 0x00898989, + 0x008a8a8a, 0x008b8b8b, 0x008c8c8c, 0x008d8d8d, 0x008e8e8e, 0x008f8f8f, + 0x00909090, 0x00919191, 0x00929292, 0x00939393, 0x00949494, 0x00959595, + 0x00969696, 0x00979797, 0x00989898, 0x00999999, 0x009a9a9a, 0x009b9b9b, + 0x009c9c9c, 0x009d9d9d, 0x009e9e9e, 0x009f9f9f, 0x00a0a0a0, 0x00a1a1a1, + 0x00a2a2a2, 0x00a3a3a3, 0x00a4a4a4, 0x00a5a5a5, 0x00a6a6a6, 0x00a7a7a7, + 0x00a8a8a8, 0x00a9a9a9, 0x00aaaaaa, 0x00ababab, 0x00acacac, 0x00adadad, + 0x00aeaeae, 0x00afafaf, 0x00b0b0b0, 0x00b1b1b1, 0x00b2b2b2, 0x00b3b3b3, + 0x00b4b4b4, 0x00b5b5b5, 0x00b6b6b6, 0x00b7b7b7, 0x00b8b8b8, 0x00b9b9b9, + 0x00bababa, 0x00bbbbbb, 0x00bcbcbc, 0x00bdbdbd, 0x00bebebe, 0x00bfbfbf, + 0x00c0c0c0, 0x00c1c1c1, 0x00c2c2c2, 0x00c3c3c3, 0x00c4c4c4, 0x00c5c5c5, + 0x00c6c6c6, 0x00c7c7c7, 0x00c8c8c8, 0x00c9c9c9, 0x00cacaca, 0x00cbcbcb, + 0x00cccccc, 0x00cdcdcd, 0x00cecece, 0x00cfcfcf, 0x00d0d0d0, 0x00d1d1d1, + 0x00d2d2d2, 0x00d3d3d3, 0x00d4d4d4, 0x00d5d5d5, 0x00d6d6d6, 0x00d7d7d7, + 0x00d8d8d8, 0x00d9d9d9, 0x00dadada, 0x00dbdbdb, 0x00dcdcdc, 0x00dddddd, + 0x00dedede, 0x00dfdfdf, 0x00e0e0e0, 0x00e1e1e1, 0x00e2e2e2, 0x00e3e3e3, + 0x00e4e4e4, 0x00e5e5e5, 0x00e6e6e6, 0x00e7e7e7, 0x00e8e8e8, 0x00e9e9e9, + 0x00eaeaea, 0x00ebebeb, 0x00ececec, 0x00ededed, 0x00eeeeee, 0x00efefef, + 0x00f0f0f0, 0x00f1f1f1, 0x00f2f2f2, 0x00f3f3f3, 0x00f4f4f4, 0x00f5f5f5, + 0x00f6f6f6, 0x00f7f7f7, 0x00f8f8f8, 0x00f9f9f9, 0x00fafafa, 0x00fbfbfb, + 0x00fcfcfc, 0x00fdfdfd, 0x00fefefe, 0x00ffffff } +}; + +static WEBP_INLINE uint32_t GetScale(uint32_t a, int inverse) { + return kMultTables[!inverse][a]; +} + +#else + +static WEBP_INLINE uint32_t GetScale(uint32_t a, int inverse) { + return inverse ? (255u << MFIX) / a : a * KINV_255; +} + +#endif // USE_TABLES_FOR_ALPHA_MULT + +void WebPMultARGBRowC(uint32_t* const ptr, int width, int inverse) { + int x; + for (x = 0; x < width; ++x) { + const uint32_t argb = ptr[x]; + if (argb < 0xff000000u) { // alpha < 255 + if (argb <= 0x00ffffffu) { // alpha == 0 + ptr[x] = 0; + } else { + const uint32_t alpha = (argb >> 24) & 0xff; + const uint32_t scale = GetScale(alpha, inverse); + uint32_t out = argb & 0xff000000u; + out |= Mult(argb >> 0, scale) << 0; + out |= Mult(argb >> 8, scale) << 8; + out |= Mult(argb >> 16, scale) << 16; + ptr[x] = out; + } + } + } +} + +void WebPMultRowC(uint8_t* const ptr, const uint8_t* const alpha, + int width, int inverse) { + int x; + for (x = 0; x < width; ++x) { + const uint32_t a = alpha[x]; + if (a != 255) { + if (a == 0) { + ptr[x] = 0; + } else { + const uint32_t scale = GetScale(a, inverse); + ptr[x] = Mult(ptr[x], scale); + } + } + } +} + +#undef KINV_255 +#undef HALF +#undef MFIX + +void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse); +void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha, + int width, int inverse); + +//------------------------------------------------------------------------------ +// Generic per-plane calls + +void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows, + int inverse) { + int n; + for (n = 0; n < num_rows; ++n) { + WebPMultARGBRow((uint32_t*)ptr, width, inverse); + ptr += stride; + } +} + +void WebPMultRows(uint8_t* ptr, int stride, + const uint8_t* alpha, int alpha_stride, + int width, int num_rows, int inverse) { + int n; + for (n = 0; n < num_rows; ++n) { + WebPMultRow(ptr, alpha, width, inverse); + ptr += stride; + alpha += alpha_stride; + } +} + +//------------------------------------------------------------------------------ +// Premultiplied modes + +// non dithered-modes + +// (x * a * 32897) >> 23 is bit-wise equivalent to (int)(x * a / 255.) +// for all 8bit x or a. For bit-wise equivalence to (int)(x * a / 255. + .5), +// one can use instead: (x * a * 65793 + (1 << 23)) >> 24 +#if 1 // (int)(x * a / 255.) +#define MULTIPLIER(a) ((a) * 32897U) +#define PREMULTIPLY(x, m) (((x) * (m)) >> 23) +#else // (int)(x * a / 255. + .5) +#define MULTIPLIER(a) ((a) * 65793U) +#define PREMULTIPLY(x, m) (((x) * (m) + (1U << 23)) >> 24) +#endif + +static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_first, + int w, int h, int stride) { + while (h-- > 0) { + uint8_t* const rgb = rgba + (alpha_first ? 1 : 0); + const uint8_t* const alpha = rgba + (alpha_first ? 0 : 3); + int i; + for (i = 0; i < w; ++i) { + const uint32_t a = alpha[4 * i]; + if (a != 0xff) { + const uint32_t mult = MULTIPLIER(a); + rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult); + rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult); + rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult); + } + } + rgba += stride; + } +} +#undef MULTIPLIER +#undef PREMULTIPLY + +// rgbA4444 + +#define MULTIPLIER(a) ((a) * 0x1111) // 0x1111 ~= (1 << 16) / 15 + +static WEBP_INLINE uint8_t dither_hi(uint8_t x) { + return (x & 0xf0) | (x >> 4); +} + +static WEBP_INLINE uint8_t dither_lo(uint8_t x) { + return (x & 0x0f) | (x << 4); +} + +static WEBP_INLINE uint8_t multiply(uint8_t x, uint32_t m) { + return (x * m) >> 16; +} + +static WEBP_INLINE void ApplyAlphaMultiply4444(uint8_t* rgba4444, + int w, int h, int stride, + int rg_byte_pos /* 0 or 1 */) { + while (h-- > 0) { + int i; + for (i = 0; i < w; ++i) { + const uint32_t rg = rgba4444[2 * i + rg_byte_pos]; + const uint32_t ba = rgba4444[2 * i + (rg_byte_pos ^ 1)]; + const uint8_t a = ba & 0x0f; + const uint32_t mult = MULTIPLIER(a); + const uint8_t r = multiply(dither_hi(rg), mult); + const uint8_t g = multiply(dither_lo(rg), mult); + const uint8_t b = multiply(dither_hi(ba), mult); + rgba4444[2 * i + rg_byte_pos] = (r & 0xf0) | ((g >> 4) & 0x0f); + rgba4444[2 * i + (rg_byte_pos ^ 1)] = (b & 0xf0) | a; + } + rgba4444 += stride; + } +} +#undef MULTIPLIER + +static void ApplyAlphaMultiply_16b(uint8_t* rgba4444, + int w, int h, int stride) { +#ifdef WEBP_SWAP_16BIT_CSP + ApplyAlphaMultiply4444(rgba4444, w, h, stride, 1); +#else + ApplyAlphaMultiply4444(rgba4444, w, h, stride, 0); +#endif +} + +static int DispatchAlpha(const uint8_t* alpha, int alpha_stride, + int width, int height, + uint8_t* dst, int dst_stride) { + uint32_t alpha_mask = 0xff; + int i, j; + + for (j = 0; j < height; ++j) { + for (i = 0; i < width; ++i) { + const uint32_t alpha_value = alpha[i]; + dst[4 * i] = alpha_value; + alpha_mask &= alpha_value; + } + alpha += alpha_stride; + dst += dst_stride; + } + + return (alpha_mask != 0xff); +} + +static void DispatchAlphaToGreen(const uint8_t* alpha, int alpha_stride, + int width, int height, + uint32_t* dst, int dst_stride) { + int i, j; + for (j = 0; j < height; ++j) { + for (i = 0; i < width; ++i) { + dst[i] = alpha[i] << 8; // leave A/R/B channels zero'd. + } + alpha += alpha_stride; + dst += dst_stride; + } +} + +static int ExtractAlpha(const uint8_t* argb, int argb_stride, + int width, int height, + uint8_t* alpha, int alpha_stride) { + uint8_t alpha_mask = 0xff; + int i, j; + + for (j = 0; j < height; ++j) { + for (i = 0; i < width; ++i) { + const uint8_t alpha_value = argb[4 * i]; + alpha[i] = alpha_value; + alpha_mask &= alpha_value; + } + argb += argb_stride; + alpha += alpha_stride; + } + return (alpha_mask == 0xff); +} + +void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int); +void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int); +int (*WebPDispatchAlpha)(const uint8_t*, int, int, int, uint8_t*, int); +void (*WebPDispatchAlphaToGreen)(const uint8_t*, int, int, int, uint32_t*, int); +int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int); + +//------------------------------------------------------------------------------ +// Init function + +extern void WebPInitAlphaProcessingMIPSdspR2(void); +extern void WebPInitAlphaProcessingSSE2(void); +extern void WebPInitAlphaProcessingSSE41(void); + +static volatile VP8CPUInfo alpha_processing_last_cpuinfo_used = + (VP8CPUInfo)&alpha_processing_last_cpuinfo_used; + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessing(void) { + if (alpha_processing_last_cpuinfo_used == VP8GetCPUInfo) return; + + WebPMultARGBRow = WebPMultARGBRowC; + WebPMultRow = WebPMultRowC; + WebPApplyAlphaMultiply = ApplyAlphaMultiply; + WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply_16b; + WebPDispatchAlpha = DispatchAlpha; + WebPDispatchAlphaToGreen = DispatchAlphaToGreen; + WebPExtractAlpha = ExtractAlpha; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + WebPInitAlphaProcessingSSE2(); +#if defined(WEBP_USE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + WebPInitAlphaProcessingSSE41(); + } +#endif + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + WebPInitAlphaProcessingMIPSdspR2(); + } +#endif + } + alpha_processing_last_cpuinfo_used = VP8GetCPUInfo; +} diff --git a/drivers/webp/dsp/alpha_processing_mips_dsp_r2.c b/drivers/webp/dsp/alpha_processing_mips_dsp_r2.c new file mode 100644 index 0000000000..c631d78905 --- /dev/null +++ b/drivers/webp/dsp/alpha_processing_mips_dsp_r2.c @@ -0,0 +1,141 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Utilities for processing transparent channel. +// +// Author(s): Branimir Vasic (branimir.vasic@imgtec.com) +// Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +static int DispatchAlpha(const uint8_t* alpha, int alpha_stride, + int width, int height, + uint8_t* dst, int dst_stride) { + uint32_t alpha_mask = 0xffffffff; + int i, j, temp0; + + for (j = 0; j < height; ++j) { + uint8_t* pdst = dst; + const uint8_t* palpha = alpha; + for (i = 0; i < (width >> 2); ++i) { + int temp1, temp2, temp3; + + __asm__ volatile ( + "ulw %[temp0], 0(%[palpha]) \n\t" + "addiu %[palpha], %[palpha], 4 \n\t" + "addiu %[pdst], %[pdst], 16 \n\t" + "srl %[temp1], %[temp0], 8 \n\t" + "srl %[temp2], %[temp0], 16 \n\t" + "srl %[temp3], %[temp0], 24 \n\t" + "and %[alpha_mask], %[alpha_mask], %[temp0] \n\t" + "sb %[temp0], -16(%[pdst]) \n\t" + "sb %[temp1], -12(%[pdst]) \n\t" + "sb %[temp2], -8(%[pdst]) \n\t" + "sb %[temp3], -4(%[pdst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [palpha]"+r"(palpha), [pdst]"+r"(pdst), + [alpha_mask]"+r"(alpha_mask) + : + : "memory" + ); + } + + for (i = 0; i < (width & 3); ++i) { + __asm__ volatile ( + "lbu %[temp0], 0(%[palpha]) \n\t" + "addiu %[palpha], %[palpha], 1 \n\t" + "sb %[temp0], 0(%[pdst]) \n\t" + "and %[alpha_mask], %[alpha_mask], %[temp0] \n\t" + "addiu %[pdst], %[pdst], 4 \n\t" + : [temp0]"=&r"(temp0), [palpha]"+r"(palpha), [pdst]"+r"(pdst), + [alpha_mask]"+r"(alpha_mask) + : + : "memory" + ); + } + alpha += alpha_stride; + dst += dst_stride; + } + + __asm__ volatile ( + "ext %[temp0], %[alpha_mask], 0, 16 \n\t" + "srl %[alpha_mask], %[alpha_mask], 16 \n\t" + "and %[alpha_mask], %[alpha_mask], %[temp0] \n\t" + "ext %[temp0], %[alpha_mask], 0, 8 \n\t" + "srl %[alpha_mask], %[alpha_mask], 8 \n\t" + "and %[alpha_mask], %[alpha_mask], %[temp0] \n\t" + : [temp0]"=&r"(temp0), [alpha_mask]"+r"(alpha_mask) + : + ); + + return (alpha_mask != 0xff); +} + +static void MultARGBRow(uint32_t* const ptr, int width, int inverse) { + int x; + const uint32_t c_00ffffff = 0x00ffffffu; + const uint32_t c_ff000000 = 0xff000000u; + const uint32_t c_8000000 = 0x00800000u; + const uint32_t c_8000080 = 0x00800080u; + for (x = 0; x < width; ++x) { + const uint32_t argb = ptr[x]; + if (argb < 0xff000000u) { // alpha < 255 + if (argb <= 0x00ffffffu) { // alpha == 0 + ptr[x] = 0; + } else { + int temp0, temp1, temp2, temp3, alpha; + __asm__ volatile ( + "srl %[alpha], %[argb], 24 \n\t" + "replv.qb %[temp0], %[alpha] \n\t" + "and %[temp0], %[temp0], %[c_00ffffff] \n\t" + "beqz %[inverse], 0f \n\t" + "divu $zero, %[c_ff000000], %[alpha] \n\t" + "mflo %[temp0] \n\t" + "0: \n\t" + "andi %[temp1], %[argb], 0xff \n\t" + "ext %[temp2], %[argb], 8, 8 \n\t" + "ext %[temp3], %[argb], 16, 8 \n\t" + "mul %[temp1], %[temp1], %[temp0] \n\t" + "mul %[temp2], %[temp2], %[temp0] \n\t" + "mul %[temp3], %[temp3], %[temp0] \n\t" + "precrq.ph.w %[temp1], %[temp2], %[temp1] \n\t" + "addu %[temp3], %[temp3], %[c_8000000] \n\t" + "addu %[temp1], %[temp1], %[c_8000080] \n\t" + "precrq.ph.w %[temp3], %[argb], %[temp3] \n\t" + "precrq.qb.ph %[temp1], %[temp3], %[temp1] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [alpha]"=&r"(alpha) + : [inverse]"r"(inverse), [c_00ffffff]"r"(c_00ffffff), + [c_8000000]"r"(c_8000000), [c_8000080]"r"(c_8000080), + [c_ff000000]"r"(c_ff000000), [argb]"r"(argb) + : "memory", "hi", "lo" + ); + ptr[x] = temp1; + } + } + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitAlphaProcessingMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingMIPSdspR2(void) { + WebPDispatchAlpha = DispatchAlpha; + WebPMultARGBRow = MultARGBRow; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/alpha_processing_sse2.c b/drivers/webp/dsp/alpha_processing_sse2.c new file mode 100644 index 0000000000..5acb481dcd --- /dev/null +++ b/drivers/webp/dsp/alpha_processing_sse2.c @@ -0,0 +1,298 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Utilities for processing transparent channel. +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE2) +#include <emmintrin.h> + +//------------------------------------------------------------------------------ + +static int DispatchAlpha(const uint8_t* alpha, int alpha_stride, + int width, int height, + uint8_t* dst, int dst_stride) { + // alpha_and stores an 'and' operation of all the alpha[] values. The final + // value is not 0xff if any of the alpha[] is not equal to 0xff. + uint32_t alpha_and = 0xff; + int i, j; + const __m128i zero = _mm_setzero_si128(); + const __m128i rgb_mask = _mm_set1_epi32(0xffffff00u); // to preserve RGB + const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u); + __m128i all_alphas = all_0xff; + + // We must be able to access 3 extra bytes after the last written byte + // 'dst[4 * width - 4]', because we don't know if alpha is the first or the + // last byte of the quadruplet. + const int limit = (width - 1) & ~7; + + for (j = 0; j < height; ++j) { + __m128i* out = (__m128i*)dst; + for (i = 0; i < limit; i += 8) { + // load 8 alpha bytes + const __m128i a0 = _mm_loadl_epi64((const __m128i*)&alpha[i]); + const __m128i a1 = _mm_unpacklo_epi8(a0, zero); + const __m128i a2_lo = _mm_unpacklo_epi16(a1, zero); + const __m128i a2_hi = _mm_unpackhi_epi16(a1, zero); + // load 8 dst pixels (32 bytes) + const __m128i b0_lo = _mm_loadu_si128(out + 0); + const __m128i b0_hi = _mm_loadu_si128(out + 1); + // mask dst alpha values + const __m128i b1_lo = _mm_and_si128(b0_lo, rgb_mask); + const __m128i b1_hi = _mm_and_si128(b0_hi, rgb_mask); + // combine + const __m128i b2_lo = _mm_or_si128(b1_lo, a2_lo); + const __m128i b2_hi = _mm_or_si128(b1_hi, a2_hi); + // store + _mm_storeu_si128(out + 0, b2_lo); + _mm_storeu_si128(out + 1, b2_hi); + // accumulate eight alpha 'and' in parallel + all_alphas = _mm_and_si128(all_alphas, a0); + out += 2; + } + for (; i < width; ++i) { + const uint32_t alpha_value = alpha[i]; + dst[4 * i] = alpha_value; + alpha_and &= alpha_value; + } + alpha += alpha_stride; + dst += dst_stride; + } + // Combine the eight alpha 'and' into a 8-bit mask. + alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff)); + return (alpha_and != 0xff); +} + +static void DispatchAlphaToGreen(const uint8_t* alpha, int alpha_stride, + int width, int height, + uint32_t* dst, int dst_stride) { + int i, j; + const __m128i zero = _mm_setzero_si128(); + const int limit = width & ~15; + for (j = 0; j < height; ++j) { + for (i = 0; i < limit; i += 16) { // process 16 alpha bytes + const __m128i a0 = _mm_loadu_si128((const __m128i*)&alpha[i]); + const __m128i a1 = _mm_unpacklo_epi8(zero, a0); // note the 'zero' first! + const __m128i b1 = _mm_unpackhi_epi8(zero, a0); + const __m128i a2_lo = _mm_unpacklo_epi16(a1, zero); + const __m128i b2_lo = _mm_unpacklo_epi16(b1, zero); + const __m128i a2_hi = _mm_unpackhi_epi16(a1, zero); + const __m128i b2_hi = _mm_unpackhi_epi16(b1, zero); + _mm_storeu_si128((__m128i*)&dst[i + 0], a2_lo); + _mm_storeu_si128((__m128i*)&dst[i + 4], a2_hi); + _mm_storeu_si128((__m128i*)&dst[i + 8], b2_lo); + _mm_storeu_si128((__m128i*)&dst[i + 12], b2_hi); + } + for (; i < width; ++i) dst[i] = alpha[i] << 8; + alpha += alpha_stride; + dst += dst_stride; + } +} + +static int ExtractAlpha(const uint8_t* argb, int argb_stride, + int width, int height, + uint8_t* alpha, int alpha_stride) { + // alpha_and stores an 'and' operation of all the alpha[] values. The final + // value is not 0xff if any of the alpha[] is not equal to 0xff. + uint32_t alpha_and = 0xff; + int i, j; + const __m128i a_mask = _mm_set1_epi32(0xffu); // to preserve alpha + const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u); + __m128i all_alphas = all_0xff; + + // We must be able to access 3 extra bytes after the last written byte + // 'src[4 * width - 4]', because we don't know if alpha is the first or the + // last byte of the quadruplet. + const int limit = (width - 1) & ~7; + + for (j = 0; j < height; ++j) { + const __m128i* src = (const __m128i*)argb; + for (i = 0; i < limit; i += 8) { + // load 32 argb bytes + const __m128i a0 = _mm_loadu_si128(src + 0); + const __m128i a1 = _mm_loadu_si128(src + 1); + const __m128i b0 = _mm_and_si128(a0, a_mask); + const __m128i b1 = _mm_and_si128(a1, a_mask); + const __m128i c0 = _mm_packs_epi32(b0, b1); + const __m128i d0 = _mm_packus_epi16(c0, c0); + // store + _mm_storel_epi64((__m128i*)&alpha[i], d0); + // accumulate eight alpha 'and' in parallel + all_alphas = _mm_and_si128(all_alphas, d0); + src += 2; + } + for (; i < width; ++i) { + const uint32_t alpha_value = argb[4 * i]; + alpha[i] = alpha_value; + alpha_and &= alpha_value; + } + argb += argb_stride; + alpha += alpha_stride; + } + // Combine the eight alpha 'and' into a 8-bit mask. + alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff)); + return (alpha_and == 0xff); +} + +//------------------------------------------------------------------------------ +// Non-dither premultiplied modes + +#define MULTIPLIER(a) ((a) * 0x8081) +#define PREMULTIPLY(x, m) (((x) * (m)) >> 23) + +// We can't use a 'const int' for the SHUFFLE value, because it has to be an +// immediate in the _mm_shufflexx_epi16() instruction. We really a macro here. +#define APPLY_ALPHA(RGBX, SHUFFLE, MASK, MULT) do { \ + const __m128i argb0 = _mm_loadl_epi64((__m128i*)&(RGBX)); \ + const __m128i argb1 = _mm_unpacklo_epi8(argb0, zero); \ + const __m128i alpha0 = _mm_and_si128(argb1, MASK); \ + const __m128i alpha1 = _mm_shufflelo_epi16(alpha0, SHUFFLE); \ + const __m128i alpha2 = _mm_shufflehi_epi16(alpha1, SHUFFLE); \ + /* alpha2 = [0 a0 a0 a0][0 a1 a1 a1] */ \ + const __m128i scale0 = _mm_mullo_epi16(alpha2, MULT); \ + const __m128i scale1 = _mm_mulhi_epu16(alpha2, MULT); \ + const __m128i argb2 = _mm_mulhi_epu16(argb1, scale0); \ + const __m128i argb3 = _mm_mullo_epi16(argb1, scale1); \ + const __m128i argb4 = _mm_adds_epu16(argb2, argb3); \ + const __m128i argb5 = _mm_srli_epi16(argb4, 7); \ + const __m128i argb6 = _mm_or_si128(argb5, alpha0); \ + const __m128i argb7 = _mm_packus_epi16(argb6, zero); \ + _mm_storel_epi64((__m128i*)&(RGBX), argb7); \ +} while (0) + +static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_first, + int w, int h, int stride) { + const __m128i zero = _mm_setzero_si128(); + const int kSpan = 2; + const int w2 = w & ~(kSpan - 1); + while (h-- > 0) { + uint32_t* const rgbx = (uint32_t*)rgba; + int i; + if (!alpha_first) { + const __m128i kMask = _mm_set_epi16(0xff, 0, 0, 0, 0xff, 0, 0, 0); + const __m128i kMult = + _mm_set_epi16(0, 0x8081, 0x8081, 0x8081, 0, 0x8081, 0x8081, 0x8081); + for (i = 0; i < w2; i += kSpan) { + APPLY_ALPHA(rgbx[i], _MM_SHUFFLE(0, 3, 3, 3), kMask, kMult); + } + } else { + const __m128i kMask = _mm_set_epi16(0, 0, 0, 0xff, 0, 0, 0, 0xff); + const __m128i kMult = + _mm_set_epi16(0x8081, 0x8081, 0x8081, 0, 0x8081, 0x8081, 0x8081, 0); + for (i = 0; i < w2; i += kSpan) { + APPLY_ALPHA(rgbx[i], _MM_SHUFFLE(0, 0, 0, 3), kMask, kMult); + } + } + // Finish with left-overs. + for (; i < w; ++i) { + uint8_t* const rgb = rgba + (alpha_first ? 1 : 0); + const uint8_t* const alpha = rgba + (alpha_first ? 0 : 3); + const uint32_t a = alpha[4 * i]; + if (a != 0xff) { + const uint32_t mult = MULTIPLIER(a); + rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult); + rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult); + rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult); + } + } + rgba += stride; + } +} +#undef MULTIPLIER +#undef PREMULTIPLY + +// ----------------------------------------------------------------------------- +// Apply alpha value to rows + +// We use: kINV255 = (1 << 24) / 255 = 0x010101 +// So: a * kINV255 = (a << 16) | [(a << 8) | a] +// -> _mm_mulhi_epu16() takes care of the (a<<16) part, +// and _mm_mullo_epu16(a * 0x0101,...) takes care of the "(a << 8) | a" one. + +static void MultARGBRow(uint32_t* const ptr, int width, int inverse) { + int x = 0; + if (!inverse) { + const int kSpan = 2; + const __m128i zero = _mm_setzero_si128(); + const __m128i kRound = + _mm_set_epi16(0, 1 << 7, 1 << 7, 1 << 7, 0, 1 << 7, 1 << 7, 1 << 7); + const __m128i kMult = + _mm_set_epi16(0, 0x0101, 0x0101, 0x0101, 0, 0x0101, 0x0101, 0x0101); + const __m128i kOne64 = _mm_set_epi16(1u << 8, 0, 0, 0, 1u << 8, 0, 0, 0); + const int w2 = width & ~(kSpan - 1); + for (x = 0; x < w2; x += kSpan) { + const __m128i argb0 = _mm_loadl_epi64((__m128i*)&ptr[x]); + const __m128i argb1 = _mm_unpacklo_epi8(argb0, zero); + const __m128i tmp0 = _mm_shufflelo_epi16(argb1, _MM_SHUFFLE(3, 3, 3, 3)); + const __m128i tmp1 = _mm_shufflehi_epi16(tmp0, _MM_SHUFFLE(3, 3, 3, 3)); + const __m128i tmp2 = _mm_srli_epi64(tmp1, 16); + const __m128i scale0 = _mm_mullo_epi16(tmp1, kMult); + const __m128i scale1 = _mm_or_si128(tmp2, kOne64); + const __m128i argb2 = _mm_mulhi_epu16(argb1, scale0); + const __m128i argb3 = _mm_mullo_epi16(argb1, scale1); + const __m128i argb4 = _mm_adds_epu16(argb2, argb3); + const __m128i argb5 = _mm_adds_epu16(argb4, kRound); + const __m128i argb6 = _mm_srli_epi16(argb5, 8); + const __m128i argb7 = _mm_packus_epi16(argb6, zero); + _mm_storel_epi64((__m128i*)&ptr[x], argb7); + } + } + width -= x; + if (width > 0) WebPMultARGBRowC(ptr + x, width, inverse); +} + +static void MultRow(uint8_t* const ptr, const uint8_t* const alpha, + int width, int inverse) { + int x = 0; + if (!inverse) { + const int kSpan = 8; + const __m128i zero = _mm_setzero_si128(); + const __m128i kRound = _mm_set1_epi16(1 << 7); + const int w2 = width & ~(kSpan - 1); + for (x = 0; x < w2; x += kSpan) { + const __m128i v0 = _mm_loadl_epi64((__m128i*)&ptr[x]); + const __m128i v1 = _mm_unpacklo_epi8(v0, zero); + const __m128i alpha0 = _mm_loadl_epi64((const __m128i*)&alpha[x]); + const __m128i alpha1 = _mm_unpacklo_epi8(alpha0, zero); + const __m128i alpha2 = _mm_unpacklo_epi8(alpha0, alpha0); + const __m128i v2 = _mm_mulhi_epu16(v1, alpha2); + const __m128i v3 = _mm_mullo_epi16(v1, alpha1); + const __m128i v4 = _mm_adds_epu16(v2, v3); + const __m128i v5 = _mm_adds_epu16(v4, kRound); + const __m128i v6 = _mm_srli_epi16(v5, 8); + const __m128i v7 = _mm_packus_epi16(v6, zero); + _mm_storel_epi64((__m128i*)&ptr[x], v7); + } + } + width -= x; + if (width > 0) WebPMultRowC(ptr + x, alpha + x, width, inverse); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitAlphaProcessingSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE2(void) { + WebPMultARGBRow = MultARGBRow; + WebPMultRow = MultRow; + WebPApplyAlphaMultiply = ApplyAlphaMultiply; + WebPDispatchAlpha = DispatchAlpha; + WebPDispatchAlphaToGreen = DispatchAlphaToGreen; + WebPExtractAlpha = ExtractAlpha; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/alpha_processing_sse41.c b/drivers/webp/dsp/alpha_processing_sse41.c new file mode 100644 index 0000000000..986fde94ed --- /dev/null +++ b/drivers/webp/dsp/alpha_processing_sse41.c @@ -0,0 +1,92 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Utilities for processing transparent channel, SSE4.1 variant. +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE41) + +#include <smmintrin.h> + +//------------------------------------------------------------------------------ + +static int ExtractAlpha(const uint8_t* argb, int argb_stride, + int width, int height, + uint8_t* alpha, int alpha_stride) { + // alpha_and stores an 'and' operation of all the alpha[] values. The final + // value is not 0xff if any of the alpha[] is not equal to 0xff. + uint32_t alpha_and = 0xff; + int i, j; + const __m128i all_0xff = _mm_set1_epi32(~0u); + __m128i all_alphas = all_0xff; + + // We must be able to access 3 extra bytes after the last written byte + // 'src[4 * width - 4]', because we don't know if alpha is the first or the + // last byte of the quadruplet. + const int limit = (width - 1) & ~15; + const __m128i kCstAlpha0 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, + -1, -1, -1, -1, 12, 8, 4, 0); + const __m128i kCstAlpha1 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, + 12, 8, 4, 0, -1, -1, -1, -1); + const __m128i kCstAlpha2 = _mm_set_epi8(-1, -1, -1, -1, 12, 8, 4, 0, + -1, -1, -1, -1, -1, -1, -1, -1); + const __m128i kCstAlpha3 = _mm_set_epi8(12, 8, 4, 0, -1, -1, -1, -1, + -1, -1, -1, -1, -1, -1, -1, -1); + for (j = 0; j < height; ++j) { + const __m128i* src = (const __m128i*)argb; + for (i = 0; i < limit; i += 16) { + // load 64 argb bytes + const __m128i a0 = _mm_loadu_si128(src + 0); + const __m128i a1 = _mm_loadu_si128(src + 1); + const __m128i a2 = _mm_loadu_si128(src + 2); + const __m128i a3 = _mm_loadu_si128(src + 3); + const __m128i b0 = _mm_shuffle_epi8(a0, kCstAlpha0); + const __m128i b1 = _mm_shuffle_epi8(a1, kCstAlpha1); + const __m128i b2 = _mm_shuffle_epi8(a2, kCstAlpha2); + const __m128i b3 = _mm_shuffle_epi8(a3, kCstAlpha3); + const __m128i c0 = _mm_or_si128(b0, b1); + const __m128i c1 = _mm_or_si128(b2, b3); + const __m128i d0 = _mm_or_si128(c0, c1); + // store + _mm_storeu_si128((__m128i*)&alpha[i], d0); + // accumulate sixteen alpha 'and' in parallel + all_alphas = _mm_and_si128(all_alphas, d0); + src += 4; + } + for (; i < width; ++i) { + const uint32_t alpha_value = argb[4 * i]; + alpha[i] = alpha_value; + alpha_and &= alpha_value; + } + argb += argb_stride; + alpha += alpha_stride; + } + // Combine the sixteen alpha 'and' into an 8-bit mask. + alpha_and |= 0xff00u; // pretend the upper bits [8..15] were tested ok. + alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff)); + return (alpha_and == 0xffffu); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitAlphaProcessingSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE41(void) { + WebPExtractAlpha = ExtractAlpha; +} + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingSSE41) + +#endif // WEBP_USE_SSE41 diff --git a/drivers/webp/dsp/argb.c b/drivers/webp/dsp/argb.c new file mode 100644 index 0000000000..cc1f9a96c3 --- /dev/null +++ b/drivers/webp/dsp/argb.c @@ -0,0 +1,68 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// ARGB making functions. +// +// Author: Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) { + return (((uint32_t)a << 24) | (r << 16) | (g << 8) | b); +} + +static void PackARGB(const uint8_t* a, const uint8_t* r, const uint8_t* g, + const uint8_t* b, int len, uint32_t* out) { + int i; + for (i = 0; i < len; ++i) { + out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]); + } +} + +static void PackRGB(const uint8_t* r, const uint8_t* g, const uint8_t* b, + int len, int step, uint32_t* out) { + int i, offset = 0; + for (i = 0; i < len; ++i) { + out[i] = MakeARGB32(0xff, r[offset], g[offset], b[offset]); + offset += step; + } +} + +void (*VP8PackARGB)(const uint8_t*, const uint8_t*, const uint8_t*, + const uint8_t*, int, uint32_t*); +void (*VP8PackRGB)(const uint8_t*, const uint8_t*, const uint8_t*, + int, int, uint32_t*); + +extern void VP8EncDspARGBInitMIPSdspR2(void); +extern void VP8EncDspARGBInitSSE2(void); + +static volatile VP8CPUInfo argb_last_cpuinfo_used = + (VP8CPUInfo)&argb_last_cpuinfo_used; + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspARGBInit(void) { + if (argb_last_cpuinfo_used == VP8GetCPUInfo) return; + + VP8PackARGB = PackARGB; + VP8PackRGB = PackRGB; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8EncDspARGBInitSSE2(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8EncDspARGBInitMIPSdspR2(); + } +#endif + } + argb_last_cpuinfo_used = VP8GetCPUInfo; +} diff --git a/drivers/webp/dsp/argb_mips_dsp_r2.c b/drivers/webp/dsp/argb_mips_dsp_r2.c new file mode 100644 index 0000000000..af65acb8ff --- /dev/null +++ b/drivers/webp/dsp/argb_mips_dsp_r2.c @@ -0,0 +1,110 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// ARGB making functions (mips version). +// +// Author: Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +static void PackARGB(const uint8_t* a, const uint8_t* r, const uint8_t* g, + const uint8_t* b, int len, uint32_t* out) { + int temp0, temp1, temp2, temp3, offset; + const int rest = len & 1; + const uint32_t* const loop_end = out + len - rest; + const int step = 4; + __asm__ volatile ( + "xor %[offset], %[offset], %[offset] \n\t" + "beq %[loop_end], %[out], 0f \n\t" + "2: \n\t" + "lbux %[temp0], %[offset](%[a]) \n\t" + "lbux %[temp1], %[offset](%[r]) \n\t" + "lbux %[temp2], %[offset](%[g]) \n\t" + "lbux %[temp3], %[offset](%[b]) \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "ins %[temp3], %[temp2], 16, 16 \n\t" + "addiu %[out], %[out], 4 \n\t" + "precr.qb.ph %[temp0], %[temp1], %[temp3] \n\t" + "sw %[temp0], -4(%[out]) \n\t" + "addu %[offset], %[offset], %[step] \n\t" + "bne %[loop_end], %[out], 2b \n\t" + "0: \n\t" + "beq %[rest], $zero, 1f \n\t" + "lbux %[temp0], %[offset](%[a]) \n\t" + "lbux %[temp1], %[offset](%[r]) \n\t" + "lbux %[temp2], %[offset](%[g]) \n\t" + "lbux %[temp3], %[offset](%[b]) \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "ins %[temp3], %[temp2], 16, 16 \n\t" + "precr.qb.ph %[temp0], %[temp1], %[temp3] \n\t" + "sw %[temp0], 0(%[out]) \n\t" + "1: \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [offset]"=&r"(offset), [out]"+&r"(out) + : [a]"r"(a), [r]"r"(r), [g]"r"(g), [b]"r"(b), [step]"r"(step), + [loop_end]"r"(loop_end), [rest]"r"(rest) + : "memory" + ); +} + +static void PackRGB(const uint8_t* r, const uint8_t* g, const uint8_t* b, + int len, int step, uint32_t* out) { + int temp0, temp1, temp2, offset; + const int rest = len & 1; + const int a = 0xff; + const uint32_t* const loop_end = out + len - rest; + __asm__ volatile ( + "xor %[offset], %[offset], %[offset] \n\t" + "beq %[loop_end], %[out], 0f \n\t" + "2: \n\t" + "lbux %[temp0], %[offset](%[r]) \n\t" + "lbux %[temp1], %[offset](%[g]) \n\t" + "lbux %[temp2], %[offset](%[b]) \n\t" + "ins %[temp0], %[a], 16, 16 \n\t" + "ins %[temp2], %[temp1], 16, 16 \n\t" + "addiu %[out], %[out], 4 \n\t" + "precr.qb.ph %[temp0], %[temp0], %[temp2] \n\t" + "sw %[temp0], -4(%[out]) \n\t" + "addu %[offset], %[offset], %[step] \n\t" + "bne %[loop_end], %[out], 2b \n\t" + "0: \n\t" + "beq %[rest], $zero, 1f \n\t" + "lbux %[temp0], %[offset](%[r]) \n\t" + "lbux %[temp1], %[offset](%[g]) \n\t" + "lbux %[temp2], %[offset](%[b]) \n\t" + "ins %[temp0], %[a], 16, 16 \n\t" + "ins %[temp2], %[temp1], 16, 16 \n\t" + "precr.qb.ph %[temp0], %[temp0], %[temp2] \n\t" + "sw %[temp0], 0(%[out]) \n\t" + "1: \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [offset]"=&r"(offset), [out]"+&r"(out) + : [a]"r"(a), [r]"r"(r), [g]"r"(g), [b]"r"(b), [step]"r"(step), + [loop_end]"r"(loop_end), [rest]"r"(rest) + : "memory" + ); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspARGBInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspARGBInitMIPSdspR2(void) { + VP8PackARGB = PackARGB; + VP8PackRGB = PackRGB; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8EncDspARGBInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/argb_sse2.c b/drivers/webp/dsp/argb_sse2.c new file mode 100644 index 0000000000..afcb1957e7 --- /dev/null +++ b/drivers/webp/dsp/argb_sse2.c @@ -0,0 +1,67 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// ARGB making functions (SSE2 version). +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE2) + +#include <assert.h> +#include <emmintrin.h> +#include <string.h> + +static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) { + return (((uint32_t)a << 24) | (r << 16) | (g << 8) | b); +} + +static void PackARGB(const uint8_t* a, const uint8_t* r, const uint8_t* g, + const uint8_t* b, int len, uint32_t* out) { + if (g == r + 1) { // RGBA input order. Need to swap R and B. + int i = 0; + const int len_max = len & ~3; // max length processed in main loop + const __m128i red_blue_mask = _mm_set1_epi32(0x00ff00ffu); + assert(b == r + 2); + assert(a == r + 3); + for (; i < len_max; i += 4) { + const __m128i A = _mm_loadu_si128((const __m128i*)(r + 4 * i)); + const __m128i B = _mm_and_si128(A, red_blue_mask); // R 0 B 0 + const __m128i C = _mm_andnot_si128(red_blue_mask, A); // 0 G 0 A + const __m128i D = _mm_shufflelo_epi16(B, _MM_SHUFFLE(2, 3, 0, 1)); + const __m128i E = _mm_shufflehi_epi16(D, _MM_SHUFFLE(2, 3, 0, 1)); + const __m128i F = _mm_or_si128(E, C); + _mm_storeu_si128((__m128i*)(out + i), F); + } + for (; i < len; ++i) { + out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]); + } + } else { + assert(g == b + 1); + assert(r == b + 2); + assert(a == b + 3); + memcpy(out, b, len * 4); + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspARGBInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspARGBInitSSE2(void) { + VP8PackARGB = PackARGB; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8EncDspARGBInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/cost.c b/drivers/webp/dsp/cost.c new file mode 100644 index 0000000000..fe72d26e79 --- /dev/null +++ b/drivers/webp/dsp/cost.c @@ -0,0 +1,412 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" +#include "../enc/cost.h" + +//------------------------------------------------------------------------------ +// Boolean-cost cost table + +const uint16_t VP8EntropyCost[256] = { + 1792, 1792, 1792, 1536, 1536, 1408, 1366, 1280, 1280, 1216, + 1178, 1152, 1110, 1076, 1061, 1024, 1024, 992, 968, 951, + 939, 911, 896, 878, 871, 854, 838, 820, 811, 794, + 786, 768, 768, 752, 740, 732, 720, 709, 704, 690, + 683, 672, 666, 655, 647, 640, 631, 622, 615, 607, + 598, 592, 586, 576, 572, 564, 559, 555, 547, 541, + 534, 528, 522, 512, 512, 504, 500, 494, 488, 483, + 477, 473, 467, 461, 458, 452, 448, 443, 438, 434, + 427, 424, 419, 415, 410, 406, 403, 399, 394, 390, + 384, 384, 377, 374, 370, 366, 362, 359, 355, 351, + 347, 342, 342, 336, 333, 330, 326, 323, 320, 316, + 312, 308, 305, 302, 299, 296, 293, 288, 287, 283, + 280, 277, 274, 272, 268, 266, 262, 256, 256, 256, + 251, 248, 245, 242, 240, 237, 234, 232, 228, 226, + 223, 221, 218, 216, 214, 211, 208, 205, 203, 201, + 198, 196, 192, 191, 188, 187, 183, 181, 179, 176, + 175, 171, 171, 168, 165, 163, 160, 159, 156, 154, + 152, 150, 148, 146, 144, 142, 139, 138, 135, 133, + 131, 128, 128, 125, 123, 121, 119, 117, 115, 113, + 111, 110, 107, 105, 103, 102, 100, 98, 96, 94, + 92, 91, 89, 86, 86, 83, 82, 80, 77, 76, + 74, 73, 71, 69, 67, 66, 64, 63, 61, 59, + 57, 55, 54, 52, 51, 49, 47, 46, 44, 43, + 41, 40, 38, 36, 35, 33, 32, 30, 29, 27, + 25, 24, 22, 21, 19, 18, 16, 15, 13, 12, + 10, 9, 7, 6, 4, 3 +}; + +//------------------------------------------------------------------------------ +// Level cost tables + +// fixed costs for coding levels, deduce from the coding tree. +// This is only the part that doesn't depend on the probability state. +const uint16_t VP8LevelFixedCosts[MAX_LEVEL + 1] = { + 0, 256, 256, 256, 256, 432, 618, 630, + 731, 640, 640, 828, 901, 948, 1021, 1101, + 1174, 1221, 1294, 1042, 1085, 1115, 1158, 1202, + 1245, 1275, 1318, 1337, 1380, 1410, 1453, 1497, + 1540, 1570, 1613, 1280, 1295, 1317, 1332, 1358, + 1373, 1395, 1410, 1454, 1469, 1491, 1506, 1532, + 1547, 1569, 1584, 1601, 1616, 1638, 1653, 1679, + 1694, 1716, 1731, 1775, 1790, 1812, 1827, 1853, + 1868, 1890, 1905, 1727, 1733, 1742, 1748, 1759, + 1765, 1774, 1780, 1800, 1806, 1815, 1821, 1832, + 1838, 1847, 1853, 1878, 1884, 1893, 1899, 1910, + 1916, 1925, 1931, 1951, 1957, 1966, 1972, 1983, + 1989, 1998, 2004, 2027, 2033, 2042, 2048, 2059, + 2065, 2074, 2080, 2100, 2106, 2115, 2121, 2132, + 2138, 2147, 2153, 2178, 2184, 2193, 2199, 2210, + 2216, 2225, 2231, 2251, 2257, 2266, 2272, 2283, + 2289, 2298, 2304, 2168, 2174, 2183, 2189, 2200, + 2206, 2215, 2221, 2241, 2247, 2256, 2262, 2273, + 2279, 2288, 2294, 2319, 2325, 2334, 2340, 2351, + 2357, 2366, 2372, 2392, 2398, 2407, 2413, 2424, + 2430, 2439, 2445, 2468, 2474, 2483, 2489, 2500, + 2506, 2515, 2521, 2541, 2547, 2556, 2562, 2573, + 2579, 2588, 2594, 2619, 2625, 2634, 2640, 2651, + 2657, 2666, 2672, 2692, 2698, 2707, 2713, 2724, + 2730, 2739, 2745, 2540, 2546, 2555, 2561, 2572, + 2578, 2587, 2593, 2613, 2619, 2628, 2634, 2645, + 2651, 2660, 2666, 2691, 2697, 2706, 2712, 2723, + 2729, 2738, 2744, 2764, 2770, 2779, 2785, 2796, + 2802, 2811, 2817, 2840, 2846, 2855, 2861, 2872, + 2878, 2887, 2893, 2913, 2919, 2928, 2934, 2945, + 2951, 2960, 2966, 2991, 2997, 3006, 3012, 3023, + 3029, 3038, 3044, 3064, 3070, 3079, 3085, 3096, + 3102, 3111, 3117, 2981, 2987, 2996, 3002, 3013, + 3019, 3028, 3034, 3054, 3060, 3069, 3075, 3086, + 3092, 3101, 3107, 3132, 3138, 3147, 3153, 3164, + 3170, 3179, 3185, 3205, 3211, 3220, 3226, 3237, + 3243, 3252, 3258, 3281, 3287, 3296, 3302, 3313, + 3319, 3328, 3334, 3354, 3360, 3369, 3375, 3386, + 3392, 3401, 3407, 3432, 3438, 3447, 3453, 3464, + 3470, 3479, 3485, 3505, 3511, 3520, 3526, 3537, + 3543, 3552, 3558, 2816, 2822, 2831, 2837, 2848, + 2854, 2863, 2869, 2889, 2895, 2904, 2910, 2921, + 2927, 2936, 2942, 2967, 2973, 2982, 2988, 2999, + 3005, 3014, 3020, 3040, 3046, 3055, 3061, 3072, + 3078, 3087, 3093, 3116, 3122, 3131, 3137, 3148, + 3154, 3163, 3169, 3189, 3195, 3204, 3210, 3221, + 3227, 3236, 3242, 3267, 3273, 3282, 3288, 3299, + 3305, 3314, 3320, 3340, 3346, 3355, 3361, 3372, + 3378, 3387, 3393, 3257, 3263, 3272, 3278, 3289, + 3295, 3304, 3310, 3330, 3336, 3345, 3351, 3362, + 3368, 3377, 3383, 3408, 3414, 3423, 3429, 3440, + 3446, 3455, 3461, 3481, 3487, 3496, 3502, 3513, + 3519, 3528, 3534, 3557, 3563, 3572, 3578, 3589, + 3595, 3604, 3610, 3630, 3636, 3645, 3651, 3662, + 3668, 3677, 3683, 3708, 3714, 3723, 3729, 3740, + 3746, 3755, 3761, 3781, 3787, 3796, 3802, 3813, + 3819, 3828, 3834, 3629, 3635, 3644, 3650, 3661, + 3667, 3676, 3682, 3702, 3708, 3717, 3723, 3734, + 3740, 3749, 3755, 3780, 3786, 3795, 3801, 3812, + 3818, 3827, 3833, 3853, 3859, 3868, 3874, 3885, + 3891, 3900, 3906, 3929, 3935, 3944, 3950, 3961, + 3967, 3976, 3982, 4002, 4008, 4017, 4023, 4034, + 4040, 4049, 4055, 4080, 4086, 4095, 4101, 4112, + 4118, 4127, 4133, 4153, 4159, 4168, 4174, 4185, + 4191, 4200, 4206, 4070, 4076, 4085, 4091, 4102, + 4108, 4117, 4123, 4143, 4149, 4158, 4164, 4175, + 4181, 4190, 4196, 4221, 4227, 4236, 4242, 4253, + 4259, 4268, 4274, 4294, 4300, 4309, 4315, 4326, + 4332, 4341, 4347, 4370, 4376, 4385, 4391, 4402, + 4408, 4417, 4423, 4443, 4449, 4458, 4464, 4475, + 4481, 4490, 4496, 4521, 4527, 4536, 4542, 4553, + 4559, 4568, 4574, 4594, 4600, 4609, 4615, 4626, + 4632, 4641, 4647, 3515, 3521, 3530, 3536, 3547, + 3553, 3562, 3568, 3588, 3594, 3603, 3609, 3620, + 3626, 3635, 3641, 3666, 3672, 3681, 3687, 3698, + 3704, 3713, 3719, 3739, 3745, 3754, 3760, 3771, + 3777, 3786, 3792, 3815, 3821, 3830, 3836, 3847, + 3853, 3862, 3868, 3888, 3894, 3903, 3909, 3920, + 3926, 3935, 3941, 3966, 3972, 3981, 3987, 3998, + 4004, 4013, 4019, 4039, 4045, 4054, 4060, 4071, + 4077, 4086, 4092, 3956, 3962, 3971, 3977, 3988, + 3994, 4003, 4009, 4029, 4035, 4044, 4050, 4061, + 4067, 4076, 4082, 4107, 4113, 4122, 4128, 4139, + 4145, 4154, 4160, 4180, 4186, 4195, 4201, 4212, + 4218, 4227, 4233, 4256, 4262, 4271, 4277, 4288, + 4294, 4303, 4309, 4329, 4335, 4344, 4350, 4361, + 4367, 4376, 4382, 4407, 4413, 4422, 4428, 4439, + 4445, 4454, 4460, 4480, 4486, 4495, 4501, 4512, + 4518, 4527, 4533, 4328, 4334, 4343, 4349, 4360, + 4366, 4375, 4381, 4401, 4407, 4416, 4422, 4433, + 4439, 4448, 4454, 4479, 4485, 4494, 4500, 4511, + 4517, 4526, 4532, 4552, 4558, 4567, 4573, 4584, + 4590, 4599, 4605, 4628, 4634, 4643, 4649, 4660, + 4666, 4675, 4681, 4701, 4707, 4716, 4722, 4733, + 4739, 4748, 4754, 4779, 4785, 4794, 4800, 4811, + 4817, 4826, 4832, 4852, 4858, 4867, 4873, 4884, + 4890, 4899, 4905, 4769, 4775, 4784, 4790, 4801, + 4807, 4816, 4822, 4842, 4848, 4857, 4863, 4874, + 4880, 4889, 4895, 4920, 4926, 4935, 4941, 4952, + 4958, 4967, 4973, 4993, 4999, 5008, 5014, 5025, + 5031, 5040, 5046, 5069, 5075, 5084, 5090, 5101, + 5107, 5116, 5122, 5142, 5148, 5157, 5163, 5174, + 5180, 5189, 5195, 5220, 5226, 5235, 5241, 5252, + 5258, 5267, 5273, 5293, 5299, 5308, 5314, 5325, + 5331, 5340, 5346, 4604, 4610, 4619, 4625, 4636, + 4642, 4651, 4657, 4677, 4683, 4692, 4698, 4709, + 4715, 4724, 4730, 4755, 4761, 4770, 4776, 4787, + 4793, 4802, 4808, 4828, 4834, 4843, 4849, 4860, + 4866, 4875, 4881, 4904, 4910, 4919, 4925, 4936, + 4942, 4951, 4957, 4977, 4983, 4992, 4998, 5009, + 5015, 5024, 5030, 5055, 5061, 5070, 5076, 5087, + 5093, 5102, 5108, 5128, 5134, 5143, 5149, 5160, + 5166, 5175, 5181, 5045, 5051, 5060, 5066, 5077, + 5083, 5092, 5098, 5118, 5124, 5133, 5139, 5150, + 5156, 5165, 5171, 5196, 5202, 5211, 5217, 5228, + 5234, 5243, 5249, 5269, 5275, 5284, 5290, 5301, + 5307, 5316, 5322, 5345, 5351, 5360, 5366, 5377, + 5383, 5392, 5398, 5418, 5424, 5433, 5439, 5450, + 5456, 5465, 5471, 5496, 5502, 5511, 5517, 5528, + 5534, 5543, 5549, 5569, 5575, 5584, 5590, 5601, + 5607, 5616, 5622, 5417, 5423, 5432, 5438, 5449, + 5455, 5464, 5470, 5490, 5496, 5505, 5511, 5522, + 5528, 5537, 5543, 5568, 5574, 5583, 5589, 5600, + 5606, 5615, 5621, 5641, 5647, 5656, 5662, 5673, + 5679, 5688, 5694, 5717, 5723, 5732, 5738, 5749, + 5755, 5764, 5770, 5790, 5796, 5805, 5811, 5822, + 5828, 5837, 5843, 5868, 5874, 5883, 5889, 5900, + 5906, 5915, 5921, 5941, 5947, 5956, 5962, 5973, + 5979, 5988, 5994, 5858, 5864, 5873, 5879, 5890, + 5896, 5905, 5911, 5931, 5937, 5946, 5952, 5963, + 5969, 5978, 5984, 6009, 6015, 6024, 6030, 6041, + 6047, 6056, 6062, 6082, 6088, 6097, 6103, 6114, + 6120, 6129, 6135, 6158, 6164, 6173, 6179, 6190, + 6196, 6205, 6211, 6231, 6237, 6246, 6252, 6263, + 6269, 6278, 6284, 6309, 6315, 6324, 6330, 6341, + 6347, 6356, 6362, 6382, 6388, 6397, 6403, 6414, + 6420, 6429, 6435, 3515, 3521, 3530, 3536, 3547, + 3553, 3562, 3568, 3588, 3594, 3603, 3609, 3620, + 3626, 3635, 3641, 3666, 3672, 3681, 3687, 3698, + 3704, 3713, 3719, 3739, 3745, 3754, 3760, 3771, + 3777, 3786, 3792, 3815, 3821, 3830, 3836, 3847, + 3853, 3862, 3868, 3888, 3894, 3903, 3909, 3920, + 3926, 3935, 3941, 3966, 3972, 3981, 3987, 3998, + 4004, 4013, 4019, 4039, 4045, 4054, 4060, 4071, + 4077, 4086, 4092, 3956, 3962, 3971, 3977, 3988, + 3994, 4003, 4009, 4029, 4035, 4044, 4050, 4061, + 4067, 4076, 4082, 4107, 4113, 4122, 4128, 4139, + 4145, 4154, 4160, 4180, 4186, 4195, 4201, 4212, + 4218, 4227, 4233, 4256, 4262, 4271, 4277, 4288, + 4294, 4303, 4309, 4329, 4335, 4344, 4350, 4361, + 4367, 4376, 4382, 4407, 4413, 4422, 4428, 4439, + 4445, 4454, 4460, 4480, 4486, 4495, 4501, 4512, + 4518, 4527, 4533, 4328, 4334, 4343, 4349, 4360, + 4366, 4375, 4381, 4401, 4407, 4416, 4422, 4433, + 4439, 4448, 4454, 4479, 4485, 4494, 4500, 4511, + 4517, 4526, 4532, 4552, 4558, 4567, 4573, 4584, + 4590, 4599, 4605, 4628, 4634, 4643, 4649, 4660, + 4666, 4675, 4681, 4701, 4707, 4716, 4722, 4733, + 4739, 4748, 4754, 4779, 4785, 4794, 4800, 4811, + 4817, 4826, 4832, 4852, 4858, 4867, 4873, 4884, + 4890, 4899, 4905, 4769, 4775, 4784, 4790, 4801, + 4807, 4816, 4822, 4842, 4848, 4857, 4863, 4874, + 4880, 4889, 4895, 4920, 4926, 4935, 4941, 4952, + 4958, 4967, 4973, 4993, 4999, 5008, 5014, 5025, + 5031, 5040, 5046, 5069, 5075, 5084, 5090, 5101, + 5107, 5116, 5122, 5142, 5148, 5157, 5163, 5174, + 5180, 5189, 5195, 5220, 5226, 5235, 5241, 5252, + 5258, 5267, 5273, 5293, 5299, 5308, 5314, 5325, + 5331, 5340, 5346, 4604, 4610, 4619, 4625, 4636, + 4642, 4651, 4657, 4677, 4683, 4692, 4698, 4709, + 4715, 4724, 4730, 4755, 4761, 4770, 4776, 4787, + 4793, 4802, 4808, 4828, 4834, 4843, 4849, 4860, + 4866, 4875, 4881, 4904, 4910, 4919, 4925, 4936, + 4942, 4951, 4957, 4977, 4983, 4992, 4998, 5009, + 5015, 5024, 5030, 5055, 5061, 5070, 5076, 5087, + 5093, 5102, 5108, 5128, 5134, 5143, 5149, 5160, + 5166, 5175, 5181, 5045, 5051, 5060, 5066, 5077, + 5083, 5092, 5098, 5118, 5124, 5133, 5139, 5150, + 5156, 5165, 5171, 5196, 5202, 5211, 5217, 5228, + 5234, 5243, 5249, 5269, 5275, 5284, 5290, 5301, + 5307, 5316, 5322, 5345, 5351, 5360, 5366, 5377, + 5383, 5392, 5398, 5418, 5424, 5433, 5439, 5450, + 5456, 5465, 5471, 5496, 5502, 5511, 5517, 5528, + 5534, 5543, 5549, 5569, 5575, 5584, 5590, 5601, + 5607, 5616, 5622, 5417, 5423, 5432, 5438, 5449, + 5455, 5464, 5470, 5490, 5496, 5505, 5511, 5522, + 5528, 5537, 5543, 5568, 5574, 5583, 5589, 5600, + 5606, 5615, 5621, 5641, 5647, 5656, 5662, 5673, + 5679, 5688, 5694, 5717, 5723, 5732, 5738, 5749, + 5755, 5764, 5770, 5790, 5796, 5805, 5811, 5822, + 5828, 5837, 5843, 5868, 5874, 5883, 5889, 5900, + 5906, 5915, 5921, 5941, 5947, 5956, 5962, 5973, + 5979, 5988, 5994, 5858, 5864, 5873, 5879, 5890, + 5896, 5905, 5911, 5931, 5937, 5946, 5952, 5963, + 5969, 5978, 5984, 6009, 6015, 6024, 6030, 6041, + 6047, 6056, 6062, 6082, 6088, 6097, 6103, 6114, + 6120, 6129, 6135, 6158, 6164, 6173, 6179, 6190, + 6196, 6205, 6211, 6231, 6237, 6246, 6252, 6263, + 6269, 6278, 6284, 6309, 6315, 6324, 6330, 6341, + 6347, 6356, 6362, 6382, 6388, 6397, 6403, 6414, + 6420, 6429, 6435, 5303, 5309, 5318, 5324, 5335, + 5341, 5350, 5356, 5376, 5382, 5391, 5397, 5408, + 5414, 5423, 5429, 5454, 5460, 5469, 5475, 5486, + 5492, 5501, 5507, 5527, 5533, 5542, 5548, 5559, + 5565, 5574, 5580, 5603, 5609, 5618, 5624, 5635, + 5641, 5650, 5656, 5676, 5682, 5691, 5697, 5708, + 5714, 5723, 5729, 5754, 5760, 5769, 5775, 5786, + 5792, 5801, 5807, 5827, 5833, 5842, 5848, 5859, + 5865, 5874, 5880, 5744, 5750, 5759, 5765, 5776, + 5782, 5791, 5797, 5817, 5823, 5832, 5838, 5849, + 5855, 5864, 5870, 5895, 5901, 5910, 5916, 5927, + 5933, 5942, 5948, 5968, 5974, 5983, 5989, 6000, + 6006, 6015, 6021, 6044, 6050, 6059, 6065, 6076, + 6082, 6091, 6097, 6117, 6123, 6132, 6138, 6149, + 6155, 6164, 6170, 6195, 6201, 6210, 6216, 6227, + 6233, 6242, 6248, 6268, 6274, 6283, 6289, 6300, + 6306, 6315, 6321, 6116, 6122, 6131, 6137, 6148, + 6154, 6163, 6169, 6189, 6195, 6204, 6210, 6221, + 6227, 6236, 6242, 6267, 6273, 6282, 6288, 6299, + 6305, 6314, 6320, 6340, 6346, 6355, 6361, 6372, + 6378, 6387, 6393, 6416, 6422, 6431, 6437, 6448, + 6454, 6463, 6469, 6489, 6495, 6504, 6510, 6521, + 6527, 6536, 6542, 6567, 6573, 6582, 6588, 6599, + 6605, 6614, 6620, 6640, 6646, 6655, 6661, 6672, + 6678, 6687, 6693, 6557, 6563, 6572, 6578, 6589, + 6595, 6604, 6610, 6630, 6636, 6645, 6651, 6662, + 6668, 6677, 6683, 6708, 6714, 6723, 6729, 6740, + 6746, 6755, 6761, 6781, 6787, 6796, 6802, 6813, + 6819, 6828, 6834, 6857, 6863, 6872, 6878, 6889, + 6895, 6904, 6910, 6930, 6936, 6945, 6951, 6962, + 6968, 6977, 6983, 7008, 7014, 7023, 7029, 7040, + 7046, 7055, 7061, 7081, 7087, 7096, 7102, 7113, + 7119, 7128, 7134, 6392, 6398, 6407, 6413, 6424, + 6430, 6439, 6445, 6465, 6471, 6480, 6486, 6497, + 6503, 6512, 6518, 6543, 6549, 6558, 6564, 6575, + 6581, 6590, 6596, 6616, 6622, 6631, 6637, 6648, + 6654, 6663, 6669, 6692, 6698, 6707, 6713, 6724, + 6730, 6739, 6745, 6765, 6771, 6780, 6786, 6797, + 6803, 6812, 6818, 6843, 6849, 6858, 6864, 6875, + 6881, 6890, 6896, 6916, 6922, 6931, 6937, 6948, + 6954, 6963, 6969, 6833, 6839, 6848, 6854, 6865, + 6871, 6880, 6886, 6906, 6912, 6921, 6927, 6938, + 6944, 6953, 6959, 6984, 6990, 6999, 7005, 7016, + 7022, 7031, 7037, 7057, 7063, 7072, 7078, 7089, + 7095, 7104, 7110, 7133, 7139, 7148, 7154, 7165, + 7171, 7180, 7186, 7206, 7212, 7221, 7227, 7238, + 7244, 7253, 7259, 7284, 7290, 7299, 7305, 7316, + 7322, 7331, 7337, 7357, 7363, 7372, 7378, 7389, + 7395, 7404, 7410, 7205, 7211, 7220, 7226, 7237, + 7243, 7252, 7258, 7278, 7284, 7293, 7299, 7310, + 7316, 7325, 7331, 7356, 7362, 7371, 7377, 7388, + 7394, 7403, 7409, 7429, 7435, 7444, 7450, 7461, + 7467, 7476, 7482, 7505, 7511, 7520, 7526, 7537, + 7543, 7552, 7558, 7578, 7584, 7593, 7599, 7610, + 7616, 7625, 7631, 7656, 7662, 7671, 7677, 7688, + 7694, 7703, 7709, 7729, 7735, 7744, 7750, 7761 +}; + +//------------------------------------------------------------------------------ +// Tables for level coding + +const uint8_t VP8EncBands[16 + 1] = { + 0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7, + 0 // sentinel +}; + +//------------------------------------------------------------------------------ +// Mode costs + +static int GetResidualCost(int ctx0, const VP8Residual* const res) { + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + const int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + for (; n < res->last; ++n) { + const int v = abs(res->coeffs[n]); + const int ctx = (v >= 2) ? 2 : v; + cost += VP8LevelCost(t, v); + t = costs[n + 1][ctx]; + } + // Last coefficient is always non-zero + { + const int v = abs(res->coeffs[n]); + assert(v != 0); + cost += VP8LevelCost(t, v); + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = (v == 1) ? 1 : 2; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +static void SetResidualCoeffs(const int16_t* const coeffs, + VP8Residual* const res) { + int n; + res->last = -1; + assert(res->first == 0 || coeffs[0] == 0); + for (n = 15; n >= 0; --n) { + if (coeffs[n]) { + res->last = n; + break; + } + } + res->coeffs = coeffs; +} + +//------------------------------------------------------------------------------ +// init function + +VP8GetResidualCostFunc VP8GetResidualCost; +VP8SetResidualCoeffsFunc VP8SetResidualCoeffs; + +extern void VP8EncDspCostInitMIPS32(void); +extern void VP8EncDspCostInitMIPSdspR2(void); +extern void VP8EncDspCostInitSSE2(void); + +static volatile VP8CPUInfo cost_last_cpuinfo_used = + (VP8CPUInfo)&cost_last_cpuinfo_used; + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInit(void) { + if (cost_last_cpuinfo_used == VP8GetCPUInfo) return; + + VP8GetResidualCost = GetResidualCost; + VP8SetResidualCoeffs = SetResidualCoeffs; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + VP8EncDspCostInitMIPS32(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8EncDspCostInitMIPSdspR2(); + } +#endif +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8EncDspCostInitSSE2(); + } +#endif + } + + cost_last_cpuinfo_used = VP8GetCPUInfo; +} + +//------------------------------------------------------------------------------ diff --git a/drivers/webp/dsp/cost_mips32.c b/drivers/webp/dsp/cost_mips32.c new file mode 100644 index 0000000000..d1e240e191 --- /dev/null +++ b/drivers/webp/dsp/cost_mips32.c @@ -0,0 +1,154 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Author: Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include "../enc/cost.h" + +static int GetResidualCost(int ctx0, const VP8Residual* const res) { + int temp0, temp1; + int v_reg, ctx_reg; + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + const int16_t* res_coeffs = res->coeffs; + const int res_last = res->last; + const int const_max_level = MAX_VARIABLE_LEVEL; + const int const_2 = 2; + const uint16_t** p_costs = &costs[n][0]; + const size_t inc_p_costs = NUM_CTX * sizeof(*p_costs); + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "subu %[temp1], %[res_last], %[n] \n\t" + "sll %[temp0], %[n], 1 \n\t" + "blez %[temp1], 2f \n\t" + " addu %[res_coeffs], %[res_coeffs], %[temp0] \n\t" + "1: \n\t" + "lh %[v_reg], 0(%[res_coeffs]) \n\t" + "addiu %[n], %[n], 1 \n\t" + "negu %[temp0], %[v_reg] \n\t" + "slti %[temp1], %[v_reg], 0 \n\t" + "movn %[v_reg], %[temp0], %[temp1] \n\t" + "sltiu %[temp0], %[v_reg], 2 \n\t" + "move %[ctx_reg], %[v_reg] \n\t" + "movz %[ctx_reg], %[const_2], %[temp0] \n\t" + "sll %[temp1], %[v_reg], 1 \n\t" + "addu %[temp1], %[temp1], %[VP8LevelFixedCosts] \n\t" + "lhu %[temp1], 0(%[temp1]) \n\t" + "slt %[temp0], %[v_reg], %[const_max_level] \n\t" + "movz %[v_reg], %[const_max_level], %[temp0] \n\t" + "addu %[cost], %[cost], %[temp1] \n\t" + "sll %[v_reg], %[v_reg], 1 \n\t" + "sll %[ctx_reg], %[ctx_reg], 2 \n\t" + "addu %[v_reg], %[v_reg], %[t] \n\t" + "lhu %[temp0], 0(%[v_reg]) \n\t" + "addu %[p_costs], %[p_costs], %[inc_p_costs] \n\t" + "addu %[t], %[p_costs], %[ctx_reg] \n\t" + "addu %[cost], %[cost], %[temp0] \n\t" + "addiu %[res_coeffs], %[res_coeffs], 2 \n\t" + "bne %[n], %[res_last], 1b \n\t" + " lw %[t], 0(%[t]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [cost]"+&r"(cost), [t]"+&r"(t), [n]"+&r"(n), [v_reg]"=&r"(v_reg), + [ctx_reg]"=&r"(ctx_reg), [p_costs]"+&r"(p_costs), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1), [res_coeffs]"+&r"(res_coeffs) + : [const_2]"r"(const_2), [const_max_level]"r"(const_max_level), + [VP8LevelFixedCosts]"r"(VP8LevelFixedCosts), [res_last]"r"(res_last), + [inc_p_costs]"r"(inc_p_costs) + : "memory" + ); + + // Last coefficient is always non-zero + { + const int v = abs(res->coeffs[n]); + assert(v != 0); + cost += VP8LevelCost(t, v); + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = (v == 1) ? 1 : 2; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +static void SetResidualCoeffs(const int16_t* const coeffs, + VP8Residual* const res) { + const int16_t* p_coeffs = (int16_t*)coeffs; + int temp0, temp1, temp2, n, n1; + assert(res->first == 0 || coeffs[0] == 0); + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "addiu %[p_coeffs], %[p_coeffs], 28 \n\t" + "li %[n], 15 \n\t" + "li %[temp2], -1 \n\t" + "0: \n\t" + "ulw %[temp0], 0(%[p_coeffs]) \n\t" + "beqz %[temp0], 1f \n\t" +#if defined(WORDS_BIGENDIAN) + " sll %[temp1], %[temp0], 16 \n\t" +#else + " srl %[temp1], %[temp0], 16 \n\t" +#endif + "addiu %[n1], %[n], -1 \n\t" + "movz %[temp0], %[n1], %[temp1] \n\t" + "movn %[temp0], %[n], %[temp1] \n\t" + "j 2f \n\t" + " addiu %[temp2], %[temp0], 0 \n\t" + "1: \n\t" + "addiu %[n], %[n], -2 \n\t" + "bgtz %[n], 0b \n\t" + " addiu %[p_coeffs], %[p_coeffs], -4 \n\t" + "2: \n\t" + ".set pop \n\t" + : [p_coeffs]"+&r"(p_coeffs), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [n]"=&r"(n), [n1]"=&r"(n1) + : + : "memory" + ); + res->last = temp2; + res->coeffs = coeffs; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspCostInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitMIPS32(void) { + VP8GetResidualCost = GetResidualCost; + VP8SetResidualCoeffs = SetResidualCoeffs; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(VP8EncDspCostInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/drivers/webp/dsp/cost_mips_dsp_r2.c b/drivers/webp/dsp/cost_mips_dsp_r2.c new file mode 100644 index 0000000000..ce64067756 --- /dev/null +++ b/drivers/webp/dsp/cost_mips_dsp_r2.c @@ -0,0 +1,107 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Author: Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../enc/cost.h" + +static int GetResidualCost(int ctx0, const VP8Residual* const res) { + int temp0, temp1; + int v_reg, ctx_reg; + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + const int16_t* res_coeffs = res->coeffs; + const int res_last = res->last; + const int const_max_level = MAX_VARIABLE_LEVEL; + const int const_2 = 2; + const uint16_t** p_costs = &costs[n][0]; + const size_t inc_p_costs = NUM_CTX * sizeof(*p_costs); + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "subu %[temp1], %[res_last], %[n] \n\t" + "blez %[temp1], 2f \n\t" + " nop \n\t" + "1: \n\t" + "sll %[temp0], %[n], 1 \n\t" + "lhx %[v_reg], %[temp0](%[res_coeffs]) \n\t" + "addiu %[n], %[n], 1 \n\t" + "absq_s.w %[v_reg], %[v_reg] \n\t" + "sltiu %[temp0], %[v_reg], 2 \n\t" + "move %[ctx_reg], %[v_reg] \n\t" + "movz %[ctx_reg], %[const_2], %[temp0] \n\t" + "sll %[temp1], %[v_reg], 1 \n\t" + "lhx %[temp1], %[temp1](%[VP8LevelFixedCosts]) \n\t" + "slt %[temp0], %[v_reg], %[const_max_level] \n\t" + "movz %[v_reg], %[const_max_level], %[temp0] \n\t" + "addu %[cost], %[cost], %[temp1] \n\t" + "sll %[v_reg], %[v_reg], 1 \n\t" + "sll %[ctx_reg], %[ctx_reg], 2 \n\t" + "lhx %[temp0], %[v_reg](%[t]) \n\t" + "addu %[p_costs], %[p_costs], %[inc_p_costs] \n\t" + "addu %[t], %[p_costs], %[ctx_reg] \n\t" + "addu %[cost], %[cost], %[temp0] \n\t" + "bne %[n], %[res_last], 1b \n\t" + " lw %[t], 0(%[t]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [cost]"+&r"(cost), [t]"+&r"(t), [n]"+&r"(n), [v_reg]"=&r"(v_reg), + [ctx_reg]"=&r"(ctx_reg), [p_costs]"+&r"(p_costs), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1) + : [const_2]"r"(const_2), [const_max_level]"r"(const_max_level), + [VP8LevelFixedCosts]"r"(VP8LevelFixedCosts), [res_last]"r"(res_last), + [res_coeffs]"r"(res_coeffs), [inc_p_costs]"r"(inc_p_costs) + : "memory" + ); + + // Last coefficient is always non-zero + { + const int v = abs(res->coeffs[n]); + assert(v != 0); + cost += VP8LevelCost(t, v); + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = (v == 1) ? 1 : 2; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspCostInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitMIPSdspR2(void) { + VP8GetResidualCost = GetResidualCost; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8EncDspCostInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/cost_sse2.c b/drivers/webp/dsp/cost_sse2.c new file mode 100644 index 0000000000..0cb1c1fa04 --- /dev/null +++ b/drivers/webp/dsp/cost_sse2.c @@ -0,0 +1,119 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 version of cost functions +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE2) +#include <emmintrin.h> + +#include "../enc/cost.h" +#include "../enc/vp8enci.h" +#include "../utils/utils.h" + +//------------------------------------------------------------------------------ + +static void SetResidualCoeffsSSE2(const int16_t* const coeffs, + VP8Residual* const res) { + const __m128i c0 = _mm_loadu_si128((const __m128i*)(coeffs + 0)); + const __m128i c1 = _mm_loadu_si128((const __m128i*)(coeffs + 8)); + // Use SSE2 to compare 16 values with a single instruction. + const __m128i zero = _mm_setzero_si128(); + const __m128i m0 = _mm_packs_epi16(c0, c1); + const __m128i m1 = _mm_cmpeq_epi8(m0, zero); + // Get the comparison results as a bitmask into 16bits. Negate the mask to get + // the position of entries that are not equal to zero. We don't need to mask + // out least significant bits according to res->first, since coeffs[0] is 0 + // if res->first > 0. + const uint32_t mask = 0x0000ffffu ^ (uint32_t)_mm_movemask_epi8(m1); + // The position of the most significant non-zero bit indicates the position of + // the last non-zero value. + assert(res->first == 0 || coeffs[0] == 0); + res->last = mask ? BitsLog2Floor(mask) : -1; + res->coeffs = coeffs; +} + +static int GetResidualCostSSE2(int ctx0, const VP8Residual* const res) { + uint8_t levels[16], ctxs[16]; + uint16_t abs_levels[16]; + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + const int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + + { // precompute clamped levels and contexts, packed to 8b. + const __m128i zero = _mm_setzero_si128(); + const __m128i kCst2 = _mm_set1_epi8(2); + const __m128i kCst67 = _mm_set1_epi8(MAX_VARIABLE_LEVEL); + const __m128i c0 = _mm_loadu_si128((const __m128i*)&res->coeffs[0]); + const __m128i c1 = _mm_loadu_si128((const __m128i*)&res->coeffs[8]); + const __m128i D0 = _mm_sub_epi16(zero, c0); + const __m128i D1 = _mm_sub_epi16(zero, c1); + const __m128i E0 = _mm_max_epi16(c0, D0); // abs(v), 16b + const __m128i E1 = _mm_max_epi16(c1, D1); + const __m128i F = _mm_packs_epi16(E0, E1); + const __m128i G = _mm_min_epu8(F, kCst2); // context = 0,1,2 + const __m128i H = _mm_min_epu8(F, kCst67); // clamp_level in [0..67] + + _mm_storeu_si128((__m128i*)&ctxs[0], G); + _mm_storeu_si128((__m128i*)&levels[0], H); + + _mm_storeu_si128((__m128i*)&abs_levels[0], E0); + _mm_storeu_si128((__m128i*)&abs_levels[8], E1); + } + for (; n < res->last; ++n) { + const int ctx = ctxs[n]; + const int level = levels[n]; + const int flevel = abs_levels[n]; // full level + cost += VP8LevelFixedCosts[flevel] + t[level]; // simplified VP8LevelCost() + t = costs[n + 1][ctx]; + } + // Last coefficient is always non-zero + { + const int level = levels[n]; + const int flevel = abs_levels[n]; + assert(flevel != 0); + cost += VP8LevelFixedCosts[flevel] + t[level]; + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = ctxs[n]; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspCostInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitSSE2(void) { + VP8SetResidualCoeffs = SetResidualCoeffsSSE2; + VP8GetResidualCost = GetResidualCostSSE2; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8EncDspCostInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/cpu.c b/drivers/webp/dsp/cpu.c index 0228734457..35c2af7f58 100644 --- a/drivers/webp/dsp/cpu.c +++ b/drivers/webp/dsp/cpu.c @@ -1,8 +1,10 @@ // Copyright 2011 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // CPU detection @@ -11,14 +13,10 @@ #include "./dsp.h" -#if defined(__ANDROID__) +#if defined(WEBP_ANDROID_NEON) #include <cpu-features.h> #endif -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif - //------------------------------------------------------------------------------ // SSE2 detection. // @@ -31,22 +29,66 @@ static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) { "cpuid\n" "xchg %%edi, %%ebx\n" : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3]) - : "a"(info_type)); + : "a"(info_type), "c"(0)); } #elif defined(__i386__) || defined(__x86_64__) static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) { __asm__ volatile ( "cpuid\n" : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3]) - : "a"(info_type)); + : "a"(info_type), "c"(0)); } +#elif (defined(_M_X64) || defined(_M_IX86)) && \ + defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 150030729 // >= VS2008 SP1 +#include <intrin.h> +#define GetCPUInfo(info, type) __cpuidex(info, type, 0) // set ecx=0 #elif defined(WEBP_MSC_SSE2) #define GetCPUInfo __cpuid #endif +// NaCl has no support for xgetbv or the raw opcode. +#if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__)) +static WEBP_INLINE uint64_t xgetbv(void) { + const uint32_t ecx = 0; + uint32_t eax, edx; + // Use the raw opcode for xgetbv for compatibility with older toolchains. + __asm__ volatile ( + ".byte 0x0f, 0x01, 0xd0\n" + : "=a"(eax), "=d"(edx) : "c" (ecx)); + return ((uint64_t)edx << 32) | eax; +} +#elif (defined(_M_X64) || defined(_M_IX86)) && \ + defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219 // >= VS2010 SP1 +#include <immintrin.h> +#define xgetbv() _xgetbv(0) +#elif defined(_MSC_VER) && defined(_M_IX86) +static WEBP_INLINE uint64_t xgetbv(void) { + uint32_t eax_, edx_; + __asm { + xor ecx, ecx // ecx = 0 + // Use the raw opcode for xgetbv for compatibility with older toolchains. + __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0 + mov eax_, eax + mov edx_, edx + } + return ((uint64_t)edx_ << 32) | eax_; +} +#else +#define xgetbv() 0U // no AVX for older x64 or unrecognized toolchains. +#endif + #if defined(__i386__) || defined(__x86_64__) || defined(WEBP_MSC_SSE2) static int x86CPUInfo(CPUFeature feature) { + int max_cpuid_value; int cpu_info[4]; + + // get the highest feature value cpuid supports + GetCPUInfo(cpu_info, 0); + max_cpuid_value = cpu_info[0]; + if (max_cpuid_value < 1) { + return 0; + } + GetCPUInfo(cpu_info, 1); if (feature == kSSE2) { return 0 != (cpu_info[3] & 0x04000000); @@ -54,10 +96,26 @@ static int x86CPUInfo(CPUFeature feature) { if (feature == kSSE3) { return 0 != (cpu_info[2] & 0x00000001); } + if (feature == kSSE4_1) { + return 0 != (cpu_info[2] & 0x00080000); + } + if (feature == kAVX) { + // bits 27 (OSXSAVE) & 28 (256-bit AVX) + if ((cpu_info[2] & 0x18000000) == 0x18000000) { + // XMM state and YMM state enabled by the OS. + return (xgetbv() & 0x6) == 0x6; + } + } + if (feature == kAVX2) { + if (x86CPUInfo(kAVX) && max_cpuid_value >= 7) { + GetCPUInfo(cpu_info, 7); + return ((cpu_info[1] & 0x00000020) == 0x00000020); + } + } return 0; } VP8CPUInfo VP8GetCPUInfo = x86CPUInfo; -#elif defined(WEBP_ANDROID_NEON) +#elif defined(WEBP_ANDROID_NEON) // NB: needs to be before generic NEON test. static int AndroidCPUInfo(CPUFeature feature) { const AndroidCpuFamily cpu_family = android_getCpuFamily(); const uint64_t cpu_features = android_getCpuFeatures(); @@ -68,7 +126,7 @@ static int AndroidCPUInfo(CPUFeature feature) { return 0; } VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo; -#elif defined(__ARM_NEON__) +#elif defined(WEBP_USE_NEON) // define a dummy function to enable turning off NEON at runtime by setting // VP8DecGetCPUInfo = NULL static int armCPUInfo(CPUFeature feature) { @@ -76,10 +134,17 @@ static int armCPUInfo(CPUFeature feature) { return 1; } VP8CPUInfo VP8GetCPUInfo = armCPUInfo; +#elif defined(WEBP_USE_MIPS32) || defined(WEBP_USE_MIPS_DSP_R2) +static int mipsCPUInfo(CPUFeature feature) { + if ((feature == kMIPS32) || (feature == kMIPSdspR2)) { + return 1; + } else { + return 0; + } + +} +VP8CPUInfo VP8GetCPUInfo = mipsCPUInfo; #else VP8CPUInfo VP8GetCPUInfo = NULL; #endif -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" -#endif diff --git a/drivers/webp/dsp/dec.c b/drivers/webp/dsp/dec.c index 9ae7b6fa76..77a00381c5 100644 --- a/drivers/webp/dsp/dec.c +++ b/drivers/webp/dsp/dec.c @@ -1,53 +1,20 @@ // Copyright 2010 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // -// Speed-critical decoding functions. +// Speed-critical decoding functions, default plain-C implementations. // // Author: Skal (pascal.massimino@gmail.com) #include "./dsp.h" #include "../dec/vp8i.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif - //------------------------------------------------------------------------------ -// run-time tables (~4k) - -static uint8_t abs0[255 + 255 + 1]; // abs(i) -static uint8_t abs1[255 + 255 + 1]; // abs(i)>>1 -static int8_t sclip1[1020 + 1020 + 1]; // clips [-1020, 1020] to [-128, 127] -static int8_t sclip2[112 + 112 + 1]; // clips [-112, 112] to [-16, 15] -static uint8_t clip1[255 + 510 + 1]; // clips [-255,510] to [0,255] - -// We declare this variable 'volatile' to prevent instruction reordering -// and make sure it's set to true _last_ (so as to be thread-safe) -static volatile int tables_ok = 0; - -static void DspInitTables(void) { - if (!tables_ok) { - int i; - for (i = -255; i <= 255; ++i) { - abs0[255 + i] = (i < 0) ? -i : i; - abs1[255 + i] = abs0[255 + i] >> 1; - } - for (i = -1020; i <= 1020; ++i) { - sclip1[1020 + i] = (i < -128) ? -128 : (i > 127) ? 127 : i; - } - for (i = -112; i <= 112; ++i) { - sclip2[112 + i] = (i < -16) ? -16 : (i > 15) ? 15 : i; - } - for (i = -255; i <= 255 + 255; ++i) { - clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i; - } - tables_ok = 1; - } -} static WEBP_INLINE uint8_t clip_8b(int v) { return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255; @@ -59,9 +26,16 @@ static WEBP_INLINE uint8_t clip_8b(int v) { #define STORE(x, y, v) \ dst[x + y * BPS] = clip_8b(dst[x + y * BPS] + ((v) >> 3)) -static const int kC1 = 20091 + (1 << 16); -static const int kC2 = 35468; -#define MUL(a, b) (((a) * (b)) >> 16) +#define STORE2(y, dc, d, c) do { \ + const int DC = (dc); \ + STORE(0, y, DC + (d)); \ + STORE(1, y, DC + (c)); \ + STORE(2, y, DC - (c)); \ + STORE(3, y, DC - (d)); \ +} while (0) + +#define MUL1(a) ((((a) * 20091) >> 16) + (a)) +#define MUL2(a) (((a) * 35468) >> 16) static void TransformOne(const int16_t* in, uint8_t* dst) { int C[4 * 4], *tmp; @@ -70,8 +44,8 @@ static void TransformOne(const int16_t* in, uint8_t* dst) { for (i = 0; i < 4; ++i) { // vertical pass const int a = in[0] + in[8]; // [-4096, 4094] const int b = in[0] - in[8]; // [-4095, 4095] - const int c = MUL(in[4], kC2) - MUL(in[12], kC1); // [-3783, 3783] - const int d = MUL(in[4], kC1) + MUL(in[12], kC2); // [-3785, 3781] + const int c = MUL2(in[4]) - MUL1(in[12]); // [-3783, 3783] + const int d = MUL1(in[4]) + MUL2(in[12]); // [-3785, 3781] tmp[0] = a + d; // [-7881, 7875] tmp[1] = b + c; // [-7878, 7878] tmp[2] = b - c; // [-7878, 7878] @@ -80,7 +54,7 @@ static void TransformOne(const int16_t* in, uint8_t* dst) { in++; } // Each pass is expanding the dynamic range by ~3.85 (upper bound). - // The exact value is (2. + (kC1 + kC2) / 65536). + // The exact value is (2. + (20091 + 35468) / 65536). // After the second pass, maximum interval is [-3794, 3794], assuming // an input in [-2048, 2047] interval. We then need to add a dst value // in the [0, 255] range. @@ -91,8 +65,8 @@ static void TransformOne(const int16_t* in, uint8_t* dst) { const int dc = tmp[0] + 4; const int a = dc + tmp[8]; const int b = dc - tmp[8]; - const int c = MUL(tmp[4], kC2) - MUL(tmp[12], kC1); - const int d = MUL(tmp[4], kC1) + MUL(tmp[12], kC2); + const int c = MUL2(tmp[4]) - MUL1(tmp[12]); + const int d = MUL1(tmp[4]) + MUL2(tmp[12]); STORE(0, 0, a + d); STORE(1, 0, b + c); STORE(2, 0, b - c); @@ -101,7 +75,22 @@ static void TransformOne(const int16_t* in, uint8_t* dst) { dst += BPS; } } -#undef MUL + +// Simplified transform when only in[0], in[1] and in[4] are non-zero +static void TransformAC3(const int16_t* in, uint8_t* dst) { + const int a = in[0] + 4; + const int c4 = MUL2(in[4]); + const int d4 = MUL1(in[4]); + const int c1 = MUL2(in[1]); + const int d1 = MUL1(in[1]); + STORE2(0, a + d4, d1, c1); + STORE2(1, a + c4, d1, c1); + STORE2(2, a - c4, d1, c1); + STORE2(3, a - d4, d1, c1); +} +#undef MUL1 +#undef MUL2 +#undef STORE2 static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) { TransformOne(in, dst); @@ -115,7 +104,7 @@ static void TransformUV(const int16_t* in, uint8_t* dst) { VP8Transform(in + 2 * 16, dst + 4 * BPS, 1); } -static void TransformDC(const int16_t *in, uint8_t* dst) { +static void TransformDC(const int16_t* in, uint8_t* dst) { const int DC = in[0] + 4; int i, j; for (j = 0; j < 4; ++j) { @@ -126,10 +115,10 @@ static void TransformDC(const int16_t *in, uint8_t* dst) { } static void TransformDCUV(const int16_t* in, uint8_t* dst) { - if (in[0 * 16]) TransformDC(in + 0 * 16, dst); - if (in[1 * 16]) TransformDC(in + 1 * 16, dst + 4); - if (in[2 * 16]) TransformDC(in + 2 * 16, dst + 4 * BPS); - if (in[3 * 16]) TransformDC(in + 3 * 16, dst + 4 * BPS + 4); + if (in[0 * 16]) VP8TransformDC(in + 0 * 16, dst); + if (in[1 * 16]) VP8TransformDC(in + 1 * 16, dst + 4); + if (in[2 * 16]) VP8TransformDC(in + 2 * 16, dst + 4 * BPS); + if (in[3 * 16]) VP8TransformDC(in + 3 * 16, dst + 4 * BPS + 4); } #undef STORE @@ -164,16 +153,16 @@ static void TransformWHT(const int16_t* in, int16_t* out) { } } -void (*VP8TransformWHT)(const int16_t* in, int16_t* out) = TransformWHT; +void (*VP8TransformWHT)(const int16_t* in, int16_t* out); //------------------------------------------------------------------------------ // Intra predictions #define DST(x, y) dst[(x) + (y) * BPS] -static WEBP_INLINE void TrueMotion(uint8_t *dst, int size) { +static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) { const uint8_t* top = dst - BPS; - const uint8_t* const clip0 = clip1 + 255 - top[-1]; + const uint8_t* const clip0 = VP8kclip1 - top[-1]; int y; for (y = 0; y < size; ++y) { const uint8_t* const clip = clip0 + dst[-1]; @@ -184,21 +173,21 @@ static WEBP_INLINE void TrueMotion(uint8_t *dst, int size) { dst += BPS; } } -static void TM4(uint8_t *dst) { TrueMotion(dst, 4); } -static void TM8uv(uint8_t *dst) { TrueMotion(dst, 8); } -static void TM16(uint8_t *dst) { TrueMotion(dst, 16); } +static void TM4(uint8_t* dst) { TrueMotion(dst, 4); } +static void TM8uv(uint8_t* dst) { TrueMotion(dst, 8); } +static void TM16(uint8_t* dst) { TrueMotion(dst, 16); } //------------------------------------------------------------------------------ // 16x16 -static void VE16(uint8_t *dst) { // vertical +static void VE16(uint8_t* dst) { // vertical int j; for (j = 0; j < 16; ++j) { memcpy(dst + j * BPS, dst - BPS, 16); } } -static void HE16(uint8_t *dst) { // horizontal +static void HE16(uint8_t* dst) { // horizontal int j; for (j = 16; j > 0; --j) { memset(dst, dst[-1], 16); @@ -213,7 +202,7 @@ static WEBP_INLINE void Put16(int v, uint8_t* dst) { } } -static void DC16(uint8_t *dst) { // DC +static void DC16(uint8_t* dst) { // DC int DC = 16; int j; for (j = 0; j < 16; ++j) { @@ -222,7 +211,7 @@ static void DC16(uint8_t *dst) { // DC Put16(DC >> 5, dst); } -static void DC16NoTop(uint8_t *dst) { // DC with top samples not available +static void DC16NoTop(uint8_t* dst) { // DC with top samples not available int DC = 8; int j; for (j = 0; j < 16; ++j) { @@ -231,7 +220,7 @@ static void DC16NoTop(uint8_t *dst) { // DC with top samples not available Put16(DC >> 4, dst); } -static void DC16NoLeft(uint8_t *dst) { // DC with left samples not available +static void DC16NoLeft(uint8_t* dst) { // DC with left samples not available int DC = 8; int i; for (i = 0; i < 16; ++i) { @@ -240,17 +229,19 @@ static void DC16NoLeft(uint8_t *dst) { // DC with left samples not available Put16(DC >> 4, dst); } -static void DC16NoTopLeft(uint8_t *dst) { // DC with no top and left samples +static void DC16NoTopLeft(uint8_t* dst) { // DC with no top and left samples Put16(0x80, dst); } +VP8PredFunc VP8PredLuma16[NUM_B_DC_MODES]; + //------------------------------------------------------------------------------ // 4x4 #define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2) #define AVG2(a, b) (((a) + (b) + 1) >> 1) -static void VE4(uint8_t *dst) { // vertical +static void VE4(uint8_t* dst) { // vertical const uint8_t* top = dst - BPS; const uint8_t vals[4] = { AVG3(top[-1], top[0], top[1]), @@ -264,7 +255,7 @@ static void VE4(uint8_t *dst) { // vertical } } -static void HE4(uint8_t *dst) { // horizontal +static void HE4(uint8_t* dst) { // horizontal const int A = dst[-1 - BPS]; const int B = dst[-1]; const int C = dst[-1 + BPS]; @@ -276,7 +267,7 @@ static void HE4(uint8_t *dst) { // horizontal *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(D, E, E); } -static void DC4(uint8_t *dst) { // DC +static void DC4(uint8_t* dst) { // DC uint32_t dc = 4; int i; for (i = 0; i < 4; ++i) dc += dst[i - BPS] + dst[-1 + i * BPS]; @@ -284,7 +275,7 @@ static void DC4(uint8_t *dst) { // DC for (i = 0; i < 4; ++i) memset(dst + i * BPS, dc, 4); } -static void RD4(uint8_t *dst) { // Down-right +static void RD4(uint8_t* dst) { // Down-right const int I = dst[-1 + 0 * BPS]; const int J = dst[-1 + 1 * BPS]; const int K = dst[-1 + 2 * BPS]; @@ -295,15 +286,15 @@ static void RD4(uint8_t *dst) { // Down-right const int C = dst[2 - BPS]; const int D = dst[3 - BPS]; DST(0, 3) = AVG3(J, K, L); - DST(0, 2) = DST(1, 3) = AVG3(I, J, K); - DST(0, 1) = DST(1, 2) = DST(2, 3) = AVG3(X, I, J); - DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I); - DST(1, 0) = DST(2, 1) = DST(3, 2) = AVG3(B, A, X); - DST(2, 0) = DST(3, 1) = AVG3(C, B, A); - DST(3, 0) = AVG3(D, C, B); + DST(1, 3) = DST(0, 2) = AVG3(I, J, K); + DST(2, 3) = DST(1, 2) = DST(0, 1) = AVG3(X, I, J); + DST(3, 3) = DST(2, 2) = DST(1, 1) = DST(0, 0) = AVG3(A, X, I); + DST(3, 2) = DST(2, 1) = DST(1, 0) = AVG3(B, A, X); + DST(3, 1) = DST(2, 0) = AVG3(C, B, A); + DST(3, 0) = AVG3(D, C, B); } -static void LD4(uint8_t *dst) { // Down-Left +static void LD4(uint8_t* dst) { // Down-Left const int A = dst[0 - BPS]; const int B = dst[1 - BPS]; const int C = dst[2 - BPS]; @@ -316,12 +307,12 @@ static void LD4(uint8_t *dst) { // Down-Left DST(1, 0) = DST(0, 1) = AVG3(B, C, D); DST(2, 0) = DST(1, 1) = DST(0, 2) = AVG3(C, D, E); DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F); - DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G); - DST(3, 2) = DST(2, 3) = AVG3(F, G, H); - DST(3, 3) = AVG3(G, H, H); + DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G); + DST(3, 2) = DST(2, 3) = AVG3(F, G, H); + DST(3, 3) = AVG3(G, H, H); } -static void VR4(uint8_t *dst) { // Vertical-Right +static void VR4(uint8_t* dst) { // Vertical-Right const int I = dst[-1 + 0 * BPS]; const int J = dst[-1 + 1 * BPS]; const int K = dst[-1 + 2 * BPS]; @@ -343,7 +334,7 @@ static void VR4(uint8_t *dst) { // Vertical-Right DST(3, 1) = AVG3(B, C, D); } -static void VL4(uint8_t *dst) { // Vertical-Left +static void VL4(uint8_t* dst) { // Vertical-Left const int A = dst[0 - BPS]; const int B = dst[1 - BPS]; const int C = dst[2 - BPS]; @@ -365,7 +356,7 @@ static void VL4(uint8_t *dst) { // Vertical-Left DST(3, 3) = AVG3(F, G, H); } -static void HU4(uint8_t *dst) { // Horizontal-Up +static void HU4(uint8_t* dst) { // Horizontal-Up const int I = dst[-1 + 0 * BPS]; const int J = dst[-1 + 1 * BPS]; const int K = dst[-1 + 2 * BPS]; @@ -380,7 +371,7 @@ static void HU4(uint8_t *dst) { // Horizontal-Up DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L; } -static void HD4(uint8_t *dst) { // Horizontal-Down +static void HD4(uint8_t* dst) { // Horizontal-Down const int I = dst[-1 + 0 * BPS]; const int J = dst[-1 + 1 * BPS]; const int K = dst[-1 + 2 * BPS]; @@ -407,17 +398,19 @@ static void HD4(uint8_t *dst) { // Horizontal-Down #undef AVG3 #undef AVG2 +VP8PredFunc VP8PredLuma4[NUM_BMODES]; + //------------------------------------------------------------------------------ // Chroma -static void VE8uv(uint8_t *dst) { // vertical +static void VE8uv(uint8_t* dst) { // vertical int j; for (j = 0; j < 8; ++j) { memcpy(dst + j * BPS, dst - BPS, 8); } } -static void HE8uv(uint8_t *dst) { // horizontal +static void HE8uv(uint8_t* dst) { // horizontal int j; for (j = 0; j < 8; ++j) { memset(dst, dst[-1], 8); @@ -426,60 +419,45 @@ static void HE8uv(uint8_t *dst) { // horizontal } // helper for chroma-DC predictions -static WEBP_INLINE void Put8x8uv(uint64_t v, uint8_t* dst) { +static WEBP_INLINE void Put8x8uv(uint8_t value, uint8_t* dst) { int j; for (j = 0; j < 8; ++j) { - *(uint64_t*)(dst + j * BPS) = v; + memset(dst + j * BPS, value, 8); } } -static void DC8uv(uint8_t *dst) { // DC +static void DC8uv(uint8_t* dst) { // DC int dc0 = 8; int i; for (i = 0; i < 8; ++i) { dc0 += dst[i - BPS] + dst[-1 + i * BPS]; } - Put8x8uv((uint64_t)((dc0 >> 4) * 0x0101010101010101ULL), dst); + Put8x8uv(dc0 >> 4, dst); } -static void DC8uvNoLeft(uint8_t *dst) { // DC with no left samples +static void DC8uvNoLeft(uint8_t* dst) { // DC with no left samples int dc0 = 4; int i; for (i = 0; i < 8; ++i) { dc0 += dst[i - BPS]; } - Put8x8uv((uint64_t)((dc0 >> 3) * 0x0101010101010101ULL), dst); + Put8x8uv(dc0 >> 3, dst); } -static void DC8uvNoTop(uint8_t *dst) { // DC with no top samples +static void DC8uvNoTop(uint8_t* dst) { // DC with no top samples int dc0 = 4; int i; for (i = 0; i < 8; ++i) { dc0 += dst[-1 + i * BPS]; } - Put8x8uv((uint64_t)((dc0 >> 3) * 0x0101010101010101ULL), dst); + Put8x8uv(dc0 >> 3, dst); } -static void DC8uvNoTopLeft(uint8_t *dst) { // DC with nothing - Put8x8uv(0x8080808080808080ULL, dst); +static void DC8uvNoTopLeft(uint8_t* dst) { // DC with nothing + Put8x8uv(0x80, dst); } -//------------------------------------------------------------------------------ -// default C implementations - -const VP8PredFunc VP8PredLuma4[NUM_BMODES] = { - DC4, TM4, VE4, HE4, RD4, VR4, LD4, VL4, HD4, HU4 -}; - -const VP8PredFunc VP8PredLuma16[NUM_B_DC_MODES] = { - DC16, TM16, VE16, HE16, - DC16NoTop, DC16NoLeft, DC16NoTopLeft -}; - -const VP8PredFunc VP8PredChroma8[NUM_B_DC_MODES] = { - DC8uv, TM8uv, VE8uv, HE8uv, - DC8uvNoTop, DC8uvNoLeft, DC8uvNoTopLeft -}; +VP8PredFunc VP8PredChroma8[NUM_B_DC_MODES]; //------------------------------------------------------------------------------ // Edge filtering functions @@ -487,61 +465,62 @@ const VP8PredFunc VP8PredChroma8[NUM_B_DC_MODES] = { // 4 pixels in, 2 pixels out static WEBP_INLINE void do_filter2(uint8_t* p, int step) { const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step]; - const int a = 3 * (q0 - p0) + sclip1[1020 + p1 - q1]; - const int a1 = sclip2[112 + ((a + 4) >> 3)]; - const int a2 = sclip2[112 + ((a + 3) >> 3)]; - p[-step] = clip1[255 + p0 + a2]; - p[ 0] = clip1[255 + q0 - a1]; + const int a = 3 * (q0 - p0) + VP8ksclip1[p1 - q1]; // in [-893,892] + const int a1 = VP8ksclip2[(a + 4) >> 3]; // in [-16,15] + const int a2 = VP8ksclip2[(a + 3) >> 3]; + p[-step] = VP8kclip1[p0 + a2]; + p[ 0] = VP8kclip1[q0 - a1]; } // 4 pixels in, 4 pixels out static WEBP_INLINE void do_filter4(uint8_t* p, int step) { const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step]; const int a = 3 * (q0 - p0); - const int a1 = sclip2[112 + ((a + 4) >> 3)]; - const int a2 = sclip2[112 + ((a + 3) >> 3)]; + const int a1 = VP8ksclip2[(a + 4) >> 3]; + const int a2 = VP8ksclip2[(a + 3) >> 3]; const int a3 = (a1 + 1) >> 1; - p[-2*step] = clip1[255 + p1 + a3]; - p[- step] = clip1[255 + p0 + a2]; - p[ 0] = clip1[255 + q0 - a1]; - p[ step] = clip1[255 + q1 - a3]; + p[-2*step] = VP8kclip1[p1 + a3]; + p[- step] = VP8kclip1[p0 + a2]; + p[ 0] = VP8kclip1[q0 - a1]; + p[ step] = VP8kclip1[q1 - a3]; } // 6 pixels in, 6 pixels out static WEBP_INLINE void do_filter6(uint8_t* p, int step) { const int p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step]; const int q0 = p[0], q1 = p[step], q2 = p[2*step]; - const int a = sclip1[1020 + 3 * (q0 - p0) + sclip1[1020 + p1 - q1]]; + const int a = VP8ksclip1[3 * (q0 - p0) + VP8ksclip1[p1 - q1]]; + // a is in [-128,127], a1 in [-27,27], a2 in [-18,18] and a3 in [-9,9] const int a1 = (27 * a + 63) >> 7; // eq. to ((3 * a + 7) * 9) >> 7 const int a2 = (18 * a + 63) >> 7; // eq. to ((2 * a + 7) * 9) >> 7 const int a3 = (9 * a + 63) >> 7; // eq. to ((1 * a + 7) * 9) >> 7 - p[-3*step] = clip1[255 + p2 + a3]; - p[-2*step] = clip1[255 + p1 + a2]; - p[- step] = clip1[255 + p0 + a1]; - p[ 0] = clip1[255 + q0 - a1]; - p[ step] = clip1[255 + q1 - a2]; - p[ 2*step] = clip1[255 + q2 - a3]; + p[-3*step] = VP8kclip1[p2 + a3]; + p[-2*step] = VP8kclip1[p1 + a2]; + p[- step] = VP8kclip1[p0 + a1]; + p[ 0] = VP8kclip1[q0 - a1]; + p[ step] = VP8kclip1[q1 - a2]; + p[ 2*step] = VP8kclip1[q2 - a3]; } static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) { const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step]; - return (abs0[255 + p1 - p0] > thresh) || (abs0[255 + q1 - q0] > thresh); + return (VP8kabs0[p1 - p0] > thresh) || (VP8kabs0[q1 - q0] > thresh); } -static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int thresh) { - const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step]; - return (2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) <= thresh; +static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int t) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + return ((4 * VP8kabs0[p0 - q0] + VP8kabs0[p1 - q1]) <= t); } static WEBP_INLINE int needs_filter2(const uint8_t* p, int step, int t, int it) { - const int p3 = p[-4*step], p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step]; - const int q0 = p[0], q1 = p[step], q2 = p[2*step], q3 = p[3*step]; - if ((2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) > t) - return 0; - return abs0[255 + p3 - p2] <= it && abs0[255 + p2 - p1] <= it && - abs0[255 + p1 - p0] <= it && abs0[255 + q3 - q2] <= it && - abs0[255 + q2 - q1] <= it && abs0[255 + q1 - q0] <= it; + const int p3 = p[-4 * step], p2 = p[-3 * step], p1 = p[-2 * step]; + const int p0 = p[-step], q0 = p[0]; + const int q1 = p[step], q2 = p[2 * step], q3 = p[3 * step]; + if ((4 * VP8kabs0[p0 - q0] + VP8kabs0[p1 - q1]) > t) return 0; + return VP8kabs0[p3 - p2] <= it && VP8kabs0[p2 - p1] <= it && + VP8kabs0[p1 - p0] <= it && VP8kabs0[q3 - q2] <= it && + VP8kabs0[q2 - q1] <= it && VP8kabs0[q1 - q0] <= it; } //------------------------------------------------------------------------------ @@ -549,8 +528,9 @@ static WEBP_INLINE int needs_filter2(const uint8_t* p, static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { int i; + const int thresh2 = 2 * thresh + 1; for (i = 0; i < 16; ++i) { - if (needs_filter(p + i, stride, thresh)) { + if (needs_filter(p + i, stride, thresh2)) { do_filter2(p + i, stride); } } @@ -558,8 +538,9 @@ static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { int i; + const int thresh2 = 2 * thresh + 1; for (i = 0; i < 16; ++i) { - if (needs_filter(p + i * stride, 1, thresh)) { + if (needs_filter(p + i * stride, 1, thresh2)) { do_filter2(p + i * stride, 1); } } @@ -587,8 +568,9 @@ static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) { static WEBP_INLINE void FilterLoop26(uint8_t* p, int hstride, int vstride, int size, int thresh, int ithresh, int hev_thresh) { + const int thresh2 = 2 * thresh + 1; while (size-- > 0) { - if (needs_filter2(p, hstride, thresh, ithresh)) { + if (needs_filter2(p, hstride, thresh2, ithresh)) { if (hev(p, hstride, hev_thresh)) { do_filter2(p, hstride); } else { @@ -602,8 +584,9 @@ static WEBP_INLINE void FilterLoop26(uint8_t* p, static WEBP_INLINE void FilterLoop24(uint8_t* p, int hstride, int vstride, int size, int thresh, int ithresh, int hev_thresh) { + const int thresh2 = 2 * thresh + 1; while (size-- > 0) { - if (needs_filter2(p, hstride, thresh, ithresh)) { + if (needs_filter2(p, hstride, thresh2, ithresh)) { if (hev(p, hstride, hev_thresh)) { do_filter2(p, hstride); } else { @@ -672,6 +655,7 @@ static void HFilter8i(uint8_t* u, uint8_t* v, int stride, //------------------------------------------------------------------------------ VP8DecIdct2 VP8Transform; +VP8DecIdct VP8TransformAC3; VP8DecIdct VP8TransformUV; VP8DecIdct VP8TransformDC; VP8DecIdct VP8TransformDCUV; @@ -690,15 +674,25 @@ VP8SimpleFilterFunc VP8SimpleVFilter16i; VP8SimpleFilterFunc VP8SimpleHFilter16i; extern void VP8DspInitSSE2(void); +extern void VP8DspInitSSE41(void); extern void VP8DspInitNEON(void); +extern void VP8DspInitMIPS32(void); +extern void VP8DspInitMIPSdspR2(void); + +static volatile VP8CPUInfo dec_last_cpuinfo_used = + (VP8CPUInfo)&dec_last_cpuinfo_used; -void VP8DspInit(void) { - DspInitTables(); +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInit(void) { + if (dec_last_cpuinfo_used == VP8GetCPUInfo) return; + VP8InitClipTables(); + + VP8TransformWHT = TransformWHT; VP8Transform = TransformTwo; VP8TransformUV = TransformUV; VP8TransformDC = TransformDC; VP8TransformDCUV = TransformDCUV; + VP8TransformAC3 = TransformAC3; VP8VFilter16 = VFilter16; VP8HFilter16 = HFilter16; @@ -713,20 +707,60 @@ void VP8DspInit(void) { VP8SimpleVFilter16i = SimpleVFilter16i; VP8SimpleHFilter16i = SimpleHFilter16i; + VP8PredLuma4[0] = DC4; + VP8PredLuma4[1] = TM4; + VP8PredLuma4[2] = VE4; + VP8PredLuma4[3] = HE4; + VP8PredLuma4[4] = RD4; + VP8PredLuma4[5] = VR4; + VP8PredLuma4[6] = LD4; + VP8PredLuma4[7] = VL4; + VP8PredLuma4[8] = HD4; + VP8PredLuma4[9] = HU4; + + VP8PredLuma16[0] = DC16; + VP8PredLuma16[1] = TM16; + VP8PredLuma16[2] = VE16; + VP8PredLuma16[3] = HE16; + VP8PredLuma16[4] = DC16NoTop; + VP8PredLuma16[5] = DC16NoLeft; + VP8PredLuma16[6] = DC16NoTopLeft; + + VP8PredChroma8[0] = DC8uv; + VP8PredChroma8[1] = TM8uv; + VP8PredChroma8[2] = VE8uv; + VP8PredChroma8[3] = HE8uv; + VP8PredChroma8[4] = DC8uvNoTop; + VP8PredChroma8[5] = DC8uvNoLeft; + VP8PredChroma8[6] = DC8uvNoTopLeft; + // If defined, use CPUInfo() to overwrite some pointers with faster versions. - if (VP8GetCPUInfo) { + if (VP8GetCPUInfo != NULL) { #if defined(WEBP_USE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8DspInitSSE2(); +#if defined(WEBP_USE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + VP8DspInitSSE41(); + } +#endif } -#elif defined(WEBP_USE_NEON) +#endif +#if defined(WEBP_USE_NEON) if (VP8GetCPUInfo(kNEON)) { VP8DspInitNEON(); } #endif +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + VP8DspInitMIPS32(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8DspInitMIPSdspR2(); + } +#endif } + dec_last_cpuinfo_used = VP8GetCPUInfo; } - -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" -#endif diff --git a/drivers/webp/dsp/dec_clip_tables.c b/drivers/webp/dsp/dec_clip_tables.c new file mode 100644 index 0000000000..3b6dde86ba --- /dev/null +++ b/drivers/webp/dsp/dec_clip_tables.c @@ -0,0 +1,366 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Clipping tables for filtering +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#define USE_STATIC_TABLES // undefine to have run-time table initialization + +#ifdef USE_STATIC_TABLES + +static const uint8_t abs0[255 + 255 + 1] = { + 0xff, 0xfe, 0xfd, 0xfc, 0xfb, 0xfa, 0xf9, 0xf8, 0xf7, 0xf6, 0xf5, 0xf4, + 0xf3, 0xf2, 0xf1, 0xf0, 0xef, 0xee, 0xed, 0xec, 0xeb, 0xea, 0xe9, 0xe8, + 0xe7, 0xe6, 0xe5, 0xe4, 0xe3, 0xe2, 0xe1, 0xe0, 0xdf, 0xde, 0xdd, 0xdc, + 0xdb, 0xda, 0xd9, 0xd8, 0xd7, 0xd6, 0xd5, 0xd4, 0xd3, 0xd2, 0xd1, 0xd0, + 0xcf, 0xce, 0xcd, 0xcc, 0xcb, 0xca, 0xc9, 0xc8, 0xc7, 0xc6, 0xc5, 0xc4, + 0xc3, 0xc2, 0xc1, 0xc0, 0xbf, 0xbe, 0xbd, 0xbc, 0xbb, 0xba, 0xb9, 0xb8, + 0xb7, 0xb6, 0xb5, 0xb4, 0xb3, 0xb2, 0xb1, 0xb0, 0xaf, 0xae, 0xad, 0xac, + 0xab, 0xaa, 0xa9, 0xa8, 0xa7, 0xa6, 0xa5, 0xa4, 0xa3, 0xa2, 0xa1, 0xa0, + 0x9f, 0x9e, 0x9d, 0x9c, 0x9b, 0x9a, 0x99, 0x98, 0x97, 0x96, 0x95, 0x94, + 0x93, 0x92, 0x91, 0x90, 0x8f, 0x8e, 0x8d, 0x8c, 0x8b, 0x8a, 0x89, 0x88, + 0x87, 0x86, 0x85, 0x84, 0x83, 0x82, 0x81, 0x80, 0x7f, 0x7e, 0x7d, 0x7c, + 0x7b, 0x7a, 0x79, 0x78, 0x77, 0x76, 0x75, 0x74, 0x73, 0x72, 0x71, 0x70, + 0x6f, 0x6e, 0x6d, 0x6c, 0x6b, 0x6a, 0x69, 0x68, 0x67, 0x66, 0x65, 0x64, + 0x63, 0x62, 0x61, 0x60, 0x5f, 0x5e, 0x5d, 0x5c, 0x5b, 0x5a, 0x59, 0x58, + 0x57, 0x56, 0x55, 0x54, 0x53, 0x52, 0x51, 0x50, 0x4f, 0x4e, 0x4d, 0x4c, + 0x4b, 0x4a, 0x49, 0x48, 0x47, 0x46, 0x45, 0x44, 0x43, 0x42, 0x41, 0x40, + 0x3f, 0x3e, 0x3d, 0x3c, 0x3b, 0x3a, 0x39, 0x38, 0x37, 0x36, 0x35, 0x34, + 0x33, 0x32, 0x31, 0x30, 0x2f, 0x2e, 0x2d, 0x2c, 0x2b, 0x2a, 0x29, 0x28, + 0x27, 0x26, 0x25, 0x24, 0x23, 0x22, 0x21, 0x20, 0x1f, 0x1e, 0x1d, 0x1c, + 0x1b, 0x1a, 0x19, 0x18, 0x17, 0x16, 0x15, 0x14, 0x13, 0x12, 0x11, 0x10, + 0x0f, 0x0e, 0x0d, 0x0c, 0x0b, 0x0a, 0x09, 0x08, 0x07, 0x06, 0x05, 0x04, + 0x03, 0x02, 0x01, 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, + 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x10, 0x11, 0x12, 0x13, 0x14, + 0x15, 0x16, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, 0x20, + 0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27, 0x28, 0x29, 0x2a, 0x2b, 0x2c, + 0x2d, 0x2e, 0x2f, 0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37, 0x38, + 0x39, 0x3a, 0x3b, 0x3c, 0x3d, 0x3e, 0x3f, 0x40, 0x41, 0x42, 0x43, 0x44, + 0x45, 0x46, 0x47, 0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f, 0x50, + 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5a, 0x5b, 0x5c, + 0x5d, 0x5e, 0x5f, 0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, + 0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x70, 0x71, 0x72, 0x73, 0x74, + 0x75, 0x76, 0x77, 0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x80, + 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89, 0x8a, 0x8b, 0x8c, + 0x8d, 0x8e, 0x8f, 0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98, + 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f, 0xa0, 0xa1, 0xa2, 0xa3, 0xa4, + 0xa5, 0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf, 0xb0, + 0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xbb, 0xbc, + 0xbd, 0xbe, 0xbf, 0xc0, 0xc1, 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc7, 0xc8, + 0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf, 0xd0, 0xd1, 0xd2, 0xd3, 0xd4, + 0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xdb, 0xdc, 0xdd, 0xde, 0xdf, 0xe0, + 0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea, 0xeb, 0xec, + 0xed, 0xee, 0xef, 0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8, + 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff +}; + +static const int8_t sclip1[1020 + 1020 + 1] = { + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87, + 0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f, 0x90, 0x91, 0x92, 0x93, + 0x94, 0x95, 0x96, 0x97, 0x98, 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f, + 0xa0, 0xa1, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xab, + 0xac, 0xad, 0xae, 0xaf, 0xb0, 0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7, + 0xb8, 0xb9, 0xba, 0xbb, 0xbc, 0xbd, 0xbe, 0xbf, 0xc0, 0xc1, 0xc2, 0xc3, + 0xc4, 0xc5, 0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf, + 0xd0, 0xd1, 0xd2, 0xd3, 0xd4, 0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xdb, + 0xdc, 0xdd, 0xde, 0xdf, 0xe0, 0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, + 0xe8, 0xe9, 0xea, 0xeb, 0xec, 0xed, 0xee, 0xef, 0xf0, 0xf1, 0xf2, 0xf3, + 0xf4, 0xf5, 0xf6, 0xf7, 0xf8, 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff, + 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b, + 0x0c, 0x0d, 0x0e, 0x0f, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, + 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, 0x20, 0x21, 0x22, 0x23, + 0x24, 0x25, 0x26, 0x27, 0x28, 0x29, 0x2a, 0x2b, 0x2c, 0x2d, 0x2e, 0x2f, + 0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37, 0x38, 0x39, 0x3a, 0x3b, + 0x3c, 0x3d, 0x3e, 0x3f, 0x40, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47, + 0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f, 0x50, 0x51, 0x52, 0x53, + 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5a, 0x5b, 0x5c, 0x5d, 0x5e, 0x5f, + 0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69, 0x6a, 0x6b, + 0x6c, 0x6d, 0x6e, 0x6f, 0x70, 0x71, 0x72, 0x73, 0x74, 0x75, 0x76, 0x77, + 0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f +}; + +static const int8_t sclip2[112 + 112 + 1] = { + 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, + 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, + 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, + 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, + 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, + 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, + 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, + 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, + 0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8, 0xf9, 0xfa, 0xfb, + 0xfc, 0xfd, 0xfe, 0xff, 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, + 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, + 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, + 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, + 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, + 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, + 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, + 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, + 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, + 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f +}; + +static const uint8_t clip1[255 + 511 + 1] = { + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, + 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x10, 0x11, 0x12, 0x13, 0x14, + 0x15, 0x16, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, 0x20, + 0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27, 0x28, 0x29, 0x2a, 0x2b, 0x2c, + 0x2d, 0x2e, 0x2f, 0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37, 0x38, + 0x39, 0x3a, 0x3b, 0x3c, 0x3d, 0x3e, 0x3f, 0x40, 0x41, 0x42, 0x43, 0x44, + 0x45, 0x46, 0x47, 0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f, 0x50, + 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5a, 0x5b, 0x5c, + 0x5d, 0x5e, 0x5f, 0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, + 0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x70, 0x71, 0x72, 0x73, 0x74, + 0x75, 0x76, 0x77, 0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x80, + 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89, 0x8a, 0x8b, 0x8c, + 0x8d, 0x8e, 0x8f, 0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98, + 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f, 0xa0, 0xa1, 0xa2, 0xa3, 0xa4, + 0xa5, 0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf, 0xb0, + 0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xbb, 0xbc, + 0xbd, 0xbe, 0xbf, 0xc0, 0xc1, 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc7, 0xc8, + 0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf, 0xd0, 0xd1, 0xd2, 0xd3, 0xd4, + 0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xdb, 0xdc, 0xdd, 0xde, 0xdf, 0xe0, + 0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea, 0xeb, 0xec, + 0xed, 0xee, 0xef, 0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8, + 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff +}; + +#else + +// uninitialized tables +static uint8_t abs0[255 + 255 + 1]; +static int8_t sclip1[1020 + 1020 + 1]; +static int8_t sclip2[112 + 112 + 1]; +static uint8_t clip1[255 + 511 + 1]; + +// We declare this variable 'volatile' to prevent instruction reordering +// and make sure it's set to true _last_ (so as to be thread-safe) +static volatile int tables_ok = 0; + +#endif + +const int8_t* const VP8ksclip1 = &sclip1[1020]; +const int8_t* const VP8ksclip2 = &sclip2[112]; +const uint8_t* const VP8kclip1 = &clip1[255]; +const uint8_t* const VP8kabs0 = &abs0[255]; + +WEBP_TSAN_IGNORE_FUNCTION void VP8InitClipTables(void) { +#if !defined(USE_STATIC_TABLES) + int i; + if (!tables_ok) { + for (i = -255; i <= 255; ++i) { + abs0[255 + i] = (i < 0) ? -i : i; + } + for (i = -1020; i <= 1020; ++i) { + sclip1[1020 + i] = (i < -128) ? -128 : (i > 127) ? 127 : i; + } + for (i = -112; i <= 112; ++i) { + sclip2[112 + i] = (i < -16) ? -16 : (i > 15) ? 15 : i; + } + for (i = -255; i <= 255 + 255; ++i) { + clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i; + } + tables_ok = 1; + } +#endif // USE_STATIC_TABLES +} diff --git a/drivers/webp/dsp/dec_mips32.c b/drivers/webp/dsp/dec_mips32.c new file mode 100644 index 0000000000..4e9ef42605 --- /dev/null +++ b/drivers/webp/dsp/dec_mips32.c @@ -0,0 +1,587 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of dsp functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include "./mips_macro.h" + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; + +static WEBP_INLINE int abs_mips32(int x) { + const int sign = x >> 31; + return (x ^ sign) - sign; +} + +// 4 pixels in, 2 pixels out +static WEBP_INLINE void do_filter2(uint8_t* p, int step) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + const int a = 3 * (q0 - p0) + VP8ksclip1[p1 - q1]; + const int a1 = VP8ksclip2[(a + 4) >> 3]; + const int a2 = VP8ksclip2[(a + 3) >> 3]; + p[-step] = VP8kclip1[p0 + a2]; + p[ 0] = VP8kclip1[q0 - a1]; +} + +// 4 pixels in, 4 pixels out +static WEBP_INLINE void do_filter4(uint8_t* p, int step) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + const int a = 3 * (q0 - p0); + const int a1 = VP8ksclip2[(a + 4) >> 3]; + const int a2 = VP8ksclip2[(a + 3) >> 3]; + const int a3 = (a1 + 1) >> 1; + p[-2 * step] = VP8kclip1[p1 + a3]; + p[- step] = VP8kclip1[p0 + a2]; + p[ 0] = VP8kclip1[q0 - a1]; + p[ step] = VP8kclip1[q1 - a3]; +} + +// 6 pixels in, 6 pixels out +static WEBP_INLINE void do_filter6(uint8_t* p, int step) { + const int p2 = p[-3 * step], p1 = p[-2 * step], p0 = p[-step]; + const int q0 = p[0], q1 = p[step], q2 = p[2 * step]; + const int a = VP8ksclip1[3 * (q0 - p0) + VP8ksclip1[p1 - q1]]; + // a is in [-128,127], a1 in [-27,27], a2 in [-18,18] and a3 in [-9,9] + const int a1 = (27 * a + 63) >> 7; // eq. to ((3 * a + 7) * 9) >> 7 + const int a2 = (18 * a + 63) >> 7; // eq. to ((2 * a + 7) * 9) >> 7 + const int a3 = (9 * a + 63) >> 7; // eq. to ((1 * a + 7) * 9) >> 7 + p[-3 * step] = VP8kclip1[p2 + a3]; + p[-2 * step] = VP8kclip1[p1 + a2]; + p[- step] = VP8kclip1[p0 + a1]; + p[ 0] = VP8kclip1[q0 - a1]; + p[ step] = VP8kclip1[q1 - a2]; + p[ 2 * step] = VP8kclip1[q2 - a3]; +} + +static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + return (abs_mips32(p1 - p0) > thresh) || (abs_mips32(q1 - q0) > thresh); +} + +static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int t) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + return ((4 * abs_mips32(p0 - q0) + abs_mips32(p1 - q1)) <= t); +} + +static WEBP_INLINE int needs_filter2(const uint8_t* p, + int step, int t, int it) { + const int p3 = p[-4 * step], p2 = p[-3 * step]; + const int p1 = p[-2 * step], p0 = p[-step]; + const int q0 = p[0], q1 = p[step], q2 = p[2 * step], q3 = p[3 * step]; + if ((4 * abs_mips32(p0 - q0) + abs_mips32(p1 - q1)) > t) { + return 0; + } + return abs_mips32(p3 - p2) <= it && abs_mips32(p2 - p1) <= it && + abs_mips32(p1 - p0) <= it && abs_mips32(q3 - q2) <= it && + abs_mips32(q2 - q1) <= it && abs_mips32(q1 - q0) <= it; +} + +static WEBP_INLINE void FilterLoop26(uint8_t* p, + int hstride, int vstride, int size, + int thresh, int ithresh, int hev_thresh) { + const int thresh2 = 2 * thresh + 1; + while (size-- > 0) { + if (needs_filter2(p, hstride, thresh2, ithresh)) { + if (hev(p, hstride, hev_thresh)) { + do_filter2(p, hstride); + } else { + do_filter6(p, hstride); + } + } + p += vstride; + } +} + +static WEBP_INLINE void FilterLoop24(uint8_t* p, + int hstride, int vstride, int size, + int thresh, int ithresh, int hev_thresh) { + const int thresh2 = 2 * thresh + 1; + while (size-- > 0) { + if (needs_filter2(p, hstride, thresh2, ithresh)) { + if (hev(p, hstride, hev_thresh)) { + do_filter2(p, hstride); + } else { + do_filter4(p, hstride); + } + } + p += vstride; + } +} + +// on macroblock edges +static void VFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(p, stride, 1, 16, thresh, ithresh, hev_thresh); +} + +static void HFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(p, 1, stride, 16, thresh, ithresh, hev_thresh); +} + +// 8-pixels wide variant, for chroma filtering +static void VFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(u, stride, 1, 8, thresh, ithresh, hev_thresh); + FilterLoop26(v, stride, 1, 8, thresh, ithresh, hev_thresh); +} + +static void HFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(u, 1, stride, 8, thresh, ithresh, hev_thresh); + FilterLoop26(v, 1, stride, 8, thresh, ithresh, hev_thresh); +} + +static void VFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh); + FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh); +} + +static void HFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh); + FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh); +} + +// on three inner edges +static void VFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4 * stride; + FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh); + } +} + +static void HFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4; + FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh); + } +} + +//------------------------------------------------------------------------------ +// Simple In-loop filtering (Paragraph 15.2) + +static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { + int i; + const int thresh2 = 2 * thresh + 1; + for (i = 0; i < 16; ++i) { + if (needs_filter(p + i, stride, thresh2)) { + do_filter2(p + i, stride); + } + } +} + +static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { + int i; + const int thresh2 = 2 * thresh + 1; + for (i = 0; i < 16; ++i) { + if (needs_filter(p + i * stride, 1, thresh2)) { + do_filter2(p + i * stride, 1); + } + } +} + +static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4 * stride; + SimpleVFilter16(p, stride, thresh); + } +} + +static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4; + SimpleHFilter16(p, stride, thresh); + } +} + +static void TransformOne(const int16_t* in, uint8_t* dst) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14; + int temp15, temp16, temp17, temp18; + int16_t* p_in = (int16_t*)in; + + // loops unrolled and merged to avoid usage of tmp buffer + // and to reduce number of stalls. MUL macro is written + // in assembler and inlined + __asm__ volatile( + "lh %[temp0], 0(%[in]) \n\t" + "lh %[temp8], 16(%[in]) \n\t" + "lh %[temp4], 8(%[in]) \n\t" + "lh %[temp12], 24(%[in]) \n\t" + "addu %[temp16], %[temp0], %[temp8] \n\t" + "subu %[temp0], %[temp0], %[temp8] \n\t" + "mul %[temp8], %[temp4], %[kC2] \n\t" + "mul %[temp17], %[temp12], %[kC1] \n\t" + "mul %[temp4], %[temp4], %[kC1] \n\t" + "mul %[temp12], %[temp12], %[kC2] \n\t" + "lh %[temp1], 2(%[in]) \n\t" + "lh %[temp5], 10(%[in]) \n\t" + "lh %[temp9], 18(%[in]) \n\t" + "lh %[temp13], 26(%[in]) \n\t" + "sra %[temp8], %[temp8], 16 \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "sra %[temp4], %[temp4], 16 \n\t" + "sra %[temp12], %[temp12], 16 \n\t" + "lh %[temp2], 4(%[in]) \n\t" + "lh %[temp6], 12(%[in]) \n\t" + "lh %[temp10], 20(%[in]) \n\t" + "lh %[temp14], 28(%[in]) \n\t" + "subu %[temp17], %[temp8], %[temp17] \n\t" + "addu %[temp4], %[temp4], %[temp12] \n\t" + "addu %[temp8], %[temp16], %[temp4] \n\t" + "subu %[temp4], %[temp16], %[temp4] \n\t" + "addu %[temp16], %[temp1], %[temp9] \n\t" + "subu %[temp1], %[temp1], %[temp9] \n\t" + "lh %[temp3], 6(%[in]) \n\t" + "lh %[temp7], 14(%[in]) \n\t" + "lh %[temp11], 22(%[in]) \n\t" + "lh %[temp15], 30(%[in]) \n\t" + "addu %[temp12], %[temp0], %[temp17] \n\t" + "subu %[temp0], %[temp0], %[temp17] \n\t" + "mul %[temp9], %[temp5], %[kC2] \n\t" + "mul %[temp17], %[temp13], %[kC1] \n\t" + "mul %[temp5], %[temp5], %[kC1] \n\t" + "mul %[temp13], %[temp13], %[kC2] \n\t" + "sra %[temp9], %[temp9], 16 \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "subu %[temp17], %[temp9], %[temp17] \n\t" + "sra %[temp5], %[temp5], 16 \n\t" + "sra %[temp13], %[temp13], 16 \n\t" + "addu %[temp5], %[temp5], %[temp13] \n\t" + "addu %[temp13], %[temp1], %[temp17] \n\t" + "subu %[temp1], %[temp1], %[temp17] \n\t" + "mul %[temp17], %[temp14], %[kC1] \n\t" + "mul %[temp14], %[temp14], %[kC2] \n\t" + "addu %[temp9], %[temp16], %[temp5] \n\t" + "subu %[temp5], %[temp16], %[temp5] \n\t" + "addu %[temp16], %[temp2], %[temp10] \n\t" + "subu %[temp2], %[temp2], %[temp10] \n\t" + "mul %[temp10], %[temp6], %[kC2] \n\t" + "mul %[temp6], %[temp6], %[kC1] \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "sra %[temp14], %[temp14], 16 \n\t" + "sra %[temp10], %[temp10], 16 \n\t" + "sra %[temp6], %[temp6], 16 \n\t" + "subu %[temp17], %[temp10], %[temp17] \n\t" + "addu %[temp6], %[temp6], %[temp14] \n\t" + "addu %[temp10], %[temp16], %[temp6] \n\t" + "subu %[temp6], %[temp16], %[temp6] \n\t" + "addu %[temp14], %[temp2], %[temp17] \n\t" + "subu %[temp2], %[temp2], %[temp17] \n\t" + "mul %[temp17], %[temp15], %[kC1] \n\t" + "mul %[temp15], %[temp15], %[kC2] \n\t" + "addu %[temp16], %[temp3], %[temp11] \n\t" + "subu %[temp3], %[temp3], %[temp11] \n\t" + "mul %[temp11], %[temp7], %[kC2] \n\t" + "mul %[temp7], %[temp7], %[kC1] \n\t" + "addiu %[temp8], %[temp8], 4 \n\t" + "addiu %[temp12], %[temp12], 4 \n\t" + "addiu %[temp0], %[temp0], 4 \n\t" + "addiu %[temp4], %[temp4], 4 \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "sra %[temp15], %[temp15], 16 \n\t" + "sra %[temp11], %[temp11], 16 \n\t" + "sra %[temp7], %[temp7], 16 \n\t" + "subu %[temp17], %[temp11], %[temp17] \n\t" + "addu %[temp7], %[temp7], %[temp15] \n\t" + "addu %[temp15], %[temp3], %[temp17] \n\t" + "subu %[temp3], %[temp3], %[temp17] \n\t" + "addu %[temp11], %[temp16], %[temp7] \n\t" + "subu %[temp7], %[temp16], %[temp7] \n\t" + "addu %[temp16], %[temp8], %[temp10] \n\t" + "subu %[temp8], %[temp8], %[temp10] \n\t" + "mul %[temp10], %[temp9], %[kC2] \n\t" + "mul %[temp17], %[temp11], %[kC1] \n\t" + "mul %[temp9], %[temp9], %[kC1] \n\t" + "mul %[temp11], %[temp11], %[kC2] \n\t" + "sra %[temp10], %[temp10], 16 \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "sra %[temp9], %[temp9], 16 \n\t" + "sra %[temp11], %[temp11], 16 \n\t" + "subu %[temp17], %[temp10], %[temp17] \n\t" + "addu %[temp11], %[temp9], %[temp11] \n\t" + "addu %[temp10], %[temp12], %[temp14] \n\t" + "subu %[temp12], %[temp12], %[temp14] \n\t" + "mul %[temp14], %[temp13], %[kC2] \n\t" + "mul %[temp9], %[temp15], %[kC1] \n\t" + "mul %[temp13], %[temp13], %[kC1] \n\t" + "mul %[temp15], %[temp15], %[kC2] \n\t" + "sra %[temp14], %[temp14], 16 \n\t" + "sra %[temp9], %[temp9], 16 \n\t" + "sra %[temp13], %[temp13], 16 \n\t" + "sra %[temp15], %[temp15], 16 \n\t" + "subu %[temp9], %[temp14], %[temp9] \n\t" + "addu %[temp15], %[temp13], %[temp15] \n\t" + "addu %[temp14], %[temp0], %[temp2] \n\t" + "subu %[temp0], %[temp0], %[temp2] \n\t" + "mul %[temp2], %[temp1], %[kC2] \n\t" + "mul %[temp13], %[temp3], %[kC1] \n\t" + "mul %[temp1], %[temp1], %[kC1] \n\t" + "mul %[temp3], %[temp3], %[kC2] \n\t" + "sra %[temp2], %[temp2], 16 \n\t" + "sra %[temp13], %[temp13], 16 \n\t" + "sra %[temp1], %[temp1], 16 \n\t" + "sra %[temp3], %[temp3], 16 \n\t" + "subu %[temp13], %[temp2], %[temp13] \n\t" + "addu %[temp3], %[temp1], %[temp3] \n\t" + "addu %[temp2], %[temp4], %[temp6] \n\t" + "subu %[temp4], %[temp4], %[temp6] \n\t" + "mul %[temp6], %[temp5], %[kC2] \n\t" + "mul %[temp1], %[temp7], %[kC1] \n\t" + "mul %[temp5], %[temp5], %[kC1] \n\t" + "mul %[temp7], %[temp7], %[kC2] \n\t" + "sra %[temp6], %[temp6], 16 \n\t" + "sra %[temp1], %[temp1], 16 \n\t" + "sra %[temp5], %[temp5], 16 \n\t" + "sra %[temp7], %[temp7], 16 \n\t" + "subu %[temp1], %[temp6], %[temp1] \n\t" + "addu %[temp7], %[temp5], %[temp7] \n\t" + "addu %[temp5], %[temp16], %[temp11] \n\t" + "subu %[temp16], %[temp16], %[temp11] \n\t" + "addu %[temp11], %[temp8], %[temp17] \n\t" + "subu %[temp8], %[temp8], %[temp17] \n\t" + "sra %[temp5], %[temp5], 3 \n\t" + "sra %[temp16], %[temp16], 3 \n\t" + "sra %[temp11], %[temp11], 3 \n\t" + "sra %[temp8], %[temp8], 3 \n\t" + "addu %[temp17], %[temp10], %[temp15] \n\t" + "subu %[temp10], %[temp10], %[temp15] \n\t" + "addu %[temp15], %[temp12], %[temp9] \n\t" + "subu %[temp12], %[temp12], %[temp9] \n\t" + "sra %[temp17], %[temp17], 3 \n\t" + "sra %[temp10], %[temp10], 3 \n\t" + "sra %[temp15], %[temp15], 3 \n\t" + "sra %[temp12], %[temp12], 3 \n\t" + "addu %[temp9], %[temp14], %[temp3] \n\t" + "subu %[temp14], %[temp14], %[temp3] \n\t" + "addu %[temp3], %[temp0], %[temp13] \n\t" + "subu %[temp0], %[temp0], %[temp13] \n\t" + "sra %[temp9], %[temp9], 3 \n\t" + "sra %[temp14], %[temp14], 3 \n\t" + "sra %[temp3], %[temp3], 3 \n\t" + "sra %[temp0], %[temp0], 3 \n\t" + "addu %[temp13], %[temp2], %[temp7] \n\t" + "subu %[temp2], %[temp2], %[temp7] \n\t" + "addu %[temp7], %[temp4], %[temp1] \n\t" + "subu %[temp4], %[temp4], %[temp1] \n\t" + "sra %[temp13], %[temp13], 3 \n\t" + "sra %[temp2], %[temp2], 3 \n\t" + "sra %[temp7], %[temp7], 3 \n\t" + "sra %[temp4], %[temp4], 3 \n\t" + "addiu %[temp6], $zero, 255 \n\t" + "lbu %[temp1], 0+0*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp1], %[temp1], %[temp5] \n\t" + "sra %[temp5], %[temp1], 8 \n\t" + "sra %[temp18], %[temp1], 31 \n\t" + "beqz %[temp5], 1f \n\t" + "xor %[temp1], %[temp1], %[temp1] \n\t" + "movz %[temp1], %[temp6], %[temp18] \n\t" + "1: \n\t" + "lbu %[temp18], 1+0*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp1], 0+0*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp18], %[temp18], %[temp11] \n\t" + "sra %[temp11], %[temp18], 8 \n\t" + "sra %[temp1], %[temp18], 31 \n\t" + "beqz %[temp11], 2f \n\t" + "xor %[temp18], %[temp18], %[temp18] \n\t" + "movz %[temp18], %[temp6], %[temp1] \n\t" + "2: \n\t" + "lbu %[temp1], 2+0*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp18], 1+0*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp1], %[temp1], %[temp8] \n\t" + "sra %[temp8], %[temp1], 8 \n\t" + "sra %[temp18], %[temp1], 31 \n\t" + "beqz %[temp8], 3f \n\t" + "xor %[temp1], %[temp1], %[temp1] \n\t" + "movz %[temp1], %[temp6], %[temp18] \n\t" + "3: \n\t" + "lbu %[temp18], 3+0*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp1], 2+0*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp18], %[temp18], %[temp16] \n\t" + "sra %[temp16], %[temp18], 8 \n\t" + "sra %[temp1], %[temp18], 31 \n\t" + "beqz %[temp16], 4f \n\t" + "xor %[temp18], %[temp18], %[temp18] \n\t" + "movz %[temp18], %[temp6], %[temp1] \n\t" + "4: \n\t" + "sb %[temp18], 3+0*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp5], 0+1*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp8], 1+1*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp11], 2+1*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp16], 3+1*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp5], %[temp5], %[temp17] \n\t" + "addu %[temp8], %[temp8], %[temp15] \n\t" + "addu %[temp11], %[temp11], %[temp12] \n\t" + "addu %[temp16], %[temp16], %[temp10] \n\t" + "sra %[temp18], %[temp5], 8 \n\t" + "sra %[temp1], %[temp5], 31 \n\t" + "beqz %[temp18], 5f \n\t" + "xor %[temp5], %[temp5], %[temp5] \n\t" + "movz %[temp5], %[temp6], %[temp1] \n\t" + "5: \n\t" + "sra %[temp18], %[temp8], 8 \n\t" + "sra %[temp1], %[temp8], 31 \n\t" + "beqz %[temp18], 6f \n\t" + "xor %[temp8], %[temp8], %[temp8] \n\t" + "movz %[temp8], %[temp6], %[temp1] \n\t" + "6: \n\t" + "sra %[temp18], %[temp11], 8 \n\t" + "sra %[temp1], %[temp11], 31 \n\t" + "sra %[temp17], %[temp16], 8 \n\t" + "sra %[temp15], %[temp16], 31 \n\t" + "beqz %[temp18], 7f \n\t" + "xor %[temp11], %[temp11], %[temp11] \n\t" + "movz %[temp11], %[temp6], %[temp1] \n\t" + "7: \n\t" + "beqz %[temp17], 8f \n\t" + "xor %[temp16], %[temp16], %[temp16] \n\t" + "movz %[temp16], %[temp6], %[temp15] \n\t" + "8: \n\t" + "sb %[temp5], 0+1*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp8], 1+1*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp11], 2+1*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp16], 3+1*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp5], 0+2*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp8], 1+2*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp11], 2+2*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp16], 3+2*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp5], %[temp5], %[temp9] \n\t" + "addu %[temp8], %[temp8], %[temp3] \n\t" + "addu %[temp11], %[temp11], %[temp0] \n\t" + "addu %[temp16], %[temp16], %[temp14] \n\t" + "sra %[temp18], %[temp5], 8 \n\t" + "sra %[temp1], %[temp5], 31 \n\t" + "sra %[temp17], %[temp8], 8 \n\t" + "sra %[temp15], %[temp8], 31 \n\t" + "sra %[temp12], %[temp11], 8 \n\t" + "sra %[temp10], %[temp11], 31 \n\t" + "sra %[temp9], %[temp16], 8 \n\t" + "sra %[temp3], %[temp16], 31 \n\t" + "beqz %[temp18], 9f \n\t" + "xor %[temp5], %[temp5], %[temp5] \n\t" + "movz %[temp5], %[temp6], %[temp1] \n\t" + "9: \n\t" + "beqz %[temp17], 10f \n\t" + "xor %[temp8], %[temp8], %[temp8] \n\t" + "movz %[temp8], %[temp6], %[temp15] \n\t" + "10: \n\t" + "beqz %[temp12], 11f \n\t" + "xor %[temp11], %[temp11], %[temp11] \n\t" + "movz %[temp11], %[temp6], %[temp10] \n\t" + "11: \n\t" + "beqz %[temp9], 12f \n\t" + "xor %[temp16], %[temp16], %[temp16] \n\t" + "movz %[temp16], %[temp6], %[temp3] \n\t" + "12: \n\t" + "sb %[temp5], 0+2*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp8], 1+2*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp11], 2+2*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp16], 3+2*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp5], 0+3*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp8], 1+3*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp11], 2+3*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp16], 3+3*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp5], %[temp5], %[temp13] \n\t" + "addu %[temp8], %[temp8], %[temp7] \n\t" + "addu %[temp11], %[temp11], %[temp4] \n\t" + "addu %[temp16], %[temp16], %[temp2] \n\t" + "sra %[temp18], %[temp5], 8 \n\t" + "sra %[temp1], %[temp5], 31 \n\t" + "sra %[temp17], %[temp8], 8 \n\t" + "sra %[temp15], %[temp8], 31 \n\t" + "sra %[temp12], %[temp11], 8 \n\t" + "sra %[temp10], %[temp11], 31 \n\t" + "sra %[temp9], %[temp16], 8 \n\t" + "sra %[temp3], %[temp16], 31 \n\t" + "beqz %[temp18], 13f \n\t" + "xor %[temp5], %[temp5], %[temp5] \n\t" + "movz %[temp5], %[temp6], %[temp1] \n\t" + "13: \n\t" + "beqz %[temp17], 14f \n\t" + "xor %[temp8], %[temp8], %[temp8] \n\t" + "movz %[temp8], %[temp6], %[temp15] \n\t" + "14: \n\t" + "beqz %[temp12], 15f \n\t" + "xor %[temp11], %[temp11], %[temp11] \n\t" + "movz %[temp11], %[temp6], %[temp10] \n\t" + "15: \n\t" + "beqz %[temp9], 16f \n\t" + "xor %[temp16], %[temp16], %[temp16] \n\t" + "movz %[temp16], %[temp6], %[temp3] \n\t" + "16: \n\t" + "sb %[temp5], 0+3*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp8], 1+3*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp11], 2+3*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp16], 3+3*" XSTR(BPS) "(%[dst]) \n\t" + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11), + [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14), + [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17), + [temp18]"=&r"(temp18) + : [in]"r"(p_in), [kC1]"r"(kC1), [kC2]"r"(kC2), [dst]"r"(dst) + : "memory", "hi", "lo" + ); +} + +static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) { + TransformOne(in, dst); + if (do_two) { + TransformOne(in + 16, dst + 4); + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8DspInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitMIPS32(void) { + VP8InitClipTables(); + + VP8Transform = TransformTwo; + + VP8VFilter16 = VFilter16; + VP8HFilter16 = HFilter16; + VP8VFilter8 = VFilter8; + VP8HFilter8 = HFilter8; + VP8VFilter16i = VFilter16i; + VP8HFilter16i = HFilter16i; + VP8VFilter8i = VFilter8i; + VP8HFilter8i = HFilter8i; + + VP8SimpleVFilter16 = SimpleVFilter16; + VP8SimpleHFilter16 = SimpleHFilter16; + VP8SimpleVFilter16i = SimpleVFilter16i; + VP8SimpleHFilter16i = SimpleHFilter16i; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(VP8DspInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/drivers/webp/dsp/dec_mips_dsp_r2.c b/drivers/webp/dsp/dec_mips_dsp_r2.c new file mode 100644 index 0000000000..db5c657228 --- /dev/null +++ b/drivers/webp/dsp/dec_mips_dsp_r2.c @@ -0,0 +1,994 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of dsp functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "./mips_macro.h" + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; + +#define MUL(a, b) (((a) * (b)) >> 16) + +static void TransformDC(const int16_t* in, uint8_t* dst) { + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9, temp10; + + __asm__ volatile ( + LOAD_WITH_OFFSET_X4(temp1, temp2, temp3, temp4, dst, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + "lh %[temp5], 0(%[in]) \n\t" + "addiu %[temp5], %[temp5], 4 \n\t" + "ins %[temp5], %[temp5], 16, 16 \n\t" + "shra.ph %[temp5], %[temp5], 3 \n\t" + CONVERT_2_BYTES_TO_HALF(temp6, temp7, temp8, temp9, temp10, temp1, temp2, + temp3, temp1, temp2, temp3, temp4) + STORE_SAT_SUM_X2(temp6, temp7, temp8, temp9, temp10, temp1, temp2, temp3, + temp5, temp5, temp5, temp5, temp5, temp5, temp5, temp5, + dst, 0, 1, 2, 3, BPS) + + OUTPUT_EARLY_CLOBBER_REGS_10() + : [in]"r"(in), [dst]"r"(dst) + : "memory" + ); +} + +static void TransformAC3(const int16_t* in, uint8_t* dst) { + const int a = in[0] + 4; + int c4 = MUL(in[4], kC2); + const int d4 = MUL(in[4], kC1); + const int c1 = MUL(in[1], kC2); + const int d1 = MUL(in[1], kC1); + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17, temp18; + + __asm__ volatile ( + "ins %[c4], %[d4], 16, 16 \n\t" + "replv.ph %[temp1], %[a] \n\t" + "replv.ph %[temp4], %[d1] \n\t" + ADD_SUB_HALVES(temp2, temp3, temp1, c4) + "replv.ph %[temp5], %[c1] \n\t" + SHIFT_R_SUM_X2(temp1, temp6, temp7, temp8, temp2, temp9, temp10, temp4, + temp2, temp2, temp3, temp3, temp4, temp5, temp4, temp5) + LOAD_WITH_OFFSET_X4(temp3, temp5, temp11, temp12, dst, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp13, temp14, temp3, temp15, temp5, temp16, + temp11, temp17, temp3, temp5, temp11, temp12) + PACK_2_HALVES_TO_WORD(temp12, temp18, temp7, temp6, temp1, temp8, temp2, + temp4, temp7, temp6, temp10, temp9) + STORE_SAT_SUM_X2(temp13, temp14, temp3, temp15, temp5, temp16, temp11, + temp17, temp12, temp18, temp1, temp8, temp2, temp4, + temp7, temp6, dst, 0, 1, 2, 3, BPS) + + OUTPUT_EARLY_CLOBBER_REGS_18(), + [c4]"+&r"(c4) + : [dst]"r"(dst), [a]"r"(a), [d1]"r"(d1), [d4]"r"(d4), [c1]"r"(c1) + : "memory" + ); +} + +static void TransformOne(const int16_t* in, uint8_t* dst) { + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17, temp18; + + __asm__ volatile ( + "ulw %[temp1], 0(%[in]) \n\t" + "ulw %[temp2], 16(%[in]) \n\t" + LOAD_IN_X2(temp5, temp6, 24, 26) + ADD_SUB_HALVES(temp3, temp4, temp1, temp2) + LOAD_IN_X2(temp1, temp2, 8, 10) + MUL_SHIFT_SUM(temp7, temp8, temp9, temp10, temp11, temp12, temp13, temp14, + temp10, temp8, temp9, temp7, temp1, temp2, temp5, temp6, + temp13, temp11, temp14, temp12) + INSERT_HALF_X2(temp8, temp7, temp10, temp9) + "ulw %[temp17], 4(%[in]) \n\t" + "ulw %[temp18], 20(%[in]) \n\t" + ADD_SUB_HALVES(temp1, temp2, temp3, temp8) + ADD_SUB_HALVES(temp5, temp6, temp4, temp7) + ADD_SUB_HALVES(temp7, temp8, temp17, temp18) + LOAD_IN_X2(temp17, temp18, 12, 14) + LOAD_IN_X2(temp9, temp10, 28, 30) + MUL_SHIFT_SUM(temp11, temp12, temp13, temp14, temp15, temp16, temp4, temp17, + temp12, temp14, temp11, temp13, temp17, temp18, temp9, temp10, + temp15, temp4, temp16, temp17) + INSERT_HALF_X2(temp11, temp12, temp13, temp14) + ADD_SUB_HALVES(temp17, temp8, temp8, temp11) + ADD_SUB_HALVES(temp3, temp4, temp7, temp12) + + // horizontal + SRA_16(temp9, temp10, temp11, temp12, temp1, temp2, temp5, temp6) + INSERT_HALF_X2(temp1, temp6, temp5, temp2) + SRA_16(temp13, temp14, temp15, temp16, temp3, temp4, temp17, temp8) + "repl.ph %[temp2], 0x4 \n\t" + INSERT_HALF_X2(temp3, temp8, temp17, temp4) + "addq.ph %[temp1], %[temp1], %[temp2] \n\t" + "addq.ph %[temp6], %[temp6], %[temp2] \n\t" + ADD_SUB_HALVES(temp2, temp4, temp1, temp3) + ADD_SUB_HALVES(temp5, temp7, temp6, temp8) + MUL_SHIFT_SUM(temp1, temp3, temp6, temp8, temp9, temp13, temp17, temp18, + temp3, temp13, temp1, temp9, temp9, temp13, temp11, temp15, + temp6, temp17, temp8, temp18) + MUL_SHIFT_SUM(temp6, temp8, temp18, temp17, temp11, temp15, temp12, temp16, + temp8, temp15, temp6, temp11, temp12, temp16, temp10, temp14, + temp18, temp12, temp17, temp16) + INSERT_HALF_X2(temp1, temp3, temp9, temp13) + INSERT_HALF_X2(temp6, temp8, temp11, temp15) + SHIFT_R_SUM_X2(temp9, temp10, temp11, temp12, temp13, temp14, temp15, + temp16, temp2, temp4, temp5, temp7, temp3, temp1, temp8, + temp6) + PACK_2_HALVES_TO_WORD(temp1, temp2, temp3, temp4, temp9, temp12, temp13, + temp16, temp11, temp10, temp15, temp14) + LOAD_WITH_OFFSET_X4(temp10, temp11, temp14, temp15, dst, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp5, temp6, temp7, temp8, temp17, temp18, temp10, + temp11, temp10, temp11, temp14, temp15) + STORE_SAT_SUM_X2(temp5, temp6, temp7, temp8, temp17, temp18, temp10, temp11, + temp9, temp12, temp1, temp2, temp13, temp16, temp3, temp4, + dst, 0, 1, 2, 3, BPS) + + OUTPUT_EARLY_CLOBBER_REGS_18() + : [dst]"r"(dst), [in]"r"(in), [kC1]"r"(kC1), [kC2]"r"(kC2) + : "memory", "hi", "lo" + ); +} + +static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) { + TransformOne(in, dst); + if (do_two) { + TransformOne(in + 16, dst + 4); + } +} + +static WEBP_INLINE void FilterLoop26(uint8_t* p, + int hstride, int vstride, int size, + int thresh, int ithresh, int hev_thresh) { + const int thresh2 = 2 * thresh + 1; + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15; + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "1: \n\t" + "negu %[temp1], %[hstride] \n\t" + "addiu %[size], %[size], -1 \n\t" + "sll %[temp2], %[hstride], 1 \n\t" + "sll %[temp3], %[temp1], 1 \n\t" + "addu %[temp4], %[temp2], %[hstride] \n\t" + "addu %[temp5], %[temp3], %[temp1] \n\t" + "lbu %[temp7], 0(%[p]) \n\t" + "sll %[temp6], %[temp3], 1 \n\t" + "lbux %[temp8], %[temp5](%[p]) \n\t" + "lbux %[temp9], %[temp3](%[p]) \n\t" + "lbux %[temp10], %[temp1](%[p]) \n\t" + "lbux %[temp11], %[temp6](%[p]) \n\t" + "lbux %[temp12], %[hstride](%[p]) \n\t" + "lbux %[temp13], %[temp2](%[p]) \n\t" + "lbux %[temp14], %[temp4](%[p]) \n\t" + "subu %[temp1], %[temp10], %[temp7] \n\t" + "subu %[temp2], %[temp9], %[temp12] \n\t" + "absq_s.w %[temp3], %[temp1] \n\t" + "absq_s.w %[temp4], %[temp2] \n\t" + "negu %[temp1], %[temp1] \n\t" + "sll %[temp3], %[temp3], 2 \n\t" + "addu %[temp15], %[temp3], %[temp4] \n\t" + "subu %[temp3], %[temp15], %[thresh2] \n\t" + "sll %[temp6], %[temp1], 1 \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp4], %[temp11], %[temp8] \n\t" + "absq_s.w %[temp4], %[temp4] \n\t" + "shll_s.w %[temp2], %[temp2], 24 \n\t" + "subu %[temp4], %[temp4], %[ithresh] \n\t" + "bgtz %[temp4], 3f \n\t" + " subu %[temp3], %[temp8], %[temp9] \n\t" + "absq_s.w %[temp3], %[temp3] \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp5], %[temp9], %[temp10] \n\t" + "absq_s.w %[temp3], %[temp5] \n\t" + "absq_s.w %[temp5], %[temp5] \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp3], %[temp14], %[temp13] \n\t" + "absq_s.w %[temp3], %[temp3] \n\t" + "slt %[temp5], %[hev_thresh], %[temp5] \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp3], %[temp13], %[temp12] \n\t" + "absq_s.w %[temp3], %[temp3] \n\t" + "sra %[temp4], %[temp2], 24 \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp15], %[temp12], %[temp7] \n\t" + "absq_s.w %[temp3], %[temp15] \n\t" + "absq_s.w %[temp15], %[temp15] \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " slt %[temp15], %[hev_thresh], %[temp15] \n\t" + "addu %[temp3], %[temp6], %[temp1] \n\t" + "or %[temp2], %[temp5], %[temp15] \n\t" + "addu %[temp5], %[temp4], %[temp3] \n\t" + "beqz %[temp2], 4f \n\t" + " shra_r.w %[temp1], %[temp5], 3 \n\t" + "addiu %[temp2], %[temp5], 3 \n\t" + "sra %[temp2], %[temp2], 3 \n\t" + "shll_s.w %[temp1], %[temp1], 27 \n\t" + "shll_s.w %[temp2], %[temp2], 27 \n\t" + "subu %[temp3], %[p], %[hstride] \n\t" + "sra %[temp1], %[temp1], 27 \n\t" + "sra %[temp2], %[temp2], 27 \n\t" + "subu %[temp1], %[temp7], %[temp1] \n\t" + "addu %[temp2], %[temp10], %[temp2] \n\t" + "lbux %[temp2], %[temp2](%[VP8kclip1]) \n\t" + "lbux %[temp1], %[temp1](%[VP8kclip1]) \n\t" + "sb %[temp2], 0(%[temp3]) \n\t" + "j 3f \n\t" + " sb %[temp1], 0(%[p]) \n\t" + "4: \n\t" + "shll_s.w %[temp5], %[temp5], 24 \n\t" + "subu %[temp14], %[p], %[hstride] \n\t" + "subu %[temp11], %[temp14], %[hstride] \n\t" + "sra %[temp6], %[temp5], 24 \n\t" + "sll %[temp1], %[temp6], 3 \n\t" + "subu %[temp15], %[temp11], %[hstride] \n\t" + "addu %[temp2], %[temp6], %[temp1] \n\t" + "sll %[temp3], %[temp2], 1 \n\t" + "addu %[temp4], %[temp3], %[temp2] \n\t" + "addiu %[temp2], %[temp2], 63 \n\t" + "addiu %[temp3], %[temp3], 63 \n\t" + "addiu %[temp4], %[temp4], 63 \n\t" + "sra %[temp2], %[temp2], 7 \n\t" + "sra %[temp3], %[temp3], 7 \n\t" + "sra %[temp4], %[temp4], 7 \n\t" + "addu %[temp1], %[temp8], %[temp2] \n\t" + "addu %[temp5], %[temp9], %[temp3] \n\t" + "addu %[temp6], %[temp10], %[temp4] \n\t" + "subu %[temp8], %[temp7], %[temp4] \n\t" + "subu %[temp7], %[temp12], %[temp3] \n\t" + "addu %[temp10], %[p], %[hstride] \n\t" + "subu %[temp9], %[temp13], %[temp2] \n\t" + "addu %[temp12], %[temp10], %[hstride] \n\t" + "lbux %[temp2], %[temp1](%[VP8kclip1]) \n\t" + "lbux %[temp3], %[temp5](%[VP8kclip1]) \n\t" + "lbux %[temp4], %[temp6](%[VP8kclip1]) \n\t" + "lbux %[temp5], %[temp8](%[VP8kclip1]) \n\t" + "lbux %[temp6], %[temp7](%[VP8kclip1]) \n\t" + "lbux %[temp8], %[temp9](%[VP8kclip1]) \n\t" + "sb %[temp2], 0(%[temp15]) \n\t" + "sb %[temp3], 0(%[temp11]) \n\t" + "sb %[temp4], 0(%[temp14]) \n\t" + "sb %[temp5], 0(%[p]) \n\t" + "sb %[temp6], 0(%[temp10]) \n\t" + "sb %[temp8], 0(%[temp12]) \n\t" + "3: \n\t" + "bgtz %[size], 1b \n\t" + " addu %[p], %[p], %[vstride] \n\t" + ".set pop \n\t" + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),[temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6), + [temp7]"=&r"(temp7),[temp8]"=&r"(temp8),[temp9]"=&r"(temp9), + [temp10]"=&r"(temp10),[temp11]"=&r"(temp11),[temp12]"=&r"(temp12), + [temp13]"=&r"(temp13),[temp14]"=&r"(temp14),[temp15]"=&r"(temp15), + [size]"+&r"(size), [p]"+&r"(p) + : [hstride]"r"(hstride), [thresh2]"r"(thresh2), + [ithresh]"r"(ithresh),[vstride]"r"(vstride), [hev_thresh]"r"(hev_thresh), + [VP8kclip1]"r"(VP8kclip1) + : "memory" + ); +} + +static WEBP_INLINE void FilterLoop24(uint8_t* p, + int hstride, int vstride, int size, + int thresh, int ithresh, int hev_thresh) { + int p0, q0, p1, q1, p2, q2, p3, q3; + int step1, step2, temp1, temp2, temp3, temp4; + uint8_t* pTemp0; + uint8_t* pTemp1; + const int thresh2 = 2 * thresh + 1; + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "bltz %[size], 3f \n\t" + " nop \n\t" + "2: \n\t" + "negu %[step1], %[hstride] \n\t" + "lbu %[q0], 0(%[p]) \n\t" + "lbux %[p0], %[step1](%[p]) \n\t" + "subu %[step1], %[step1], %[hstride] \n\t" + "lbux %[q1], %[hstride](%[p]) \n\t" + "subu %[temp1], %[p0], %[q0] \n\t" + "lbux %[p1], %[step1](%[p]) \n\t" + "addu %[step2], %[hstride], %[hstride] \n\t" + "absq_s.w %[temp2], %[temp1] \n\t" + "subu %[temp3], %[p1], %[q1] \n\t" + "absq_s.w %[temp4], %[temp3] \n\t" + "sll %[temp2], %[temp2], 2 \n\t" + "addu %[temp2], %[temp2], %[temp4] \n\t" + "subu %[temp4], %[temp2], %[thresh2] \n\t" + "subu %[step1], %[step1], %[hstride] \n\t" + "bgtz %[temp4], 0f \n\t" + " lbux %[p2], %[step1](%[p]) \n\t" + "subu %[step1], %[step1], %[hstride] \n\t" + "lbux %[q2], %[step2](%[p]) \n\t" + "lbux %[p3], %[step1](%[p]) \n\t" + "subu %[temp4], %[p2], %[p1] \n\t" + "addu %[step2], %[step2], %[hstride] \n\t" + "subu %[temp2], %[p3], %[p2] \n\t" + "absq_s.w %[temp4], %[temp4] \n\t" + "absq_s.w %[temp2], %[temp2] \n\t" + "lbux %[q3], %[step2](%[p]) \n\t" + "subu %[temp4], %[temp4], %[ithresh] \n\t" + "negu %[temp1], %[temp1] \n\t" + "bgtz %[temp4], 0f \n\t" + " subu %[temp2], %[temp2], %[ithresh] \n\t" + "subu %[p3], %[p1], %[p0] \n\t" + "bgtz %[temp2], 0f \n\t" + " absq_s.w %[p3], %[p3] \n\t" + "subu %[temp4], %[q3], %[q2] \n\t" + "subu %[pTemp0], %[p], %[hstride] \n\t" + "absq_s.w %[temp4], %[temp4] \n\t" + "subu %[temp2], %[p3], %[ithresh] \n\t" + "sll %[step1], %[temp1], 1 \n\t" + "bgtz %[temp2], 0f \n\t" + " subu %[temp4], %[temp4], %[ithresh] \n\t" + "subu %[temp2], %[q2], %[q1] \n\t" + "bgtz %[temp4], 0f \n\t" + " absq_s.w %[temp2], %[temp2] \n\t" + "subu %[q3], %[q1], %[q0] \n\t" + "absq_s.w %[q3], %[q3] \n\t" + "subu %[temp2], %[temp2], %[ithresh] \n\t" + "addu %[temp1], %[temp1], %[step1] \n\t" + "bgtz %[temp2], 0f \n\t" + " subu %[temp4], %[q3], %[ithresh] \n\t" + "slt %[p3], %[hev_thresh], %[p3] \n\t" + "bgtz %[temp4], 0f \n\t" + " slt %[q3], %[hev_thresh], %[q3] \n\t" + "or %[q3], %[q3], %[p3] \n\t" + "bgtz %[q3], 1f \n\t" + " shra_r.w %[temp2], %[temp1], 3 \n\t" + "addiu %[temp1], %[temp1], 3 \n\t" + "sra %[temp1], %[temp1], 3 \n\t" + "shll_s.w %[temp2], %[temp2], 27 \n\t" + "shll_s.w %[temp1], %[temp1], 27 \n\t" + "addu %[pTemp1], %[p], %[hstride] \n\t" + "sra %[temp2], %[temp2], 27 \n\t" + "sra %[temp1], %[temp1], 27 \n\t" + "addiu %[step1], %[temp2], 1 \n\t" + "sra %[step1], %[step1], 1 \n\t" + "addu %[p0], %[p0], %[temp1] \n\t" + "addu %[p1], %[p1], %[step1] \n\t" + "subu %[q0], %[q0], %[temp2] \n\t" + "subu %[q1], %[q1], %[step1] \n\t" + "lbux %[temp2], %[p0](%[VP8kclip1]) \n\t" + "lbux %[temp3], %[q0](%[VP8kclip1]) \n\t" + "lbux %[temp4], %[q1](%[VP8kclip1]) \n\t" + "sb %[temp2], 0(%[pTemp0]) \n\t" + "lbux %[temp1], %[p1](%[VP8kclip1]) \n\t" + "subu %[pTemp0], %[pTemp0], %[hstride] \n\t" + "sb %[temp3], 0(%[p]) \n\t" + "sb %[temp4], 0(%[pTemp1]) \n\t" + "j 0f \n\t" + " sb %[temp1], 0(%[pTemp0]) \n\t" + "1: \n\t" + "shll_s.w %[temp3], %[temp3], 24 \n\t" + "sra %[temp3], %[temp3], 24 \n\t" + "addu %[temp1], %[temp1], %[temp3] \n\t" + "shra_r.w %[temp2], %[temp1], 3 \n\t" + "addiu %[temp1], %[temp1], 3 \n\t" + "shll_s.w %[temp2], %[temp2], 27 \n\t" + "sra %[temp1], %[temp1], 3 \n\t" + "shll_s.w %[temp1], %[temp1], 27 \n\t" + "sra %[temp2], %[temp2], 27 \n\t" + "sra %[temp1], %[temp1], 27 \n\t" + "addu %[p0], %[p0], %[temp1] \n\t" + "subu %[q0], %[q0], %[temp2] \n\t" + "lbux %[temp1], %[p0](%[VP8kclip1]) \n\t" + "lbux %[temp2], %[q0](%[VP8kclip1]) \n\t" + "sb %[temp2], 0(%[p]) \n\t" + "sb %[temp1], 0(%[pTemp0]) \n\t" + "0: \n\t" + "subu %[size], %[size], 1 \n\t" + "bgtz %[size], 2b \n\t" + " addu %[p], %[p], %[vstride] \n\t" + "3: \n\t" + ".set pop \n\t" + : [p0]"=&r"(p0), [q0]"=&r"(q0), [p1]"=&r"(p1), [q1]"=&r"(q1), + [p2]"=&r"(p2), [q2]"=&r"(q2), [p3]"=&r"(p3), [q3]"=&r"(q3), + [step2]"=&r"(step2), [step1]"=&r"(step1), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), + [pTemp0]"=&r"(pTemp0), [pTemp1]"=&r"(pTemp1), [p]"+&r"(p), + [size]"+&r"(size) + : [vstride]"r"(vstride), [ithresh]"r"(ithresh), + [hev_thresh]"r"(hev_thresh), [hstride]"r"(hstride), + [VP8kclip1]"r"(VP8kclip1), [thresh2]"r"(thresh2) + : "memory" + ); +} + +// on macroblock edges +static void VFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(p, stride, 1, 16, thresh, ithresh, hev_thresh); +} + +static void HFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(p, 1, stride, 16, thresh, ithresh, hev_thresh); +} + +// 8-pixels wide variant, for chroma filtering +static void VFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(u, stride, 1, 8, thresh, ithresh, hev_thresh); + FilterLoop26(v, stride, 1, 8, thresh, ithresh, hev_thresh); +} + +static void HFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(u, 1, stride, 8, thresh, ithresh, hev_thresh); + FilterLoop26(v, 1, stride, 8, thresh, ithresh, hev_thresh); +} + +// on three inner edges +static void VFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4 * stride; + FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh); + } +} + +static void HFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4; + FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh); + } +} + +static void VFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh); + FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh); +} + +static void HFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh); + FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh); +} + +#undef MUL + +//------------------------------------------------------------------------------ +// Simple In-loop filtering (Paragraph 15.2) + +static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { + int i; + const int thresh2 = 2 * thresh + 1; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + uint8_t* p1 = p - stride; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "li %[i], 16 \n\t" + "0: \n\t" + "negu %[temp4], %[stride] \n\t" + "sll %[temp5], %[temp4], 1 \n\t" + "lbu %[temp2], 0(%[p]) \n\t" + "lbux %[temp3], %[stride](%[p]) \n\t" + "lbux %[temp1], %[temp4](%[p]) \n\t" + "lbux %[temp0], %[temp5](%[p]) \n\t" + "subu %[temp7], %[temp1], %[temp2] \n\t" + "subu %[temp6], %[temp0], %[temp3] \n\t" + "absq_s.w %[temp4], %[temp7] \n\t" + "absq_s.w %[temp5], %[temp6] \n\t" + "sll %[temp4], %[temp4], 2 \n\t" + "subu %[temp5], %[temp5], %[thresh2] \n\t" + "addu %[temp5], %[temp4], %[temp5] \n\t" + "negu %[temp8], %[temp7] \n\t" + "bgtz %[temp5], 1f \n\t" + " addiu %[i], %[i], -1 \n\t" + "sll %[temp4], %[temp8], 1 \n\t" + "shll_s.w %[temp5], %[temp6], 24 \n\t" + "addu %[temp3], %[temp4], %[temp8] \n\t" + "sra %[temp5], %[temp5], 24 \n\t" + "addu %[temp3], %[temp3], %[temp5] \n\t" + "addiu %[temp7], %[temp3], 3 \n\t" + "sra %[temp7], %[temp7], 3 \n\t" + "shra_r.w %[temp8], %[temp3], 3 \n\t" + "shll_s.w %[temp0], %[temp7], 27 \n\t" + "shll_s.w %[temp4], %[temp8], 27 \n\t" + "sra %[temp0], %[temp0], 27 \n\t" + "sra %[temp4], %[temp4], 27 \n\t" + "addu %[temp7], %[temp1], %[temp0] \n\t" + "subu %[temp2], %[temp2], %[temp4] \n\t" + "lbux %[temp3], %[temp7](%[VP8kclip1]) \n\t" + "lbux %[temp4], %[temp2](%[VP8kclip1]) \n\t" + "sb %[temp3], 0(%[p1]) \n\t" + "sb %[temp4], 0(%[p]) \n\t" + "1: \n\t" + "addiu %[p1], %[p1], 1 \n\t" + "bgtz %[i], 0b \n\t" + " addiu %[p], %[p], 1 \n\t" + " .set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [p]"+&r"(p), [i]"=&r"(i), [p1]"+&r"(p1) + : [stride]"r"(stride), [VP8kclip1]"r"(VP8kclip1), [thresh2]"r"(thresh2) + : "memory" + ); +} + +// TEMP0 = SRC[A + A1 * BPS] +// TEMP1 = SRC[B + B1 * BPS] +// TEMP2 = SRC[C + C1 * BPS] +// TEMP3 = SRC[D + D1 * BPS] +#define LOAD_4_BYTES(TEMP0, TEMP1, TEMP2, TEMP3, \ + A, A1, B, B1, C, C1, D, D1, SRC) \ + "lbu %[" #TEMP0 "], " #A "+" #A1 "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + "lbu %[" #TEMP1 "], " #B "+" #B1 "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + "lbu %[" #TEMP2 "], " #C "+" #C1 "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + "lbu %[" #TEMP3 "], " #D "+" #D1 "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + +static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { + int i; + const int thresh2 = 2 * thresh + 1; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "li %[i], 16 \n\t" + "0: \n\t" + LOAD_4_BYTES(temp0, temp1, temp2, temp3, -2, 0, -1, 0, 0, 0, 1, 0, p) + "subu %[temp7], %[temp1], %[temp2] \n\t" + "subu %[temp6], %[temp0], %[temp3] \n\t" + "absq_s.w %[temp4], %[temp7] \n\t" + "absq_s.w %[temp5], %[temp6] \n\t" + "sll %[temp4], %[temp4], 2 \n\t" + "addu %[temp5], %[temp4], %[temp5] \n\t" + "subu %[temp5], %[temp5], %[thresh2] \n\t" + "negu %[temp8], %[temp7] \n\t" + "bgtz %[temp5], 1f \n\t" + " addiu %[i], %[i], -1 \n\t" + "sll %[temp4], %[temp8], 1 \n\t" + "shll_s.w %[temp5], %[temp6], 24 \n\t" + "addu %[temp3], %[temp4], %[temp8] \n\t" + "sra %[temp5], %[temp5], 24 \n\t" + "addu %[temp3], %[temp3], %[temp5] \n\t" + "addiu %[temp7], %[temp3], 3 \n\t" + "sra %[temp7], %[temp7], 3 \n\t" + "shra_r.w %[temp8], %[temp3], 3 \n\t" + "shll_s.w %[temp0], %[temp7], 27 \n\t" + "shll_s.w %[temp4], %[temp8], 27 \n\t" + "sra %[temp0], %[temp0], 27 \n\t" + "sra %[temp4], %[temp4], 27 \n\t" + "addu %[temp7], %[temp1], %[temp0] \n\t" + "subu %[temp2], %[temp2], %[temp4] \n\t" + "lbux %[temp3], %[temp7](%[VP8kclip1]) \n\t" + "lbux %[temp4], %[temp2](%[VP8kclip1]) \n\t" + "sb %[temp3], -1(%[p]) \n\t" + "sb %[temp4], 0(%[p]) \n\t" + "1: \n\t" + "bgtz %[i], 0b \n\t" + " addu %[p], %[p], %[stride] \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [p]"+&r"(p), [i]"=&r"(i) + : [stride]"r"(stride), [VP8kclip1]"r"(VP8kclip1), [thresh2]"r"(thresh2) + : "memory" + ); +} + +static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4 * stride; + SimpleVFilter16(p, stride, thresh); + } +} + +static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4; + SimpleHFilter16(p, stride, thresh); + } +} + +// DST[A * BPS] = TEMP0 +// DST[B + C * BPS] = TEMP1 +#define STORE_8_BYTES(TEMP0, TEMP1, A, B, C, DST) \ + "usw %[" #TEMP0 "], " #A "*" XSTR(BPS) "(%[" #DST "]) \n\t" \ + "usw %[" #TEMP1 "], " #B "+" #C "*" XSTR(BPS) "(%[" #DST "]) \n\t" + +static void VE4(uint8_t* dst) { // vertical + const uint8_t* top = dst - BPS; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + __asm__ volatile ( + "ulw %[temp0], -1(%[top]) \n\t" + "ulh %[temp1], 3(%[top]) \n\t" + "preceu.ph.qbr %[temp2], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "packrl.ph %[temp5], %[temp3], %[temp2] \n\t" + "packrl.ph %[temp6], %[temp4], %[temp3] \n\t" + "shll.ph %[temp5], %[temp5], 1 \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp2], %[temp5], %[temp2] \n\t" + "addq.ph %[temp6], %[temp6], %[temp4] \n\t" + "addq.ph %[temp2], %[temp2], %[temp3] \n\t" + "addq.ph %[temp6], %[temp6], %[temp3] \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precr.qb.ph %[temp4], %[temp6], %[temp2] \n\t" + STORE_8_BYTES(temp4, temp4, 0, 0, 1, dst) + STORE_8_BYTES(temp4, temp4, 2, 0, 3, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void DC4(uint8_t* dst) { // DC + int temp0, temp1, temp2, temp3, temp4; + __asm__ volatile ( + "ulw %[temp0], -1*" XSTR(BPS) "(%[dst]) \n\t" + LOAD_4_BYTES(temp1, temp2, temp3, temp4, -1, 0, -1, 1, -1, 2, -1, 3, dst) + "ins %[temp1], %[temp2], 8, 8 \n\t" + "ins %[temp1], %[temp3], 16, 8 \n\t" + "ins %[temp1], %[temp4], 24, 8 \n\t" + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "shra_r.w %[temp0], %[temp0], 3 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + STORE_8_BYTES(temp0, temp0, 0, 0, 1, dst) + STORE_8_BYTES(temp0, temp0, 2, 0, 3, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4) + : [dst]"r"(dst) + : "memory" + ); +} + +static void RD4(uint8_t* dst) { // Down-right + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8; + __asm__ volatile ( + LOAD_4_BYTES(temp0, temp1, temp2, temp3, -1, 0, -1, 1, -1, 2, -1, 3, dst) + "ulw %[temp7], -1-" XSTR(BPS) "(%[dst]) \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "preceu.ph.qbr %[temp5], %[temp7] \n\t" + "ins %[temp2], %[temp1], 16, 16 \n\t" + "preceu.ph.qbl %[temp4], %[temp7] \n\t" + "ins %[temp3], %[temp2], 16, 16 \n\t" + "shll.ph %[temp2], %[temp2], 1 \n\t" + "addq.ph %[temp3], %[temp3], %[temp1] \n\t" + "packrl.ph %[temp6], %[temp5], %[temp1] \n\t" + "addq.ph %[temp3], %[temp3], %[temp2] \n\t" + "addq.ph %[temp1], %[temp1], %[temp5] \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp1], %[temp1], %[temp6] \n\t" + "packrl.ph %[temp0], %[temp4], %[temp5] \n\t" + "addq.ph %[temp8], %[temp5], %[temp4] \n\t" + "shra_r.ph %[temp3], %[temp3], 2 \n\t" + "shll.ph %[temp0], %[temp0], 1 \n\t" + "shra_r.ph %[temp1], %[temp1], 2 \n\t" + "addq.ph %[temp8], %[temp0], %[temp8] \n\t" + "lbu %[temp5], 3-" XSTR(BPS) "(%[dst]) \n\t" + "precrq.ph.w %[temp7], %[temp7], %[temp7] \n\t" + "shra_r.ph %[temp8], %[temp8], 2 \n\t" + "ins %[temp7], %[temp5], 0, 8 \n\t" + "precr.qb.ph %[temp2], %[temp1], %[temp3] \n\t" + "raddu.w.qb %[temp4], %[temp7] \n\t" + "precr.qb.ph %[temp6], %[temp8], %[temp1] \n\t" + "shra_r.w %[temp4], %[temp4], 2 \n\t" + STORE_8_BYTES(temp2, temp6, 3, 0, 1, dst) + "prepend %[temp2], %[temp8], 8 \n\t" + "prepend %[temp6], %[temp4], 8 \n\t" + STORE_8_BYTES(temp2, temp6, 2, 0, 0, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8) + : [dst]"r"(dst) + : "memory" + ); +} + +// TEMP0 = SRC[A * BPS] +// TEMP1 = SRC[B + C * BPS] +#define LOAD_8_BYTES(TEMP0, TEMP1, A, B, C, SRC) \ + "ulw %[" #TEMP0 "], " #A "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + "ulw %[" #TEMP1 "], " #B "+" #C "*" XSTR(BPS) "(%[" #SRC "]) \n\t" + +static void LD4(uint8_t* dst) { // Down-Left + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + LOAD_8_BYTES(temp0, temp1, -1, 4, -1, dst) + "preceu.ph.qbl %[temp2], %[temp0] \n\t" + "preceu.ph.qbr %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "preceu.ph.qbl %[temp5], %[temp1] \n\t" + "packrl.ph %[temp6], %[temp2], %[temp3] \n\t" + "packrl.ph %[temp7], %[temp4], %[temp2] \n\t" + "packrl.ph %[temp8], %[temp5], %[temp4] \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp9], %[temp2], %[temp6] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "addq.ph %[temp9], %[temp9], %[temp3] \n\t" + "shll.ph %[temp8], %[temp8], 1 \n\t" + "shra_r.ph %[temp9], %[temp9], 2 \n\t" + "addq.ph %[temp3], %[temp4], %[temp7] \n\t" + "addq.ph %[temp0], %[temp5], %[temp8] \n\t" + "addq.ph %[temp3], %[temp3], %[temp2] \n\t" + "addq.ph %[temp0], %[temp0], %[temp4] \n\t" + "shra_r.ph %[temp3], %[temp3], 2 \n\t" + "shra_r.ph %[temp0], %[temp0], 2 \n\t" + "srl %[temp1], %[temp1], 24 \n\t" + "sll %[temp1], %[temp1], 1 \n\t" + "raddu.w.qb %[temp5], %[temp5] \n\t" + "precr.qb.ph %[temp9], %[temp3], %[temp9] \n\t" + "precr.qb.ph %[temp3], %[temp0], %[temp3] \n\t" + "addu %[temp1], %[temp1], %[temp5] \n\t" + "shra_r.w %[temp1], %[temp1], 2 \n\t" + STORE_8_BYTES(temp9, temp3, 0, 0, 2, dst) + "prepend %[temp9], %[temp0], 8 \n\t" + "prepend %[temp3], %[temp1], 8 \n\t" + STORE_8_BYTES(temp9, temp3, 1, 0, 3, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [dst]"r"(dst) + : "memory" + ); +} + +//------------------------------------------------------------------------------ +// Chroma + +static void DC8uv(uint8_t* dst) { // DC + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + LOAD_8_BYTES(temp0, temp1, -1, 4, -1, dst) + LOAD_4_BYTES(temp2, temp3, temp4, temp5, -1, 0, -1, 1, -1, 2, -1, 3, dst) + LOAD_4_BYTES(temp6, temp7, temp8, temp9, -1, 4, -1, 5, -1, 6, -1, 7, dst) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[temp4], %[temp4], %[temp5] \n\t" + "addu %[temp6], %[temp6], %[temp7] \n\t" + "addu %[temp8], %[temp8], %[temp9] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp4] \n\t" + "addu %[temp6], %[temp6], %[temp8] \n\t" + "addu %[temp0], %[temp0], %[temp2] \n\t" + "addu %[temp0], %[temp0], %[temp6] \n\t" + "shra_r.w %[temp0], %[temp0], 4 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + STORE_8_BYTES(temp0, temp0, 0, 4, 0, dst) + STORE_8_BYTES(temp0, temp0, 1, 4, 1, dst) + STORE_8_BYTES(temp0, temp0, 2, 4, 2, dst) + STORE_8_BYTES(temp0, temp0, 3, 4, 3, dst) + STORE_8_BYTES(temp0, temp0, 4, 4, 4, dst) + STORE_8_BYTES(temp0, temp0, 5, 4, 5, dst) + STORE_8_BYTES(temp0, temp0, 6, 4, 6, dst) + STORE_8_BYTES(temp0, temp0, 7, 4, 7, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [dst]"r"(dst) + : "memory" + ); +} + +static void DC8uvNoLeft(uint8_t* dst) { // DC with no left samples + int temp0, temp1; + __asm__ volatile ( + LOAD_8_BYTES(temp0, temp1, -1, 4, -1, dst) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "shra_r.w %[temp0], %[temp0], 3 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + STORE_8_BYTES(temp0, temp0, 0, 4, 0, dst) + STORE_8_BYTES(temp0, temp0, 1, 4, 1, dst) + STORE_8_BYTES(temp0, temp0, 2, 4, 2, dst) + STORE_8_BYTES(temp0, temp0, 3, 4, 3, dst) + STORE_8_BYTES(temp0, temp0, 4, 4, 4, dst) + STORE_8_BYTES(temp0, temp0, 5, 4, 5, dst) + STORE_8_BYTES(temp0, temp0, 6, 4, 6, dst) + STORE_8_BYTES(temp0, temp0, 7, 4, 7, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1) + : [dst]"r"(dst) + : "memory" + ); +} + +static void DC8uvNoTop(uint8_t* dst) { // DC with no top samples + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8; + __asm__ volatile ( + LOAD_4_BYTES(temp2, temp3, temp4, temp5, -1, 0, -1, 1, -1, 2, -1, 3, dst) + LOAD_4_BYTES(temp6, temp7, temp8, temp1, -1, 4, -1, 5, -1, 6, -1, 7, dst) + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[temp4], %[temp4], %[temp5] \n\t" + "addu %[temp6], %[temp6], %[temp7] \n\t" + "addu %[temp8], %[temp8], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp4] \n\t" + "addu %[temp6], %[temp6], %[temp8] \n\t" + "addu %[temp0], %[temp6], %[temp2] \n\t" + "shra_r.w %[temp0], %[temp0], 3 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + STORE_8_BYTES(temp0, temp0, 0, 4, 0, dst) + STORE_8_BYTES(temp0, temp0, 1, 4, 1, dst) + STORE_8_BYTES(temp0, temp0, 2, 4, 2, dst) + STORE_8_BYTES(temp0, temp0, 3, 4, 3, dst) + STORE_8_BYTES(temp0, temp0, 4, 4, 4, dst) + STORE_8_BYTES(temp0, temp0, 5, 4, 5, dst) + STORE_8_BYTES(temp0, temp0, 6, 4, 6, dst) + STORE_8_BYTES(temp0, temp0, 7, 4, 7, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8) + : [dst]"r"(dst) + : "memory" + ); +} + +#undef LOAD_8_BYTES +#undef STORE_8_BYTES +#undef LOAD_4_BYTES + +#define CLIPPING(SIZE) \ + "preceu.ph.qbl %[temp2], %[temp0] \n\t" \ + "preceu.ph.qbr %[temp0], %[temp0] \n\t" \ +".if " #SIZE " == 8 \n\t" \ + "preceu.ph.qbl %[temp3], %[temp1] \n\t" \ + "preceu.ph.qbr %[temp1], %[temp1] \n\t" \ +".endif \n\t" \ + "addu.ph %[temp2], %[temp2], %[dst_1] \n\t" \ + "addu.ph %[temp0], %[temp0], %[dst_1] \n\t" \ +".if " #SIZE " == 8 \n\t" \ + "addu.ph %[temp3], %[temp3], %[dst_1] \n\t" \ + "addu.ph %[temp1], %[temp1], %[dst_1] \n\t" \ +".endif \n\t" \ + "shll_s.ph %[temp2], %[temp2], 7 \n\t" \ + "shll_s.ph %[temp0], %[temp0], 7 \n\t" \ +".if " #SIZE " == 8 \n\t" \ + "shll_s.ph %[temp3], %[temp3], 7 \n\t" \ + "shll_s.ph %[temp1], %[temp1], 7 \n\t" \ +".endif \n\t" \ + "precrqu_s.qb.ph %[temp0], %[temp2], %[temp0] \n\t" \ +".if " #SIZE " == 8 \n\t" \ + "precrqu_s.qb.ph %[temp1], %[temp3], %[temp1] \n\t" \ +".endif \n\t" + + +#define CLIP_8B_TO_DST(DST, TOP, SIZE) do { \ + int dst_1 = ((int)(DST)[-1] << 16) + (DST)[-1]; \ + int temp0, temp1, temp2, temp3; \ + __asm__ volatile ( \ + ".if " #SIZE " < 8 \n\t" \ + "ulw %[temp0], 0(%[top]) \n\t" \ + "subu.ph %[dst_1], %[dst_1], %[top_1] \n\t" \ + CLIPPING(4) \ + "usw %[temp0], 0(%[dst]) \n\t" \ + ".else \n\t" \ + "ulw %[temp0], 0(%[top]) \n\t" \ + "ulw %[temp1], 4(%[top]) \n\t" \ + "subu.ph %[dst_1], %[dst_1], %[top_1] \n\t" \ + CLIPPING(8) \ + "usw %[temp0], 0(%[dst]) \n\t" \ + "usw %[temp1], 4(%[dst]) \n\t" \ + ".if " #SIZE " == 16 \n\t" \ + "ulw %[temp0], 8(%[top]) \n\t" \ + "ulw %[temp1], 12(%[top]) \n\t" \ + CLIPPING(8) \ + "usw %[temp0], 8(%[dst]) \n\t" \ + "usw %[temp1], 12(%[dst]) \n\t" \ + ".endif \n\t" \ + ".endif \n\t" \ + : [dst_1]"+&r"(dst_1), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), \ + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3) \ + : [top_1]"r"(top_1), [top]"r"((TOP)), [dst]"r"((DST)) \ + : "memory" \ + ); \ +} while (0) + +#define CLIP_TO_DST(DST, SIZE) do { \ + int y; \ + const uint8_t* top = (DST) - BPS; \ + const int top_1 = ((int)top[-1] << 16) + top[-1]; \ + for (y = 0; y < (SIZE); ++y) { \ + CLIP_8B_TO_DST((DST), top, (SIZE)); \ + (DST) += BPS; \ + } \ +} while (0) + +#define TRUE_MOTION(DST, SIZE) \ +static void TrueMotion##SIZE(uint8_t* (DST)) { \ + CLIP_TO_DST((DST), (SIZE)); \ +} + +TRUE_MOTION(dst, 4) +TRUE_MOTION(dst, 8) +TRUE_MOTION(dst, 16) + +#undef TRUE_MOTION +#undef CLIP_TO_DST +#undef CLIP_8B_TO_DST +#undef CLIPPING + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8DspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitMIPSdspR2(void) { + VP8TransformDC = TransformDC; + VP8TransformAC3 = TransformAC3; + VP8Transform = TransformTwo; + + VP8VFilter16 = VFilter16; + VP8HFilter16 = HFilter16; + VP8VFilter8 = VFilter8; + VP8HFilter8 = HFilter8; + VP8VFilter16i = VFilter16i; + VP8HFilter16i = HFilter16i; + VP8VFilter8i = VFilter8i; + VP8HFilter8i = HFilter8i; + VP8SimpleVFilter16 = SimpleVFilter16; + VP8SimpleHFilter16 = SimpleHFilter16; + VP8SimpleVFilter16i = SimpleVFilter16i; + VP8SimpleHFilter16i = SimpleHFilter16i; + + VP8PredLuma4[0] = DC4; + VP8PredLuma4[1] = TrueMotion4; + VP8PredLuma4[2] = VE4; + VP8PredLuma4[4] = RD4; + VP8PredLuma4[6] = LD4; + + VP8PredChroma8[0] = DC8uv; + VP8PredChroma8[1] = TrueMotion8; + VP8PredChroma8[4] = DC8uvNoTop; + VP8PredChroma8[5] = DC8uvNoLeft; + + VP8PredLuma16[1] = TrueMotion16; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8DspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/dec_neon.c b/drivers/webp/dsp/dec_neon.c index ec824b790b..a63f43fe17 100644 --- a/drivers/webp/dsp/dec_neon.c +++ b/drivers/webp/dsp/dec_neon.c @@ -1,8 +1,10 @@ // Copyright 2012 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // ARM NEON version of dsp functions and loop filtering. @@ -14,13 +16,535 @@ #if defined(WEBP_USE_NEON) +#include "./neon.h" #include "../dec/vp8i.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif +//------------------------------------------------------------------------------ +// NxM Loading functions + +// Load/Store vertical edge +#define LOAD8x4(c1, c2, c3, c4, b1, b2, stride) \ + "vld4.8 {" #c1 "[0]," #c2 "[0]," #c3 "[0]," #c4 "[0]}," #b1 "," #stride "\n" \ + "vld4.8 {" #c1 "[1]," #c2 "[1]," #c3 "[1]," #c4 "[1]}," #b2 "," #stride "\n" \ + "vld4.8 {" #c1 "[2]," #c2 "[2]," #c3 "[2]," #c4 "[2]}," #b1 "," #stride "\n" \ + "vld4.8 {" #c1 "[3]," #c2 "[3]," #c3 "[3]," #c4 "[3]}," #b2 "," #stride "\n" \ + "vld4.8 {" #c1 "[4]," #c2 "[4]," #c3 "[4]," #c4 "[4]}," #b1 "," #stride "\n" \ + "vld4.8 {" #c1 "[5]," #c2 "[5]," #c3 "[5]," #c4 "[5]}," #b2 "," #stride "\n" \ + "vld4.8 {" #c1 "[6]," #c2 "[6]," #c3 "[6]," #c4 "[6]}," #b1 "," #stride "\n" \ + "vld4.8 {" #c1 "[7]," #c2 "[7]," #c3 "[7]," #c4 "[7]}," #b2 "," #stride "\n" + +#define STORE8x2(c1, c2, p, stride) \ + "vst2.8 {" #c1 "[0], " #c2 "[0]}," #p "," #stride " \n" \ + "vst2.8 {" #c1 "[1], " #c2 "[1]}," #p "," #stride " \n" \ + "vst2.8 {" #c1 "[2], " #c2 "[2]}," #p "," #stride " \n" \ + "vst2.8 {" #c1 "[3], " #c2 "[3]}," #p "," #stride " \n" \ + "vst2.8 {" #c1 "[4], " #c2 "[4]}," #p "," #stride " \n" \ + "vst2.8 {" #c1 "[5], " #c2 "[5]}," #p "," #stride " \n" \ + "vst2.8 {" #c1 "[6], " #c2 "[6]}," #p "," #stride " \n" \ + "vst2.8 {" #c1 "[7], " #c2 "[7]}," #p "," #stride " \n" + +#if !defined(WORK_AROUND_GCC) + +// This intrinsics version makes gcc-4.6.3 crash during Load4x??() compilation +// (register alloc, probably). The variants somewhat mitigate the problem, but +// not quite. HFilter16i() remains problematic. +static WEBP_INLINE uint8x8x4_t Load4x8(const uint8_t* const src, int stride) { + const uint8x8_t zero = vdup_n_u8(0); + uint8x8x4_t out; + INIT_VECTOR4(out, zero, zero, zero, zero); + out = vld4_lane_u8(src + 0 * stride, out, 0); + out = vld4_lane_u8(src + 1 * stride, out, 1); + out = vld4_lane_u8(src + 2 * stride, out, 2); + out = vld4_lane_u8(src + 3 * stride, out, 3); + out = vld4_lane_u8(src + 4 * stride, out, 4); + out = vld4_lane_u8(src + 5 * stride, out, 5); + out = vld4_lane_u8(src + 6 * stride, out, 6); + out = vld4_lane_u8(src + 7 * stride, out, 7); + return out; +} + +static WEBP_INLINE void Load4x16(const uint8_t* const src, int stride, + uint8x16_t* const p1, uint8x16_t* const p0, + uint8x16_t* const q0, uint8x16_t* const q1) { + // row0 = p1[0..7]|p0[0..7]|q0[0..7]|q1[0..7] + // row8 = p1[8..15]|p0[8..15]|q0[8..15]|q1[8..15] + const uint8x8x4_t row0 = Load4x8(src - 2 + 0 * stride, stride); + const uint8x8x4_t row8 = Load4x8(src - 2 + 8 * stride, stride); + *p1 = vcombine_u8(row0.val[0], row8.val[0]); + *p0 = vcombine_u8(row0.val[1], row8.val[1]); + *q0 = vcombine_u8(row0.val[2], row8.val[2]); + *q1 = vcombine_u8(row0.val[3], row8.val[3]); +} + +#else // WORK_AROUND_GCC + +#define LOADQ_LANE_32b(VALUE, LANE) do { \ + (VALUE) = vld1q_lane_u32((const uint32_t*)src, (VALUE), (LANE)); \ + src += stride; \ +} while (0) + +static WEBP_INLINE void Load4x16(const uint8_t* src, int stride, + uint8x16_t* const p1, uint8x16_t* const p0, + uint8x16_t* const q0, uint8x16_t* const q1) { + const uint32x4_t zero = vdupq_n_u32(0); + uint32x4x4_t in; + INIT_VECTOR4(in, zero, zero, zero, zero); + src -= 2; + LOADQ_LANE_32b(in.val[0], 0); + LOADQ_LANE_32b(in.val[1], 0); + LOADQ_LANE_32b(in.val[2], 0); + LOADQ_LANE_32b(in.val[3], 0); + LOADQ_LANE_32b(in.val[0], 1); + LOADQ_LANE_32b(in.val[1], 1); + LOADQ_LANE_32b(in.val[2], 1); + LOADQ_LANE_32b(in.val[3], 1); + LOADQ_LANE_32b(in.val[0], 2); + LOADQ_LANE_32b(in.val[1], 2); + LOADQ_LANE_32b(in.val[2], 2); + LOADQ_LANE_32b(in.val[3], 2); + LOADQ_LANE_32b(in.val[0], 3); + LOADQ_LANE_32b(in.val[1], 3); + LOADQ_LANE_32b(in.val[2], 3); + LOADQ_LANE_32b(in.val[3], 3); + // Transpose four 4x4 parts: + { + const uint8x16x2_t row01 = vtrnq_u8(vreinterpretq_u8_u32(in.val[0]), + vreinterpretq_u8_u32(in.val[1])); + const uint8x16x2_t row23 = vtrnq_u8(vreinterpretq_u8_u32(in.val[2]), + vreinterpretq_u8_u32(in.val[3])); + const uint16x8x2_t row02 = vtrnq_u16(vreinterpretq_u16_u8(row01.val[0]), + vreinterpretq_u16_u8(row23.val[0])); + const uint16x8x2_t row13 = vtrnq_u16(vreinterpretq_u16_u8(row01.val[1]), + vreinterpretq_u16_u8(row23.val[1])); + *p1 = vreinterpretq_u8_u16(row02.val[0]); + *p0 = vreinterpretq_u8_u16(row13.val[0]); + *q0 = vreinterpretq_u8_u16(row02.val[1]); + *q1 = vreinterpretq_u8_u16(row13.val[1]); + } +} +#undef LOADQ_LANE_32b + +#endif // !WORK_AROUND_GCC + +static WEBP_INLINE void Load8x16(const uint8_t* const src, int stride, + uint8x16_t* const p3, uint8x16_t* const p2, + uint8x16_t* const p1, uint8x16_t* const p0, + uint8x16_t* const q0, uint8x16_t* const q1, + uint8x16_t* const q2, uint8x16_t* const q3) { + Load4x16(src - 2, stride, p3, p2, p1, p0); + Load4x16(src + 2, stride, q0, q1, q2, q3); +} + +static WEBP_INLINE void Load16x4(const uint8_t* const src, int stride, + uint8x16_t* const p1, uint8x16_t* const p0, + uint8x16_t* const q0, uint8x16_t* const q1) { + *p1 = vld1q_u8(src - 2 * stride); + *p0 = vld1q_u8(src - 1 * stride); + *q0 = vld1q_u8(src + 0 * stride); + *q1 = vld1q_u8(src + 1 * stride); +} + +static WEBP_INLINE void Load16x8(const uint8_t* const src, int stride, + uint8x16_t* const p3, uint8x16_t* const p2, + uint8x16_t* const p1, uint8x16_t* const p0, + uint8x16_t* const q0, uint8x16_t* const q1, + uint8x16_t* const q2, uint8x16_t* const q3) { + Load16x4(src - 2 * stride, stride, p3, p2, p1, p0); + Load16x4(src + 2 * stride, stride, q0, q1, q2, q3); +} + +static WEBP_INLINE void Load8x8x2(const uint8_t* const u, + const uint8_t* const v, + int stride, + uint8x16_t* const p3, uint8x16_t* const p2, + uint8x16_t* const p1, uint8x16_t* const p0, + uint8x16_t* const q0, uint8x16_t* const q1, + uint8x16_t* const q2, uint8x16_t* const q3) { + // We pack the 8x8 u-samples in the lower half of the uint8x16_t destination + // and the v-samples on the higher half. + *p3 = vcombine_u8(vld1_u8(u - 4 * stride), vld1_u8(v - 4 * stride)); + *p2 = vcombine_u8(vld1_u8(u - 3 * stride), vld1_u8(v - 3 * stride)); + *p1 = vcombine_u8(vld1_u8(u - 2 * stride), vld1_u8(v - 2 * stride)); + *p0 = vcombine_u8(vld1_u8(u - 1 * stride), vld1_u8(v - 1 * stride)); + *q0 = vcombine_u8(vld1_u8(u + 0 * stride), vld1_u8(v + 0 * stride)); + *q1 = vcombine_u8(vld1_u8(u + 1 * stride), vld1_u8(v + 1 * stride)); + *q2 = vcombine_u8(vld1_u8(u + 2 * stride), vld1_u8(v + 2 * stride)); + *q3 = vcombine_u8(vld1_u8(u + 3 * stride), vld1_u8(v + 3 * stride)); +} + +#if !defined(WORK_AROUND_GCC) + +#define LOAD_UV_8(ROW) \ + vcombine_u8(vld1_u8(u - 4 + (ROW) * stride), vld1_u8(v - 4 + (ROW) * stride)) + +static WEBP_INLINE void Load8x8x2T(const uint8_t* const u, + const uint8_t* const v, + int stride, + uint8x16_t* const p3, uint8x16_t* const p2, + uint8x16_t* const p1, uint8x16_t* const p0, + uint8x16_t* const q0, uint8x16_t* const q1, + uint8x16_t* const q2, uint8x16_t* const q3) { + // We pack the 8x8 u-samples in the lower half of the uint8x16_t destination + // and the v-samples on the higher half. + const uint8x16_t row0 = LOAD_UV_8(0); + const uint8x16_t row1 = LOAD_UV_8(1); + const uint8x16_t row2 = LOAD_UV_8(2); + const uint8x16_t row3 = LOAD_UV_8(3); + const uint8x16_t row4 = LOAD_UV_8(4); + const uint8x16_t row5 = LOAD_UV_8(5); + const uint8x16_t row6 = LOAD_UV_8(6); + const uint8x16_t row7 = LOAD_UV_8(7); + // Perform two side-by-side 8x8 transposes + // u00 u01 u02 u03 u04 u05 u06 u07 | v00 v01 v02 v03 v04 v05 v06 v07 + // u10 u11 u12 u13 u14 u15 u16 u17 | v10 v11 v12 ... + // u20 u21 u22 u23 u24 u25 u26 u27 | v20 v21 ... + // u30 u31 u32 u33 u34 u35 u36 u37 | ... + // u40 u41 u42 u43 u44 u45 u46 u47 | ... + // u50 u51 u52 u53 u54 u55 u56 u57 | ... + // u60 u61 u62 u63 u64 u65 u66 u67 | v60 ... + // u70 u71 u72 u73 u74 u75 u76 u77 | v70 v71 v72 ... + const uint8x16x2_t row01 = vtrnq_u8(row0, row1); // u00 u10 u02 u12 ... + // u01 u11 u03 u13 ... + const uint8x16x2_t row23 = vtrnq_u8(row2, row3); // u20 u30 u22 u32 ... + // u21 u31 u23 u33 ... + const uint8x16x2_t row45 = vtrnq_u8(row4, row5); // ... + const uint8x16x2_t row67 = vtrnq_u8(row6, row7); // ... + const uint16x8x2_t row02 = vtrnq_u16(vreinterpretq_u16_u8(row01.val[0]), + vreinterpretq_u16_u8(row23.val[0])); + const uint16x8x2_t row13 = vtrnq_u16(vreinterpretq_u16_u8(row01.val[1]), + vreinterpretq_u16_u8(row23.val[1])); + const uint16x8x2_t row46 = vtrnq_u16(vreinterpretq_u16_u8(row45.val[0]), + vreinterpretq_u16_u8(row67.val[0])); + const uint16x8x2_t row57 = vtrnq_u16(vreinterpretq_u16_u8(row45.val[1]), + vreinterpretq_u16_u8(row67.val[1])); + const uint32x4x2_t row04 = vtrnq_u32(vreinterpretq_u32_u16(row02.val[0]), + vreinterpretq_u32_u16(row46.val[0])); + const uint32x4x2_t row26 = vtrnq_u32(vreinterpretq_u32_u16(row02.val[1]), + vreinterpretq_u32_u16(row46.val[1])); + const uint32x4x2_t row15 = vtrnq_u32(vreinterpretq_u32_u16(row13.val[0]), + vreinterpretq_u32_u16(row57.val[0])); + const uint32x4x2_t row37 = vtrnq_u32(vreinterpretq_u32_u16(row13.val[1]), + vreinterpretq_u32_u16(row57.val[1])); + *p3 = vreinterpretq_u8_u32(row04.val[0]); + *p2 = vreinterpretq_u8_u32(row15.val[0]); + *p1 = vreinterpretq_u8_u32(row26.val[0]); + *p0 = vreinterpretq_u8_u32(row37.val[0]); + *q0 = vreinterpretq_u8_u32(row04.val[1]); + *q1 = vreinterpretq_u8_u32(row15.val[1]); + *q2 = vreinterpretq_u8_u32(row26.val[1]); + *q3 = vreinterpretq_u8_u32(row37.val[1]); +} +#undef LOAD_UV_8 + +#endif // !WORK_AROUND_GCC + +static WEBP_INLINE void Store2x8(const uint8x8x2_t v, + uint8_t* const dst, int stride) { + vst2_lane_u8(dst + 0 * stride, v, 0); + vst2_lane_u8(dst + 1 * stride, v, 1); + vst2_lane_u8(dst + 2 * stride, v, 2); + vst2_lane_u8(dst + 3 * stride, v, 3); + vst2_lane_u8(dst + 4 * stride, v, 4); + vst2_lane_u8(dst + 5 * stride, v, 5); + vst2_lane_u8(dst + 6 * stride, v, 6); + vst2_lane_u8(dst + 7 * stride, v, 7); +} -#define QRegs "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", \ +static WEBP_INLINE void Store2x16(const uint8x16_t p0, const uint8x16_t q0, + uint8_t* const dst, int stride) { + uint8x8x2_t lo, hi; + lo.val[0] = vget_low_u8(p0); + lo.val[1] = vget_low_u8(q0); + hi.val[0] = vget_high_u8(p0); + hi.val[1] = vget_high_u8(q0); + Store2x8(lo, dst - 1 + 0 * stride, stride); + Store2x8(hi, dst - 1 + 8 * stride, stride); +} + +#if !defined(WORK_AROUND_GCC) +static WEBP_INLINE void Store4x8(const uint8x8x4_t v, + uint8_t* const dst, int stride) { + vst4_lane_u8(dst + 0 * stride, v, 0); + vst4_lane_u8(dst + 1 * stride, v, 1); + vst4_lane_u8(dst + 2 * stride, v, 2); + vst4_lane_u8(dst + 3 * stride, v, 3); + vst4_lane_u8(dst + 4 * stride, v, 4); + vst4_lane_u8(dst + 5 * stride, v, 5); + vst4_lane_u8(dst + 6 * stride, v, 6); + vst4_lane_u8(dst + 7 * stride, v, 7); +} + +static WEBP_INLINE void Store4x16(const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + uint8_t* const dst, int stride) { + uint8x8x4_t lo, hi; + INIT_VECTOR4(lo, + vget_low_u8(p1), vget_low_u8(p0), + vget_low_u8(q0), vget_low_u8(q1)); + INIT_VECTOR4(hi, + vget_high_u8(p1), vget_high_u8(p0), + vget_high_u8(q0), vget_high_u8(q1)); + Store4x8(lo, dst - 2 + 0 * stride, stride); + Store4x8(hi, dst - 2 + 8 * stride, stride); +} +#endif // !WORK_AROUND_GCC + +static WEBP_INLINE void Store16x2(const uint8x16_t p0, const uint8x16_t q0, + uint8_t* const dst, int stride) { + vst1q_u8(dst - stride, p0); + vst1q_u8(dst, q0); +} + +static WEBP_INLINE void Store16x4(const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + uint8_t* const dst, int stride) { + Store16x2(p1, p0, dst - stride, stride); + Store16x2(q0, q1, dst + stride, stride); +} + +static WEBP_INLINE void Store8x2x2(const uint8x16_t p0, const uint8x16_t q0, + uint8_t* const u, uint8_t* const v, + int stride) { + // p0 and q0 contain the u+v samples packed in low/high halves. + vst1_u8(u - stride, vget_low_u8(p0)); + vst1_u8(u, vget_low_u8(q0)); + vst1_u8(v - stride, vget_high_u8(p0)); + vst1_u8(v, vget_high_u8(q0)); +} + +static WEBP_INLINE void Store8x4x2(const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + uint8_t* const u, uint8_t* const v, + int stride) { + // The p1...q1 registers contain the u+v samples packed in low/high halves. + Store8x2x2(p1, p0, u - stride, v - stride, stride); + Store8x2x2(q0, q1, u + stride, v + stride, stride); +} + +#if !defined(WORK_AROUND_GCC) + +#define STORE6_LANE(DST, VAL0, VAL1, LANE) do { \ + vst3_lane_u8((DST) - 3, (VAL0), (LANE)); \ + vst3_lane_u8((DST) + 0, (VAL1), (LANE)); \ + (DST) += stride; \ +} while (0) + +static WEBP_INLINE void Store6x8x2(const uint8x16_t p2, const uint8x16_t p1, + const uint8x16_t p0, const uint8x16_t q0, + const uint8x16_t q1, const uint8x16_t q2, + uint8_t* u, uint8_t* v, + int stride) { + uint8x8x3_t u0, u1, v0, v1; + INIT_VECTOR3(u0, vget_low_u8(p2), vget_low_u8(p1), vget_low_u8(p0)); + INIT_VECTOR3(u1, vget_low_u8(q0), vget_low_u8(q1), vget_low_u8(q2)); + INIT_VECTOR3(v0, vget_high_u8(p2), vget_high_u8(p1), vget_high_u8(p0)); + INIT_VECTOR3(v1, vget_high_u8(q0), vget_high_u8(q1), vget_high_u8(q2)); + STORE6_LANE(u, u0, u1, 0); + STORE6_LANE(u, u0, u1, 1); + STORE6_LANE(u, u0, u1, 2); + STORE6_LANE(u, u0, u1, 3); + STORE6_LANE(u, u0, u1, 4); + STORE6_LANE(u, u0, u1, 5); + STORE6_LANE(u, u0, u1, 6); + STORE6_LANE(u, u0, u1, 7); + STORE6_LANE(v, v0, v1, 0); + STORE6_LANE(v, v0, v1, 1); + STORE6_LANE(v, v0, v1, 2); + STORE6_LANE(v, v0, v1, 3); + STORE6_LANE(v, v0, v1, 4); + STORE6_LANE(v, v0, v1, 5); + STORE6_LANE(v, v0, v1, 6); + STORE6_LANE(v, v0, v1, 7); +} +#undef STORE6_LANE + +static WEBP_INLINE void Store4x8x2(const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + uint8_t* const u, uint8_t* const v, + int stride) { + uint8x8x4_t u0, v0; + INIT_VECTOR4(u0, + vget_low_u8(p1), vget_low_u8(p0), + vget_low_u8(q0), vget_low_u8(q1)); + INIT_VECTOR4(v0, + vget_high_u8(p1), vget_high_u8(p0), + vget_high_u8(q0), vget_high_u8(q1)); + vst4_lane_u8(u - 2 + 0 * stride, u0, 0); + vst4_lane_u8(u - 2 + 1 * stride, u0, 1); + vst4_lane_u8(u - 2 + 2 * stride, u0, 2); + vst4_lane_u8(u - 2 + 3 * stride, u0, 3); + vst4_lane_u8(u - 2 + 4 * stride, u0, 4); + vst4_lane_u8(u - 2 + 5 * stride, u0, 5); + vst4_lane_u8(u - 2 + 6 * stride, u0, 6); + vst4_lane_u8(u - 2 + 7 * stride, u0, 7); + vst4_lane_u8(v - 2 + 0 * stride, v0, 0); + vst4_lane_u8(v - 2 + 1 * stride, v0, 1); + vst4_lane_u8(v - 2 + 2 * stride, v0, 2); + vst4_lane_u8(v - 2 + 3 * stride, v0, 3); + vst4_lane_u8(v - 2 + 4 * stride, v0, 4); + vst4_lane_u8(v - 2 + 5 * stride, v0, 5); + vst4_lane_u8(v - 2 + 6 * stride, v0, 6); + vst4_lane_u8(v - 2 + 7 * stride, v0, 7); +} + +#endif // !WORK_AROUND_GCC + +// Zero extend 'v' to an int16x8_t. +static WEBP_INLINE int16x8_t ConvertU8ToS16(uint8x8_t v) { + return vreinterpretq_s16_u16(vmovl_u8(v)); +} + +// Performs unsigned 8b saturation on 'dst01' and 'dst23' storing the result +// to the corresponding rows of 'dst'. +static WEBP_INLINE void SaturateAndStore4x4(uint8_t* const dst, + const int16x8_t dst01, + const int16x8_t dst23) { + // Unsigned saturate to 8b. + const uint8x8_t dst01_u8 = vqmovun_s16(dst01); + const uint8x8_t dst23_u8 = vqmovun_s16(dst23); + + // Store the results. + vst1_lane_u32((uint32_t*)(dst + 0 * BPS), vreinterpret_u32_u8(dst01_u8), 0); + vst1_lane_u32((uint32_t*)(dst + 1 * BPS), vreinterpret_u32_u8(dst01_u8), 1); + vst1_lane_u32((uint32_t*)(dst + 2 * BPS), vreinterpret_u32_u8(dst23_u8), 0); + vst1_lane_u32((uint32_t*)(dst + 3 * BPS), vreinterpret_u32_u8(dst23_u8), 1); +} + +static WEBP_INLINE void Add4x4(const int16x8_t row01, const int16x8_t row23, + uint8_t* const dst) { + uint32x2_t dst01 = vdup_n_u32(0); + uint32x2_t dst23 = vdup_n_u32(0); + + // Load the source pixels. + dst01 = vld1_lane_u32((uint32_t*)(dst + 0 * BPS), dst01, 0); + dst23 = vld1_lane_u32((uint32_t*)(dst + 2 * BPS), dst23, 0); + dst01 = vld1_lane_u32((uint32_t*)(dst + 1 * BPS), dst01, 1); + dst23 = vld1_lane_u32((uint32_t*)(dst + 3 * BPS), dst23, 1); + + { + // Convert to 16b. + const int16x8_t dst01_s16 = ConvertU8ToS16(vreinterpret_u8_u32(dst01)); + const int16x8_t dst23_s16 = ConvertU8ToS16(vreinterpret_u8_u32(dst23)); + + // Descale with rounding. + const int16x8_t out01 = vrsraq_n_s16(dst01_s16, row01, 3); + const int16x8_t out23 = vrsraq_n_s16(dst23_s16, row23, 3); + // Add the inverse transform. + SaturateAndStore4x4(dst, out01, out23); + } +} + +//----------------------------------------------------------------------------- +// Simple In-loop filtering (Paragraph 15.2) + +static uint8x16_t NeedsFilter(const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + int thresh) { + const uint8x16_t thresh_v = vdupq_n_u8((uint8_t)thresh); + const uint8x16_t a_p0_q0 = vabdq_u8(p0, q0); // abs(p0-q0) + const uint8x16_t a_p1_q1 = vabdq_u8(p1, q1); // abs(p1-q1) + const uint8x16_t a_p0_q0_2 = vqaddq_u8(a_p0_q0, a_p0_q0); // 2 * abs(p0-q0) + const uint8x16_t a_p1_q1_2 = vshrq_n_u8(a_p1_q1, 1); // abs(p1-q1) / 2 + const uint8x16_t sum = vqaddq_u8(a_p0_q0_2, a_p1_q1_2); + const uint8x16_t mask = vcgeq_u8(thresh_v, sum); + return mask; +} + +static int8x16_t FlipSign(const uint8x16_t v) { + const uint8x16_t sign_bit = vdupq_n_u8(0x80); + return vreinterpretq_s8_u8(veorq_u8(v, sign_bit)); +} + +static uint8x16_t FlipSignBack(const int8x16_t v) { + const int8x16_t sign_bit = vdupq_n_s8(0x80); + return vreinterpretq_u8_s8(veorq_s8(v, sign_bit)); +} + +static int8x16_t GetBaseDelta(const int8x16_t p1, const int8x16_t p0, + const int8x16_t q0, const int8x16_t q1) { + const int8x16_t q0_p0 = vqsubq_s8(q0, p0); // (q0-p0) + const int8x16_t p1_q1 = vqsubq_s8(p1, q1); // (p1-q1) + const int8x16_t s1 = vqaddq_s8(p1_q1, q0_p0); // (p1-q1) + 1 * (q0 - p0) + const int8x16_t s2 = vqaddq_s8(q0_p0, s1); // (p1-q1) + 2 * (q0 - p0) + const int8x16_t s3 = vqaddq_s8(q0_p0, s2); // (p1-q1) + 3 * (q0 - p0) + return s3; +} + +static int8x16_t GetBaseDelta0(const int8x16_t p0, const int8x16_t q0) { + const int8x16_t q0_p0 = vqsubq_s8(q0, p0); // (q0-p0) + const int8x16_t s1 = vqaddq_s8(q0_p0, q0_p0); // 2 * (q0 - p0) + const int8x16_t s2 = vqaddq_s8(q0_p0, s1); // 3 * (q0 - p0) + return s2; +} + +//------------------------------------------------------------------------------ + +static void ApplyFilter2NoFlip(const int8x16_t p0s, const int8x16_t q0s, + const int8x16_t delta, + int8x16_t* const op0, int8x16_t* const oq0) { + const int8x16_t kCst3 = vdupq_n_s8(0x03); + const int8x16_t kCst4 = vdupq_n_s8(0x04); + const int8x16_t delta_p3 = vqaddq_s8(delta, kCst3); + const int8x16_t delta_p4 = vqaddq_s8(delta, kCst4); + const int8x16_t delta3 = vshrq_n_s8(delta_p3, 3); + const int8x16_t delta4 = vshrq_n_s8(delta_p4, 3); + *op0 = vqaddq_s8(p0s, delta3); + *oq0 = vqsubq_s8(q0s, delta4); +} + +#if defined(WEBP_USE_INTRINSICS) + +static void ApplyFilter2(const int8x16_t p0s, const int8x16_t q0s, + const int8x16_t delta, + uint8x16_t* const op0, uint8x16_t* const oq0) { + const int8x16_t kCst3 = vdupq_n_s8(0x03); + const int8x16_t kCst4 = vdupq_n_s8(0x04); + const int8x16_t delta_p3 = vqaddq_s8(delta, kCst3); + const int8x16_t delta_p4 = vqaddq_s8(delta, kCst4); + const int8x16_t delta3 = vshrq_n_s8(delta_p3, 3); + const int8x16_t delta4 = vshrq_n_s8(delta_p4, 3); + const int8x16_t sp0 = vqaddq_s8(p0s, delta3); + const int8x16_t sq0 = vqsubq_s8(q0s, delta4); + *op0 = FlipSignBack(sp0); + *oq0 = FlipSignBack(sq0); +} + +static void DoFilter2(const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + const uint8x16_t mask, + uint8x16_t* const op0, uint8x16_t* const oq0) { + const int8x16_t p1s = FlipSign(p1); + const int8x16_t p0s = FlipSign(p0); + const int8x16_t q0s = FlipSign(q0); + const int8x16_t q1s = FlipSign(q1); + const int8x16_t delta0 = GetBaseDelta(p1s, p0s, q0s, q1s); + const int8x16_t delta1 = vandq_s8(delta0, vreinterpretq_s8_u8(mask)); + ApplyFilter2(p0s, q0s, delta1, op0, oq0); +} + +static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { + uint8x16_t p1, p0, q0, q1, op0, oq0; + Load16x4(p, stride, &p1, &p0, &q0, &q1); + { + const uint8x16_t mask = NeedsFilter(p1, p0, q0, q1, thresh); + DoFilter2(p1, p0, q0, q1, mask, &op0, &oq0); + } + Store16x2(op0, oq0, p, stride); +} + +static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { + uint8x16_t p1, p0, q0, q1, oq0, op0; + Load4x16(p, stride, &p1, &p0, &q0, &q1); + { + const uint8x16_t mask = NeedsFilter(p1, p0, q0, q1, thresh); + DoFilter2(p1, p0, q0, q1, mask, &op0, &oq0); + } + Store2x16(op0, oq0, p, stride); +} + +#else + +#define QRegs "q0", "q1", "q2", "q3", \ "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" #define FLIP_SIGN_BIT2(a, b, s) \ @@ -68,40 +592,16 @@ extern "C" { DO_SIMPLE_FILTER(p0, q0, q9) /* apply filter */ \ FLIP_SIGN_BIT2(p0, q0, q10) -// Load/Store vertical edge -#define LOAD8x4(c1, c2, c3, c4, b1, b2, stride) \ - "vld4.8 {" #c1"[0], " #c2"[0], " #c3"[0], " #c4"[0]}," #b1 "," #stride"\n" \ - "vld4.8 {" #c1"[1], " #c2"[1], " #c3"[1], " #c4"[1]}," #b2 "," #stride"\n" \ - "vld4.8 {" #c1"[2], " #c2"[2], " #c3"[2], " #c4"[2]}," #b1 "," #stride"\n" \ - "vld4.8 {" #c1"[3], " #c2"[3], " #c3"[3], " #c4"[3]}," #b2 "," #stride"\n" \ - "vld4.8 {" #c1"[4], " #c2"[4], " #c3"[4], " #c4"[4]}," #b1 "," #stride"\n" \ - "vld4.8 {" #c1"[5], " #c2"[5], " #c3"[5], " #c4"[5]}," #b2 "," #stride"\n" \ - "vld4.8 {" #c1"[6], " #c2"[6], " #c3"[6], " #c4"[6]}," #b1 "," #stride"\n" \ - "vld4.8 {" #c1"[7], " #c2"[7], " #c3"[7], " #c4"[7]}," #b2 "," #stride"\n" - -#define STORE8x2(c1, c2, p,stride) \ - "vst2.8 {" #c1"[0], " #c2"[0]}," #p "," #stride " \n" \ - "vst2.8 {" #c1"[1], " #c2"[1]}," #p "," #stride " \n" \ - "vst2.8 {" #c1"[2], " #c2"[2]}," #p "," #stride " \n" \ - "vst2.8 {" #c1"[3], " #c2"[3]}," #p "," #stride " \n" \ - "vst2.8 {" #c1"[4], " #c2"[4]}," #p "," #stride " \n" \ - "vst2.8 {" #c1"[5], " #c2"[5]}," #p "," #stride " \n" \ - "vst2.8 {" #c1"[6], " #c2"[6]}," #p "," #stride " \n" \ - "vst2.8 {" #c1"[7], " #c2"[7]}," #p "," #stride " \n" - -//----------------------------------------------------------------------------- -// Simple In-loop filtering (Paragraph 15.2) - -static void SimpleVFilter16NEON(uint8_t* p, int stride, int thresh) { +static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { __asm__ volatile ( "sub %[p], %[p], %[stride], lsl #1 \n" // p -= 2 * stride "vld1.u8 {q1}, [%[p]], %[stride] \n" // p1 "vld1.u8 {q2}, [%[p]], %[stride] \n" // p0 "vld1.u8 {q3}, [%[p]], %[stride] \n" // q0 - "vld1.u8 {q4}, [%[p]] \n" // q1 + "vld1.u8 {q12}, [%[p]] \n" // q1 - DO_FILTER2(q1, q2, q3, q4, %[thresh]) + DO_FILTER2(q1, q2, q3, q12, %[thresh]) "sub %[p], %[p], %[stride], lsl #1 \n" // p -= 2 * stride @@ -113,25 +613,25 @@ static void SimpleVFilter16NEON(uint8_t* p, int stride, int thresh) { ); } -static void SimpleHFilter16NEON(uint8_t* p, int stride, int thresh) { +static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { __asm__ volatile ( "sub r4, %[p], #2 \n" // base1 = p - 2 "lsl r6, %[stride], #1 \n" // r6 = 2 * stride "add r5, r4, %[stride] \n" // base2 = base1 + stride LOAD8x4(d2, d3, d4, d5, [r4], [r5], r6) - LOAD8x4(d6, d7, d8, d9, [r4], [r5], r6) - "vswp d3, d6 \n" // p1:q1 p0:q3 - "vswp d5, d8 \n" // q0:q2 q1:q4 - "vswp q2, q3 \n" // p1:q1 p0:q2 q0:q3 q1:q4 + LOAD8x4(d24, d25, d26, d27, [r4], [r5], r6) + "vswp d3, d24 \n" // p1:q1 p0:q3 + "vswp d5, d26 \n" // q0:q2 q1:q4 + "vswp q2, q12 \n" // p1:q1 p0:q2 q0:q3 q1:q4 - DO_FILTER2(q1, q2, q3, q4, %[thresh]) + DO_FILTER2(q1, q2, q12, q13, %[thresh]) "sub %[p], %[p], #1 \n" // p - 1 - "vswp d5, d6 \n" + "vswp d5, d24 \n" STORE8x2(d4, d5, [%[p]], %[stride]) - STORE8x2(d6, d7, [%[p]], %[stride]) + STORE8x2(d24, d25, [%[p]], %[stride]) : [p] "+r"(p) : [stride] "r"(stride), [thresh] "r"(thresh) @@ -139,44 +639,408 @@ static void SimpleHFilter16NEON(uint8_t* p, int stride, int thresh) { ); } -static void SimpleVFilter16iNEON(uint8_t* p, int stride, int thresh) { - int k; - for (k = 3; k > 0; --k) { +#endif // WEBP_USE_INTRINSICS + +static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) { + uint32_t k; + for (k = 3; k != 0; --k) { p += 4 * stride; - SimpleVFilter16NEON(p, stride, thresh); + SimpleVFilter16(p, stride, thresh); } } -static void SimpleHFilter16iNEON(uint8_t* p, int stride, int thresh) { - int k; - for (k = 3; k > 0; --k) { +static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) { + uint32_t k; + for (k = 3; k != 0; --k) { p += 4; - SimpleHFilter16NEON(p, stride, thresh); + SimpleHFilter16(p, stride, thresh); } } -static void TransformOneNEON(const int16_t *in, uint8_t *dst) { - const int kBPS = BPS; - const int16_t constants[] = {20091, 17734, 0, 0}; - /* kC1, kC2. Padded because vld1.16 loads 8 bytes - * Technically these are unsigned but vqdmulh is only available in signed. - * vqdmulh returns high half (effectively >> 16) but also doubles the value, - * changing the >> 16 to >> 15 and requiring an additional >> 1. - * We use this to our advantage with kC2. The canonical value is 35468. - * However, the high bit is set so treating it as signed will give incorrect - * results. We avoid this by down shifting by 1 here to clear the highest bit. - * Combined with the doubling effect of vqdmulh we get >> 16. - * This can not be applied to kC1 because the lowest bit is set. Down shifting - * the constant would reduce precision. - */ - - /* libwebp uses a trick to avoid some extra addition that libvpx does. - * Instead of: - * temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1) >> 16); - * libwebp adds 1 << 16 to cospi8sqrt2minus1 (kC1). However, this causes the - * same issue with kC1 and vqdmulh that we work around by down shifting kC2 - */ +//------------------------------------------------------------------------------ +// Complex In-loop filtering (Paragraph 15.3) + +static uint8x16_t NeedsHev(const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + int hev_thresh) { + const uint8x16_t hev_thresh_v = vdupq_n_u8((uint8_t)hev_thresh); + const uint8x16_t a_p1_p0 = vabdq_u8(p1, p0); // abs(p1 - p0) + const uint8x16_t a_q1_q0 = vabdq_u8(q1, q0); // abs(q1 - q0) + const uint8x16_t mask1 = vcgtq_u8(a_p1_p0, hev_thresh_v); + const uint8x16_t mask2 = vcgtq_u8(a_q1_q0, hev_thresh_v); + const uint8x16_t mask = vorrq_u8(mask1, mask2); + return mask; +} + +static uint8x16_t NeedsFilter2(const uint8x16_t p3, const uint8x16_t p2, + const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + const uint8x16_t q2, const uint8x16_t q3, + int ithresh, int thresh) { + const uint8x16_t ithresh_v = vdupq_n_u8((uint8_t)ithresh); + const uint8x16_t a_p3_p2 = vabdq_u8(p3, p2); // abs(p3 - p2) + const uint8x16_t a_p2_p1 = vabdq_u8(p2, p1); // abs(p2 - p1) + const uint8x16_t a_p1_p0 = vabdq_u8(p1, p0); // abs(p1 - p0) + const uint8x16_t a_q3_q2 = vabdq_u8(q3, q2); // abs(q3 - q2) + const uint8x16_t a_q2_q1 = vabdq_u8(q2, q1); // abs(q2 - q1) + const uint8x16_t a_q1_q0 = vabdq_u8(q1, q0); // abs(q1 - q0) + const uint8x16_t max1 = vmaxq_u8(a_p3_p2, a_p2_p1); + const uint8x16_t max2 = vmaxq_u8(a_p1_p0, a_q3_q2); + const uint8x16_t max3 = vmaxq_u8(a_q2_q1, a_q1_q0); + const uint8x16_t max12 = vmaxq_u8(max1, max2); + const uint8x16_t max123 = vmaxq_u8(max12, max3); + const uint8x16_t mask2 = vcgeq_u8(ithresh_v, max123); + const uint8x16_t mask1 = NeedsFilter(p1, p0, q0, q1, thresh); + const uint8x16_t mask = vandq_u8(mask1, mask2); + return mask; +} + +// 4-points filter + +static void ApplyFilter4( + const int8x16_t p1, const int8x16_t p0, + const int8x16_t q0, const int8x16_t q1, + const int8x16_t delta0, + uint8x16_t* const op1, uint8x16_t* const op0, + uint8x16_t* const oq0, uint8x16_t* const oq1) { + const int8x16_t kCst3 = vdupq_n_s8(0x03); + const int8x16_t kCst4 = vdupq_n_s8(0x04); + const int8x16_t delta1 = vqaddq_s8(delta0, kCst4); + const int8x16_t delta2 = vqaddq_s8(delta0, kCst3); + const int8x16_t a1 = vshrq_n_s8(delta1, 3); + const int8x16_t a2 = vshrq_n_s8(delta2, 3); + const int8x16_t a3 = vrshrq_n_s8(a1, 1); // a3 = (a1 + 1) >> 1 + *op0 = FlipSignBack(vqaddq_s8(p0, a2)); // clip(p0 + a2) + *oq0 = FlipSignBack(vqsubq_s8(q0, a1)); // clip(q0 - a1) + *op1 = FlipSignBack(vqaddq_s8(p1, a3)); // clip(p1 + a3) + *oq1 = FlipSignBack(vqsubq_s8(q1, a3)); // clip(q1 - a3) +} + +static void DoFilter4( + const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, + const uint8x16_t mask, const uint8x16_t hev_mask, + uint8x16_t* const op1, uint8x16_t* const op0, + uint8x16_t* const oq0, uint8x16_t* const oq1) { + // This is a fused version of DoFilter2() calling ApplyFilter2 directly + const int8x16_t p1s = FlipSign(p1); + int8x16_t p0s = FlipSign(p0); + int8x16_t q0s = FlipSign(q0); + const int8x16_t q1s = FlipSign(q1); + const uint8x16_t simple_lf_mask = vandq_u8(mask, hev_mask); + + // do_filter2 part (simple loopfilter on pixels with hev) + { + const int8x16_t delta = GetBaseDelta(p1s, p0s, q0s, q1s); + const int8x16_t simple_lf_delta = + vandq_s8(delta, vreinterpretq_s8_u8(simple_lf_mask)); + ApplyFilter2NoFlip(p0s, q0s, simple_lf_delta, &p0s, &q0s); + } + + // do_filter4 part (complex loopfilter on pixels without hev) + { + const int8x16_t delta0 = GetBaseDelta0(p0s, q0s); + // we use: (mask & hev_mask) ^ mask = mask & !hev_mask + const uint8x16_t complex_lf_mask = veorq_u8(simple_lf_mask, mask); + const int8x16_t complex_lf_delta = + vandq_s8(delta0, vreinterpretq_s8_u8(complex_lf_mask)); + ApplyFilter4(p1s, p0s, q0s, q1s, complex_lf_delta, op1, op0, oq0, oq1); + } +} + +// 6-points filter + +static void ApplyFilter6( + const int8x16_t p2, const int8x16_t p1, const int8x16_t p0, + const int8x16_t q0, const int8x16_t q1, const int8x16_t q2, + const int8x16_t delta, + uint8x16_t* const op2, uint8x16_t* const op1, uint8x16_t* const op0, + uint8x16_t* const oq0, uint8x16_t* const oq1, uint8x16_t* const oq2) { + const int16x8_t kCst63 = vdupq_n_s16(63); + const int8x8_t kCst27 = vdup_n_s8(27); + const int8x8_t kCst18 = vdup_n_s8(18); + const int8x8_t kCst9 = vdup_n_s8(9); + const int8x8_t delta_lo = vget_low_s8(delta); + const int8x8_t delta_hi = vget_high_s8(delta); + const int16x8_t s1_lo = vmlal_s8(kCst63, kCst27, delta_lo); // 63 + 27 * a + const int16x8_t s1_hi = vmlal_s8(kCst63, kCst27, delta_hi); // 63 + 27 * a + const int16x8_t s2_lo = vmlal_s8(kCst63, kCst18, delta_lo); // 63 + 18 * a + const int16x8_t s2_hi = vmlal_s8(kCst63, kCst18, delta_hi); // 63 + 18 * a + const int16x8_t s3_lo = vmlal_s8(kCst63, kCst9, delta_lo); // 63 + 9 * a + const int16x8_t s3_hi = vmlal_s8(kCst63, kCst9, delta_hi); // 63 + 9 * a + const int8x8_t a1_lo = vqshrn_n_s16(s1_lo, 7); + const int8x8_t a1_hi = vqshrn_n_s16(s1_hi, 7); + const int8x8_t a2_lo = vqshrn_n_s16(s2_lo, 7); + const int8x8_t a2_hi = vqshrn_n_s16(s2_hi, 7); + const int8x8_t a3_lo = vqshrn_n_s16(s3_lo, 7); + const int8x8_t a3_hi = vqshrn_n_s16(s3_hi, 7); + const int8x16_t a1 = vcombine_s8(a1_lo, a1_hi); + const int8x16_t a2 = vcombine_s8(a2_lo, a2_hi); + const int8x16_t a3 = vcombine_s8(a3_lo, a3_hi); + + *op0 = FlipSignBack(vqaddq_s8(p0, a1)); // clip(p0 + a1) + *oq0 = FlipSignBack(vqsubq_s8(q0, a1)); // clip(q0 - q1) + *oq1 = FlipSignBack(vqsubq_s8(q1, a2)); // clip(q1 - a2) + *op1 = FlipSignBack(vqaddq_s8(p1, a2)); // clip(p1 + a2) + *oq2 = FlipSignBack(vqsubq_s8(q2, a3)); // clip(q2 - a3) + *op2 = FlipSignBack(vqaddq_s8(p2, a3)); // clip(p2 + a3) +} + +static void DoFilter6( + const uint8x16_t p2, const uint8x16_t p1, const uint8x16_t p0, + const uint8x16_t q0, const uint8x16_t q1, const uint8x16_t q2, + const uint8x16_t mask, const uint8x16_t hev_mask, + uint8x16_t* const op2, uint8x16_t* const op1, uint8x16_t* const op0, + uint8x16_t* const oq0, uint8x16_t* const oq1, uint8x16_t* const oq2) { + // This is a fused version of DoFilter2() calling ApplyFilter2 directly + const int8x16_t p2s = FlipSign(p2); + const int8x16_t p1s = FlipSign(p1); + int8x16_t p0s = FlipSign(p0); + int8x16_t q0s = FlipSign(q0); + const int8x16_t q1s = FlipSign(q1); + const int8x16_t q2s = FlipSign(q2); + const uint8x16_t simple_lf_mask = vandq_u8(mask, hev_mask); + const int8x16_t delta0 = GetBaseDelta(p1s, p0s, q0s, q1s); + + // do_filter2 part (simple loopfilter on pixels with hev) + { + const int8x16_t simple_lf_delta = + vandq_s8(delta0, vreinterpretq_s8_u8(simple_lf_mask)); + ApplyFilter2NoFlip(p0s, q0s, simple_lf_delta, &p0s, &q0s); + } + + // do_filter6 part (complex loopfilter on pixels without hev) + { + // we use: (mask & hev_mask) ^ mask = mask & !hev_mask + const uint8x16_t complex_lf_mask = veorq_u8(simple_lf_mask, mask); + const int8x16_t complex_lf_delta = + vandq_s8(delta0, vreinterpretq_s8_u8(complex_lf_mask)); + ApplyFilter6(p2s, p1s, p0s, q0s, q1s, q2s, complex_lf_delta, + op2, op1, op0, oq0, oq1, oq2); + } +} + +// on macroblock edges + +static void VFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3; + Load16x8(p, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3); + { + const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, + ithresh, thresh); + const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh); + uint8x16_t op2, op1, op0, oq0, oq1, oq2; + DoFilter6(p2, p1, p0, q0, q1, q2, mask, hev_mask, + &op2, &op1, &op0, &oq0, &oq1, &oq2); + Store16x2(op2, op1, p - 2 * stride, stride); + Store16x2(op0, oq0, p + 0 * stride, stride); + Store16x2(oq1, oq2, p + 2 * stride, stride); + } +} + +static void HFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3; + Load8x16(p, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3); + { + const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, + ithresh, thresh); + const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh); + uint8x16_t op2, op1, op0, oq0, oq1, oq2; + DoFilter6(p2, p1, p0, q0, q1, q2, mask, hev_mask, + &op2, &op1, &op0, &oq0, &oq1, &oq2); + Store2x16(op2, op1, p - 2, stride); + Store2x16(op0, oq0, p + 0, stride); + Store2x16(oq1, oq2, p + 2, stride); + } +} + +// on three inner edges +static void VFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + uint32_t k; + uint8x16_t p3, p2, p1, p0; + Load16x4(p + 2 * stride, stride, &p3, &p2, &p1, &p0); + for (k = 3; k != 0; --k) { + uint8x16_t q0, q1, q2, q3; + p += 4 * stride; + Load16x4(p + 2 * stride, stride, &q0, &q1, &q2, &q3); + { + const uint8x16_t mask = + NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, ithresh, thresh); + const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh); + // p3 and p2 are not just temporary variables here: they will be + // re-used for next span. And q2/q3 will become p1/p0 accordingly. + DoFilter4(p1, p0, q0, q1, mask, hev_mask, &p1, &p0, &p3, &p2); + Store16x4(p1, p0, p3, p2, p, stride); + p1 = q2; + p0 = q3; + } + } +} + +#if !defined(WORK_AROUND_GCC) +static void HFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + uint32_t k; + uint8x16_t p3, p2, p1, p0; + Load4x16(p + 2, stride, &p3, &p2, &p1, &p0); + for (k = 3; k != 0; --k) { + uint8x16_t q0, q1, q2, q3; + p += 4; + Load4x16(p + 2, stride, &q0, &q1, &q2, &q3); + { + const uint8x16_t mask = + NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, ithresh, thresh); + const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh); + DoFilter4(p1, p0, q0, q1, mask, hev_mask, &p1, &p0, &p3, &p2); + Store4x16(p1, p0, p3, p2, p, stride); + p1 = q2; + p0 = q3; + } + } +} +#endif // !WORK_AROUND_GCC + +// 8-pixels wide variant, for chroma filtering +static void VFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3; + Load8x8x2(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3); + { + const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, + ithresh, thresh); + const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh); + uint8x16_t op2, op1, op0, oq0, oq1, oq2; + DoFilter6(p2, p1, p0, q0, q1, q2, mask, hev_mask, + &op2, &op1, &op0, &oq0, &oq1, &oq2); + Store8x2x2(op2, op1, u - 2 * stride, v - 2 * stride, stride); + Store8x2x2(op0, oq0, u + 0 * stride, v + 0 * stride, stride); + Store8x2x2(oq1, oq2, u + 2 * stride, v + 2 * stride, stride); + } +} +static void VFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3; + u += 4 * stride; + v += 4 * stride; + Load8x8x2(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3); + { + const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, + ithresh, thresh); + const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh); + uint8x16_t op1, op0, oq0, oq1; + DoFilter4(p1, p0, q0, q1, mask, hev_mask, &op1, &op0, &oq0, &oq1); + Store8x4x2(op1, op0, oq0, oq1, u, v, stride); + } +} + +#if !defined(WORK_AROUND_GCC) +static void HFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3; + Load8x8x2T(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3); + { + const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, + ithresh, thresh); + const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh); + uint8x16_t op2, op1, op0, oq0, oq1, oq2; + DoFilter6(p2, p1, p0, q0, q1, q2, mask, hev_mask, + &op2, &op1, &op0, &oq0, &oq1, &oq2); + Store6x8x2(op2, op1, op0, oq0, oq1, oq2, u, v, stride); + } +} + +static void HFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3; + u += 4; + v += 4; + Load8x8x2T(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3); + { + const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, + ithresh, thresh); + const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh); + uint8x16_t op1, op0, oq0, oq1; + DoFilter4(p1, p0, q0, q1, mask, hev_mask, &op1, &op0, &oq0, &oq1); + Store4x8x2(op1, op0, oq0, oq1, u, v, stride); + } +} +#endif // !WORK_AROUND_GCC + +//----------------------------------------------------------------------------- +// Inverse transforms (Paragraph 14.4) + +// Technically these are unsigned but vqdmulh is only available in signed. +// vqdmulh returns high half (effectively >> 16) but also doubles the value, +// changing the >> 16 to >> 15 and requiring an additional >> 1. +// We use this to our advantage with kC2. The canonical value is 35468. +// However, the high bit is set so treating it as signed will give incorrect +// results. We avoid this by down shifting by 1 here to clear the highest bit. +// Combined with the doubling effect of vqdmulh we get >> 16. +// This can not be applied to kC1 because the lowest bit is set. Down shifting +// the constant would reduce precision. + +// libwebp uses a trick to avoid some extra addition that libvpx does. +// Instead of: +// temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1) >> 16); +// libwebp adds 1 << 16 to cospi8sqrt2minus1 (kC1). However, this causes the +// same issue with kC1 and vqdmulh that we work around by down shifting kC2 +static const int16_t kC1 = 20091; +static const int16_t kC2 = 17734; // half of kC2, actually. See comment above. + +#if defined(WEBP_USE_INTRINSICS) +static WEBP_INLINE void Transpose8x2(const int16x8_t in0, const int16x8_t in1, + int16x8x2_t* const out) { + // a0 a1 a2 a3 | b0 b1 b2 b3 => a0 b0 c0 d0 | a1 b1 c1 d1 + // c0 c1 c2 c3 | d0 d1 d2 d3 a2 b2 c2 d2 | a3 b3 c3 d3 + const int16x8x2_t tmp0 = vzipq_s16(in0, in1); // a0 c0 a1 c1 a2 c2 ... + // b0 d0 b1 d1 b2 d2 ... + *out = vzipq_s16(tmp0.val[0], tmp0.val[1]); +} + +static WEBP_INLINE void TransformPass(int16x8x2_t* const rows) { + // {rows} = in0 | in4 + // in8 | in12 + // B1 = in4 | in12 + const int16x8_t B1 = + vcombine_s16(vget_high_s16(rows->val[0]), vget_high_s16(rows->val[1])); + // C0 = kC1 * in4 | kC1 * in12 + // C1 = kC2 * in4 | kC2 * in12 + const int16x8_t C0 = vsraq_n_s16(B1, vqdmulhq_n_s16(B1, kC1), 1); + const int16x8_t C1 = vqdmulhq_n_s16(B1, kC2); + const int16x4_t a = vqadd_s16(vget_low_s16(rows->val[0]), + vget_low_s16(rows->val[1])); // in0 + in8 + const int16x4_t b = vqsub_s16(vget_low_s16(rows->val[0]), + vget_low_s16(rows->val[1])); // in0 - in8 + // c = kC2 * in4 - kC1 * in12 + // d = kC1 * in4 + kC2 * in12 + const int16x4_t c = vqsub_s16(vget_low_s16(C1), vget_high_s16(C0)); + const int16x4_t d = vqadd_s16(vget_low_s16(C0), vget_high_s16(C1)); + const int16x8_t D0 = vcombine_s16(a, b); // D0 = a | b + const int16x8_t D1 = vcombine_s16(d, c); // D1 = d | c + const int16x8_t E0 = vqaddq_s16(D0, D1); // a+d | b+c + const int16x8_t E_tmp = vqsubq_s16(D0, D1); // a-d | b-c + const int16x8_t E1 = vcombine_s16(vget_high_s16(E_tmp), vget_low_s16(E_tmp)); + Transpose8x2(E0, E1, rows); +} + +static void TransformOne(const int16_t* in, uint8_t* dst) { + int16x8x2_t rows; + INIT_VECTOR2(rows, vld1q_s16(in + 0), vld1q_s16(in + 8)); + TransformPass(&rows); + TransformPass(&rows); + Add4x4(rows.val[0], rows.val[1], dst); +} + +#else + +static void TransformOne(const int16_t* in, uint8_t* dst) { + const int kBPS = BPS; + // kC1, kC2. Padded because vld1.16 loads 8 bytes + const int16_t constants[4] = { kC1, kC2, 0, 0 }; /* Adapted from libvpx: vp8/common/arm/neon/shortidct4x4llm_neon.asm */ __asm__ volatile ( "vld1.16 {q1, q2}, [%[in]] \n" @@ -304,26 +1168,472 @@ static void TransformOneNEON(const int16_t *in, uint8_t *dst) { ); } -static void TransformTwoNEON(const int16_t* in, uint8_t* dst, int do_two) { - TransformOneNEON(in, dst); +#endif // WEBP_USE_INTRINSICS + +static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) { + TransformOne(in, dst); if (do_two) { - TransformOneNEON(in + 16, dst + 4); + TransformOne(in + 16, dst + 4); } } -extern void VP8DspInitNEON(void); +static void TransformDC(const int16_t* in, uint8_t* dst) { + const int16x8_t DC = vdupq_n_s16(in[0]); + Add4x4(DC, DC, dst); +} + +//------------------------------------------------------------------------------ + +#define STORE_WHT(dst, col, rows) do { \ + *dst = vgetq_lane_s32(rows.val[0], col); (dst) += 16; \ + *dst = vgetq_lane_s32(rows.val[1], col); (dst) += 16; \ + *dst = vgetq_lane_s32(rows.val[2], col); (dst) += 16; \ + *dst = vgetq_lane_s32(rows.val[3], col); (dst) += 16; \ +} while (0) + +static void TransformWHT(const int16_t* in, int16_t* out) { + int32x4x4_t tmp; -void VP8DspInitNEON(void) { - VP8Transform = TransformTwoNEON; + { + // Load the source. + const int16x4_t in00_03 = vld1_s16(in + 0); + const int16x4_t in04_07 = vld1_s16(in + 4); + const int16x4_t in08_11 = vld1_s16(in + 8); + const int16x4_t in12_15 = vld1_s16(in + 12); + const int32x4_t a0 = vaddl_s16(in00_03, in12_15); // in[0..3] + in[12..15] + const int32x4_t a1 = vaddl_s16(in04_07, in08_11); // in[4..7] + in[8..11] + const int32x4_t a2 = vsubl_s16(in04_07, in08_11); // in[4..7] - in[8..11] + const int32x4_t a3 = vsubl_s16(in00_03, in12_15); // in[0..3] - in[12..15] + tmp.val[0] = vaddq_s32(a0, a1); + tmp.val[1] = vaddq_s32(a3, a2); + tmp.val[2] = vsubq_s32(a0, a1); + tmp.val[3] = vsubq_s32(a3, a2); + // Arrange the temporary results column-wise. + tmp = Transpose4x4(tmp); + } + + { + const int32x4_t kCst3 = vdupq_n_s32(3); + const int32x4_t dc = vaddq_s32(tmp.val[0], kCst3); // add rounder + const int32x4_t a0 = vaddq_s32(dc, tmp.val[3]); + const int32x4_t a1 = vaddq_s32(tmp.val[1], tmp.val[2]); + const int32x4_t a2 = vsubq_s32(tmp.val[1], tmp.val[2]); + const int32x4_t a3 = vsubq_s32(dc, tmp.val[3]); + + tmp.val[0] = vaddq_s32(a0, a1); + tmp.val[1] = vaddq_s32(a3, a2); + tmp.val[2] = vsubq_s32(a0, a1); + tmp.val[3] = vsubq_s32(a3, a2); + + // right shift the results by 3. + tmp.val[0] = vshrq_n_s32(tmp.val[0], 3); + tmp.val[1] = vshrq_n_s32(tmp.val[1], 3); + tmp.val[2] = vshrq_n_s32(tmp.val[2], 3); + tmp.val[3] = vshrq_n_s32(tmp.val[3], 3); - VP8SimpleVFilter16 = SimpleVFilter16NEON; - VP8SimpleHFilter16 = SimpleHFilter16NEON; - VP8SimpleVFilter16i = SimpleVFilter16iNEON; - VP8SimpleHFilter16i = SimpleHFilter16iNEON; + STORE_WHT(out, 0, tmp); + STORE_WHT(out, 1, tmp); + STORE_WHT(out, 2, tmp); + STORE_WHT(out, 3, tmp); + } } -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" +#undef STORE_WHT + +//------------------------------------------------------------------------------ + +#define MUL(a, b) (((a) * (b)) >> 16) +static void TransformAC3(const int16_t* in, uint8_t* dst) { + static const int kC1_full = 20091 + (1 << 16); + static const int kC2_full = 35468; + const int16x4_t A = vld1_dup_s16(in); + const int16x4_t c4 = vdup_n_s16(MUL(in[4], kC2_full)); + const int16x4_t d4 = vdup_n_s16(MUL(in[4], kC1_full)); + const int c1 = MUL(in[1], kC2_full); + const int d1 = MUL(in[1], kC1_full); + const uint64_t cd = (uint64_t)( d1 & 0xffff) << 0 | + (uint64_t)( c1 & 0xffff) << 16 | + (uint64_t)(-c1 & 0xffff) << 32 | + (uint64_t)(-d1 & 0xffff) << 48; + const int16x4_t CD = vcreate_s16(cd); + const int16x4_t B = vqadd_s16(A, CD); + const int16x8_t m0_m1 = vcombine_s16(vqadd_s16(B, d4), vqadd_s16(B, c4)); + const int16x8_t m2_m3 = vcombine_s16(vqsub_s16(B, c4), vqsub_s16(B, d4)); + Add4x4(m0_m1, m2_m3, dst); +} +#undef MUL + +//------------------------------------------------------------------------------ +// 4x4 + +static void DC4(uint8_t* dst) { // DC + const uint8x8_t A = vld1_u8(dst - BPS); // top row + const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top + const uint16x4_t p1 = vpadd_u16(p0, p0); + const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + 0 * BPS - 1)); + const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + 1 * BPS - 1)); + const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + 2 * BPS - 1)); + const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + 3 * BPS - 1)); + const uint16x8_t s0 = vaddq_u16(L0, L1); + const uint16x8_t s1 = vaddq_u16(L2, L3); + const uint16x8_t s01 = vaddq_u16(s0, s1); + const uint16x8_t sum = vaddq_u16(s01, vcombine_u16(p1, p1)); + const uint8x8_t dc0 = vrshrn_n_u16(sum, 3); // (sum + 4) >> 3 + const uint8x8_t dc = vdup_lane_u8(dc0, 0); + int i; + for (i = 0; i < 4; ++i) { + vst1_lane_u32((uint32_t*)(dst + i * BPS), vreinterpret_u32_u8(dc), 0); + } +} + +// TrueMotion (4x4 + 8x8) +static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) { + const uint8x8_t TL = vld1_dup_u8(dst - BPS - 1); // top-left pixel 'A[-1]' + const uint8x8_t T = vld1_u8(dst - BPS); // top row 'A[0..3]' + const int16x8_t d = vreinterpretq_s16_u16(vsubl_u8(T, TL)); // A[c] - A[-1] + int y; + for (y = 0; y < size; y += 4) { + // left edge + const int16x8_t L0 = ConvertU8ToS16(vld1_dup_u8(dst + 0 * BPS - 1)); + const int16x8_t L1 = ConvertU8ToS16(vld1_dup_u8(dst + 1 * BPS - 1)); + const int16x8_t L2 = ConvertU8ToS16(vld1_dup_u8(dst + 2 * BPS - 1)); + const int16x8_t L3 = ConvertU8ToS16(vld1_dup_u8(dst + 3 * BPS - 1)); + const int16x8_t r0 = vaddq_s16(L0, d); // L[r] + A[c] - A[-1] + const int16x8_t r1 = vaddq_s16(L1, d); + const int16x8_t r2 = vaddq_s16(L2, d); + const int16x8_t r3 = vaddq_s16(L3, d); + // Saturate and store the result. + const uint32x2_t r0_u32 = vreinterpret_u32_u8(vqmovun_s16(r0)); + const uint32x2_t r1_u32 = vreinterpret_u32_u8(vqmovun_s16(r1)); + const uint32x2_t r2_u32 = vreinterpret_u32_u8(vqmovun_s16(r2)); + const uint32x2_t r3_u32 = vreinterpret_u32_u8(vqmovun_s16(r3)); + if (size == 4) { + vst1_lane_u32((uint32_t*)(dst + 0 * BPS), r0_u32, 0); + vst1_lane_u32((uint32_t*)(dst + 1 * BPS), r1_u32, 0); + vst1_lane_u32((uint32_t*)(dst + 2 * BPS), r2_u32, 0); + vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3_u32, 0); + } else { + vst1_u32((uint32_t*)(dst + 0 * BPS), r0_u32); + vst1_u32((uint32_t*)(dst + 1 * BPS), r1_u32); + vst1_u32((uint32_t*)(dst + 2 * BPS), r2_u32); + vst1_u32((uint32_t*)(dst + 3 * BPS), r3_u32); + } + dst += 4 * BPS; + } +} + +static void TM4(uint8_t* dst) { TrueMotion(dst, 4); } + +static void VE4(uint8_t* dst) { // vertical + // NB: avoid vld1_u64 here as an alignment hint may be added -> SIGBUS. + const uint64x1_t A0 = vreinterpret_u64_u8(vld1_u8(dst - BPS - 1)); // top row + const uint64x1_t A1 = vshr_n_u64(A0, 8); + const uint64x1_t A2 = vshr_n_u64(A0, 16); + const uint8x8_t ABCDEFGH = vreinterpret_u8_u64(A0); + const uint8x8_t BCDEFGH0 = vreinterpret_u8_u64(A1); + const uint8x8_t CDEFGH00 = vreinterpret_u8_u64(A2); + const uint8x8_t b = vhadd_u8(ABCDEFGH, CDEFGH00); + const uint8x8_t avg = vrhadd_u8(b, BCDEFGH0); + int i; + for (i = 0; i < 4; ++i) { + vst1_lane_u32((uint32_t*)(dst + i * BPS), vreinterpret_u32_u8(avg), 0); + } +} + +static void RD4(uint8_t* dst) { // Down-right + const uint8x8_t XABCD_u8 = vld1_u8(dst - BPS - 1); + const uint64x1_t XABCD = vreinterpret_u64_u8(XABCD_u8); + const uint64x1_t ____XABC = vshl_n_u64(XABCD, 32); + const uint32_t I = dst[-1 + 0 * BPS]; + const uint32_t J = dst[-1 + 1 * BPS]; + const uint32_t K = dst[-1 + 2 * BPS]; + const uint32_t L = dst[-1 + 3 * BPS]; + const uint64x1_t LKJI____ = vcreate_u64(L | (K << 8) | (J << 16) | (I << 24)); + const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC); + const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8)); + const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16)); + const uint8_t D = vget_lane_u8(XABCD_u8, 4); + const uint8x8_t JIXABCD_ = vset_lane_u8(D, JIXABC__, 6); + const uint8x8_t LKJIXABC_u8 = vreinterpret_u8_u64(LKJIXABC); + const uint8x8_t avg1 = vhadd_u8(JIXABCD_, LKJIXABC_u8); + const uint8x8_t avg2 = vrhadd_u8(avg1, KJIXABC_); + const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2); + const uint32x2_t r3 = vreinterpret_u32_u8(avg2); + const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8)); + const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16)); + const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24)); + vst1_lane_u32((uint32_t*)(dst + 0 * BPS), r0, 0); + vst1_lane_u32((uint32_t*)(dst + 1 * BPS), r1, 0); + vst1_lane_u32((uint32_t*)(dst + 2 * BPS), r2, 0); + vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3, 0); +} + +static void LD4(uint8_t* dst) { // Down-left + // Note using the same shift trick as VE4() is slower here. + const uint8x8_t ABCDEFGH = vld1_u8(dst - BPS + 0); + const uint8x8_t BCDEFGH0 = vld1_u8(dst - BPS + 1); + const uint8x8_t CDEFGH00 = vld1_u8(dst - BPS + 2); + const uint8x8_t CDEFGHH0 = vset_lane_u8(dst[-BPS + 7], CDEFGH00, 6); + const uint8x8_t avg1 = vhadd_u8(ABCDEFGH, CDEFGHH0); + const uint8x8_t avg2 = vrhadd_u8(avg1, BCDEFGH0); + const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2); + const uint32x2_t r0 = vreinterpret_u32_u8(avg2); + const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8)); + const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16)); + const uint32x2_t r3 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24)); + vst1_lane_u32((uint32_t*)(dst + 0 * BPS), r0, 0); + vst1_lane_u32((uint32_t*)(dst + 1 * BPS), r1, 0); + vst1_lane_u32((uint32_t*)(dst + 2 * BPS), r2, 0); + vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3, 0); +} + +//------------------------------------------------------------------------------ +// Chroma + +static void VE8uv(uint8_t* dst) { // vertical + const uint8x8_t top = vld1_u8(dst - BPS); + int j; + for (j = 0; j < 8; ++j) { + vst1_u8(dst + j * BPS, top); + } +} + +static void HE8uv(uint8_t* dst) { // horizontal + int j; + for (j = 0; j < 8; ++j) { + const uint8x8_t left = vld1_dup_u8(dst - 1); + vst1_u8(dst, left); + dst += BPS; + } +} + +static WEBP_INLINE void DC8(uint8_t* dst, int do_top, int do_left) { + uint16x8_t sum_top; + uint16x8_t sum_left; + uint8x8_t dc0; + + if (do_top) { + const uint8x8_t A = vld1_u8(dst - BPS); // top row + const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top + const uint16x4_t p1 = vpadd_u16(p0, p0); + const uint16x4_t p2 = vpadd_u16(p1, p1); + sum_top = vcombine_u16(p2, p2); + } + + if (do_left) { + const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + 0 * BPS - 1)); + const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + 1 * BPS - 1)); + const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + 2 * BPS - 1)); + const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + 3 * BPS - 1)); + const uint16x8_t L4 = vmovl_u8(vld1_u8(dst + 4 * BPS - 1)); + const uint16x8_t L5 = vmovl_u8(vld1_u8(dst + 5 * BPS - 1)); + const uint16x8_t L6 = vmovl_u8(vld1_u8(dst + 6 * BPS - 1)); + const uint16x8_t L7 = vmovl_u8(vld1_u8(dst + 7 * BPS - 1)); + const uint16x8_t s0 = vaddq_u16(L0, L1); + const uint16x8_t s1 = vaddq_u16(L2, L3); + const uint16x8_t s2 = vaddq_u16(L4, L5); + const uint16x8_t s3 = vaddq_u16(L6, L7); + const uint16x8_t s01 = vaddq_u16(s0, s1); + const uint16x8_t s23 = vaddq_u16(s2, s3); + sum_left = vaddq_u16(s01, s23); + } + + if (do_top && do_left) { + const uint16x8_t sum = vaddq_u16(sum_left, sum_top); + dc0 = vrshrn_n_u16(sum, 4); + } else if (do_top) { + dc0 = vrshrn_n_u16(sum_top, 3); + } else if (do_left) { + dc0 = vrshrn_n_u16(sum_left, 3); + } else { + dc0 = vdup_n_u8(0x80); + } + + { + const uint8x8_t dc = vdup_lane_u8(dc0, 0); + int i; + for (i = 0; i < 8; ++i) { + vst1_u32((uint32_t*)(dst + i * BPS), vreinterpret_u32_u8(dc)); + } + } +} + +static void DC8uv(uint8_t* dst) { DC8(dst, 1, 1); } +static void DC8uvNoTop(uint8_t* dst) { DC8(dst, 0, 1); } +static void DC8uvNoLeft(uint8_t* dst) { DC8(dst, 1, 0); } +static void DC8uvNoTopLeft(uint8_t* dst) { DC8(dst, 0, 0); } + +static void TM8uv(uint8_t* dst) { TrueMotion(dst, 8); } + +//------------------------------------------------------------------------------ +// 16x16 + +static void VE16(uint8_t* dst) { // vertical + const uint8x16_t top = vld1q_u8(dst - BPS); + int j; + for (j = 0; j < 16; ++j) { + vst1q_u8(dst + j * BPS, top); + } +} + +static void HE16(uint8_t* dst) { // horizontal + int j; + for (j = 0; j < 16; ++j) { + const uint8x16_t left = vld1q_dup_u8(dst - 1); + vst1q_u8(dst, left); + dst += BPS; + } +} + +static WEBP_INLINE void DC16(uint8_t* dst, int do_top, int do_left) { + uint16x8_t sum_top; + uint16x8_t sum_left; + uint8x8_t dc0; + + if (do_top) { + const uint8x16_t A = vld1q_u8(dst - BPS); // top row + const uint16x8_t p0 = vpaddlq_u8(A); // cascading summation of the top + const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0)); + const uint16x4_t p2 = vpadd_u16(p1, p1); + const uint16x4_t p3 = vpadd_u16(p2, p2); + sum_top = vcombine_u16(p3, p3); + } + + if (do_left) { + int i; + sum_left = vdupq_n_u16(0); + for (i = 0; i < 16; i += 8) { + const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + (i + 0) * BPS - 1)); + const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + (i + 1) * BPS - 1)); + const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + (i + 2) * BPS - 1)); + const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + (i + 3) * BPS - 1)); + const uint16x8_t L4 = vmovl_u8(vld1_u8(dst + (i + 4) * BPS - 1)); + const uint16x8_t L5 = vmovl_u8(vld1_u8(dst + (i + 5) * BPS - 1)); + const uint16x8_t L6 = vmovl_u8(vld1_u8(dst + (i + 6) * BPS - 1)); + const uint16x8_t L7 = vmovl_u8(vld1_u8(dst + (i + 7) * BPS - 1)); + const uint16x8_t s0 = vaddq_u16(L0, L1); + const uint16x8_t s1 = vaddq_u16(L2, L3); + const uint16x8_t s2 = vaddq_u16(L4, L5); + const uint16x8_t s3 = vaddq_u16(L6, L7); + const uint16x8_t s01 = vaddq_u16(s0, s1); + const uint16x8_t s23 = vaddq_u16(s2, s3); + const uint16x8_t sum = vaddq_u16(s01, s23); + sum_left = vaddq_u16(sum_left, sum); + } + } + + if (do_top && do_left) { + const uint16x8_t sum = vaddq_u16(sum_left, sum_top); + dc0 = vrshrn_n_u16(sum, 5); + } else if (do_top) { + dc0 = vrshrn_n_u16(sum_top, 4); + } else if (do_left) { + dc0 = vrshrn_n_u16(sum_left, 4); + } else { + dc0 = vdup_n_u8(0x80); + } + + { + const uint8x16_t dc = vdupq_lane_u8(dc0, 0); + int i; + for (i = 0; i < 16; ++i) { + vst1q_u8(dst + i * BPS, dc); + } + } +} + +static void DC16TopLeft(uint8_t* dst) { DC16(dst, 1, 1); } +static void DC16NoTop(uint8_t* dst) { DC16(dst, 0, 1); } +static void DC16NoLeft(uint8_t* dst) { DC16(dst, 1, 0); } +static void DC16NoTopLeft(uint8_t* dst) { DC16(dst, 0, 0); } + +static void TM16(uint8_t* dst) { + const uint8x8_t TL = vld1_dup_u8(dst - BPS - 1); // top-left pixel 'A[-1]' + const uint8x16_t T = vld1q_u8(dst - BPS); // top row 'A[0..15]' + // A[c] - A[-1] + const int16x8_t d_lo = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(T), TL)); + const int16x8_t d_hi = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(T), TL)); + int y; + for (y = 0; y < 16; y += 4) { + // left edge + const int16x8_t L0 = ConvertU8ToS16(vld1_dup_u8(dst + 0 * BPS - 1)); + const int16x8_t L1 = ConvertU8ToS16(vld1_dup_u8(dst + 1 * BPS - 1)); + const int16x8_t L2 = ConvertU8ToS16(vld1_dup_u8(dst + 2 * BPS - 1)); + const int16x8_t L3 = ConvertU8ToS16(vld1_dup_u8(dst + 3 * BPS - 1)); + const int16x8_t r0_lo = vaddq_s16(L0, d_lo); // L[r] + A[c] - A[-1] + const int16x8_t r1_lo = vaddq_s16(L1, d_lo); + const int16x8_t r2_lo = vaddq_s16(L2, d_lo); + const int16x8_t r3_lo = vaddq_s16(L3, d_lo); + const int16x8_t r0_hi = vaddq_s16(L0, d_hi); + const int16x8_t r1_hi = vaddq_s16(L1, d_hi); + const int16x8_t r2_hi = vaddq_s16(L2, d_hi); + const int16x8_t r3_hi = vaddq_s16(L3, d_hi); + // Saturate and store the result. + const uint8x16_t row0 = vcombine_u8(vqmovun_s16(r0_lo), vqmovun_s16(r0_hi)); + const uint8x16_t row1 = vcombine_u8(vqmovun_s16(r1_lo), vqmovun_s16(r1_hi)); + const uint8x16_t row2 = vcombine_u8(vqmovun_s16(r2_lo), vqmovun_s16(r2_hi)); + const uint8x16_t row3 = vcombine_u8(vqmovun_s16(r3_lo), vqmovun_s16(r3_hi)); + vst1q_u8(dst + 0 * BPS, row0); + vst1q_u8(dst + 1 * BPS, row1); + vst1q_u8(dst + 2 * BPS, row2); + vst1q_u8(dst + 3 * BPS, row3); + dst += 4 * BPS; + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8DspInitNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitNEON(void) { + VP8Transform = TransformTwo; + VP8TransformAC3 = TransformAC3; + VP8TransformDC = TransformDC; + VP8TransformWHT = TransformWHT; + + VP8VFilter16 = VFilter16; + VP8VFilter16i = VFilter16i; + VP8HFilter16 = HFilter16; +#if !defined(WORK_AROUND_GCC) + VP8HFilter16i = HFilter16i; #endif + VP8VFilter8 = VFilter8; + VP8VFilter8i = VFilter8i; +#if !defined(WORK_AROUND_GCC) + VP8HFilter8 = HFilter8; + VP8HFilter8i = HFilter8i; +#endif + VP8SimpleVFilter16 = SimpleVFilter16; + VP8SimpleHFilter16 = SimpleHFilter16; + VP8SimpleVFilter16i = SimpleVFilter16i; + VP8SimpleHFilter16i = SimpleHFilter16i; + + VP8PredLuma4[0] = DC4; + VP8PredLuma4[1] = TM4; + VP8PredLuma4[2] = VE4; + VP8PredLuma4[4] = RD4; + VP8PredLuma4[6] = LD4; + + VP8PredLuma16[0] = DC16TopLeft; + VP8PredLuma16[1] = TM16; + VP8PredLuma16[2] = VE16; + VP8PredLuma16[3] = HE16; + VP8PredLuma16[4] = DC16NoTop; + VP8PredLuma16[5] = DC16NoLeft; + VP8PredLuma16[6] = DC16NoTopLeft; + + VP8PredChroma8[0] = DC8uv; + VP8PredChroma8[1] = TM8uv; + VP8PredChroma8[2] = VE8uv; + VP8PredChroma8[3] = HE8uv; + VP8PredChroma8[4] = DC8uvNoTop; + VP8PredChroma8[5] = DC8uvNoLeft; + VP8PredChroma8[6] = DC8uvNoTopLeft; +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(VP8DspInitNEON) -#endif // WEBP_USE_NEON +#endif // WEBP_USE_NEON diff --git a/drivers/webp/dsp/dec_sse2.c b/drivers/webp/dsp/dec_sse2.c index 472b68ecb8..d4838b9210 100644 --- a/drivers/webp/dsp/dec_sse2.c +++ b/drivers/webp/dsp/dec_sse2.c @@ -1,8 +1,10 @@ // Copyright 2011 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // SSE2 version of some decoding functions (idct, loop filtering). @@ -14,17 +16,17 @@ #if defined(WEBP_USE_SSE2) +// The 3-coeff sparse transform in SSE2 is not really faster than the plain-C +// one it seems => disable it by default. Uncomment the following to enable: +// #define USE_TRANSFORM_AC3 + #include <emmintrin.h> #include "../dec/vp8i.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif - //------------------------------------------------------------------------------ // Transforms (Paragraph 14.4) -static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) { +static void Transform(const int16_t* in, uint8_t* dst, int do_two) { // This implementation makes use of 16-bit fixed point versions of two // multiply constants: // K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16 @@ -50,19 +52,19 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) { // vectors will just contain random value we'll never use nor store. __m128i in0, in1, in2, in3; { - in0 = _mm_loadl_epi64((__m128i*)&in[0]); - in1 = _mm_loadl_epi64((__m128i*)&in[4]); - in2 = _mm_loadl_epi64((__m128i*)&in[8]); - in3 = _mm_loadl_epi64((__m128i*)&in[12]); + in0 = _mm_loadl_epi64((const __m128i*)&in[0]); + in1 = _mm_loadl_epi64((const __m128i*)&in[4]); + in2 = _mm_loadl_epi64((const __m128i*)&in[8]); + in3 = _mm_loadl_epi64((const __m128i*)&in[12]); // a00 a10 a20 a30 x x x x // a01 a11 a21 a31 x x x x // a02 a12 a22 a32 x x x x // a03 a13 a23 a33 x x x x if (do_two) { - const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]); - const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]); - const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]); - const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]); + const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]); + const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]); + const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]); + const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]); in0 = _mm_unpacklo_epi64(in0, inB0); in1 = _mm_unpacklo_epi64(in1, inB1); in2 = _mm_unpacklo_epi64(in2, inB2); @@ -194,21 +196,21 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) { // Add inverse transform to 'dst' and store. { - const __m128i zero = _mm_set1_epi16(0); + const __m128i zero = _mm_setzero_si128(); // Load the reference(s). __m128i dst0, dst1, dst2, dst3; if (do_two) { // Load eight bytes/pixels per line. - dst0 = _mm_loadl_epi64((__m128i*)&dst[0 * BPS]); - dst1 = _mm_loadl_epi64((__m128i*)&dst[1 * BPS]); - dst2 = _mm_loadl_epi64((__m128i*)&dst[2 * BPS]); - dst3 = _mm_loadl_epi64((__m128i*)&dst[3 * BPS]); + dst0 = _mm_loadl_epi64((__m128i*)(dst + 0 * BPS)); + dst1 = _mm_loadl_epi64((__m128i*)(dst + 1 * BPS)); + dst2 = _mm_loadl_epi64((__m128i*)(dst + 2 * BPS)); + dst3 = _mm_loadl_epi64((__m128i*)(dst + 3 * BPS)); } else { // Load four bytes/pixels per line. - dst0 = _mm_cvtsi32_si128(*(int*)&dst[0 * BPS]); - dst1 = _mm_cvtsi32_si128(*(int*)&dst[1 * BPS]); - dst2 = _mm_cvtsi32_si128(*(int*)&dst[2 * BPS]); - dst3 = _mm_cvtsi32_si128(*(int*)&dst[3 * BPS]); + dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS)); + dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS)); + dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS)); + dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS)); } // Convert to 16b. dst0 = _mm_unpacklo_epi8(dst0, zero); @@ -228,20 +230,66 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) { // Store the results. if (do_two) { // Store eight bytes/pixels per line. - _mm_storel_epi64((__m128i*)&dst[0 * BPS], dst0); - _mm_storel_epi64((__m128i*)&dst[1 * BPS], dst1); - _mm_storel_epi64((__m128i*)&dst[2 * BPS], dst2); - _mm_storel_epi64((__m128i*)&dst[3 * BPS], dst3); + _mm_storel_epi64((__m128i*)(dst + 0 * BPS), dst0); + _mm_storel_epi64((__m128i*)(dst + 1 * BPS), dst1); + _mm_storel_epi64((__m128i*)(dst + 2 * BPS), dst2); + _mm_storel_epi64((__m128i*)(dst + 3 * BPS), dst3); } else { // Store four bytes/pixels per line. - *((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(dst0); - *((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(dst1); - *((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(dst2); - *((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(dst3); + *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0); + *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1); + *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2); + *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3); } } } +#if defined(USE_TRANSFORM_AC3) +#define MUL(a, b) (((a) * (b)) >> 16) +static void TransformAC3(const int16_t* in, uint8_t* dst) { + static const int kC1 = 20091 + (1 << 16); + static const int kC2 = 35468; + const __m128i A = _mm_set1_epi16(in[0] + 4); + const __m128i c4 = _mm_set1_epi16(MUL(in[4], kC2)); + const __m128i d4 = _mm_set1_epi16(MUL(in[4], kC1)); + const int c1 = MUL(in[1], kC2); + const int d1 = MUL(in[1], kC1); + const __m128i CD = _mm_set_epi16(0, 0, 0, 0, -d1, -c1, c1, d1); + const __m128i B = _mm_adds_epi16(A, CD); + const __m128i m0 = _mm_adds_epi16(B, d4); + const __m128i m1 = _mm_adds_epi16(B, c4); + const __m128i m2 = _mm_subs_epi16(B, c4); + const __m128i m3 = _mm_subs_epi16(B, d4); + const __m128i zero = _mm_setzero_si128(); + // Load the source pixels. + __m128i dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS)); + __m128i dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS)); + __m128i dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS)); + __m128i dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS)); + // Convert to 16b. + dst0 = _mm_unpacklo_epi8(dst0, zero); + dst1 = _mm_unpacklo_epi8(dst1, zero); + dst2 = _mm_unpacklo_epi8(dst2, zero); + dst3 = _mm_unpacklo_epi8(dst3, zero); + // Add the inverse transform. + dst0 = _mm_adds_epi16(dst0, _mm_srai_epi16(m0, 3)); + dst1 = _mm_adds_epi16(dst1, _mm_srai_epi16(m1, 3)); + dst2 = _mm_adds_epi16(dst2, _mm_srai_epi16(m2, 3)); + dst3 = _mm_adds_epi16(dst3, _mm_srai_epi16(m3, 3)); + // Unsigned saturate to 8b. + dst0 = _mm_packus_epi16(dst0, dst0); + dst1 = _mm_packus_epi16(dst1, dst1); + dst2 = _mm_packus_epi16(dst2, dst2); + dst3 = _mm_packus_epi16(dst3, dst3); + // Store the results. + *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0); + *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1); + *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2); + *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3); +} +#undef MUL +#endif // USE_TRANSFORM_AC3 + //------------------------------------------------------------------------------ // Loop Filter (Paragraph 15) @@ -250,20 +298,14 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) { _mm_subs_epu8((q), (p)), \ _mm_subs_epu8((p), (q))) -// Shift each byte of "a" by N bits while preserving by the sign bit. -// -// It first shifts the lower bytes of the words and then the upper bytes and -// then merges the results together. -#define SIGNED_SHIFT_N(a, N) { \ - __m128i t = a; \ - t = _mm_slli_epi16(t, 8); \ - t = _mm_srai_epi16(t, N); \ - t = _mm_srli_epi16(t, 8); \ - \ - a = _mm_srai_epi16(a, N + 8); \ - a = _mm_slli_epi16(a, 8); \ - \ - a = _mm_or_si128(t, a); \ +// Shift each byte of "x" by 3 bits while preserving by the sign bit. +static WEBP_INLINE void SignedShift8b(__m128i* const x) { + const __m128i zero = _mm_setzero_si128(); + const __m128i lo_0 = _mm_unpacklo_epi8(zero, *x); + const __m128i hi_0 = _mm_unpackhi_epi8(zero, *x); + const __m128i lo_1 = _mm_srai_epi16(lo_0, 3 + 8); + const __m128i hi_1 = _mm_srai_epi16(hi_0, 3 + 8); + *x = _mm_packs_epi16(lo_1, hi_1); } #define FLIP_SIGN_BIT2(a, b) { \ @@ -276,103 +318,124 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) { FLIP_SIGN_BIT2(c, d); \ } -#define GET_NOTHEV(p1, p0, q0, q1, hev_thresh, not_hev) { \ - const __m128i zero = _mm_setzero_si128(); \ - const __m128i t1 = MM_ABS(p1, p0); \ - const __m128i t2 = MM_ABS(q1, q0); \ - \ - const __m128i h = _mm_set1_epi8(hev_thresh); \ - const __m128i t3 = _mm_subs_epu8(t1, h); /* abs(p1 - p0) - hev_tresh */ \ - const __m128i t4 = _mm_subs_epu8(t2, h); /* abs(q1 - q0) - hev_tresh */ \ - \ - not_hev = _mm_or_si128(t3, t4); \ - not_hev = _mm_cmpeq_epi8(not_hev, zero); /* not_hev <= t1 && not_hev <= t2 */\ -} - -#define GET_BASE_DELTA(p1, p0, q0, q1, o) { \ - const __m128i qp0 = _mm_subs_epi8(q0, p0); /* q0 - p0 */ \ - o = _mm_subs_epi8(p1, q1); /* p1 - q1 */ \ - o = _mm_adds_epi8(o, qp0); /* p1 - q1 + 1 * (q0 - p0) */ \ - o = _mm_adds_epi8(o, qp0); /* p1 - q1 + 2 * (q0 - p0) */ \ - o = _mm_adds_epi8(o, qp0); /* p1 - q1 + 3 * (q0 - p0) */ \ -} - -#define DO_SIMPLE_FILTER(p0, q0, fl) { \ - const __m128i three = _mm_set1_epi8(3); \ - const __m128i four = _mm_set1_epi8(4); \ - __m128i v3 = _mm_adds_epi8(fl, three); \ - __m128i v4 = _mm_adds_epi8(fl, four); \ - \ - /* Do +4 side */ \ - SIGNED_SHIFT_N(v4, 3); /* v4 >> 3 */ \ - q0 = _mm_subs_epi8(q0, v4); /* q0 -= v4 */ \ - \ - /* Now do +3 side */ \ - SIGNED_SHIFT_N(v3, 3); /* v3 >> 3 */ \ - p0 = _mm_adds_epi8(p0, v3); /* p0 += v3 */ \ +// input/output is uint8_t +static WEBP_INLINE void GetNotHEV(const __m128i* const p1, + const __m128i* const p0, + const __m128i* const q0, + const __m128i* const q1, + int hev_thresh, __m128i* const not_hev) { + const __m128i zero = _mm_setzero_si128(); + const __m128i t_1 = MM_ABS(*p1, *p0); + const __m128i t_2 = MM_ABS(*q1, *q0); + + const __m128i h = _mm_set1_epi8(hev_thresh); + const __m128i t_max = _mm_max_epu8(t_1, t_2); + + const __m128i t_max_h = _mm_subs_epu8(t_max, h); + *not_hev = _mm_cmpeq_epi8(t_max_h, zero); // not_hev <= t1 && not_hev <= t2 } -// Updates values of 2 pixels at MB edge during complex filtering. -// Update operations: -// q = q - a and p = p + a; where a = [(a_hi >> 7), (a_lo >> 7)] -#define UPDATE_2PIXELS(pi, qi, a_lo, a_hi) { \ - const __m128i a_lo7 = _mm_srai_epi16(a_lo, 7); \ - const __m128i a_hi7 = _mm_srai_epi16(a_hi, 7); \ - const __m128i a = _mm_packs_epi16(a_lo7, a_hi7); \ - pi = _mm_adds_epi8(pi, a); \ - qi = _mm_subs_epi8(qi, a); \ +// input pixels are int8_t +static WEBP_INLINE void GetBaseDelta(const __m128i* const p1, + const __m128i* const p0, + const __m128i* const q0, + const __m128i* const q1, + __m128i* const delta) { + // beware of addition order, for saturation! + const __m128i p1_q1 = _mm_subs_epi8(*p1, *q1); // p1 - q1 + const __m128i q0_p0 = _mm_subs_epi8(*q0, *p0); // q0 - p0 + const __m128i s1 = _mm_adds_epi8(p1_q1, q0_p0); // p1 - q1 + 1 * (q0 - p0) + const __m128i s2 = _mm_adds_epi8(q0_p0, s1); // p1 - q1 + 2 * (q0 - p0) + const __m128i s3 = _mm_adds_epi8(q0_p0, s2); // p1 - q1 + 3 * (q0 - p0) + *delta = s3; } -static void NeedsFilter(const __m128i* p1, const __m128i* p0, const __m128i* q0, - const __m128i* q1, int thresh, __m128i *mask) { - __m128i t1 = MM_ABS(*p1, *q1); // abs(p1 - q1) - *mask = _mm_set1_epi8(0xFE); - t1 = _mm_and_si128(t1, *mask); // set lsb of each byte to zero - t1 = _mm_srli_epi16(t1, 1); // abs(p1 - q1) / 2 +// input and output are int8_t +static WEBP_INLINE void DoSimpleFilter(__m128i* const p0, __m128i* const q0, + const __m128i* const fl) { + const __m128i k3 = _mm_set1_epi8(3); + const __m128i k4 = _mm_set1_epi8(4); + __m128i v3 = _mm_adds_epi8(*fl, k3); + __m128i v4 = _mm_adds_epi8(*fl, k4); + + SignedShift8b(&v4); // v4 >> 3 + SignedShift8b(&v3); // v3 >> 3 + *q0 = _mm_subs_epi8(*q0, v4); // q0 -= v4 + *p0 = _mm_adds_epi8(*p0, v3); // p0 += v3 +} - *mask = MM_ABS(*p0, *q0); // abs(p0 - q0) - *mask = _mm_adds_epu8(*mask, *mask); // abs(p0 - q0) * 2 - *mask = _mm_adds_epu8(*mask, t1); // abs(p0 - q0) * 2 + abs(p1 - q1) / 2 +// Updates values of 2 pixels at MB edge during complex filtering. +// Update operations: +// q = q - delta and p = p + delta; where delta = [(a_hi >> 7), (a_lo >> 7)] +// Pixels 'pi' and 'qi' are int8_t on input, uint8_t on output (sign flip). +static WEBP_INLINE void Update2Pixels(__m128i* const pi, __m128i* const qi, + const __m128i* const a0_lo, + const __m128i* const a0_hi) { + const __m128i a1_lo = _mm_srai_epi16(*a0_lo, 7); + const __m128i a1_hi = _mm_srai_epi16(*a0_hi, 7); + const __m128i delta = _mm_packs_epi16(a1_lo, a1_hi); + const __m128i sign_bit = _mm_set1_epi8(0x80); + *pi = _mm_adds_epi8(*pi, delta); + *qi = _mm_subs_epi8(*qi, delta); + FLIP_SIGN_BIT2(*pi, *qi); +} - t1 = _mm_set1_epi8(thresh); - *mask = _mm_subs_epu8(*mask, t1); // mask <= thresh - *mask = _mm_cmpeq_epi8(*mask, _mm_setzero_si128()); +// input pixels are uint8_t +static WEBP_INLINE void NeedsFilter(const __m128i* const p1, + const __m128i* const p0, + const __m128i* const q0, + const __m128i* const q1, + int thresh, __m128i* const mask) { + const __m128i m_thresh = _mm_set1_epi8(thresh); + const __m128i t1 = MM_ABS(*p1, *q1); // abs(p1 - q1) + const __m128i kFE = _mm_set1_epi8(0xFE); + const __m128i t2 = _mm_and_si128(t1, kFE); // set lsb of each byte to zero + const __m128i t3 = _mm_srli_epi16(t2, 1); // abs(p1 - q1) / 2 + + const __m128i t4 = MM_ABS(*p0, *q0); // abs(p0 - q0) + const __m128i t5 = _mm_adds_epu8(t4, t4); // abs(p0 - q0) * 2 + const __m128i t6 = _mm_adds_epu8(t5, t3); // abs(p0-q0)*2 + abs(p1-q1)/2 + + const __m128i t7 = _mm_subs_epu8(t6, m_thresh); // mask <= m_thresh + *mask = _mm_cmpeq_epi8(t7, _mm_setzero_si128()); } //------------------------------------------------------------------------------ // Edge filtering functions // Applies filter on 2 pixels (p0 and q0) -static WEBP_INLINE void DoFilter2(const __m128i* p1, __m128i* p0, __m128i* q0, - const __m128i* q1, int thresh) { +static WEBP_INLINE void DoFilter2(__m128i* const p1, __m128i* const p0, + __m128i* const q0, __m128i* const q1, + int thresh) { __m128i a, mask; const __m128i sign_bit = _mm_set1_epi8(0x80); + // convert p1/q1 to int8_t (for GetBaseDelta) const __m128i p1s = _mm_xor_si128(*p1, sign_bit); const __m128i q1s = _mm_xor_si128(*q1, sign_bit); NeedsFilter(p1, p0, q0, q1, thresh, &mask); - // convert to signed values FLIP_SIGN_BIT2(*p0, *q0); - - GET_BASE_DELTA(p1s, *p0, *q0, q1s, a); + GetBaseDelta(&p1s, p0, q0, &q1s, &a); a = _mm_and_si128(a, mask); // mask filter values we don't care about - DO_SIMPLE_FILTER(*p0, *q0, a); - - // unoffset + DoSimpleFilter(p0, q0, &a); FLIP_SIGN_BIT2(*p0, *q0); } // Applies filter on 4 pixels (p1, p0, q0 and q1) -static WEBP_INLINE void DoFilter4(__m128i* p1, __m128i *p0, - __m128i* q0, __m128i* q1, - const __m128i* mask, int hev_thresh) { +static WEBP_INLINE void DoFilter4(__m128i* const p1, __m128i* const p0, + __m128i* const q0, __m128i* const q1, + const __m128i* const mask, int hev_thresh) { + const __m128i zero = _mm_setzero_si128(); + const __m128i sign_bit = _mm_set1_epi8(0x80); + const __m128i k64 = _mm_set1_epi8(64); + const __m128i k3 = _mm_set1_epi8(3); + const __m128i k4 = _mm_set1_epi8(4); __m128i not_hev; __m128i t1, t2, t3; - const __m128i sign_bit = _mm_set1_epi8(0x80); // compute hev mask - GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev); + GetNotHEV(p1, p0, q0, q1, hev_thresh, ¬_hev); // convert to signed values FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1); @@ -385,135 +448,115 @@ static WEBP_INLINE void DoFilter4(__m128i* p1, __m128i *p0, t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 3 * (q0 - p0) t1 = _mm_and_si128(t1, *mask); // mask filter values we don't care about - // Do +4 side - t2 = _mm_set1_epi8(4); - t2 = _mm_adds_epi8(t1, t2); // 3 * (q0 - p0) + (p1 - q1) + 4 - SIGNED_SHIFT_N(t2, 3); // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3 - t3 = t2; // save t2 - *q0 = _mm_subs_epi8(*q0, t2); // q0 -= t2 - - // Now do +3 side - t2 = _mm_set1_epi8(3); - t2 = _mm_adds_epi8(t1, t2); // +3 instead of +4 - SIGNED_SHIFT_N(t2, 3); // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3 + t2 = _mm_adds_epi8(t1, k3); // 3 * (q0 - p0) + hev(p1 - q1) + 3 + t3 = _mm_adds_epi8(t1, k4); // 3 * (q0 - p0) + hev(p1 - q1) + 4 + SignedShift8b(&t2); // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3 + SignedShift8b(&t3); // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3 *p0 = _mm_adds_epi8(*p0, t2); // p0 += t2 + *q0 = _mm_subs_epi8(*q0, t3); // q0 -= t3 + FLIP_SIGN_BIT2(*p0, *q0); - t2 = _mm_set1_epi8(1); - t3 = _mm_adds_epi8(t3, t2); - SIGNED_SHIFT_N(t3, 1); // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 4 + // this is equivalent to signed (a + 1) >> 1 calculation + t2 = _mm_add_epi8(t3, sign_bit); + t3 = _mm_avg_epu8(t2, zero); + t3 = _mm_sub_epi8(t3, k64); t3 = _mm_and_si128(not_hev, t3); // if !hev *q1 = _mm_subs_epi8(*q1, t3); // q1 -= t3 *p1 = _mm_adds_epi8(*p1, t3); // p1 += t3 - - // unoffset - FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1); + FLIP_SIGN_BIT2(*p1, *q1); } // Applies filter on 6 pixels (p2, p1, p0, q0, q1 and q2) -static WEBP_INLINE void DoFilter6(__m128i *p2, __m128i* p1, __m128i *p0, - __m128i* q0, __m128i* q1, __m128i *q2, - const __m128i* mask, int hev_thresh) { - __m128i a, not_hev; +static WEBP_INLINE void DoFilter6(__m128i* const p2, __m128i* const p1, + __m128i* const p0, __m128i* const q0, + __m128i* const q1, __m128i* const q2, + const __m128i* const mask, int hev_thresh) { + const __m128i zero = _mm_setzero_si128(); const __m128i sign_bit = _mm_set1_epi8(0x80); + __m128i a, not_hev; // compute hev mask - GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev); + GetNotHEV(p1, p0, q0, q1, hev_thresh, ¬_hev); - // convert to signed values FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1); FLIP_SIGN_BIT2(*p2, *q2); - - GET_BASE_DELTA(*p1, *p0, *q0, *q1, a); + GetBaseDelta(p1, p0, q0, q1, &a); { // do simple filter on pixels with hev const __m128i m = _mm_andnot_si128(not_hev, *mask); const __m128i f = _mm_and_si128(a, m); - DO_SIMPLE_FILTER(*p0, *q0, f); + DoSimpleFilter(p0, q0, &f); } + { // do strong filter on pixels with not hev - const __m128i zero = _mm_setzero_si128(); - const __m128i nine = _mm_set1_epi16(0x0900); - const __m128i sixty_three = _mm_set1_epi16(63); + const __m128i k9 = _mm_set1_epi16(0x0900); + const __m128i k63 = _mm_set1_epi16(63); const __m128i m = _mm_and_si128(not_hev, *mask); const __m128i f = _mm_and_si128(a, m); + const __m128i f_lo = _mm_unpacklo_epi8(zero, f); const __m128i f_hi = _mm_unpackhi_epi8(zero, f); - const __m128i f9_lo = _mm_mulhi_epi16(f_lo, nine); // Filter (lo) * 9 - const __m128i f9_hi = _mm_mulhi_epi16(f_hi, nine); // Filter (hi) * 9 - const __m128i f18_lo = _mm_add_epi16(f9_lo, f9_lo); // Filter (lo) * 18 - const __m128i f18_hi = _mm_add_epi16(f9_hi, f9_hi); // Filter (hi) * 18 + const __m128i f9_lo = _mm_mulhi_epi16(f_lo, k9); // Filter (lo) * 9 + const __m128i f9_hi = _mm_mulhi_epi16(f_hi, k9); // Filter (hi) * 9 - const __m128i a2_lo = _mm_add_epi16(f9_lo, sixty_three); // Filter * 9 + 63 - const __m128i a2_hi = _mm_add_epi16(f9_hi, sixty_three); // Filter * 9 + 63 + const __m128i a2_lo = _mm_add_epi16(f9_lo, k63); // Filter * 9 + 63 + const __m128i a2_hi = _mm_add_epi16(f9_hi, k63); // Filter * 9 + 63 - const __m128i a1_lo = _mm_add_epi16(f18_lo, sixty_three); // F... * 18 + 63 - const __m128i a1_hi = _mm_add_epi16(f18_hi, sixty_three); // F... * 18 + 63 + const __m128i a1_lo = _mm_add_epi16(a2_lo, f9_lo); // Filter * 18 + 63 + const __m128i a1_hi = _mm_add_epi16(a2_hi, f9_hi); // Filter * 18 + 63 - const __m128i a0_lo = _mm_add_epi16(f18_lo, a2_lo); // Filter * 27 + 63 - const __m128i a0_hi = _mm_add_epi16(f18_hi, a2_hi); // Filter * 27 + 63 + const __m128i a0_lo = _mm_add_epi16(a1_lo, f9_lo); // Filter * 27 + 63 + const __m128i a0_hi = _mm_add_epi16(a1_hi, f9_hi); // Filter * 27 + 63 - UPDATE_2PIXELS(*p2, *q2, a2_lo, a2_hi); - UPDATE_2PIXELS(*p1, *q1, a1_lo, a1_hi); - UPDATE_2PIXELS(*p0, *q0, a0_lo, a0_hi); + Update2Pixels(p2, q2, &a2_lo, &a2_hi); + Update2Pixels(p1, q1, &a1_lo, &a1_hi); + Update2Pixels(p0, q0, &a0_lo, &a0_hi); } +} - // unoffset - FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1); - FLIP_SIGN_BIT2(*p2, *q2); +// memcpy() is the safe way of moving potentially unaligned 32b memory. +static WEBP_INLINE uint32_t MemToUint32(const uint8_t* const ptr) { + uint32_t A; + memcpy(&A, (const int*)ptr, sizeof(A)); + return A; } // reads 8 rows across a vertical edge. -// -// TODO(somnath): Investigate _mm_shuffle* also see if it can be broken into -// two Load4x4() to avoid code duplication. -static WEBP_INLINE void Load8x4(const uint8_t* b, int stride, - __m128i* p, __m128i* q) { - __m128i t1, t2; - - // Load 0th, 1st, 4th and 5th rows - __m128i r0 = _mm_cvtsi32_si128(*((int*)&b[0 * stride])); // 03 02 01 00 - __m128i r1 = _mm_cvtsi32_si128(*((int*)&b[1 * stride])); // 13 12 11 10 - __m128i r4 = _mm_cvtsi32_si128(*((int*)&b[4 * stride])); // 43 42 41 40 - __m128i r5 = _mm_cvtsi32_si128(*((int*)&b[5 * stride])); // 53 52 51 50 - - r0 = _mm_unpacklo_epi32(r0, r4); // 43 42 41 40 03 02 01 00 - r1 = _mm_unpacklo_epi32(r1, r5); // 53 52 51 50 13 12 11 10 - - // t1 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00 - t1 = _mm_unpacklo_epi8(r0, r1); - - // Load 2nd, 3rd, 6th and 7th rows - r0 = _mm_cvtsi32_si128(*((int*)&b[2 * stride])); // 23 22 21 22 - r1 = _mm_cvtsi32_si128(*((int*)&b[3 * stride])); // 33 32 31 30 - r4 = _mm_cvtsi32_si128(*((int*)&b[6 * stride])); // 63 62 61 60 - r5 = _mm_cvtsi32_si128(*((int*)&b[7 * stride])); // 73 72 71 70 - - r0 = _mm_unpacklo_epi32(r0, r4); // 63 62 61 60 23 22 21 20 - r1 = _mm_unpacklo_epi32(r1, r5); // 73 72 71 70 33 32 31 30 - - // t2 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20 - t2 = _mm_unpacklo_epi8(r0, r1); - - // t1 = 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00 - // t2 = 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40 - r0 = t1; - t1 = _mm_unpacklo_epi16(t1, t2); - t2 = _mm_unpackhi_epi16(r0, t2); +static WEBP_INLINE void Load8x4(const uint8_t* const b, int stride, + __m128i* const p, __m128i* const q) { + // A0 = 63 62 61 60 23 22 21 20 43 42 41 40 03 02 01 00 + // A1 = 73 72 71 70 33 32 31 30 53 52 51 50 13 12 11 10 + const __m128i A0 = _mm_set_epi32( + MemToUint32(&b[6 * stride]), MemToUint32(&b[2 * stride]), + MemToUint32(&b[4 * stride]), MemToUint32(&b[0 * stride])); + const __m128i A1 = _mm_set_epi32( + MemToUint32(&b[7 * stride]), MemToUint32(&b[3 * stride]), + MemToUint32(&b[5 * stride]), MemToUint32(&b[1 * stride])); + + // B0 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00 + // B1 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20 + const __m128i B0 = _mm_unpacklo_epi8(A0, A1); + const __m128i B1 = _mm_unpackhi_epi8(A0, A1); + + // C0 = 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00 + // C1 = 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40 + const __m128i C0 = _mm_unpacklo_epi16(B0, B1); + const __m128i C1 = _mm_unpackhi_epi16(B0, B1); // *p = 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00 // *q = 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02 - *p = _mm_unpacklo_epi32(t1, t2); - *q = _mm_unpackhi_epi32(t1, t2); + *p = _mm_unpacklo_epi32(C0, C1); + *q = _mm_unpackhi_epi32(C0, C1); } -static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8, +static WEBP_INLINE void Load16x4(const uint8_t* const r0, + const uint8_t* const r8, int stride, - __m128i* p1, __m128i* p0, - __m128i* q0, __m128i* q1) { - __m128i t1, t2; + __m128i* const p1, __m128i* const p0, + __m128i* const q0, __m128i* const q1) { // Assume the pixels around the edge (|) are numbered as follows // 00 01 | 02 03 // 10 11 | 12 13 @@ -532,19 +575,21 @@ static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8, Load8x4(r0, stride, p1, q0); Load8x4(r8, stride, p0, q1); - t1 = *p1; - t2 = *q0; - // p1 = f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00 - // p0 = f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01 - // q0 = f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02 - // q1 = f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03 - *p1 = _mm_unpacklo_epi64(t1, *p0); - *p0 = _mm_unpackhi_epi64(t1, *p0); - *q0 = _mm_unpacklo_epi64(t2, *q1); - *q1 = _mm_unpackhi_epi64(t2, *q1); + { + // p1 = f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00 + // p0 = f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01 + // q0 = f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02 + // q1 = f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03 + const __m128i t1 = *p1; + const __m128i t2 = *q0; + *p1 = _mm_unpacklo_epi64(t1, *p0); + *p0 = _mm_unpackhi_epi64(t1, *p0); + *q0 = _mm_unpacklo_epi64(t2, *q1); + *q1 = _mm_unpackhi_epi64(t2, *q1); + } } -static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) { +static WEBP_INLINE void Store4x4(__m128i* const x, uint8_t* dst, int stride) { int i; for (i = 0; i < 4; ++i, dst += stride) { *((int32_t*)dst) = _mm_cvtsi128_si32(*x); @@ -553,48 +598,51 @@ static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) { } // Transpose back and store -static WEBP_INLINE void Store16x4(uint8_t* r0, uint8_t* r8, int stride, - __m128i* p1, __m128i* p0, - __m128i* q0, __m128i* q1) { - __m128i t1; +static WEBP_INLINE void Store16x4(const __m128i* const p1, + const __m128i* const p0, + const __m128i* const q0, + const __m128i* const q1, + uint8_t* r0, uint8_t* r8, + int stride) { + __m128i t1, p1_s, p0_s, q0_s, q1_s; // p0 = 71 70 61 60 51 50 41 40 31 30 21 20 11 10 01 00 // p1 = f1 f0 e1 e0 d1 d0 c1 c0 b1 b0 a1 a0 91 90 81 80 t1 = *p0; - *p0 = _mm_unpacklo_epi8(*p1, t1); - *p1 = _mm_unpackhi_epi8(*p1, t1); + p0_s = _mm_unpacklo_epi8(*p1, t1); + p1_s = _mm_unpackhi_epi8(*p1, t1); // q0 = 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02 // q1 = f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82 t1 = *q0; - *q0 = _mm_unpacklo_epi8(t1, *q1); - *q1 = _mm_unpackhi_epi8(t1, *q1); + q0_s = _mm_unpacklo_epi8(t1, *q1); + q1_s = _mm_unpackhi_epi8(t1, *q1); // p0 = 33 32 31 30 23 22 21 20 13 12 11 10 03 02 01 00 // q0 = 73 72 71 70 63 62 61 60 53 52 51 50 43 42 41 40 - t1 = *p0; - *p0 = _mm_unpacklo_epi16(t1, *q0); - *q0 = _mm_unpackhi_epi16(t1, *q0); + t1 = p0_s; + p0_s = _mm_unpacklo_epi16(t1, q0_s); + q0_s = _mm_unpackhi_epi16(t1, q0_s); // p1 = b3 b2 b1 b0 a3 a2 a1 a0 93 92 91 90 83 82 81 80 // q1 = f3 f2 f1 f0 e3 e2 e1 e0 d3 d2 d1 d0 c3 c2 c1 c0 - t1 = *p1; - *p1 = _mm_unpacklo_epi16(t1, *q1); - *q1 = _mm_unpackhi_epi16(t1, *q1); + t1 = p1_s; + p1_s = _mm_unpacklo_epi16(t1, q1_s); + q1_s = _mm_unpackhi_epi16(t1, q1_s); - Store4x4(p0, r0, stride); + Store4x4(&p0_s, r0, stride); r0 += 4 * stride; - Store4x4(q0, r0, stride); + Store4x4(&q0_s, r0, stride); - Store4x4(p1, r8, stride); + Store4x4(&p1_s, r8, stride); r8 += 4 * stride; - Store4x4(q1, r8, stride); + Store4x4(&q1_s, r8, stride); } //------------------------------------------------------------------------------ // Simple In-loop filtering (Paragraph 15.2) -static void SimpleVFilter16SSE2(uint8_t* p, int stride, int thresh) { +static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { // Load __m128i p1 = _mm_loadu_si128((__m128i*)&p[-2 * stride]); __m128i p0 = _mm_loadu_si128((__m128i*)&p[-stride]); @@ -605,49 +653,49 @@ static void SimpleVFilter16SSE2(uint8_t* p, int stride, int thresh) { // Store _mm_storeu_si128((__m128i*)&p[-stride], p0); - _mm_storeu_si128((__m128i*)p, q0); + _mm_storeu_si128((__m128i*)&p[0], q0); } -static void SimpleHFilter16SSE2(uint8_t* p, int stride, int thresh) { +static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { __m128i p1, p0, q0, q1; p -= 2; // beginning of p1 - Load16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1); + Load16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1); DoFilter2(&p1, &p0, &q0, &q1, thresh); - Store16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1); + Store16x4(&p1, &p0, &q0, &q1, p, p + 8 * stride, stride); } -static void SimpleVFilter16iSSE2(uint8_t* p, int stride, int thresh) { +static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) { int k; for (k = 3; k > 0; --k) { p += 4 * stride; - SimpleVFilter16SSE2(p, stride, thresh); + SimpleVFilter16(p, stride, thresh); } } -static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) { +static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) { int k; for (k = 3; k > 0; --k) { p += 4; - SimpleHFilter16SSE2(p, stride, thresh); + SimpleHFilter16(p, stride, thresh); } } //------------------------------------------------------------------------------ // Complex In-loop filtering (Paragraph 15.3) -#define MAX_DIFF1(p3, p2, p1, p0, m) { \ - m = MM_ABS(p3, p2); \ +#define MAX_DIFF1(p3, p2, p1, p0, m) do { \ + m = MM_ABS(p1, p0); \ + m = _mm_max_epu8(m, MM_ABS(p3, p2)); \ m = _mm_max_epu8(m, MM_ABS(p2, p1)); \ - m = _mm_max_epu8(m, MM_ABS(p1, p0)); \ -} +} while (0) -#define MAX_DIFF2(p3, p2, p1, p0, m) { \ +#define MAX_DIFF2(p3, p2, p1, p0, m) do { \ + m = _mm_max_epu8(m, MM_ABS(p1, p0)); \ m = _mm_max_epu8(m, MM_ABS(p3, p2)); \ m = _mm_max_epu8(m, MM_ABS(p2, p1)); \ - m = _mm_max_epu8(m, MM_ABS(p1, p0)); \ -} +} while (0) #define LOAD_H_EDGES4(p, stride, e1, e2, e3, e4) { \ e1 = _mm_loadu_si128((__m128i*)&(p)[0 * stride]); \ @@ -656,10 +704,11 @@ static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) { e4 = _mm_loadu_si128((__m128i*)&(p)[3 * stride]); \ } -#define LOADUV_H_EDGE(p, u, v, stride) { \ - p = _mm_loadl_epi64((__m128i*)&(u)[(stride)]); \ - p = _mm_unpacklo_epi64(p, _mm_loadl_epi64((__m128i*)&(v)[(stride)])); \ -} +#define LOADUV_H_EDGE(p, u, v, stride) do { \ + const __m128i U = _mm_loadl_epi64((__m128i*)&(u)[(stride)]); \ + const __m128i V = _mm_loadl_epi64((__m128i*)&(v)[(stride)]); \ + p = _mm_unpacklo_epi64(U, V); \ +} while (0) #define LOADUV_H_EDGES4(u, v, stride, e1, e2, e3, e4) { \ LOADUV_H_EDGE(e1, u, v, 0 * stride); \ @@ -674,18 +723,23 @@ static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) { _mm_storel_epi64((__m128i*)&v[(stride)], p); \ } -#define COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask) { \ - __m128i fl_yes; \ - const __m128i it = _mm_set1_epi8(ithresh); \ - mask = _mm_subs_epu8(mask, it); \ - mask = _mm_cmpeq_epi8(mask, _mm_setzero_si128()); \ - NeedsFilter(&p1, &p0, &q0, &q1, thresh, &fl_yes); \ - mask = _mm_and_si128(mask, fl_yes); \ +static WEBP_INLINE void ComplexMask(const __m128i* const p1, + const __m128i* const p0, + const __m128i* const q0, + const __m128i* const q1, + int thresh, int ithresh, + __m128i* const mask) { + const __m128i it = _mm_set1_epi8(ithresh); + const __m128i diff = _mm_subs_epu8(*mask, it); + const __m128i thresh_mask = _mm_cmpeq_epi8(diff, _mm_setzero_si128()); + __m128i filter_mask; + NeedsFilter(p1, p0, q0, q1, thresh, &filter_mask); + *mask = _mm_and_si128(thresh_mask, filter_mask); } // on macroblock edges -static void VFilter16SSE2(uint8_t* p, int stride, - int thresh, int ithresh, int hev_thresh) { +static void VFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i t1; __m128i mask; __m128i p2, p1, p0, q0, q1, q2; @@ -698,20 +752,20 @@ static void VFilter16SSE2(uint8_t* p, int stride, LOAD_H_EDGES4(p, stride, q0, q1, q2, t1); MAX_DIFF2(t1, q2, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh); // Store _mm_storeu_si128((__m128i*)&p[-3 * stride], p2); _mm_storeu_si128((__m128i*)&p[-2 * stride], p1); _mm_storeu_si128((__m128i*)&p[-1 * stride], p0); - _mm_storeu_si128((__m128i*)&p[0 * stride], q0); - _mm_storeu_si128((__m128i*)&p[1 * stride], q1); - _mm_storeu_si128((__m128i*)&p[2 * stride], q2); + _mm_storeu_si128((__m128i*)&p[+0 * stride], q0); + _mm_storeu_si128((__m128i*)&p[+1 * stride], q1); + _mm_storeu_si128((__m128i*)&p[+2 * stride], q2); } -static void HFilter16SSE2(uint8_t* p, int stride, - int thresh, int ithresh, int hev_thresh) { +static void HFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i p3, p2, p1, p0, q0, q1, q2, q3; @@ -722,71 +776,78 @@ static void HFilter16SSE2(uint8_t* p, int stride, Load16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3); // q0, q1, q2, q3 MAX_DIFF2(q3, q2, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh); - Store16x4(b, b + 8 * stride, stride, &p3, &p2, &p1, &p0); - Store16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3); + Store16x4(&p3, &p2, &p1, &p0, b, b + 8 * stride, stride); + Store16x4(&q0, &q1, &q2, &q3, p, p + 8 * stride, stride); } // on three inner edges -static void VFilter16iSSE2(uint8_t* p, int stride, - int thresh, int ithresh, int hev_thresh) { +static void VFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { int k; - __m128i mask; - __m128i t1, t2, p1, p0, q0, q1; + __m128i p3, p2, p1, p0; // loop invariants - for (k = 3; k > 0; --k) { - // Load p3, p2, p1, p0 - LOAD_H_EDGES4(p, stride, t2, t1, p1, p0); - MAX_DIFF1(t2, t1, p1, p0, mask); + LOAD_H_EDGES4(p, stride, p3, p2, p1, p0); // prologue + for (k = 3; k > 0; --k) { + __m128i mask, tmp1, tmp2; + uint8_t* const b = p + 2 * stride; // beginning of p1 p += 4 * stride; - // Load q0, q1, q2, q3 - LOAD_H_EDGES4(p, stride, q0, q1, t1, t2); - MAX_DIFF2(t2, t1, q1, q0, mask); + MAX_DIFF1(p3, p2, p1, p0, mask); // compute partial mask + LOAD_H_EDGES4(p, stride, p3, p2, tmp1, tmp2); + MAX_DIFF2(p3, p2, tmp1, tmp2, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); - DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh); + // p3 and p2 are not just temporary variables here: they will be + // re-used for next span. And q2/q3 will become p1/p0 accordingly. + ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask); + DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh); // Store - _mm_storeu_si128((__m128i*)&p[-2 * stride], p1); - _mm_storeu_si128((__m128i*)&p[-1 * stride], p0); - _mm_storeu_si128((__m128i*)&p[0 * stride], q0); - _mm_storeu_si128((__m128i*)&p[1 * stride], q1); + _mm_storeu_si128((__m128i*)&b[0 * stride], p1); + _mm_storeu_si128((__m128i*)&b[1 * stride], p0); + _mm_storeu_si128((__m128i*)&b[2 * stride], p3); + _mm_storeu_si128((__m128i*)&b[3 * stride], p2); + + // rotate samples + p1 = tmp1; + p0 = tmp2; } } -static void HFilter16iSSE2(uint8_t* p, int stride, - int thresh, int ithresh, int hev_thresh) { +static void HFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { int k; - uint8_t* b; - __m128i mask; - __m128i t1, t2, p1, p0, q0, q1; + __m128i p3, p2, p1, p0; // loop invariants + + Load16x4(p, p + 8 * stride, stride, &p3, &p2, &p1, &p0); // prologue for (k = 3; k > 0; --k) { - b = p; - Load16x4(b, b + 8 * stride, stride, &t2, &t1, &p1, &p0); // p3, p2, p1, p0 - MAX_DIFF1(t2, t1, p1, p0, mask); + __m128i mask, tmp1, tmp2; + uint8_t* const b = p + 2; // beginning of p1 - b += 4; // beginning of q0 - Load16x4(b, b + 8 * stride, stride, &q0, &q1, &t1, &t2); // q0, q1, q2, q3 - MAX_DIFF2(t2, t1, q1, q0, mask); + p += 4; // beginning of q0 (and next span) - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); - DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh); + MAX_DIFF1(p3, p2, p1, p0, mask); // compute partial mask + Load16x4(p, p + 8 * stride, stride, &p3, &p2, &tmp1, &tmp2); + MAX_DIFF2(p3, p2, tmp1, tmp2, mask); - b -= 2; // beginning of p1 - Store16x4(b, b + 8 * stride, stride, &p1, &p0, &q0, &q1); + ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask); + DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh); - p += 4; + Store16x4(&p1, &p0, &p3, &p2, b, b + 8 * stride, stride); + + // rotate samples + p1 = tmp1; + p0 = tmp2; } } // 8-pixels wide variant, for chroma filtering -static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride, - int thresh, int ithresh, int hev_thresh) { +static void VFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i t1, p2, p1, p0, q0, q1, q2; @@ -798,7 +859,7 @@ static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride, LOADUV_H_EDGES4(u, v, stride, q0, q1, q2, t1); MAX_DIFF2(t1, q2, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh); // Store @@ -810,8 +871,8 @@ static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride, STOREUV(q2, u, v, 2 * stride); } -static void HFilter8SSE2(uint8_t* u, uint8_t* v, int stride, - int thresh, int ithresh, int hev_thresh) { +static void HFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i p3, p2, p1, p0, q0, q1, q2, q3; @@ -823,15 +884,15 @@ static void HFilter8SSE2(uint8_t* u, uint8_t* v, int stride, Load16x4(u, v, stride, &q0, &q1, &q2, &q3); // q0, q1, q2, q3 MAX_DIFF2(q3, q2, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh); - Store16x4(tu, tv, stride, &p3, &p2, &p1, &p0); - Store16x4(u, v, stride, &q0, &q1, &q2, &q3); + Store16x4(&p3, &p2, &p1, &p0, tu, tv, stride); + Store16x4(&q0, &q1, &q2, &q3, u, v, stride); } -static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, - int thresh, int ithresh, int hev_thresh) { +static void VFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i t1, t2, p1, p0, q0, q1; @@ -846,7 +907,7 @@ static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, LOADUV_H_EDGES4(u, v, stride, q0, q1, t1, t2); MAX_DIFF2(t2, t1, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh); // Store @@ -856,8 +917,8 @@ static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, STOREUV(q1, u, v, 1 * stride); } -static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, - int thresh, int ithresh, int hev_thresh) { +static void HFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i t1, t2, p1, p0, q0, q1; Load16x4(u, v, stride, &t2, &t1, &p1, &p0); // p3, p2, p1, p0 @@ -868,36 +929,361 @@ static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, Load16x4(u, v, stride, &q0, &q1, &t1, &t2); // q0, q1, q2, q3 MAX_DIFF2(t2, t1, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh); u -= 2; // beginning of p1 v -= 2; - Store16x4(u, v, stride, &p1, &p0, &q0, &q1); + Store16x4(&p1, &p0, &q0, &q1, u, v, stride); } -extern void VP8DspInitSSE2(void); +//------------------------------------------------------------------------------ +// 4x4 predictions + +#define DST(x, y) dst[(x) + (y) * BPS] +#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2) + +// We use the following 8b-arithmetic tricks: +// (a + 2 * b + c + 2) >> 2 = (AC + b + 1) >> 1 +// where: AC = (a + c) >> 1 = [(a + c + 1) >> 1] - [(a^c) & 1] +// and: +// (a + 2 * b + c + 2) >> 2 = (AB + BC + 1) >> 1 - (ab|bc)&lsb +// where: AC = (a + b + 1) >> 1, BC = (b + c + 1) >> 1 +// and ab = a ^ b, bc = b ^ c, lsb = (AC^BC)&1 + +static void VE4(uint8_t* dst) { // vertical + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS - 1)); + const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2); + const __m128i a = _mm_avg_epu8(ABCDEFGH, CDEFGH00); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one); + const __m128i b = _mm_subs_epu8(a, lsb); + const __m128i avg = _mm_avg_epu8(b, BCDEFGH0); + const uint32_t vals = _mm_cvtsi128_si32(avg); + int i; + for (i = 0; i < 4; ++i) { + *(uint32_t*)(dst + i * BPS) = vals; + } +} + +static void LD4(uint8_t* dst) { // Down-Left + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS)); + const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2); + const __m128i CDEFGHH0 = _mm_insert_epi16(CDEFGH00, dst[-BPS + 7], 3); + const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, CDEFGHH0); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0); + *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( abcdefg ); + *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)); + *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)); + *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)); +} + +static void VR4(uint8_t* dst) { // Vertical-Right + const __m128i one = _mm_set1_epi8(1); + const int I = dst[-1 + 0 * BPS]; + const int J = dst[-1 + 1 * BPS]; + const int K = dst[-1 + 2 * BPS]; + const int X = dst[-1 - BPS]; + const __m128i XABCD = _mm_loadl_epi64((__m128i*)(dst - BPS - 1)); + const __m128i ABCD0 = _mm_srli_si128(XABCD, 1); + const __m128i abcd = _mm_avg_epu8(XABCD, ABCD0); + const __m128i _XABCD = _mm_slli_si128(XABCD, 1); + const __m128i IXABCD = _mm_insert_epi16(_XABCD, I | (X << 8), 0); + const __m128i avg1 = _mm_avg_epu8(IXABCD, ABCD0); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i efgh = _mm_avg_epu8(avg2, XABCD); + *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( abcd ); + *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32( efgh ); + *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)); + *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)); + + // these two are hard to implement in SSE2, so we keep the C-version: + DST(0, 2) = AVG3(J, I, X); + DST(0, 3) = AVG3(K, J, I); +} -void VP8DspInitSSE2(void) { - VP8Transform = TransformSSE2; +static void VL4(uint8_t* dst) { // Vertical-Left + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS)); + const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH__ = _mm_srli_si128(ABCDEFGH, 2); + const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, BCDEFGH_); + const __m128i avg2 = _mm_avg_epu8(CDEFGH__, BCDEFGH_); + const __m128i avg3 = _mm_avg_epu8(avg1, avg2); + const __m128i lsb1 = _mm_and_si128(_mm_xor_si128(avg1, avg2), one); + const __m128i ab = _mm_xor_si128(ABCDEFGH, BCDEFGH_); + const __m128i bc = _mm_xor_si128(CDEFGH__, BCDEFGH_); + const __m128i abbc = _mm_or_si128(ab, bc); + const __m128i lsb2 = _mm_and_si128(abbc, lsb1); + const __m128i avg4 = _mm_subs_epu8(avg3, lsb2); + const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4)); + *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( avg1 ); + *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32( avg4 ); + *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)); + *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)); + + // these two are hard to get and irregular + DST(3, 2) = (extra_out >> 0) & 0xff; + DST(3, 3) = (extra_out >> 8) & 0xff; +} + +static void RD4(uint8_t* dst) { // Down-right + const __m128i one = _mm_set1_epi8(1); + const __m128i XABCD = _mm_loadl_epi64((__m128i*)(dst - BPS - 1)); + const __m128i ____XABCD = _mm_slli_si128(XABCD, 4); + const uint32_t I = dst[-1 + 0 * BPS]; + const uint32_t J = dst[-1 + 1 * BPS]; + const uint32_t K = dst[-1 + 2 * BPS]; + const uint32_t L = dst[-1 + 3 * BPS]; + const __m128i LKJI_____ = + _mm_cvtsi32_si128(L | (K << 8) | (J << 16) | (I << 24)); + const __m128i LKJIXABCD = _mm_or_si128(LKJI_____, ____XABCD); + const __m128i KJIXABCD_ = _mm_srli_si128(LKJIXABCD, 1); + const __m128i JIXABCD__ = _mm_srli_si128(LKJIXABCD, 2); + const __m128i avg1 = _mm_avg_epu8(JIXABCD__, LKJIXABCD); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_); + *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32( abcdefg ); + *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)); + *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)); + *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)); +} + +#undef DST +#undef AVG3 + +//------------------------------------------------------------------------------ +// Luma 16x16 + +static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) { + const uint8_t* top = dst - BPS; + const __m128i zero = _mm_setzero_si128(); + int y; + if (size == 4) { + const __m128i top_values = _mm_cvtsi32_si128(MemToUint32(top)); + const __m128i top_base = _mm_unpacklo_epi8(top_values, zero); + for (y = 0; y < 4; ++y, dst += BPS) { + const int val = dst[-1] - top[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero); + *(int*)dst = _mm_cvtsi128_si32(out); + } + } else if (size == 8) { + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + const __m128i top_base = _mm_unpacklo_epi8(top_values, zero); + for (y = 0; y < 8; ++y, dst += BPS) { + const int val = dst[-1] - top[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero); + _mm_storel_epi64((__m128i*)dst, out); + } + } else { + const __m128i top_values = _mm_loadu_si128((const __m128i*)top); + const __m128i top_base_0 = _mm_unpacklo_epi8(top_values, zero); + const __m128i top_base_1 = _mm_unpackhi_epi8(top_values, zero); + for (y = 0; y < 16; ++y, dst += BPS) { + const int val = dst[-1] - top[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out_0 = _mm_add_epi16(base, top_base_0); + const __m128i out_1 = _mm_add_epi16(base, top_base_1); + const __m128i out = _mm_packus_epi16(out_0, out_1); + _mm_storeu_si128((__m128i*)dst, out); + } + } +} + +static void TM4(uint8_t* dst) { TrueMotion(dst, 4); } +static void TM8uv(uint8_t* dst) { TrueMotion(dst, 8); } +static void TM16(uint8_t* dst) { TrueMotion(dst, 16); } + +static void VE16(uint8_t* dst) { + const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS)); + int j; + for (j = 0; j < 16; ++j) { + _mm_storeu_si128((__m128i*)(dst + j * BPS), top); + } +} + +static void HE16(uint8_t* dst) { // horizontal + int j; + for (j = 16; j > 0; --j) { + const __m128i values = _mm_set1_epi8(dst[-1]); + _mm_storeu_si128((__m128i*)dst, values); + dst += BPS; + } +} + +static WEBP_INLINE void Put16(uint8_t v, uint8_t* dst) { + int j; + const __m128i values = _mm_set1_epi8(v); + for (j = 0; j < 16; ++j) { + _mm_storeu_si128((__m128i*)(dst + j * BPS), values); + } +} + +static void DC16(uint8_t* dst) { // DC + const __m128i zero = _mm_setzero_si128(); + const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS)); + const __m128i sad8x2 = _mm_sad_epu8(top, zero); + // sum the two sads: sad8x2[0:1] + sad8x2[8:9] + const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2)); + int left = 0; + int j; + for (j = 0; j < 16; ++j) { + left += dst[-1 + j * BPS]; + } + { + const int DC = _mm_cvtsi128_si32(sum) + left + 16; + Put16(DC >> 5, dst); + } +} + +static void DC16NoTop(uint8_t* dst) { // DC with top samples not available + int DC = 8; + int j; + for (j = 0; j < 16; ++j) { + DC += dst[-1 + j * BPS]; + } + Put16(DC >> 4, dst); +} + +static void DC16NoLeft(uint8_t* dst) { // DC with left samples not available + const __m128i zero = _mm_setzero_si128(); + const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS)); + const __m128i sad8x2 = _mm_sad_epu8(top, zero); + // sum the two sads: sad8x2[0:1] + sad8x2[8:9] + const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2)); + const int DC = _mm_cvtsi128_si32(sum) + 8; + Put16(DC >> 4, dst); +} - VP8VFilter16 = VFilter16SSE2; - VP8HFilter16 = HFilter16SSE2; - VP8VFilter8 = VFilter8SSE2; - VP8HFilter8 = HFilter8SSE2; - VP8VFilter16i = VFilter16iSSE2; - VP8HFilter16i = HFilter16iSSE2; - VP8VFilter8i = VFilter8iSSE2; - VP8HFilter8i = HFilter8iSSE2; +static void DC16NoTopLeft(uint8_t* dst) { // DC with no top and left samples + Put16(0x80, dst); +} + +//------------------------------------------------------------------------------ +// Chroma + +static void VE8uv(uint8_t* dst) { // vertical + int j; + const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS)); + for (j = 0; j < 8; ++j) { + _mm_storel_epi64((__m128i*)(dst + j * BPS), top); + } +} + +static void HE8uv(uint8_t* dst) { // horizontal + int j; + for (j = 0; j < 8; ++j) { + const __m128i values = _mm_set1_epi8(dst[-1]); + _mm_storel_epi64((__m128i*)dst, values); + dst += BPS; + } +} + +// helper for chroma-DC predictions +static WEBP_INLINE void Put8x8uv(uint8_t v, uint8_t* dst) { + int j; + const __m128i values = _mm_set1_epi8(v); + for (j = 0; j < 8; ++j) { + _mm_storel_epi64((__m128i*)(dst + j * BPS), values); + } +} - VP8SimpleVFilter16 = SimpleVFilter16SSE2; - VP8SimpleHFilter16 = SimpleHFilter16SSE2; - VP8SimpleVFilter16i = SimpleVFilter16iSSE2; - VP8SimpleHFilter16i = SimpleHFilter16iSSE2; +static void DC8uv(uint8_t* dst) { // DC + const __m128i zero = _mm_setzero_si128(); + const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS)); + const __m128i sum = _mm_sad_epu8(top, zero); + int left = 0; + int j; + for (j = 0; j < 8; ++j) { + left += dst[-1 + j * BPS]; + } + { + const int DC = _mm_cvtsi128_si32(sum) + left + 8; + Put8x8uv(DC >> 4, dst); + } } -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" +static void DC8uvNoLeft(uint8_t* dst) { // DC with no left samples + const __m128i zero = _mm_setzero_si128(); + const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS)); + const __m128i sum = _mm_sad_epu8(top, zero); + const int DC = _mm_cvtsi128_si32(sum) + 4; + Put8x8uv(DC >> 3, dst); +} + +static void DC8uvNoTop(uint8_t* dst) { // DC with no top samples + int dc0 = 4; + int i; + for (i = 0; i < 8; ++i) { + dc0 += dst[-1 + i * BPS]; + } + Put8x8uv(dc0 >> 3, dst); +} + +static void DC8uvNoTopLeft(uint8_t* dst) { // DC with nothing + Put8x8uv(0x80, dst); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8DspInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitSSE2(void) { + VP8Transform = Transform; +#if defined(USE_TRANSFORM_AC3) + VP8TransformAC3 = TransformAC3; #endif -#endif // WEBP_USE_SSE2 + VP8VFilter16 = VFilter16; + VP8HFilter16 = HFilter16; + VP8VFilter8 = VFilter8; + VP8HFilter8 = HFilter8; + VP8VFilter16i = VFilter16i; + VP8HFilter16i = HFilter16i; + VP8VFilter8i = VFilter8i; + VP8HFilter8i = HFilter8i; + + VP8SimpleVFilter16 = SimpleVFilter16; + VP8SimpleHFilter16 = SimpleHFilter16; + VP8SimpleVFilter16i = SimpleVFilter16i; + VP8SimpleHFilter16i = SimpleHFilter16i; + + VP8PredLuma4[1] = TM4; + VP8PredLuma4[2] = VE4; + VP8PredLuma4[4] = RD4; + VP8PredLuma4[5] = VR4; + VP8PredLuma4[6] = LD4; + VP8PredLuma4[7] = VL4; + + VP8PredLuma16[0] = DC16; + VP8PredLuma16[1] = TM16; + VP8PredLuma16[2] = VE16; + VP8PredLuma16[3] = HE16; + VP8PredLuma16[4] = DC16NoTop; + VP8PredLuma16[5] = DC16NoLeft; + VP8PredLuma16[6] = DC16NoTopLeft; + + VP8PredChroma8[0] = DC8uv; + VP8PredChroma8[1] = TM8uv; + VP8PredChroma8[2] = VE8uv; + VP8PredChroma8[3] = HE8uv; + VP8PredChroma8[4] = DC8uvNoTop; + VP8PredChroma8[5] = DC8uvNoLeft; + VP8PredChroma8[6] = DC8uvNoTopLeft; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8DspInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/dec_sse41.c b/drivers/webp/dsp/dec_sse41.c new file mode 100644 index 0000000000..dc1e70428d --- /dev/null +++ b/drivers/webp/dsp/dec_sse41.c @@ -0,0 +1,45 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE4 version of some decoding functions. +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE41) + +#include <smmintrin.h> +#include "../dec/vp8i.h" + +static void HE16(uint8_t* dst) { // horizontal + int j; + const __m128i kShuffle3 = _mm_set1_epi8(3); + for (j = 16; j > 0; --j) { + const __m128i in = _mm_cvtsi32_si128(*(int*)(dst - 4)); + const __m128i values = _mm_shuffle_epi8(in, kShuffle3); + _mm_storeu_si128((__m128i*)dst, values); + dst += BPS; + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8DspInitSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitSSE41(void) { + VP8PredLuma16[3] = HE16; +} + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(VP8DspInitSSE41) + +#endif // WEBP_USE_SSE41 diff --git a/drivers/webp/dsp/dsp.h b/drivers/webp/dsp/dsp.h index fd686a8532..4613d9c3ff 100644 --- a/drivers/webp/dsp/dsp.h +++ b/drivers/webp/dsp/dsp.h @@ -1,8 +1,10 @@ // Copyright 2011 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // Speed-critical functions. @@ -12,44 +14,121 @@ #ifndef WEBP_DSP_DSP_H_ #define WEBP_DSP_DSP_H_ -#include "../types.h" +#ifdef HAVE_CONFIG_H +#include "webp/config.h" +#endif + +#include "webp/types.h" -#if defined(__cplusplus) || defined(c_plusplus) +#ifdef __cplusplus extern "C" { #endif +#define BPS 32 // this is the common stride for enc/dec + //------------------------------------------------------------------------------ // CPU detection -#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) +#if defined(__GNUC__) +# define LOCAL_GCC_VERSION ((__GNUC__ << 8) | __GNUC_MINOR__) +# define LOCAL_GCC_PREREQ(maj, min) \ + (LOCAL_GCC_VERSION >= (((maj) << 8) | (min))) +#else +# define LOCAL_GCC_VERSION 0 +# define LOCAL_GCC_PREREQ(maj, min) 0 +#endif + +#ifndef __has_builtin +# define __has_builtin(x) 0 +#endif + +#if defined(_MSC_VER) && _MSC_VER > 1310 && \ + (defined(_M_X64) || defined(_M_IX86)) #define WEBP_MSC_SSE2 // Visual C++ SSE2 targets #endif -#if defined(__SSE2__) || defined(WEBP_MSC_SSE2) +#if defined(_MSC_VER) && _MSC_VER >= 1500 && \ + (defined(_M_X64) || defined(_M_IX86)) +#define WEBP_MSC_SSE41 // Visual C++ SSE4.1 targets +#endif + +// WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp +// files without intrinsics, allowing the corresponding Init() to be called. +// Files containing intrinsics will need to be built targeting the instruction +// set so should succeed on one of the earlier tests. +#if defined(__SSE2__) || defined(WEBP_MSC_SSE2) || defined(WEBP_HAVE_SSE2) #define WEBP_USE_SSE2 #endif -#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__) && defined(__ARM_NEON__) +#if defined(__SSE4_1__) || defined(WEBP_MSC_SSE41) || defined(WEBP_HAVE_SSE41) +#define WEBP_USE_SSE41 +#endif + +#if defined(__AVX2__) || defined(WEBP_HAVE_AVX2) +#define WEBP_USE_AVX2 +#endif + +#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__) #define WEBP_ANDROID_NEON // Android targets that might support NEON #endif -#if ( (defined(__ARM_NEON__) && !defined(__aarch64__)) || defined(WEBP_ANDROID_NEON)) && !defined(PSP2_ENABLED) +// The intrinsics currently cause compiler errors with arm-nacl-gcc and the +// inline assembly would need to be modified for use with Native Client. +#if (defined(__ARM_NEON__) || defined(WEBP_ANDROID_NEON) || \ + defined(__aarch64__)) && !defined(__native_client__) +#define WEBP_USE_NEON +#endif + +#if defined(_MSC_VER) && _MSC_VER >= 1700 && defined(_M_ARM) #define WEBP_USE_NEON +#define WEBP_USE_INTRINSICS +#endif + +#if defined(__mips__) && !defined(__mips64) && \ + defined(__mips_isa_rev) && (__mips_isa_rev >= 1) && (__mips_isa_rev < 6) +#define WEBP_USE_MIPS32 +#if (__mips_isa_rev >= 2) +#define WEBP_USE_MIPS32_R2 +#if defined(__mips_dspr2) || (__mips_dsp_rev >= 2) +#define WEBP_USE_MIPS_DSP_R2 +#endif +#endif +#endif + +// This macro prevents thread_sanitizer from reporting known concurrent writes. +#define WEBP_TSAN_IGNORE_FUNCTION +#if defined(__has_feature) +#if __has_feature(thread_sanitizer) +#undef WEBP_TSAN_IGNORE_FUNCTION +#define WEBP_TSAN_IGNORE_FUNCTION __attribute__((no_sanitize_thread)) +#endif #endif typedef enum { kSSE2, kSSE3, - kNEON + kSSE4_1, + kAVX, + kAVX2, + kNEON, + kMIPS32, + kMIPSdspR2 } CPUFeature; // returns true if the CPU supports the feature. typedef int (*VP8CPUInfo)(CPUFeature feature); -extern VP8CPUInfo VP8GetCPUInfo; +WEBP_EXTERN(VP8CPUInfo) VP8GetCPUInfo; //------------------------------------------------------------------------------ -// Encoding +// Init stub generator -int VP8GetAlpha(const int histo[]); +// Defines an init function stub to ensure each module exposes a symbol, +// avoiding a compiler warning. +#define WEBP_DSP_INIT_STUB(func) \ + extern void func(void); \ + WEBP_TSAN_IGNORE_FUNCTION void func(void) {} + +//------------------------------------------------------------------------------ +// Encoding // Transforms // VP8Idct: Does one of two inverse transforms. If do_two is set, the transforms @@ -60,7 +139,7 @@ typedef void (*VP8Fdct)(const uint8_t* src, const uint8_t* ref, int16_t* out); typedef void (*VP8WHT)(const int16_t* in, int16_t* out); extern VP8Idct VP8ITransform; extern VP8Fdct VP8FTransform; -extern VP8WHT VP8ITransformWHT; +extern VP8Fdct VP8FTransform2; // performs two transforms at a time extern VP8WHT VP8FTransformWHT; // Predictions // *dst is the destination block. *top and *left can be NULL. @@ -79,20 +158,63 @@ extern VP8WMetric VP8TDisto4x4, VP8TDisto16x16; typedef void (*VP8BlockCopy)(const uint8_t* src, uint8_t* dst); extern VP8BlockCopy VP8Copy4x4; +extern VP8BlockCopy VP8Copy16x8; // Quantization struct VP8Matrix; // forward declaration typedef int (*VP8QuantizeBlock)(int16_t in[16], int16_t out[16], - int n, const struct VP8Matrix* const mtx); + const struct VP8Matrix* const mtx); +// Same as VP8QuantizeBlock, but quantizes two consecutive blocks. +typedef int (*VP8Quantize2Blocks)(int16_t in[32], int16_t out[32], + const struct VP8Matrix* const mtx); + extern VP8QuantizeBlock VP8EncQuantizeBlock; +extern VP8Quantize2Blocks VP8EncQuantize2Blocks; + +// specific to 2nd transform: +typedef int (*VP8QuantizeBlockWHT)(int16_t in[16], int16_t out[16], + const struct VP8Matrix* const mtx); +extern VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT; -// Compute susceptibility based on DCT-coeff histograms: -// the higher, the "easier" the macroblock is to compress. -typedef int (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred, - int start_block, int end_block); extern const int VP8DspScan[16 + 4 + 4]; + +// Collect histogram for susceptibility calculation. +#define MAX_COEFF_THRESH 31 // size of histogram used by CollectHistogram. +typedef struct { + // We only need to store max_value and last_non_zero, not the distribution. + int max_value; + int last_non_zero; +} VP8Histogram; +typedef void (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo); extern VP8CHisto VP8CollectHistogram; +// General-purpose util function to help VP8CollectHistogram(). +void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1], + VP8Histogram* const histo); + +// must be called before using any of the above +void VP8EncDspInit(void); + +//------------------------------------------------------------------------------ +// cost functions (encoding) + +extern const uint16_t VP8EntropyCost[256]; // 8bit fixed-point log(p) +// approximate cost per level: +extern const uint16_t VP8LevelFixedCosts[2047 /*MAX_LEVEL*/ + 1]; +extern const uint8_t VP8EncBands[16 + 1]; + +struct VP8Residual; +typedef void (*VP8SetResidualCoeffsFunc)(const int16_t* const coeffs, + struct VP8Residual* const res); +extern VP8SetResidualCoeffsFunc VP8SetResidualCoeffs; -void VP8EncDspInit(void); // must be called before using any of the above +// Cost calculation function. +typedef int (*VP8GetResidualCostFunc)(int ctx0, + const struct VP8Residual* const res); +extern VP8GetResidualCostFunc VP8GetResidualCost; + +// must be called before anything using the above +void VP8EncDspCostInit(void); //------------------------------------------------------------------------------ // Decoding @@ -101,17 +223,26 @@ typedef void (*VP8DecIdct)(const int16_t* coeffs, uint8_t* dst); // when doing two transforms, coeffs is actually int16_t[2][16]. typedef void (*VP8DecIdct2)(const int16_t* coeffs, uint8_t* dst, int do_two); extern VP8DecIdct2 VP8Transform; +extern VP8DecIdct VP8TransformAC3; extern VP8DecIdct VP8TransformUV; extern VP8DecIdct VP8TransformDC; extern VP8DecIdct VP8TransformDCUV; -extern void (*VP8TransformWHT)(const int16_t* in, int16_t* out); +extern VP8WHT VP8TransformWHT; // *dst is the destination block, with stride BPS. Boundary samples are // assumed accessible when needed. typedef void (*VP8PredFunc)(uint8_t* dst); -extern const VP8PredFunc VP8PredLuma16[/* NUM_B_DC_MODES */]; -extern const VP8PredFunc VP8PredChroma8[/* NUM_B_DC_MODES */]; -extern const VP8PredFunc VP8PredLuma4[/* NUM_BMODES */]; +extern VP8PredFunc VP8PredLuma16[/* NUM_B_DC_MODES */]; +extern VP8PredFunc VP8PredChroma8[/* NUM_B_DC_MODES */]; +extern VP8PredFunc VP8PredLuma4[/* NUM_BMODES */]; + +// clipping tables (for filtering) +extern const int8_t* const VP8ksclip1; // clips [-1020, 1020] to [-128, 127] +extern const int8_t* const VP8ksclip2; // clips [-112, 112] to [-16, 15] +extern const uint8_t* const VP8kclip1; // clips [-255,511] to [0,255] +extern const uint8_t* const VP8kabs0; // abs(x) for x in [-255,255] +// must be called first +void VP8InitClipTables(void); // simple filter (only for luma) typedef void (*VP8SimpleFilterFunc)(uint8_t* p, int stride, int thresh); @@ -145,6 +276,8 @@ void VP8DspInit(void); #define FANCY_UPSAMPLING // undefined to remove fancy upsampling support +// Convert a pair of y/u/v lines together to the output rgb/a colorspace. +// bottom_y can be NULL if only one line of output is needed (at top/bottom). typedef void (*WebPUpsampleLinePairFunc)( const uint8_t* top_y, const uint8_t* bottom_y, const uint8_t* top_u, const uint8_t* top_v, @@ -156,18 +289,20 @@ typedef void (*WebPUpsampleLinePairFunc)( // Fancy upsampling functions to convert YUV to RGB(A) modes extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; -// Initializes SSE2 version of the fancy upsamplers. -void WebPInitUpsamplersSSE2(void); - #endif // FANCY_UPSAMPLING -// Point-sampling methods. -typedef void (*WebPSampleLinePairFunc)( - const uint8_t* top_y, const uint8_t* bottom_y, - const uint8_t* u, const uint8_t* v, - uint8_t* top_dst, uint8_t* bottom_dst, int len); +// Per-row point-sampling methods. +typedef void (*WebPSamplerRowFunc)(const uint8_t* y, + const uint8_t* u, const uint8_t* v, + uint8_t* dst, int len); +// Generic function to apply 'WebPSamplerRowFunc' to the whole plane: +void WebPSamplerProcessPlane(const uint8_t* y, int y_stride, + const uint8_t* u, const uint8_t* v, int uv_stride, + uint8_t* dst, int dst_stride, + int width, int height, WebPSamplerRowFunc func); -extern const WebPSampleLinePairFunc WebPSamplers[/* MODE_LAST */]; +// Sampling functions to convert rows of YUV to RGB(A) +extern WebPSamplerRowFunc WebPSamplers[/* MODE_LAST */]; // General function for converting two lines of ARGB or RGBA. // 'alpha_is_last' should be true if 0xff000000 is stored in memory as @@ -179,13 +314,84 @@ typedef void (*WebPYUV444Converter)(const uint8_t* y, const uint8_t* u, const uint8_t* v, uint8_t* dst, int len); -extern const WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */]; +extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */]; -// Main function to be called +// Must be called before using the WebPUpsamplers[] (and for premultiplied +// colorspaces like rgbA, rgbA4444, etc) void WebPInitUpsamplers(void); +// Must be called before using WebPSamplers[] +void WebPInitSamplers(void); +// Must be called before using WebPYUV444Converters[] +void WebPInitYUV444Converters(void); //------------------------------------------------------------------------------ -// Pre-multiply planes with alpha values +// ARGB -> YUV converters + +// Convert ARGB samples to luma Y. +extern void (*WebPConvertARGBToY)(const uint32_t* argb, uint8_t* y, int width); +// Convert ARGB samples to U/V with downsampling. do_store should be '1' for +// even lines and '0' for odd ones. 'src_width' is the original width, not +// the U/V one. +extern void (*WebPConvertARGBToUV)(const uint32_t* argb, uint8_t* u, uint8_t* v, + int src_width, int do_store); + +// Convert a row of accumulated (four-values) of rgba32 toward U/V +extern void (*WebPConvertRGBA32ToUV)(const uint16_t* rgb, + uint8_t* u, uint8_t* v, int width); + +// Convert RGB or BGR to Y +extern void (*WebPConvertRGB24ToY)(const uint8_t* rgb, uint8_t* y, int width); +extern void (*WebPConvertBGR24ToY)(const uint8_t* bgr, uint8_t* y, int width); + +// used for plain-C fallback. +extern void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v, + int src_width, int do_store); +extern void WebPConvertRGBA32ToUV_C(const uint16_t* rgb, + uint8_t* u, uint8_t* v, int width); + +// Must be called before using the above. +void WebPInitConvertARGBToYUV(void); + +//------------------------------------------------------------------------------ +// Rescaler + +struct WebPRescaler; + +// Import a row of data and save its contribution in the rescaler. +// 'channel' denotes the channel number to be imported. 'Expand' corresponds to +// the wrk->x_expand case. Otherwise, 'Shrink' is to be used. +typedef void (*WebPRescalerImportRowFunc)(struct WebPRescaler* const wrk, + const uint8_t* src); + +extern WebPRescalerImportRowFunc WebPRescalerImportRowExpand; +extern WebPRescalerImportRowFunc WebPRescalerImportRowShrink; + +// Export one row (starting at x_out position) from rescaler. +// 'Expand' corresponds to the wrk->y_expand case. +// Otherwise 'Shrink' is to be used +typedef void (*WebPRescalerExportRowFunc)(struct WebPRescaler* const wrk); +extern WebPRescalerExportRowFunc WebPRescalerExportRowExpand; +extern WebPRescalerExportRowFunc WebPRescalerExportRowShrink; + +// Plain-C implementation, as fall-back. +extern void WebPRescalerImportRowExpandC(struct WebPRescaler* const wrk, + const uint8_t* src); +extern void WebPRescalerImportRowShrinkC(struct WebPRescaler* const wrk, + const uint8_t* src); +extern void WebPRescalerExportRowExpandC(struct WebPRescaler* const wrk); +extern void WebPRescalerExportRowShrinkC(struct WebPRescaler* const wrk); + +// Main entry calls: +extern void WebPRescalerImportRow(struct WebPRescaler* const wrk, + const uint8_t* src); +// Export one row (starting at x_out position) from rescaler. +extern void WebPRescalerExportRow(struct WebPRescaler* const wrk); + +// Must be called first before using the above. +void WebPRescalerDspInit(void); + +//------------------------------------------------------------------------------ +// Utilities for processing transparent channel. // Apply alpha pre-multiply on an rgba, bgra or argb plane of size w * h. // alpha_first should be 0 for argb, 1 for rgba or bgra (where alpha is last). @@ -196,14 +402,98 @@ extern void (*WebPApplyAlphaMultiply)( extern void (*WebPApplyAlphaMultiply4444)( uint8_t* rgba4444, int w, int h, int stride); +// Dispatch the values from alpha[] plane to the ARGB destination 'dst'. +// Returns true if alpha[] plane has non-trivial values different from 0xff. +extern int (*WebPDispatchAlpha)(const uint8_t* alpha, int alpha_stride, + int width, int height, + uint8_t* dst, int dst_stride); + +// Transfer packed 8b alpha[] values to green channel in dst[], zero'ing the +// A/R/B values. 'dst_stride' is the stride for dst[] in uint32_t units. +extern void (*WebPDispatchAlphaToGreen)(const uint8_t* alpha, int alpha_stride, + int width, int height, + uint32_t* dst, int dst_stride); + +// Extract the alpha values from 32b values in argb[] and pack them into alpha[] +// (this is the opposite of WebPDispatchAlpha). +// Returns true if there's only trivial 0xff alpha values. +extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride, + int width, int height, + uint8_t* alpha, int alpha_stride); + +// Pre-Multiply operation transforms x into x * A / 255 (where x=Y,R,G or B). +// Un-Multiply operation transforms x into x * 255 / A. + +// Pre-Multiply or Un-Multiply (if 'inverse' is true) argb values in a row. +extern void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse); + +// Same a WebPMultARGBRow(), but for several rows. +void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows, + int inverse); + +// Same for a row of single values, with side alpha values. +extern void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha, + int width, int inverse); + +// Same a WebPMultRow(), but for several 'num_rows' rows. +void WebPMultRows(uint8_t* ptr, int stride, + const uint8_t* alpha, int alpha_stride, + int width, int num_rows, int inverse); + +// Plain-C versions, used as fallback by some implementations. +void WebPMultRowC(uint8_t* const ptr, const uint8_t* const alpha, + int width, int inverse); +void WebPMultARGBRowC(uint32_t* const ptr, int width, int inverse); + // To be called first before using the above. -void WebPInitPremultiply(void); +void WebPInitAlphaProcessing(void); + +// ARGB packing function: a/r/g/b input is rgba or bgra order. +extern void (*VP8PackARGB)(const uint8_t* a, const uint8_t* r, + const uint8_t* g, const uint8_t* b, int len, + uint32_t* out); + +// RGB packing function. 'step' can be 3 or 4. r/g/b input is rgb or bgr order. +extern void (*VP8PackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b, + int len, int step, uint32_t* out); -void WebPInitPremultiplySSE2(void); // should not be called directly. +// To be called first before using the above. +void VP8EncDspARGBInit(void); //------------------------------------------------------------------------------ +// Filter functions + +typedef enum { // Filter types. + WEBP_FILTER_NONE = 0, + WEBP_FILTER_HORIZONTAL, + WEBP_FILTER_VERTICAL, + WEBP_FILTER_GRADIENT, + WEBP_FILTER_LAST = WEBP_FILTER_GRADIENT + 1, // end marker + WEBP_FILTER_BEST, // meta-types + WEBP_FILTER_FAST +} WEBP_FILTER_TYPE; + +typedef void (*WebPFilterFunc)(const uint8_t* in, int width, int height, + int stride, uint8_t* out); +typedef void (*WebPUnfilterFunc)(int width, int height, int stride, + int row, int num_rows, uint8_t* data); + +// Filter the given data using the given predictor. +// 'in' corresponds to a 2-dimensional pixel array of size (stride * height) +// in raster order. +// 'stride' is number of bytes per scan line (with possible padding). +// 'out' should be pre-allocated. +extern WebPFilterFunc WebPFilters[WEBP_FILTER_LAST]; + +// In-place reconstruct the original data from the given filtered data. +// The reconstruction will be done for 'num_rows' rows starting from 'row' +// (assuming rows upto 'row - 1' are already reconstructed). +extern WebPUnfilterFunc WebPUnfilters[WEBP_FILTER_LAST]; + +// To be called first before using the above. +void VP8FiltersInit(void); -#if defined(__cplusplus) || defined(c_plusplus) +#ifdef __cplusplus } // extern "C" #endif diff --git a/drivers/webp/dsp/enc.c b/drivers/webp/dsp/enc.c index 02234564be..95e63f89ab 100644 --- a/drivers/webp/dsp/enc.c +++ b/drivers/webp/dsp/enc.c @@ -1,47 +1,34 @@ // Copyright 2011 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // Speed-critical encoding functions. // // Author: Skal (pascal.massimino@gmail.com) +#include <assert.h> #include <stdlib.h> // for abs() + #include "./dsp.h" #include "../enc/vp8enci.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif +static WEBP_INLINE uint8_t clip_8b(int v) { + return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255; +} + +static WEBP_INLINE int clip_max(int v, int max) { + return (v > max) ? max : v; +} //------------------------------------------------------------------------------ // Compute susceptibility based on DCT-coeff histograms: // the higher, the "easier" the macroblock is to compress. -static int ClipAlpha(int alpha) { - return alpha < 0 ? 0 : alpha > 255 ? 255 : alpha; -} - -int VP8GetAlpha(const int histo[MAX_COEFF_THRESH + 1]) { - int num = 0, den = 0, val = 0; - int k; - int alpha; - // note: changing this loop to avoid the numerous "k + 1" slows things down. - for (k = 0; k < MAX_COEFF_THRESH; ++k) { - if (histo[k + 1]) { - val += histo[k + 1]; - num += val * (k + 1); - den += (k + 1) * (k + 1); - } - } - // we scale the value to a usable [0..255] range - alpha = den ? 10 * num / den - 5 : 0; - return ClipAlpha(alpha); -} - const int VP8DspScan[16 + 4 + 4] = { // Luma 0 + 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS, @@ -53,27 +40,41 @@ const int VP8DspScan[16 + 4 + 4] = { 8 + 0 * BPS, 12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS // V }; -static int CollectHistogram(const uint8_t* ref, const uint8_t* pred, - int start_block, int end_block) { - int histo[MAX_COEFF_THRESH + 1] = { 0 }; - int16_t out[16]; - int j, k; +// general-purpose util function +void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1], + VP8Histogram* const histo) { + int max_value = 0, last_non_zero = 1; + int k; + for (k = 0; k <= MAX_COEFF_THRESH; ++k) { + const int value = distribution[k]; + if (value > 0) { + if (value > max_value) max_value = value; + last_non_zero = k; + } + } + histo->max_value = max_value; + histo->last_non_zero = last_non_zero; +} + +static void CollectHistogram(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; for (j = start_block; j < end_block; ++j) { - VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + int k; + int16_t out[16]; - // Convert coefficients to bin (within out[]). - for (k = 0; k < 16; ++k) { - const int v = abs(out[k]) >> 2; - out[k] = (v > MAX_COEFF_THRESH) ? MAX_COEFF_THRESH : v; - } + VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); - // Use bin to update histogram. + // Convert coefficients to bin. for (k = 0; k < 16; ++k) { - histo[out[k]]++; + const int v = abs(out[k]) >> 3; // TODO(skal): add rounding? + const int clipped_value = clip_max(v, MAX_COEFF_THRESH); + ++distribution[clipped_value]; } } - - return VP8GetAlpha(histo); + VP8SetHistogramData(distribution, histo); } //------------------------------------------------------------------------------ @@ -85,19 +86,16 @@ static uint8_t clip1[255 + 510 + 1]; // clips [-255,510] to [0,255] // and make sure it's set to true _last_ (so as to be thread-safe) static volatile int tables_ok = 0; -static void InitTables(void) { +static WEBP_TSAN_IGNORE_FUNCTION void InitTables(void) { if (!tables_ok) { int i; for (i = -255; i <= 255 + 255; ++i) { - clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i; + clip1[255 + i] = clip_8b(i); } tables_ok = 1; } } -static WEBP_INLINE uint8_t clip_8b(int v) { - return (!(v & ~0xff)) ? v : v < 0 ? 0 : 255; -} //------------------------------------------------------------------------------ // Transforms (Paragraph 14.4) @@ -154,84 +152,63 @@ static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) { int i; int tmp[16]; for (i = 0; i < 4; ++i, src += BPS, ref += BPS) { - const int d0 = src[0] - ref[0]; + const int d0 = src[0] - ref[0]; // 9bit dynamic range ([-255,255]) const int d1 = src[1] - ref[1]; const int d2 = src[2] - ref[2]; const int d3 = src[3] - ref[3]; - const int a0 = (d0 + d3) << 3; - const int a1 = (d1 + d2) << 3; - const int a2 = (d1 - d2) << 3; - const int a3 = (d0 - d3) << 3; - tmp[0 + i * 4] = (a0 + a1); - tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 14500) >> 12; - tmp[2 + i * 4] = (a0 - a1); - tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 + 7500) >> 12; + const int a0 = (d0 + d3); // 10b [-510,510] + const int a1 = (d1 + d2); + const int a2 = (d1 - d2); + const int a3 = (d0 - d3); + tmp[0 + i * 4] = (a0 + a1) * 8; // 14b [-8160,8160] + tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9; // [-7536,7542] + tmp[2 + i * 4] = (a0 - a1) * 8; + tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 + 937) >> 9; } for (i = 0; i < 4; ++i) { - const int a0 = (tmp[0 + i] + tmp[12 + i]); + const int a0 = (tmp[0 + i] + tmp[12 + i]); // 15b const int a1 = (tmp[4 + i] + tmp[ 8 + i]); const int a2 = (tmp[4 + i] - tmp[ 8 + i]); const int a3 = (tmp[0 + i] - tmp[12 + i]); - out[0 + i] = (a0 + a1 + 7) >> 4; + out[0 + i] = (a0 + a1 + 7) >> 4; // 12b out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0); out[8 + i] = (a0 - a1 + 7) >> 4; out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16); } } -static void ITransformWHT(const int16_t* in, int16_t* out) { - int tmp[16]; - int i; - for (i = 0; i < 4; ++i) { - const int a0 = in[0 + i] + in[12 + i]; - const int a1 = in[4 + i] + in[ 8 + i]; - const int a2 = in[4 + i] - in[ 8 + i]; - const int a3 = in[0 + i] - in[12 + i]; - tmp[0 + i] = a0 + a1; - tmp[8 + i] = a0 - a1; - tmp[4 + i] = a3 + a2; - tmp[12 + i] = a3 - a2; - } - for (i = 0; i < 4; ++i) { - const int dc = tmp[0 + i * 4] + 3; // w/ rounder - const int a0 = dc + tmp[3 + i * 4]; - const int a1 = tmp[1 + i * 4] + tmp[2 + i * 4]; - const int a2 = tmp[1 + i * 4] - tmp[2 + i * 4]; - const int a3 = dc - tmp[3 + i * 4]; - out[ 0] = (a0 + a1) >> 3; - out[16] = (a3 + a2) >> 3; - out[32] = (a0 - a1) >> 3; - out[48] = (a3 - a2) >> 3; - out += 64; - } +static void FTransform2(const uint8_t* src, const uint8_t* ref, int16_t* out) { + VP8FTransform(src, ref, out); + VP8FTransform(src + 4, ref + 4, out + 16); } static void FTransformWHT(const int16_t* in, int16_t* out) { - int tmp[16]; + // input is 12b signed + int32_t tmp[16]; int i; for (i = 0; i < 4; ++i, in += 64) { - const int a0 = (in[0 * 16] + in[2 * 16]) << 2; - const int a1 = (in[1 * 16] + in[3 * 16]) << 2; - const int a2 = (in[1 * 16] - in[3 * 16]) << 2; - const int a3 = (in[0 * 16] - in[2 * 16]) << 2; - tmp[0 + i * 4] = (a0 + a1) + (a0 != 0); + const int a0 = (in[0 * 16] + in[2 * 16]); // 13b + const int a1 = (in[1 * 16] + in[3 * 16]); + const int a2 = (in[1 * 16] - in[3 * 16]); + const int a3 = (in[0 * 16] - in[2 * 16]); + tmp[0 + i * 4] = a0 + a1; // 14b tmp[1 + i * 4] = a3 + a2; tmp[2 + i * 4] = a3 - a2; tmp[3 + i * 4] = a0 - a1; } for (i = 0; i < 4; ++i) { - const int a0 = (tmp[0 + i] + tmp[8 + i]); + const int a0 = (tmp[0 + i] + tmp[8 + i]); // 15b const int a1 = (tmp[4 + i] + tmp[12+ i]); const int a2 = (tmp[4 + i] - tmp[12+ i]); const int a3 = (tmp[0 + i] - tmp[8 + i]); - const int b0 = a0 + a1; + const int b0 = a0 + a1; // 16b const int b1 = a3 + a2; const int b2 = a3 - a2; const int b3 = a0 - a1; - out[ 0 + i] = (b0 + (b0 > 0) + 3) >> 3; - out[ 4 + i] = (b1 + (b1 > 0) + 3) >> 3; - out[ 8 + i] = (b2 + (b2 > 0) + 3) >> 3; - out[12 + i] = (b3 + (b3 > 0) + 3) >> 3; + out[ 0 + i] = b0 >> 1; // 15b + out[ 4 + i] = b1 >> 1; + out[ 8 + i] = b2 >> 1; + out[12 + i] = b3 >> 1; } } @@ -241,8 +218,6 @@ static void FTransformWHT(const int16_t* in, int16_t* out) { //------------------------------------------------------------------------------ // Intra predictions -#define DST(x, y) dst[(x) + (y) * BPS] - static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) { int j; for (j = 0; j < size; ++j) { @@ -253,7 +228,7 @@ static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) { static WEBP_INLINE void VerticalPred(uint8_t* dst, const uint8_t* top, int size) { int j; - if (top) { + if (top != NULL) { for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size); } else { Fill(dst, 127, size); @@ -262,7 +237,7 @@ static WEBP_INLINE void VerticalPred(uint8_t* dst, static WEBP_INLINE void HorizontalPred(uint8_t* dst, const uint8_t* left, int size) { - if (left) { + if (left != NULL) { int j; for (j = 0; j < size; ++j) { memset(dst + j * BPS, left[j], size); @@ -275,8 +250,8 @@ static WEBP_INLINE void HorizontalPred(uint8_t* dst, static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left, const uint8_t* top, int size) { int y; - if (left) { - if (top) { + if (left != NULL) { + if (top != NULL) { const uint8_t* const clip = clip1 + 255 - left[-1]; for (y = 0; y < size; ++y) { const uint8_t* const clip_table = clip + left[y]; @@ -294,7 +269,7 @@ static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left, // is equivalent to VE prediction where you just copy the top samples. // Note that if top samples are not available, the default value is // then 129, and not 127 as in the VerticalPred case. - if (top) { + if (top != NULL) { VerticalPred(dst, top, size); } else { Fill(dst, 129, size); @@ -307,15 +282,15 @@ static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left, int size, int round, int shift) { int DC = 0; int j; - if (top) { + if (top != NULL) { for (j = 0; j < size; ++j) DC += top[j]; - if (left) { // top and left present + if (left != NULL) { // top and left present for (j = 0; j < size; ++j) DC += left[j]; } else { // top, but no left DC += DC; } DC = (DC + round) >> shift; - } else if (left) { // left but no top + } else if (left != NULL) { // left but no top for (j = 0; j < size; ++j) DC += left[j]; DC += DC; DC = (DC + round) >> shift; @@ -337,8 +312,8 @@ static void IntraChromaPreds(uint8_t* dst, const uint8_t* left, TrueMotion(C8TM8 + dst, left, top, 8); // V block dst += 8; - if (top) top += 8; - if (left) left += 16; + if (top != NULL) top += 8; + if (left != NULL) left += 16; DCMode(C8DC8 + dst, left, top, 8, 8, 4); VerticalPred(C8VE8 + dst, top, 8); HorizontalPred(C8HE8 + dst, left, 8); @@ -359,6 +334,7 @@ static void Intra16Preds(uint8_t* dst, //------------------------------------------------------------------------------ // luma 4x4 prediction +#define DST(x, y) dst[(x) + (y) * BPS] #define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2) #define AVG2(a, b) (((a) + (b) + 1) >> 1) @@ -589,30 +565,30 @@ static int TTransform(const uint8_t* in, const uint16_t* w) { int i; // horizontal pass for (i = 0; i < 4; ++i, in += BPS) { - const int a0 = (in[0] + in[2]) << 2; - const int a1 = (in[1] + in[3]) << 2; - const int a2 = (in[1] - in[3]) << 2; - const int a3 = (in[0] - in[2]) << 2; - tmp[0 + i * 4] = a0 + a1 + (a0 != 0); + const int a0 = in[0] + in[2]; + const int a1 = in[1] + in[3]; + const int a2 = in[1] - in[3]; + const int a3 = in[0] - in[2]; + tmp[0 + i * 4] = a0 + a1; tmp[1 + i * 4] = a3 + a2; tmp[2 + i * 4] = a3 - a2; tmp[3 + i * 4] = a0 - a1; } // vertical pass for (i = 0; i < 4; ++i, ++w) { - const int a0 = (tmp[0 + i] + tmp[8 + i]); - const int a1 = (tmp[4 + i] + tmp[12+ i]); - const int a2 = (tmp[4 + i] - tmp[12+ i]); - const int a3 = (tmp[0 + i] - tmp[8 + i]); + const int a0 = tmp[0 + i] + tmp[8 + i]; + const int a1 = tmp[4 + i] + tmp[12+ i]; + const int a2 = tmp[4 + i] - tmp[12+ i]; + const int a3 = tmp[0 + i] - tmp[8 + i]; const int b0 = a0 + a1; const int b1 = a3 + a2; const int b2 = a3 - a2; const int b3 = a0 - a1; - // abs((b + (b<0) + 3) >> 3) = (abs(b) + 3) >> 3 - sum += w[ 0] * ((abs(b0) + 3) >> 3); - sum += w[ 4] * ((abs(b1) + 3) >> 3); - sum += w[ 8] * ((abs(b2) + 3) >> 3); - sum += w[12] * ((abs(b3) + 3) >> 3); + + sum += w[ 0] * abs(b0); + sum += w[ 4] * abs(b1); + sum += w[ 8] * abs(b2); + sum += w[12] * abs(b3); } return sum; } @@ -621,7 +597,7 @@ static int Disto4x4(const uint8_t* const a, const uint8_t* const b, const uint16_t* const w) { const int sum1 = TTransform(a, w); const int sum2 = TTransform(b, w); - return (abs(sum2 - sum1) + 8) >> 4; + return abs(sum2 - sum1) >> 5; } static int Disto16x16(const uint8_t* const a, const uint8_t* const b, @@ -646,21 +622,57 @@ static const uint8_t kZigzag[16] = { // Simple quantization static int QuantizeBlock(int16_t in[16], int16_t out[16], - int n, const VP8Matrix* const mtx) { + const VP8Matrix* const mtx) { int last = -1; - for (; n < 16; ++n) { + int n; + for (n = 0; n < 16; ++n) { + const int j = kZigzag[n]; + const int sign = (in[j] < 0); + const uint32_t coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j]; + if (coeff > mtx->zthresh_[j]) { + const uint32_t Q = mtx->q_[j]; + const uint32_t iQ = mtx->iq_[j]; + const uint32_t B = mtx->bias_[j]; + int level = QUANTDIV(coeff, iQ, B); + if (level > MAX_LEVEL) level = MAX_LEVEL; + if (sign) level = -level; + in[j] = level * Q; + out[n] = level; + if (level) last = n; + } else { + out[n] = 0; + in[j] = 0; + } + } + return (last >= 0); +} + +static int Quantize2Blocks(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} + +static int QuantizeBlockWHT(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + int n, last = -1; + for (n = 0; n < 16; ++n) { const int j = kZigzag[n]; const int sign = (in[j] < 0); - int coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j]; - if (coeff > 2047) coeff = 2047; + const uint32_t coeff = sign ? -in[j] : in[j]; + assert(mtx->sharpen_[j] == 0); if (coeff > mtx->zthresh_[j]) { - const int Q = mtx->q_[j]; - const int iQ = mtx->iq_[j]; - const int B = mtx->bias_[j]; - out[n] = QUANTDIV(coeff, iQ, B); - if (sign) out[n] = -out[n]; - in[j] = out[n] * Q; - if (out[n]) last = n; + const uint32_t Q = mtx->q_[j]; + const uint32_t iQ = mtx->iq_[j]; + const uint32_t B = mtx->bias_[j]; + int level = QUANTDIV(coeff, iQ, B); + if (level > MAX_LEVEL) level = MAX_LEVEL; + if (sign) level = -level; + in[j] = level * Q; + out[n] = level; + if (level) last = n; } else { out[n] = 0; in[j] = 0; @@ -672,16 +684,22 @@ static int QuantizeBlock(int16_t in[16], int16_t out[16], //------------------------------------------------------------------------------ // Block copy -static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int size) { +static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int w, int h) { int y; - for (y = 0; y < size; ++y) { - memcpy(dst, src, size); + for (y = 0; y < h; ++y) { + memcpy(dst, src, w); src += BPS; dst += BPS; } } -static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); } +static void Copy4x4(const uint8_t* src, uint8_t* dst) { + Copy(src, dst, 4, 4); +} + +static void Copy16x8(const uint8_t* src, uint8_t* dst) { + Copy(src, dst, 16, 8); +} //------------------------------------------------------------------------------ // Initialization @@ -691,7 +709,7 @@ static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); } VP8CHisto VP8CollectHistogram; VP8Idct VP8ITransform; VP8Fdct VP8FTransform; -VP8WHT VP8ITransformWHT; +VP8Fdct VP8FTransform2; VP8WHT VP8FTransformWHT; VP8Intra4Preds VP8EncPredLuma4; VP8IntraPreds VP8EncPredLuma16; @@ -703,18 +721,32 @@ VP8Metric VP8SSE4x4; VP8WMetric VP8TDisto4x4; VP8WMetric VP8TDisto16x16; VP8QuantizeBlock VP8EncQuantizeBlock; +VP8Quantize2Blocks VP8EncQuantize2Blocks; +VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT; VP8BlockCopy VP8Copy4x4; +VP8BlockCopy VP8Copy16x8; extern void VP8EncDspInitSSE2(void); +extern void VP8EncDspInitSSE41(void); +extern void VP8EncDspInitAVX2(void); +extern void VP8EncDspInitNEON(void); +extern void VP8EncDspInitMIPS32(void); +extern void VP8EncDspInitMIPSdspR2(void); + +static volatile VP8CPUInfo enc_last_cpuinfo_used = + (VP8CPUInfo)&enc_last_cpuinfo_used; -void VP8EncDspInit(void) { +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInit(void) { + if (enc_last_cpuinfo_used == VP8GetCPUInfo) return; + + VP8DspInit(); // common inverse transforms InitTables(); // default C implementations VP8CollectHistogram = CollectHistogram; VP8ITransform = ITransform; VP8FTransform = FTransform; - VP8ITransformWHT = ITransformWHT; + VP8FTransform2 = FTransform2; VP8FTransformWHT = FTransformWHT; VP8EncPredLuma4 = Intra4Preds; VP8EncPredLuma16 = Intra16Preds; @@ -726,18 +758,43 @@ void VP8EncDspInit(void) { VP8TDisto4x4 = Disto4x4; VP8TDisto16x16 = Disto16x16; VP8EncQuantizeBlock = QuantizeBlock; + VP8EncQuantize2Blocks = Quantize2Blocks; + VP8EncQuantizeBlockWHT = QuantizeBlockWHT; VP8Copy4x4 = Copy4x4; + VP8Copy16x8 = Copy16x8; // If defined, use CPUInfo() to overwrite some pointers with faster versions. - if (VP8GetCPUInfo) { + if (VP8GetCPUInfo != NULL) { #if defined(WEBP_USE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8EncDspInitSSE2(); +#if defined(WEBP_USE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + VP8EncDspInitSSE41(); + } +#endif + } +#endif +#if defined(WEBP_USE_AVX2) + if (VP8GetCPUInfo(kAVX2)) { + VP8EncDspInitAVX2(); + } +#endif +#if defined(WEBP_USE_NEON) + if (VP8GetCPUInfo(kNEON)) { + VP8EncDspInitNEON(); + } +#endif +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + VP8EncDspInitMIPS32(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8EncDspInitMIPSdspR2(); } #endif } + enc_last_cpuinfo_used = VP8GetCPUInfo; } - -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" -#endif diff --git a/drivers/webp/dsp/enc_avx2.c b/drivers/webp/dsp/enc_avx2.c new file mode 100644 index 0000000000..93efb30b10 --- /dev/null +++ b/drivers/webp/dsp/enc_avx2.c @@ -0,0 +1,21 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// AVX2 version of speed-critical encoding functions. + +#include "./dsp.h" + +#if defined(WEBP_USE_AVX2) + +#endif // WEBP_USE_AVX2 + +//------------------------------------------------------------------------------ +// Entry point + +WEBP_DSP_INIT_STUB(VP8EncDspInitAVX2) diff --git a/drivers/webp/dsp/enc_mips32.c b/drivers/webp/dsp/enc_mips32.c new file mode 100644 index 0000000000..fd10143de9 --- /dev/null +++ b/drivers/webp/dsp/enc_mips32.c @@ -0,0 +1,672 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of speed-critical encoding functions. +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) +// Slobodan Prijic (slobodan.prijic@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include "./mips_macro.h" +#include "../enc/vp8enci.h" +#include "../enc/cost.h" + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; + +// macro for one vertical pass in ITransformOne +// MUL macro inlined +// temp0..temp15 holds tmp[0]..tmp[15] +// A..D - offsets in bytes to load from in buffer +// TEMP0..TEMP3 - registers for corresponding tmp elements +// TEMP4..TEMP5 - temporary registers +#define VERTICAL_PASS(A, B, C, D, TEMP4, TEMP0, TEMP1, TEMP2, TEMP3) \ + "lh %[temp16], " #A "(%[temp20]) \n\t" \ + "lh %[temp18], " #B "(%[temp20]) \n\t" \ + "lh %[temp17], " #C "(%[temp20]) \n\t" \ + "lh %[temp19], " #D "(%[temp20]) \n\t" \ + "addu %[" #TEMP4 "], %[temp16], %[temp18] \n\t" \ + "subu %[temp16], %[temp16], %[temp18] \n\t" \ + "mul %[" #TEMP0 "], %[temp17], %[kC2] \n\t" \ + "mul %[temp18], %[temp19], %[kC1] \n\t" \ + "mul %[temp17], %[temp17], %[kC1] \n\t" \ + "mul %[temp19], %[temp19], %[kC2] \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 16 \n\n" \ + "sra %[temp18], %[temp18], 16 \n\n" \ + "sra %[temp17], %[temp17], 16 \n\n" \ + "sra %[temp19], %[temp19], 16 \n\n" \ + "subu %[" #TEMP2 "], %[" #TEMP0 "], %[temp18] \n\t" \ + "addu %[" #TEMP3 "], %[temp17], %[temp19] \n\t" \ + "addu %[" #TEMP0 "], %[" #TEMP4 "], %[" #TEMP3 "] \n\t" \ + "addu %[" #TEMP1 "], %[temp16], %[" #TEMP2 "] \n\t" \ + "subu %[" #TEMP2 "], %[temp16], %[" #TEMP2 "] \n\t" \ + "subu %[" #TEMP3 "], %[" #TEMP4 "], %[" #TEMP3 "] \n\t" + +// macro for one horizontal pass in ITransformOne +// MUL and STORE macros inlined +// a = clip_8b(a) is replaced with: a = max(a, 0); a = min(a, 255) +// temp0..temp15 holds tmp[0]..tmp[15] +// A - offset in bytes to load from ref and store to dst buffer +// TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements +#define HORIZONTAL_PASS(A, TEMP0, TEMP4, TEMP8, TEMP12) \ + "addiu %[" #TEMP0 "], %[" #TEMP0 "], 4 \n\t" \ + "addu %[temp16], %[" #TEMP0 "], %[" #TEMP8 "] \n\t" \ + "subu %[temp17], %[" #TEMP0 "], %[" #TEMP8 "] \n\t" \ + "mul %[" #TEMP0 "], %[" #TEMP4 "], %[kC2] \n\t" \ + "mul %[" #TEMP8 "], %[" #TEMP12 "], %[kC1] \n\t" \ + "mul %[" #TEMP4 "], %[" #TEMP4 "], %[kC1] \n\t" \ + "mul %[" #TEMP12 "], %[" #TEMP12 "], %[kC2] \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 16 \n\t" \ + "sra %[" #TEMP8 "], %[" #TEMP8 "], 16 \n\t" \ + "sra %[" #TEMP4 "], %[" #TEMP4 "], 16 \n\t" \ + "sra %[" #TEMP12 "], %[" #TEMP12 "], 16 \n\t" \ + "subu %[temp18], %[" #TEMP0 "], %[" #TEMP8 "] \n\t" \ + "addu %[temp19], %[" #TEMP4 "], %[" #TEMP12 "] \n\t" \ + "addu %[" #TEMP0 "], %[temp16], %[temp19] \n\t" \ + "addu %[" #TEMP4 "], %[temp17], %[temp18] \n\t" \ + "subu %[" #TEMP8 "], %[temp17], %[temp18] \n\t" \ + "subu %[" #TEMP12 "], %[temp16], %[temp19] \n\t" \ + "lw %[temp20], 0(%[args]) \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 3 \n\t" \ + "sra %[" #TEMP4 "], %[" #TEMP4 "], 3 \n\t" \ + "sra %[" #TEMP8 "], %[" #TEMP8 "], 3 \n\t" \ + "sra %[" #TEMP12 "], %[" #TEMP12 "], 3 \n\t" \ + "lbu %[temp16], 0+" XSTR(BPS) "*" #A "(%[temp20]) \n\t" \ + "lbu %[temp17], 1+" XSTR(BPS) "*" #A "(%[temp20]) \n\t" \ + "lbu %[temp18], 2+" XSTR(BPS) "*" #A "(%[temp20]) \n\t" \ + "lbu %[temp19], 3+" XSTR(BPS) "*" #A "(%[temp20]) \n\t" \ + "addu %[" #TEMP0 "], %[temp16], %[" #TEMP0 "] \n\t" \ + "addu %[" #TEMP4 "], %[temp17], %[" #TEMP4 "] \n\t" \ + "addu %[" #TEMP8 "], %[temp18], %[" #TEMP8 "] \n\t" \ + "addu %[" #TEMP12 "], %[temp19], %[" #TEMP12 "] \n\t" \ + "slt %[temp16], %[" #TEMP0 "], $zero \n\t" \ + "slt %[temp17], %[" #TEMP4 "], $zero \n\t" \ + "slt %[temp18], %[" #TEMP8 "], $zero \n\t" \ + "slt %[temp19], %[" #TEMP12 "], $zero \n\t" \ + "movn %[" #TEMP0 "], $zero, %[temp16] \n\t" \ + "movn %[" #TEMP4 "], $zero, %[temp17] \n\t" \ + "movn %[" #TEMP8 "], $zero, %[temp18] \n\t" \ + "movn %[" #TEMP12 "], $zero, %[temp19] \n\t" \ + "addiu %[temp20], $zero, 255 \n\t" \ + "slt %[temp16], %[" #TEMP0 "], %[temp20] \n\t" \ + "slt %[temp17], %[" #TEMP4 "], %[temp20] \n\t" \ + "slt %[temp18], %[" #TEMP8 "], %[temp20] \n\t" \ + "slt %[temp19], %[" #TEMP12 "], %[temp20] \n\t" \ + "movz %[" #TEMP0 "], %[temp20], %[temp16] \n\t" \ + "movz %[" #TEMP4 "], %[temp20], %[temp17] \n\t" \ + "lw %[temp16], 8(%[args]) \n\t" \ + "movz %[" #TEMP8 "], %[temp20], %[temp18] \n\t" \ + "movz %[" #TEMP12 "], %[temp20], %[temp19] \n\t" \ + "sb %[" #TEMP0 "], 0+" XSTR(BPS) "*" #A "(%[temp16]) \n\t" \ + "sb %[" #TEMP4 "], 1+" XSTR(BPS) "*" #A "(%[temp16]) \n\t" \ + "sb %[" #TEMP8 "], 2+" XSTR(BPS) "*" #A "(%[temp16]) \n\t" \ + "sb %[" #TEMP12 "], 3+" XSTR(BPS) "*" #A "(%[temp16]) \n\t" + +// Does one or two inverse transforms. +static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in, + uint8_t* dst) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + int temp7, temp8, temp9, temp10, temp11, temp12, temp13; + int temp14, temp15, temp16, temp17, temp18, temp19, temp20; + const int* args[3] = {(const int*)ref, (const int*)in, (const int*)dst}; + + __asm__ volatile( + "lw %[temp20], 4(%[args]) \n\t" + VERTICAL_PASS(0, 16, 8, 24, temp4, temp0, temp1, temp2, temp3) + VERTICAL_PASS(2, 18, 10, 26, temp8, temp4, temp5, temp6, temp7) + VERTICAL_PASS(4, 20, 12, 28, temp12, temp8, temp9, temp10, temp11) + VERTICAL_PASS(6, 22, 14, 30, temp20, temp12, temp13, temp14, temp15) + + HORIZONTAL_PASS(0, temp0, temp4, temp8, temp12) + HORIZONTAL_PASS(1, temp1, temp5, temp9, temp13) + HORIZONTAL_PASS(2, temp2, temp6, temp10, temp14) + HORIZONTAL_PASS(3, temp3, temp7, temp11, temp15) + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11), + [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14), + [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17), + [temp18]"=&r"(temp18), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20) + : [args]"r"(args), [kC1]"r"(kC1), [kC2]"r"(kC2) + : "memory", "hi", "lo" + ); +} + +static void ITransform(const uint8_t* ref, const int16_t* in, + uint8_t* dst, int do_two) { + ITransformOne(ref, in, dst); + if (do_two) { + ITransformOne(ref + 4, in + 16, dst + 4); + } +} + +#undef VERTICAL_PASS +#undef HORIZONTAL_PASS + +// macro for one pass through for loop in QuantizeBlock +// QUANTDIV macro inlined +// J - offset in bytes (kZigzag[n] * 2) +// K - offset in bytes (kZigzag[n] * 4) +// N - offset in bytes (n * 2) +#define QUANTIZE_ONE(J, K, N) \ + "lh %[temp0], " #J "(%[ppin]) \n\t" \ + "lhu %[temp1], " #J "(%[ppsharpen]) \n\t" \ + "lw %[temp2], " #K "(%[ppzthresh]) \n\t" \ + "sra %[sign], %[temp0], 15 \n\t" \ + "xor %[coeff], %[temp0], %[sign] \n\t" \ + "subu %[coeff], %[coeff], %[sign] \n\t" \ + "addu %[coeff], %[coeff], %[temp1] \n\t" \ + "slt %[temp4], %[temp2], %[coeff] \n\t" \ + "addiu %[temp5], $zero, 0 \n\t" \ + "addiu %[level], $zero, 0 \n\t" \ + "beqz %[temp4], 2f \n\t" \ + "lhu %[temp1], " #J "(%[ppiq]) \n\t" \ + "lw %[temp2], " #K "(%[ppbias]) \n\t" \ + "lhu %[temp3], " #J "(%[ppq]) \n\t" \ + "mul %[level], %[coeff], %[temp1] \n\t" \ + "addu %[level], %[level], %[temp2] \n\t" \ + "sra %[level], %[level], 17 \n\t" \ + "slt %[temp4], %[max_level], %[level] \n\t" \ + "movn %[level], %[max_level], %[temp4] \n\t" \ + "xor %[level], %[level], %[sign] \n\t" \ + "subu %[level], %[level], %[sign] \n\t" \ + "mul %[temp5], %[level], %[temp3] \n\t" \ +"2: \n\t" \ + "sh %[temp5], " #J "(%[ppin]) \n\t" \ + "sh %[level], " #N "(%[pout]) \n\t" + +static int QuantizeBlock(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + int temp0, temp1, temp2, temp3, temp4, temp5; + int sign, coeff, level, i; + int max_level = MAX_LEVEL; + + int16_t* ppin = &in[0]; + int16_t* pout = &out[0]; + const uint16_t* ppsharpen = &mtx->sharpen_[0]; + const uint32_t* ppzthresh = &mtx->zthresh_[0]; + const uint16_t* ppq = &mtx->q_[0]; + const uint16_t* ppiq = &mtx->iq_[0]; + const uint32_t* ppbias = &mtx->bias_[0]; + + __asm__ volatile( + QUANTIZE_ONE( 0, 0, 0) + QUANTIZE_ONE( 2, 4, 2) + QUANTIZE_ONE( 8, 16, 4) + QUANTIZE_ONE(16, 32, 6) + QUANTIZE_ONE(10, 20, 8) + QUANTIZE_ONE( 4, 8, 10) + QUANTIZE_ONE( 6, 12, 12) + QUANTIZE_ONE(12, 24, 14) + QUANTIZE_ONE(18, 36, 16) + QUANTIZE_ONE(24, 48, 18) + QUANTIZE_ONE(26, 52, 20) + QUANTIZE_ONE(20, 40, 22) + QUANTIZE_ONE(14, 28, 24) + QUANTIZE_ONE(22, 44, 26) + QUANTIZE_ONE(28, 56, 28) + QUANTIZE_ONE(30, 60, 30) + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [sign]"=&r"(sign), [coeff]"=&r"(coeff), + [level]"=&r"(level) + : [pout]"r"(pout), [ppin]"r"(ppin), + [ppiq]"r"(ppiq), [max_level]"r"(max_level), + [ppbias]"r"(ppbias), [ppzthresh]"r"(ppzthresh), + [ppsharpen]"r"(ppsharpen), [ppq]"r"(ppq) + : "memory", "hi", "lo" + ); + + // moved out from macro to increase possibility for earlier breaking + for (i = 15; i >= 0; i--) { + if (out[i]) return 1; + } + return 0; +} + +static int Quantize2Blocks(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = QuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= QuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} + +#undef QUANTIZE_ONE + +// macro for one horizontal pass in Disto4x4 (TTransform) +// two calls of function TTransform are merged into single one +// A - offset in bytes to load from a and b buffers +// E..H - offsets in bytes to store first results to tmp buffer +// E1..H1 - offsets in bytes to store second results to tmp buffer +#define HORIZONTAL_PASS(A, E, F, G, H, E1, F1, G1, H1) \ + "lbu %[temp0], 0+" XSTR(BPS) "*" #A "(%[a]) \n\t" \ + "lbu %[temp1], 1+" XSTR(BPS) "*" #A "(%[a]) \n\t" \ + "lbu %[temp2], 2+" XSTR(BPS) "*" #A "(%[a]) \n\t" \ + "lbu %[temp3], 3+" XSTR(BPS) "*" #A "(%[a]) \n\t" \ + "lbu %[temp4], 0+" XSTR(BPS) "*" #A "(%[b]) \n\t" \ + "lbu %[temp5], 1+" XSTR(BPS) "*" #A "(%[b]) \n\t" \ + "lbu %[temp6], 2+" XSTR(BPS) "*" #A "(%[b]) \n\t" \ + "lbu %[temp7], 3+" XSTR(BPS) "*" #A "(%[b]) \n\t" \ + "addu %[temp8], %[temp0], %[temp2] \n\t" \ + "subu %[temp0], %[temp0], %[temp2] \n\t" \ + "addu %[temp2], %[temp1], %[temp3] \n\t" \ + "subu %[temp1], %[temp1], %[temp3] \n\t" \ + "addu %[temp3], %[temp4], %[temp6] \n\t" \ + "subu %[temp4], %[temp4], %[temp6] \n\t" \ + "addu %[temp6], %[temp5], %[temp7] \n\t" \ + "subu %[temp5], %[temp5], %[temp7] \n\t" \ + "addu %[temp7], %[temp8], %[temp2] \n\t" \ + "subu %[temp2], %[temp8], %[temp2] \n\t" \ + "addu %[temp8], %[temp0], %[temp1] \n\t" \ + "subu %[temp0], %[temp0], %[temp1] \n\t" \ + "addu %[temp1], %[temp3], %[temp6] \n\t" \ + "subu %[temp3], %[temp3], %[temp6] \n\t" \ + "addu %[temp6], %[temp4], %[temp5] \n\t" \ + "subu %[temp4], %[temp4], %[temp5] \n\t" \ + "sw %[temp7], " #E "(%[tmp]) \n\t" \ + "sw %[temp2], " #H "(%[tmp]) \n\t" \ + "sw %[temp8], " #F "(%[tmp]) \n\t" \ + "sw %[temp0], " #G "(%[tmp]) \n\t" \ + "sw %[temp1], " #E1 "(%[tmp]) \n\t" \ + "sw %[temp3], " #H1 "(%[tmp]) \n\t" \ + "sw %[temp6], " #F1 "(%[tmp]) \n\t" \ + "sw %[temp4], " #G1 "(%[tmp]) \n\t" + +// macro for one vertical pass in Disto4x4 (TTransform) +// two calls of function TTransform are merged into single one +// since only one accu is available in mips32r1 instruction set +// first is done second call of function TTransform and after +// that first one. +// const int sum1 = TTransform(a, w); +// const int sum2 = TTransform(b, w); +// return abs(sum2 - sum1) >> 5; +// (sum2 - sum1) is calculated with madds (sub2) and msubs (sub1) +// A..D - offsets in bytes to load first results from tmp buffer +// A1..D1 - offsets in bytes to load second results from tmp buffer +// E..H - offsets in bytes to load from w buffer +#define VERTICAL_PASS(A, B, C, D, A1, B1, C1, D1, E, F, G, H) \ + "lw %[temp0], " #A1 "(%[tmp]) \n\t" \ + "lw %[temp1], " #C1 "(%[tmp]) \n\t" \ + "lw %[temp2], " #B1 "(%[tmp]) \n\t" \ + "lw %[temp3], " #D1 "(%[tmp]) \n\t" \ + "addu %[temp8], %[temp0], %[temp1] \n\t" \ + "subu %[temp0], %[temp0], %[temp1] \n\t" \ + "addu %[temp1], %[temp2], %[temp3] \n\t" \ + "subu %[temp2], %[temp2], %[temp3] \n\t" \ + "addu %[temp3], %[temp8], %[temp1] \n\t" \ + "subu %[temp8], %[temp8], %[temp1] \n\t" \ + "addu %[temp1], %[temp0], %[temp2] \n\t" \ + "subu %[temp0], %[temp0], %[temp2] \n\t" \ + "sra %[temp4], %[temp3], 31 \n\t" \ + "sra %[temp5], %[temp1], 31 \n\t" \ + "sra %[temp6], %[temp0], 31 \n\t" \ + "sra %[temp7], %[temp8], 31 \n\t" \ + "xor %[temp3], %[temp3], %[temp4] \n\t" \ + "xor %[temp1], %[temp1], %[temp5] \n\t" \ + "xor %[temp0], %[temp0], %[temp6] \n\t" \ + "xor %[temp8], %[temp8], %[temp7] \n\t" \ + "subu %[temp3], %[temp3], %[temp4] \n\t" \ + "subu %[temp1], %[temp1], %[temp5] \n\t" \ + "subu %[temp0], %[temp0], %[temp6] \n\t" \ + "subu %[temp8], %[temp8], %[temp7] \n\t" \ + "lhu %[temp4], " #E "(%[w]) \n\t" \ + "lhu %[temp5], " #F "(%[w]) \n\t" \ + "lhu %[temp6], " #G "(%[w]) \n\t" \ + "lhu %[temp7], " #H "(%[w]) \n\t" \ + "madd %[temp4], %[temp3] \n\t" \ + "madd %[temp5], %[temp1] \n\t" \ + "madd %[temp6], %[temp0] \n\t" \ + "madd %[temp7], %[temp8] \n\t" \ + "lw %[temp0], " #A "(%[tmp]) \n\t" \ + "lw %[temp1], " #C "(%[tmp]) \n\t" \ + "lw %[temp2], " #B "(%[tmp]) \n\t" \ + "lw %[temp3], " #D "(%[tmp]) \n\t" \ + "addu %[temp8], %[temp0], %[temp1] \n\t" \ + "subu %[temp0], %[temp0], %[temp1] \n\t" \ + "addu %[temp1], %[temp2], %[temp3] \n\t" \ + "subu %[temp2], %[temp2], %[temp3] \n\t" \ + "addu %[temp3], %[temp8], %[temp1] \n\t" \ + "subu %[temp1], %[temp8], %[temp1] \n\t" \ + "addu %[temp8], %[temp0], %[temp2] \n\t" \ + "subu %[temp0], %[temp0], %[temp2] \n\t" \ + "sra %[temp2], %[temp3], 31 \n\t" \ + "xor %[temp3], %[temp3], %[temp2] \n\t" \ + "subu %[temp3], %[temp3], %[temp2] \n\t" \ + "msub %[temp4], %[temp3] \n\t" \ + "sra %[temp2], %[temp8], 31 \n\t" \ + "sra %[temp3], %[temp0], 31 \n\t" \ + "sra %[temp4], %[temp1], 31 \n\t" \ + "xor %[temp8], %[temp8], %[temp2] \n\t" \ + "xor %[temp0], %[temp0], %[temp3] \n\t" \ + "xor %[temp1], %[temp1], %[temp4] \n\t" \ + "subu %[temp8], %[temp8], %[temp2] \n\t" \ + "subu %[temp0], %[temp0], %[temp3] \n\t" \ + "subu %[temp1], %[temp1], %[temp4] \n\t" \ + "msub %[temp5], %[temp8] \n\t" \ + "msub %[temp6], %[temp0] \n\t" \ + "msub %[temp7], %[temp1] \n\t" + +static int Disto4x4(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int tmp[32]; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + + __asm__ volatile( + HORIZONTAL_PASS(0, 0, 4, 8, 12, 64, 68, 72, 76) + HORIZONTAL_PASS(1, 16, 20, 24, 28, 80, 84, 88, 92) + HORIZONTAL_PASS(2, 32, 36, 40, 44, 96, 100, 104, 108) + HORIZONTAL_PASS(3, 48, 52, 56, 60, 112, 116, 120, 124) + "mthi $zero \n\t" + "mtlo $zero \n\t" + VERTICAL_PASS( 0, 16, 32, 48, 64, 80, 96, 112, 0, 8, 16, 24) + VERTICAL_PASS( 4, 20, 36, 52, 68, 84, 100, 116, 2, 10, 18, 26) + VERTICAL_PASS( 8, 24, 40, 56, 72, 88, 104, 120, 4, 12, 20, 28) + VERTICAL_PASS(12, 28, 44, 60, 76, 92, 108, 124, 6, 14, 22, 30) + "mflo %[temp0] \n\t" + "sra %[temp1], %[temp0], 31 \n\t" + "xor %[temp0], %[temp0], %[temp1] \n\t" + "subu %[temp0], %[temp0], %[temp1] \n\t" + "sra %[temp0], %[temp0], 5 \n\t" + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8) + : [a]"r"(a), [b]"r"(b), [w]"r"(w), [tmp]"r"(tmp) + : "memory", "hi", "lo" + ); + + return temp0; +} + +#undef VERTICAL_PASS +#undef HORIZONTAL_PASS + +static int Disto16x16(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4(a + x + y, b + x + y, w); + } + } + return D; +} + +// macro for one horizontal pass in FTransform +// temp0..temp15 holds tmp[0]..tmp[15] +// A - offset in bytes to load from src and ref buffers +// TEMP0..TEMP3 - registers for corresponding tmp elements +#define HORIZONTAL_PASS(A, TEMP0, TEMP1, TEMP2, TEMP3) \ + "lw %[" #TEMP1 "], 0(%[args]) \n\t" \ + "lw %[" #TEMP2 "], 4(%[args]) \n\t" \ + "lbu %[temp16], 0+" XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "lbu %[temp17], 0+" XSTR(BPS) "*" #A "(%[" #TEMP2 "]) \n\t" \ + "lbu %[temp18], 1+" XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "lbu %[temp19], 1+" XSTR(BPS) "*" #A "(%[" #TEMP2 "]) \n\t" \ + "subu %[temp20], %[temp16], %[temp17] \n\t" \ + "lbu %[temp16], 2+" XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "lbu %[temp17], 2+" XSTR(BPS) "*" #A "(%[" #TEMP2 "]) \n\t" \ + "subu %[" #TEMP0 "], %[temp18], %[temp19] \n\t" \ + "lbu %[temp18], 3+" XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "lbu %[temp19], 3+" XSTR(BPS) "*" #A "(%[" #TEMP2 "]) \n\t" \ + "subu %[" #TEMP1 "], %[temp16], %[temp17] \n\t" \ + "subu %[" #TEMP2 "], %[temp18], %[temp19] \n\t" \ + "addu %[" #TEMP3 "], %[temp20], %[" #TEMP2 "] \n\t" \ + "subu %[" #TEMP2 "], %[temp20], %[" #TEMP2 "] \n\t" \ + "addu %[temp20], %[" #TEMP0 "], %[" #TEMP1 "] \n\t" \ + "subu %[" #TEMP0 "], %[" #TEMP0 "], %[" #TEMP1 "] \n\t" \ + "mul %[temp16], %[" #TEMP2 "], %[c5352] \n\t" \ + "mul %[temp17], %[" #TEMP2 "], %[c2217] \n\t" \ + "mul %[temp18], %[" #TEMP0 "], %[c5352] \n\t" \ + "mul %[temp19], %[" #TEMP0 "], %[c2217] \n\t" \ + "addu %[" #TEMP1 "], %[" #TEMP3 "], %[temp20] \n\t" \ + "subu %[temp20], %[" #TEMP3 "], %[temp20] \n\t" \ + "sll %[" #TEMP0 "], %[" #TEMP1 "], 3 \n\t" \ + "sll %[" #TEMP2 "], %[temp20], 3 \n\t" \ + "addiu %[temp16], %[temp16], 1812 \n\t" \ + "addiu %[temp17], %[temp17], 937 \n\t" \ + "addu %[temp16], %[temp16], %[temp19] \n\t" \ + "subu %[temp17], %[temp17], %[temp18] \n\t" \ + "sra %[" #TEMP1 "], %[temp16], 9 \n\t" \ + "sra %[" #TEMP3 "], %[temp17], 9 \n\t" + +// macro for one vertical pass in FTransform +// temp0..temp15 holds tmp[0]..tmp[15] +// A..D - offsets in bytes to store to out buffer +// TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements +#define VERTICAL_PASS(A, B, C, D, TEMP0, TEMP4, TEMP8, TEMP12) \ + "addu %[temp16], %[" #TEMP0 "], %[" #TEMP12 "] \n\t" \ + "subu %[temp19], %[" #TEMP0 "], %[" #TEMP12 "] \n\t" \ + "addu %[temp17], %[" #TEMP4 "], %[" #TEMP8 "] \n\t" \ + "subu %[temp18], %[" #TEMP4 "], %[" #TEMP8 "] \n\t" \ + "mul %[" #TEMP8 "], %[temp19], %[c2217] \n\t" \ + "mul %[" #TEMP12 "], %[temp18], %[c2217] \n\t" \ + "mul %[" #TEMP4 "], %[temp19], %[c5352] \n\t" \ + "mul %[temp18], %[temp18], %[c5352] \n\t" \ + "addiu %[temp16], %[temp16], 7 \n\t" \ + "addu %[" #TEMP0 "], %[temp16], %[temp17] \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 4 \n\t" \ + "addu %[" #TEMP12 "], %[" #TEMP12 "], %[" #TEMP4 "] \n\t" \ + "subu %[" #TEMP4 "], %[temp16], %[temp17] \n\t" \ + "sra %[" #TEMP4 "], %[" #TEMP4 "], 4 \n\t" \ + "addiu %[" #TEMP8 "], %[" #TEMP8 "], 30000 \n\t" \ + "addiu %[" #TEMP12 "], %[" #TEMP12 "], 12000 \n\t" \ + "addiu %[" #TEMP8 "], %[" #TEMP8 "], 21000 \n\t" \ + "subu %[" #TEMP8 "], %[" #TEMP8 "], %[temp18] \n\t" \ + "sra %[" #TEMP12 "], %[" #TEMP12 "], 16 \n\t" \ + "sra %[" #TEMP8 "], %[" #TEMP8 "], 16 \n\t" \ + "addiu %[temp16], %[" #TEMP12 "], 1 \n\t" \ + "movn %[" #TEMP12 "], %[temp16], %[temp19] \n\t" \ + "sh %[" #TEMP0 "], " #A "(%[temp20]) \n\t" \ + "sh %[" #TEMP4 "], " #C "(%[temp20]) \n\t" \ + "sh %[" #TEMP8 "], " #D "(%[temp20]) \n\t" \ + "sh %[" #TEMP12 "], " #B "(%[temp20]) \n\t" + +static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + int temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16; + int temp17, temp18, temp19, temp20; + const int c2217 = 2217; + const int c5352 = 5352; + const int* const args[3] = + { (const int*)src, (const int*)ref, (const int*)out }; + + __asm__ volatile( + HORIZONTAL_PASS(0, temp0, temp1, temp2, temp3) + HORIZONTAL_PASS(1, temp4, temp5, temp6, temp7) + HORIZONTAL_PASS(2, temp8, temp9, temp10, temp11) + HORIZONTAL_PASS(3, temp12, temp13, temp14, temp15) + "lw %[temp20], 8(%[args]) \n\t" + VERTICAL_PASS(0, 8, 16, 24, temp0, temp4, temp8, temp12) + VERTICAL_PASS(2, 10, 18, 26, temp1, temp5, temp9, temp13) + VERTICAL_PASS(4, 12, 20, 28, temp2, temp6, temp10, temp14) + VERTICAL_PASS(6, 14, 22, 30, temp3, temp7, temp11, temp15) + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11), + [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14), + [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17), + [temp18]"=&r"(temp18), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20) + : [args]"r"(args), [c2217]"r"(c2217), [c5352]"r"(c5352) + : "memory", "hi", "lo" + ); +} + +#undef VERTICAL_PASS +#undef HORIZONTAL_PASS + +#if !defined(WORK_AROUND_GCC) + +#define GET_SSE_INNER(A, B, C, D) \ + "lbu %[temp0], " #A "(%[a]) \n\t" \ + "lbu %[temp1], " #A "(%[b]) \n\t" \ + "lbu %[temp2], " #B "(%[a]) \n\t" \ + "lbu %[temp3], " #B "(%[b]) \n\t" \ + "lbu %[temp4], " #C "(%[a]) \n\t" \ + "lbu %[temp5], " #C "(%[b]) \n\t" \ + "lbu %[temp6], " #D "(%[a]) \n\t" \ + "lbu %[temp7], " #D "(%[b]) \n\t" \ + "subu %[temp0], %[temp0], %[temp1] \n\t" \ + "subu %[temp2], %[temp2], %[temp3] \n\t" \ + "subu %[temp4], %[temp4], %[temp5] \n\t" \ + "subu %[temp6], %[temp6], %[temp7] \n\t" \ + "madd %[temp0], %[temp0] \n\t" \ + "madd %[temp2], %[temp2] \n\t" \ + "madd %[temp4], %[temp4] \n\t" \ + "madd %[temp6], %[temp6] \n\t" + +#define GET_SSE(A, B, C, D) \ + GET_SSE_INNER(A, A + 1, A + 2, A + 3) \ + GET_SSE_INNER(B, B + 1, B + 2, B + 3) \ + GET_SSE_INNER(C, C + 1, C + 2, C + 3) \ + GET_SSE_INNER(D, D + 1, D + 2, D + 3) + +static int SSE16x16(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + + __asm__ volatile( + "mult $zero, $zero \n\t" + + GET_SSE( 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS) + GET_SSE( 1 * BPS, 4 + 1 * BPS, 8 + 1 * BPS, 12 + 1 * BPS) + GET_SSE( 2 * BPS, 4 + 2 * BPS, 8 + 2 * BPS, 12 + 2 * BPS) + GET_SSE( 3 * BPS, 4 + 3 * BPS, 8 + 3 * BPS, 12 + 3 * BPS) + GET_SSE( 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS) + GET_SSE( 5 * BPS, 4 + 5 * BPS, 8 + 5 * BPS, 12 + 5 * BPS) + GET_SSE( 6 * BPS, 4 + 6 * BPS, 8 + 6 * BPS, 12 + 6 * BPS) + GET_SSE( 7 * BPS, 4 + 7 * BPS, 8 + 7 * BPS, 12 + 7 * BPS) + GET_SSE( 8 * BPS, 4 + 8 * BPS, 8 + 8 * BPS, 12 + 8 * BPS) + GET_SSE( 9 * BPS, 4 + 9 * BPS, 8 + 9 * BPS, 12 + 9 * BPS) + GET_SSE(10 * BPS, 4 + 10 * BPS, 8 + 10 * BPS, 12 + 10 * BPS) + GET_SSE(11 * BPS, 4 + 11 * BPS, 8 + 11 * BPS, 12 + 11 * BPS) + GET_SSE(12 * BPS, 4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS) + GET_SSE(13 * BPS, 4 + 13 * BPS, 8 + 13 * BPS, 12 + 13 * BPS) + GET_SSE(14 * BPS, 4 + 14 * BPS, 8 + 14 * BPS, 12 + 14 * BPS) + GET_SSE(15 * BPS, 4 + 15 * BPS, 8 + 15 * BPS, 12 + 15 * BPS) + + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE16x8(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + + __asm__ volatile( + "mult $zero, $zero \n\t" + + GET_SSE( 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS) + GET_SSE( 1 * BPS, 4 + 1 * BPS, 8 + 1 * BPS, 12 + 1 * BPS) + GET_SSE( 2 * BPS, 4 + 2 * BPS, 8 + 2 * BPS, 12 + 2 * BPS) + GET_SSE( 3 * BPS, 4 + 3 * BPS, 8 + 3 * BPS, 12 + 3 * BPS) + GET_SSE( 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS) + GET_SSE( 5 * BPS, 4 + 5 * BPS, 8 + 5 * BPS, 12 + 5 * BPS) + GET_SSE( 6 * BPS, 4 + 6 * BPS, 8 + 6 * BPS, 12 + 6 * BPS) + GET_SSE( 7 * BPS, 4 + 7 * BPS, 8 + 7 * BPS, 12 + 7 * BPS) + + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE8x8(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + + __asm__ volatile( + "mult $zero, $zero \n\t" + + GET_SSE(0 * BPS, 4 + 0 * BPS, 1 * BPS, 4 + 1 * BPS) + GET_SSE(2 * BPS, 4 + 2 * BPS, 3 * BPS, 4 + 3 * BPS) + GET_SSE(4 * BPS, 4 + 4 * BPS, 5 * BPS, 4 + 5 * BPS) + GET_SSE(6 * BPS, 4 + 6 * BPS, 7 * BPS, 4 + 7 * BPS) + + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE4x4(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + + __asm__ volatile( + "mult $zero, $zero \n\t" + + GET_SSE(0 * BPS, 1 * BPS, 2 * BPS, 3 * BPS) + + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +#undef GET_SSE +#undef GET_SSE_INNER + +#endif // !WORK_AROUND_GCC + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitMIPS32(void) { + VP8ITransform = ITransform; + VP8FTransform = FTransform; + VP8EncQuantizeBlock = QuantizeBlock; + VP8EncQuantize2Blocks = Quantize2Blocks; + VP8TDisto4x4 = Disto4x4; + VP8TDisto16x16 = Disto16x16; +#if !defined(WORK_AROUND_GCC) + VP8SSE16x16 = SSE16x16; + VP8SSE8x8 = SSE8x8; + VP8SSE16x8 = SSE16x8; + VP8SSE4x4 = SSE4x4; +#endif +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(VP8EncDspInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/drivers/webp/dsp/enc_mips_dsp_r2.c b/drivers/webp/dsp/enc_mips_dsp_r2.c new file mode 100644 index 0000000000..7c814fa04a --- /dev/null +++ b/drivers/webp/dsp/enc_mips_dsp_r2.c @@ -0,0 +1,1512 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of speed-critical encoding functions. +// +// Author(s): Darko Laus (darko.laus@imgtec.com) +// Mirko Raus (mirko.raus@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "./mips_macro.h" +#include "../enc/cost.h" +#include "../enc/vp8enci.h" + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; + +// O - output +// I - input (macro doesn't change it) +#define ADD_SUB_HALVES_X4(O0, O1, O2, O3, O4, O5, O6, O7, \ + I0, I1, I2, I3, I4, I5, I6, I7) \ + "addq.ph %[" #O0 "], %[" #I0 "], %[" #I1 "] \n\t" \ + "subq.ph %[" #O1 "], %[" #I0 "], %[" #I1 "] \n\t" \ + "addq.ph %[" #O2 "], %[" #I2 "], %[" #I3 "] \n\t" \ + "subq.ph %[" #O3 "], %[" #I2 "], %[" #I3 "] \n\t" \ + "addq.ph %[" #O4 "], %[" #I4 "], %[" #I5 "] \n\t" \ + "subq.ph %[" #O5 "], %[" #I4 "], %[" #I5 "] \n\t" \ + "addq.ph %[" #O6 "], %[" #I6 "], %[" #I7 "] \n\t" \ + "subq.ph %[" #O7 "], %[" #I6 "], %[" #I7 "] \n\t" + +// IO - input/output +#define ABS_X8(IO0, IO1, IO2, IO3, IO4, IO5, IO6, IO7) \ + "absq_s.ph %[" #IO0 "], %[" #IO0 "] \n\t" \ + "absq_s.ph %[" #IO1 "], %[" #IO1 "] \n\t" \ + "absq_s.ph %[" #IO2 "], %[" #IO2 "] \n\t" \ + "absq_s.ph %[" #IO3 "], %[" #IO3 "] \n\t" \ + "absq_s.ph %[" #IO4 "], %[" #IO4 "] \n\t" \ + "absq_s.ph %[" #IO5 "], %[" #IO5 "] \n\t" \ + "absq_s.ph %[" #IO6 "], %[" #IO6 "] \n\t" \ + "absq_s.ph %[" #IO7 "], %[" #IO7 "] \n\t" + +// dpa.w.ph $ac0 temp0 ,temp1 +// $ac += temp0[31..16] * temp1[31..16] + temp0[15..0] * temp1[15..0] +// dpax.w.ph $ac0 temp0 ,temp1 +// $ac += temp0[31..16] * temp1[15..0] + temp0[15..0] * temp1[31..16] +// O - output +// I - input (macro doesn't change it) +#define MUL_HALF(O0, I0, I1, I2, I3, I4, I5, I6, I7, \ + I8, I9, I10, I11, I12, I13, I14, I15) \ + "mult $ac0, $zero, $zero \n\t" \ + "dpa.w.ph $ac0, %[" #I2 "], %[" #I0 "] \n\t" \ + "dpax.w.ph $ac0, %[" #I5 "], %[" #I6 "] \n\t" \ + "dpa.w.ph $ac0, %[" #I8 "], %[" #I9 "] \n\t" \ + "dpax.w.ph $ac0, %[" #I11 "], %[" #I4 "] \n\t" \ + "dpa.w.ph $ac0, %[" #I12 "], %[" #I7 "] \n\t" \ + "dpax.w.ph $ac0, %[" #I13 "], %[" #I1 "] \n\t" \ + "dpa.w.ph $ac0, %[" #I14 "], %[" #I3 "] \n\t" \ + "dpax.w.ph $ac0, %[" #I15 "], %[" #I10 "] \n\t" \ + "mflo %[" #O0 "], $ac0 \n\t" + +#define OUTPUT_EARLY_CLOBBER_REGS_17() \ + OUTPUT_EARLY_CLOBBER_REGS_10(), \ + [temp11]"=&r"(temp11), [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), \ + [temp14]"=&r"(temp14), [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), \ + [temp17]"=&r"(temp17) + +// macro for one horizontal pass in FTransform +// temp0..temp15 holds tmp[0]..tmp[15] +// A - offset in bytes to load from src and ref buffers +// TEMP0..TEMP3 - registers for corresponding tmp elements +#define HORIZONTAL_PASS(A, TEMP0, TEMP1, TEMP2, TEMP3) \ + "lw %[" #TEMP0 "], 0(%[args]) \n\t" \ + "lw %[" #TEMP1 "], 4(%[args]) \n\t" \ + "lw %[" #TEMP2 "], " XSTR(BPS) "*" #A "(%[" #TEMP0 "]) \n\t" \ + "lw %[" #TEMP3 "], " XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "preceu.ph.qbl %[" #TEMP0 "], %[" #TEMP2 "] \n\t" \ + "preceu.ph.qbl %[" #TEMP1 "], %[" #TEMP3 "] \n\t" \ + "preceu.ph.qbr %[" #TEMP2 "], %[" #TEMP2 "] \n\t" \ + "preceu.ph.qbr %[" #TEMP3 "], %[" #TEMP3 "] \n\t" \ + "subq.ph %[" #TEMP0 "], %[" #TEMP0 "], %[" #TEMP1 "] \n\t" \ + "subq.ph %[" #TEMP2 "], %[" #TEMP2 "], %[" #TEMP3 "] \n\t" \ + "rotr %[" #TEMP0 "], %[" #TEMP0 "], 16 \n\t" \ + "addq.ph %[" #TEMP1 "], %[" #TEMP2 "], %[" #TEMP0 "] \n\t" \ + "subq.ph %[" #TEMP3 "], %[" #TEMP2 "], %[" #TEMP0 "] \n\t" \ + "seh %[" #TEMP0 "], %[" #TEMP1 "] \n\t" \ + "sra %[temp16], %[" #TEMP1 "], 16 \n\t" \ + "seh %[temp19], %[" #TEMP3 "] \n\t" \ + "sra %[" #TEMP3 "], %[" #TEMP3 "], 16 \n\t" \ + "subu %[" #TEMP2 "], %[" #TEMP0 "], %[temp16] \n\t" \ + "addu %[" #TEMP0 "], %[" #TEMP0 "], %[temp16] \n\t" \ + "mul %[temp17], %[temp19], %[c2217] \n\t" \ + "mul %[temp18], %[" #TEMP3 "], %[c5352] \n\t" \ + "mul %[" #TEMP1 "], %[temp19], %[c5352] \n\t" \ + "mul %[temp16], %[" #TEMP3 "], %[c2217] \n\t" \ + "sll %[" #TEMP2 "], %[" #TEMP2 "], 3 \n\t" \ + "sll %[" #TEMP0 "], %[" #TEMP0 "], 3 \n\t" \ + "subu %[" #TEMP3 "], %[temp17], %[temp18] \n\t" \ + "addu %[" #TEMP1 "], %[temp16], %[" #TEMP1 "] \n\t" \ + "addiu %[" #TEMP3 "], %[" #TEMP3 "], 937 \n\t" \ + "addiu %[" #TEMP1 "], %[" #TEMP1 "], 1812 \n\t" \ + "sra %[" #TEMP3 "], %[" #TEMP3 "], 9 \n\t" \ + "sra %[" #TEMP1 "], %[" #TEMP1 "], 9 \n\t" + +// macro for one vertical pass in FTransform +// temp0..temp15 holds tmp[0]..tmp[15] +// A..D - offsets in bytes to store to out buffer +// TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements +#define VERTICAL_PASS(A, B, C, D, TEMP0, TEMP4, TEMP8, TEMP12) \ + "addu %[temp16], %[" #TEMP0 "], %[" #TEMP12 "] \n\t" \ + "subu %[temp19], %[" #TEMP0 "], %[" #TEMP12 "] \n\t" \ + "addu %[temp17], %[" #TEMP4 "], %[" #TEMP8 "] \n\t" \ + "subu %[temp18], %[" #TEMP4 "], %[" #TEMP8 "] \n\t" \ + "mul %[" #TEMP8 "], %[temp19], %[c2217] \n\t" \ + "mul %[" #TEMP12 "], %[temp18], %[c2217] \n\t" \ + "mul %[" #TEMP4 "], %[temp19], %[c5352] \n\t" \ + "mul %[temp18], %[temp18], %[c5352] \n\t" \ + "addiu %[temp16], %[temp16], 7 \n\t" \ + "addu %[" #TEMP0 "], %[temp16], %[temp17] \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 4 \n\t" \ + "addu %[" #TEMP12 "], %[" #TEMP12 "], %[" #TEMP4 "] \n\t" \ + "subu %[" #TEMP4 "], %[temp16], %[temp17] \n\t" \ + "sra %[" #TEMP4 "], %[" #TEMP4 "], 4 \n\t" \ + "addiu %[" #TEMP8 "], %[" #TEMP8 "], 30000 \n\t" \ + "addiu %[" #TEMP12 "], %[" #TEMP12 "], 12000 \n\t" \ + "addiu %[" #TEMP8 "], %[" #TEMP8 "], 21000 \n\t" \ + "subu %[" #TEMP8 "], %[" #TEMP8 "], %[temp18] \n\t" \ + "sra %[" #TEMP12 "], %[" #TEMP12 "], 16 \n\t" \ + "sra %[" #TEMP8 "], %[" #TEMP8 "], 16 \n\t" \ + "addiu %[temp16], %[" #TEMP12 "], 1 \n\t" \ + "movn %[" #TEMP12 "], %[temp16], %[temp19] \n\t" \ + "sh %[" #TEMP0 "], " #A "(%[temp20]) \n\t" \ + "sh %[" #TEMP4 "], " #C "(%[temp20]) \n\t" \ + "sh %[" #TEMP8 "], " #D "(%[temp20]) \n\t" \ + "sh %[" #TEMP12 "], " #B "(%[temp20]) \n\t" + +static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) { + const int c2217 = 2217; + const int c5352 = 5352; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + int temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16; + int temp17, temp18, temp19, temp20; + const int* const args[3] = + { (const int*)src, (const int*)ref, (const int*)out }; + + __asm__ volatile ( + HORIZONTAL_PASS(0, temp0, temp1, temp2, temp3) + HORIZONTAL_PASS(1, temp4, temp5, temp6, temp7) + HORIZONTAL_PASS(2, temp8, temp9, temp10, temp11) + HORIZONTAL_PASS(3, temp12, temp13, temp14, temp15) + "lw %[temp20], 8(%[args]) \n\t" + VERTICAL_PASS(0, 8, 16, 24, temp0, temp4, temp8, temp12) + VERTICAL_PASS(2, 10, 18, 26, temp1, temp5, temp9, temp13) + VERTICAL_PASS(4, 12, 20, 28, temp2, temp6, temp10, temp14) + VERTICAL_PASS(6, 14, 22, 30, temp3, temp7, temp11, temp15) + OUTPUT_EARLY_CLOBBER_REGS_18(), + [temp0]"=&r"(temp0), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20) + : [args]"r"(args), [c2217]"r"(c2217), [c5352]"r"(c5352) + : "memory", "hi", "lo" + ); +} + +#undef VERTICAL_PASS +#undef HORIZONTAL_PASS + +static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in, + uint8_t* dst) { + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17, temp18; + + __asm__ volatile ( + "ulw %[temp1], 0(%[in]) \n\t" + "ulw %[temp2], 16(%[in]) \n\t" + LOAD_IN_X2(temp5, temp6, 24, 26) + ADD_SUB_HALVES(temp3, temp4, temp1, temp2) + LOAD_IN_X2(temp1, temp2, 8, 10) + MUL_SHIFT_SUM(temp7, temp8, temp9, temp10, temp11, temp12, temp13, temp14, + temp10, temp8, temp9, temp7, temp1, temp2, temp5, temp6, + temp13, temp11, temp14, temp12) + INSERT_HALF_X2(temp8, temp7, temp10, temp9) + "ulw %[temp17], 4(%[in]) \n\t" + "ulw %[temp18], 20(%[in]) \n\t" + ADD_SUB_HALVES(temp1, temp2, temp3, temp8) + ADD_SUB_HALVES(temp5, temp6, temp4, temp7) + ADD_SUB_HALVES(temp7, temp8, temp17, temp18) + LOAD_IN_X2(temp17, temp18, 12, 14) + LOAD_IN_X2(temp9, temp10, 28, 30) + MUL_SHIFT_SUM(temp11, temp12, temp13, temp14, temp15, temp16, temp4, temp17, + temp12, temp14, temp11, temp13, temp17, temp18, temp9, temp10, + temp15, temp4, temp16, temp17) + INSERT_HALF_X2(temp11, temp12, temp13, temp14) + ADD_SUB_HALVES(temp17, temp8, temp8, temp11) + ADD_SUB_HALVES(temp3, temp4, temp7, temp12) + + // horizontal + SRA_16(temp9, temp10, temp11, temp12, temp1, temp2, temp5, temp6) + INSERT_HALF_X2(temp1, temp6, temp5, temp2) + SRA_16(temp13, temp14, temp15, temp16, temp3, temp4, temp17, temp8) + "repl.ph %[temp2], 0x4 \n\t" + INSERT_HALF_X2(temp3, temp8, temp17, temp4) + "addq.ph %[temp1], %[temp1], %[temp2] \n\t" + "addq.ph %[temp6], %[temp6], %[temp2] \n\t" + ADD_SUB_HALVES(temp2, temp4, temp1, temp3) + ADD_SUB_HALVES(temp5, temp7, temp6, temp8) + MUL_SHIFT_SUM(temp1, temp3, temp6, temp8, temp9, temp13, temp17, temp18, + temp3, temp13, temp1, temp9, temp9, temp13, temp11, temp15, + temp6, temp17, temp8, temp18) + MUL_SHIFT_SUM(temp6, temp8, temp18, temp17, temp11, temp15, temp12, temp16, + temp8, temp15, temp6, temp11, temp12, temp16, temp10, temp14, + temp18, temp12, temp17, temp16) + INSERT_HALF_X2(temp1, temp3, temp9, temp13) + INSERT_HALF_X2(temp6, temp8, temp11, temp15) + SHIFT_R_SUM_X2(temp9, temp10, temp11, temp12, temp13, temp14, temp15, + temp16, temp2, temp4, temp5, temp7, temp3, temp1, temp8, + temp6) + PACK_2_HALVES_TO_WORD(temp1, temp2, temp3, temp4, temp9, temp12, temp13, + temp16, temp11, temp10, temp15, temp14) + LOAD_WITH_OFFSET_X4(temp10, temp11, temp14, temp15, ref, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp5, temp6, temp7, temp8, temp17, temp18, temp10, + temp11, temp10, temp11, temp14, temp15) + STORE_SAT_SUM_X2(temp5, temp6, temp7, temp8, temp17, temp18, temp10, temp11, + temp9, temp12, temp1, temp2, temp13, temp16, temp3, temp4, + dst, 0, 1, 2, 3, BPS) + + OUTPUT_EARLY_CLOBBER_REGS_18() + : [dst]"r"(dst), [in]"r"(in), [kC1]"r"(kC1), [kC2]"r"(kC2), [ref]"r"(ref) + : "memory", "hi", "lo" + ); +} + +static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst, + int do_two) { + ITransformOne(ref, in, dst); + if (do_two) { + ITransformOne(ref + 4, in + 16, dst + 4); + } +} + +static int Disto4x4(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17; + + __asm__ volatile ( + LOAD_WITH_OFFSET_X4(temp1, temp2, temp3, temp4, a, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp5, temp6, temp7, temp8, temp9,temp10, temp11, + temp12, temp1, temp2, temp3, temp4) + ADD_SUB_HALVES_X4(temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, + temp5, temp6, temp7, temp8, temp9, temp10, temp11, temp12) + PACK_2_HALVES_TO_WORD(temp9, temp10, temp11, temp12, temp1, temp3, temp5, + temp7, temp2, temp4, temp6, temp8) + ADD_SUB_HALVES_X4(temp2, temp4, temp6, temp8, temp9, temp1, temp3, temp10, + temp1, temp9, temp3, temp10, temp5, temp11, temp7, temp12) + ADD_SUB_HALVES_X4(temp5, temp11, temp7, temp2, temp9, temp3, temp6, temp12, + temp2, temp9, temp6, temp3, temp4, temp1, temp8, temp10) + ADD_SUB_HALVES_X4(temp1, temp4, temp10, temp8, temp7, temp11, temp5, temp2, + temp5, temp7, temp11, temp2, temp9, temp6, temp3, temp12) + ABS_X8(temp1, temp4, temp10, temp8, temp7, temp11, temp5, temp2) + LOAD_WITH_OFFSET_X4(temp3, temp6, temp9, temp12, w, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + LOAD_WITH_OFFSET_X4(temp13, temp14, temp15, temp16, w, + 0, 4, 8, 12, + 1, 1, 1, 1, + 16) + MUL_HALF(temp17, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, + temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16) + LOAD_WITH_OFFSET_X4(temp1, temp2, temp3, temp4, b, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp5,temp6, temp7, temp8, temp9,temp10, temp11, + temp12, temp1, temp2, temp3, temp4) + ADD_SUB_HALVES_X4(temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, + temp5, temp6, temp7, temp8, temp9, temp10, temp11, temp12) + PACK_2_HALVES_TO_WORD(temp9, temp10, temp11, temp12, temp1, temp3, temp5, + temp7, temp2, temp4, temp6, temp8) + ADD_SUB_HALVES_X4(temp2, temp4, temp6, temp8, temp9, temp1, temp3, temp10, + temp1, temp9, temp3, temp10, temp5, temp11, temp7, temp12) + ADD_SUB_HALVES_X4(temp5, temp11, temp7, temp2, temp9, temp3, temp6, temp12, + temp2, temp9, temp6, temp3, temp4, temp1, temp8, temp10) + ADD_SUB_HALVES_X4(temp1, temp4, temp10, temp8, temp7, temp11, temp5, temp2, + temp5, temp7, temp11, temp2, temp9, temp6, temp3, temp12) + ABS_X8(temp1, temp4, temp10, temp8, temp7, temp11, temp5, temp2) + LOAD_WITH_OFFSET_X4(temp3, temp6, temp9, temp12, w, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + LOAD_WITH_OFFSET_X4(temp13, temp14, temp15, temp16, w, + 0, 4, 8, 12, + 1, 1, 1, 1, + 16) + MUL_HALF(temp3, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, + temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16) + OUTPUT_EARLY_CLOBBER_REGS_17() + : [a]"r"(a), [b]"r"(b), [w]"r"(w) + : "memory", "hi", "lo" + ); + return abs(temp3 - temp17) >> 5; +} + +static int Disto16x16(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4(a + x + y, b + x + y, w); + } + } + return D; +} + +//------------------------------------------------------------------------------ +// Intra predictions + +#define FILL_PART(J, SIZE) \ + "usw %[value], 0+" #J "*" XSTR(BPS) "(%[dst]) \n\t" \ + "usw %[value], 4+" #J "*" XSTR(BPS) "(%[dst]) \n\t" \ + ".if " #SIZE " == 16 \n\t" \ + "usw %[value], 8+" #J "*" XSTR(BPS) "(%[dst]) \n\t" \ + "usw %[value], 12+" #J "*" XSTR(BPS) "(%[dst]) \n\t" \ + ".endif \n\t" + +#define FILL_8_OR_16(DST, VALUE, SIZE) do { \ + int value = (VALUE); \ + __asm__ volatile ( \ + "replv.qb %[value], %[value] \n\t" \ + FILL_PART( 0, SIZE) \ + FILL_PART( 1, SIZE) \ + FILL_PART( 2, SIZE) \ + FILL_PART( 3, SIZE) \ + FILL_PART( 4, SIZE) \ + FILL_PART( 5, SIZE) \ + FILL_PART( 6, SIZE) \ + FILL_PART( 7, SIZE) \ + ".if " #SIZE " == 16 \n\t" \ + FILL_PART( 8, 16) \ + FILL_PART( 9, 16) \ + FILL_PART(10, 16) \ + FILL_PART(11, 16) \ + FILL_PART(12, 16) \ + FILL_PART(13, 16) \ + FILL_PART(14, 16) \ + FILL_PART(15, 16) \ + ".endif \n\t" \ + : [value]"+&r"(value) \ + : [dst]"r"((DST)) \ + : "memory" \ + ); \ +} while (0) + +#define VERTICAL_PRED(DST, TOP, SIZE) \ +static WEBP_INLINE void VerticalPred##SIZE(uint8_t* (DST), \ + const uint8_t* (TOP)) { \ + int j; \ + if ((TOP)) { \ + for (j = 0; j < (SIZE); ++j) memcpy((DST) + j * BPS, (TOP), (SIZE)); \ + } else { \ + FILL_8_OR_16((DST), 127, (SIZE)); \ + } \ +} + +VERTICAL_PRED(dst, top, 8) +VERTICAL_PRED(dst, top, 16) + +#undef VERTICAL_PRED + +#define HORIZONTAL_PRED(DST, LEFT, SIZE) \ +static WEBP_INLINE void HorizontalPred##SIZE(uint8_t* (DST), \ + const uint8_t* (LEFT)) { \ + if (LEFT) { \ + int j; \ + for (j = 0; j < (SIZE); ++j) { \ + memset((DST) + j * BPS, (LEFT)[j], (SIZE)); \ + } \ + } else { \ + FILL_8_OR_16((DST), 129, (SIZE)); \ + } \ +} + +HORIZONTAL_PRED(dst, left, 8) +HORIZONTAL_PRED(dst, left, 16) + +#undef HORIZONTAL_PRED + +#define CLIPPING() \ + "preceu.ph.qbl %[temp2], %[temp0] \n\t" \ + "preceu.ph.qbr %[temp0], %[temp0] \n\t" \ + "preceu.ph.qbl %[temp3], %[temp1] \n\t" \ + "preceu.ph.qbr %[temp1], %[temp1] \n\t" \ + "addu.ph %[temp2], %[temp2], %[leftY_1] \n\t" \ + "addu.ph %[temp0], %[temp0], %[leftY_1] \n\t" \ + "addu.ph %[temp3], %[temp3], %[leftY_1] \n\t" \ + "addu.ph %[temp1], %[temp1], %[leftY_1] \n\t" \ + "shll_s.ph %[temp2], %[temp2], 7 \n\t" \ + "shll_s.ph %[temp0], %[temp0], 7 \n\t" \ + "shll_s.ph %[temp3], %[temp3], 7 \n\t" \ + "shll_s.ph %[temp1], %[temp1], 7 \n\t" \ + "precrqu_s.qb.ph %[temp0], %[temp2], %[temp0] \n\t" \ + "precrqu_s.qb.ph %[temp1], %[temp3], %[temp1] \n\t" + +#define CLIP_8B_TO_DST(DST, LEFT, TOP, SIZE) do { \ + int leftY_1 = ((int)(LEFT)[y] << 16) + (LEFT)[y]; \ + int temp0, temp1, temp2, temp3; \ + __asm__ volatile ( \ + "replv.ph %[leftY_1], %[leftY_1] \n\t" \ + "ulw %[temp0], 0(%[top]) \n\t" \ + "ulw %[temp1], 4(%[top]) \n\t" \ + "subu.ph %[leftY_1], %[leftY_1], %[left_1] \n\t" \ + CLIPPING() \ + "usw %[temp0], 0(%[dst]) \n\t" \ + "usw %[temp1], 4(%[dst]) \n\t" \ + ".if " #SIZE " == 16 \n\t" \ + "ulw %[temp0], 8(%[top]) \n\t" \ + "ulw %[temp1], 12(%[top]) \n\t" \ + CLIPPING() \ + "usw %[temp0], 8(%[dst]) \n\t" \ + "usw %[temp1], 12(%[dst]) \n\t" \ + ".endif \n\t" \ + : [leftY_1]"+&r"(leftY_1), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), \ + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3) \ + : [left_1]"r"(left_1), [top]"r"((TOP)), [dst]"r"((DST)) \ + : "memory" \ + ); \ +} while (0) + +#define CLIP_TO_DST(DST, LEFT, TOP, SIZE) do { \ + int y; \ + const int left_1 = ((int)(LEFT)[-1] << 16) + (LEFT)[-1]; \ + for (y = 0; y < (SIZE); ++y) { \ + CLIP_8B_TO_DST((DST), (LEFT), (TOP), (SIZE)); \ + (DST) += BPS; \ + } \ +} while (0) + +#define TRUE_MOTION(DST, LEFT, TOP, SIZE) \ +static WEBP_INLINE void TrueMotion##SIZE(uint8_t* (DST), const uint8_t* (LEFT),\ + const uint8_t* (TOP)) { \ + if ((LEFT) != NULL) { \ + if ((TOP) != NULL) { \ + CLIP_TO_DST((DST), (LEFT), (TOP), (SIZE)); \ + } else { \ + HorizontalPred##SIZE((DST), (LEFT)); \ + } \ + } else { \ + /* true motion without left samples (hence: with default 129 value) */ \ + /* is equivalent to VE prediction where you just copy the top samples. */ \ + /* Note that if top samples are not available, the default value is */ \ + /* then 129, and not 127 as in the VerticalPred case. */ \ + if ((TOP) != NULL) { \ + VerticalPred##SIZE((DST), (TOP)); \ + } else { \ + FILL_8_OR_16((DST), 129, (SIZE)); \ + } \ + } \ +} + +TRUE_MOTION(dst, left, top, 8) +TRUE_MOTION(dst, left, top, 16) + +#undef TRUE_MOTION +#undef CLIP_TO_DST +#undef CLIP_8B_TO_DST +#undef CLIPPING + +static WEBP_INLINE void DCMode16(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + int DC, DC1; + int temp0, temp1, temp2, temp3; + + __asm__ volatile( + "beqz %[top], 2f \n\t" + LOAD_WITH_OFFSET_X4(temp0, temp1, temp2, temp3, top, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[DC], %[temp0], %[temp2] \n\t" + "move %[DC1], %[DC] \n\t" + "beqz %[left], 1f \n\t" + LOAD_WITH_OFFSET_X4(temp0, temp1, temp2, temp3, left, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[DC1], %[temp0], %[temp2] \n\t" + "1: \n\t" + "addu %[DC], %[DC], %[DC1] \n\t" + "j 3f \n\t" + "2: \n\t" + "beqz %[left], 4f \n\t" + LOAD_WITH_OFFSET_X4(temp0, temp1, temp2, temp3, left, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[DC], %[temp0], %[temp2] \n\t" + "addu %[DC], %[DC], %[DC] \n\t" + "3: \n\t" + "shra_r.w %[DC], %[DC], 5 \n\t" + "j 5f \n\t" + "4: \n\t" + "li %[DC], 0x80 \n\t" + "5: \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [DC]"=&r"(DC), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [DC1]"=&r"(DC1) + : [left]"r"(left), [top]"r"(top) + : "memory" + ); + + FILL_8_OR_16(dst, DC, 16); +} + +static WEBP_INLINE void DCMode8(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + int DC, DC1; + int temp0, temp1, temp2, temp3; + + __asm__ volatile( + "beqz %[top], 2f \n\t" + "ulw %[temp0], 0(%[top]) \n\t" + "ulw %[temp1], 4(%[top]) \n\t" + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[DC], %[temp0], %[temp1] \n\t" + "move %[DC1], %[DC] \n\t" + "beqz %[left], 1f \n\t" + "ulw %[temp2], 0(%[left]) \n\t" + "ulw %[temp3], 4(%[left]) \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[DC1], %[temp2], %[temp3] \n\t" + "1: \n\t" + "addu %[DC], %[DC], %[DC1] \n\t" + "j 3f \n\t" + "2: \n\t" + "beqz %[left], 4f \n\t" + "ulw %[temp2], 0(%[left]) \n\t" + "ulw %[temp3], 4(%[left]) \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[DC], %[temp2], %[temp3] \n\t" + "addu %[DC], %[DC], %[DC] \n\t" + "3: \n\t" + "shra_r.w %[DC], %[DC], 4 \n\t" + "j 5f \n\t" + "4: \n\t" + "li %[DC], 0x80 \n\t" + "5: \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [DC]"=&r"(DC), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [DC1]"=&r"(DC1) + : [left]"r"(left), [top]"r"(top) + : "memory" + ); + + FILL_8_OR_16(dst, DC, 8); +} + +static void DC4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1; + __asm__ volatile( + "ulw %[temp0], 0(%[top]) \n\t" + "ulw %[temp1], -5(%[top]) \n\t" + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addiu %[temp0], %[temp0], 4 \n\t" + "srl %[temp0], %[temp0], 3 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + "usw %[temp0], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void TM4(uint8_t* dst, const uint8_t* top) { + int a10, a32, temp0, temp1, temp2, temp3, temp4, temp5; + const int c35 = 0xff00ff; + __asm__ volatile ( + "lbu %[temp1], 0(%[top]) \n\t" + "lbu %[a10], 1(%[top]) \n\t" + "lbu %[temp2], 2(%[top]) \n\t" + "lbu %[a32], 3(%[top]) \n\t" + "ulw %[temp0], -5(%[top]) \n\t" + "lbu %[temp4], -1(%[top]) \n\t" + "append %[a10], %[temp1], 16 \n\t" + "append %[a32], %[temp2], 16 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "shrl.ph %[temp1], %[temp0], 8 \n\t" + "and %[temp0], %[temp0], %[c35] \n\t" + "subu.ph %[temp1], %[temp1], %[temp4] \n\t" + "subu.ph %[temp0], %[temp0], %[temp4] \n\t" + "srl %[temp2], %[temp1], 16 \n\t" + "srl %[temp3], %[temp0], 16 \n\t" + "replv.ph %[temp2], %[temp2] \n\t" + "replv.ph %[temp3], %[temp3] \n\t" + "replv.ph %[temp4], %[temp1] \n\t" + "replv.ph %[temp5], %[temp0] \n\t" + "addu.ph %[temp0], %[temp3], %[a10] \n\t" + "addu.ph %[temp1], %[temp3], %[a32] \n\t" + "addu.ph %[temp3], %[temp2], %[a10] \n\t" + "addu.ph %[temp2], %[temp2], %[a32] \n\t" + "shll_s.ph %[temp0], %[temp0], 7 \n\t" + "shll_s.ph %[temp1], %[temp1], 7 \n\t" + "shll_s.ph %[temp3], %[temp3], 7 \n\t" + "shll_s.ph %[temp2], %[temp2], 7 \n\t" + "precrqu_s.qb.ph %[temp0], %[temp1], %[temp0] \n\t" + "precrqu_s.qb.ph %[temp1], %[temp2], %[temp3] \n\t" + "addu.ph %[temp2], %[temp5], %[a10] \n\t" + "addu.ph %[temp3], %[temp5], %[a32] \n\t" + "addu.ph %[temp5], %[temp4], %[a10] \n\t" + "addu.ph %[temp4], %[temp4], %[a32] \n\t" + "shll_s.ph %[temp2], %[temp2], 7 \n\t" + "shll_s.ph %[temp3], %[temp3], 7 \n\t" + "shll_s.ph %[temp4], %[temp4], 7 \n\t" + "shll_s.ph %[temp5], %[temp5], 7 \n\t" + "precrqu_s.qb.ph %[temp2], %[temp3], %[temp2] \n\t" + "precrqu_s.qb.ph %[temp3], %[temp4], %[temp5] \n\t" + "usw %[temp1], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp3], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp2], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [a10]"=&r"(a10), [a32]"=&r"(a32) + : [c35]"r"(c35), [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void VE4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + __asm__ volatile( + "ulw %[temp0], -1(%[top]) \n\t" + "ulh %[temp1], 3(%[top]) \n\t" + "preceu.ph.qbr %[temp2], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "packrl.ph %[temp5], %[temp3], %[temp2] \n\t" + "packrl.ph %[temp6], %[temp4], %[temp3] \n\t" + "shll.ph %[temp5], %[temp5], 1 \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp2], %[temp5], %[temp2] \n\t" + "addq.ph %[temp6], %[temp6], %[temp4] \n\t" + "addq.ph %[temp2], %[temp2], %[temp3] \n\t" + "addq.ph %[temp6], %[temp6], %[temp3] \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precr.qb.ph %[temp4], %[temp6], %[temp2] \n\t" + "usw %[temp4], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp4], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp4], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp4], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void HE4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + __asm__ volatile( + "ulw %[temp0], -4(%[top]) \n\t" + "lbu %[temp1], -5(%[top]) \n\t" + "preceu.ph.qbr %[temp2], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp0] \n\t" + "replv.ph %[temp4], %[temp1] \n\t" + "packrl.ph %[temp5], %[temp3], %[temp2] \n\t" + "packrl.ph %[temp6], %[temp2], %[temp4] \n\t" + "shll.ph %[temp5], %[temp5], 1 \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp3], %[temp3], %[temp5] \n\t" + "addq.ph %[temp3], %[temp3], %[temp2] \n\t" + "addq.ph %[temp2], %[temp2], %[temp6] \n\t" + "addq.ph %[temp2], %[temp2], %[temp4] \n\t" + "shra_r.ph %[temp3], %[temp3], 2 \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "replv.qb %[temp0], %[temp3] \n\t" + "replv.qb %[temp1], %[temp2] \n\t" + "srl %[temp3], %[temp3], 16 \n\t" + "srl %[temp2], %[temp2], 16 \n\t" + "replv.qb %[temp3], %[temp3] \n\t" + "replv.qb %[temp2], %[temp2] \n\t" + "usw %[temp3], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp2], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp1], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void RD4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5; + int temp6, temp7, temp8, temp9, temp10, temp11; + __asm__ volatile( + "ulw %[temp0], -5(%[top]) \n\t" + "ulw %[temp1], -1(%[top]) \n\t" + "preceu.ph.qbl %[temp2], %[temp0] \n\t" + "preceu.ph.qbr %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "preceu.ph.qbl %[temp5], %[temp1] \n\t" + "packrl.ph %[temp6], %[temp2], %[temp3] \n\t" + "packrl.ph %[temp7], %[temp4], %[temp2] \n\t" + "packrl.ph %[temp8], %[temp5], %[temp4] \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp9], %[temp2], %[temp6] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "addq.ph %[temp9], %[temp9], %[temp3] \n\t" + "shll.ph %[temp8], %[temp8], 1 \n\t" + "shra_r.ph %[temp9], %[temp9], 2 \n\t" + "addq.ph %[temp10], %[temp4], %[temp7] \n\t" + "addq.ph %[temp11], %[temp5], %[temp8] \n\t" + "addq.ph %[temp10], %[temp10], %[temp2] \n\t" + "addq.ph %[temp11], %[temp11], %[temp4] \n\t" + "shra_r.ph %[temp10], %[temp10], 2 \n\t" + "shra_r.ph %[temp11], %[temp11], 2 \n\t" + "lbu %[temp0], 3(%[top]) \n\t" + "lbu %[temp1], 2(%[top]) \n\t" + "lbu %[temp2], 1(%[top]) \n\t" + "sll %[temp1], %[temp1], 1 \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp0], %[temp0], %[temp2] \n\t" + "precr.qb.ph %[temp9], %[temp10], %[temp9] \n\t" + "shra_r.w %[temp0], %[temp0], 2 \n\t" + "precr.qb.ph %[temp10], %[temp11], %[temp10] \n\t" + "usw %[temp9], 3*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp10], 1*" XSTR(BPS) "(%[dst]) \n\t" + "prepend %[temp9], %[temp11], 8 \n\t" + "prepend %[temp10], %[temp0], 8 \n\t" + "usw %[temp9], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp10], 0*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void VR4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + "ulw %[temp0], -4(%[top]) \n\t" + "ulw %[temp1], 0(%[top]) \n\t" + "preceu.ph.qbl %[temp2], %[temp0] \n\t" + "preceu.ph.qbr %[temp0], %[temp0] \n\t" + "preceu.ph.qbla %[temp3], %[temp1] \n\t" + "preceu.ph.qbra %[temp1], %[temp1] \n\t" + "packrl.ph %[temp7], %[temp3], %[temp2] \n\t" + "addqh_r.ph %[temp4], %[temp1], %[temp3] \n\t" + "move %[temp6], %[temp1] \n\t" + "append %[temp1], %[temp2], 16 \n\t" + "shll.ph %[temp9], %[temp6], 1 \n\t" + "addqh_r.ph %[temp5], %[temp7], %[temp6] \n\t" + "shll.ph %[temp8], %[temp7], 1 \n\t" + "addu.ph %[temp3], %[temp7], %[temp3] \n\t" + "addu.ph %[temp1], %[temp1], %[temp6] \n\t" + "packrl.ph %[temp7], %[temp2], %[temp0] \n\t" + "addu.ph %[temp6], %[temp0], %[temp2] \n\t" + "addu.ph %[temp3], %[temp3], %[temp9] \n\t" + "addu.ph %[temp1], %[temp1], %[temp8] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "shra_r.ph %[temp3], %[temp3], 2 \n\t" + "shra_r.ph %[temp1], %[temp1], 2 \n\t" + "addu.ph %[temp6], %[temp6], %[temp7] \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precrq.ph.w %[temp8], %[temp4], %[temp5] \n\t" + "append %[temp4], %[temp5], 16 \n\t" + "precrq.ph.w %[temp2], %[temp3], %[temp1] \n\t" + "append %[temp3], %[temp1], 16 \n\t" + "precr.qb.ph %[temp8], %[temp8], %[temp4] \n\t" + "precr.qb.ph %[temp3], %[temp2], %[temp3] \n\t" + "usw %[temp8], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp3], 1*" XSTR(BPS) "(%[dst]) \n\t" + "append %[temp3], %[temp6], 8 \n\t" + "srl %[temp6], %[temp6], 16 \n\t" + "append %[temp8], %[temp6], 8 \n\t" + "usw %[temp3], 3*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp8], 2*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void LD4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5; + int temp6, temp7, temp8, temp9, temp10, temp11; + __asm__ volatile( + "ulw %[temp0], 0(%[top]) \n\t" + "ulw %[temp1], 4(%[top]) \n\t" + "preceu.ph.qbl %[temp2], %[temp0] \n\t" + "preceu.ph.qbr %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "preceu.ph.qbl %[temp5], %[temp1] \n\t" + "packrl.ph %[temp6], %[temp2], %[temp3] \n\t" + "packrl.ph %[temp7], %[temp4], %[temp2] \n\t" + "packrl.ph %[temp8], %[temp5], %[temp4] \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp9], %[temp2], %[temp6] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "addq.ph %[temp9], %[temp9], %[temp3] \n\t" + "shll.ph %[temp8], %[temp8], 1 \n\t" + "shra_r.ph %[temp9], %[temp9], 2 \n\t" + "addq.ph %[temp10], %[temp4], %[temp7] \n\t" + "addq.ph %[temp11], %[temp5], %[temp8] \n\t" + "addq.ph %[temp10], %[temp10], %[temp2] \n\t" + "addq.ph %[temp11], %[temp11], %[temp4] \n\t" + "shra_r.ph %[temp10], %[temp10], 2 \n\t" + "shra_r.ph %[temp11], %[temp11], 2 \n\t" + "srl %[temp1], %[temp1], 24 \n\t" + "sll %[temp1], %[temp1], 1 \n\t" + "raddu.w.qb %[temp5], %[temp5] \n\t" + "precr.qb.ph %[temp9], %[temp10], %[temp9] \n\t" + "precr.qb.ph %[temp10], %[temp11], %[temp10] \n\t" + "addu %[temp1], %[temp1], %[temp5] \n\t" + "shra_r.w %[temp1], %[temp1], 2 \n\t" + "usw %[temp9], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp10], 2*" XSTR(BPS) "(%[dst]) \n\t" + "prepend %[temp9], %[temp11], 8 \n\t" + "prepend %[temp10], %[temp1], 8 \n\t" + "usw %[temp9], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp10], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void VL4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + "ulw %[temp0], 0(%[top]) \n\t" + "ulw %[temp1], 4(%[top]) \n\t" + "preceu.ph.qbla %[temp2], %[temp0] \n\t" + "preceu.ph.qbra %[temp0], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp1] \n\t" + "preceu.ph.qbr %[temp1], %[temp1] \n\t" + "addqh_r.ph %[temp4], %[temp0], %[temp2] \n\t" + "packrl.ph %[temp7], %[temp1], %[temp0] \n\t" + "precrq.ph.w %[temp6], %[temp1], %[temp2] \n\t" + "shll.ph %[temp9], %[temp2], 1 \n\t" + "addqh_r.ph %[temp5], %[temp7], %[temp2] \n\t" + "shll.ph %[temp8], %[temp7], 1 \n\t" + "addu.ph %[temp2], %[temp2], %[temp6] \n\t" + "addu.ph %[temp0], %[temp0], %[temp7] \n\t" + "packrl.ph %[temp7], %[temp3], %[temp1] \n\t" + "addu.ph %[temp6], %[temp1], %[temp3] \n\t" + "addu.ph %[temp2], %[temp2], %[temp8] \n\t" + "addu.ph %[temp0], %[temp0], %[temp9] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "shra_r.ph %[temp0], %[temp0], 2 \n\t" + "addu.ph %[temp6], %[temp6], %[temp7] \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precrq.ph.w %[temp8], %[temp5], %[temp4] \n\t" + "append %[temp5], %[temp4], 16 \n\t" + "precrq.ph.w %[temp3], %[temp2], %[temp0] \n\t" + "append %[temp2], %[temp0], 16 \n\t" + "precr.qb.ph %[temp8], %[temp8], %[temp5] \n\t" + "precr.qb.ph %[temp3], %[temp3], %[temp2] \n\t" + "usw %[temp8], 0*" XSTR(BPS) "(%[dst]) \n\t" + "prepend %[temp8], %[temp6], 8 \n\t" + "usw %[temp3], 1*" XSTR(BPS) "(%[dst]) \n\t" + "srl %[temp6], %[temp6], 16 \n\t" + "prepend %[temp3], %[temp6], 8 \n\t" + "usw %[temp8], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp3], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void HD4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + "ulw %[temp0], -5(%[top]) \n\t" + "ulw %[temp1], -1(%[top]) \n\t" + "preceu.ph.qbla %[temp2], %[temp0] \n\t" + "preceu.ph.qbra %[temp0], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp1] \n\t" + "preceu.ph.qbr %[temp1], %[temp1] \n\t" + "addqh_r.ph %[temp4], %[temp0], %[temp2] \n\t" + "packrl.ph %[temp7], %[temp1], %[temp0] \n\t" + "precrq.ph.w %[temp6], %[temp1], %[temp2] \n\t" + "shll.ph %[temp9], %[temp2], 1 \n\t" + "addqh_r.ph %[temp5], %[temp7], %[temp2] \n\t" + "shll.ph %[temp8], %[temp7], 1 \n\t" + "addu.ph %[temp2], %[temp2], %[temp6] \n\t" + "addu.ph %[temp0], %[temp0], %[temp7] \n\t" + "packrl.ph %[temp7], %[temp3], %[temp1] \n\t" + "addu.ph %[temp6], %[temp1], %[temp3] \n\t" + "addu.ph %[temp2], %[temp2], %[temp8] \n\t" + "addu.ph %[temp0], %[temp0], %[temp9] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "shra_r.ph %[temp0], %[temp0], 2 \n\t" + "addu.ph %[temp6], %[temp6], %[temp7] \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precrq.ph.w %[temp1], %[temp2], %[temp5] \n\t" + "precrq.ph.w %[temp3], %[temp0], %[temp4] \n\t" + "precr.qb.ph %[temp7], %[temp6], %[temp1] \n\t" + "precr.qb.ph %[temp6], %[temp1], %[temp3] \n\t" + "usw %[temp7], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp6], 1*" XSTR(BPS) "(%[dst]) \n\t" + "append %[temp2], %[temp5], 16 \n\t" + "append %[temp0], %[temp4], 16 \n\t" + "precr.qb.ph %[temp5], %[temp3], %[temp2] \n\t" + "precr.qb.ph %[temp4], %[temp2], %[temp0] \n\t" + "usw %[temp5], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp4], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void HU4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + __asm__ volatile ( + "ulw %[temp0], -5(%[top]) \n\t" + "preceu.ph.qbl %[temp1], %[temp0] \n\t" + "preceu.ph.qbr %[temp2], %[temp0] \n\t" + "packrl.ph %[temp3], %[temp1], %[temp2] \n\t" + "replv.qb %[temp7], %[temp2] \n\t" + "addqh_r.ph %[temp4], %[temp1], %[temp3] \n\t" + "addqh_r.ph %[temp5], %[temp3], %[temp2] \n\t" + "shll.ph %[temp6], %[temp3], 1 \n\t" + "addu.ph %[temp3], %[temp2], %[temp3] \n\t" + "addu.ph %[temp6], %[temp1], %[temp6] \n\t" + "shll.ph %[temp0], %[temp2], 1 \n\t" + "addu.ph %[temp6], %[temp6], %[temp2] \n\t" + "addu.ph %[temp0], %[temp3], %[temp0] \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "shra_r.ph %[temp0], %[temp0], 2 \n\t" + "packrl.ph %[temp3], %[temp6], %[temp5] \n\t" + "precrq.ph.w %[temp2], %[temp6], %[temp4] \n\t" + "append %[temp0], %[temp5], 16 \n\t" + "precr.qb.ph %[temp3], %[temp3], %[temp2] \n\t" + "usw %[temp3], 0*" XSTR(BPS) "(%[dst]) \n\t" + "precr.qb.ph %[temp1], %[temp7], %[temp0] \n\t" + "usw %[temp7], 3*" XSTR(BPS) "(%[dst]) \n\t" + "packrl.ph %[temp2], %[temp1], %[temp3] \n\t" + "usw %[temp1], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp2], 1*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +//------------------------------------------------------------------------------ +// Chroma 8x8 prediction (paragraph 12.2) + +static void IntraChromaPreds(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + // U block + DCMode8(C8DC8 + dst, left, top); + VerticalPred8(C8VE8 + dst, top); + HorizontalPred8(C8HE8 + dst, left); + TrueMotion8(C8TM8 + dst, left, top); + // V block + dst += 8; + if (top) top += 8; + if (left) left += 16; + DCMode8(C8DC8 + dst, left, top); + VerticalPred8(C8VE8 + dst, top); + HorizontalPred8(C8HE8 + dst, left); + TrueMotion8(C8TM8 + dst, left, top); +} + +//------------------------------------------------------------------------------ +// luma 16x16 prediction (paragraph 12.3) + +static void Intra16Preds(uint8_t* dst, + const uint8_t* left, const uint8_t* top) { + DCMode16(I16DC16 + dst, left, top); + VerticalPred16(I16VE16 + dst, top); + HorizontalPred16(I16HE16 + dst, left); + TrueMotion16(I16TM16 + dst, left, top); +} + +// Left samples are top[-5 .. -2], top_left is top[-1], top are +// located at top[0..3], and top right is top[4..7] +static void Intra4Preds(uint8_t* dst, const uint8_t* top) { + DC4(I4DC4 + dst, top); + TM4(I4TM4 + dst, top); + VE4(I4VE4 + dst, top); + HE4(I4HE4 + dst, top); + RD4(I4RD4 + dst, top); + VR4(I4VR4 + dst, top); + LD4(I4LD4 + dst, top); + VL4(I4VL4 + dst, top); + HD4(I4HD4 + dst, top); + HU4(I4HU4 + dst, top); +} + +//------------------------------------------------------------------------------ +// Metric + +#if !defined(WORK_AROUND_GCC) + +#define GET_SSE_INNER(A) \ + "lw %[temp0], " #A "(%[a]) \n\t" \ + "lw %[temp1], " #A "(%[b]) \n\t" \ + "preceu.ph.qbr %[temp2], %[temp0] \n\t" \ + "preceu.ph.qbl %[temp0], %[temp0] \n\t" \ + "preceu.ph.qbr %[temp3], %[temp1] \n\t" \ + "preceu.ph.qbl %[temp1], %[temp1] \n\t" \ + "subq.ph %[temp2], %[temp2], %[temp3] \n\t" \ + "subq.ph %[temp0], %[temp0], %[temp1] \n\t" \ + "dpa.w.ph $ac0, %[temp2], %[temp2] \n\t" \ + "dpa.w.ph $ac0, %[temp0], %[temp0] \n\t" + +#define GET_SSE(A, B, C, D) \ + GET_SSE_INNER(A) \ + GET_SSE_INNER(B) \ + GET_SSE_INNER(C) \ + GET_SSE_INNER(D) + +static int SSE16x16(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3; + __asm__ volatile ( + "mult $zero, $zero \n\t" + GET_SSE( 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS) + GET_SSE( 1 * BPS, 4 + 1 * BPS, 8 + 1 * BPS, 12 + 1 * BPS) + GET_SSE( 2 * BPS, 4 + 2 * BPS, 8 + 2 * BPS, 12 + 2 * BPS) + GET_SSE( 3 * BPS, 4 + 3 * BPS, 8 + 3 * BPS, 12 + 3 * BPS) + GET_SSE( 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS) + GET_SSE( 5 * BPS, 4 + 5 * BPS, 8 + 5 * BPS, 12 + 5 * BPS) + GET_SSE( 6 * BPS, 4 + 6 * BPS, 8 + 6 * BPS, 12 + 6 * BPS) + GET_SSE( 7 * BPS, 4 + 7 * BPS, 8 + 7 * BPS, 12 + 7 * BPS) + GET_SSE( 8 * BPS, 4 + 8 * BPS, 8 + 8 * BPS, 12 + 8 * BPS) + GET_SSE( 9 * BPS, 4 + 9 * BPS, 8 + 9 * BPS, 12 + 9 * BPS) + GET_SSE(10 * BPS, 4 + 10 * BPS, 8 + 10 * BPS, 12 + 10 * BPS) + GET_SSE(11 * BPS, 4 + 11 * BPS, 8 + 11 * BPS, 12 + 11 * BPS) + GET_SSE(12 * BPS, 4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS) + GET_SSE(13 * BPS, 4 + 13 * BPS, 8 + 13 * BPS, 12 + 13 * BPS) + GET_SSE(14 * BPS, 4 + 14 * BPS, 8 + 14 * BPS, 12 + 14 * BPS) + GET_SSE(15 * BPS, 4 + 15 * BPS, 8 + 15 * BPS, 12 + 15 * BPS) + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE16x8(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3; + __asm__ volatile ( + "mult $zero, $zero \n\t" + GET_SSE( 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS) + GET_SSE( 1 * BPS, 4 + 1 * BPS, 8 + 1 * BPS, 12 + 1 * BPS) + GET_SSE( 2 * BPS, 4 + 2 * BPS, 8 + 2 * BPS, 12 + 2 * BPS) + GET_SSE( 3 * BPS, 4 + 3 * BPS, 8 + 3 * BPS, 12 + 3 * BPS) + GET_SSE( 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS) + GET_SSE( 5 * BPS, 4 + 5 * BPS, 8 + 5 * BPS, 12 + 5 * BPS) + GET_SSE( 6 * BPS, 4 + 6 * BPS, 8 + 6 * BPS, 12 + 6 * BPS) + GET_SSE( 7 * BPS, 4 + 7 * BPS, 8 + 7 * BPS, 12 + 7 * BPS) + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE8x8(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3; + __asm__ volatile ( + "mult $zero, $zero \n\t" + GET_SSE(0 * BPS, 4 + 0 * BPS, 1 * BPS, 4 + 1 * BPS) + GET_SSE(2 * BPS, 4 + 2 * BPS, 3 * BPS, 4 + 3 * BPS) + GET_SSE(4 * BPS, 4 + 4 * BPS, 5 * BPS, 4 + 5 * BPS) + GET_SSE(6 * BPS, 4 + 6 * BPS, 7 * BPS, 4 + 7 * BPS) + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE4x4(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3; + __asm__ volatile ( + "mult $zero, $zero \n\t" + GET_SSE(0 * BPS, 1 * BPS, 2 * BPS, 3 * BPS) + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +#undef GET_SSE +#undef GET_SSE_INNER + +#endif // !WORK_AROUND_GCC + +#undef FILL_8_OR_16 +#undef FILL_PART +#undef OUTPUT_EARLY_CLOBBER_REGS_17 +#undef MUL_HALF +#undef ABS_X8 +#undef ADD_SUB_HALVES_X4 + +//------------------------------------------------------------------------------ +// Quantization +// + +// macro for one pass through for loop in QuantizeBlock reading 2 values at time +// QUANTDIV macro inlined +// J - offset in bytes (kZigzag[n] * 2) +// K - offset in bytes (kZigzag[n] * 4) +// N - offset in bytes (n * 2) +// N1 - offset in bytes ((n + 1) * 2) +#define QUANTIZE_ONE(J, K, N, N1) \ + "ulw %[temp1], " #J "(%[ppin]) \n\t" \ + "ulw %[temp2], " #J "(%[ppsharpen]) \n\t" \ + "lhu %[temp3], " #K "(%[ppzthresh]) \n\t" \ + "lhu %[temp6], " #K "+4(%[ppzthresh]) \n\t" \ + "absq_s.ph %[temp4], %[temp1] \n\t" \ + "ins %[temp3], %[temp6], 16, 16 \n\t" \ + "addu.ph %[coeff], %[temp4], %[temp2] \n\t" \ + "shra.ph %[sign], %[temp1], 15 \n\t" \ + "li %[level], 0x10001 \n\t" \ + "cmp.lt.ph %[temp3], %[coeff] \n\t" \ + "lhu %[temp1], " #J "(%[ppiq]) \n\t" \ + "pick.ph %[temp5], %[level], $0 \n\t" \ + "lw %[temp2], " #K "(%[ppbias]) \n\t" \ + "beqz %[temp5], 0f \n\t" \ + "lhu %[temp3], " #J "(%[ppq]) \n\t" \ + "beq %[temp5], %[level], 1f \n\t" \ + "andi %[temp5], %[temp5], 0x1 \n\t" \ + "andi %[temp4], %[coeff], 0xffff \n\t" \ + "beqz %[temp5], 2f \n\t" \ + "mul %[level], %[temp4], %[temp1] \n\t" \ + "sh $0, " #J "+2(%[ppin]) \n\t" \ + "sh $0, " #N1 "(%[pout]) \n\t" \ + "addu %[level], %[level], %[temp2] \n\t" \ + "sra %[level], %[level], 17 \n\t" \ + "slt %[temp4], %[max_level], %[level] \n\t" \ + "movn %[level], %[max_level], %[temp4] \n\t" \ + "andi %[temp6], %[sign], 0xffff \n\t" \ + "xor %[level], %[level], %[temp6] \n\t" \ + "subu %[level], %[level], %[temp6] \n\t" \ + "mul %[temp5], %[level], %[temp3] \n\t" \ + "or %[ret], %[ret], %[level] \n\t" \ + "sh %[level], " #N "(%[pout]) \n\t" \ + "sh %[temp5], " #J "(%[ppin]) \n\t" \ + "j 3f \n\t" \ +"2: \n\t" \ + "lhu %[temp1], " #J "+2(%[ppiq]) \n\t" \ + "srl %[temp5], %[coeff], 16 \n\t" \ + "mul %[level], %[temp5], %[temp1] \n\t" \ + "lw %[temp2], " #K "+4(%[ppbias]) \n\t" \ + "lhu %[temp3], " #J "+2(%[ppq]) \n\t" \ + "addu %[level], %[level], %[temp2] \n\t" \ + "sra %[level], %[level], 17 \n\t" \ + "srl %[temp6], %[sign], 16 \n\t" \ + "slt %[temp4], %[max_level], %[level] \n\t" \ + "movn %[level], %[max_level], %[temp4] \n\t" \ + "xor %[level], %[level], %[temp6] \n\t" \ + "subu %[level], %[level], %[temp6] \n\t" \ + "mul %[temp5], %[level], %[temp3] \n\t" \ + "sh $0, " #J "(%[ppin]) \n\t" \ + "sh $0, " #N "(%[pout]) \n\t" \ + "or %[ret], %[ret], %[level] \n\t" \ + "sh %[temp5], " #J "+2(%[ppin]) \n\t" \ + "sh %[level], " #N1 "(%[pout]) \n\t" \ + "j 3f \n\t" \ +"1: \n\t" \ + "lhu %[temp1], " #J "(%[ppiq]) \n\t" \ + "lw %[temp2], " #K "(%[ppbias]) \n\t" \ + "ulw %[temp3], " #J "(%[ppq]) \n\t" \ + "andi %[temp5], %[coeff], 0xffff \n\t" \ + "srl %[temp0], %[coeff], 16 \n\t" \ + "lhu %[temp6], " #J "+2(%[ppiq]) \n\t" \ + "lw %[coeff], " #K "+4(%[ppbias]) \n\t" \ + "mul %[level], %[temp5], %[temp1] \n\t" \ + "mul %[temp4], %[temp0], %[temp6] \n\t" \ + "addu %[level], %[level], %[temp2] \n\t" \ + "addu %[temp4], %[temp4], %[coeff] \n\t" \ + "precrq.ph.w %[level], %[temp4], %[level] \n\t" \ + "shra.ph %[level], %[level], 1 \n\t" \ + "cmp.lt.ph %[max_level1],%[level] \n\t" \ + "pick.ph %[level], %[max_level], %[level] \n\t" \ + "xor %[level], %[level], %[sign] \n\t" \ + "subu.ph %[level], %[level], %[sign] \n\t" \ + "mul.ph %[temp3], %[level], %[temp3] \n\t" \ + "or %[ret], %[ret], %[level] \n\t" \ + "sh %[level], " #N "(%[pout]) \n\t" \ + "srl %[level], %[level], 16 \n\t" \ + "sh %[level], " #N1 "(%[pout]) \n\t" \ + "usw %[temp3], " #J "(%[ppin]) \n\t" \ + "j 3f \n\t" \ +"0: \n\t" \ + "sh $0, " #N "(%[pout]) \n\t" \ + "sh $0, " #N1 "(%[pout]) \n\t" \ + "usw $0, " #J "(%[ppin]) \n\t" \ +"3: \n\t" + +static int QuantizeBlock(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + int temp0, temp1, temp2, temp3, temp4, temp5,temp6; + int sign, coeff, level; + int max_level = MAX_LEVEL; + int max_level1 = max_level << 16 | max_level; + int ret = 0; + + int16_t* ppin = &in[0]; + int16_t* pout = &out[0]; + const uint16_t* ppsharpen = &mtx->sharpen_[0]; + const uint32_t* ppzthresh = &mtx->zthresh_[0]; + const uint16_t* ppq = &mtx->q_[0]; + const uint16_t* ppiq = &mtx->iq_[0]; + const uint32_t* ppbias = &mtx->bias_[0]; + + __asm__ volatile ( + QUANTIZE_ONE( 0, 0, 0, 2) + QUANTIZE_ONE( 4, 8, 10, 12) + QUANTIZE_ONE( 8, 16, 4, 8) + QUANTIZE_ONE(12, 24, 14, 24) + QUANTIZE_ONE(16, 32, 6, 16) + QUANTIZE_ONE(20, 40, 22, 26) + QUANTIZE_ONE(24, 48, 18, 20) + QUANTIZE_ONE(28, 56, 28, 30) + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [sign]"=&r"(sign), [coeff]"=&r"(coeff), + [level]"=&r"(level), [temp6]"=&r"(temp6), [ret]"+&r"(ret) + : [ppin]"r"(ppin), [pout]"r"(pout), [max_level1]"r"(max_level1), + [ppiq]"r"(ppiq), [max_level]"r"(max_level), + [ppbias]"r"(ppbias), [ppzthresh]"r"(ppzthresh), + [ppsharpen]"r"(ppsharpen), [ppq]"r"(ppq) + : "memory", "hi", "lo" + ); + + return (ret != 0); +} + +static int Quantize2Blocks(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = QuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= QuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} + +#undef QUANTIZE_ONE + +// macro for one horizontal pass in FTransformWHT +// temp0..temp7 holds tmp[0]..tmp[15] +// A, B, C, D - offset in bytes to load from in buffer +// TEMP0, TEMP1 - registers for corresponding tmp elements +#define HORIZONTAL_PASS_WHT(A, B, C, D, TEMP0, TEMP1) \ + "lh %[" #TEMP0 "], " #A "(%[in]) \n\t" \ + "lh %[" #TEMP1 "], " #B "(%[in]) \n\t" \ + "lh %[temp8], " #C "(%[in]) \n\t" \ + "lh %[temp9], " #D "(%[in]) \n\t" \ + "ins %[" #TEMP1 "], %[" #TEMP0 "], 16, 16 \n\t" \ + "ins %[temp9], %[temp8], 16, 16 \n\t" \ + "subq.ph %[temp8], %[" #TEMP1 "], %[temp9] \n\t" \ + "addq.ph %[temp9], %[" #TEMP1 "], %[temp9] \n\t" \ + "precrq.ph.w %[" #TEMP0 "], %[temp8], %[temp9] \n\t" \ + "append %[temp8], %[temp9], 16 \n\t" \ + "subq.ph %[" #TEMP1 "], %[" #TEMP0 "], %[temp8] \n\t" \ + "addq.ph %[" #TEMP0 "], %[" #TEMP0 "], %[temp8] \n\t" \ + "rotr %[" #TEMP1 "], %[" #TEMP1 "], 16 \n\t" + +// macro for one vertical pass in FTransformWHT +// temp0..temp7 holds tmp[0]..tmp[15] +// A, B, C, D - offsets in bytes to store to out buffer +// TEMP0, TEMP2, TEMP4 and TEMP6 - registers for corresponding tmp elements +#define VERTICAL_PASS_WHT(A, B, C, D, TEMP0, TEMP2, TEMP4, TEMP6) \ + "addq.ph %[temp8], %[" #TEMP0 "], %[" #TEMP4 "] \n\t" \ + "addq.ph %[temp9], %[" #TEMP2 "], %[" #TEMP6 "] \n\t" \ + "subq.ph %[" #TEMP2 "], %[" #TEMP2 "], %[" #TEMP6 "] \n\t" \ + "subq.ph %[" #TEMP6 "], %[" #TEMP0 "], %[" #TEMP4 "] \n\t" \ + "addqh.ph %[" #TEMP0 "], %[temp8], %[temp9] \n\t" \ + "subqh.ph %[" #TEMP4 "], %[" #TEMP6 "], %[" #TEMP2 "] \n\t" \ + "addqh.ph %[" #TEMP2 "], %[" #TEMP2 "], %[" #TEMP6 "] \n\t" \ + "subqh.ph %[" #TEMP6 "], %[temp8], %[temp9] \n\t" \ + "usw %[" #TEMP0 "], " #A "(%[out]) \n\t" \ + "usw %[" #TEMP2 "], " #B "(%[out]) \n\t" \ + "usw %[" #TEMP4 "], " #C "(%[out]) \n\t" \ + "usw %[" #TEMP6 "], " #D "(%[out]) \n\t" + +static void FTransformWHT(const int16_t* in, int16_t* out) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + + __asm__ volatile ( + HORIZONTAL_PASS_WHT( 0, 32, 64, 96, temp0, temp1) + HORIZONTAL_PASS_WHT(128, 160, 192, 224, temp2, temp3) + HORIZONTAL_PASS_WHT(256, 288, 320, 352, temp4, temp5) + HORIZONTAL_PASS_WHT(384, 416, 448, 480, temp6, temp7) + VERTICAL_PASS_WHT(0, 8, 16, 24, temp0, temp2, temp4, temp6) + VERTICAL_PASS_WHT(4, 12, 20, 28, temp1, temp3, temp5, temp7) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [in]"r"(in), [out]"r"(out) + : "memory" + ); +} + +#undef VERTICAL_PASS_WHT +#undef HORIZONTAL_PASS_WHT + +// macro for converting coefficients to bin +// convert 8 coeffs at time +// A, B, C, D - offsets in bytes to load from out buffer +#define CONVERT_COEFFS_TO_BIN(A, B, C, D) \ + "ulw %[temp0], " #A "(%[out]) \n\t" \ + "ulw %[temp1], " #B "(%[out]) \n\t" \ + "ulw %[temp2], " #C "(%[out]) \n\t" \ + "ulw %[temp3], " #D "(%[out]) \n\t" \ + "absq_s.ph %[temp0], %[temp0] \n\t" \ + "absq_s.ph %[temp1], %[temp1] \n\t" \ + "absq_s.ph %[temp2], %[temp2] \n\t" \ + "absq_s.ph %[temp3], %[temp3] \n\t" \ + /* TODO(skal): add rounding ? shra_r.ph : shra.ph */ \ + /* for following 4 instructions */ \ + "shra.ph %[temp0], %[temp0], 3 \n\t" \ + "shra.ph %[temp1], %[temp1], 3 \n\t" \ + "shra.ph %[temp2], %[temp2], 3 \n\t" \ + "shra.ph %[temp3], %[temp3], 3 \n\t" \ + "shll_s.ph %[temp0], %[temp0], 10 \n\t" \ + "shll_s.ph %[temp1], %[temp1], 10 \n\t" \ + "shll_s.ph %[temp2], %[temp2], 10 \n\t" \ + "shll_s.ph %[temp3], %[temp3], 10 \n\t" \ + "shrl.ph %[temp0], %[temp0], 10 \n\t" \ + "shrl.ph %[temp1], %[temp1], 10 \n\t" \ + "shrl.ph %[temp2], %[temp2], 10 \n\t" \ + "shrl.ph %[temp3], %[temp3], 10 \n\t" \ + "shll.ph %[temp0], %[temp0], 2 \n\t" \ + "shll.ph %[temp1], %[temp1], 2 \n\t" \ + "shll.ph %[temp2], %[temp2], 2 \n\t" \ + "shll.ph %[temp3], %[temp3], 2 \n\t" \ + "ext %[temp4], %[temp0], 0, 16 \n\t" \ + "ext %[temp0], %[temp0], 16, 16 \n\t" \ + "addu %[temp4], %[temp4], %[dist] \n\t" \ + "addu %[temp0], %[temp0], %[dist] \n\t" \ + "ext %[temp5], %[temp1], 0, 16 \n\t" \ + "lw %[temp8], 0(%[temp4]) \n\t" \ + "ext %[temp1], %[temp1], 16, 16 \n\t" \ + "addu %[temp5], %[temp5], %[dist] \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp4]) \n\t" \ + "lw %[temp8], 0(%[temp0]) \n\t" \ + "addu %[temp1], %[temp1], %[dist] \n\t" \ + "ext %[temp6], %[temp2], 0, 16 \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp0]) \n\t" \ + "lw %[temp8], 0(%[temp5]) \n\t" \ + "ext %[temp2], %[temp2], 16, 16 \n\t" \ + "addu %[temp6], %[temp6], %[dist] \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp5]) \n\t" \ + "lw %[temp8], 0(%[temp1]) \n\t" \ + "addu %[temp2], %[temp2], %[dist] \n\t" \ + "ext %[temp7], %[temp3], 0, 16 \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp1]) \n\t" \ + "lw %[temp8], 0(%[temp6]) \n\t" \ + "ext %[temp3], %[temp3], 16, 16 \n\t" \ + "addu %[temp7], %[temp7], %[dist] \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp6]) \n\t" \ + "lw %[temp8], 0(%[temp2]) \n\t" \ + "addu %[temp3], %[temp3], %[dist] \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp2]) \n\t" \ + "lw %[temp8], 0(%[temp7]) \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp7]) \n\t" \ + "lw %[temp8], 0(%[temp3]) \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp3]) \n\t" + +static void CollectHistogram(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + const int max_coeff = (MAX_COEFF_THRESH << 16) + MAX_COEFF_THRESH; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + + VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + + // Convert coefficients to bin. + __asm__ volatile ( + CONVERT_COEFFS_TO_BIN( 0, 4, 8, 12) + CONVERT_COEFFS_TO_BIN(16, 20, 24, 28) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8) + : [dist]"r"(distribution), [out]"r"(out), [max_coeff]"r"(max_coeff) + : "memory" + ); + } + VP8SetHistogramData(distribution, histo); +} + +#undef CONVERT_COEFFS_TO_BIN + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitMIPSdspR2(void) { + VP8FTransform = FTransform; + VP8ITransform = ITransform; + VP8TDisto4x4 = Disto4x4; + VP8TDisto16x16 = Disto16x16; + VP8EncPredLuma16 = Intra16Preds; + VP8EncPredChroma8 = IntraChromaPreds; + VP8EncPredLuma4 = Intra4Preds; +#if !defined(WORK_AROUND_GCC) + VP8SSE16x16 = SSE16x16; + VP8SSE8x8 = SSE8x8; + VP8SSE16x8 = SSE16x8; + VP8SSE4x4 = SSE4x4; +#endif + VP8EncQuantizeBlock = QuantizeBlock; + VP8EncQuantize2Blocks = Quantize2Blocks; + VP8FTransformWHT = FTransformWHT; + VP8CollectHistogram = CollectHistogram; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8EncDspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/enc_neon.c b/drivers/webp/dsp/enc_neon.c new file mode 100644 index 0000000000..c2aef58e70 --- /dev/null +++ b/drivers/webp/dsp/enc_neon.c @@ -0,0 +1,934 @@ +// Copyright 2012 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// ARM NEON version of speed-critical encoding functions. +// +// adapted from libvpx (http://www.webmproject.org/code/) + +#include "./dsp.h" + +#if defined(WEBP_USE_NEON) + +#include <assert.h> + +#include "./neon.h" +#include "../enc/vp8enci.h" + +//------------------------------------------------------------------------------ +// Transforms (Paragraph 14.4) + +// Inverse transform. +// This code is pretty much the same as TransformOne in the dec_neon.c, except +// for subtraction to *ref. See the comments there for algorithmic explanations. + +static const int16_t kC1 = 20091; +static const int16_t kC2 = 17734; // half of kC2, actually. See comment above. + +// This code works but is *slower* than the inlined-asm version below +// (with gcc-4.6). So we disable it for now. Later, it'll be conditional to +// WEBP_USE_INTRINSICS define. +// With gcc-4.8, it's a little faster speed than inlined-assembly. +#if defined(WEBP_USE_INTRINSICS) + +// Treats 'v' as an uint8x8_t and zero extends to an int16x8_t. +static WEBP_INLINE int16x8_t ConvertU8ToS16(uint32x2_t v) { + return vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(v))); +} + +// Performs unsigned 8b saturation on 'dst01' and 'dst23' storing the result +// to the corresponding rows of 'dst'. +static WEBP_INLINE void SaturateAndStore4x4(uint8_t* const dst, + const int16x8_t dst01, + const int16x8_t dst23) { + // Unsigned saturate to 8b. + const uint8x8_t dst01_u8 = vqmovun_s16(dst01); + const uint8x8_t dst23_u8 = vqmovun_s16(dst23); + + // Store the results. + vst1_lane_u32((uint32_t*)(dst + 0 * BPS), vreinterpret_u32_u8(dst01_u8), 0); + vst1_lane_u32((uint32_t*)(dst + 1 * BPS), vreinterpret_u32_u8(dst01_u8), 1); + vst1_lane_u32((uint32_t*)(dst + 2 * BPS), vreinterpret_u32_u8(dst23_u8), 0); + vst1_lane_u32((uint32_t*)(dst + 3 * BPS), vreinterpret_u32_u8(dst23_u8), 1); +} + +static WEBP_INLINE void Add4x4(const int16x8_t row01, const int16x8_t row23, + const uint8_t* const ref, uint8_t* const dst) { + uint32x2_t dst01 = vdup_n_u32(0); + uint32x2_t dst23 = vdup_n_u32(0); + + // Load the source pixels. + dst01 = vld1_lane_u32((uint32_t*)(ref + 0 * BPS), dst01, 0); + dst23 = vld1_lane_u32((uint32_t*)(ref + 2 * BPS), dst23, 0); + dst01 = vld1_lane_u32((uint32_t*)(ref + 1 * BPS), dst01, 1); + dst23 = vld1_lane_u32((uint32_t*)(ref + 3 * BPS), dst23, 1); + + { + // Convert to 16b. + const int16x8_t dst01_s16 = ConvertU8ToS16(dst01); + const int16x8_t dst23_s16 = ConvertU8ToS16(dst23); + + // Descale with rounding. + const int16x8_t out01 = vrsraq_n_s16(dst01_s16, row01, 3); + const int16x8_t out23 = vrsraq_n_s16(dst23_s16, row23, 3); + // Add the inverse transform. + SaturateAndStore4x4(dst, out01, out23); + } +} + +static WEBP_INLINE void Transpose8x2(const int16x8_t in0, const int16x8_t in1, + int16x8x2_t* const out) { + // a0 a1 a2 a3 | b0 b1 b2 b3 => a0 b0 c0 d0 | a1 b1 c1 d1 + // c0 c1 c2 c3 | d0 d1 d2 d3 a2 b2 c2 d2 | a3 b3 c3 d3 + const int16x8x2_t tmp0 = vzipq_s16(in0, in1); // a0 c0 a1 c1 a2 c2 ... + // b0 d0 b1 d1 b2 d2 ... + *out = vzipq_s16(tmp0.val[0], tmp0.val[1]); +} + +static WEBP_INLINE void TransformPass(int16x8x2_t* const rows) { + // {rows} = in0 | in4 + // in8 | in12 + // B1 = in4 | in12 + const int16x8_t B1 = + vcombine_s16(vget_high_s16(rows->val[0]), vget_high_s16(rows->val[1])); + // C0 = kC1 * in4 | kC1 * in12 + // C1 = kC2 * in4 | kC2 * in12 + const int16x8_t C0 = vsraq_n_s16(B1, vqdmulhq_n_s16(B1, kC1), 1); + const int16x8_t C1 = vqdmulhq_n_s16(B1, kC2); + const int16x4_t a = vqadd_s16(vget_low_s16(rows->val[0]), + vget_low_s16(rows->val[1])); // in0 + in8 + const int16x4_t b = vqsub_s16(vget_low_s16(rows->val[0]), + vget_low_s16(rows->val[1])); // in0 - in8 + // c = kC2 * in4 - kC1 * in12 + // d = kC1 * in4 + kC2 * in12 + const int16x4_t c = vqsub_s16(vget_low_s16(C1), vget_high_s16(C0)); + const int16x4_t d = vqadd_s16(vget_low_s16(C0), vget_high_s16(C1)); + const int16x8_t D0 = vcombine_s16(a, b); // D0 = a | b + const int16x8_t D1 = vcombine_s16(d, c); // D1 = d | c + const int16x8_t E0 = vqaddq_s16(D0, D1); // a+d | b+c + const int16x8_t E_tmp = vqsubq_s16(D0, D1); // a-d | b-c + const int16x8_t E1 = vcombine_s16(vget_high_s16(E_tmp), vget_low_s16(E_tmp)); + Transpose8x2(E0, E1, rows); +} + +static void ITransformOne(const uint8_t* ref, + const int16_t* in, uint8_t* dst) { + int16x8x2_t rows; + INIT_VECTOR2(rows, vld1q_s16(in + 0), vld1q_s16(in + 8)); + TransformPass(&rows); + TransformPass(&rows); + Add4x4(rows.val[0], rows.val[1], ref, dst); +} + +#else + +static void ITransformOne(const uint8_t* ref, + const int16_t* in, uint8_t* dst) { + const int kBPS = BPS; + const int16_t kC1C2[] = { kC1, kC2, 0, 0 }; + + __asm__ volatile ( + "vld1.16 {q1, q2}, [%[in]] \n" + "vld1.16 {d0}, [%[kC1C2]] \n" + + // d2: in[0] + // d3: in[8] + // d4: in[4] + // d5: in[12] + "vswp d3, d4 \n" + + // q8 = {in[4], in[12]} * kC1 * 2 >> 16 + // q9 = {in[4], in[12]} * kC2 >> 16 + "vqdmulh.s16 q8, q2, d0[0] \n" + "vqdmulh.s16 q9, q2, d0[1] \n" + + // d22 = a = in[0] + in[8] + // d23 = b = in[0] - in[8] + "vqadd.s16 d22, d2, d3 \n" + "vqsub.s16 d23, d2, d3 \n" + + // q8 = in[4]/[12] * kC1 >> 16 + "vshr.s16 q8, q8, #1 \n" + + // Add {in[4], in[12]} back after the multiplication. + "vqadd.s16 q8, q2, q8 \n" + + // d20 = c = in[4]*kC2 - in[12]*kC1 + // d21 = d = in[4]*kC1 + in[12]*kC2 + "vqsub.s16 d20, d18, d17 \n" + "vqadd.s16 d21, d19, d16 \n" + + // d2 = tmp[0] = a + d + // d3 = tmp[1] = b + c + // d4 = tmp[2] = b - c + // d5 = tmp[3] = a - d + "vqadd.s16 d2, d22, d21 \n" + "vqadd.s16 d3, d23, d20 \n" + "vqsub.s16 d4, d23, d20 \n" + "vqsub.s16 d5, d22, d21 \n" + + "vzip.16 q1, q2 \n" + "vzip.16 q1, q2 \n" + + "vswp d3, d4 \n" + + // q8 = {tmp[4], tmp[12]} * kC1 * 2 >> 16 + // q9 = {tmp[4], tmp[12]} * kC2 >> 16 + "vqdmulh.s16 q8, q2, d0[0] \n" + "vqdmulh.s16 q9, q2, d0[1] \n" + + // d22 = a = tmp[0] + tmp[8] + // d23 = b = tmp[0] - tmp[8] + "vqadd.s16 d22, d2, d3 \n" + "vqsub.s16 d23, d2, d3 \n" + + "vshr.s16 q8, q8, #1 \n" + "vqadd.s16 q8, q2, q8 \n" + + // d20 = c = in[4]*kC2 - in[12]*kC1 + // d21 = d = in[4]*kC1 + in[12]*kC2 + "vqsub.s16 d20, d18, d17 \n" + "vqadd.s16 d21, d19, d16 \n" + + // d2 = tmp[0] = a + d + // d3 = tmp[1] = b + c + // d4 = tmp[2] = b - c + // d5 = tmp[3] = a - d + "vqadd.s16 d2, d22, d21 \n" + "vqadd.s16 d3, d23, d20 \n" + "vqsub.s16 d4, d23, d20 \n" + "vqsub.s16 d5, d22, d21 \n" + + "vld1.32 d6[0], [%[ref]], %[kBPS] \n" + "vld1.32 d6[1], [%[ref]], %[kBPS] \n" + "vld1.32 d7[0], [%[ref]], %[kBPS] \n" + "vld1.32 d7[1], [%[ref]], %[kBPS] \n" + + "sub %[ref], %[ref], %[kBPS], lsl #2 \n" + + // (val) + 4 >> 3 + "vrshr.s16 d2, d2, #3 \n" + "vrshr.s16 d3, d3, #3 \n" + "vrshr.s16 d4, d4, #3 \n" + "vrshr.s16 d5, d5, #3 \n" + + "vzip.16 q1, q2 \n" + "vzip.16 q1, q2 \n" + + // Must accumulate before saturating + "vmovl.u8 q8, d6 \n" + "vmovl.u8 q9, d7 \n" + + "vqadd.s16 q1, q1, q8 \n" + "vqadd.s16 q2, q2, q9 \n" + + "vqmovun.s16 d0, q1 \n" + "vqmovun.s16 d1, q2 \n" + + "vst1.32 d0[0], [%[dst]], %[kBPS] \n" + "vst1.32 d0[1], [%[dst]], %[kBPS] \n" + "vst1.32 d1[0], [%[dst]], %[kBPS] \n" + "vst1.32 d1[1], [%[dst]] \n" + + : [in] "+r"(in), [dst] "+r"(dst) // modified registers + : [kBPS] "r"(kBPS), [kC1C2] "r"(kC1C2), [ref] "r"(ref) // constants + : "memory", "q0", "q1", "q2", "q8", "q9", "q10", "q11" // clobbered + ); +} + +#endif // WEBP_USE_INTRINSICS + +static void ITransform(const uint8_t* ref, + const int16_t* in, uint8_t* dst, int do_two) { + ITransformOne(ref, in, dst); + if (do_two) { + ITransformOne(ref + 4, in + 16, dst + 4); + } +} + +// Load all 4x4 pixels into a single uint8x16_t variable. +static uint8x16_t Load4x4(const uint8_t* src) { + uint32x4_t out = vdupq_n_u32(0); + out = vld1q_lane_u32((const uint32_t*)(src + 0 * BPS), out, 0); + out = vld1q_lane_u32((const uint32_t*)(src + 1 * BPS), out, 1); + out = vld1q_lane_u32((const uint32_t*)(src + 2 * BPS), out, 2); + out = vld1q_lane_u32((const uint32_t*)(src + 3 * BPS), out, 3); + return vreinterpretq_u8_u32(out); +} + +// Forward transform. + +#if defined(WEBP_USE_INTRINSICS) + +static WEBP_INLINE void Transpose4x4_S16(const int16x4_t A, const int16x4_t B, + const int16x4_t C, const int16x4_t D, + int16x8_t* const out01, + int16x8_t* const out32) { + const int16x4x2_t AB = vtrn_s16(A, B); + const int16x4x2_t CD = vtrn_s16(C, D); + const int32x2x2_t tmp02 = vtrn_s32(vreinterpret_s32_s16(AB.val[0]), + vreinterpret_s32_s16(CD.val[0])); + const int32x2x2_t tmp13 = vtrn_s32(vreinterpret_s32_s16(AB.val[1]), + vreinterpret_s32_s16(CD.val[1])); + *out01 = vreinterpretq_s16_s64( + vcombine_s64(vreinterpret_s64_s32(tmp02.val[0]), + vreinterpret_s64_s32(tmp13.val[0]))); + *out32 = vreinterpretq_s16_s64( + vcombine_s64(vreinterpret_s64_s32(tmp13.val[1]), + vreinterpret_s64_s32(tmp02.val[1]))); +} + +static WEBP_INLINE int16x8_t DiffU8ToS16(const uint8x8_t a, + const uint8x8_t b) { + return vreinterpretq_s16_u16(vsubl_u8(a, b)); +} + +static void FTransform(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + int16x8_t d0d1, d3d2; // working 4x4 int16 variables + { + const uint8x16_t S0 = Load4x4(src); + const uint8x16_t R0 = Load4x4(ref); + const int16x8_t D0D1 = DiffU8ToS16(vget_low_u8(S0), vget_low_u8(R0)); + const int16x8_t D2D3 = DiffU8ToS16(vget_high_u8(S0), vget_high_u8(R0)); + const int16x4_t D0 = vget_low_s16(D0D1); + const int16x4_t D1 = vget_high_s16(D0D1); + const int16x4_t D2 = vget_low_s16(D2D3); + const int16x4_t D3 = vget_high_s16(D2D3); + Transpose4x4_S16(D0, D1, D2, D3, &d0d1, &d3d2); + } + { // 1rst pass + const int32x4_t kCst937 = vdupq_n_s32(937); + const int32x4_t kCst1812 = vdupq_n_s32(1812); + const int16x8_t a0a1 = vaddq_s16(d0d1, d3d2); // d0+d3 | d1+d2 (=a0|a1) + const int16x8_t a3a2 = vsubq_s16(d0d1, d3d2); // d0-d3 | d1-d2 (=a3|a2) + const int16x8_t a0a1_2 = vshlq_n_s16(a0a1, 3); + const int16x4_t tmp0 = vadd_s16(vget_low_s16(a0a1_2), + vget_high_s16(a0a1_2)); + const int16x4_t tmp2 = vsub_s16(vget_low_s16(a0a1_2), + vget_high_s16(a0a1_2)); + const int32x4_t a3_2217 = vmull_n_s16(vget_low_s16(a3a2), 2217); + const int32x4_t a2_2217 = vmull_n_s16(vget_high_s16(a3a2), 2217); + const int32x4_t a2_p_a3 = vmlal_n_s16(a2_2217, vget_low_s16(a3a2), 5352); + const int32x4_t a3_m_a2 = vmlsl_n_s16(a3_2217, vget_high_s16(a3a2), 5352); + const int16x4_t tmp1 = vshrn_n_s32(vaddq_s32(a2_p_a3, kCst1812), 9); + const int16x4_t tmp3 = vshrn_n_s32(vaddq_s32(a3_m_a2, kCst937), 9); + Transpose4x4_S16(tmp0, tmp1, tmp2, tmp3, &d0d1, &d3d2); + } + { // 2nd pass + // the (1<<16) addition is for the replacement: a3!=0 <-> 1-(a3==0) + const int32x4_t kCst12000 = vdupq_n_s32(12000 + (1 << 16)); + const int32x4_t kCst51000 = vdupq_n_s32(51000); + const int16x8_t a0a1 = vaddq_s16(d0d1, d3d2); // d0+d3 | d1+d2 (=a0|a1) + const int16x8_t a3a2 = vsubq_s16(d0d1, d3d2); // d0-d3 | d1-d2 (=a3|a2) + const int16x4_t a0_k7 = vadd_s16(vget_low_s16(a0a1), vdup_n_s16(7)); + const int16x4_t out0 = vshr_n_s16(vadd_s16(a0_k7, vget_high_s16(a0a1)), 4); + const int16x4_t out2 = vshr_n_s16(vsub_s16(a0_k7, vget_high_s16(a0a1)), 4); + const int32x4_t a3_2217 = vmull_n_s16(vget_low_s16(a3a2), 2217); + const int32x4_t a2_2217 = vmull_n_s16(vget_high_s16(a3a2), 2217); + const int32x4_t a2_p_a3 = vmlal_n_s16(a2_2217, vget_low_s16(a3a2), 5352); + const int32x4_t a3_m_a2 = vmlsl_n_s16(a3_2217, vget_high_s16(a3a2), 5352); + const int16x4_t tmp1 = vaddhn_s32(a2_p_a3, kCst12000); + const int16x4_t out3 = vaddhn_s32(a3_m_a2, kCst51000); + const int16x4_t a3_eq_0 = + vreinterpret_s16_u16(vceq_s16(vget_low_s16(a3a2), vdup_n_s16(0))); + const int16x4_t out1 = vadd_s16(tmp1, a3_eq_0); + vst1_s16(out + 0, out0); + vst1_s16(out + 4, out1); + vst1_s16(out + 8, out2); + vst1_s16(out + 12, out3); + } +} + +#else + +// adapted from vp8/encoder/arm/neon/shortfdct_neon.asm +static const int16_t kCoeff16[] = { + 5352, 5352, 5352, 5352, 2217, 2217, 2217, 2217 +}; +static const int32_t kCoeff32[] = { + 1812, 1812, 1812, 1812, + 937, 937, 937, 937, + 12000, 12000, 12000, 12000, + 51000, 51000, 51000, 51000 +}; + +static void FTransform(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + const int kBPS = BPS; + const uint8_t* src_ptr = src; + const uint8_t* ref_ptr = ref; + const int16_t* coeff16 = kCoeff16; + const int32_t* coeff32 = kCoeff32; + + __asm__ volatile ( + // load src into q4, q5 in high half + "vld1.8 {d8}, [%[src_ptr]], %[kBPS] \n" + "vld1.8 {d10}, [%[src_ptr]], %[kBPS] \n" + "vld1.8 {d9}, [%[src_ptr]], %[kBPS] \n" + "vld1.8 {d11}, [%[src_ptr]] \n" + + // load ref into q6, q7 in high half + "vld1.8 {d12}, [%[ref_ptr]], %[kBPS] \n" + "vld1.8 {d14}, [%[ref_ptr]], %[kBPS] \n" + "vld1.8 {d13}, [%[ref_ptr]], %[kBPS] \n" + "vld1.8 {d15}, [%[ref_ptr]] \n" + + // Pack the high values in to q4 and q6 + "vtrn.32 q4, q5 \n" + "vtrn.32 q6, q7 \n" + + // d[0-3] = src - ref + "vsubl.u8 q0, d8, d12 \n" + "vsubl.u8 q1, d9, d13 \n" + + // load coeff16 into q8(d16=5352, d17=2217) + "vld1.16 {q8}, [%[coeff16]] \n" + + // load coeff32 high half into q9 = 1812, q10 = 937 + "vld1.32 {q9, q10}, [%[coeff32]]! \n" + + // load coeff32 low half into q11=12000, q12=51000 + "vld1.32 {q11,q12}, [%[coeff32]] \n" + + // part 1 + // Transpose. Register dN is the same as dN in C + "vtrn.32 d0, d2 \n" + "vtrn.32 d1, d3 \n" + "vtrn.16 d0, d1 \n" + "vtrn.16 d2, d3 \n" + + "vadd.s16 d4, d0, d3 \n" // a0 = d0 + d3 + "vadd.s16 d5, d1, d2 \n" // a1 = d1 + d2 + "vsub.s16 d6, d1, d2 \n" // a2 = d1 - d2 + "vsub.s16 d7, d0, d3 \n" // a3 = d0 - d3 + + "vadd.s16 d0, d4, d5 \n" // a0 + a1 + "vshl.s16 d0, d0, #3 \n" // temp[0+i*4] = (a0+a1) << 3 + "vsub.s16 d2, d4, d5 \n" // a0 - a1 + "vshl.s16 d2, d2, #3 \n" // (temp[2+i*4] = (a0-a1) << 3 + + "vmlal.s16 q9, d7, d16 \n" // a3*5352 + 1812 + "vmlal.s16 q10, d7, d17 \n" // a3*2217 + 937 + "vmlal.s16 q9, d6, d17 \n" // a2*2217 + a3*5352 + 1812 + "vmlsl.s16 q10, d6, d16 \n" // a3*2217 + 937 - a2*5352 + + // temp[1+i*4] = (d2*2217 + d3*5352 + 1812) >> 9 + // temp[3+i*4] = (d3*2217 + 937 - d2*5352) >> 9 + "vshrn.s32 d1, q9, #9 \n" + "vshrn.s32 d3, q10, #9 \n" + + // part 2 + // transpose d0=ip[0], d1=ip[4], d2=ip[8], d3=ip[12] + "vtrn.32 d0, d2 \n" + "vtrn.32 d1, d3 \n" + "vtrn.16 d0, d1 \n" + "vtrn.16 d2, d3 \n" + + "vmov.s16 d26, #7 \n" + + "vadd.s16 d4, d0, d3 \n" // a1 = ip[0] + ip[12] + "vadd.s16 d5, d1, d2 \n" // b1 = ip[4] + ip[8] + "vsub.s16 d6, d1, d2 \n" // c1 = ip[4] - ip[8] + "vadd.s16 d4, d4, d26 \n" // a1 + 7 + "vsub.s16 d7, d0, d3 \n" // d1 = ip[0] - ip[12] + + "vadd.s16 d0, d4, d5 \n" // op[0] = a1 + b1 + 7 + "vsub.s16 d2, d4, d5 \n" // op[8] = a1 - b1 + 7 + + "vmlal.s16 q11, d7, d16 \n" // d1*5352 + 12000 + "vmlal.s16 q12, d7, d17 \n" // d1*2217 + 51000 + + "vceq.s16 d4, d7, #0 \n" + + "vshr.s16 d0, d0, #4 \n" + "vshr.s16 d2, d2, #4 \n" + + "vmlal.s16 q11, d6, d17 \n" // c1*2217 + d1*5352 + 12000 + "vmlsl.s16 q12, d6, d16 \n" // d1*2217 - c1*5352 + 51000 + + "vmvn d4, d4 \n" // !(d1 == 0) + // op[4] = (c1*2217 + d1*5352 + 12000)>>16 + "vshrn.s32 d1, q11, #16 \n" + // op[4] += (d1!=0) + "vsub.s16 d1, d1, d4 \n" + // op[12]= (d1*2217 - c1*5352 + 51000)>>16 + "vshrn.s32 d3, q12, #16 \n" + + // set result to out array + "vst1.16 {q0, q1}, [%[out]] \n" + : [src_ptr] "+r"(src_ptr), [ref_ptr] "+r"(ref_ptr), + [coeff32] "+r"(coeff32) // modified registers + : [kBPS] "r"(kBPS), [coeff16] "r"(coeff16), + [out] "r"(out) // constants + : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", + "q10", "q11", "q12", "q13" // clobbered + ); +} + +#endif + +#define LOAD_LANE_16b(VALUE, LANE) do { \ + (VALUE) = vld1_lane_s16(src, (VALUE), (LANE)); \ + src += stride; \ +} while (0) + +static void FTransformWHT(const int16_t* src, int16_t* out) { + const int stride = 16; + const int16x4_t zero = vdup_n_s16(0); + int32x4x4_t tmp0; + int16x4x4_t in; + INIT_VECTOR4(in, zero, zero, zero, zero); + LOAD_LANE_16b(in.val[0], 0); + LOAD_LANE_16b(in.val[1], 0); + LOAD_LANE_16b(in.val[2], 0); + LOAD_LANE_16b(in.val[3], 0); + LOAD_LANE_16b(in.val[0], 1); + LOAD_LANE_16b(in.val[1], 1); + LOAD_LANE_16b(in.val[2], 1); + LOAD_LANE_16b(in.val[3], 1); + LOAD_LANE_16b(in.val[0], 2); + LOAD_LANE_16b(in.val[1], 2); + LOAD_LANE_16b(in.val[2], 2); + LOAD_LANE_16b(in.val[3], 2); + LOAD_LANE_16b(in.val[0], 3); + LOAD_LANE_16b(in.val[1], 3); + LOAD_LANE_16b(in.val[2], 3); + LOAD_LANE_16b(in.val[3], 3); + + { + // a0 = in[0 * 16] + in[2 * 16] + // a1 = in[1 * 16] + in[3 * 16] + // a2 = in[1 * 16] - in[3 * 16] + // a3 = in[0 * 16] - in[2 * 16] + const int32x4_t a0 = vaddl_s16(in.val[0], in.val[2]); + const int32x4_t a1 = vaddl_s16(in.val[1], in.val[3]); + const int32x4_t a2 = vsubl_s16(in.val[1], in.val[3]); + const int32x4_t a3 = vsubl_s16(in.val[0], in.val[2]); + tmp0.val[0] = vaddq_s32(a0, a1); + tmp0.val[1] = vaddq_s32(a3, a2); + tmp0.val[2] = vsubq_s32(a3, a2); + tmp0.val[3] = vsubq_s32(a0, a1); + } + { + const int32x4x4_t tmp1 = Transpose4x4(tmp0); + // a0 = tmp[0 + i] + tmp[ 8 + i] + // a1 = tmp[4 + i] + tmp[12 + i] + // a2 = tmp[4 + i] - tmp[12 + i] + // a3 = tmp[0 + i] - tmp[ 8 + i] + const int32x4_t a0 = vaddq_s32(tmp1.val[0], tmp1.val[2]); + const int32x4_t a1 = vaddq_s32(tmp1.val[1], tmp1.val[3]); + const int32x4_t a2 = vsubq_s32(tmp1.val[1], tmp1.val[3]); + const int32x4_t a3 = vsubq_s32(tmp1.val[0], tmp1.val[2]); + const int32x4_t b0 = vhaddq_s32(a0, a1); // (a0 + a1) >> 1 + const int32x4_t b1 = vhaddq_s32(a3, a2); // (a3 + a2) >> 1 + const int32x4_t b2 = vhsubq_s32(a3, a2); // (a3 - a2) >> 1 + const int32x4_t b3 = vhsubq_s32(a0, a1); // (a0 - a1) >> 1 + const int16x4_t out0 = vmovn_s32(b0); + const int16x4_t out1 = vmovn_s32(b1); + const int16x4_t out2 = vmovn_s32(b2); + const int16x4_t out3 = vmovn_s32(b3); + + vst1_s16(out + 0, out0); + vst1_s16(out + 4, out1); + vst1_s16(out + 8, out2); + vst1_s16(out + 12, out3); + } +} +#undef LOAD_LANE_16b + +//------------------------------------------------------------------------------ +// Texture distortion +// +// We try to match the spectral content (weighted) between source and +// reconstructed samples. + +// a 0123, b 0123 +// a 4567, b 4567 +// a 89ab, b 89ab +// a cdef, b cdef +// +// transpose +// +// a 048c, b 048c +// a 159d, b 159d +// a 26ae, b 26ae +// a 37bf, b 37bf +// +static WEBP_INLINE uint8x8x4_t DistoTranspose4x4U8(uint8x8x4_t d4_in) { + const uint8x8x2_t d2_tmp0 = vtrn_u8(d4_in.val[0], d4_in.val[1]); + const uint8x8x2_t d2_tmp1 = vtrn_u8(d4_in.val[2], d4_in.val[3]); + const uint16x4x2_t d2_tmp2 = vtrn_u16(vreinterpret_u16_u8(d2_tmp0.val[0]), + vreinterpret_u16_u8(d2_tmp1.val[0])); + const uint16x4x2_t d2_tmp3 = vtrn_u16(vreinterpret_u16_u8(d2_tmp0.val[1]), + vreinterpret_u16_u8(d2_tmp1.val[1])); + + d4_in.val[0] = vreinterpret_u8_u16(d2_tmp2.val[0]); + d4_in.val[2] = vreinterpret_u8_u16(d2_tmp2.val[1]); + d4_in.val[1] = vreinterpret_u8_u16(d2_tmp3.val[0]); + d4_in.val[3] = vreinterpret_u8_u16(d2_tmp3.val[1]); + return d4_in; +} + +static WEBP_INLINE int16x8x4_t DistoTranspose4x4S16(int16x8x4_t q4_in) { + const int16x8x2_t q2_tmp0 = vtrnq_s16(q4_in.val[0], q4_in.val[1]); + const int16x8x2_t q2_tmp1 = vtrnq_s16(q4_in.val[2], q4_in.val[3]); + const int32x4x2_t q2_tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q2_tmp0.val[0]), + vreinterpretq_s32_s16(q2_tmp1.val[0])); + const int32x4x2_t q2_tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q2_tmp0.val[1]), + vreinterpretq_s32_s16(q2_tmp1.val[1])); + q4_in.val[0] = vreinterpretq_s16_s32(q2_tmp2.val[0]); + q4_in.val[2] = vreinterpretq_s16_s32(q2_tmp2.val[1]); + q4_in.val[1] = vreinterpretq_s16_s32(q2_tmp3.val[0]); + q4_in.val[3] = vreinterpretq_s16_s32(q2_tmp3.val[1]); + return q4_in; +} + +static WEBP_INLINE int16x8x4_t DistoHorizontalPass(const uint8x8x4_t d4_in) { + // {a0, a1} = {in[0] + in[2], in[1] + in[3]} + // {a3, a2} = {in[0] - in[2], in[1] - in[3]} + const int16x8_t q_a0 = vreinterpretq_s16_u16(vaddl_u8(d4_in.val[0], + d4_in.val[2])); + const int16x8_t q_a1 = vreinterpretq_s16_u16(vaddl_u8(d4_in.val[1], + d4_in.val[3])); + const int16x8_t q_a3 = vreinterpretq_s16_u16(vsubl_u8(d4_in.val[0], + d4_in.val[2])); + const int16x8_t q_a2 = vreinterpretq_s16_u16(vsubl_u8(d4_in.val[1], + d4_in.val[3])); + int16x8x4_t q4_out; + // tmp[0] = a0 + a1 + // tmp[1] = a3 + a2 + // tmp[2] = a3 - a2 + // tmp[3] = a0 - a1 + INIT_VECTOR4(q4_out, + vaddq_s16(q_a0, q_a1), vaddq_s16(q_a3, q_a2), + vsubq_s16(q_a3, q_a2), vsubq_s16(q_a0, q_a1)); + return q4_out; +} + +static WEBP_INLINE int16x8x4_t DistoVerticalPass(int16x8x4_t q4_in) { + const int16x8_t q_a0 = vaddq_s16(q4_in.val[0], q4_in.val[2]); + const int16x8_t q_a1 = vaddq_s16(q4_in.val[1], q4_in.val[3]); + const int16x8_t q_a2 = vsubq_s16(q4_in.val[1], q4_in.val[3]); + const int16x8_t q_a3 = vsubq_s16(q4_in.val[0], q4_in.val[2]); + + q4_in.val[0] = vaddq_s16(q_a0, q_a1); + q4_in.val[1] = vaddq_s16(q_a3, q_a2); + q4_in.val[2] = vabdq_s16(q_a3, q_a2); + q4_in.val[3] = vabdq_s16(q_a0, q_a1); + q4_in.val[0] = vabsq_s16(q4_in.val[0]); + q4_in.val[1] = vabsq_s16(q4_in.val[1]); + return q4_in; +} + +static WEBP_INLINE int16x4x4_t DistoLoadW(const uint16_t* w) { + const uint16x8_t q_w07 = vld1q_u16(&w[0]); + const uint16x8_t q_w8f = vld1q_u16(&w[8]); + int16x4x4_t d4_w; + INIT_VECTOR4(d4_w, + vget_low_s16(vreinterpretq_s16_u16(q_w07)), + vget_high_s16(vreinterpretq_s16_u16(q_w07)), + vget_low_s16(vreinterpretq_s16_u16(q_w8f)), + vget_high_s16(vreinterpretq_s16_u16(q_w8f))); + return d4_w; +} + +static WEBP_INLINE int32x2_t DistoSum(const int16x8x4_t q4_in, + const int16x4x4_t d4_w) { + int32x2_t d_sum; + // sum += w[ 0] * abs(b0); + // sum += w[ 4] * abs(b1); + // sum += w[ 8] * abs(b2); + // sum += w[12] * abs(b3); + int32x4_t q_sum0 = vmull_s16(d4_w.val[0], vget_low_s16(q4_in.val[0])); + int32x4_t q_sum1 = vmull_s16(d4_w.val[1], vget_low_s16(q4_in.val[1])); + int32x4_t q_sum2 = vmull_s16(d4_w.val[2], vget_low_s16(q4_in.val[2])); + int32x4_t q_sum3 = vmull_s16(d4_w.val[3], vget_low_s16(q4_in.val[3])); + q_sum0 = vmlsl_s16(q_sum0, d4_w.val[0], vget_high_s16(q4_in.val[0])); + q_sum1 = vmlsl_s16(q_sum1, d4_w.val[1], vget_high_s16(q4_in.val[1])); + q_sum2 = vmlsl_s16(q_sum2, d4_w.val[2], vget_high_s16(q4_in.val[2])); + q_sum3 = vmlsl_s16(q_sum3, d4_w.val[3], vget_high_s16(q4_in.val[3])); + + q_sum0 = vaddq_s32(q_sum0, q_sum1); + q_sum2 = vaddq_s32(q_sum2, q_sum3); + q_sum2 = vaddq_s32(q_sum0, q_sum2); + d_sum = vpadd_s32(vget_low_s32(q_sum2), vget_high_s32(q_sum2)); + d_sum = vpadd_s32(d_sum, d_sum); + return d_sum; +} + +#define LOAD_LANE_32b(src, VALUE, LANE) \ + (VALUE) = vld1_lane_u32((const uint32_t*)(src), (VALUE), (LANE)) + +// Hadamard transform +// Returns the weighted sum of the absolute value of transformed coefficients. +static int Disto4x4(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + uint32x2_t d_in_ab_0123 = vdup_n_u32(0); + uint32x2_t d_in_ab_4567 = vdup_n_u32(0); + uint32x2_t d_in_ab_89ab = vdup_n_u32(0); + uint32x2_t d_in_ab_cdef = vdup_n_u32(0); + uint8x8x4_t d4_in; + + // load data a, b + LOAD_LANE_32b(a + 0 * BPS, d_in_ab_0123, 0); + LOAD_LANE_32b(a + 1 * BPS, d_in_ab_4567, 0); + LOAD_LANE_32b(a + 2 * BPS, d_in_ab_89ab, 0); + LOAD_LANE_32b(a + 3 * BPS, d_in_ab_cdef, 0); + LOAD_LANE_32b(b + 0 * BPS, d_in_ab_0123, 1); + LOAD_LANE_32b(b + 1 * BPS, d_in_ab_4567, 1); + LOAD_LANE_32b(b + 2 * BPS, d_in_ab_89ab, 1); + LOAD_LANE_32b(b + 3 * BPS, d_in_ab_cdef, 1); + INIT_VECTOR4(d4_in, + vreinterpret_u8_u32(d_in_ab_0123), + vreinterpret_u8_u32(d_in_ab_4567), + vreinterpret_u8_u32(d_in_ab_89ab), + vreinterpret_u8_u32(d_in_ab_cdef)); + + { + // horizontal pass + const uint8x8x4_t d4_t = DistoTranspose4x4U8(d4_in); + const int16x8x4_t q4_h = DistoHorizontalPass(d4_t); + const int16x4x4_t d4_w = DistoLoadW(w); + // vertical pass + const int16x8x4_t q4_t = DistoTranspose4x4S16(q4_h); + const int16x8x4_t q4_v = DistoVerticalPass(q4_t); + int32x2_t d_sum = DistoSum(q4_v, d4_w); + + // abs(sum2 - sum1) >> 5 + d_sum = vabs_s32(d_sum); + d_sum = vshr_n_s32(d_sum, 5); + return vget_lane_s32(d_sum, 0); + } +} +#undef LOAD_LANE_32b + +static int Disto16x16(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4(a + x + y, b + x + y, w); + } + } + return D; +} + +//------------------------------------------------------------------------------ + +static void CollectHistogram(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + const uint16x8_t max_coeff_thresh = vdupq_n_u16(MAX_COEFF_THRESH); + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + { + int k; + const int16x8_t a0 = vld1q_s16(out + 0); + const int16x8_t b0 = vld1q_s16(out + 8); + const uint16x8_t a1 = vreinterpretq_u16_s16(vabsq_s16(a0)); + const uint16x8_t b1 = vreinterpretq_u16_s16(vabsq_s16(b0)); + const uint16x8_t a2 = vshrq_n_u16(a1, 3); + const uint16x8_t b2 = vshrq_n_u16(b1, 3); + const uint16x8_t a3 = vminq_u16(a2, max_coeff_thresh); + const uint16x8_t b3 = vminq_u16(b2, max_coeff_thresh); + vst1q_s16(out + 0, vreinterpretq_s16_u16(a3)); + vst1q_s16(out + 8, vreinterpretq_s16_u16(b3)); + // Convert coefficients to bin. + for (k = 0; k < 16; ++k) { + ++distribution[out[k]]; + } + } + } + VP8SetHistogramData(distribution, histo); +} + +//------------------------------------------------------------------------------ + +static WEBP_INLINE void AccumulateSSE16(const uint8_t* const a, + const uint8_t* const b, + uint32x4_t* const sum) { + const uint8x16_t a0 = vld1q_u8(a); + const uint8x16_t b0 = vld1q_u8(b); + const uint8x16_t abs_diff = vabdq_u8(a0, b0); + uint16x8_t prod = vmull_u8(vget_low_u8(abs_diff), vget_low_u8(abs_diff)); + prod = vmlal_u8(prod, vget_high_u8(abs_diff), vget_high_u8(abs_diff)); + *sum = vpadalq_u16(*sum, prod); // pair-wise add and accumulate +} + +// Horizontal sum of all four uint32_t values in 'sum'. +static int SumToInt(uint32x4_t sum) { + const uint64x2_t sum2 = vpaddlq_u32(sum); + const uint64_t sum3 = vgetq_lane_u64(sum2, 0) + vgetq_lane_u64(sum2, 1); + return (int)sum3; +} + +static int SSE16x16(const uint8_t* a, const uint8_t* b) { + uint32x4_t sum = vdupq_n_u32(0); + int y; + for (y = 0; y < 16; ++y) { + AccumulateSSE16(a + y * BPS, b + y * BPS, &sum); + } + return SumToInt(sum); +} + +static int SSE16x8(const uint8_t* a, const uint8_t* b) { + uint32x4_t sum = vdupq_n_u32(0); + int y; + for (y = 0; y < 8; ++y) { + AccumulateSSE16(a + y * BPS, b + y * BPS, &sum); + } + return SumToInt(sum); +} + +static int SSE8x8(const uint8_t* a, const uint8_t* b) { + uint32x4_t sum = vdupq_n_u32(0); + int y; + for (y = 0; y < 8; ++y) { + const uint8x8_t a0 = vld1_u8(a + y * BPS); + const uint8x8_t b0 = vld1_u8(b + y * BPS); + const uint8x8_t abs_diff = vabd_u8(a0, b0); + const uint16x8_t prod = vmull_u8(abs_diff, abs_diff); + sum = vpadalq_u16(sum, prod); + } + return SumToInt(sum); +} + +static int SSE4x4(const uint8_t* a, const uint8_t* b) { + const uint8x16_t a0 = Load4x4(a); + const uint8x16_t b0 = Load4x4(b); + const uint8x16_t abs_diff = vabdq_u8(a0, b0); + uint16x8_t prod = vmull_u8(vget_low_u8(abs_diff), vget_low_u8(abs_diff)); + prod = vmlal_u8(prod, vget_high_u8(abs_diff), vget_high_u8(abs_diff)); + return SumToInt(vpaddlq_u16(prod)); +} + +//------------------------------------------------------------------------------ + +// Compilation with gcc-4.6.x is problematic for now. +#if !defined(WORK_AROUND_GCC) + +static int16x8_t Quantize(int16_t* const in, + const VP8Matrix* const mtx, int offset) { + const uint16x8_t sharp = vld1q_u16(&mtx->sharpen_[offset]); + const uint16x8_t q = vld1q_u16(&mtx->q_[offset]); + const uint16x8_t iq = vld1q_u16(&mtx->iq_[offset]); + const uint32x4_t bias0 = vld1q_u32(&mtx->bias_[offset + 0]); + const uint32x4_t bias1 = vld1q_u32(&mtx->bias_[offset + 4]); + + const int16x8_t a = vld1q_s16(in + offset); // in + const uint16x8_t b = vreinterpretq_u16_s16(vabsq_s16(a)); // coeff = abs(in) + const int16x8_t sign = vshrq_n_s16(a, 15); // sign + const uint16x8_t c = vaddq_u16(b, sharp); // + sharpen + const uint32x4_t m0 = vmull_u16(vget_low_u16(c), vget_low_u16(iq)); + const uint32x4_t m1 = vmull_u16(vget_high_u16(c), vget_high_u16(iq)); + const uint32x4_t m2 = vhaddq_u32(m0, bias0); + const uint32x4_t m3 = vhaddq_u32(m1, bias1); // (coeff * iQ + bias) >> 1 + const uint16x8_t c0 = vcombine_u16(vshrn_n_u32(m2, 16), + vshrn_n_u32(m3, 16)); // QFIX=17 = 16+1 + const uint16x8_t c1 = vminq_u16(c0, vdupq_n_u16(MAX_LEVEL)); + const int16x8_t c2 = veorq_s16(vreinterpretq_s16_u16(c1), sign); + const int16x8_t c3 = vsubq_s16(c2, sign); // restore sign + const int16x8_t c4 = vmulq_s16(c3, vreinterpretq_s16_u16(q)); + vst1q_s16(in + offset, c4); + assert(QFIX == 17); // this function can't work as is if QFIX != 16+1 + return c3; +} + +static const uint8_t kShuffles[4][8] = { + { 0, 1, 2, 3, 8, 9, 16, 17 }, + { 10, 11, 4, 5, 6, 7, 12, 13 }, + { 18, 19, 24, 25, 26, 27, 20, 21 }, + { 14, 15, 22, 23, 28, 29, 30, 31 } +}; + +static int QuantizeBlock(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + const int16x8_t out0 = Quantize(in, mtx, 0); + const int16x8_t out1 = Quantize(in, mtx, 8); + uint8x8x4_t shuffles; + // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use + // non-standard versions there. +#if defined(__APPLE__) && defined(__aarch64__) && \ + defined(__apple_build_version__) && (__apple_build_version__< 6020037) + uint8x16x2_t all_out; + INIT_VECTOR2(all_out, vreinterpretq_u8_s16(out0), vreinterpretq_u8_s16(out1)); + INIT_VECTOR4(shuffles, + vtbl2q_u8(all_out, vld1_u8(kShuffles[0])), + vtbl2q_u8(all_out, vld1_u8(kShuffles[1])), + vtbl2q_u8(all_out, vld1_u8(kShuffles[2])), + vtbl2q_u8(all_out, vld1_u8(kShuffles[3]))); +#else + uint8x8x4_t all_out; + INIT_VECTOR4(all_out, + vreinterpret_u8_s16(vget_low_s16(out0)), + vreinterpret_u8_s16(vget_high_s16(out0)), + vreinterpret_u8_s16(vget_low_s16(out1)), + vreinterpret_u8_s16(vget_high_s16(out1))); + INIT_VECTOR4(shuffles, + vtbl4_u8(all_out, vld1_u8(kShuffles[0])), + vtbl4_u8(all_out, vld1_u8(kShuffles[1])), + vtbl4_u8(all_out, vld1_u8(kShuffles[2])), + vtbl4_u8(all_out, vld1_u8(kShuffles[3]))); +#endif + // Zigzag reordering + vst1_u8((uint8_t*)(out + 0), shuffles.val[0]); + vst1_u8((uint8_t*)(out + 4), shuffles.val[1]); + vst1_u8((uint8_t*)(out + 8), shuffles.val[2]); + vst1_u8((uint8_t*)(out + 12), shuffles.val[3]); + // test zeros + if (*(uint64_t*)(out + 0) != 0) return 1; + if (*(uint64_t*)(out + 4) != 0) return 1; + if (*(uint64_t*)(out + 8) != 0) return 1; + if (*(uint64_t*)(out + 12) != 0) return 1; + return 0; +} + +static int Quantize2Blocks(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = QuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= QuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} + +#endif // !WORK_AROUND_GCC + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitNEON(void) { + VP8ITransform = ITransform; + VP8FTransform = FTransform; + + VP8FTransformWHT = FTransformWHT; + + VP8TDisto4x4 = Disto4x4; + VP8TDisto16x16 = Disto16x16; + VP8CollectHistogram = CollectHistogram; + VP8SSE16x16 = SSE16x16; + VP8SSE16x8 = SSE16x8; + VP8SSE8x8 = SSE8x8; + VP8SSE4x4 = SSE4x4; +#if !defined(WORK_AROUND_GCC) + VP8EncQuantizeBlock = QuantizeBlock; + VP8EncQuantize2Blocks = Quantize2Blocks; +#endif +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(VP8EncDspInitNEON) + +#endif // WEBP_USE_NEON diff --git a/drivers/webp/dsp/enc_sse2.c b/drivers/webp/dsp/enc_sse2.c index b046761dc1..63d9cecd85 100644 --- a/drivers/webp/dsp/enc_sse2.c +++ b/drivers/webp/dsp/enc_sse2.c @@ -1,8 +1,10 @@ // Copyright 2011 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // SSE2 version of speed-critical encoding functions. @@ -15,64 +17,55 @@ #include <stdlib.h> // for abs() #include <emmintrin.h> +#include "../enc/cost.h" #include "../enc/vp8enci.h" +#include "../utils/utils.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { +//------------------------------------------------------------------------------ +// Quite useful macro for debugging. Left here for convenience. + +#if 0 +#include <stdio.h> +static void PrintReg(const __m128i r, const char* const name, int size) { + int n; + union { + __m128i r; + uint8_t i8[16]; + uint16_t i16[8]; + uint32_t i32[4]; + uint64_t i64[2]; + } tmp; + tmp.r = r; + fprintf(stderr, "%s\t: ", name); + if (size == 8) { + for (n = 0; n < 16; ++n) fprintf(stderr, "%.2x ", tmp.i8[n]); + } else if (size == 16) { + for (n = 0; n < 8; ++n) fprintf(stderr, "%.4x ", tmp.i16[n]); + } else if (size == 32) { + for (n = 0; n < 4; ++n) fprintf(stderr, "%.8x ", tmp.i32[n]); + } else { + for (n = 0; n < 2; ++n) fprintf(stderr, "%.16lx ", tmp.i64[n]); + } + fprintf(stderr, "\n"); +} #endif //------------------------------------------------------------------------------ -// Compute susceptibility based on DCT-coeff histograms: -// the higher, the "easier" the macroblock is to compress. - -static int CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred, - int start_block, int end_block) { - int histo[MAX_COEFF_THRESH + 1] = { 0 }; - int16_t out[16]; - int j, k; - const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); - for (j = start_block; j < end_block; ++j) { - VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); - - // Convert coefficients to bin (within out[]). - { - // Load. - const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); - const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); - // sign(out) = out >> 15 (0x0000 if positive, 0xffff if negative) - const __m128i sign0 = _mm_srai_epi16(out0, 15); - const __m128i sign1 = _mm_srai_epi16(out1, 15); - // abs(out) = (out ^ sign) - sign - const __m128i xor0 = _mm_xor_si128(out0, sign0); - const __m128i xor1 = _mm_xor_si128(out1, sign1); - const __m128i abs0 = _mm_sub_epi16(xor0, sign0); - const __m128i abs1 = _mm_sub_epi16(xor1, sign1); - // v = abs(out) >> 2 - const __m128i v0 = _mm_srai_epi16(abs0, 2); - const __m128i v1 = _mm_srai_epi16(abs1, 2); - // bin = min(v, MAX_COEFF_THRESH) - const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); - const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); - // Store. - _mm_storeu_si128((__m128i*)&out[0], bin0); - _mm_storeu_si128((__m128i*)&out[8], bin1); - } +// util for unaligned loads. - // Use bin to update histogram. - for (k = 0; k < 16; ++k) { - histo[out[k]]++; - } - } - - return VP8GetAlpha(histo); +// memcpy() is the safe way of moving potentially unaligned 32b memory. +static WEBP_INLINE uint32_t MemToUint32(const uint8_t* const ptr) { + uint32_t A; + memcpy(&A, (const int*)ptr, sizeof(A)); + return A; } //------------------------------------------------------------------------------ // Transforms (Paragraph 14.4) // Does one or two inverse transforms. -static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst, - int do_two) { +static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst, + int do_two) { // This implementation makes use of 16-bit fixed point versions of two // multiply constants: // K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16 @@ -99,19 +92,19 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst, // use nor store. __m128i in0, in1, in2, in3; { - in0 = _mm_loadl_epi64((__m128i*)&in[0]); - in1 = _mm_loadl_epi64((__m128i*)&in[4]); - in2 = _mm_loadl_epi64((__m128i*)&in[8]); - in3 = _mm_loadl_epi64((__m128i*)&in[12]); + in0 = _mm_loadl_epi64((const __m128i*)&in[0]); + in1 = _mm_loadl_epi64((const __m128i*)&in[4]); + in2 = _mm_loadl_epi64((const __m128i*)&in[8]); + in3 = _mm_loadl_epi64((const __m128i*)&in[12]); // a00 a10 a20 a30 x x x x // a01 a11 a21 a31 x x x x // a02 a12 a22 a32 x x x x // a03 a13 a23 a33 x x x x if (do_two) { - const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]); - const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]); - const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]); - const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]); + const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]); + const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]); + const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]); + const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]); in0 = _mm_unpacklo_epi64(in0, inB0); in1 = _mm_unpacklo_epi64(in1, inB1); in2 = _mm_unpacklo_epi64(in2, inB2); @@ -243,21 +236,21 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst, // Add inverse transform to 'ref' and store. { - const __m128i zero = _mm_set1_epi16(0); + const __m128i zero = _mm_setzero_si128(); // Load the reference(s). __m128i ref0, ref1, ref2, ref3; if (do_two) { // Load eight bytes/pixels per line. - ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]); - ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]); - ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]); - ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]); + ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]); + ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]); + ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]); + ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]); } else { // Load four bytes/pixels per line. - ref0 = _mm_cvtsi32_si128(*(int*)&ref[0 * BPS]); - ref1 = _mm_cvtsi32_si128(*(int*)&ref[1 * BPS]); - ref2 = _mm_cvtsi32_si128(*(int*)&ref[2 * BPS]); - ref3 = _mm_cvtsi32_si128(*(int*)&ref[3 * BPS]); + ref0 = _mm_cvtsi32_si128(MemToUint32(&ref[0 * BPS])); + ref1 = _mm_cvtsi32_si128(MemToUint32(&ref[1 * BPS])); + ref2 = _mm_cvtsi32_si128(MemToUint32(&ref[2 * BPS])); + ref3 = _mm_cvtsi32_si128(MemToUint32(&ref[3 * BPS])); } // Convert to 16b. ref0 = _mm_unpacklo_epi8(ref0, zero); @@ -291,200 +284,865 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst, } } -static void FTransformSSE2(const uint8_t* src, const uint8_t* ref, - int16_t* out) { +static void FTransformPass1(const __m128i* const in01, + const __m128i* const in23, + __m128i* const out01, + __m128i* const out32) { + const __m128i k937 = _mm_set1_epi32(937); + const __m128i k1812 = _mm_set1_epi32(1812); + + const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8); + const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8); + const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352, + 2217, 5352, 2217, 5352); + const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217, + -5352, 2217, -5352, 2217); + + // *in01 = 00 01 10 11 02 03 12 13 + // *in23 = 20 21 30 31 22 23 32 33 + const __m128i shuf01_p = _mm_shufflehi_epi16(*in01, _MM_SHUFFLE(2, 3, 0, 1)); + const __m128i shuf23_p = _mm_shufflehi_epi16(*in23, _MM_SHUFFLE(2, 3, 0, 1)); + // 00 01 10 11 03 02 13 12 + // 20 21 30 31 23 22 33 32 + const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p); + const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p); + // 00 01 10 11 20 21 30 31 + // 03 02 13 12 23 22 33 32 + const __m128i a01 = _mm_add_epi16(s01, s32); + const __m128i a32 = _mm_sub_epi16(s01, s32); + // [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ] + // [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ] + + const __m128i tmp0 = _mm_madd_epi16(a01, k88p); // [ (a0 + a1) << 3, ... ] + const __m128i tmp2 = _mm_madd_epi16(a01, k88m); // [ (a0 - a1) << 3, ... ] + const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p); + const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m); + const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812); + const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937); + const __m128i tmp1 = _mm_srai_epi32(tmp1_2, 9); + const __m128i tmp3 = _mm_srai_epi32(tmp3_2, 9); + const __m128i s03 = _mm_packs_epi32(tmp0, tmp2); + const __m128i s12 = _mm_packs_epi32(tmp1, tmp3); + const __m128i s_lo = _mm_unpacklo_epi16(s03, s12); // 0 1 0 1 0 1... + const __m128i s_hi = _mm_unpackhi_epi16(s03, s12); // 2 3 2 3 2 3 + const __m128i v23 = _mm_unpackhi_epi32(s_lo, s_hi); + *out01 = _mm_unpacklo_epi32(s_lo, s_hi); + *out32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2)); // 3 2 3 2 3 2.. +} + +static void FTransformPass2(const __m128i* const v01, const __m128i* const v32, + int16_t* out) { const __m128i zero = _mm_setzero_si128(); const __m128i seven = _mm_set1_epi16(7); - const __m128i k7500 = _mm_set1_epi32(7500); - const __m128i k14500 = _mm_set1_epi32(14500); - const __m128i k51000 = _mm_set1_epi32(51000); - const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16)); const __m128i k5352_2217 = _mm_set_epi16(5352, 2217, 5352, 2217, 5352, 2217, 5352, 2217); const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352, 2217, -5352, 2217, -5352); + const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16)); + const __m128i k51000 = _mm_set1_epi32(51000); + + // Same operations are done on the (0,3) and (1,2) pairs. + // a0 = v0 + v3 + // a1 = v1 + v2 + // a3 = v0 - v3 + // a2 = v1 - v2 + const __m128i a01 = _mm_add_epi16(*v01, *v32); + const __m128i a32 = _mm_sub_epi16(*v01, *v32); + const __m128i a11 = _mm_unpackhi_epi64(a01, a01); + const __m128i a22 = _mm_unpackhi_epi64(a32, a32); + const __m128i a01_plus_7 = _mm_add_epi16(a01, seven); + + // d0 = (a0 + a1 + 7) >> 4; + // d2 = (a0 - a1 + 7) >> 4; + const __m128i c0 = _mm_add_epi16(a01_plus_7, a11); + const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11); + const __m128i d0 = _mm_srai_epi16(c0, 4); + const __m128i d2 = _mm_srai_epi16(c2, 4); + + // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16) + // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16) + const __m128i b23 = _mm_unpacklo_epi16(a22, a32); + const __m128i c1 = _mm_madd_epi16(b23, k5352_2217); + const __m128i c3 = _mm_madd_epi16(b23, k2217_5352); + const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one); + const __m128i d3 = _mm_add_epi32(c3, k51000); + const __m128i e1 = _mm_srai_epi32(d1, 16); + const __m128i e3 = _mm_srai_epi32(d3, 16); + const __m128i f1 = _mm_packs_epi32(e1, e1); + const __m128i f3 = _mm_packs_epi32(e3, e3); + // f1 = f1 + (a3 != 0); + // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the + // desired (0, 1), we add one earlier through k12000_plus_one. + // -> f1 = f1 + 1 - (a3 == 0) + const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero)); + + const __m128i d0_g1 = _mm_unpacklo_epi64(d0, g1); + const __m128i d2_f3 = _mm_unpacklo_epi64(d2, f3); + _mm_storeu_si128((__m128i*)&out[0], d0_g1); + _mm_storeu_si128((__m128i*)&out[8], d2_f3); +} + +static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) { + const __m128i zero = _mm_setzero_si128(); + // Load src and convert to 16b. + const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]); + const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]); + const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]); + const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]); + const __m128i src_0 = _mm_unpacklo_epi8(src0, zero); + const __m128i src_1 = _mm_unpacklo_epi8(src1, zero); + const __m128i src_2 = _mm_unpacklo_epi8(src2, zero); + const __m128i src_3 = _mm_unpacklo_epi8(src3, zero); + // Load ref and convert to 16b. + const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]); + const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]); + const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]); + const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]); + const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero); + const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero); + const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero); + const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero); + // Compute difference. -> 00 01 02 03 00 00 00 00 + const __m128i diff0 = _mm_sub_epi16(src_0, ref_0); + const __m128i diff1 = _mm_sub_epi16(src_1, ref_1); + const __m128i diff2 = _mm_sub_epi16(src_2, ref_2); + const __m128i diff3 = _mm_sub_epi16(src_3, ref_3); + + // Unpack and shuffle + // 00 01 02 03 0 0 0 0 + // 10 11 12 13 0 0 0 0 + // 20 21 22 23 0 0 0 0 + // 30 31 32 33 0 0 0 0 + const __m128i shuf01 = _mm_unpacklo_epi32(diff0, diff1); + const __m128i shuf23 = _mm_unpacklo_epi32(diff2, diff3); __m128i v01, v32; - // Difference between src and ref and initial transpose. - { - // Load src and convert to 16b. - const __m128i src0 = _mm_loadl_epi64((__m128i*)&src[0 * BPS]); - const __m128i src1 = _mm_loadl_epi64((__m128i*)&src[1 * BPS]); - const __m128i src2 = _mm_loadl_epi64((__m128i*)&src[2 * BPS]); - const __m128i src3 = _mm_loadl_epi64((__m128i*)&src[3 * BPS]); - const __m128i src_0 = _mm_unpacklo_epi8(src0, zero); - const __m128i src_1 = _mm_unpacklo_epi8(src1, zero); - const __m128i src_2 = _mm_unpacklo_epi8(src2, zero); - const __m128i src_3 = _mm_unpacklo_epi8(src3, zero); - // Load ref and convert to 16b. - const __m128i ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]); - const __m128i ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]); - const __m128i ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]); - const __m128i ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]); - const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero); - const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero); - const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero); - const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero); - // Compute difference. - const __m128i diff0 = _mm_sub_epi16(src_0, ref_0); - const __m128i diff1 = _mm_sub_epi16(src_1, ref_1); - const __m128i diff2 = _mm_sub_epi16(src_2, ref_2); - const __m128i diff3 = _mm_sub_epi16(src_3, ref_3); - - // Transpose. - // 00 01 02 03 0 0 0 0 - // 10 11 12 13 0 0 0 0 - // 20 21 22 23 0 0 0 0 - // 30 31 32 33 0 0 0 0 - const __m128i transpose0_0 = _mm_unpacklo_epi16(diff0, diff1); - const __m128i transpose0_1 = _mm_unpacklo_epi16(diff2, diff3); - // 00 10 01 11 02 12 03 13 - // 20 30 21 31 22 32 23 33 - const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); - v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); - v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2)); - // a02 a12 a22 a32 a03 a13 a23 a33 - // a00 a10 a20 a30 a01 a11 a21 a31 - // a03 a13 a23 a33 a02 a12 a22 a32 - } - - // First pass and subsequent transpose. - { - // Same operations are done on the (0,3) and (1,2) pairs. - // b0 = (a0 + a3) << 3 - // b1 = (a1 + a2) << 3 - // b3 = (a0 - a3) << 3 - // b2 = (a1 - a2) << 3 - const __m128i a01 = _mm_add_epi16(v01, v32); - const __m128i a32 = _mm_sub_epi16(v01, v32); - const __m128i b01 = _mm_slli_epi16(a01, 3); - const __m128i b32 = _mm_slli_epi16(a32, 3); - const __m128i b11 = _mm_unpackhi_epi64(b01, b01); - const __m128i b22 = _mm_unpackhi_epi64(b32, b32); - - // e0 = b0 + b1 - // e2 = b0 - b1 - const __m128i e0 = _mm_add_epi16(b01, b11); - const __m128i e2 = _mm_sub_epi16(b01, b11); - const __m128i e02 = _mm_unpacklo_epi64(e0, e2); - - // e1 = (b3 * 5352 + b2 * 2217 + 14500) >> 12 - // e3 = (b3 * 2217 - b2 * 5352 + 7500) >> 12 - const __m128i b23 = _mm_unpacklo_epi16(b22, b32); - const __m128i c1 = _mm_madd_epi16(b23, k5352_2217); - const __m128i c3 = _mm_madd_epi16(b23, k2217_5352); - const __m128i d1 = _mm_add_epi32(c1, k14500); - const __m128i d3 = _mm_add_epi32(c3, k7500); - const __m128i e1 = _mm_srai_epi32(d1, 12); - const __m128i e3 = _mm_srai_epi32(d3, 12); - const __m128i e13 = _mm_packs_epi32(e1, e3); - - // Transpose. - // 00 01 02 03 20 21 22 23 - // 10 11 12 13 30 31 32 33 - const __m128i transpose0_0 = _mm_unpacklo_epi16(e02, e13); - const __m128i transpose0_1 = _mm_unpackhi_epi16(e02, e13); - // 00 10 01 11 02 12 03 13 - // 20 30 21 31 22 32 23 33 - const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); - v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); - v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2)); - // 02 12 22 32 03 13 23 33 - // 00 10 20 30 01 11 21 31 - // 03 13 23 33 02 12 22 32 - } + // First pass + FTransformPass1(&shuf01, &shuf23, &v01, &v32); // Second pass + FTransformPass2(&v01, &v32, out); +} + +static void FTransform2(const uint8_t* src, const uint8_t* ref, int16_t* out) { + const __m128i zero = _mm_setzero_si128(); + + // Load src and convert to 16b. + const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]); + const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]); + const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]); + const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]); + const __m128i src_0 = _mm_unpacklo_epi8(src0, zero); + const __m128i src_1 = _mm_unpacklo_epi8(src1, zero); + const __m128i src_2 = _mm_unpacklo_epi8(src2, zero); + const __m128i src_3 = _mm_unpacklo_epi8(src3, zero); + // Load ref and convert to 16b. + const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]); + const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]); + const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]); + const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]); + const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero); + const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero); + const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero); + const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero); + // Compute difference. -> 00 01 02 03 00' 01' 02' 03' + const __m128i diff0 = _mm_sub_epi16(src_0, ref_0); + const __m128i diff1 = _mm_sub_epi16(src_1, ref_1); + const __m128i diff2 = _mm_sub_epi16(src_2, ref_2); + const __m128i diff3 = _mm_sub_epi16(src_3, ref_3); + + // Unpack and shuffle + // 00 01 02 03 0 0 0 0 + // 10 11 12 13 0 0 0 0 + // 20 21 22 23 0 0 0 0 + // 30 31 32 33 0 0 0 0 + const __m128i shuf01l = _mm_unpacklo_epi32(diff0, diff1); + const __m128i shuf23l = _mm_unpacklo_epi32(diff2, diff3); + const __m128i shuf01h = _mm_unpackhi_epi32(diff0, diff1); + const __m128i shuf23h = _mm_unpackhi_epi32(diff2, diff3); + __m128i v01l, v32l; + __m128i v01h, v32h; + + // First pass + FTransformPass1(&shuf01l, &shuf23l, &v01l, &v32l); + FTransformPass1(&shuf01h, &shuf23h, &v01h, &v32h); + + // Second pass + FTransformPass2(&v01l, &v32l, out + 0); + FTransformPass2(&v01h, &v32h, out + 16); +} + +static void FTransformWHTRow(const int16_t* const in, __m128i* const out) { + const __m128i kMult1 = _mm_set_epi16(0, 0, 0, 0, 1, 1, 1, 1); + const __m128i kMult2 = _mm_set_epi16(0, 0, 0, 0, -1, 1, -1, 1); + const __m128i src0 = _mm_loadl_epi64((__m128i*)&in[0 * 16]); + const __m128i src1 = _mm_loadl_epi64((__m128i*)&in[1 * 16]); + const __m128i src2 = _mm_loadl_epi64((__m128i*)&in[2 * 16]); + const __m128i src3 = _mm_loadl_epi64((__m128i*)&in[3 * 16]); + const __m128i A01 = _mm_unpacklo_epi16(src0, src1); // A0 A1 | ... + const __m128i A23 = _mm_unpacklo_epi16(src2, src3); // A2 A3 | ... + const __m128i B0 = _mm_adds_epi16(A01, A23); // a0 | a1 | ... + const __m128i B1 = _mm_subs_epi16(A01, A23); // a3 | a2 | ... + const __m128i C0 = _mm_unpacklo_epi32(B0, B1); // a0 | a1 | a3 | a2 + const __m128i C1 = _mm_unpacklo_epi32(B1, B0); // a3 | a2 | a0 | a1 + const __m128i D0 = _mm_madd_epi16(C0, kMult1); // out0, out1 + const __m128i D1 = _mm_madd_epi16(C1, kMult2); // out2, out3 + *out = _mm_unpacklo_epi64(D0, D1); +} + +static void FTransformWHT(const int16_t* in, int16_t* out) { + __m128i row0, row1, row2, row3; + FTransformWHTRow(in + 0 * 64, &row0); + FTransformWHTRow(in + 1 * 64, &row1); + FTransformWHTRow(in + 2 * 64, &row2); + FTransformWHTRow(in + 3 * 64, &row3); + { - // Same operations are done on the (0,3) and (1,2) pairs. - // a0 = v0 + v3 - // a1 = v1 + v2 - // a3 = v0 - v3 - // a2 = v1 - v2 - const __m128i a01 = _mm_add_epi16(v01, v32); - const __m128i a32 = _mm_sub_epi16(v01, v32); - const __m128i a11 = _mm_unpackhi_epi64(a01, a01); - const __m128i a22 = _mm_unpackhi_epi64(a32, a32); - - // d0 = (a0 + a1 + 7) >> 4; - // d2 = (a0 - a1 + 7) >> 4; - const __m128i b0 = _mm_add_epi16(a01, a11); - const __m128i b2 = _mm_sub_epi16(a01, a11); - const __m128i c0 = _mm_add_epi16(b0, seven); - const __m128i c2 = _mm_add_epi16(b2, seven); - const __m128i d0 = _mm_srai_epi16(c0, 4); - const __m128i d2 = _mm_srai_epi16(c2, 4); - - // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16) - // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16) - const __m128i b23 = _mm_unpacklo_epi16(a22, a32); - const __m128i c1 = _mm_madd_epi16(b23, k5352_2217); - const __m128i c3 = _mm_madd_epi16(b23, k2217_5352); - const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one); - const __m128i d3 = _mm_add_epi32(c3, k51000); - const __m128i e1 = _mm_srai_epi32(d1, 16); - const __m128i e3 = _mm_srai_epi32(d3, 16); - const __m128i f1 = _mm_packs_epi32(e1, e1); - const __m128i f3 = _mm_packs_epi32(e3, e3); - // f1 = f1 + (a3 != 0); - // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the - // desired (0, 1), we add one earlier through k12000_plus_one. - const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero)); - - _mm_storel_epi64((__m128i*)&out[ 0], d0); - _mm_storel_epi64((__m128i*)&out[ 4], g1); - _mm_storel_epi64((__m128i*)&out[ 8], d2); - _mm_storel_epi64((__m128i*)&out[12], f3); + const __m128i a0 = _mm_add_epi32(row0, row2); + const __m128i a1 = _mm_add_epi32(row1, row3); + const __m128i a2 = _mm_sub_epi32(row1, row3); + const __m128i a3 = _mm_sub_epi32(row0, row2); + const __m128i b0 = _mm_srai_epi32(_mm_add_epi32(a0, a1), 1); + const __m128i b1 = _mm_srai_epi32(_mm_add_epi32(a3, a2), 1); + const __m128i b2 = _mm_srai_epi32(_mm_sub_epi32(a3, a2), 1); + const __m128i b3 = _mm_srai_epi32(_mm_sub_epi32(a0, a1), 1); + const __m128i out0 = _mm_packs_epi32(b0, b1); + const __m128i out1 = _mm_packs_epi32(b2, b3); + _mm_storeu_si128((__m128i*)&out[0], out0); + _mm_storeu_si128((__m128i*)&out[8], out1); + } +} + +//------------------------------------------------------------------------------ +// Compute susceptibility based on DCT-coeff histograms: +// the higher, the "easier" the macroblock is to compress. + +static void CollectHistogram(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + const __m128i zero = _mm_setzero_si128(); + const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + int k; + + FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + + // Convert coefficients to bin (within out[]). + { + // Load. + const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); + const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); + const __m128i d0 = _mm_sub_epi16(zero, out0); + const __m128i d1 = _mm_sub_epi16(zero, out1); + const __m128i abs0 = _mm_max_epi16(out0, d0); // abs(v), 16b + const __m128i abs1 = _mm_max_epi16(out1, d1); + // v = abs(out) >> 3 + const __m128i v0 = _mm_srai_epi16(abs0, 3); + const __m128i v1 = _mm_srai_epi16(abs1, 3); + // bin = min(v, MAX_COEFF_THRESH) + const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); + const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); + // Store. + _mm_storeu_si128((__m128i*)&out[0], bin0); + _mm_storeu_si128((__m128i*)&out[8], bin1); + } + + // Convert coefficients to bin. + for (k = 0; k < 16; ++k) { + ++distribution[out[k]]; + } } + VP8SetHistogramData(distribution, histo); +} + +//------------------------------------------------------------------------------ +// Intra predictions + +// helper for chroma-DC predictions +static WEBP_INLINE void Put8x8uv(uint8_t v, uint8_t* dst) { + int j; + const __m128i values = _mm_set1_epi8(v); + for (j = 0; j < 8; ++j) { + _mm_storel_epi64((__m128i*)(dst + j * BPS), values); + } +} + +static WEBP_INLINE void Put16(uint8_t v, uint8_t* dst) { + int j; + const __m128i values = _mm_set1_epi8(v); + for (j = 0; j < 16; ++j) { + _mm_store_si128((__m128i*)(dst + j * BPS), values); + } +} + +static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) { + if (size == 4) { + int j; + for (j = 0; j < 4; ++j) { + memset(dst + j * BPS, value, 4); + } + } else if (size == 8) { + Put8x8uv(value, dst); + } else { + Put16(value, dst); + } +} + +static WEBP_INLINE void VE8uv(uint8_t* dst, const uint8_t* top) { + int j; + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + for (j = 0; j < 8; ++j) { + _mm_storel_epi64((__m128i*)(dst + j * BPS), top_values); + } +} + +static WEBP_INLINE void VE16(uint8_t* dst, const uint8_t* top) { + const __m128i top_values = _mm_load_si128((const __m128i*)top); + int j; + for (j = 0; j < 16; ++j) { + _mm_store_si128((__m128i*)(dst + j * BPS), top_values); + } +} + +static WEBP_INLINE void VerticalPred(uint8_t* dst, + const uint8_t* top, int size) { + if (top != NULL) { + if (size == 8) { + VE8uv(dst, top); + } else { + VE16(dst, top); + } + } else { + Fill(dst, 127, size); + } +} + +static WEBP_INLINE void HE8uv(uint8_t* dst, const uint8_t* left) { + int j; + for (j = 0; j < 8; ++j) { + const __m128i values = _mm_set1_epi8(left[j]); + _mm_storel_epi64((__m128i*)dst, values); + dst += BPS; + } +} + +static WEBP_INLINE void HE16(uint8_t* dst, const uint8_t* left) { + int j; + for (j = 0; j < 16; ++j) { + const __m128i values = _mm_set1_epi8(left[j]); + _mm_store_si128((__m128i*)dst, values); + dst += BPS; + } +} + +static WEBP_INLINE void HorizontalPred(uint8_t* dst, + const uint8_t* left, int size) { + if (left != NULL) { + if (size == 8) { + HE8uv(dst, left); + } else { + HE16(dst, left); + } + } else { + Fill(dst, 129, size); + } +} + +static WEBP_INLINE void TM(uint8_t* dst, const uint8_t* left, + const uint8_t* top, int size) { + const __m128i zero = _mm_setzero_si128(); + int y; + if (size == 8) { + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + const __m128i top_base = _mm_unpacklo_epi8(top_values, zero); + for (y = 0; y < 8; ++y, dst += BPS) { + const int val = left[y] - left[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero); + _mm_storel_epi64((__m128i*)dst, out); + } + } else { + const __m128i top_values = _mm_load_si128((const __m128i*)top); + const __m128i top_base_0 = _mm_unpacklo_epi8(top_values, zero); + const __m128i top_base_1 = _mm_unpackhi_epi8(top_values, zero); + for (y = 0; y < 16; ++y, dst += BPS) { + const int val = left[y] - left[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out_0 = _mm_add_epi16(base, top_base_0); + const __m128i out_1 = _mm_add_epi16(base, top_base_1); + const __m128i out = _mm_packus_epi16(out_0, out_1); + _mm_store_si128((__m128i*)dst, out); + } + } +} + +static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left, + const uint8_t* top, int size) { + if (left != NULL) { + if (top != NULL) { + TM(dst, left, top, size); + } else { + HorizontalPred(dst, left, size); + } + } else { + // true motion without left samples (hence: with default 129 value) + // is equivalent to VE prediction where you just copy the top samples. + // Note that if top samples are not available, the default value is + // then 129, and not 127 as in the VerticalPred case. + if (top != NULL) { + VerticalPred(dst, top, size); + } else { + Fill(dst, 129, size); + } + } +} + +static WEBP_INLINE void DC8uv(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + const __m128i zero = _mm_setzero_si128(); + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + const __m128i left_values = _mm_loadl_epi64((const __m128i*)left); + const __m128i sum_top = _mm_sad_epu8(top_values, zero); + const __m128i sum_left = _mm_sad_epu8(left_values, zero); + const int DC = _mm_cvtsi128_si32(sum_top) + _mm_cvtsi128_si32(sum_left) + 8; + Put8x8uv(DC >> 4, dst); +} + +static WEBP_INLINE void DC8uvNoLeft(uint8_t* dst, const uint8_t* top) { + const __m128i zero = _mm_setzero_si128(); + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + const __m128i sum = _mm_sad_epu8(top_values, zero); + const int DC = _mm_cvtsi128_si32(sum) + 4; + Put8x8uv(DC >> 3, dst); +} + +static WEBP_INLINE void DC8uvNoTop(uint8_t* dst, const uint8_t* left) { + // 'left' is contiguous so we can reuse the top summation. + DC8uvNoLeft(dst, left); +} + +static WEBP_INLINE void DC8uvNoTopLeft(uint8_t* dst) { + Put8x8uv(0x80, dst); +} + +static WEBP_INLINE void DC8uvMode(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + if (top != NULL) { + if (left != NULL) { // top and left present + DC8uv(dst, left, top); + } else { // top, but no left + DC8uvNoLeft(dst, top); + } + } else if (left != NULL) { // left but no top + DC8uvNoTop(dst, left); + } else { // no top, no left, nothing. + DC8uvNoTopLeft(dst); + } +} + +static WEBP_INLINE void DC16(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + const __m128i zero = _mm_setzero_si128(); + const __m128i top_row = _mm_load_si128((const __m128i*)top); + const __m128i left_row = _mm_load_si128((const __m128i*)left); + const __m128i sad8x2 = _mm_sad_epu8(top_row, zero); + // sum the two sads: sad8x2[0:1] + sad8x2[8:9] + const __m128i sum_top = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2)); + const __m128i sad8x2_left = _mm_sad_epu8(left_row, zero); + // sum the two sads: sad8x2[0:1] + sad8x2[8:9] + const __m128i sum_left = + _mm_add_epi16(sad8x2_left, _mm_shuffle_epi32(sad8x2_left, 2)); + const int DC = _mm_cvtsi128_si32(sum_top) + _mm_cvtsi128_si32(sum_left) + 16; + Put16(DC >> 5, dst); +} + +static WEBP_INLINE void DC16NoLeft(uint8_t* dst, const uint8_t* top) { + const __m128i zero = _mm_setzero_si128(); + const __m128i top_row = _mm_load_si128((const __m128i*)top); + const __m128i sad8x2 = _mm_sad_epu8(top_row, zero); + // sum the two sads: sad8x2[0:1] + sad8x2[8:9] + const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2)); + const int DC = _mm_cvtsi128_si32(sum) + 8; + Put16(DC >> 4, dst); +} + +static WEBP_INLINE void DC16NoTop(uint8_t* dst, const uint8_t* left) { + // 'left' is contiguous so we can reuse the top summation. + DC16NoLeft(dst, left); +} + +static WEBP_INLINE void DC16NoTopLeft(uint8_t* dst) { + Put16(0x80, dst); +} + +static WEBP_INLINE void DC16Mode(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + if (top != NULL) { + if (left != NULL) { // top and left present + DC16(dst, left, top); + } else { // top, but no left + DC16NoLeft(dst, top); + } + } else if (left != NULL) { // left but no top + DC16NoTop(dst, left); + } else { // no top, no left, nothing. + DC16NoTopLeft(dst); + } +} + +//------------------------------------------------------------------------------ +// 4x4 predictions + +#define DST(x, y) dst[(x) + (y) * BPS] +#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2) +#define AVG2(a, b) (((a) + (b) + 1) >> 1) + +// We use the following 8b-arithmetic tricks: +// (a + 2 * b + c + 2) >> 2 = (AC + b + 1) >> 1 +// where: AC = (a + c) >> 1 = [(a + c + 1) >> 1] - [(a^c) & 1] +// and: +// (a + 2 * b + c + 2) >> 2 = (AB + BC + 1) >> 1 - (ab|bc)&lsb +// where: AC = (a + b + 1) >> 1, BC = (b + c + 1) >> 1 +// and ab = a ^ b, bc = b ^ c, lsb = (AC^BC)&1 + +static WEBP_INLINE void VE4(uint8_t* dst, const uint8_t* top) { // vertical + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(top - 1)); + const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2); + const __m128i a = _mm_avg_epu8(ABCDEFGH, CDEFGH00); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one); + const __m128i b = _mm_subs_epu8(a, lsb); + const __m128i avg = _mm_avg_epu8(b, BCDEFGH0); + const uint32_t vals = _mm_cvtsi128_si32(avg); + int i; + for (i = 0; i < 4; ++i) { + *(uint32_t*)(dst + i * BPS) = vals; + } +} + +static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) { // horizontal + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + *(uint32_t*)(dst + 0 * BPS) = 0x01010101U * AVG3(X, I, J); + *(uint32_t*)(dst + 1 * BPS) = 0x01010101U * AVG3(I, J, K); + *(uint32_t*)(dst + 2 * BPS) = 0x01010101U * AVG3(J, K, L); + *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(K, L, L); +} + +static WEBP_INLINE void DC4(uint8_t* dst, const uint8_t* top) { + uint32_t dc = 4; + int i; + for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i]; + Fill(dst, dc >> 3, 4); +} + +static WEBP_INLINE void LD4(uint8_t* dst, const uint8_t* top) { // Down-Left + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top); + const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2); + const __m128i CDEFGHH0 = _mm_insert_epi16(CDEFGH00, top[7], 3); + const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, CDEFGHH0); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0); + *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( abcdefg ); + *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)); + *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)); + *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)); +} + +static WEBP_INLINE void VR4(uint8_t* dst, + const uint8_t* top) { // Vertical-Right + const __m128i one = _mm_set1_epi8(1); + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int X = top[-1]; + const __m128i XABCD = _mm_loadl_epi64((const __m128i*)(top - 1)); + const __m128i ABCD0 = _mm_srli_si128(XABCD, 1); + const __m128i abcd = _mm_avg_epu8(XABCD, ABCD0); + const __m128i _XABCD = _mm_slli_si128(XABCD, 1); + const __m128i IXABCD = _mm_insert_epi16(_XABCD, I | (X << 8), 0); + const __m128i avg1 = _mm_avg_epu8(IXABCD, ABCD0); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i efgh = _mm_avg_epu8(avg2, XABCD); + *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( abcd ); + *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32( efgh ); + *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)); + *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)); + + // these two are hard to implement in SSE2, so we keep the C-version: + DST(0, 2) = AVG3(J, I, X); + DST(0, 3) = AVG3(K, J, I); +} + +static WEBP_INLINE void VL4(uint8_t* dst, + const uint8_t* top) { // Vertical-Left + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top); + const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH__ = _mm_srli_si128(ABCDEFGH, 2); + const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, BCDEFGH_); + const __m128i avg2 = _mm_avg_epu8(CDEFGH__, BCDEFGH_); + const __m128i avg3 = _mm_avg_epu8(avg1, avg2); + const __m128i lsb1 = _mm_and_si128(_mm_xor_si128(avg1, avg2), one); + const __m128i ab = _mm_xor_si128(ABCDEFGH, BCDEFGH_); + const __m128i bc = _mm_xor_si128(CDEFGH__, BCDEFGH_); + const __m128i abbc = _mm_or_si128(ab, bc); + const __m128i lsb2 = _mm_and_si128(abbc, lsb1); + const __m128i avg4 = _mm_subs_epu8(avg3, lsb2); + const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4)); + *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( avg1 ); + *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32( avg4 ); + *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)); + *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)); + + // these two are hard to get and irregular + DST(3, 2) = (extra_out >> 0) & 0xff; + DST(3, 3) = (extra_out >> 8) & 0xff; +} + +static WEBP_INLINE void RD4(uint8_t* dst, const uint8_t* top) { // Down-right + const __m128i one = _mm_set1_epi8(1); + const __m128i LKJIXABC = _mm_loadl_epi64((const __m128i*)(top - 5)); + const __m128i LKJIXABCD = _mm_insert_epi16(LKJIXABC, top[3], 4); + const __m128i KJIXABCD_ = _mm_srli_si128(LKJIXABCD, 1); + const __m128i JIXABCD__ = _mm_srli_si128(LKJIXABCD, 2); + const __m128i avg1 = _mm_avg_epu8(JIXABCD__, LKJIXABCD); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_); + *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32( abcdefg ); + *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)); + *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)); + *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)); +} + +static WEBP_INLINE void HU4(uint8_t* dst, const uint8_t* top) { + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + DST(0, 0) = AVG2(I, J); + DST(2, 0) = DST(0, 1) = AVG2(J, K); + DST(2, 1) = DST(0, 2) = AVG2(K, L); + DST(1, 0) = AVG3(I, J, K); + DST(3, 0) = DST(1, 1) = AVG3(J, K, L); + DST(3, 1) = DST(1, 2) = AVG3(K, L, L); + DST(3, 2) = DST(2, 2) = + DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L; +} + +static WEBP_INLINE void HD4(uint8_t* dst, const uint8_t* top) { + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + + DST(0, 0) = DST(2, 1) = AVG2(I, X); + DST(0, 1) = DST(2, 2) = AVG2(J, I); + DST(0, 2) = DST(2, 3) = AVG2(K, J); + DST(0, 3) = AVG2(L, K); + + DST(3, 0) = AVG3(A, B, C); + DST(2, 0) = AVG3(X, A, B); + DST(1, 0) = DST(3, 1) = AVG3(I, X, A); + DST(1, 1) = DST(3, 2) = AVG3(J, I, X); + DST(1, 2) = DST(3, 3) = AVG3(K, J, I); + DST(1, 3) = AVG3(L, K, J); +} + +static WEBP_INLINE void TM4(uint8_t* dst, const uint8_t* top) { + const __m128i zero = _mm_setzero_si128(); + const __m128i top_values = _mm_cvtsi32_si128(MemToUint32(top)); + const __m128i top_base = _mm_unpacklo_epi8(top_values, zero); + int y; + for (y = 0; y < 4; ++y, dst += BPS) { + const int val = top[-2 - y] - top[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero); + *(int*)dst = _mm_cvtsi128_si32(out); + } +} + +#undef DST +#undef AVG3 +#undef AVG2 + +//------------------------------------------------------------------------------ +// luma 4x4 prediction + +// Left samples are top[-5 .. -2], top_left is top[-1], top are +// located at top[0..3], and top right is top[4..7] +static void Intra4Preds(uint8_t* dst, const uint8_t* top) { + DC4(I4DC4 + dst, top); + TM4(I4TM4 + dst, top); + VE4(I4VE4 + dst, top); + HE4(I4HE4 + dst, top); + RD4(I4RD4 + dst, top); + VR4(I4VR4 + dst, top); + LD4(I4LD4 + dst, top); + VL4(I4VL4 + dst, top); + HD4(I4HD4 + dst, top); + HU4(I4HU4 + dst, top); +} + +//------------------------------------------------------------------------------ +// Chroma 8x8 prediction (paragraph 12.2) + +static void IntraChromaPreds(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + // U block + DC8uvMode(C8DC8 + dst, left, top); + VerticalPred(C8VE8 + dst, top, 8); + HorizontalPred(C8HE8 + dst, left, 8); + TrueMotion(C8TM8 + dst, left, top, 8); + // V block + dst += 8; + if (top != NULL) top += 8; + if (left != NULL) left += 16; + DC8uvMode(C8DC8 + dst, left, top); + VerticalPred(C8VE8 + dst, top, 8); + HorizontalPred(C8HE8 + dst, left, 8); + TrueMotion(C8TM8 + dst, left, top, 8); +} + +//------------------------------------------------------------------------------ +// luma 16x16 prediction (paragraph 12.3) + +static void Intra16Preds(uint8_t* dst, + const uint8_t* left, const uint8_t* top) { + DC16Mode(I16DC16 + dst, left, top); + VerticalPred(I16VE16 + dst, top, 16); + HorizontalPred(I16HE16 + dst, left, 16); + TrueMotion(I16TM16 + dst, left, top, 16); } //------------------------------------------------------------------------------ // Metric -static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) { - const __m128i zero = _mm_set1_epi16(0); +static WEBP_INLINE void SubtractAndAccumulate(const __m128i a, const __m128i b, + __m128i* const sum) { + // take abs(a-b) in 8b + const __m128i a_b = _mm_subs_epu8(a, b); + const __m128i b_a = _mm_subs_epu8(b, a); + const __m128i abs_a_b = _mm_or_si128(a_b, b_a); + // zero-extend to 16b + const __m128i zero = _mm_setzero_si128(); + const __m128i C0 = _mm_unpacklo_epi8(abs_a_b, zero); + const __m128i C1 = _mm_unpackhi_epi8(abs_a_b, zero); + // multiply with self + const __m128i sum1 = _mm_madd_epi16(C0, C0); + const __m128i sum2 = _mm_madd_epi16(C1, C1); + *sum = _mm_add_epi32(sum1, sum2); +} - // Load values. - const __m128i a0 = _mm_loadl_epi64((__m128i*)&a[BPS * 0]); - const __m128i a1 = _mm_loadl_epi64((__m128i*)&a[BPS * 1]); - const __m128i a2 = _mm_loadl_epi64((__m128i*)&a[BPS * 2]); - const __m128i a3 = _mm_loadl_epi64((__m128i*)&a[BPS * 3]); - const __m128i b0 = _mm_loadl_epi64((__m128i*)&b[BPS * 0]); - const __m128i b1 = _mm_loadl_epi64((__m128i*)&b[BPS * 1]); - const __m128i b2 = _mm_loadl_epi64((__m128i*)&b[BPS * 2]); - const __m128i b3 = _mm_loadl_epi64((__m128i*)&b[BPS * 3]); +static WEBP_INLINE int SSE_16xN(const uint8_t* a, const uint8_t* b, + int num_pairs) { + __m128i sum = _mm_setzero_si128(); + int32_t tmp[4]; + int i; + + for (i = 0; i < num_pairs; ++i) { + const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[BPS * 0]); + const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[BPS * 0]); + const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[BPS * 1]); + const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[BPS * 1]); + __m128i sum1, sum2; + SubtractAndAccumulate(a0, b0, &sum1); + SubtractAndAccumulate(a1, b1, &sum2); + sum = _mm_add_epi32(sum, _mm_add_epi32(sum1, sum2)); + a += 2 * BPS; + b += 2 * BPS; + } + _mm_storeu_si128((__m128i*)tmp, sum); + return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); +} + +static int SSE16x16(const uint8_t* a, const uint8_t* b) { + return SSE_16xN(a, b, 8); +} - // Combine pair of lines and convert to 16b. +static int SSE16x8(const uint8_t* a, const uint8_t* b) { + return SSE_16xN(a, b, 4); +} + +#define LOAD_8x16b(ptr) \ + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i*)(ptr)), zero) + +static int SSE8x8(const uint8_t* a, const uint8_t* b) { + const __m128i zero = _mm_setzero_si128(); + int num_pairs = 4; + __m128i sum = zero; + int32_t tmp[4]; + while (num_pairs-- > 0) { + const __m128i a0 = LOAD_8x16b(&a[BPS * 0]); + const __m128i a1 = LOAD_8x16b(&a[BPS * 1]); + const __m128i b0 = LOAD_8x16b(&b[BPS * 0]); + const __m128i b1 = LOAD_8x16b(&b[BPS * 1]); + // subtract + const __m128i c0 = _mm_subs_epi16(a0, b0); + const __m128i c1 = _mm_subs_epi16(a1, b1); + // multiply/accumulate with self + const __m128i d0 = _mm_madd_epi16(c0, c0); + const __m128i d1 = _mm_madd_epi16(c1, c1); + // collect + const __m128i sum01 = _mm_add_epi32(d0, d1); + sum = _mm_add_epi32(sum, sum01); + a += 2 * BPS; + b += 2 * BPS; + } + _mm_storeu_si128((__m128i*)tmp, sum); + return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); +} +#undef LOAD_8x16b + +static int SSE4x4(const uint8_t* a, const uint8_t* b) { + const __m128i zero = _mm_setzero_si128(); + + // Load values. Note that we read 8 pixels instead of 4, + // but the a/b buffers are over-allocated to that effect. + const __m128i a0 = _mm_loadl_epi64((const __m128i*)&a[BPS * 0]); + const __m128i a1 = _mm_loadl_epi64((const __m128i*)&a[BPS * 1]); + const __m128i a2 = _mm_loadl_epi64((const __m128i*)&a[BPS * 2]); + const __m128i a3 = _mm_loadl_epi64((const __m128i*)&a[BPS * 3]); + const __m128i b0 = _mm_loadl_epi64((const __m128i*)&b[BPS * 0]); + const __m128i b1 = _mm_loadl_epi64((const __m128i*)&b[BPS * 1]); + const __m128i b2 = _mm_loadl_epi64((const __m128i*)&b[BPS * 2]); + const __m128i b3 = _mm_loadl_epi64((const __m128i*)&b[BPS * 3]); + // Combine pair of lines. const __m128i a01 = _mm_unpacklo_epi32(a0, a1); const __m128i a23 = _mm_unpacklo_epi32(a2, a3); const __m128i b01 = _mm_unpacklo_epi32(b0, b1); const __m128i b23 = _mm_unpacklo_epi32(b2, b3); + // Convert to 16b. const __m128i a01s = _mm_unpacklo_epi8(a01, zero); const __m128i a23s = _mm_unpacklo_epi8(a23, zero); const __m128i b01s = _mm_unpacklo_epi8(b01, zero); const __m128i b23s = _mm_unpacklo_epi8(b23, zero); + // subtract, square and accumulate + const __m128i d0 = _mm_subs_epi16(a01s, b01s); + const __m128i d1 = _mm_subs_epi16(a23s, b23s); + const __m128i e0 = _mm_madd_epi16(d0, d0); + const __m128i e1 = _mm_madd_epi16(d1, d1); + const __m128i sum = _mm_add_epi32(e0, e1); - // Compute differences; (a-b)^2 = (abs(a-b))^2 = (sat8(a-b) + sat8(b-a))^2 - // TODO(cduvivier): Dissassemble and figure out why this is fastest. We don't - // need absolute values, there is no need to do calculation - // in 8bit as we are already in 16bit, ... Yet this is what - // benchmarks the fastest! - const __m128i d0 = _mm_subs_epu8(a01s, b01s); - const __m128i d1 = _mm_subs_epu8(b01s, a01s); - const __m128i d2 = _mm_subs_epu8(a23s, b23s); - const __m128i d3 = _mm_subs_epu8(b23s, a23s); - - // Square and add them all together. - const __m128i madd0 = _mm_madd_epi16(d0, d0); - const __m128i madd1 = _mm_madd_epi16(d1, d1); - const __m128i madd2 = _mm_madd_epi16(d2, d2); - const __m128i madd3 = _mm_madd_epi16(d3, d3); - const __m128i sum0 = _mm_add_epi32(madd0, madd1); - const __m128i sum1 = _mm_add_epi32(madd2, madd3); - const __m128i sum2 = _mm_add_epi32(sum0, sum1); int32_t tmp[4]; - _mm_storeu_si128((__m128i*)tmp, sum2); + _mm_storeu_si128((__m128i*)tmp, sum); return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); } @@ -497,24 +1155,22 @@ static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) { // Hadamard transform // Returns the difference between the weighted sum of the absolute value of // transformed coefficients. -static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB, - const uint16_t* const w) { +static int TTransform(const uint8_t* inA, const uint8_t* inB, + const uint16_t* const w) { int32_t sum[4]; __m128i tmp_0, tmp_1, tmp_2, tmp_3; const __m128i zero = _mm_setzero_si128(); - const __m128i one = _mm_set1_epi16(1); - const __m128i three = _mm_set1_epi16(3); - // Load, combine and tranpose inputs. + // Load, combine and transpose inputs. { - const __m128i inA_0 = _mm_loadl_epi64((__m128i*)&inA[BPS * 0]); - const __m128i inA_1 = _mm_loadl_epi64((__m128i*)&inA[BPS * 1]); - const __m128i inA_2 = _mm_loadl_epi64((__m128i*)&inA[BPS * 2]); - const __m128i inA_3 = _mm_loadl_epi64((__m128i*)&inA[BPS * 3]); - const __m128i inB_0 = _mm_loadl_epi64((__m128i*)&inB[BPS * 0]); - const __m128i inB_1 = _mm_loadl_epi64((__m128i*)&inB[BPS * 1]); - const __m128i inB_2 = _mm_loadl_epi64((__m128i*)&inB[BPS * 2]); - const __m128i inB_3 = _mm_loadl_epi64((__m128i*)&inB[BPS * 3]); + const __m128i inA_0 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 0]); + const __m128i inA_1 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 1]); + const __m128i inA_2 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 2]); + const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]); + const __m128i inB_0 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 0]); + const __m128i inB_1 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 1]); + const __m128i inB_2 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 2]); + const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]); // Combine inA and inB (we'll do two transforms in parallel). const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0); @@ -550,17 +1206,14 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB, // Horizontal pass and subsequent transpose. { // Calculate a and b (two 4x4 at once). - const __m128i a0 = _mm_slli_epi16(_mm_add_epi16(tmp_0, tmp_2), 2); - const __m128i a1 = _mm_slli_epi16(_mm_add_epi16(tmp_1, tmp_3), 2); - const __m128i a2 = _mm_slli_epi16(_mm_sub_epi16(tmp_1, tmp_3), 2); - const __m128i a3 = _mm_slli_epi16(_mm_sub_epi16(tmp_0, tmp_2), 2); - // b0_extra = (a0 != 0); - const __m128i b0_extra = _mm_andnot_si128(_mm_cmpeq_epi16 (a0, zero), one); - const __m128i b0_base = _mm_add_epi16(a0, a1); + const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); + const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); + const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); + const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); + const __m128i b0 = _mm_add_epi16(a0, a1); const __m128i b1 = _mm_add_epi16(a3, a2); const __m128i b2 = _mm_sub_epi16(a3, a2); const __m128i b3 = _mm_sub_epi16(a0, a1); - const __m128i b0 = _mm_add_epi16(b0_base, b0_extra); // a00 a01 a02 a03 b00 b01 b02 b03 // a10 a11 a12 a13 b10 b11 b12 b13 // a20 a21 a22 a23 b20 b21 b22 b23 @@ -598,8 +1251,8 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB, // Load all inputs. // TODO(cduvivier): Make variable declarations and allocations aligned so // we can use _mm_load_si128 instead of _mm_loadu_si128. - const __m128i w_0 = _mm_loadu_si128((__m128i*)&w[0]); - const __m128i w_8 = _mm_loadu_si128((__m128i*)&w[8]); + const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]); + const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]); // Calculate a and b (two 4x4 at once). const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); @@ -618,36 +1271,16 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB, __m128i B_b2 = _mm_unpackhi_epi64(b2, b3); { - // sign(b) = b >> 15 (0x0000 if positive, 0xffff if negative) - const __m128i sign_A_b0 = _mm_srai_epi16(A_b0, 15); - const __m128i sign_A_b2 = _mm_srai_epi16(A_b2, 15); - const __m128i sign_B_b0 = _mm_srai_epi16(B_b0, 15); - const __m128i sign_B_b2 = _mm_srai_epi16(B_b2, 15); - - // b = abs(b) = (b ^ sign) - sign - A_b0 = _mm_xor_si128(A_b0, sign_A_b0); - A_b2 = _mm_xor_si128(A_b2, sign_A_b2); - B_b0 = _mm_xor_si128(B_b0, sign_B_b0); - B_b2 = _mm_xor_si128(B_b2, sign_B_b2); - A_b0 = _mm_sub_epi16(A_b0, sign_A_b0); - A_b2 = _mm_sub_epi16(A_b2, sign_A_b2); - B_b0 = _mm_sub_epi16(B_b0, sign_B_b0); - B_b2 = _mm_sub_epi16(B_b2, sign_B_b2); + const __m128i d0 = _mm_sub_epi16(zero, A_b0); + const __m128i d1 = _mm_sub_epi16(zero, A_b2); + const __m128i d2 = _mm_sub_epi16(zero, B_b0); + const __m128i d3 = _mm_sub_epi16(zero, B_b2); + A_b0 = _mm_max_epi16(A_b0, d0); // abs(v), 16b + A_b2 = _mm_max_epi16(A_b2, d1); + B_b0 = _mm_max_epi16(B_b0, d2); + B_b2 = _mm_max_epi16(B_b2, d3); } - // b = abs(b) + 3 - A_b0 = _mm_add_epi16(A_b0, three); - A_b2 = _mm_add_epi16(A_b2, three); - B_b0 = _mm_add_epi16(B_b0, three); - B_b2 = _mm_add_epi16(B_b2, three); - - // abs((b + (b<0) + 3) >> 3) = (abs(b) + 3) >> 3 - // b = (abs(b) + 3) >> 3 - A_b0 = _mm_srai_epi16(A_b0, 3); - A_b2 = _mm_srai_epi16(A_b2, 3); - B_b0 = _mm_srai_epi16(B_b0, 3); - B_b2 = _mm_srai_epi16(B_b2, 3); - // weighted sums A_b0 = _mm_madd_epi16(A_b0, w_0); A_b2 = _mm_madd_epi16(A_b2, w_8); @@ -663,35 +1296,33 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB, return sum[0] + sum[1] + sum[2] + sum[3]; } -static int Disto4x4SSE2(const uint8_t* const a, const uint8_t* const b, - const uint16_t* const w) { - const int diff_sum = TTransformSSE2(a, b, w); - return (abs(diff_sum) + 8) >> 4; +static int Disto4x4(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + const int diff_sum = TTransform(a, b, w); + return abs(diff_sum) >> 5; } -static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b, - const uint16_t* const w) { +static int Disto16x16(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { int D = 0; int x, y; for (y = 0; y < 16 * BPS; y += 4 * BPS) { for (x = 0; x < 16; x += 4) { - D += Disto4x4SSE2(a + x + y, b + x + y, w); + D += Disto4x4(a + x + y, b + x + y, w); } } return D; } - //------------------------------------------------------------------------------ // Quantization // -// Simple quantization -static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16], - int n, const VP8Matrix* const mtx) { - const __m128i max_coeff_2047 = _mm_set1_epi16(2047); - const __m128i zero = _mm_set1_epi16(0); - __m128i sign0, sign8; +static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16], + const uint16_t* const sharpen, + const VP8Matrix* const mtx) { + const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL); + const __m128i zero = _mm_setzero_si128(); __m128i coeff0, coeff8; __m128i out0, out8; __m128i packed_out; @@ -701,20 +1332,14 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16], // we can use _mm_load_si128 instead of _mm_loadu_si128. __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]); __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]); - const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[0]); - const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[8]); - const __m128i iq0 = _mm_loadu_si128((__m128i*)&mtx->iq_[0]); - const __m128i iq8 = _mm_loadu_si128((__m128i*)&mtx->iq_[8]); - const __m128i bias0 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]); - const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]); - const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]); - const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]); - const __m128i zthresh0 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[0]); - const __m128i zthresh8 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[8]); - - // sign(in) = in >> 15 (0x0000 if positive, 0xffff if negative) - sign0 = _mm_srai_epi16(in0, 15); - sign8 = _mm_srai_epi16(in8, 15); + const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]); + const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]); + const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]); + const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]); + + // extract sign(in) (0x0000 if positive, 0xffff if negative) + const __m128i sign0 = _mm_cmpgt_epi16(zero, in0); + const __m128i sign8 = _mm_cmpgt_epi16(zero, in8); // coeff = abs(in) = (in ^ sign) - sign coeff0 = _mm_xor_si128(in0, sign0); @@ -723,43 +1348,47 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16], coeff8 = _mm_sub_epi16(coeff8, sign8); // coeff = abs(in) + sharpen - coeff0 = _mm_add_epi16(coeff0, sharpen0); - coeff8 = _mm_add_epi16(coeff8, sharpen8); - - // if (coeff > 2047) coeff = 2047 - coeff0 = _mm_min_epi16(coeff0, max_coeff_2047); - coeff8 = _mm_min_epi16(coeff8, max_coeff_2047); + if (sharpen != NULL) { + const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]); + const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]); + coeff0 = _mm_add_epi16(coeff0, sharpen0); + coeff8 = _mm_add_epi16(coeff8, sharpen8); + } - // out = (coeff * iQ + B) >> QFIX; + // out = (coeff * iQ + B) >> QFIX { // doing calculations with 32b precision (QFIX=17) // out = (coeff * iQ) - __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0); - __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0); - __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8); - __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8); + const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0); + const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0); + const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8); + const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8); __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H); __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H); __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H); __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H); - // expand bias from 16b to 32b - __m128i bias_00 = _mm_unpacklo_epi16(bias0, zero); - __m128i bias_04 = _mm_unpackhi_epi16(bias0, zero); - __m128i bias_08 = _mm_unpacklo_epi16(bias8, zero); - __m128i bias_12 = _mm_unpackhi_epi16(bias8, zero); // out = (coeff * iQ + B) + const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]); + const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]); + const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]); + const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]); out_00 = _mm_add_epi32(out_00, bias_00); out_04 = _mm_add_epi32(out_04, bias_04); out_08 = _mm_add_epi32(out_08, bias_08); out_12 = _mm_add_epi32(out_12, bias_12); - // out = (coeff * iQ + B) >> QFIX; + // out = QUANTDIV(coeff, iQ, B, QFIX) out_00 = _mm_srai_epi32(out_00, QFIX); out_04 = _mm_srai_epi32(out_04, QFIX); out_08 = _mm_srai_epi32(out_08, QFIX); out_12 = _mm_srai_epi32(out_12, QFIX); + // pack result as 16b out0 = _mm_packs_epi32(out_00, out_04); out8 = _mm_packs_epi32(out_08, out_12); + + // if (coeff > 2047) coeff = 2047 + out0 = _mm_min_epi16(out0, max_coeff_2047); + out8 = _mm_min_epi16(out8, max_coeff_2047); } // get sign back (if (sign[j]) out_n = -out_n) @@ -772,17 +1401,8 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16], in0 = _mm_mullo_epi16(out0, q0); in8 = _mm_mullo_epi16(out8, q8); - // if (coeff <= mtx->zthresh_) {in=0; out=0;} - { - __m128i cmp0 = _mm_cmpgt_epi16(coeff0, zthresh0); - __m128i cmp8 = _mm_cmpgt_epi16(coeff8, zthresh8); - in0 = _mm_and_si128(in0, cmp0); - in8 = _mm_and_si128(in8, cmp8); - _mm_storeu_si128((__m128i*)&in[0], in0); - _mm_storeu_si128((__m128i*)&in[8], in8); - out0 = _mm_and_si128(out0, cmp0); - out8 = _mm_and_si128(out8, cmp8); - } + _mm_storeu_si128((__m128i*)&in[0], in0); + _mm_storeu_si128((__m128i*)&in[8], in8); // zigzag the output before storing it. // @@ -809,29 +1429,55 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16], } // detect if all 'out' values are zeroes or not - { - int32_t tmp[4]; - _mm_storeu_si128((__m128i*)tmp, packed_out); - if (n) { - tmp[0] &= ~0xff; - } - return (tmp[3] || tmp[2] || tmp[1] || tmp[0]); - } + return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff); +} + +static int QuantizeBlock(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + return DoQuantizeBlock(in, out, &mtx->sharpen_[0], mtx); +} + +static int QuantizeBlockWHT(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + return DoQuantizeBlock(in, out, NULL, mtx); +} + +static int Quantize2Blocks(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + const uint16_t* const sharpen = &mtx->sharpen_[0]; + nz = DoQuantizeBlock(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0; + nz |= DoQuantizeBlock(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1; + return nz; } +//------------------------------------------------------------------------------ +// Entry point + extern void VP8EncDspInitSSE2(void); -void VP8EncDspInitSSE2(void) { - VP8CollectHistogram = CollectHistogramSSE2; - VP8EncQuantizeBlock = QuantizeBlockSSE2; - VP8ITransform = ITransformSSE2; - VP8FTransform = FTransformSSE2; - VP8SSE4x4 = SSE4x4SSE2; - VP8TDisto4x4 = Disto4x4SSE2; - VP8TDisto16x16 = Disto16x16SSE2; -} - -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" -#endif -#endif // WEBP_USE_SSE2 +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE2(void) { + VP8CollectHistogram = CollectHistogram; + VP8EncPredLuma16 = Intra16Preds; + VP8EncPredChroma8 = IntraChromaPreds; + VP8EncPredLuma4 = Intra4Preds; + VP8EncQuantizeBlock = QuantizeBlock; + VP8EncQuantize2Blocks = Quantize2Blocks; + VP8EncQuantizeBlockWHT = QuantizeBlockWHT; + VP8ITransform = ITransform; + VP8FTransform = FTransform; + VP8FTransform2 = FTransform2; + VP8FTransformWHT = FTransformWHT; + VP8SSE16x16 = SSE16x16; + VP8SSE16x8 = SSE16x8; + VP8SSE8x8 = SSE8x8; + VP8SSE4x4 = SSE4x4; + VP8TDisto4x4 = Disto4x4; + VP8TDisto16x16 = Disto16x16; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8EncDspInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/enc_sse41.c b/drivers/webp/dsp/enc_sse41.c new file mode 100644 index 0000000000..27f4189833 --- /dev/null +++ b/drivers/webp/dsp/enc_sse41.c @@ -0,0 +1,375 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE4 version of some encoding functions. +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE41) +#include <smmintrin.h> +#include <stdlib.h> // for abs() + +#include "../enc/vp8enci.h" + +//------------------------------------------------------------------------------ +// Compute susceptibility based on DCT-coeff histograms. + +static void CollectHistogram(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + int k; + + VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + + // Convert coefficients to bin (within out[]). + { + // Load. + const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); + const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); + // v = abs(out) >> 3 + const __m128i abs0 = _mm_abs_epi16(out0); + const __m128i abs1 = _mm_abs_epi16(out1); + const __m128i v0 = _mm_srai_epi16(abs0, 3); + const __m128i v1 = _mm_srai_epi16(abs1, 3); + // bin = min(v, MAX_COEFF_THRESH) + const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); + const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); + // Store. + _mm_storeu_si128((__m128i*)&out[0], bin0); + _mm_storeu_si128((__m128i*)&out[8], bin1); + } + + // Convert coefficients to bin. + for (k = 0; k < 16; ++k) { + ++distribution[out[k]]; + } + } + VP8SetHistogramData(distribution, histo); +} + +//------------------------------------------------------------------------------ +// Texture distortion +// +// We try to match the spectral content (weighted) between source and +// reconstructed samples. + +// Hadamard transform +// Returns the difference between the weighted sum of the absolute value of +// transformed coefficients. +static int TTransform(const uint8_t* inA, const uint8_t* inB, + const uint16_t* const w) { + __m128i tmp_0, tmp_1, tmp_2, tmp_3; + + // Load, combine and transpose inputs. + { + const __m128i inA_0 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 0]); + const __m128i inA_1 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 1]); + const __m128i inA_2 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 2]); + const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]); + const __m128i inB_0 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 0]); + const __m128i inB_1 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 1]); + const __m128i inB_2 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 2]); + const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]); + + // Combine inA and inB (we'll do two transforms in parallel). + const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0); + const __m128i inAB_1 = _mm_unpacklo_epi8(inA_1, inB_1); + const __m128i inAB_2 = _mm_unpacklo_epi8(inA_2, inB_2); + const __m128i inAB_3 = _mm_unpacklo_epi8(inA_3, inB_3); + // a00 b00 a01 b01 a02 b03 a03 b03 0 0 0 0 0 0 0 0 + // a10 b10 a11 b11 a12 b12 a13 b13 0 0 0 0 0 0 0 0 + // a20 b20 a21 b21 a22 b22 a23 b23 0 0 0 0 0 0 0 0 + // a30 b30 a31 b31 a32 b32 a33 b33 0 0 0 0 0 0 0 0 + + // Transpose the two 4x4, discarding the filling zeroes. + const __m128i transpose0_0 = _mm_unpacklo_epi8(inAB_0, inAB_2); + const __m128i transpose0_1 = _mm_unpacklo_epi8(inAB_1, inAB_3); + // a00 a20 b00 b20 a01 a21 b01 b21 a02 a22 b02 b22 a03 a23 b03 b23 + // a10 a30 b10 b30 a11 a31 b11 b31 a12 a32 b12 b32 a13 a33 b13 b33 + const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1); + const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1); + // a00 a10 a20 a30 b00 b10 b20 b30 a01 a11 a21 a31 b01 b11 b21 b31 + // a02 a12 a22 a32 b02 b12 b22 b32 a03 a13 a23 a33 b03 b13 b23 b33 + + // Convert to 16b. + tmp_0 = _mm_cvtepu8_epi16(transpose1_0); + tmp_1 = _mm_cvtepu8_epi16(_mm_srli_si128(transpose1_0, 8)); + tmp_2 = _mm_cvtepu8_epi16(transpose1_1); + tmp_3 = _mm_cvtepu8_epi16(_mm_srli_si128(transpose1_1, 8)); + // a00 a10 a20 a30 b00 b10 b20 b30 + // a01 a11 a21 a31 b01 b11 b21 b31 + // a02 a12 a22 a32 b02 b12 b22 b32 + // a03 a13 a23 a33 b03 b13 b23 b33 + } + + // Horizontal pass and subsequent transpose. + { + // Calculate a and b (two 4x4 at once). + const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); + const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); + const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); + const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); + const __m128i b0 = _mm_add_epi16(a0, a1); + const __m128i b1 = _mm_add_epi16(a3, a2); + const __m128i b2 = _mm_sub_epi16(a3, a2); + const __m128i b3 = _mm_sub_epi16(a0, a1); + // a00 a01 a02 a03 b00 b01 b02 b03 + // a10 a11 a12 a13 b10 b11 b12 b13 + // a20 a21 a22 a23 b20 b21 b22 b23 + // a30 a31 a32 a33 b30 b31 b32 b33 + + // Transpose the two 4x4. + const __m128i transpose0_0 = _mm_unpacklo_epi16(b0, b1); + const __m128i transpose0_1 = _mm_unpacklo_epi16(b2, b3); + const __m128i transpose0_2 = _mm_unpackhi_epi16(b0, b1); + const __m128i transpose0_3 = _mm_unpackhi_epi16(b2, b3); + // a00 a10 a01 a11 a02 a12 a03 a13 + // a20 a30 a21 a31 a22 a32 a23 a33 + // b00 b10 b01 b11 b02 b12 b03 b13 + // b20 b30 b21 b31 b22 b32 b23 b33 + const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); + const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3); + const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); + const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3); + // a00 a10 a20 a30 a01 a11 a21 a31 + // b00 b10 b20 b30 b01 b11 b21 b31 + // a02 a12 a22 a32 a03 a13 a23 a33 + // b02 b12 a22 b32 b03 b13 b23 b33 + tmp_0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); + tmp_1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); + tmp_2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); + tmp_3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); + // a00 a10 a20 a30 b00 b10 b20 b30 + // a01 a11 a21 a31 b01 b11 b21 b31 + // a02 a12 a22 a32 b02 b12 b22 b32 + // a03 a13 a23 a33 b03 b13 b23 b33 + } + + // Vertical pass and difference of weighted sums. + { + // Load all inputs. + const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]); + const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]); + + // Calculate a and b (two 4x4 at once). + const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); + const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); + const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); + const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); + const __m128i b0 = _mm_add_epi16(a0, a1); + const __m128i b1 = _mm_add_epi16(a3, a2); + const __m128i b2 = _mm_sub_epi16(a3, a2); + const __m128i b3 = _mm_sub_epi16(a0, a1); + + // Separate the transforms of inA and inB. + __m128i A_b0 = _mm_unpacklo_epi64(b0, b1); + __m128i A_b2 = _mm_unpacklo_epi64(b2, b3); + __m128i B_b0 = _mm_unpackhi_epi64(b0, b1); + __m128i B_b2 = _mm_unpackhi_epi64(b2, b3); + + A_b0 = _mm_abs_epi16(A_b0); + A_b2 = _mm_abs_epi16(A_b2); + B_b0 = _mm_abs_epi16(B_b0); + B_b2 = _mm_abs_epi16(B_b2); + + // weighted sums + A_b0 = _mm_madd_epi16(A_b0, w_0); + A_b2 = _mm_madd_epi16(A_b2, w_8); + B_b0 = _mm_madd_epi16(B_b0, w_0); + B_b2 = _mm_madd_epi16(B_b2, w_8); + A_b0 = _mm_add_epi32(A_b0, A_b2); + B_b0 = _mm_add_epi32(B_b0, B_b2); + + // difference of weighted sums + A_b2 = _mm_sub_epi32(A_b0, B_b0); + // cascading summation of the differences + B_b0 = _mm_hadd_epi32(A_b2, A_b2); + B_b2 = _mm_hadd_epi32(B_b0, B_b0); + return _mm_cvtsi128_si32(B_b2); + } +} + +static int Disto4x4(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + const int diff_sum = TTransform(a, b, w); + return abs(diff_sum) >> 5; +} + +static int Disto16x16(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4(a + x + y, b + x + y, w); + } + } + return D; +} + +//------------------------------------------------------------------------------ +// Quantization +// + +// Generates a pshufb constant for shuffling 16b words. +#define PSHUFB_CST(A,B,C,D,E,F,G,H) \ + _mm_set_epi8(2 * (H) + 1, 2 * (H) + 0, 2 * (G) + 1, 2 * (G) + 0, \ + 2 * (F) + 1, 2 * (F) + 0, 2 * (E) + 1, 2 * (E) + 0, \ + 2 * (D) + 1, 2 * (D) + 0, 2 * (C) + 1, 2 * (C) + 0, \ + 2 * (B) + 1, 2 * (B) + 0, 2 * (A) + 1, 2 * (A) + 0) + +static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16], + const uint16_t* const sharpen, + const VP8Matrix* const mtx) { + const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL); + const __m128i zero = _mm_setzero_si128(); + __m128i out0, out8; + __m128i packed_out; + + // Load all inputs. + // TODO(cduvivier): Make variable declarations and allocations aligned so that + // we can use _mm_load_si128 instead of _mm_loadu_si128. + __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]); + __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]); + const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]); + const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]); + const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]); + const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]); + + // coeff = abs(in) + __m128i coeff0 = _mm_abs_epi16(in0); + __m128i coeff8 = _mm_abs_epi16(in8); + + // coeff = abs(in) + sharpen + if (sharpen != NULL) { + const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]); + const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]); + coeff0 = _mm_add_epi16(coeff0, sharpen0); + coeff8 = _mm_add_epi16(coeff8, sharpen8); + } + + // out = (coeff * iQ + B) >> QFIX + { + // doing calculations with 32b precision (QFIX=17) + // out = (coeff * iQ) + const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0); + const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0); + const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8); + const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8); + __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H); + __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H); + __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H); + __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H); + // out = (coeff * iQ + B) + const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]); + const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]); + const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]); + const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]); + out_00 = _mm_add_epi32(out_00, bias_00); + out_04 = _mm_add_epi32(out_04, bias_04); + out_08 = _mm_add_epi32(out_08, bias_08); + out_12 = _mm_add_epi32(out_12, bias_12); + // out = QUANTDIV(coeff, iQ, B, QFIX) + out_00 = _mm_srai_epi32(out_00, QFIX); + out_04 = _mm_srai_epi32(out_04, QFIX); + out_08 = _mm_srai_epi32(out_08, QFIX); + out_12 = _mm_srai_epi32(out_12, QFIX); + + // pack result as 16b + out0 = _mm_packs_epi32(out_00, out_04); + out8 = _mm_packs_epi32(out_08, out_12); + + // if (coeff > 2047) coeff = 2047 + out0 = _mm_min_epi16(out0, max_coeff_2047); + out8 = _mm_min_epi16(out8, max_coeff_2047); + } + + // put sign back + out0 = _mm_sign_epi16(out0, in0); + out8 = _mm_sign_epi16(out8, in8); + + // in = out * Q + in0 = _mm_mullo_epi16(out0, q0); + in8 = _mm_mullo_epi16(out8, q8); + + _mm_storeu_si128((__m128i*)&in[0], in0); + _mm_storeu_si128((__m128i*)&in[8], in8); + + // zigzag the output before storing it. The re-ordering is: + // 0 1 2 3 4 5 6 7 | 8 9 10 11 12 13 14 15 + // -> 0 1 4[8]5 2 3 6 | 9 12 13 10 [7]11 14 15 + // There's only two misplaced entries ([8] and [7]) that are crossing the + // reg's boundaries. + // We use pshufb instead of pshuflo/pshufhi. + { + const __m128i kCst_lo = PSHUFB_CST(0, 1, 4, -1, 5, 2, 3, 6); + const __m128i kCst_7 = PSHUFB_CST(-1, -1, -1, -1, 7, -1, -1, -1); + const __m128i tmp_lo = _mm_shuffle_epi8(out0, kCst_lo); + const __m128i tmp_7 = _mm_shuffle_epi8(out0, kCst_7); // extract #7 + const __m128i kCst_hi = PSHUFB_CST(1, 4, 5, 2, -1, 3, 6, 7); + const __m128i kCst_8 = PSHUFB_CST(-1, -1, -1, 0, -1, -1, -1, -1); + const __m128i tmp_hi = _mm_shuffle_epi8(out8, kCst_hi); + const __m128i tmp_8 = _mm_shuffle_epi8(out8, kCst_8); // extract #8 + const __m128i out_z0 = _mm_or_si128(tmp_lo, tmp_8); + const __m128i out_z8 = _mm_or_si128(tmp_hi, tmp_7); + _mm_storeu_si128((__m128i*)&out[0], out_z0); + _mm_storeu_si128((__m128i*)&out[8], out_z8); + packed_out = _mm_packs_epi16(out_z0, out_z8); + } + + // detect if all 'out' values are zeroes or not + return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff); +} + +#undef PSHUFB_CST + +static int QuantizeBlock(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + return DoQuantizeBlock(in, out, &mtx->sharpen_[0], mtx); +} + +static int QuantizeBlockWHT(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + return DoQuantizeBlock(in, out, NULL, mtx); +} + +static int Quantize2Blocks(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + const uint16_t* const sharpen = &mtx->sharpen_[0]; + nz = DoQuantizeBlock(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0; + nz |= DoQuantizeBlock(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1; + return nz; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitSSE41(void); +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE41(void) { + VP8CollectHistogram = CollectHistogram; + VP8EncQuantizeBlock = QuantizeBlock; + VP8EncQuantize2Blocks = Quantize2Blocks; + VP8EncQuantizeBlockWHT = QuantizeBlockWHT; + VP8TDisto4x4 = Disto4x4; + VP8TDisto16x16 = Disto16x16; +} + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(VP8EncDspInitSSE41) + +#endif // WEBP_USE_SSE41 diff --git a/drivers/webp/dsp/filters.c b/drivers/webp/dsp/filters.c new file mode 100644 index 0000000000..5c30f2e457 --- /dev/null +++ b/drivers/webp/dsp/filters.c @@ -0,0 +1,240 @@ +// Copyright 2011 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Spatial prediction using various filters +// +// Author: Urvang (urvang@google.com) + +#include "./dsp.h" +#include <assert.h> +#include <stdlib.h> +#include <string.h> + +//------------------------------------------------------------------------------ +// Helpful macro. + +# define SANITY_CHECK(in, out) \ + assert(in != NULL); \ + assert(out != NULL); \ + assert(width > 0); \ + assert(height > 0); \ + assert(stride >= width); \ + assert(row >= 0 && num_rows > 0 && row + num_rows <= height); \ + (void)height; // Silence unused warning. + +static WEBP_INLINE void PredictLine(const uint8_t* src, const uint8_t* pred, + uint8_t* dst, int length, int inverse) { + int i; + if (inverse) { + for (i = 0; i < length; ++i) dst[i] = src[i] + pred[i]; + } else { + for (i = 0; i < length; ++i) dst[i] = src[i] - pred[i]; + } +} + +//------------------------------------------------------------------------------ +// Horizontal filter. + +static WEBP_INLINE void DoHorizontalFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = inverse ? out : in; + + if (row == 0) { + // Leftmost pixel is the same as input for topmost scanline. + out[0] = in[0]; + PredictLine(in + 1, preds, out + 1, width - 1, inverse); + row = 1; + preds += stride; + in += stride; + out += stride; + } + + // Filter line-by-line. + while (row < last_row) { + // Leftmost pixel is predicted from above. + PredictLine(in, preds - stride, out, 1, inverse); + PredictLine(in + 1, preds, out + 1, width - 1, inverse); + ++row; + preds += stride; + in += stride; + out += stride; + } +} + +//------------------------------------------------------------------------------ +// Vertical filter. + +static WEBP_INLINE void DoVerticalFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = inverse ? out : in; + + if (row == 0) { + // Very first top-left pixel is copied. + out[0] = in[0]; + // Rest of top scan-line is left-predicted. + PredictLine(in + 1, preds, out + 1, width - 1, inverse); + row = 1; + in += stride; + out += stride; + } else { + // We are starting from in-between. Make sure 'preds' points to prev row. + preds -= stride; + } + + // Filter line-by-line. + while (row < last_row) { + PredictLine(in, preds, out, width, inverse); + ++row; + preds += stride; + in += stride; + out += stride; + } +} + +//------------------------------------------------------------------------------ +// Gradient filter. + +static WEBP_INLINE int GradientPredictor(uint8_t a, uint8_t b, uint8_t c) { + const int g = a + b - c; + return ((g & ~0xff) == 0) ? g : (g < 0) ? 0 : 255; // clip to 8bit +} + +static WEBP_INLINE void DoGradientFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = inverse ? out : in; + + // left prediction for top scan-line + if (row == 0) { + out[0] = in[0]; + PredictLine(in + 1, preds, out + 1, width - 1, inverse); + row = 1; + preds += stride; + in += stride; + out += stride; + } + + // Filter line-by-line. + while (row < last_row) { + int w; + // leftmost pixel: predict from above. + PredictLine(in, preds - stride, out, 1, inverse); + for (w = 1; w < width; ++w) { + const int pred = GradientPredictor(preds[w - 1], + preds[w - stride], + preds[w - stride - 1]); + out[w] = in[w] + (inverse ? pred : -pred); + } + ++row; + preds += stride; + in += stride; + out += stride; + } +} + +#undef SANITY_CHECK + +//------------------------------------------------------------------------------ + +static void HorizontalFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoHorizontalFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + +static void VerticalFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoVerticalFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + + +static void GradientFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoGradientFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + + +//------------------------------------------------------------------------------ + +static void VerticalUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoVerticalFilter(data, width, height, stride, row, num_rows, 1, data); +} + +static void HorizontalUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoHorizontalFilter(data, width, height, stride, row, num_rows, 1, data); +} + +static void GradientUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoGradientFilter(data, width, height, stride, row, num_rows, 1, data); +} + +//------------------------------------------------------------------------------ +// Init function + +WebPFilterFunc WebPFilters[WEBP_FILTER_LAST]; +WebPUnfilterFunc WebPUnfilters[WEBP_FILTER_LAST]; + +extern void VP8FiltersInitMIPSdspR2(void); +extern void VP8FiltersInitSSE2(void); + +static volatile VP8CPUInfo filters_last_cpuinfo_used = + (VP8CPUInfo)&filters_last_cpuinfo_used; + +WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInit(void) { + if (filters_last_cpuinfo_used == VP8GetCPUInfo) return; + + WebPUnfilters[WEBP_FILTER_NONE] = NULL; + WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter; + WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter; + WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter; + + WebPFilters[WEBP_FILTER_NONE] = NULL; + WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter; + WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter; + WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter; + + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8FiltersInitSSE2(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8FiltersInitMIPSdspR2(); + } +#endif + } + filters_last_cpuinfo_used = VP8GetCPUInfo; +} diff --git a/drivers/webp/dsp/filters_mips_dsp_r2.c b/drivers/webp/dsp/filters_mips_dsp_r2.c new file mode 100644 index 0000000000..8134af511b --- /dev/null +++ b/drivers/webp/dsp/filters_mips_dsp_r2.c @@ -0,0 +1,405 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Spatial prediction using various filters +// +// Author(s): Branimir Vasic (branimir.vasic@imgtec.com) +// Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../dsp/dsp.h" +#include <assert.h> +#include <stdlib.h> +#include <string.h> + +//------------------------------------------------------------------------------ +// Helpful macro. + +# define SANITY_CHECK(in, out) \ + assert(in != NULL); \ + assert(out != NULL); \ + assert(width > 0); \ + assert(height > 0); \ + assert(stride >= width); \ + assert(row >= 0 && num_rows > 0 && row + num_rows <= height); \ + (void)height; // Silence unused warning. + +// if INVERSE +// preds == &dst[-1] == &src[-1] +// else +// preds == &src[-1] != &dst[-1] +#define DO_PREDICT_LINE(SRC, DST, LENGTH, INVERSE) do { \ + const uint8_t* psrc = (uint8_t*)(SRC); \ + uint8_t* pdst = (uint8_t*)(DST); \ + const int ilength = (int)(LENGTH); \ + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; \ + __asm__ volatile ( \ + ".set push \n\t" \ + ".set noreorder \n\t" \ + "srl %[temp0], %[length], 0x2 \n\t" \ + "beqz %[temp0], 4f \n\t" \ + " andi %[temp6], %[length], 0x3 \n\t" \ + ".if " #INVERSE " \n\t" \ + "lbu %[temp1], -1(%[src]) \n\t" \ + "1: \n\t" \ + "lbu %[temp2], 0(%[src]) \n\t" \ + "lbu %[temp3], 1(%[src]) \n\t" \ + "lbu %[temp4], 2(%[src]) \n\t" \ + "lbu %[temp5], 3(%[src]) \n\t" \ + "addiu %[src], %[src], 4 \n\t" \ + "addiu %[temp0], %[temp0], -1 \n\t" \ + "addu %[temp2], %[temp2], %[temp1] \n\t" \ + "addu %[temp3], %[temp3], %[temp2] \n\t" \ + "addu %[temp4], %[temp4], %[temp3] \n\t" \ + "addu %[temp1], %[temp5], %[temp4] \n\t" \ + "sb %[temp2], -4(%[src]) \n\t" \ + "sb %[temp3], -3(%[src]) \n\t" \ + "sb %[temp4], -2(%[src]) \n\t" \ + "bnez %[temp0], 1b \n\t" \ + " sb %[temp1], -1(%[src]) \n\t" \ + ".else \n\t" \ + "1: \n\t" \ + "ulw %[temp1], -1(%[src]) \n\t" \ + "ulw %[temp2], 0(%[src]) \n\t" \ + "addiu %[src], %[src], 4 \n\t" \ + "addiu %[temp0], %[temp0], -1 \n\t" \ + "subu.qb %[temp3], %[temp2], %[temp1] \n\t" \ + "usw %[temp3], 0(%[dst]) \n\t" \ + "bnez %[temp0], 1b \n\t" \ + " addiu %[dst], %[dst], 4 \n\t" \ + ".endif \n\t" \ + "4: \n\t" \ + "beqz %[temp6], 3f \n\t" \ + " nop \n\t" \ + "2: \n\t" \ + "lbu %[temp1], -1(%[src]) \n\t" \ + "lbu %[temp2], 0(%[src]) \n\t" \ + "addiu %[src], %[src], 1 \n\t" \ + ".if " #INVERSE " \n\t" \ + "addu %[temp3], %[temp1], %[temp2] \n\t" \ + "sb %[temp3], -1(%[src]) \n\t" \ + ".else \n\t" \ + "subu %[temp3], %[temp1], %[temp2] \n\t" \ + "sb %[temp3], 0(%[dst]) \n\t" \ + ".endif \n\t" \ + "addiu %[temp6], %[temp6], -1 \n\t" \ + "bnez %[temp6], 2b \n\t" \ + " addiu %[dst], %[dst], 1 \n\t" \ + "3: \n\t" \ + ".set pop \n\t" \ + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), \ + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), \ + [temp6]"=&r"(temp6), [dst]"+&r"(pdst), [src]"+&r"(psrc) \ + : [length]"r"(ilength) \ + : "memory" \ + ); \ + } while (0) + +static WEBP_INLINE void PredictLine(const uint8_t* src, uint8_t* dst, + int length, int inverse) { + if (inverse) { + DO_PREDICT_LINE(src, dst, length, 1); + } else { + DO_PREDICT_LINE(src, dst, length, 0); + } +} + +#define DO_PREDICT_LINE_VERTICAL(SRC, PRED, DST, LENGTH, INVERSE) do { \ + const uint8_t* psrc = (uint8_t*)(SRC); \ + const uint8_t* ppred = (uint8_t*)(PRED); \ + uint8_t* pdst = (uint8_t*)(DST); \ + const int ilength = (int)(LENGTH); \ + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; \ + __asm__ volatile ( \ + ".set push \n\t" \ + ".set noreorder \n\t" \ + "srl %[temp0], %[length], 0x3 \n\t" \ + "beqz %[temp0], 4f \n\t" \ + " andi %[temp7], %[length], 0x7 \n\t" \ + "1: \n\t" \ + "ulw %[temp1], 0(%[src]) \n\t" \ + "ulw %[temp2], 0(%[pred]) \n\t" \ + "ulw %[temp3], 4(%[src]) \n\t" \ + "ulw %[temp4], 4(%[pred]) \n\t" \ + "addiu %[src], %[src], 8 \n\t" \ + ".if " #INVERSE " \n\t" \ + "addu.qb %[temp5], %[temp1], %[temp2] \n\t" \ + "addu.qb %[temp6], %[temp3], %[temp4] \n\t" \ + ".else \n\t" \ + "subu.qb %[temp5], %[temp1], %[temp2] \n\t" \ + "subu.qb %[temp6], %[temp3], %[temp4] \n\t" \ + ".endif \n\t" \ + "addiu %[pred], %[pred], 8 \n\t" \ + "usw %[temp5], 0(%[dst]) \n\t" \ + "usw %[temp6], 4(%[dst]) \n\t" \ + "addiu %[temp0], %[temp0], -1 \n\t" \ + "bnez %[temp0], 1b \n\t" \ + " addiu %[dst], %[dst], 8 \n\t" \ + "4: \n\t" \ + "beqz %[temp7], 3f \n\t" \ + " nop \n\t" \ + "2: \n\t" \ + "lbu %[temp1], 0(%[src]) \n\t" \ + "lbu %[temp2], 0(%[pred]) \n\t" \ + "addiu %[src], %[src], 1 \n\t" \ + "addiu %[pred], %[pred], 1 \n\t" \ + ".if " #INVERSE " \n\t" \ + "addu %[temp3], %[temp1], %[temp2] \n\t" \ + ".else \n\t" \ + "subu %[temp3], %[temp1], %[temp2] \n\t" \ + ".endif \n\t" \ + "sb %[temp3], 0(%[dst]) \n\t" \ + "addiu %[temp7], %[temp7], -1 \n\t" \ + "bnez %[temp7], 2b \n\t" \ + " addiu %[dst], %[dst], 1 \n\t" \ + "3: \n\t" \ + ".set pop \n\t" \ + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), \ + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), \ + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [pred]"+&r"(ppred), \ + [dst]"+&r"(pdst), [src]"+&r"(psrc) \ + : [length]"r"(ilength) \ + : "memory" \ + ); \ + } while (0) + +#define PREDICT_LINE_ONE_PASS(SRC, PRED, DST, INVERSE) do { \ + int temp1, temp2, temp3; \ + __asm__ volatile ( \ + "lbu %[temp1], 0(%[src]) \n\t" \ + "lbu %[temp2], 0(%[pred]) \n\t" \ + ".if " #INVERSE " \n\t" \ + "addu %[temp3], %[temp1], %[temp2] \n\t" \ + ".else \n\t" \ + "subu %[temp3], %[temp1], %[temp2] \n\t" \ + ".endif \n\t" \ + "sb %[temp3], 0(%[dst]) \n\t" \ + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3) \ + : [pred]"r"((PRED)), [dst]"r"((DST)), [src]"r"((SRC)) \ + : "memory" \ + ); \ + } while (0) + +//------------------------------------------------------------------------------ +// Horizontal filter. + +#define FILTER_LINE_BY_LINE(INVERSE) do { \ + while (row < last_row) { \ + PREDICT_LINE_ONE_PASS(in, preds - stride, out, INVERSE); \ + DO_PREDICT_LINE(in + 1, out + 1, width - 1, INVERSE); \ + ++row; \ + preds += stride; \ + in += stride; \ + out += stride; \ + } \ + } while (0) + +static WEBP_INLINE void DoHorizontalFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = inverse ? out : in; + + if (row == 0) { + // Leftmost pixel is the same as input for topmost scanline. + out[0] = in[0]; + PredictLine(in + 1, out + 1, width - 1, inverse); + row = 1; + preds += stride; + in += stride; + out += stride; + } + + // Filter line-by-line. + if (inverse) { + FILTER_LINE_BY_LINE(1); + } else { + FILTER_LINE_BY_LINE(0); + } +} + +#undef FILTER_LINE_BY_LINE + +static void HorizontalFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoHorizontalFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + +static void HorizontalUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoHorizontalFilter(data, width, height, stride, row, num_rows, 1, data); +} + +//------------------------------------------------------------------------------ +// Vertical filter. + +#define FILTER_LINE_BY_LINE(INVERSE) do { \ + while (row < last_row) { \ + DO_PREDICT_LINE_VERTICAL(in, preds, out, width, INVERSE); \ + ++row; \ + preds += stride; \ + in += stride; \ + out += stride; \ + } \ + } while (0) + +static WEBP_INLINE void DoVerticalFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = inverse ? out : in; + + if (row == 0) { + // Very first top-left pixel is copied. + out[0] = in[0]; + // Rest of top scan-line is left-predicted. + PredictLine(in + 1, out + 1, width - 1, inverse); + row = 1; + in += stride; + out += stride; + } else { + // We are starting from in-between. Make sure 'preds' points to prev row. + preds -= stride; + } + + // Filter line-by-line. + if (inverse) { + FILTER_LINE_BY_LINE(1); + } else { + FILTER_LINE_BY_LINE(0); + } +} + +#undef FILTER_LINE_BY_LINE +#undef DO_PREDICT_LINE_VERTICAL + +static void VerticalFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoVerticalFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + +static void VerticalUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoVerticalFilter(data, width, height, stride, row, num_rows, 1, data); +} + +//------------------------------------------------------------------------------ +// Gradient filter. + +static WEBP_INLINE int GradientPredictor(uint8_t a, uint8_t b, uint8_t c) { + int temp0; + __asm__ volatile ( + "addu %[temp0], %[a], %[b] \n\t" + "subu %[temp0], %[temp0], %[c] \n\t" + "shll_s.w %[temp0], %[temp0], 23 \n\t" + "precrqu_s.qb.ph %[temp0], %[temp0], $zero \n\t" + "srl %[temp0], %[temp0], 24 \n\t" + : [temp0]"=&r"(temp0) + : [a]"r"(a),[b]"r"(b),[c]"r"(c) + ); + return temp0; +} + +#define FILTER_LINE_BY_LINE(INVERSE, PREDS, OPERATION) do { \ + while (row < last_row) { \ + int w; \ + PREDICT_LINE_ONE_PASS(in, PREDS - stride, out, INVERSE); \ + for (w = 1; w < width; ++w) { \ + const int pred = GradientPredictor(PREDS[w - 1], \ + PREDS[w - stride], \ + PREDS[w - stride - 1]); \ + out[w] = in[w] OPERATION pred; \ + } \ + ++row; \ + in += stride; \ + out += stride; \ + } \ + } while (0) + +static WEBP_INLINE void DoGradientFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = inverse ? out : in; + + // left prediction for top scan-line + if (row == 0) { + out[0] = in[0]; + PredictLine(in + 1, out + 1, width - 1, inverse); + row = 1; + preds += stride; + in += stride; + out += stride; + } + + // Filter line-by-line. + if (inverse) { + FILTER_LINE_BY_LINE(1, out, +); + } else { + FILTER_LINE_BY_LINE(0, in, -); + } +} + +#undef FILTER_LINE_BY_LINE + +static void GradientFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoGradientFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + +static void GradientUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoGradientFilter(data, width, height, stride, row, num_rows, 1, data); +} + +#undef PREDICT_LINE_ONE_PASS +#undef DO_PREDICT_LINE +#undef SANITY_CHECK + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8FiltersInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitMIPSdspR2(void) { + WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter; + WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter; + WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter; + + WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter; + WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter; + WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8FiltersInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/filters_sse2.c b/drivers/webp/dsp/filters_sse2.c new file mode 100644 index 0000000000..bf93342eb7 --- /dev/null +++ b/drivers/webp/dsp/filters_sse2.c @@ -0,0 +1,352 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 variant of alpha filters +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE2) + +#include <assert.h> +#include <emmintrin.h> +#include <stdlib.h> +#include <string.h> + +//------------------------------------------------------------------------------ +// Helpful macro. + +# define SANITY_CHECK(in, out) \ + assert(in != NULL); \ + assert(out != NULL); \ + assert(width > 0); \ + assert(height > 0); \ + assert(stride >= width); \ + assert(row >= 0 && num_rows > 0 && row + num_rows <= height); \ + (void)height; // Silence unused warning. + +static void PredictLineTop(const uint8_t* src, const uint8_t* pred, + uint8_t* dst, int length, int inverse) { + int i; + const int max_pos = length & ~31; + assert(length >= 0); + if (inverse) { + for (i = 0; i < max_pos; i += 32) { + const __m128i A0 = _mm_loadu_si128((const __m128i*)&src[i + 0]); + const __m128i A1 = _mm_loadu_si128((const __m128i*)&src[i + 16]); + const __m128i B0 = _mm_loadu_si128((const __m128i*)&pred[i + 0]); + const __m128i B1 = _mm_loadu_si128((const __m128i*)&pred[i + 16]); + const __m128i C0 = _mm_add_epi8(A0, B0); + const __m128i C1 = _mm_add_epi8(A1, B1); + _mm_storeu_si128((__m128i*)&dst[i + 0], C0); + _mm_storeu_si128((__m128i*)&dst[i + 16], C1); + } + for (; i < length; ++i) dst[i] = src[i] + pred[i]; + } else { + for (i = 0; i < max_pos; i += 32) { + const __m128i A0 = _mm_loadu_si128((const __m128i*)&src[i + 0]); + const __m128i A1 = _mm_loadu_si128((const __m128i*)&src[i + 16]); + const __m128i B0 = _mm_loadu_si128((const __m128i*)&pred[i + 0]); + const __m128i B1 = _mm_loadu_si128((const __m128i*)&pred[i + 16]); + const __m128i C0 = _mm_sub_epi8(A0, B0); + const __m128i C1 = _mm_sub_epi8(A1, B1); + _mm_storeu_si128((__m128i*)&dst[i + 0], C0); + _mm_storeu_si128((__m128i*)&dst[i + 16], C1); + } + for (; i < length; ++i) dst[i] = src[i] - pred[i]; + } +} + +// Special case for left-based prediction (when preds==dst-1 or preds==src-1). +static void PredictLineLeft(const uint8_t* src, uint8_t* dst, int length, + int inverse) { + int i; + if (length <= 0) return; + if (inverse) { + const int max_pos = length & ~7; + __m128i last = _mm_set_epi32(0, 0, 0, dst[-1]); + for (i = 0; i < max_pos; i += 8) { + const __m128i A0 = _mm_loadl_epi64((const __m128i*)(src + i)); + const __m128i A1 = _mm_add_epi8(A0, last); + const __m128i A2 = _mm_slli_si128(A1, 1); + const __m128i A3 = _mm_add_epi8(A1, A2); + const __m128i A4 = _mm_slli_si128(A3, 2); + const __m128i A5 = _mm_add_epi8(A3, A4); + const __m128i A6 = _mm_slli_si128(A5, 4); + const __m128i A7 = _mm_add_epi8(A5, A6); + _mm_storel_epi64((__m128i*)(dst + i), A7); + last = _mm_srli_epi64(A7, 56); + } + for (; i < length; ++i) dst[i] = src[i] + dst[i - 1]; + } else { + const int max_pos = length & ~31; + for (i = 0; i < max_pos; i += 32) { + const __m128i A0 = _mm_loadu_si128((const __m128i*)(src + i + 0 )); + const __m128i B0 = _mm_loadu_si128((const __m128i*)(src + i + 0 - 1)); + const __m128i A1 = _mm_loadu_si128((const __m128i*)(src + i + 16 )); + const __m128i B1 = _mm_loadu_si128((const __m128i*)(src + i + 16 - 1)); + const __m128i C0 = _mm_sub_epi8(A0, B0); + const __m128i C1 = _mm_sub_epi8(A1, B1); + _mm_storeu_si128((__m128i*)(dst + i + 0), C0); + _mm_storeu_si128((__m128i*)(dst + i + 16), C1); + } + for (; i < length; ++i) dst[i] = src[i] - src[i - 1]; + } +} + +static void PredictLineC(const uint8_t* src, const uint8_t* pred, + uint8_t* dst, int length, int inverse) { + int i; + if (inverse) { + for (i = 0; i < length; ++i) dst[i] = src[i] + pred[i]; + } else { + for (i = 0; i < length; ++i) dst[i] = src[i] - pred[i]; + } +} + +//------------------------------------------------------------------------------ +// Horizontal filter. + +static WEBP_INLINE void DoHorizontalFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = inverse ? out : in; + + if (row == 0) { + // Leftmost pixel is the same as input for topmost scanline. + out[0] = in[0]; + PredictLineLeft(in + 1, out + 1, width - 1, inverse); + row = 1; + preds += stride; + in += stride; + out += stride; + } + + // Filter line-by-line. + while (row < last_row) { + // Leftmost pixel is predicted from above. + PredictLineC(in, preds - stride, out, 1, inverse); + PredictLineLeft(in + 1, out + 1, width - 1, inverse); + ++row; + preds += stride; + in += stride; + out += stride; + } +} + +//------------------------------------------------------------------------------ +// Vertical filter. + +static WEBP_INLINE void DoVerticalFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = inverse ? out : in; + + if (row == 0) { + // Very first top-left pixel is copied. + out[0] = in[0]; + // Rest of top scan-line is left-predicted. + PredictLineLeft(in + 1, out + 1, width - 1, inverse); + row = 1; + in += stride; + out += stride; + } else { + // We are starting from in-between. Make sure 'preds' points to prev row. + preds -= stride; + } + + // Filter line-by-line. + while (row < last_row) { + PredictLineTop(in, preds, out, width, inverse); + ++row; + preds += stride; + in += stride; + out += stride; + } +} + +//------------------------------------------------------------------------------ +// Gradient filter. + +static WEBP_INLINE int GradientPredictorC(uint8_t a, uint8_t b, uint8_t c) { + const int g = a + b - c; + return ((g & ~0xff) == 0) ? g : (g < 0) ? 0 : 255; // clip to 8bit +} + +static void GradientPredictDirect(const uint8_t* const row, + const uint8_t* const top, + uint8_t* const out, int length) { + const int max_pos = length & ~7; + int i; + const __m128i zero = _mm_setzero_si128(); + for (i = 0; i < max_pos; i += 8) { + const __m128i A0 = _mm_loadl_epi64((const __m128i*)&row[i - 1]); + const __m128i B0 = _mm_loadl_epi64((const __m128i*)&top[i]); + const __m128i C0 = _mm_loadl_epi64((const __m128i*)&top[i - 1]); + const __m128i D = _mm_loadl_epi64((const __m128i*)&row[i]); + const __m128i A1 = _mm_unpacklo_epi8(A0, zero); + const __m128i B1 = _mm_unpacklo_epi8(B0, zero); + const __m128i C1 = _mm_unpacklo_epi8(C0, zero); + const __m128i E = _mm_add_epi16(A1, B1); + const __m128i F = _mm_sub_epi16(E, C1); + const __m128i G = _mm_packus_epi16(F, zero); + const __m128i H = _mm_sub_epi8(D, G); + _mm_storel_epi64((__m128i*)(out + i), H); + } + for (; i < length; ++i) { + out[i] = row[i] - GradientPredictorC(row[i - 1], top[i], top[i - 1]); + } +} + +static void GradientPredictInverse(const uint8_t* const in, + const uint8_t* const top, + uint8_t* const row, int length) { + if (length > 0) { + int i; + const int max_pos = length & ~7; + const __m128i zero = _mm_setzero_si128(); + __m128i A = _mm_set_epi32(0, 0, 0, row[-1]); // left sample + for (i = 0; i < max_pos; i += 8) { + const __m128i tmp0 = _mm_loadl_epi64((const __m128i*)&top[i]); + const __m128i tmp1 = _mm_loadl_epi64((const __m128i*)&top[i - 1]); + const __m128i B = _mm_unpacklo_epi8(tmp0, zero); + const __m128i C = _mm_unpacklo_epi8(tmp1, zero); + const __m128i tmp2 = _mm_loadl_epi64((const __m128i*)&in[i]); + const __m128i D = _mm_unpacklo_epi8(tmp2, zero); // base input + const __m128i E = _mm_sub_epi16(B, C); // unclipped gradient basis B - C + __m128i out = zero; // accumulator for output + __m128i mask_hi = _mm_set_epi32(0, 0, 0, 0xff); + int k = 8; + while (1) { + const __m128i tmp3 = _mm_add_epi16(A, E); // delta = A + B - C + const __m128i tmp4 = _mm_min_epi16(tmp3, mask_hi); + const __m128i tmp5 = _mm_max_epi16(tmp4, zero); // clipped delta + const __m128i tmp6 = _mm_add_epi16(tmp5, D); // add to in[] values + A = _mm_and_si128(tmp6, mask_hi); // 1-complement clip + out = _mm_or_si128(out, A); // accumulate output + if (--k == 0) break; + A = _mm_slli_si128(A, 2); // rotate left sample + mask_hi = _mm_slli_si128(mask_hi, 2); // rotate mask + } + A = _mm_srli_si128(A, 14); // prepare left sample for next iteration + _mm_storel_epi64((__m128i*)&row[i], _mm_packus_epi16(out, zero)); + } + for (; i < length; ++i) { + row[i] = in[i] + GradientPredictorC(row[i - 1], top[i], top[i - 1]); + } + } +} + +static WEBP_INLINE void DoGradientFilter(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, + int inverse, uint8_t* out) { + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + + // left prediction for top scan-line + if (row == 0) { + out[0] = in[0]; + PredictLineLeft(in + 1, out + 1, width - 1, inverse); + row = 1; + in += stride; + out += stride; + } + + // Filter line-by-line. + while (row < last_row) { + if (inverse) { + PredictLineC(in, out - stride, out, 1, inverse); // predict from above + GradientPredictInverse(in + 1, out + 1 - stride, out + 1, width - 1); + } else { + PredictLineC(in, in - stride, out, 1, inverse); + GradientPredictDirect(in + 1, in + 1 - stride, out + 1, width - 1); + } + ++row; + in += stride; + out += stride; + } +} + +#undef SANITY_CHECK + +//------------------------------------------------------------------------------ + +static void HorizontalFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoHorizontalFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + +static void VerticalFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoVerticalFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + + +static void GradientFilter(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoGradientFilter(data, width, height, stride, 0, height, 0, filtered_data); +} + + +//------------------------------------------------------------------------------ + +static void VerticalUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoVerticalFilter(data, width, height, stride, row, num_rows, 1, data); +} + +static void HorizontalUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoHorizontalFilter(data, width, height, stride, row, num_rows, 1, data); +} + +static void GradientUnfilter(int width, int height, int stride, int row, + int num_rows, uint8_t* data) { + DoGradientFilter(data, width, height, stride, row, num_rows, 1, data); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8FiltersInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitSSE2(void) { + WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter; + WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter; + WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter; + + WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter; + WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter; + WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8FiltersInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/lossless.c b/drivers/webp/dsp/lossless.c index 62a6b7b15a..5702eb3b17 100644 --- a/drivers/webp/dsp/lossless.c +++ b/drivers/webp/dsp/lossless.c @@ -1,8 +1,10 @@ // Copyright 2012 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // Image transforms and color space conversion methods for lossless decoder. @@ -11,170 +13,16 @@ // Jyrki Alakuijala (jyrki@google.com) // Urvang Joshi (urvang@google.com) -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif +#include "./dsp.h" #include <math.h> #include <stdlib.h> -#include "./lossless.h" #include "../dec/vp8li.h" -#include "../dsp/yuv.h" -#include "../dsp/dsp.h" -#include "../enc/histogram.h" +#include "../utils/endian_inl.h" +#include "./lossless.h" #define MAX_DIFF_COST (1e30f) -// lookup table for small values of log2(int) -#define APPROX_LOG_MAX 4096 -#define LOG_2_RECIPROCAL 1.44269504088896338700465094007086 -#define LOG_LOOKUP_IDX_MAX 256 -static const float kLog2Table[LOG_LOOKUP_IDX_MAX] = { - 0.0000000000000000f, 0.0000000000000000f, - 1.0000000000000000f, 1.5849625007211560f, - 2.0000000000000000f, 2.3219280948873621f, - 2.5849625007211560f, 2.8073549220576041f, - 3.0000000000000000f, 3.1699250014423121f, - 3.3219280948873621f, 3.4594316186372973f, - 3.5849625007211560f, 3.7004397181410921f, - 3.8073549220576041f, 3.9068905956085187f, - 4.0000000000000000f, 4.0874628412503390f, - 4.1699250014423121f, 4.2479275134435852f, - 4.3219280948873626f, 4.3923174227787606f, - 4.4594316186372973f, 4.5235619560570130f, - 4.5849625007211560f, 4.6438561897747243f, - 4.7004397181410917f, 4.7548875021634682f, - 4.8073549220576037f, 4.8579809951275718f, - 4.9068905956085187f, 4.9541963103868749f, - 5.0000000000000000f, 5.0443941193584533f, - 5.0874628412503390f, 5.1292830169449663f, - 5.1699250014423121f, 5.2094533656289501f, - 5.2479275134435852f, 5.2854022188622487f, - 5.3219280948873626f, 5.3575520046180837f, - 5.3923174227787606f, 5.4262647547020979f, - 5.4594316186372973f, 5.4918530963296747f, - 5.5235619560570130f, 5.5545888516776376f, - 5.5849625007211560f, 5.6147098441152083f, - 5.6438561897747243f, 5.6724253419714951f, - 5.7004397181410917f, 5.7279204545631987f, - 5.7548875021634682f, 5.7813597135246599f, - 5.8073549220576037f, 5.8328900141647412f, - 5.8579809951275718f, 5.8826430493618415f, - 5.9068905956085187f, 5.9307373375628866f, - 5.9541963103868749f, 5.9772799234999167f, - 6.0000000000000000f, 6.0223678130284543f, - 6.0443941193584533f, 6.0660891904577720f, - 6.0874628412503390f, 6.1085244567781691f, - 6.1292830169449663f, 6.1497471195046822f, - 6.1699250014423121f, 6.1898245588800175f, - 6.2094533656289501f, 6.2288186904958804f, - 6.2479275134435852f, 6.2667865406949010f, - 6.2854022188622487f, 6.3037807481771030f, - 6.3219280948873626f, 6.3398500028846243f, - 6.3575520046180837f, 6.3750394313469245f, - 6.3923174227787606f, 6.4093909361377017f, - 6.4262647547020979f, 6.4429434958487279f, - 6.4594316186372973f, 6.4757334309663976f, - 6.4918530963296747f, 6.5077946401986963f, - 6.5235619560570130f, 6.5391588111080309f, - 6.5545888516776376f, 6.5698556083309478f, - 6.5849625007211560f, 6.5999128421871278f, - 6.6147098441152083f, 6.6293566200796094f, - 6.6438561897747243f, 6.6582114827517946f, - 6.6724253419714951f, 6.6865005271832185f, - 6.7004397181410917f, 6.7142455176661224f, - 6.7279204545631987f, 6.7414669864011464f, - 6.7548875021634682f, 6.7681843247769259f, - 6.7813597135246599f, 6.7944158663501061f, - 6.8073549220576037f, 6.8201789624151878f, - 6.8328900141647412f, 6.8454900509443747f, - 6.8579809951275718f, 6.8703647195834047f, - 6.8826430493618415f, 6.8948177633079437f, - 6.9068905956085187f, 6.9188632372745946f, - 6.9307373375628866f, 6.9425145053392398f, - 6.9541963103868749f, 6.9657842846620869f, - 6.9772799234999167f, 6.9886846867721654f, - 7.0000000000000000f, 7.0112272554232539f, - 7.0223678130284543f, 7.0334230015374501f, - 7.0443941193584533f, 7.0552824355011898f, - 7.0660891904577720f, 7.0768155970508308f, - 7.0874628412503390f, 7.0980320829605263f, - 7.1085244567781691f, 7.1189410727235076f, - 7.1292830169449663f, 7.1395513523987936f, - 7.1497471195046822f, 7.1598713367783890f, - 7.1699250014423121f, 7.1799090900149344f, - 7.1898245588800175f, 7.1996723448363644f, - 7.2094533656289501f, 7.2191685204621611f, - 7.2288186904958804f, 7.2384047393250785f, - 7.2479275134435852f, 7.2573878426926521f, - 7.2667865406949010f, 7.2761244052742375f, - 7.2854022188622487f, 7.2946207488916270f, - 7.3037807481771030f, 7.3128829552843557f, - 7.3219280948873626f, 7.3309168781146167f, - 7.3398500028846243f, 7.3487281542310771f, - 7.3575520046180837f, 7.3663222142458160f, - 7.3750394313469245f, 7.3837042924740519f, - 7.3923174227787606f, 7.4008794362821843f, - 7.4093909361377017f, 7.4178525148858982f, - 7.4262647547020979f, 7.4346282276367245f, - 7.4429434958487279f, 7.4512111118323289f, - 7.4594316186372973f, 7.4676055500829976f, - 7.4757334309663976f, 7.4838157772642563f, - 7.4918530963296747f, 7.4998458870832056f, - 7.5077946401986963f, 7.5156998382840427f, - 7.5235619560570130f, 7.5313814605163118f, - 7.5391588111080309f, 7.5468944598876364f, - 7.5545888516776376f, 7.5622424242210728f, - 7.5698556083309478f, 7.5774288280357486f, - 7.5849625007211560f, 7.5924570372680806f, - 7.5999128421871278f, 7.6073303137496104f, - 7.6147098441152083f, 7.6220518194563764f, - 7.6293566200796094f, 7.6366246205436487f, - 7.6438561897747243f, 7.6510516911789281f, - 7.6582114827517946f, 7.6653359171851764f, - 7.6724253419714951f, 7.6794800995054464f, - 7.6865005271832185f, 7.6934869574993252f, - 7.7004397181410917f, 7.7073591320808825f, - 7.7142455176661224f, 7.7210991887071855f, - 7.7279204545631987f, 7.7347096202258383f, - 7.7414669864011464f, 7.7481928495894605f, - 7.7548875021634682f, 7.7615512324444795f, - 7.7681843247769259f, 7.7747870596011736f, - 7.7813597135246599f, 7.7879025593914317f, - 7.7944158663501061f, 7.8008998999203047f, - 7.8073549220576037f, 7.8137811912170374f, - 7.8201789624151878f, 7.8265484872909150f, - 7.8328900141647412f, 7.8392037880969436f, - 7.8454900509443747f, 7.8517490414160571f, - 7.8579809951275718f, 7.8641861446542797f, - 7.8703647195834047f, 7.8765169465649993f, - 7.8826430493618415f, 7.8887432488982591f, - 7.8948177633079437f, 7.9008668079807486f, - 7.9068905956085187f, 7.9128893362299619f, - 7.9188632372745946f, 7.9248125036057812f, - 7.9307373375628866f, 7.9366379390025709f, - 7.9425145053392398f, 7.9483672315846778f, - 7.9541963103868749f, 7.9600019320680805f, - 7.9657842846620869f, 7.9715435539507719f, - 7.9772799234999167f, 7.9829935746943103f, - 7.9886846867721654f, 7.9943534368588577f -}; - -float VP8LFastLog2(int v) { - if (v < LOG_LOOKUP_IDX_MAX) { - return kLog2Table[v]; - } else if (v < APPROX_LOG_MAX) { - int log_cnt = 0; - while (v >= LOG_LOOKUP_IDX_MAX) { - ++log_cnt; - v = v >> 1; - } - return kLog2Table[v] + (float)log_cnt; - } else { - return (float)(LOG_2_RECIPROCAL * log((double)v)); - } -} - //------------------------------------------------------------------------------ // Image transforms. @@ -186,7 +34,7 @@ static WEBP_INLINE void AddPixelsEq(uint32_t* a, uint32_t b) { } static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) { - return (((a0 ^ a1) & 0xfefefefeL) >> 1) + (a0 & a1); + return (((a0 ^ a1) & 0xfefefefeu) >> 1) + (a0 & a1); } static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) { @@ -221,7 +69,7 @@ static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, (c1 >> 8) & 0xff, (c2 >> 8) & 0xff); const int b = AddSubtractComponentFull(c0 & 0xff, c1 & 0xff, c2 & 0xff); - return (a << 24) | (r << 16) | (g << 8) | b; + return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b; } static WEBP_INLINE int AddSubtractComponentHalf(int a, int b) { @@ -235,22 +83,30 @@ static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, const int r = AddSubtractComponentHalf((ave >> 16) & 0xff, (c2 >> 16) & 0xff); const int g = AddSubtractComponentHalf((ave >> 8) & 0xff, (c2 >> 8) & 0xff); const int b = AddSubtractComponentHalf((ave >> 0) & 0xff, (c2 >> 0) & 0xff); - return (a << 24) | (r << 16) | (g << 8) | b; + return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b; } -static WEBP_INLINE int Sub3(int a, int b, int c) { - const int pa = b - c; - const int pb = a - c; - return abs(pa) - abs(pb); +// gcc-4.9 on ARM generates incorrect code in Select() when Sub3() is inlined. +#if defined(__arm__) && LOCAL_GCC_VERSION == 0x409 +# define LOCAL_INLINE __attribute__ ((noinline)) +#else +# define LOCAL_INLINE WEBP_INLINE +#endif + +static LOCAL_INLINE int Sub3(int a, int b, int c) { + const int pb = b - c; + const int pa = a - c; + return abs(pb) - abs(pa); } +#undef LOCAL_INLINE + static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { const int pa_minus_pb = Sub3((a >> 24) , (b >> 24) , (c >> 24) ) + Sub3((a >> 16) & 0xff, (b >> 16) & 0xff, (c >> 16) & 0xff) + Sub3((a >> 8) & 0xff, (b >> 8) & 0xff, (c >> 8) & 0xff) + Sub3((a ) & 0xff, (b ) & 0xff, (c ) & 0xff); - return (pa_minus_pb <= 0) ? a : b; } @@ -317,208 +173,7 @@ static uint32_t Predictor13(uint32_t left, const uint32_t* const top) { return pred; } -typedef uint32_t (*PredictorFunc)(uint32_t left, const uint32_t* const top); -static const PredictorFunc kPredictors[16] = { - Predictor0, Predictor1, Predictor2, Predictor3, - Predictor4, Predictor5, Predictor6, Predictor7, - Predictor8, Predictor9, Predictor10, Predictor11, - Predictor12, Predictor13, - Predictor0, Predictor0 // <- padding security sentinels -}; - -// TODO(vikasa): Replace 256 etc with defines. -static float PredictionCostSpatial(const int* counts, - int weight_0, double exp_val) { - const int significant_symbols = 16; - const double exp_decay_factor = 0.6; - double bits = weight_0 * counts[0]; - int i; - for (i = 1; i < significant_symbols; ++i) { - bits += exp_val * (counts[i] + counts[256 - i]); - exp_val *= exp_decay_factor; - } - return (float)(-0.1 * bits); -} - -// Compute the Shanon's entropy: Sum(p*log2(p)) -static float ShannonEntropy(const int* const array, int n) { - int i; - float retval = 0.f; - int sum = 0; - for (i = 0; i < n; ++i) { - if (array[i] != 0) { - sum += array[i]; - retval -= VP8LFastSLog2(array[i]); - } - } - retval += VP8LFastSLog2(sum); - return retval; -} - -static float PredictionCostSpatialHistogram(int accumulated[4][256], - int tile[4][256]) { - int i; - int k; - int combo[256]; - double retval = 0; - for (i = 0; i < 4; ++i) { - const double exp_val = 0.94; - retval += PredictionCostSpatial(&tile[i][0], 1, exp_val); - retval += ShannonEntropy(&tile[i][0], 256); - for (k = 0; k < 256; ++k) { - combo[k] = accumulated[i][k] + tile[i][k]; - } - retval += ShannonEntropy(&combo[0], 256); - } - return (float)retval; -} - -static int GetBestPredictorForTile(int width, int height, - int tile_x, int tile_y, int bits, - int accumulated[4][256], - const uint32_t* const argb_scratch) { - const int kNumPredModes = 14; - const int col_start = tile_x << bits; - const int row_start = tile_y << bits; - const int tile_size = 1 << bits; - const int ymax = (tile_size <= height - row_start) ? - tile_size : height - row_start; - const int xmax = (tile_size <= width - col_start) ? - tile_size : width - col_start; - int histo[4][256]; - float best_diff = MAX_DIFF_COST; - int best_mode = 0; - - int mode; - for (mode = 0; mode < kNumPredModes; ++mode) { - const uint32_t* current_row = argb_scratch; - const PredictorFunc pred_func = kPredictors[mode]; - float cur_diff; - int y; - memset(&histo[0][0], 0, sizeof(histo)); - for (y = 0; y < ymax; ++y) { - int x; - const int row = row_start + y; - const uint32_t* const upper_row = current_row; - current_row = upper_row + width; - for (x = 0; x < xmax; ++x) { - const int col = col_start + x; - uint32_t predict; - uint32_t predict_diff; - if (row == 0) { - predict = (col == 0) ? ARGB_BLACK : current_row[col - 1]; // Left. - } else if (col == 0) { - predict = upper_row[col]; // Top. - } else { - predict = pred_func(current_row[col - 1], upper_row + col); - } - predict_diff = VP8LSubPixels(current_row[col], predict); - ++histo[0][predict_diff >> 24]; - ++histo[1][((predict_diff >> 16) & 0xff)]; - ++histo[2][((predict_diff >> 8) & 0xff)]; - ++histo[3][(predict_diff & 0xff)]; - } - } - cur_diff = PredictionCostSpatialHistogram(accumulated, histo); - if (cur_diff < best_diff) { - best_diff = cur_diff; - best_mode = mode; - } - } - - return best_mode; -} - -static void CopyTileWithPrediction(int width, int height, - int tile_x, int tile_y, int bits, int mode, - const uint32_t* const argb_scratch, - uint32_t* const argb) { - const int col_start = tile_x << bits; - const int row_start = tile_y << bits; - const int tile_size = 1 << bits; - const int ymax = (tile_size <= height - row_start) ? - tile_size : height - row_start; - const int xmax = (tile_size <= width - col_start) ? - tile_size : width - col_start; - const PredictorFunc pred_func = kPredictors[mode]; - const uint32_t* current_row = argb_scratch; - - int y; - for (y = 0; y < ymax; ++y) { - int x; - const int row = row_start + y; - const uint32_t* const upper_row = current_row; - current_row = upper_row + width; - for (x = 0; x < xmax; ++x) { - const int col = col_start + x; - const int pix = row * width + col; - uint32_t predict; - if (row == 0) { - predict = (col == 0) ? ARGB_BLACK : current_row[col - 1]; // Left. - } else if (col == 0) { - predict = upper_row[col]; // Top. - } else { - predict = pred_func(current_row[col - 1], upper_row + col); - } - argb[pix] = VP8LSubPixels(current_row[col], predict); - } - } -} - -void VP8LResidualImage(int width, int height, int bits, - uint32_t* const argb, uint32_t* const argb_scratch, - uint32_t* const image) { - const int max_tile_size = 1 << bits; - const int tiles_per_row = VP8LSubSampleSize(width, bits); - const int tiles_per_col = VP8LSubSampleSize(height, bits); - uint32_t* const upper_row = argb_scratch; - uint32_t* const current_tile_rows = argb_scratch + width; - int tile_y; - int histo[4][256]; - memset(histo, 0, sizeof(histo)); - for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) { - const int tile_y_offset = tile_y * max_tile_size; - const int this_tile_height = - (tile_y < tiles_per_col - 1) ? max_tile_size : height - tile_y_offset; - int tile_x; - if (tile_y > 0) { - memcpy(upper_row, current_tile_rows + (max_tile_size - 1) * width, - width * sizeof(*upper_row)); - } - memcpy(current_tile_rows, &argb[tile_y_offset * width], - this_tile_height * width * sizeof(*current_tile_rows)); - for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) { - int pred; - int y; - const int tile_x_offset = tile_x * max_tile_size; - int all_x_max = tile_x_offset + max_tile_size; - if (all_x_max > width) { - all_x_max = width; - } - pred = GetBestPredictorForTile(width, height, tile_x, tile_y, bits, histo, - argb_scratch); - image[tile_y * tiles_per_row + tile_x] = 0xff000000u | (pred << 8); - CopyTileWithPrediction(width, height, tile_x, tile_y, bits, pred, - argb_scratch, argb); - for (y = 0; y < max_tile_size; ++y) { - int ix; - int all_x; - int all_y = tile_y_offset + y; - if (all_y >= height) { - break; - } - ix = all_y * width + tile_x_offset; - for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) { - const uint32_t a = argb[ix]; - ++histo[0][a >> 24]; - ++histo[1][((a >> 16) & 0xff)]; - ++histo[2][((a >> 8) & 0xff)]; - ++histo[3][(a & 0xff)]; - } - } - } - } -} +//------------------------------------------------------------------------------ // Inverse prediction. static void PredictorInverseTransform(const VP8LTransform* const transform, @@ -538,29 +193,36 @@ static void PredictorInverseTransform(const VP8LTransform* const transform, { int y = y_start; - const int mask = (1 << transform->bits_) - 1; + const int tile_width = 1 << transform->bits_; + const int mask = tile_width - 1; + const int safe_width = width & ~mask; const int tiles_per_row = VP8LSubSampleSize(width, transform->bits_); const uint32_t* pred_mode_base = transform->data_ + (y >> transform->bits_) * tiles_per_row; while (y < y_end) { - int x; const uint32_t pred2 = Predictor2(data[-1], data - width); const uint32_t* pred_mode_src = pred_mode_base; - PredictorFunc pred_func; - + VP8LPredictorFunc pred_func; + int x = 1; + int t = 1; // First pixel follows the T (mode=2) mode. AddPixelsEq(data, pred2); - // .. the rest: - pred_func = kPredictors[((*pred_mode_src++) >> 8) & 0xf]; - for (x = 1; x < width; ++x) { - uint32_t pred; - if ((x & mask) == 0) { // start of tile. Read predictor function. - pred_func = kPredictors[((*pred_mode_src++) >> 8) & 0xf]; + while (x < safe_width) { + pred_func = VP8LPredictors[((*pred_mode_src++) >> 8) & 0xf]; + for (; t < tile_width; ++t, ++x) { + const uint32_t pred = pred_func(data[x - 1], data + x - width); + AddPixelsEq(data + x, pred); + } + t = 0; + } + if (x < width) { + pred_func = VP8LPredictors[((*pred_mode_src++) >> 8) & 0xf]; + for (; x < width; ++x) { + const uint32_t pred = pred_func(data[x - 1], data + x - width); + AddPixelsEq(data + x, pred); } - pred = pred_func(data[x - 1], data + x - width); - AddPixelsEq(data + x, pred); } data += width; ++y; @@ -571,326 +233,47 @@ static void PredictorInverseTransform(const VP8LTransform* const transform, } } -void VP8LSubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs) { - int i; - for (i = 0; i < num_pixs; ++i) { - const uint32_t argb = argb_data[i]; - const uint32_t green = (argb >> 8) & 0xff; - const uint32_t new_r = (((argb >> 16) & 0xff) - green) & 0xff; - const uint32_t new_b = ((argb & 0xff) - green) & 0xff; - argb_data[i] = (argb & 0xff00ff00) | (new_r << 16) | new_b; - } -} - // Add green to blue and red channels (i.e. perform the inverse transform of // 'subtract green'). -static void AddGreenToBlueAndRed(const VP8LTransform* const transform, - int y_start, int y_end, uint32_t* data) { - const int width = transform->xsize_; - const uint32_t* const data_end = data + (y_end - y_start) * width; - while (data < data_end) { - const uint32_t argb = *data; - // "* 0001001u" is equivalent to "(green << 16) + green)" +void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels) { + int i; + for (i = 0; i < num_pixels; ++i) { + const uint32_t argb = data[i]; const uint32_t green = ((argb >> 8) & 0xff); uint32_t red_blue = (argb & 0x00ff00ffu); red_blue += (green << 16) | green; red_blue &= 0x00ff00ffu; - *data++ = (argb & 0xff00ff00u) | red_blue; + data[i] = (argb & 0xff00ff00u) | red_blue; } } -typedef struct { - // Note: the members are uint8_t, so that any negative values are - // automatically converted to "mod 256" values. - uint8_t green_to_red_; - uint8_t green_to_blue_; - uint8_t red_to_blue_; -} Multipliers; - -static WEBP_INLINE void MultipliersClear(Multipliers* m) { - m->green_to_red_ = 0; - m->green_to_blue_ = 0; - m->red_to_blue_ = 0; -} - static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred, int8_t color) { return (uint32_t)((int)(color_pred) * color) >> 5; } static WEBP_INLINE void ColorCodeToMultipliers(uint32_t color_code, - Multipliers* const m) { + VP8LMultipliers* const m) { m->green_to_red_ = (color_code >> 0) & 0xff; m->green_to_blue_ = (color_code >> 8) & 0xff; m->red_to_blue_ = (color_code >> 16) & 0xff; } -static WEBP_INLINE uint32_t MultipliersToColorCode(Multipliers* const m) { - return 0xff000000u | - ((uint32_t)(m->red_to_blue_) << 16) | - ((uint32_t)(m->green_to_blue_) << 8) | - m->green_to_red_; -} - -static WEBP_INLINE uint32_t TransformColor(const Multipliers* const m, - uint32_t argb, int inverse) { - const uint32_t green = argb >> 8; - const uint32_t red = argb >> 16; - uint32_t new_red = red; - uint32_t new_blue = argb; - - if (inverse) { +void VP8LTransformColorInverse_C(const VP8LMultipliers* const m, uint32_t* data, + int num_pixels) { + int i; + for (i = 0; i < num_pixels; ++i) { + const uint32_t argb = data[i]; + const uint32_t green = argb >> 8; + const uint32_t red = argb >> 16; + uint32_t new_red = red; + uint32_t new_blue = argb; new_red += ColorTransformDelta(m->green_to_red_, green); new_red &= 0xff; new_blue += ColorTransformDelta(m->green_to_blue_, green); new_blue += ColorTransformDelta(m->red_to_blue_, new_red); new_blue &= 0xff; - } else { - new_red -= ColorTransformDelta(m->green_to_red_, green); - new_red &= 0xff; - new_blue -= ColorTransformDelta(m->green_to_blue_, green); - new_blue -= ColorTransformDelta(m->red_to_blue_, red); - new_blue &= 0xff; - } - return (argb & 0xff00ff00u) | (new_red << 16) | (new_blue); -} - -static WEBP_INLINE int SkipRepeatedPixels(const uint32_t* const argb, - int ix, int xsize) { - const uint32_t v = argb[ix]; - if (ix >= xsize + 3) { - if (v == argb[ix - xsize] && - argb[ix - 1] == argb[ix - xsize - 1] && - argb[ix - 2] == argb[ix - xsize - 2] && - argb[ix - 3] == argb[ix - xsize - 3]) { - return 1; - } - return v == argb[ix - 3] && v == argb[ix - 2] && v == argb[ix - 1]; - } else if (ix >= 3) { - return v == argb[ix - 3] && v == argb[ix - 2] && v == argb[ix - 1]; - } - return 0; -} - -static float PredictionCostCrossColor(const int accumulated[256], - const int counts[256]) { - // Favor low entropy, locally and globally. - int i; - int combo[256]; - for (i = 0; i < 256; ++i) { - combo[i] = accumulated[i] + counts[i]; - } - return ShannonEntropy(combo, 256) + - ShannonEntropy(counts, 256) + - PredictionCostSpatial(counts, 3, 2.4); // Favor small absolute values. -} - -static Multipliers GetBestColorTransformForTile( - int tile_x, int tile_y, int bits, - Multipliers prevX, - Multipliers prevY, - int step, int xsize, int ysize, - int* accumulated_red_histo, - int* accumulated_blue_histo, - const uint32_t* const argb) { - float best_diff = MAX_DIFF_COST; - float cur_diff; - const int halfstep = step / 2; - const int max_tile_size = 1 << bits; - const int tile_y_offset = tile_y * max_tile_size; - const int tile_x_offset = tile_x * max_tile_size; - int green_to_red; - int green_to_blue; - int red_to_blue; - int all_x_max = tile_x_offset + max_tile_size; - int all_y_max = tile_y_offset + max_tile_size; - Multipliers best_tx; - MultipliersClear(&best_tx); - if (all_x_max > xsize) { - all_x_max = xsize; - } - if (all_y_max > ysize) { - all_y_max = ysize; - } - for (green_to_red = -64; green_to_red <= 64; green_to_red += halfstep) { - int histo[256] = { 0 }; - int all_y; - Multipliers tx; - MultipliersClear(&tx); - tx.green_to_red_ = green_to_red & 0xff; - - for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) { - uint32_t predict; - int ix = all_y * xsize + tile_x_offset; - int all_x; - for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) { - if (SkipRepeatedPixels(argb, ix, xsize)) { - continue; - } - predict = TransformColor(&tx, argb[ix], 0); - ++histo[(predict >> 16) & 0xff]; // red. - } - } - cur_diff = PredictionCostCrossColor(&accumulated_red_histo[0], &histo[0]); - if (tx.green_to_red_ == prevX.green_to_red_) { - cur_diff -= 3; // favor keeping the areas locally similar - } - if (tx.green_to_red_ == prevY.green_to_red_) { - cur_diff -= 3; // favor keeping the areas locally similar - } - if (tx.green_to_red_ == 0) { - cur_diff -= 3; - } - if (cur_diff < best_diff) { - best_diff = cur_diff; - best_tx = tx; - } - } - best_diff = MAX_DIFF_COST; - green_to_red = best_tx.green_to_red_; - for (green_to_blue = -32; green_to_blue <= 32; green_to_blue += step) { - for (red_to_blue = -32; red_to_blue <= 32; red_to_blue += step) { - int all_y; - int histo[256] = { 0 }; - Multipliers tx; - tx.green_to_red_ = green_to_red; - tx.green_to_blue_ = green_to_blue; - tx.red_to_blue_ = red_to_blue; - for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) { - uint32_t predict; - int all_x; - int ix = all_y * xsize + tile_x_offset; - for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) { - if (SkipRepeatedPixels(argb, ix, xsize)) { - continue; - } - predict = TransformColor(&tx, argb[ix], 0); - ++histo[predict & 0xff]; // blue. - } - } - cur_diff = - PredictionCostCrossColor(&accumulated_blue_histo[0], &histo[0]); - if (tx.green_to_blue_ == prevX.green_to_blue_) { - cur_diff -= 3; // favor keeping the areas locally similar - } - if (tx.green_to_blue_ == prevY.green_to_blue_) { - cur_diff -= 3; // favor keeping the areas locally similar - } - if (tx.red_to_blue_ == prevX.red_to_blue_) { - cur_diff -= 3; // favor keeping the areas locally similar - } - if (tx.red_to_blue_ == prevY.red_to_blue_) { - cur_diff -= 3; // favor keeping the areas locally similar - } - if (tx.green_to_blue_ == 0) { - cur_diff -= 3; - } - if (tx.red_to_blue_ == 0) { - cur_diff -= 3; - } - if (cur_diff < best_diff) { - best_diff = cur_diff; - best_tx = tx; - } - } - } - return best_tx; -} - -static void CopyTileWithColorTransform(int xsize, int ysize, - int tile_x, int tile_y, int bits, - Multipliers color_transform, - uint32_t* const argb) { - int y; - int xscan = 1 << bits; - int yscan = 1 << bits; - tile_x <<= bits; - tile_y <<= bits; - if (xscan > xsize - tile_x) { - xscan = xsize - tile_x; - } - if (yscan > ysize - tile_y) { - yscan = ysize - tile_y; - } - yscan += tile_y; - for (y = tile_y; y < yscan; ++y) { - int ix = y * xsize + tile_x; - const int end_ix = ix + xscan; - for (; ix < end_ix; ++ix) { - argb[ix] = TransformColor(&color_transform, argb[ix], 0); - } - } -} - -void VP8LColorSpaceTransform(int width, int height, int bits, int step, - uint32_t* const argb, uint32_t* image) { - const int max_tile_size = 1 << bits; - int tile_xsize = VP8LSubSampleSize(width, bits); - int tile_ysize = VP8LSubSampleSize(height, bits); - int accumulated_red_histo[256] = { 0 }; - int accumulated_blue_histo[256] = { 0 }; - int tile_y; - int tile_x; - Multipliers prevX; - Multipliers prevY; - MultipliersClear(&prevY); - MultipliersClear(&prevX); - for (tile_y = 0; tile_y < tile_ysize; ++tile_y) { - for (tile_x = 0; tile_x < tile_xsize; ++tile_x) { - Multipliers color_transform; - int all_x_max; - int y; - const int tile_y_offset = tile_y * max_tile_size; - const int tile_x_offset = tile_x * max_tile_size; - if (tile_y != 0) { - ColorCodeToMultipliers(image[tile_y * tile_xsize + tile_x - 1], &prevX); - ColorCodeToMultipliers(image[(tile_y - 1) * tile_xsize + tile_x], - &prevY); - } else if (tile_x != 0) { - ColorCodeToMultipliers(image[tile_y * tile_xsize + tile_x - 1], &prevX); - } - color_transform = - GetBestColorTransformForTile(tile_x, tile_y, bits, - prevX, prevY, - step, width, height, - &accumulated_red_histo[0], - &accumulated_blue_histo[0], - argb); - image[tile_y * tile_xsize + tile_x] = - MultipliersToColorCode(&color_transform); - CopyTileWithColorTransform(width, height, tile_x, tile_y, bits, - color_transform, argb); - - // Gather accumulated histogram data. - all_x_max = tile_x_offset + max_tile_size; - if (all_x_max > width) { - all_x_max = width; - } - for (y = 0; y < max_tile_size; ++y) { - int ix; - int all_x; - int all_y = tile_y_offset + y; - if (all_y >= height) { - break; - } - ix = all_y * width + tile_x_offset; - for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) { - if (ix >= 2 && - argb[ix] == argb[ix - 2] && - argb[ix] == argb[ix - 1]) { - continue; // repeated pixels are handled by backward references - } - if (ix >= width + 2 && - argb[ix - 2] == argb[ix - width - 2] && - argb[ix - 1] == argb[ix - width - 1] && - argb[ix] == argb[ix - width]) { - continue; // repeated pixels are handled by backward references - } - ++accumulated_red_histo[(argb[ix] >> 16) & 0xff]; - ++accumulated_blue_histo[argb[ix] & 0xff]; - } - } - } + data[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue); } } @@ -898,7 +281,10 @@ void VP8LColorSpaceTransform(int width, int height, int bits, int step, static void ColorSpaceInverseTransform(const VP8LTransform* const transform, int y_start, int y_end, uint32_t* data) { const int width = transform->xsize_; - const int mask = (1 << transform->bits_) - 1; + const int tile_width = 1 << transform->bits_; + const int mask = tile_width - 1; + const int safe_width = width & ~mask; + const int remaining_width = width - safe_width; const int tiles_per_row = VP8LSubSampleSize(width, transform->bits_); int y = y_start; const uint32_t* pred_row = @@ -906,68 +292,89 @@ static void ColorSpaceInverseTransform(const VP8LTransform* const transform, while (y < y_end) { const uint32_t* pred = pred_row; - Multipliers m = { 0, 0, 0 }; - int x; - - for (x = 0; x < width; ++x) { - if ((x & mask) == 0) ColorCodeToMultipliers(*pred++, &m); - data[x] = TransformColor(&m, data[x], 1); + VP8LMultipliers m = { 0, 0, 0 }; + const uint32_t* const data_safe_end = data + safe_width; + const uint32_t* const data_end = data + width; + while (data < data_safe_end) { + ColorCodeToMultipliers(*pred++, &m); + VP8LTransformColorInverse(&m, data, tile_width); + data += tile_width; + } + if (data < data_end) { // Left-overs using C-version. + ColorCodeToMultipliers(*pred++, &m); + VP8LTransformColorInverse(&m, data, remaining_width); + data += remaining_width; } - data += width; ++y; - if ((y & mask) == 0) pred_row += tiles_per_row;; + if ((y & mask) == 0) pred_row += tiles_per_row; } } // Separate out pixels packed together using pixel-bundling. -static void ColorIndexInverseTransform( - const VP8LTransform* const transform, - int y_start, int y_end, const uint32_t* src, uint32_t* dst) { - int y; - const int bits_per_pixel = 8 >> transform->bits_; - const int width = transform->xsize_; - const uint32_t* const color_map = transform->data_; - if (bits_per_pixel < 8) { - const int pixels_per_byte = 1 << transform->bits_; - const int count_mask = pixels_per_byte - 1; - const uint32_t bit_mask = (1 << bits_per_pixel) - 1; - for (y = y_start; y < y_end; ++y) { - uint32_t packed_pixels = 0; - int x; - for (x = 0; x < width; ++x) { - // We need to load fresh 'packed_pixels' once every 'pixels_per_byte' - // increments of x. Fortunately, pixels_per_byte is a power of 2, so - // can just use a mask for that, instead of decrementing a counter. - if ((x & count_mask) == 0) packed_pixels = ((*src++) >> 8) & 0xff; - *dst++ = color_map[packed_pixels & bit_mask]; - packed_pixels >>= bits_per_pixel; - } - } - } else { - for (y = y_start; y < y_end; ++y) { - int x; - for (x = 0; x < width; ++x) { - *dst++ = color_map[((*src++) >> 8) & 0xff]; - } - } - } -} +// We define two methods for ARGB data (uint32_t) and alpha-only data (uint8_t). +#define COLOR_INDEX_INVERSE(FUNC_NAME, F_NAME, STATIC_DECL, TYPE, BIT_SUFFIX, \ + GET_INDEX, GET_VALUE) \ +static void F_NAME(const TYPE* src, const uint32_t* const color_map, \ + TYPE* dst, int y_start, int y_end, int width) { \ + int y; \ + for (y = y_start; y < y_end; ++y) { \ + int x; \ + for (x = 0; x < width; ++x) { \ + *dst++ = GET_VALUE(color_map[GET_INDEX(*src++)]); \ + } \ + } \ +} \ +STATIC_DECL void FUNC_NAME(const VP8LTransform* const transform, \ + int y_start, int y_end, const TYPE* src, \ + TYPE* dst) { \ + int y; \ + const int bits_per_pixel = 8 >> transform->bits_; \ + const int width = transform->xsize_; \ + const uint32_t* const color_map = transform->data_; \ + if (bits_per_pixel < 8) { \ + const int pixels_per_byte = 1 << transform->bits_; \ + const int count_mask = pixels_per_byte - 1; \ + const uint32_t bit_mask = (1 << bits_per_pixel) - 1; \ + for (y = y_start; y < y_end; ++y) { \ + uint32_t packed_pixels = 0; \ + int x; \ + for (x = 0; x < width; ++x) { \ + /* We need to load fresh 'packed_pixels' once every */ \ + /* 'pixels_per_byte' increments of x. Fortunately, pixels_per_byte */ \ + /* is a power of 2, so can just use a mask for that, instead of */ \ + /* decrementing a counter. */ \ + if ((x & count_mask) == 0) packed_pixels = GET_INDEX(*src++); \ + *dst++ = GET_VALUE(color_map[packed_pixels & bit_mask]); \ + packed_pixels >>= bits_per_pixel; \ + } \ + } \ + } else { \ + VP8LMapColor##BIT_SUFFIX(src, color_map, dst, y_start, y_end, width); \ + } \ +} + +COLOR_INDEX_INVERSE(ColorIndexInverseTransform, MapARGB, static, uint32_t, 32b, + VP8GetARGBIndex, VP8GetARGBValue) +COLOR_INDEX_INVERSE(VP8LColorIndexInverseTransformAlpha, MapAlpha, , uint8_t, + 8b, VP8GetAlphaIndex, VP8GetAlphaValue) + +#undef COLOR_INDEX_INVERSE void VP8LInverseTransform(const VP8LTransform* const transform, int row_start, int row_end, const uint32_t* const in, uint32_t* const out) { + const int width = transform->xsize_; assert(row_start < row_end); assert(row_end <= transform->ysize_); switch (transform->type_) { case SUBTRACT_GREEN: - AddGreenToBlueAndRed(transform, row_start, row_end, out); + VP8LAddGreenToBlueAndRed(out, (row_end - row_start) * width); break; case PREDICTOR_TRANSFORM: PredictorInverseTransform(transform, row_start, row_end, out); if (row_end != transform->ysize_) { // The last predicted row in this iteration will be the top-pred row // for the first row in next iteration. - const int width = transform->xsize_; memcpy(out - width, out + (row_end - row_start - 1) * width, width * sizeof(*out)); } @@ -982,7 +389,7 @@ void VP8LInverseTransform(const VP8LTransform* const transform, // Also, note that this is the only transform that applies on // the effective width of VP8LSubSampleSize(xsize_, bits_). All other // transforms work on effective width of xsize_. - const int out_stride = (row_end - row_start) * transform->xsize_; + const int out_stride = (row_end - row_start) * width; const int in_stride = (row_end - row_start) * VP8LSubSampleSize(transform->xsize_, transform->bits_); uint32_t* const src = out + out_stride - in_stride; @@ -1006,8 +413,8 @@ static int is_big_endian(void) { return (tmp.b[0] != 1); } -static void ConvertBGRAToRGB(const uint32_t* src, - int num_pixels, uint8_t* dst) { +void VP8LConvertBGRAToRGB_C(const uint32_t* src, + int num_pixels, uint8_t* dst) { const uint32_t* const src_end = src + num_pixels; while (src < src_end) { const uint32_t argb = *src++; @@ -1017,8 +424,8 @@ static void ConvertBGRAToRGB(const uint32_t* src, } } -static void ConvertBGRAToRGBA(const uint32_t* src, - int num_pixels, uint8_t* dst) { +void VP8LConvertBGRAToRGBA_C(const uint32_t* src, + int num_pixels, uint8_t* dst) { const uint32_t* const src_end = src + num_pixels; while (src < src_end) { const uint32_t argb = *src++; @@ -1029,28 +436,42 @@ static void ConvertBGRAToRGBA(const uint32_t* src, } } -static void ConvertBGRAToRGBA4444(const uint32_t* src, - int num_pixels, uint8_t* dst) { +void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src, + int num_pixels, uint8_t* dst) { const uint32_t* const src_end = src + num_pixels; while (src < src_end) { const uint32_t argb = *src++; - *dst++ = ((argb >> 16) & 0xf0) | ((argb >> 12) & 0xf); - *dst++ = ((argb >> 0) & 0xf0) | ((argb >> 28) & 0xf); + const uint8_t rg = ((argb >> 16) & 0xf0) | ((argb >> 12) & 0xf); + const uint8_t ba = ((argb >> 0) & 0xf0) | ((argb >> 28) & 0xf); +#ifdef WEBP_SWAP_16BIT_CSP + *dst++ = ba; + *dst++ = rg; +#else + *dst++ = rg; + *dst++ = ba; +#endif } } -static void ConvertBGRAToRGB565(const uint32_t* src, - int num_pixels, uint8_t* dst) { +void VP8LConvertBGRAToRGB565_C(const uint32_t* src, + int num_pixels, uint8_t* dst) { const uint32_t* const src_end = src + num_pixels; while (src < src_end) { const uint32_t argb = *src++; - *dst++ = ((argb >> 16) & 0xf8) | ((argb >> 13) & 0x7); - *dst++ = ((argb >> 5) & 0xe0) | ((argb >> 3) & 0x1f); + const uint8_t rg = ((argb >> 16) & 0xf8) | ((argb >> 13) & 0x7); + const uint8_t gb = ((argb >> 5) & 0xe0) | ((argb >> 3) & 0x1f); +#ifdef WEBP_SWAP_16BIT_CSP + *dst++ = gb; + *dst++ = rg; +#else + *dst++ = rg; + *dst++ = gb; +#endif } } -static void ConvertBGRAToBGR(const uint32_t* src, - int num_pixels, uint8_t* dst) { +void VP8LConvertBGRAToBGR_C(const uint32_t* src, + int num_pixels, uint8_t* dst) { const uint32_t* const src_end = src + num_pixels; while (src < src_end) { const uint32_t argb = *src++; @@ -1065,21 +486,24 @@ static void CopyOrSwap(const uint32_t* src, int num_pixels, uint8_t* dst, if (is_big_endian() == swap_on_big_endian) { const uint32_t* const src_end = src + num_pixels; while (src < src_end) { - uint32_t argb = *src++; -#if !defined(__BIG_ENDIAN__) && (defined(__i386__) || defined(__x86_64__)) - __asm__ volatile("bswap %0" : "=r"(argb) : "0"(argb)); - *(uint32_t*)dst = argb; - dst += sizeof(argb); -#elif !defined(__BIG_ENDIAN__) && defined(_MSC_VER) - argb = _byteswap_ulong(argb); - *(uint32_t*)dst = argb; - dst += sizeof(argb); -#else - *dst++ = (argb >> 24) & 0xff; - *dst++ = (argb >> 16) & 0xff; - *dst++ = (argb >> 8) & 0xff; - *dst++ = (argb >> 0) & 0xff; + const uint32_t argb = *src++; + +#if !defined(WORDS_BIGENDIAN) +#if !defined(WEBP_REFERENCE_IMPLEMENTATION) + *(uint32_t*)dst = BSwap32(argb); +#else // WEBP_REFERENCE_IMPLEMENTATION + dst[0] = (argb >> 24) & 0xff; + dst[1] = (argb >> 16) & 0xff; + dst[2] = (argb >> 8) & 0xff; + dst[3] = (argb >> 0) & 0xff; #endif +#else // WORDS_BIGENDIAN + dst[0] = (argb >> 0) & 0xff; + dst[1] = (argb >> 8) & 0xff; + dst[2] = (argb >> 16) & 0xff; + dst[3] = (argb >> 24) & 0xff; +#endif + dst += sizeof(argb); } } else { memcpy(dst, src, num_pixels * sizeof(*src)); @@ -1090,17 +514,17 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels, WEBP_CSP_MODE out_colorspace, uint8_t* const rgba) { switch (out_colorspace) { case MODE_RGB: - ConvertBGRAToRGB(in_data, num_pixels, rgba); + VP8LConvertBGRAToRGB(in_data, num_pixels, rgba); break; case MODE_RGBA: - ConvertBGRAToRGBA(in_data, num_pixels, rgba); + VP8LConvertBGRAToRGBA(in_data, num_pixels, rgba); break; case MODE_rgbA: - ConvertBGRAToRGBA(in_data, num_pixels, rgba); + VP8LConvertBGRAToRGBA(in_data, num_pixels, rgba); WebPApplyAlphaMultiply(rgba, 0, num_pixels, 1, 0); break; case MODE_BGR: - ConvertBGRAToBGR(in_data, num_pixels, rgba); + VP8LConvertBGRAToBGR(in_data, num_pixels, rgba); break; case MODE_BGRA: CopyOrSwap(in_data, num_pixels, rgba, 1); @@ -1117,14 +541,14 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels, WebPApplyAlphaMultiply(rgba, 1, num_pixels, 1, 0); break; case MODE_RGBA_4444: - ConvertBGRAToRGBA4444(in_data, num_pixels, rgba); + VP8LConvertBGRAToRGBA4444(in_data, num_pixels, rgba); break; case MODE_rgbA_4444: - ConvertBGRAToRGBA4444(in_data, num_pixels, rgba); + VP8LConvertBGRAToRGBA4444(in_data, num_pixels, rgba); WebPApplyAlphaMultiply4444(rgba, num_pixels, 1, 0); break; case MODE_RGB_565: - ConvertBGRAToRGB565(in_data, num_pixels, rgba); + VP8LConvertBGRAToRGB565(in_data, num_pixels, rgba); break; default: assert(0); // Code flow should not reach here. @@ -1133,6 +557,79 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels, //------------------------------------------------------------------------------ -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" +VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed; +VP8LPredictorFunc VP8LPredictors[16]; + +VP8LTransformColorFunc VP8LTransformColorInverse; + +VP8LConvertFunc VP8LConvertBGRAToRGB; +VP8LConvertFunc VP8LConvertBGRAToRGBA; +VP8LConvertFunc VP8LConvertBGRAToRGBA4444; +VP8LConvertFunc VP8LConvertBGRAToRGB565; +VP8LConvertFunc VP8LConvertBGRAToBGR; + +VP8LMapARGBFunc VP8LMapColor32b; +VP8LMapAlphaFunc VP8LMapColor8b; + +extern void VP8LDspInitSSE2(void); +extern void VP8LDspInitNEON(void); +extern void VP8LDspInitMIPSdspR2(void); + +static volatile VP8CPUInfo lossless_last_cpuinfo_used = + (VP8CPUInfo)&lossless_last_cpuinfo_used; + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInit(void) { + if (lossless_last_cpuinfo_used == VP8GetCPUInfo) return; + + VP8LPredictors[0] = Predictor0; + VP8LPredictors[1] = Predictor1; + VP8LPredictors[2] = Predictor2; + VP8LPredictors[3] = Predictor3; + VP8LPredictors[4] = Predictor4; + VP8LPredictors[5] = Predictor5; + VP8LPredictors[6] = Predictor6; + VP8LPredictors[7] = Predictor7; + VP8LPredictors[8] = Predictor8; + VP8LPredictors[9] = Predictor9; + VP8LPredictors[10] = Predictor10; + VP8LPredictors[11] = Predictor11; + VP8LPredictors[12] = Predictor12; + VP8LPredictors[13] = Predictor13; + VP8LPredictors[14] = Predictor0; // <- padding security sentinels + VP8LPredictors[15] = Predictor0; + + VP8LAddGreenToBlueAndRed = VP8LAddGreenToBlueAndRed_C; + + VP8LTransformColorInverse = VP8LTransformColorInverse_C; + + VP8LConvertBGRAToRGB = VP8LConvertBGRAToRGB_C; + VP8LConvertBGRAToRGBA = VP8LConvertBGRAToRGBA_C; + VP8LConvertBGRAToRGBA4444 = VP8LConvertBGRAToRGBA4444_C; + VP8LConvertBGRAToRGB565 = VP8LConvertBGRAToRGB565_C; + VP8LConvertBGRAToBGR = VP8LConvertBGRAToBGR_C; + + VP8LMapColor32b = MapARGB; + VP8LMapColor8b = MapAlpha; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8LDspInitSSE2(); + } +#endif +#if defined(WEBP_USE_NEON) + if (VP8GetCPUInfo(kNEON)) { + VP8LDspInitNEON(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8LDspInitMIPSdspR2(); + } #endif + } + lossless_last_cpuinfo_used = VP8GetCPUInfo; +} + +//------------------------------------------------------------------------------ diff --git a/drivers/webp/dsp/lossless.h b/drivers/webp/dsp/lossless.h index 7c7d5555ed..149c6a01d3 100644 --- a/drivers/webp/dsp/lossless.h +++ b/drivers/webp/dsp/lossless.h @@ -1,8 +1,10 @@ // Copyright 2012 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // Image transforms and color space conversion methods for lossless decoder. @@ -13,15 +15,42 @@ #ifndef WEBP_DSP_LOSSLESS_H_ #define WEBP_DSP_LOSSLESS_H_ -#include "../types.h" -#include "../decode.h" +#include "webp/types.h" +#include "webp/decode.h" -#if defined(__cplusplus) || defined(c_plusplus) +#include "../enc/histogram.h" +#include "../utils/utils.h" + +#ifdef __cplusplus extern "C" { #endif +#ifdef WEBP_EXPERIMENTAL_FEATURES +#include "../enc/delta_palettization.h" +#endif // WEBP_EXPERIMENTAL_FEATURES + +// Not a trivial literal symbol. +#define VP8L_NON_TRIVIAL_SYM (0xffffffff) + //------------------------------------------------------------------------------ -// Image transforms. +// Decoding + +typedef uint32_t (*VP8LPredictorFunc)(uint32_t left, const uint32_t* const top); +extern VP8LPredictorFunc VP8LPredictors[16]; + +typedef void (*VP8LProcessBlueAndRedFunc)(uint32_t* argb_data, int num_pixels); +extern VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed; + +typedef struct { + // Note: the members are uint8_t, so that any negative values are + // automatically converted to "mod 256" values. + uint8_t green_to_red_; + uint8_t green_to_blue_; + uint8_t red_to_blue_; +} VP8LMultipliers; +typedef void (*VP8LTransformColorFunc)(const VP8LMultipliers* const m, + uint32_t* argb_data, int num_pixels); +extern VP8LTransformColorFunc VP8LTransformColorInverse; struct VP8LTransform; // Defined in dec/vp8li.h. @@ -33,23 +62,110 @@ void VP8LInverseTransform(const struct VP8LTransform* const transform, int row_start, int row_end, const uint32_t* const in, uint32_t* const out); -// Subtracts green from blue and red channels. -void VP8LSubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs); - -void VP8LResidualImage(int width, int height, int bits, - uint32_t* const argb, uint32_t* const argb_scratch, - uint32_t* const image); - -void VP8LColorSpaceTransform(int width, int height, int bits, int step, - uint32_t* const argb, uint32_t* image); - -//------------------------------------------------------------------------------ // Color space conversion. +typedef void (*VP8LConvertFunc)(const uint32_t* src, int num_pixels, + uint8_t* dst); +extern VP8LConvertFunc VP8LConvertBGRAToRGB; +extern VP8LConvertFunc VP8LConvertBGRAToRGBA; +extern VP8LConvertFunc VP8LConvertBGRAToRGBA4444; +extern VP8LConvertFunc VP8LConvertBGRAToRGB565; +extern VP8LConvertFunc VP8LConvertBGRAToBGR; // Converts from BGRA to other color spaces. void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels, WEBP_CSP_MODE out_colorspace, uint8_t* const rgba); +// color mapping related functions. +static WEBP_INLINE uint32_t VP8GetARGBIndex(uint32_t idx) { + return (idx >> 8) & 0xff; +} + +static WEBP_INLINE uint8_t VP8GetAlphaIndex(uint8_t idx) { + return idx; +} + +static WEBP_INLINE uint32_t VP8GetARGBValue(uint32_t val) { + return val; +} + +static WEBP_INLINE uint8_t VP8GetAlphaValue(uint32_t val) { + return (val >> 8) & 0xff; +} + +typedef void (*VP8LMapARGBFunc)(const uint32_t* src, + const uint32_t* const color_map, + uint32_t* dst, int y_start, + int y_end, int width); +typedef void (*VP8LMapAlphaFunc)(const uint8_t* src, + const uint32_t* const color_map, + uint8_t* dst, int y_start, + int y_end, int width); + +extern VP8LMapARGBFunc VP8LMapColor32b; +extern VP8LMapAlphaFunc VP8LMapColor8b; + +// Similar to the static method ColorIndexInverseTransform() that is part of +// lossless.c, but used only for alpha decoding. It takes uint8_t (rather than +// uint32_t) arguments for 'src' and 'dst'. +void VP8LColorIndexInverseTransformAlpha( + const struct VP8LTransform* const transform, int y_start, int y_end, + const uint8_t* src, uint8_t* dst); + +// Expose some C-only fallback functions +void VP8LTransformColorInverse_C(const VP8LMultipliers* const m, + uint32_t* data, int num_pixels); + +void VP8LConvertBGRAToRGB_C(const uint32_t* src, int num_pixels, uint8_t* dst); +void VP8LConvertBGRAToRGBA_C(const uint32_t* src, int num_pixels, uint8_t* dst); +void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src, + int num_pixels, uint8_t* dst); +void VP8LConvertBGRAToRGB565_C(const uint32_t* src, + int num_pixels, uint8_t* dst); +void VP8LConvertBGRAToBGR_C(const uint32_t* src, int num_pixels, uint8_t* dst); +void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels); + +// Must be called before calling any of the above methods. +void VP8LDspInit(void); + +//------------------------------------------------------------------------------ +// Encoding + +extern VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed; +extern VP8LTransformColorFunc VP8LTransformColor; +typedef void (*VP8LCollectColorBlueTransformsFunc)( + const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_blue, int red_to_blue, int histo[]); +extern VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms; + +typedef void (*VP8LCollectColorRedTransformsFunc)( + const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]); +extern VP8LCollectColorRedTransformsFunc VP8LCollectColorRedTransforms; + +// Expose some C-only fallback functions +void VP8LTransformColor_C(const VP8LMultipliers* const m, + uint32_t* data, int num_pixels); +void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels); +void VP8LCollectColorRedTransforms_C(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]); +void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_blue, int red_to_blue, + int histo[]); + +//------------------------------------------------------------------------------ +// Image transforms. + +void VP8LResidualImage(int width, int height, int bits, int low_effort, + uint32_t* const argb, uint32_t* const argb_scratch, + uint32_t* const image); + +void VP8LColorSpaceTransform(int width, int height, int bits, int quality, + uint32_t* const argb, uint32_t* image); + //------------------------------------------------------------------------------ // Misc methods. @@ -59,10 +175,136 @@ static WEBP_INLINE uint32_t VP8LSubSampleSize(uint32_t size, return (size + (1 << sampling_bits) - 1) >> sampling_bits; } -// Faster logarithm for integers, with the property of log2(0) == 0. -float VP8LFastLog2(int v); +// ----------------------------------------------------------------------------- +// Faster logarithm for integers. Small values use a look-up table. +#define LOG_LOOKUP_IDX_MAX 256 +extern const float kLog2Table[LOG_LOOKUP_IDX_MAX]; +extern const float kSLog2Table[LOG_LOOKUP_IDX_MAX]; +typedef float (*VP8LFastLog2SlowFunc)(uint32_t v); + +extern VP8LFastLog2SlowFunc VP8LFastLog2Slow; +extern VP8LFastLog2SlowFunc VP8LFastSLog2Slow; + +static WEBP_INLINE float VP8LFastLog2(uint32_t v) { + return (v < LOG_LOOKUP_IDX_MAX) ? kLog2Table[v] : VP8LFastLog2Slow(v); +} // Fast calculation of v * log2(v) for integer input. -static WEBP_INLINE float VP8LFastSLog2(int v) { return VP8LFastLog2(v) * v; } +static WEBP_INLINE float VP8LFastSLog2(uint32_t v) { + return (v < LOG_LOOKUP_IDX_MAX) ? kSLog2Table[v] : VP8LFastSLog2Slow(v); +} + +// ----------------------------------------------------------------------------- +// Huffman-cost related functions. + +typedef double (*VP8LCostFunc)(const uint32_t* population, int length); +typedef double (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y, + int length); + +extern VP8LCostFunc VP8LExtraCost; +extern VP8LCostCombinedFunc VP8LExtraCostCombined; + +typedef struct { // small struct to hold counters + int counts[2]; // index: 0=zero steak, 1=non-zero streak + int streaks[2][2]; // [zero/non-zero][streak<3 / streak>=3] +} VP8LStreaks; + +typedef VP8LStreaks (*VP8LCostCountFunc)(const uint32_t* population, + int length); +typedef VP8LStreaks (*VP8LCostCombinedCountFunc)(const uint32_t* X, + const uint32_t* Y, int length); + +extern VP8LCostCountFunc VP8LHuffmanCostCount; +extern VP8LCostCombinedCountFunc VP8LHuffmanCostCombinedCount; + +// Get the symbol entropy for the distribution 'population'. +// Set 'trivial_sym', if there's only one symbol present in the distribution. +double VP8LPopulationCost(const uint32_t* const population, int length, + uint32_t* const trivial_sym); + +// Get the combined symbol entropy for the distributions 'X' and 'Y'. +double VP8LGetCombinedEntropy(const uint32_t* const X, + const uint32_t* const Y, int length); + +double VP8LBitsEntropy(const uint32_t* const array, int n, + uint32_t* const trivial_symbol); + +// Estimate how many bits the combined entropy of literals and distance +// approximately maps to. +double VP8LHistogramEstimateBits(const VP8LHistogram* const p); + +// This function estimates the cost in bits excluding the bits needed to +// represent the entropy code itself. +double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p); + +typedef void (*VP8LHistogramAddFunc)(const VP8LHistogram* const a, + const VP8LHistogram* const b, + VP8LHistogram* const out); +extern VP8LHistogramAddFunc VP8LHistogramAdd; + +// ----------------------------------------------------------------------------- +// PrefixEncode() + +static WEBP_INLINE int VP8LBitsLog2Ceiling(uint32_t n) { + const int log_floor = BitsLog2Floor(n); + if (n == (n & ~(n - 1))) // zero or a power of two. + return log_floor; + else + return log_floor + 1; +} + +// Splitting of distance and length codes into prefixes and +// extra bits. The prefixes are encoded with an entropy code +// while the extra bits are stored just as normal bits. +static WEBP_INLINE void VP8LPrefixEncodeBitsNoLUT(int distance, int* const code, + int* const extra_bits) { + const int highest_bit = BitsLog2Floor(--distance); + const int second_highest_bit = (distance >> (highest_bit - 1)) & 1; + *extra_bits = highest_bit - 1; + *code = 2 * highest_bit + second_highest_bit; +} + +static WEBP_INLINE void VP8LPrefixEncodeNoLUT(int distance, int* const code, + int* const extra_bits, + int* const extra_bits_value) { + const int highest_bit = BitsLog2Floor(--distance); + const int second_highest_bit = (distance >> (highest_bit - 1)) & 1; + *extra_bits = highest_bit - 1; + *extra_bits_value = distance & ((1 << *extra_bits) - 1); + *code = 2 * highest_bit + second_highest_bit; +} + +#define PREFIX_LOOKUP_IDX_MAX 512 +typedef struct { + int8_t code_; + int8_t extra_bits_; +} VP8LPrefixCode; + +// These tables are derived using VP8LPrefixEncodeNoLUT. +extern const VP8LPrefixCode kPrefixEncodeCode[PREFIX_LOOKUP_IDX_MAX]; +extern const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX]; +static WEBP_INLINE void VP8LPrefixEncodeBits(int distance, int* const code, + int* const extra_bits) { + if (distance < PREFIX_LOOKUP_IDX_MAX) { + const VP8LPrefixCode prefix_code = kPrefixEncodeCode[distance]; + *code = prefix_code.code_; + *extra_bits = prefix_code.extra_bits_; + } else { + VP8LPrefixEncodeBitsNoLUT(distance, code, extra_bits); + } +} + +static WEBP_INLINE void VP8LPrefixEncode(int distance, int* const code, + int* const extra_bits, + int* const extra_bits_value) { + if (distance < PREFIX_LOOKUP_IDX_MAX) { + const VP8LPrefixCode prefix_code = kPrefixEncodeCode[distance]; + *code = prefix_code.code_; + *extra_bits = prefix_code.extra_bits_; + *extra_bits_value = kPrefixEncodeExtraBitsValue[distance]; + } else { + VP8LPrefixEncodeNoLUT(distance, code, extra_bits, extra_bits_value); + } +} // In-place difference of each component with mod 256. static WEBP_INLINE uint32_t VP8LSubPixels(uint32_t a, uint32_t b) { @@ -73,9 +315,15 @@ static WEBP_INLINE uint32_t VP8LSubPixels(uint32_t a, uint32_t b) { return (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu); } +void VP8LBundleColorMap(const uint8_t* const row, int width, + int xbits, uint32_t* const dst); + +// Must be called before calling any of the above methods. +void VP8LEncDspInit(void); + //------------------------------------------------------------------------------ -#if defined(__cplusplus) || defined(c_plusplus) +#ifdef __cplusplus } // extern "C" #endif diff --git a/drivers/webp/dsp/lossless_enc.c b/drivers/webp/dsp/lossless_enc.c new file mode 100644 index 0000000000..b3036f5384 --- /dev/null +++ b/drivers/webp/dsp/lossless_enc.c @@ -0,0 +1,1305 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Image transform methods for lossless encoder. +// +// Authors: Vikas Arora (vikaas.arora@gmail.com) +// Jyrki Alakuijala (jyrki@google.com) +// Urvang Joshi (urvang@google.com) + +#include "./dsp.h" + +#include <math.h> +#include <stdlib.h> +#include "../dec/vp8li.h" +#include "../utils/endian_inl.h" +#include "./lossless.h" +#include "./yuv.h" + +#define MAX_DIFF_COST (1e30f) + +// lookup table for small values of log2(int) +const float kLog2Table[LOG_LOOKUP_IDX_MAX] = { + 0.0000000000000000f, 0.0000000000000000f, + 1.0000000000000000f, 1.5849625007211560f, + 2.0000000000000000f, 2.3219280948873621f, + 2.5849625007211560f, 2.8073549220576041f, + 3.0000000000000000f, 3.1699250014423121f, + 3.3219280948873621f, 3.4594316186372973f, + 3.5849625007211560f, 3.7004397181410921f, + 3.8073549220576041f, 3.9068905956085187f, + 4.0000000000000000f, 4.0874628412503390f, + 4.1699250014423121f, 4.2479275134435852f, + 4.3219280948873626f, 4.3923174227787606f, + 4.4594316186372973f, 4.5235619560570130f, + 4.5849625007211560f, 4.6438561897747243f, + 4.7004397181410917f, 4.7548875021634682f, + 4.8073549220576037f, 4.8579809951275718f, + 4.9068905956085187f, 4.9541963103868749f, + 5.0000000000000000f, 5.0443941193584533f, + 5.0874628412503390f, 5.1292830169449663f, + 5.1699250014423121f, 5.2094533656289501f, + 5.2479275134435852f, 5.2854022188622487f, + 5.3219280948873626f, 5.3575520046180837f, + 5.3923174227787606f, 5.4262647547020979f, + 5.4594316186372973f, 5.4918530963296747f, + 5.5235619560570130f, 5.5545888516776376f, + 5.5849625007211560f, 5.6147098441152083f, + 5.6438561897747243f, 5.6724253419714951f, + 5.7004397181410917f, 5.7279204545631987f, + 5.7548875021634682f, 5.7813597135246599f, + 5.8073549220576037f, 5.8328900141647412f, + 5.8579809951275718f, 5.8826430493618415f, + 5.9068905956085187f, 5.9307373375628866f, + 5.9541963103868749f, 5.9772799234999167f, + 6.0000000000000000f, 6.0223678130284543f, + 6.0443941193584533f, 6.0660891904577720f, + 6.0874628412503390f, 6.1085244567781691f, + 6.1292830169449663f, 6.1497471195046822f, + 6.1699250014423121f, 6.1898245588800175f, + 6.2094533656289501f, 6.2288186904958804f, + 6.2479275134435852f, 6.2667865406949010f, + 6.2854022188622487f, 6.3037807481771030f, + 6.3219280948873626f, 6.3398500028846243f, + 6.3575520046180837f, 6.3750394313469245f, + 6.3923174227787606f, 6.4093909361377017f, + 6.4262647547020979f, 6.4429434958487279f, + 6.4594316186372973f, 6.4757334309663976f, + 6.4918530963296747f, 6.5077946401986963f, + 6.5235619560570130f, 6.5391588111080309f, + 6.5545888516776376f, 6.5698556083309478f, + 6.5849625007211560f, 6.5999128421871278f, + 6.6147098441152083f, 6.6293566200796094f, + 6.6438561897747243f, 6.6582114827517946f, + 6.6724253419714951f, 6.6865005271832185f, + 6.7004397181410917f, 6.7142455176661224f, + 6.7279204545631987f, 6.7414669864011464f, + 6.7548875021634682f, 6.7681843247769259f, + 6.7813597135246599f, 6.7944158663501061f, + 6.8073549220576037f, 6.8201789624151878f, + 6.8328900141647412f, 6.8454900509443747f, + 6.8579809951275718f, 6.8703647195834047f, + 6.8826430493618415f, 6.8948177633079437f, + 6.9068905956085187f, 6.9188632372745946f, + 6.9307373375628866f, 6.9425145053392398f, + 6.9541963103868749f, 6.9657842846620869f, + 6.9772799234999167f, 6.9886846867721654f, + 7.0000000000000000f, 7.0112272554232539f, + 7.0223678130284543f, 7.0334230015374501f, + 7.0443941193584533f, 7.0552824355011898f, + 7.0660891904577720f, 7.0768155970508308f, + 7.0874628412503390f, 7.0980320829605263f, + 7.1085244567781691f, 7.1189410727235076f, + 7.1292830169449663f, 7.1395513523987936f, + 7.1497471195046822f, 7.1598713367783890f, + 7.1699250014423121f, 7.1799090900149344f, + 7.1898245588800175f, 7.1996723448363644f, + 7.2094533656289501f, 7.2191685204621611f, + 7.2288186904958804f, 7.2384047393250785f, + 7.2479275134435852f, 7.2573878426926521f, + 7.2667865406949010f, 7.2761244052742375f, + 7.2854022188622487f, 7.2946207488916270f, + 7.3037807481771030f, 7.3128829552843557f, + 7.3219280948873626f, 7.3309168781146167f, + 7.3398500028846243f, 7.3487281542310771f, + 7.3575520046180837f, 7.3663222142458160f, + 7.3750394313469245f, 7.3837042924740519f, + 7.3923174227787606f, 7.4008794362821843f, + 7.4093909361377017f, 7.4178525148858982f, + 7.4262647547020979f, 7.4346282276367245f, + 7.4429434958487279f, 7.4512111118323289f, + 7.4594316186372973f, 7.4676055500829976f, + 7.4757334309663976f, 7.4838157772642563f, + 7.4918530963296747f, 7.4998458870832056f, + 7.5077946401986963f, 7.5156998382840427f, + 7.5235619560570130f, 7.5313814605163118f, + 7.5391588111080309f, 7.5468944598876364f, + 7.5545888516776376f, 7.5622424242210728f, + 7.5698556083309478f, 7.5774288280357486f, + 7.5849625007211560f, 7.5924570372680806f, + 7.5999128421871278f, 7.6073303137496104f, + 7.6147098441152083f, 7.6220518194563764f, + 7.6293566200796094f, 7.6366246205436487f, + 7.6438561897747243f, 7.6510516911789281f, + 7.6582114827517946f, 7.6653359171851764f, + 7.6724253419714951f, 7.6794800995054464f, + 7.6865005271832185f, 7.6934869574993252f, + 7.7004397181410917f, 7.7073591320808825f, + 7.7142455176661224f, 7.7210991887071855f, + 7.7279204545631987f, 7.7347096202258383f, + 7.7414669864011464f, 7.7481928495894605f, + 7.7548875021634682f, 7.7615512324444795f, + 7.7681843247769259f, 7.7747870596011736f, + 7.7813597135246599f, 7.7879025593914317f, + 7.7944158663501061f, 7.8008998999203047f, + 7.8073549220576037f, 7.8137811912170374f, + 7.8201789624151878f, 7.8265484872909150f, + 7.8328900141647412f, 7.8392037880969436f, + 7.8454900509443747f, 7.8517490414160571f, + 7.8579809951275718f, 7.8641861446542797f, + 7.8703647195834047f, 7.8765169465649993f, + 7.8826430493618415f, 7.8887432488982591f, + 7.8948177633079437f, 7.9008668079807486f, + 7.9068905956085187f, 7.9128893362299619f, + 7.9188632372745946f, 7.9248125036057812f, + 7.9307373375628866f, 7.9366379390025709f, + 7.9425145053392398f, 7.9483672315846778f, + 7.9541963103868749f, 7.9600019320680805f, + 7.9657842846620869f, 7.9715435539507719f, + 7.9772799234999167f, 7.9829935746943103f, + 7.9886846867721654f, 7.9943534368588577f +}; + +const float kSLog2Table[LOG_LOOKUP_IDX_MAX] = { + 0.00000000f, 0.00000000f, 2.00000000f, 4.75488750f, + 8.00000000f, 11.60964047f, 15.50977500f, 19.65148445f, + 24.00000000f, 28.52932501f, 33.21928095f, 38.05374781f, + 43.01955001f, 48.10571634f, 53.30296891f, 58.60335893f, + 64.00000000f, 69.48686830f, 75.05865003f, 80.71062276f, + 86.43856190f, 92.23866588f, 98.10749561f, 104.04192499f, + 110.03910002f, 116.09640474f, 122.21143267f, 128.38196256f, + 134.60593782f, 140.88144886f, 147.20671787f, 153.58008562f, + 160.00000000f, 166.46500594f, 172.97373660f, 179.52490559f, + 186.11730005f, 192.74977453f, 199.42124551f, 206.13068654f, + 212.87712380f, 219.65963219f, 226.47733176f, 233.32938445f, + 240.21499122f, 247.13338933f, 254.08384998f, 261.06567603f, + 268.07820003f, 275.12078236f, 282.19280949f, 289.29369244f, + 296.42286534f, 303.57978409f, 310.76392512f, 317.97478424f, + 325.21187564f, 332.47473081f, 339.76289772f, 347.07593991f, + 354.41343574f, 361.77497759f, 369.16017124f, 376.56863518f, + 384.00000000f, 391.45390785f, 398.93001188f, 406.42797576f, + 413.94747321f, 421.48818752f, 429.04981119f, 436.63204548f, + 444.23460010f, 451.85719280f, 459.49954906f, 467.16140179f, + 474.84249102f, 482.54256363f, 490.26137307f, 497.99867911f, + 505.75424759f, 513.52785023f, 521.31926438f, 529.12827280f, + 536.95466351f, 544.79822957f, 552.65876890f, 560.53608414f, + 568.42998244f, 576.34027536f, 584.26677867f, 592.20931226f, + 600.16769996f, 608.14176943f, 616.13135206f, 624.13628279f, + 632.15640007f, 640.19154569f, 648.24156472f, 656.30630539f, + 664.38561898f, 672.47935976f, 680.58738488f, 688.70955430f, + 696.84573069f, 704.99577935f, 713.15956818f, 721.33696754f, + 729.52785023f, 737.73209140f, 745.94956849f, 754.18016116f, + 762.42375127f, 770.68022275f, 778.94946161f, 787.23135586f, + 795.52579543f, 803.83267219f, 812.15187982f, 820.48331383f, + 828.82687147f, 837.18245171f, 845.54995518f, 853.92928416f, + 862.32034249f, 870.72303558f, 879.13727036f, 887.56295522f, + 896.00000000f, 904.44831595f, 912.90781569f, 921.37841320f, + 929.86002376f, 938.35256392f, 946.85595152f, 955.37010560f, + 963.89494641f, 972.43039537f, 980.97637504f, 989.53280911f, + 998.09962237f, 1006.67674069f, 1015.26409097f, 1023.86160116f, + 1032.46920021f, 1041.08681805f, 1049.71438560f, 1058.35183469f, + 1066.99909811f, 1075.65610955f, 1084.32280357f, 1092.99911564f, + 1101.68498204f, 1110.38033993f, 1119.08512727f, 1127.79928282f, + 1136.52274614f, 1145.25545758f, 1153.99735821f, 1162.74838989f, + 1171.50849518f, 1180.27761738f, 1189.05570047f, 1197.84268914f, + 1206.63852876f, 1215.44316535f, 1224.25654560f, 1233.07861684f, + 1241.90932703f, 1250.74862473f, 1259.59645914f, 1268.45278005f, + 1277.31753781f, 1286.19068338f, 1295.07216828f, 1303.96194457f, + 1312.85996488f, 1321.76618236f, 1330.68055071f, 1339.60302413f, + 1348.53355734f, 1357.47210556f, 1366.41862452f, 1375.37307041f, + 1384.33539991f, 1393.30557020f, 1402.28353887f, 1411.26926400f, + 1420.26270412f, 1429.26381818f, 1438.27256558f, 1447.28890615f, + 1456.31280014f, 1465.34420819f, 1474.38309138f, 1483.42941118f, + 1492.48312945f, 1501.54420843f, 1510.61261078f, 1519.68829949f, + 1528.77123795f, 1537.86138993f, 1546.95871952f, 1556.06319119f, + 1565.17476976f, 1574.29342040f, 1583.41910860f, 1592.55180020f, + 1601.69146137f, 1610.83805860f, 1619.99155871f, 1629.15192882f, + 1638.31913637f, 1647.49314911f, 1656.67393509f, 1665.86146266f, + 1675.05570047f, 1684.25661744f, 1693.46418280f, 1702.67836605f, + 1711.89913698f, 1721.12646563f, 1730.36032233f, 1739.60067768f, + 1748.84750254f, 1758.10076802f, 1767.36044551f, 1776.62650662f, + 1785.89892323f, 1795.17766747f, 1804.46271172f, 1813.75402857f, + 1823.05159087f, 1832.35537170f, 1841.66534438f, 1850.98148244f, + 1860.30375965f, 1869.63214999f, 1878.96662767f, 1888.30716711f, + 1897.65374295f, 1907.00633003f, 1916.36490342f, 1925.72943838f, + 1935.09991037f, 1944.47629506f, 1953.85856831f, 1963.24670620f, + 1972.64068498f, 1982.04048108f, 1991.44607117f, 2000.85743204f, + 2010.27454072f, 2019.69737440f, 2029.12591044f, 2038.56012640f +}; + +const VP8LPrefixCode kPrefixEncodeCode[PREFIX_LOOKUP_IDX_MAX] = { + { 0, 0}, { 0, 0}, { 1, 0}, { 2, 0}, { 3, 0}, { 4, 1}, { 4, 1}, { 5, 1}, + { 5, 1}, { 6, 2}, { 6, 2}, { 6, 2}, { 6, 2}, { 7, 2}, { 7, 2}, { 7, 2}, + { 7, 2}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, + { 8, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, + { 9, 3}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, + {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, + {10, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, + {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, + {11, 4}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, + {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, + {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, + {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, + {12, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, + {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, + {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, + {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, + {13, 5}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, +}; + +const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX] = { + 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 2, 3, 0, 1, 2, 3, + 0, 1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4, 5, 6, 7, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, + 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, + 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, + 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, + 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, + 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, + 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, + 127, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, + 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, + 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, + 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, + 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126 +}; + +// The threshold till approximate version of log_2 can be used. +// Practically, we can get rid of the call to log() as the two values match to +// very high degree (the ratio of these two is 0.99999x). +// Keeping a high threshold for now. +#define APPROX_LOG_WITH_CORRECTION_MAX 65536 +#define APPROX_LOG_MAX 4096 +#define LOG_2_RECIPROCAL 1.44269504088896338700465094007086 +static float FastSLog2Slow(uint32_t v) { + assert(v >= LOG_LOOKUP_IDX_MAX); + if (v < APPROX_LOG_WITH_CORRECTION_MAX) { + int log_cnt = 0; + uint32_t y = 1; + int correction = 0; + const float v_f = (float)v; + const uint32_t orig_v = v; + do { + ++log_cnt; + v = v >> 1; + y = y << 1; + } while (v >= LOG_LOOKUP_IDX_MAX); + // vf = (2^log_cnt) * Xf; where y = 2^log_cnt and Xf < 256 + // Xf = floor(Xf) * (1 + (v % y) / v) + // log2(Xf) = log2(floor(Xf)) + log2(1 + (v % y) / v) + // The correction factor: log(1 + d) ~ d; for very small d values, so + // log2(1 + (v % y) / v) ~ LOG_2_RECIPROCAL * (v % y)/v + // LOG_2_RECIPROCAL ~ 23/16 + correction = (23 * (orig_v & (y - 1))) >> 4; + return v_f * (kLog2Table[v] + log_cnt) + correction; + } else { + return (float)(LOG_2_RECIPROCAL * v * log((double)v)); + } +} + +static float FastLog2Slow(uint32_t v) { + assert(v >= LOG_LOOKUP_IDX_MAX); + if (v < APPROX_LOG_WITH_CORRECTION_MAX) { + int log_cnt = 0; + uint32_t y = 1; + const uint32_t orig_v = v; + double log_2; + do { + ++log_cnt; + v = v >> 1; + y = y << 1; + } while (v >= LOG_LOOKUP_IDX_MAX); + log_2 = kLog2Table[v] + log_cnt; + if (orig_v >= APPROX_LOG_MAX) { + // Since the division is still expensive, add this correction factor only + // for large values of 'v'. + const int correction = (23 * (orig_v & (y - 1))) >> 4; + log_2 += (double)correction / orig_v; + } + return (float)log_2; + } else { + return (float)(LOG_2_RECIPROCAL * log((double)v)); + } +} + +// Mostly used to reduce code size + readability +static WEBP_INLINE int GetMin(int a, int b) { return (a > b) ? b : a; } + +//------------------------------------------------------------------------------ +// Methods to calculate Entropy (Shannon). + +static float PredictionCostSpatial(const int counts[256], int weight_0, + double exp_val) { + const int significant_symbols = 256 >> 4; + const double exp_decay_factor = 0.6; + double bits = weight_0 * counts[0]; + int i; + for (i = 1; i < significant_symbols; ++i) { + bits += exp_val * (counts[i] + counts[256 - i]); + exp_val *= exp_decay_factor; + } + return (float)(-0.1 * bits); +} + +// Compute the combined Shanon's entropy for distribution {X} and {X+Y} +static float CombinedShannonEntropy(const int X[256], const int Y[256]) { + int i; + double retval = 0.; + int sumX = 0, sumXY = 0; + for (i = 0; i < 256; ++i) { + const int x = X[i]; + const int xy = x + Y[i]; + if (x != 0) { + sumX += x; + retval -= VP8LFastSLog2(x); + sumXY += xy; + retval -= VP8LFastSLog2(xy); + } else if (xy != 0) { + sumXY += xy; + retval -= VP8LFastSLog2(xy); + } + } + retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY); + return (float)retval; +} + +static float PredictionCostSpatialHistogram(const int accumulated[4][256], + const int tile[4][256]) { + int i; + double retval = 0; + for (i = 0; i < 4; ++i) { + const double kExpValue = 0.94; + retval += PredictionCostSpatial(tile[i], 1, kExpValue); + retval += CombinedShannonEntropy(tile[i], accumulated[i]); + } + return (float)retval; +} + +static WEBP_INLINE double BitsEntropyRefine(int nonzeros, int sum, int max_val, + double retval) { + double mix; + if (nonzeros < 5) { + if (nonzeros <= 1) { + return 0; + } + // Two symbols, they will be 0 and 1 in a Huffman code. + // Let's mix in a bit of entropy to favor good clustering when + // distributions of these are combined. + if (nonzeros == 2) { + return 0.99 * sum + 0.01 * retval; + } + // No matter what the entropy says, we cannot be better than min_limit + // with Huffman coding. I am mixing a bit of entropy into the + // min_limit since it produces much better (~0.5 %) compression results + // perhaps because of better entropy clustering. + if (nonzeros == 3) { + mix = 0.95; + } else { + mix = 0.7; // nonzeros == 4. + } + } else { + mix = 0.627; + } + + { + double min_limit = 2 * sum - max_val; + min_limit = mix * min_limit + (1.0 - mix) * retval; + return (retval < min_limit) ? min_limit : retval; + } +} + +// Returns the entropy for the symbols in the input array. +// Also sets trivial_symbol to the code value, if the array has only one code +// value. Otherwise, set it to VP8L_NON_TRIVIAL_SYM. +double VP8LBitsEntropy(const uint32_t* const array, int n, + uint32_t* const trivial_symbol) { + double retval = 0.; + uint32_t sum = 0; + uint32_t nonzero_code = VP8L_NON_TRIVIAL_SYM; + int nonzeros = 0; + uint32_t max_val = 0; + int i; + for (i = 0; i < n; ++i) { + if (array[i] != 0) { + sum += array[i]; + nonzero_code = i; + ++nonzeros; + retval -= VP8LFastSLog2(array[i]); + if (max_val < array[i]) { + max_val = array[i]; + } + } + } + retval += VP8LFastSLog2(sum); + if (trivial_symbol != NULL) { + *trivial_symbol = (nonzeros == 1) ? nonzero_code : VP8L_NON_TRIVIAL_SYM; + } + return BitsEntropyRefine(nonzeros, sum, max_val, retval); +} + +static double InitialHuffmanCost(void) { + // Small bias because Huffman code length is typically not stored in + // full length. + static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3; + static const double kSmallBias = 9.1; + return kHuffmanCodeOfHuffmanCodeSize - kSmallBias; +} + +// Finalize the Huffman cost based on streak numbers and length type (<3 or >=3) +static double FinalHuffmanCost(const VP8LStreaks* const stats) { + double retval = InitialHuffmanCost(); + retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1]; + retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1]; + retval += 1.796875 * stats->streaks[0][0]; + retval += 3.28125 * stats->streaks[1][0]; + return retval; +} + +// Trampolines +static double HuffmanCost(const uint32_t* const population, int length) { + const VP8LStreaks stats = VP8LHuffmanCostCount(population, length); + return FinalHuffmanCost(&stats); +} + +// Aggregated costs +double VP8LPopulationCost(const uint32_t* const population, int length, + uint32_t* const trivial_sym) { + return + VP8LBitsEntropy(population, length, trivial_sym) + + HuffmanCost(population, length); +} + +double VP8LGetCombinedEntropy(const uint32_t* const X, + const uint32_t* const Y, int length) { + double bits_entropy_combined; + double huffman_cost_combined; + int i; + + // Bit entropy variables. + double retval = 0.; + int sum = 0; + int nonzeros = 0; + uint32_t max_val = 0; + int i_prev; + uint32_t xy; + + // Huffman cost variables. + int streak = 0; + uint32_t xy_prev; + VP8LStreaks stats; + memset(&stats, 0, sizeof(stats)); + + // Treat the first value for the huffman cost: this is keeping the original + // behavior, even though there is no first streak. + // TODO(vrabaud): study proper behavior + xy = X[0] + Y[0]; + ++stats.streaks[xy != 0][0]; + xy_prev = xy; + i_prev = 0; + + for (i = 1; i < length; ++i) { + xy = X[i] + Y[i]; + + // Process data by streaks for both bit entropy and huffman cost. + if (xy != xy_prev) { + streak = i - i_prev; + + // Gather info for the bit entropy. + if (xy_prev != 0) { + sum += xy_prev * streak; + nonzeros += streak; + retval -= VP8LFastSLog2(xy_prev) * streak; + if (max_val < xy_prev) { + max_val = xy_prev; + } + } + + // Gather info for the huffman cost. + stats.counts[xy != 0] += (streak > 3); + stats.streaks[xy != 0][(streak > 3)] += streak; + + xy_prev = xy; + i_prev = i; + } + } + + // Finish off the last streak for bit entropy. + if (xy != 0) { + streak = i - i_prev; + sum += xy * streak; + nonzeros += streak; + retval -= VP8LFastSLog2(xy) * streak; + if (max_val < xy) { + max_val = xy; + } + } + // Huffman cost is not updated with the last streak to keep original behavior. + // TODO(vrabaud): study proper behavior + + retval += VP8LFastSLog2(sum); + bits_entropy_combined = BitsEntropyRefine(nonzeros, sum, max_val, retval); + + huffman_cost_combined = FinalHuffmanCost(&stats); + + return bits_entropy_combined + huffman_cost_combined; +} + +// Estimates the Entropy + Huffman + other block overhead size cost. +double VP8LHistogramEstimateBits(const VP8LHistogram* const p) { + return + VP8LPopulationCost( + p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_), NULL) + + VP8LPopulationCost(p->red_, NUM_LITERAL_CODES, NULL) + + VP8LPopulationCost(p->blue_, NUM_LITERAL_CODES, NULL) + + VP8LPopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL) + + VP8LPopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL) + + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES) + + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES); +} + +double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p) { + return + VP8LBitsEntropy(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_), + NULL) + + VP8LBitsEntropy(p->red_, NUM_LITERAL_CODES, NULL) + + VP8LBitsEntropy(p->blue_, NUM_LITERAL_CODES, NULL) + + VP8LBitsEntropy(p->alpha_, NUM_LITERAL_CODES, NULL) + + VP8LBitsEntropy(p->distance_, NUM_DISTANCE_CODES, NULL) + + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES) + + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES); +} + +static WEBP_INLINE void UpdateHisto(int histo_argb[4][256], uint32_t argb) { + ++histo_argb[0][argb >> 24]; + ++histo_argb[1][(argb >> 16) & 0xff]; + ++histo_argb[2][(argb >> 8) & 0xff]; + ++histo_argb[3][argb & 0xff]; +} + +//------------------------------------------------------------------------------ + +// Returns best predictor and updates the accumulated histogram. +static int GetBestPredictorForTile(int width, int height, + int tile_x, int tile_y, int bits, + int accumulated[4][256], + const uint32_t* const argb_scratch) { + const int kNumPredModes = 14; + const int col_start = tile_x << bits; + const int row_start = tile_y << bits; + const int tile_size = 1 << bits; + const int max_y = GetMin(tile_size, height - row_start); + const int max_x = GetMin(tile_size, width - col_start); + float best_diff = MAX_DIFF_COST; + int best_mode = 0; + int mode; + int histo_stack_1[4][256]; + int histo_stack_2[4][256]; + // Need pointers to be able to swap arrays. + int (*histo_argb)[256] = histo_stack_1; + int (*best_histo)[256] = histo_stack_2; + + int i, j; + for (mode = 0; mode < kNumPredModes; ++mode) { + const uint32_t* current_row = argb_scratch; + const VP8LPredictorFunc pred_func = VP8LPredictors[mode]; + float cur_diff; + int y; + memset(histo_argb, 0, sizeof(histo_stack_1)); + for (y = 0; y < max_y; ++y) { + int x; + const int row = row_start + y; + const uint32_t* const upper_row = current_row; + current_row = upper_row + width; + for (x = 0; x < max_x; ++x) { + const int col = col_start + x; + uint32_t predict; + if (row == 0) { + predict = (col == 0) ? ARGB_BLACK : current_row[col - 1]; // Left. + } else if (col == 0) { + predict = upper_row[col]; // Top. + } else { + predict = pred_func(current_row[col - 1], upper_row + col); + } + UpdateHisto(histo_argb, VP8LSubPixels(current_row[col], predict)); + } + } + cur_diff = PredictionCostSpatialHistogram( + (const int (*)[256])accumulated, (const int (*)[256])histo_argb); + if (cur_diff < best_diff) { + int (*tmp)[256] = histo_argb; + histo_argb = best_histo; + best_histo = tmp; + best_diff = cur_diff; + best_mode = mode; + } + } + + for (i = 0; i < 4; i++) { + for (j = 0; j < 256; j++) { + accumulated[i][j] += best_histo[i][j]; + } + } + + return best_mode; +} + +static void CopyImageWithPrediction(int width, int height, + int bits, uint32_t* const modes, + uint32_t* const argb_scratch, + uint32_t* const argb) { + const int tiles_per_row = VP8LSubSampleSize(width, bits); + const int mask = (1 << bits) - 1; + // The row size is one pixel longer to allow the top right pixel to point to + // the leftmost pixel of the next row when at the right edge. + uint32_t* current_row = argb_scratch; + uint32_t* upper_row = argb_scratch + width + 1; + int y; + VP8LPredictorFunc pred_func = 0; + + for (y = 0; y < height; ++y) { + int x; + uint32_t* tmp = upper_row; + upper_row = current_row; + current_row = tmp; + memcpy(current_row, argb + y * width, sizeof(*current_row) * width); + current_row[width] = (y + 1 < height) ? argb[(y + 1) * width] : ARGB_BLACK; + for (x = 0; x < width; ++x) { + uint32_t predict; + if ((x & mask) == 0) { + const int mode = + (modes[(y >> bits) * tiles_per_row + (x >> bits)] >> 8) & 0xff; + pred_func = VP8LPredictors[mode]; + } + if (y == 0) { + predict = (x == 0) ? ARGB_BLACK : current_row[x - 1]; // Left. + } else if (x == 0) { + predict = upper_row[x]; // Top. + } else { + predict = pred_func(current_row[x - 1], upper_row + x); + } + argb[y * width + x] = VP8LSubPixels(current_row[x], predict); + } + } +} + +void VP8LResidualImage(int width, int height, int bits, int low_effort, + uint32_t* const argb, uint32_t* const argb_scratch, + uint32_t* const image) { + const int max_tile_size = 1 << bits; + const int tiles_per_row = VP8LSubSampleSize(width, bits); + const int tiles_per_col = VP8LSubSampleSize(height, bits); + const int kPredLowEffort = 11; + uint32_t* const upper_row = argb_scratch; + uint32_t* const current_tile_rows = argb_scratch + width; + int tile_y; + int histo[4][256]; + if (!low_effort) memset(histo, 0, sizeof(histo)); + for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) { + const int tile_y_offset = tile_y * max_tile_size; + const int this_tile_height = + (tile_y < tiles_per_col - 1) ? max_tile_size : height - tile_y_offset; + int tile_x; + if (tile_y > 0) { + memcpy(upper_row, current_tile_rows + (max_tile_size - 1) * width, + width * sizeof(*upper_row)); + } + memcpy(current_tile_rows, &argb[tile_y_offset * width], + this_tile_height * width * sizeof(*current_tile_rows)); + for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) { + const int pred = + low_effort ? kPredLowEffort : + GetBestPredictorForTile(width, height, + tile_x, tile_y, bits, + (int (*)[256])histo, + argb_scratch); + image[tile_y * tiles_per_row + tile_x] = 0xff000000u | (pred << 8); + } + } + + CopyImageWithPrediction(width, height, bits, image, argb_scratch, argb); +} + +void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels) { + int i; + for (i = 0; i < num_pixels; ++i) { + const uint32_t argb = argb_data[i]; + const uint32_t green = (argb >> 8) & 0xff; + const uint32_t new_r = (((argb >> 16) & 0xff) - green) & 0xff; + const uint32_t new_b = ((argb & 0xff) - green) & 0xff; + argb_data[i] = (argb & 0xff00ff00) | (new_r << 16) | new_b; + } +} + +static WEBP_INLINE void MultipliersClear(VP8LMultipliers* const m) { + m->green_to_red_ = 0; + m->green_to_blue_ = 0; + m->red_to_blue_ = 0; +} + +static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred, + int8_t color) { + return (uint32_t)((int)(color_pred) * color) >> 5; +} + +static WEBP_INLINE void ColorCodeToMultipliers(uint32_t color_code, + VP8LMultipliers* const m) { + m->green_to_red_ = (color_code >> 0) & 0xff; + m->green_to_blue_ = (color_code >> 8) & 0xff; + m->red_to_blue_ = (color_code >> 16) & 0xff; +} + +static WEBP_INLINE uint32_t MultipliersToColorCode( + const VP8LMultipliers* const m) { + return 0xff000000u | + ((uint32_t)(m->red_to_blue_) << 16) | + ((uint32_t)(m->green_to_blue_) << 8) | + m->green_to_red_; +} + +void VP8LTransformColor_C(const VP8LMultipliers* const m, uint32_t* data, + int num_pixels) { + int i; + for (i = 0; i < num_pixels; ++i) { + const uint32_t argb = data[i]; + const uint32_t green = argb >> 8; + const uint32_t red = argb >> 16; + uint32_t new_red = red; + uint32_t new_blue = argb; + new_red -= ColorTransformDelta(m->green_to_red_, green); + new_red &= 0xff; + new_blue -= ColorTransformDelta(m->green_to_blue_, green); + new_blue -= ColorTransformDelta(m->red_to_blue_, red); + new_blue &= 0xff; + data[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue); + } +} + +static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red, + uint32_t argb) { + const uint32_t green = argb >> 8; + uint32_t new_red = argb >> 16; + new_red -= ColorTransformDelta(green_to_red, green); + return (new_red & 0xff); +} + +static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue, + uint8_t red_to_blue, + uint32_t argb) { + const uint32_t green = argb >> 8; + const uint32_t red = argb >> 16; + uint8_t new_blue = argb; + new_blue -= ColorTransformDelta(green_to_blue, green); + new_blue -= ColorTransformDelta(red_to_blue, red); + return (new_blue & 0xff); +} + +static float PredictionCostCrossColor(const int accumulated[256], + const int counts[256]) { + // Favor low entropy, locally and globally. + // Favor small absolute values for PredictionCostSpatial + static const double kExpValue = 2.4; + return CombinedShannonEntropy(counts, accumulated) + + PredictionCostSpatial(counts, 3, kExpValue); +} + +void VP8LCollectColorRedTransforms_C(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]) { + while (tile_height-- > 0) { + int x; + for (x = 0; x < tile_width; ++x) { + ++histo[TransformColorRed(green_to_red, argb[x])]; + } + argb += stride; + } +} + +static float GetPredictionCostCrossColorRed( + const uint32_t* argb, int stride, int tile_width, int tile_height, + VP8LMultipliers prev_x, VP8LMultipliers prev_y, int green_to_red, + const int accumulated_red_histo[256]) { + int histo[256] = { 0 }; + float cur_diff; + + VP8LCollectColorRedTransforms(argb, stride, tile_width, tile_height, + green_to_red, histo); + + cur_diff = PredictionCostCrossColor(accumulated_red_histo, histo); + if ((uint8_t)green_to_red == prev_x.green_to_red_) { + cur_diff -= 3; // favor keeping the areas locally similar + } + if ((uint8_t)green_to_red == prev_y.green_to_red_) { + cur_diff -= 3; // favor keeping the areas locally similar + } + if (green_to_red == 0) { + cur_diff -= 3; + } + return cur_diff; +} + +static void GetBestGreenToRed( + const uint32_t* argb, int stride, int tile_width, int tile_height, + VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality, + const int accumulated_red_histo[256], VP8LMultipliers* const best_tx) { + const int kMaxIters = 4 + ((7 * quality) >> 8); // in range [4..6] + int green_to_red_best = 0; + int iter, offset; + float best_diff = GetPredictionCostCrossColorRed( + argb, stride, tile_width, tile_height, prev_x, prev_y, + green_to_red_best, accumulated_red_histo); + for (iter = 0; iter < kMaxIters; ++iter) { + // ColorTransformDelta is a 3.5 bit fixed point, so 32 is equal to + // one in color computation. Having initial delta here as 1 is sufficient + // to explore the range of (-2, 2). + const int delta = 32 >> iter; + // Try a negative and a positive delta from the best known value. + for (offset = -delta; offset <= delta; offset += 2 * delta) { + const int green_to_red_cur = offset + green_to_red_best; + const float cur_diff = GetPredictionCostCrossColorRed( + argb, stride, tile_width, tile_height, prev_x, prev_y, + green_to_red_cur, accumulated_red_histo); + if (cur_diff < best_diff) { + best_diff = cur_diff; + green_to_red_best = green_to_red_cur; + } + } + } + best_tx->green_to_red_ = green_to_red_best; +} + +void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_blue, int red_to_blue, + int histo[]) { + while (tile_height-- > 0) { + int x; + for (x = 0; x < tile_width; ++x) { + ++histo[TransformColorBlue(green_to_blue, red_to_blue, argb[x])]; + } + argb += stride; + } +} + +static float GetPredictionCostCrossColorBlue( + const uint32_t* argb, int stride, int tile_width, int tile_height, + VP8LMultipliers prev_x, VP8LMultipliers prev_y, + int green_to_blue, int red_to_blue, const int accumulated_blue_histo[256]) { + int histo[256] = { 0 }; + float cur_diff; + + VP8LCollectColorBlueTransforms(argb, stride, tile_width, tile_height, + green_to_blue, red_to_blue, histo); + + cur_diff = PredictionCostCrossColor(accumulated_blue_histo, histo); + if ((uint8_t)green_to_blue == prev_x.green_to_blue_) { + cur_diff -= 3; // favor keeping the areas locally similar + } + if ((uint8_t)green_to_blue == prev_y.green_to_blue_) { + cur_diff -= 3; // favor keeping the areas locally similar + } + if ((uint8_t)red_to_blue == prev_x.red_to_blue_) { + cur_diff -= 3; // favor keeping the areas locally similar + } + if ((uint8_t)red_to_blue == prev_y.red_to_blue_) { + cur_diff -= 3; // favor keeping the areas locally similar + } + if (green_to_blue == 0) { + cur_diff -= 3; + } + if (red_to_blue == 0) { + cur_diff -= 3; + } + return cur_diff; +} + +#define kGreenRedToBlueNumAxis 8 +#define kGreenRedToBlueMaxIters 7 +static void GetBestGreenRedToBlue( + const uint32_t* argb, int stride, int tile_width, int tile_height, + VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality, + const int accumulated_blue_histo[256], + VP8LMultipliers* const best_tx) { + const int8_t offset[kGreenRedToBlueNumAxis][2] = + {{0, -1}, {0, 1}, {-1, 0}, {1, 0}, {-1, -1}, {-1, 1}, {1, -1}, {1, 1}}; + const int8_t delta_lut[kGreenRedToBlueMaxIters] = { 16, 16, 8, 4, 2, 2, 2 }; + const int iters = + (quality < 25) ? 1 : (quality > 50) ? kGreenRedToBlueMaxIters : 4; + int green_to_blue_best = 0; + int red_to_blue_best = 0; + int iter; + // Initial value at origin: + float best_diff = GetPredictionCostCrossColorBlue( + argb, stride, tile_width, tile_height, prev_x, prev_y, + green_to_blue_best, red_to_blue_best, accumulated_blue_histo); + for (iter = 0; iter < iters; ++iter) { + const int delta = delta_lut[iter]; + int axis; + for (axis = 0; axis < kGreenRedToBlueNumAxis; ++axis) { + const int green_to_blue_cur = + offset[axis][0] * delta + green_to_blue_best; + const int red_to_blue_cur = offset[axis][1] * delta + red_to_blue_best; + const float cur_diff = GetPredictionCostCrossColorBlue( + argb, stride, tile_width, tile_height, prev_x, prev_y, + green_to_blue_cur, red_to_blue_cur, accumulated_blue_histo); + if (cur_diff < best_diff) { + best_diff = cur_diff; + green_to_blue_best = green_to_blue_cur; + red_to_blue_best = red_to_blue_cur; + } + if (quality < 25 && iter == 4) { + // Only axis aligned diffs for lower quality. + break; // next iter. + } + } + if (delta == 2 && green_to_blue_best == 0 && red_to_blue_best == 0) { + // Further iterations would not help. + break; // out of iter-loop. + } + } + best_tx->green_to_blue_ = green_to_blue_best; + best_tx->red_to_blue_ = red_to_blue_best; +} +#undef kGreenRedToBlueMaxIters +#undef kGreenRedToBlueNumAxis + +static VP8LMultipliers GetBestColorTransformForTile( + int tile_x, int tile_y, int bits, + VP8LMultipliers prev_x, + VP8LMultipliers prev_y, + int quality, int xsize, int ysize, + const int accumulated_red_histo[256], + const int accumulated_blue_histo[256], + const uint32_t* const argb) { + const int max_tile_size = 1 << bits; + const int tile_y_offset = tile_y * max_tile_size; + const int tile_x_offset = tile_x * max_tile_size; + const int all_x_max = GetMin(tile_x_offset + max_tile_size, xsize); + const int all_y_max = GetMin(tile_y_offset + max_tile_size, ysize); + const int tile_width = all_x_max - tile_x_offset; + const int tile_height = all_y_max - tile_y_offset; + const uint32_t* const tile_argb = argb + tile_y_offset * xsize + + tile_x_offset; + VP8LMultipliers best_tx; + MultipliersClear(&best_tx); + + GetBestGreenToRed(tile_argb, xsize, tile_width, tile_height, + prev_x, prev_y, quality, accumulated_red_histo, &best_tx); + GetBestGreenRedToBlue(tile_argb, xsize, tile_width, tile_height, + prev_x, prev_y, quality, accumulated_blue_histo, + &best_tx); + return best_tx; +} + +static void CopyTileWithColorTransform(int xsize, int ysize, + int tile_x, int tile_y, + int max_tile_size, + VP8LMultipliers color_transform, + uint32_t* argb) { + const int xscan = GetMin(max_tile_size, xsize - tile_x); + int yscan = GetMin(max_tile_size, ysize - tile_y); + argb += tile_y * xsize + tile_x; + while (yscan-- > 0) { + VP8LTransformColor(&color_transform, argb, xscan); + argb += xsize; + } +} + +void VP8LColorSpaceTransform(int width, int height, int bits, int quality, + uint32_t* const argb, uint32_t* image) { + const int max_tile_size = 1 << bits; + const int tile_xsize = VP8LSubSampleSize(width, bits); + const int tile_ysize = VP8LSubSampleSize(height, bits); + int accumulated_red_histo[256] = { 0 }; + int accumulated_blue_histo[256] = { 0 }; + int tile_x, tile_y; + VP8LMultipliers prev_x, prev_y; + MultipliersClear(&prev_y); + MultipliersClear(&prev_x); + for (tile_y = 0; tile_y < tile_ysize; ++tile_y) { + for (tile_x = 0; tile_x < tile_xsize; ++tile_x) { + int y; + const int tile_x_offset = tile_x * max_tile_size; + const int tile_y_offset = tile_y * max_tile_size; + const int all_x_max = GetMin(tile_x_offset + max_tile_size, width); + const int all_y_max = GetMin(tile_y_offset + max_tile_size, height); + const int offset = tile_y * tile_xsize + tile_x; + if (tile_y != 0) { + ColorCodeToMultipliers(image[offset - tile_xsize], &prev_y); + } + prev_x = GetBestColorTransformForTile(tile_x, tile_y, bits, + prev_x, prev_y, + quality, width, height, + accumulated_red_histo, + accumulated_blue_histo, + argb); + image[offset] = MultipliersToColorCode(&prev_x); + CopyTileWithColorTransform(width, height, tile_x_offset, tile_y_offset, + max_tile_size, prev_x, argb); + + // Gather accumulated histogram data. + for (y = tile_y_offset; y < all_y_max; ++y) { + int ix = y * width + tile_x_offset; + const int ix_end = ix + all_x_max - tile_x_offset; + for (; ix < ix_end; ++ix) { + const uint32_t pix = argb[ix]; + if (ix >= 2 && + pix == argb[ix - 2] && + pix == argb[ix - 1]) { + continue; // repeated pixels are handled by backward references + } + if (ix >= width + 2 && + argb[ix - 2] == argb[ix - width - 2] && + argb[ix - 1] == argb[ix - width - 1] && + pix == argb[ix - width]) { + continue; // repeated pixels are handled by backward references + } + ++accumulated_red_histo[(pix >> 16) & 0xff]; + ++accumulated_blue_histo[(pix >> 0) & 0xff]; + } + } + } + } +} + +//------------------------------------------------------------------------------ +// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel. +void VP8LBundleColorMap(const uint8_t* const row, int width, + int xbits, uint32_t* const dst) { + int x; + if (xbits > 0) { + const int bit_depth = 1 << (3 - xbits); + const int mask = (1 << xbits) - 1; + uint32_t code = 0xff000000; + for (x = 0; x < width; ++x) { + const int xsub = x & mask; + if (xsub == 0) { + code = 0xff000000; + } + code |= row[x] << (8 + bit_depth * xsub); + dst[x >> xbits] = code; + } + } else { + for (x = 0; x < width; ++x) dst[x] = 0xff000000 | (row[x] << 8); + } +} + +//------------------------------------------------------------------------------ + +static double ExtraCost(const uint32_t* population, int length) { + int i; + double cost = 0.; + for (i = 2; i < length - 2; ++i) cost += (i >> 1) * population[i + 2]; + return cost; +} + +static double ExtraCostCombined(const uint32_t* X, const uint32_t* Y, + int length) { + int i; + double cost = 0.; + for (i = 2; i < length - 2; ++i) { + const int xy = X[i + 2] + Y[i + 2]; + cost += (i >> 1) * xy; + } + return cost; +} + +// Returns the various RLE counts +static VP8LStreaks HuffmanCostCount(const uint32_t* population, int length) { + int i; + int streak = 0; + VP8LStreaks stats; + memset(&stats, 0, sizeof(stats)); + for (i = 0; i < length - 1; ++i) { + ++streak; + if (population[i] == population[i + 1]) { + continue; + } + stats.counts[population[i] != 0] += (streak > 3); + stats.streaks[population[i] != 0][(streak > 3)] += streak; + streak = 0; + } + ++streak; + stats.counts[population[i] != 0] += (streak > 3); + stats.streaks[population[i] != 0][(streak > 3)] += streak; + return stats; +} + +//------------------------------------------------------------------------------ + +static void HistogramAdd(const VP8LHistogram* const a, + const VP8LHistogram* const b, + VP8LHistogram* const out) { + int i; + const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_); + assert(a->palette_code_bits_ == b->palette_code_bits_); + if (b != out) { + for (i = 0; i < literal_size; ++i) { + out->literal_[i] = a->literal_[i] + b->literal_[i]; + } + for (i = 0; i < NUM_DISTANCE_CODES; ++i) { + out->distance_[i] = a->distance_[i] + b->distance_[i]; + } + for (i = 0; i < NUM_LITERAL_CODES; ++i) { + out->red_[i] = a->red_[i] + b->red_[i]; + out->blue_[i] = a->blue_[i] + b->blue_[i]; + out->alpha_[i] = a->alpha_[i] + b->alpha_[i]; + } + } else { + for (i = 0; i < literal_size; ++i) { + out->literal_[i] += a->literal_[i]; + } + for (i = 0; i < NUM_DISTANCE_CODES; ++i) { + out->distance_[i] += a->distance_[i]; + } + for (i = 0; i < NUM_LITERAL_CODES; ++i) { + out->red_[i] += a->red_[i]; + out->blue_[i] += a->blue_[i]; + out->alpha_[i] += a->alpha_[i]; + } + } +} + +//------------------------------------------------------------------------------ + +VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed; + +VP8LTransformColorFunc VP8LTransformColor; + +VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms; +VP8LCollectColorRedTransformsFunc VP8LCollectColorRedTransforms; + +VP8LFastLog2SlowFunc VP8LFastLog2Slow; +VP8LFastLog2SlowFunc VP8LFastSLog2Slow; + +VP8LCostFunc VP8LExtraCost; +VP8LCostCombinedFunc VP8LExtraCostCombined; + +VP8LCostCountFunc VP8LHuffmanCostCount; + +VP8LHistogramAddFunc VP8LHistogramAdd; + +extern void VP8LEncDspInitSSE2(void); +extern void VP8LEncDspInitSSE41(void); +extern void VP8LEncDspInitNEON(void); +extern void VP8LEncDspInitMIPS32(void); +extern void VP8LEncDspInitMIPSdspR2(void); + +static volatile VP8CPUInfo lossless_enc_last_cpuinfo_used = + (VP8CPUInfo)&lossless_enc_last_cpuinfo_used; + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInit(void) { + if (lossless_enc_last_cpuinfo_used == VP8GetCPUInfo) return; + + VP8LDspInit(); + + VP8LSubtractGreenFromBlueAndRed = VP8LSubtractGreenFromBlueAndRed_C; + + VP8LTransformColor = VP8LTransformColor_C; + + VP8LCollectColorBlueTransforms = VP8LCollectColorBlueTransforms_C; + VP8LCollectColorRedTransforms = VP8LCollectColorRedTransforms_C; + + VP8LFastLog2Slow = FastLog2Slow; + VP8LFastSLog2Slow = FastSLog2Slow; + + VP8LExtraCost = ExtraCost; + VP8LExtraCostCombined = ExtraCostCombined; + + VP8LHuffmanCostCount = HuffmanCostCount; + + VP8LHistogramAdd = HistogramAdd; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8LEncDspInitSSE2(); +#if defined(WEBP_USE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + VP8LEncDspInitSSE41(); + } +#endif + } +#endif +#if defined(WEBP_USE_NEON) + if (VP8GetCPUInfo(kNEON)) { + VP8LEncDspInitNEON(); + } +#endif +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + VP8LEncDspInitMIPS32(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8LEncDspInitMIPSdspR2(); + } +#endif + } + lossless_enc_last_cpuinfo_used = VP8GetCPUInfo; +} + +//------------------------------------------------------------------------------ diff --git a/drivers/webp/dsp/lossless_enc_mips32.c b/drivers/webp/dsp/lossless_enc_mips32.c new file mode 100644 index 0000000000..0468a5aac2 --- /dev/null +++ b/drivers/webp/dsp/lossless_enc_mips32.c @@ -0,0 +1,417 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of lossless functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "./dsp.h" +#include "./lossless.h" + +#if defined(WEBP_USE_MIPS32) + +#include <assert.h> +#include <math.h> +#include <stdlib.h> +#include <string.h> + +#define APPROX_LOG_WITH_CORRECTION_MAX 65536 +#define APPROX_LOG_MAX 4096 +#define LOG_2_RECIPROCAL 1.44269504088896338700465094007086 + +static float FastSLog2Slow(uint32_t v) { + assert(v >= LOG_LOOKUP_IDX_MAX); + if (v < APPROX_LOG_WITH_CORRECTION_MAX) { + uint32_t log_cnt, y, correction; + const int c24 = 24; + const float v_f = (float)v; + uint32_t temp; + + // Xf = 256 = 2^8 + // log_cnt is index of leading one in upper 24 bits + __asm__ volatile( + "clz %[log_cnt], %[v] \n\t" + "addiu %[y], $zero, 1 \n\t" + "subu %[log_cnt], %[c24], %[log_cnt] \n\t" + "sllv %[y], %[y], %[log_cnt] \n\t" + "srlv %[temp], %[v], %[log_cnt] \n\t" + : [log_cnt]"=&r"(log_cnt), [y]"=&r"(y), + [temp]"=r"(temp) + : [c24]"r"(c24), [v]"r"(v) + ); + + // vf = (2^log_cnt) * Xf; where y = 2^log_cnt and Xf < 256 + // Xf = floor(Xf) * (1 + (v % y) / v) + // log2(Xf) = log2(floor(Xf)) + log2(1 + (v % y) / v) + // The correction factor: log(1 + d) ~ d; for very small d values, so + // log2(1 + (v % y) / v) ~ LOG_2_RECIPROCAL * (v % y)/v + // LOG_2_RECIPROCAL ~ 23/16 + + // (v % y) = (v % 2^log_cnt) = v & (2^log_cnt - 1) + correction = (23 * (v & (y - 1))) >> 4; + return v_f * (kLog2Table[temp] + log_cnt) + correction; + } else { + return (float)(LOG_2_RECIPROCAL * v * log((double)v)); + } +} + +static float FastLog2Slow(uint32_t v) { + assert(v >= LOG_LOOKUP_IDX_MAX); + if (v < APPROX_LOG_WITH_CORRECTION_MAX) { + uint32_t log_cnt, y; + const int c24 = 24; + double log_2; + uint32_t temp; + + __asm__ volatile( + "clz %[log_cnt], %[v] \n\t" + "addiu %[y], $zero, 1 \n\t" + "subu %[log_cnt], %[c24], %[log_cnt] \n\t" + "sllv %[y], %[y], %[log_cnt] \n\t" + "srlv %[temp], %[v], %[log_cnt] \n\t" + : [log_cnt]"=&r"(log_cnt), [y]"=&r"(y), + [temp]"=r"(temp) + : [c24]"r"(c24), [v]"r"(v) + ); + + log_2 = kLog2Table[temp] + log_cnt; + if (v >= APPROX_LOG_MAX) { + // Since the division is still expensive, add this correction factor only + // for large values of 'v'. + + const uint32_t correction = (23 * (v & (y - 1))) >> 4; + log_2 += (double)correction / v; + } + return (float)log_2; + } else { + return (float)(LOG_2_RECIPROCAL * log((double)v)); + } +} + +// C version of this function: +// int i = 0; +// int64_t cost = 0; +// const uint32_t* pop = &population[4]; +// const uint32_t* LoopEnd = &population[length]; +// while (pop != LoopEnd) { +// ++i; +// cost += i * *pop; +// cost += i * *(pop + 1); +// pop += 2; +// } +// return (double)cost; +static double ExtraCost(const uint32_t* const population, int length) { + int i, temp0, temp1; + const uint32_t* pop = &population[4]; + const uint32_t* const LoopEnd = &population[length]; + + __asm__ volatile( + "mult $zero, $zero \n\t" + "xor %[i], %[i], %[i] \n\t" + "beq %[pop], %[LoopEnd], 2f \n\t" + "1: \n\t" + "lw %[temp0], 0(%[pop]) \n\t" + "lw %[temp1], 4(%[pop]) \n\t" + "addiu %[i], %[i], 1 \n\t" + "addiu %[pop], %[pop], 8 \n\t" + "madd %[i], %[temp0] \n\t" + "madd %[i], %[temp1] \n\t" + "bne %[pop], %[LoopEnd], 1b \n\t" + "2: \n\t" + "mfhi %[temp0] \n\t" + "mflo %[temp1] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [i]"=&r"(i), [pop]"+r"(pop) + : [LoopEnd]"r"(LoopEnd) + : "memory", "hi", "lo" + ); + + return (double)((int64_t)temp0 << 32 | temp1); +} + +// C version of this function: +// int i = 0; +// int64_t cost = 0; +// const uint32_t* pX = &X[4]; +// const uint32_t* pY = &Y[4]; +// const uint32_t* LoopEnd = &X[length]; +// while (pX != LoopEnd) { +// const uint32_t xy0 = *pX + *pY; +// const uint32_t xy1 = *(pX + 1) + *(pY + 1); +// ++i; +// cost += i * xy0; +// cost += i * xy1; +// pX += 2; +// pY += 2; +// } +// return (double)cost; +static double ExtraCostCombined(const uint32_t* const X, + const uint32_t* const Y, int length) { + int i, temp0, temp1, temp2, temp3; + const uint32_t* pX = &X[4]; + const uint32_t* pY = &Y[4]; + const uint32_t* const LoopEnd = &X[length]; + + __asm__ volatile( + "mult $zero, $zero \n\t" + "xor %[i], %[i], %[i] \n\t" + "beq %[pX], %[LoopEnd], 2f \n\t" + "1: \n\t" + "lw %[temp0], 0(%[pX]) \n\t" + "lw %[temp1], 0(%[pY]) \n\t" + "lw %[temp2], 4(%[pX]) \n\t" + "lw %[temp3], 4(%[pY]) \n\t" + "addiu %[i], %[i], 1 \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addiu %[pX], %[pX], 8 \n\t" + "addiu %[pY], %[pY], 8 \n\t" + "madd %[i], %[temp0] \n\t" + "madd %[i], %[temp2] \n\t" + "bne %[pX], %[LoopEnd], 1b \n\t" + "2: \n\t" + "mfhi %[temp0] \n\t" + "mflo %[temp1] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [i]"=&r"(i), [pX]"+r"(pX), [pY]"+r"(pY) + : [LoopEnd]"r"(LoopEnd) + : "memory", "hi", "lo" + ); + + return (double)((int64_t)temp0 << 32 | temp1); +} + +#define HUFFMAN_COST_PASS \ + __asm__ volatile( \ + "sll %[temp1], %[temp0], 3 \n\t" \ + "addiu %[temp3], %[streak], -3 \n\t" \ + "addu %[temp2], %[pstreaks], %[temp1] \n\t" \ + "blez %[temp3], 1f \n\t" \ + "srl %[temp1], %[temp1], 1 \n\t" \ + "addu %[temp3], %[pcnts], %[temp1] \n\t" \ + "lw %[temp0], 4(%[temp2]) \n\t" \ + "lw %[temp1], 0(%[temp3]) \n\t" \ + "addu %[temp0], %[temp0], %[streak] \n\t" \ + "addiu %[temp1], %[temp1], 1 \n\t" \ + "sw %[temp0], 4(%[temp2]) \n\t" \ + "sw %[temp1], 0(%[temp3]) \n\t" \ + "b 2f \n\t" \ + "1: \n\t" \ + "lw %[temp0], 0(%[temp2]) \n\t" \ + "addu %[temp0], %[temp0], %[streak] \n\t" \ + "sw %[temp0], 0(%[temp2]) \n\t" \ + "2: \n\t" \ + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), \ + [temp3]"=&r"(temp3), [temp0]"+r"(temp0) \ + : [pstreaks]"r"(pstreaks), [pcnts]"r"(pcnts), \ + [streak]"r"(streak) \ + : "memory" \ + ); + +// Returns the various RLE counts +static VP8LStreaks HuffmanCostCount(const uint32_t* population, int length) { + int i; + int streak = 0; + VP8LStreaks stats; + int* const pstreaks = &stats.streaks[0][0]; + int* const pcnts = &stats.counts[0]; + int temp0, temp1, temp2, temp3; + memset(&stats, 0, sizeof(stats)); + for (i = 0; i < length - 1; ++i) { + ++streak; + if (population[i] == population[i + 1]) { + continue; + } + temp0 = (population[i] != 0); + HUFFMAN_COST_PASS + streak = 0; + } + ++streak; + temp0 = (population[i] != 0); + HUFFMAN_COST_PASS + + return stats; +} + +static VP8LStreaks HuffmanCostCombinedCount(const uint32_t* X, + const uint32_t* Y, int length) { + int i; + int streak = 0; + uint32_t xy_prev = 0xffffffff; + VP8LStreaks stats; + int* const pstreaks = &stats.streaks[0][0]; + int* const pcnts = &stats.counts[0]; + int temp0, temp1, temp2, temp3; + memset(&stats, 0, sizeof(stats)); + for (i = 0; i < length; ++i) { + const uint32_t xy = X[i] + Y[i]; + ++streak; + if (xy != xy_prev) { + temp0 = (xy != 0); + HUFFMAN_COST_PASS + streak = 0; + xy_prev = xy; + } + } + return stats; +} + +#define ASM_START \ + __asm__ volatile( \ + ".set push \n\t" \ + ".set at \n\t" \ + ".set macro \n\t" \ + "1: \n\t" + +// P2 = P0 + P1 +// A..D - offsets +// E - temp variable to tell macro +// if pointer should be incremented +// literal_ and successive histograms could be unaligned +// so we must use ulw and usw +#define ADD_TO_OUT(A, B, C, D, E, P0, P1, P2) \ + "ulw %[temp0], " #A "(%[" #P0 "]) \n\t" \ + "ulw %[temp1], " #B "(%[" #P0 "]) \n\t" \ + "ulw %[temp2], " #C "(%[" #P0 "]) \n\t" \ + "ulw %[temp3], " #D "(%[" #P0 "]) \n\t" \ + "ulw %[temp4], " #A "(%[" #P1 "]) \n\t" \ + "ulw %[temp5], " #B "(%[" #P1 "]) \n\t" \ + "ulw %[temp6], " #C "(%[" #P1 "]) \n\t" \ + "ulw %[temp7], " #D "(%[" #P1 "]) \n\t" \ + "addu %[temp4], %[temp4], %[temp0] \n\t" \ + "addu %[temp5], %[temp5], %[temp1] \n\t" \ + "addu %[temp6], %[temp6], %[temp2] \n\t" \ + "addu %[temp7], %[temp7], %[temp3] \n\t" \ + "addiu %[" #P0 "], %[" #P0 "], 16 \n\t" \ + ".if " #E " == 1 \n\t" \ + "addiu %[" #P1 "], %[" #P1 "], 16 \n\t" \ + ".endif \n\t" \ + "usw %[temp4], " #A "(%[" #P2 "]) \n\t" \ + "usw %[temp5], " #B "(%[" #P2 "]) \n\t" \ + "usw %[temp6], " #C "(%[" #P2 "]) \n\t" \ + "usw %[temp7], " #D "(%[" #P2 "]) \n\t" \ + "addiu %[" #P2 "], %[" #P2 "], 16 \n\t" \ + "bne %[" #P0 "], %[LoopEnd], 1b \n\t" \ + ".set pop \n\t" \ + +#define ASM_END_COMMON_0 \ + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), \ + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), \ + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), \ + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), \ + [pa]"+r"(pa), [pout]"+r"(pout) + +#define ASM_END_COMMON_1 \ + : [LoopEnd]"r"(LoopEnd) \ + : "memory", "at" \ + ); + +#define ASM_END_0 \ + ASM_END_COMMON_0 \ + , [pb]"+r"(pb) \ + ASM_END_COMMON_1 + +#define ASM_END_1 \ + ASM_END_COMMON_0 \ + ASM_END_COMMON_1 + +#define ADD_VECTOR(A, B, OUT, SIZE, EXTRA_SIZE) do { \ + const uint32_t* pa = (const uint32_t*)(A); \ + const uint32_t* pb = (const uint32_t*)(B); \ + uint32_t* pout = (uint32_t*)(OUT); \ + const uint32_t* const LoopEnd = pa + (SIZE); \ + assert((SIZE) % 4 == 0); \ + ASM_START \ + ADD_TO_OUT(0, 4, 8, 12, 1, pa, pb, pout) \ + ASM_END_0 \ + if ((EXTRA_SIZE) > 0) { \ + const int last = (EXTRA_SIZE); \ + int i; \ + for (i = 0; i < last; ++i) pout[i] = pa[i] + pb[i]; \ + } \ +} while (0) + +#define ADD_VECTOR_EQ(A, OUT, SIZE, EXTRA_SIZE) do { \ + const uint32_t* pa = (const uint32_t*)(A); \ + uint32_t* pout = (uint32_t*)(OUT); \ + const uint32_t* const LoopEnd = pa + (SIZE); \ + assert((SIZE) % 4 == 0); \ + ASM_START \ + ADD_TO_OUT(0, 4, 8, 12, 0, pa, pout, pout) \ + ASM_END_1 \ + if ((EXTRA_SIZE) > 0) { \ + const int last = (EXTRA_SIZE); \ + int i; \ + for (i = 0; i < last; ++i) pout[i] += pa[i]; \ + } \ +} while (0) + +static void HistogramAdd(const VP8LHistogram* const a, + const VP8LHistogram* const b, + VP8LHistogram* const out) { + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + const int extra_cache_size = VP8LHistogramNumCodes(a->palette_code_bits_) + - (NUM_LITERAL_CODES + NUM_LENGTH_CODES); + assert(a->palette_code_bits_ == b->palette_code_bits_); + + if (b != out) { + ADD_VECTOR(a->literal_, b->literal_, out->literal_, + NUM_LITERAL_CODES + NUM_LENGTH_CODES, extra_cache_size); + ADD_VECTOR(a->distance_, b->distance_, out->distance_, + NUM_DISTANCE_CODES, 0); + ADD_VECTOR(a->red_, b->red_, out->red_, NUM_LITERAL_CODES, 0); + ADD_VECTOR(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES, 0); + ADD_VECTOR(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES, 0); + } else { + ADD_VECTOR_EQ(a->literal_, out->literal_, + NUM_LITERAL_CODES + NUM_LENGTH_CODES, extra_cache_size); + ADD_VECTOR_EQ(a->distance_, out->distance_, NUM_DISTANCE_CODES, 0); + ADD_VECTOR_EQ(a->red_, out->red_, NUM_LITERAL_CODES, 0); + ADD_VECTOR_EQ(a->blue_, out->blue_, NUM_LITERAL_CODES, 0); + ADD_VECTOR_EQ(a->alpha_, out->alpha_, NUM_LITERAL_CODES, 0); + } +} + +#undef ADD_VECTOR_EQ +#undef ADD_VECTOR +#undef ASM_END_1 +#undef ASM_END_0 +#undef ASM_END_COMMON_1 +#undef ASM_END_COMMON_0 +#undef ADD_TO_OUT +#undef ASM_START + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPS32(void) { + VP8LFastSLog2Slow = FastSLog2Slow; + VP8LFastLog2Slow = FastLog2Slow; + VP8LExtraCost = ExtraCost; + VP8LExtraCostCombined = ExtraCostCombined; + VP8LHuffmanCostCount = HuffmanCostCount; +// TODO(mips team): rewrite VP8LGetCombinedEntropy (which used to use +// HuffmanCostCombinedCount) with MIPS optimizations +#if 0 + VP8LHuffmanCostCombinedCount = HuffmanCostCombinedCount; +#else + (void)HuffmanCostCombinedCount; +#endif + VP8LHistogramAdd = HistogramAdd; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(VP8LEncDspInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/drivers/webp/dsp/lossless_enc_mips_dsp_r2.c b/drivers/webp/dsp/lossless_enc_mips_dsp_r2.c new file mode 100644 index 0000000000..0abf3c4f36 --- /dev/null +++ b/drivers/webp/dsp/lossless_enc_mips_dsp_r2.c @@ -0,0 +1,275 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Image transform methods for lossless encoder. +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "./lossless.h" + +static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, + int num_pixels) { + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + uint32_t* const p_loop1_end = argb_data + (num_pixels & ~3); + uint32_t* const p_loop2_end = p_loop1_end + (num_pixels & 3); + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[argb_data], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[argb_data]) \n\t" + "lw %[temp1], 4(%[argb_data]) \n\t" + "lw %[temp2], 8(%[argb_data]) \n\t" + "lw %[temp3], 12(%[argb_data]) \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "ext %[temp5], %[temp1], 8, 8 \n\t" + "ext %[temp6], %[temp2], 8, 8 \n\t" + "ext %[temp7], %[temp3], 8, 8 \n\t" + "addiu %[argb_data], %[argb_data], 16 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "replv.ph %[temp5], %[temp5] \n\t" + "replv.ph %[temp6], %[temp6] \n\t" + "replv.ph %[temp7], %[temp7] \n\t" + "subu.qb %[temp0], %[temp0], %[temp4] \n\t" + "subu.qb %[temp1], %[temp1], %[temp5] \n\t" + "subu.qb %[temp2], %[temp2], %[temp6] \n\t" + "subu.qb %[temp3], %[temp3], %[temp7] \n\t" + "sw %[temp0], -16(%[argb_data]) \n\t" + "sw %[temp1], -12(%[argb_data]) \n\t" + "sw %[temp2], -8(%[argb_data]) \n\t" + "bne %[argb_data], %[p_loop1_end], 0b \n\t" + " sw %[temp3], -4(%[argb_data]) \n\t" + "3: \n\t" + "beq %[argb_data], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[argb_data]) \n\t" + "addiu %[argb_data], %[argb_data], 4 \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "subu.qb %[temp0], %[temp0], %[temp4] \n\t" + "bne %[argb_data], %[p_loop2_end], 1b \n\t" + " sw %[temp0], -4(%[argb_data]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [argb_data]"+&r"(argb_data), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6), + [temp7]"=&r"(temp7) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred, + int8_t color) { + return (uint32_t)((int)(color_pred) * color) >> 5; +} + +static void TransformColor(const VP8LMultipliers* const m, uint32_t* data, + int num_pixels) { + int temp0, temp1, temp2, temp3, temp4, temp5; + uint32_t argb, argb1, new_red, new_red1; + const uint32_t G_to_R = m->green_to_red_; + const uint32_t G_to_B = m->green_to_blue_; + const uint32_t R_to_B = m->red_to_blue_; + uint32_t* const p_loop_end = data + (num_pixels & ~1); + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[data], %[p_loop_end], 1f \n\t" + " nop \n\t" + "replv.ph %[temp0], %[G_to_R] \n\t" + "replv.ph %[temp1], %[G_to_B] \n\t" + "replv.ph %[temp2], %[R_to_B] \n\t" + "shll.ph %[temp0], %[temp0], 8 \n\t" + "shll.ph %[temp1], %[temp1], 8 \n\t" + "shll.ph %[temp2], %[temp2], 8 \n\t" + "shra.ph %[temp0], %[temp0], 8 \n\t" + "shra.ph %[temp1], %[temp1], 8 \n\t" + "shra.ph %[temp2], %[temp2], 8 \n\t" + "0: \n\t" + "lw %[argb], 0(%[data]) \n\t" + "lw %[argb1], 4(%[data]) \n\t" + "lhu %[new_red], 2(%[data]) \n\t" + "lhu %[new_red1], 6(%[data]) \n\t" + "precrq.qb.ph %[temp3], %[argb], %[argb1] \n\t" + "precr.qb.ph %[temp4], %[argb], %[argb1] \n\t" + "preceu.ph.qbra %[temp3], %[temp3] \n\t" + "preceu.ph.qbla %[temp4], %[temp4] \n\t" + "shll.ph %[temp3], %[temp3], 8 \n\t" + "shll.ph %[temp4], %[temp4], 8 \n\t" + "shra.ph %[temp3], %[temp3], 8 \n\t" + "shra.ph %[temp4], %[temp4], 8 \n\t" + "mul.ph %[temp5], %[temp3], %[temp0] \n\t" + "mul.ph %[temp3], %[temp3], %[temp1] \n\t" + "mul.ph %[temp4], %[temp4], %[temp2] \n\t" + "addiu %[data], %[data], 8 \n\t" + "ins %[new_red1], %[new_red], 16, 16 \n\t" + "ins %[argb1], %[argb], 16, 16 \n\t" + "shra.ph %[temp5], %[temp5], 5 \n\t" + "shra.ph %[temp3], %[temp3], 5 \n\t" + "shra.ph %[temp4], %[temp4], 5 \n\t" + "subu.ph %[new_red1], %[new_red1], %[temp5] \n\t" + "subu.ph %[argb1], %[argb1], %[temp3] \n\t" + "preceu.ph.qbra %[temp5], %[new_red1] \n\t" + "subu.ph %[argb1], %[argb1], %[temp4] \n\t" + "preceu.ph.qbra %[temp3], %[argb1] \n\t" + "sb %[temp5], -2(%[data]) \n\t" + "sb %[temp3], -4(%[data]) \n\t" + "sra %[temp5], %[temp5], 16 \n\t" + "sra %[temp3], %[temp3], 16 \n\t" + "sb %[temp5], -6(%[data]) \n\t" + "bne %[data], %[p_loop_end], 0b \n\t" + " sb %[temp3], -8(%[data]) \n\t" + "1: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [new_red1]"=&r"(new_red1), [new_red]"=&r"(new_red), + [argb]"=&r"(argb), [argb1]"=&r"(argb1), [data]"+&r"(data) + : [G_to_R]"r"(G_to_R), [R_to_B]"r"(R_to_B), + [G_to_B]"r"(G_to_B), [p_loop_end]"r"(p_loop_end) + : "memory", "hi", "lo" + ); + + if (num_pixels & 1) { + const uint32_t argb_ = data[0]; + const uint32_t green = argb_ >> 8; + const uint32_t red = argb_ >> 16; + uint32_t new_blue = argb_; + new_red = red; + new_red -= ColorTransformDelta(m->green_to_red_, green); + new_red &= 0xff; + new_blue -= ColorTransformDelta(m->green_to_blue_, green); + new_blue -= ColorTransformDelta(m->red_to_blue_, red); + new_blue &= 0xff; + data[0] = (argb_ & 0xff00ff00u) | (new_red << 16) | (new_blue); + } +} + +static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue, + uint8_t red_to_blue, + uint32_t argb) { + const uint32_t green = argb >> 8; + const uint32_t red = argb >> 16; + uint8_t new_blue = argb; + new_blue -= ColorTransformDelta(green_to_blue, green); + new_blue -= ColorTransformDelta(red_to_blue, red); + return (new_blue & 0xff); +} + +static void CollectColorBlueTransforms(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_blue, int red_to_blue, + int histo[]) { + const int rtb = (red_to_blue << 16) | (red_to_blue & 0xffff); + const int gtb = (green_to_blue << 16) | (green_to_blue & 0xffff); + const uint32_t mask = 0xff00ffu; + while (tile_height-- > 0) { + int x; + const uint32_t* p_argb = argb; + argb += stride; + for (x = 0; x < (tile_width >> 1); ++x) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + __asm__ volatile ( + "lw %[temp0], 0(%[p_argb]) \n\t" + "lw %[temp1], 4(%[p_argb]) \n\t" + "precr.qb.ph %[temp2], %[temp0], %[temp1] \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "shra.ph %[temp2], %[temp2], 8 \n\t" + "shra.ph %[temp3], %[temp1], 8 \n\t" + "mul.ph %[temp5], %[temp2], %[rtb] \n\t" + "mul.ph %[temp6], %[temp3], %[gtb] \n\t" + "and %[temp4], %[temp1], %[mask] \n\t" + "addiu %[p_argb], %[p_argb], 8 \n\t" + "shra.ph %[temp5], %[temp5], 5 \n\t" + "shra.ph %[temp6], %[temp6], 5 \n\t" + "subu.qb %[temp2], %[temp4], %[temp5] \n\t" + "subu.qb %[temp2], %[temp2], %[temp6] \n\t" + : [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), + [temp5]"=&r"(temp5), [temp6]"=&r"(temp6) + : [rtb]"r"(rtb), [gtb]"r"(gtb), [mask]"r"(mask) + : "memory", "hi", "lo" + ); + ++histo[(uint8_t)(temp2 >> 16)]; + ++histo[(uint8_t)temp2]; + } + if (tile_width & 1) { + ++histo[TransformColorBlue(green_to_blue, red_to_blue, *p_argb)]; + } + } +} + +static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red, + uint32_t argb) { + const uint32_t green = argb >> 8; + uint32_t new_red = argb >> 16; + new_red -= ColorTransformDelta(green_to_red, green); + return (new_red & 0xff); +} + +static void CollectColorRedTransforms(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]) { + const int gtr = (green_to_red << 16) | (green_to_red & 0xffff); + while (tile_height-- > 0) { + int x; + const uint32_t* p_argb = argb; + argb += stride; + for (x = 0; x < (tile_width >> 1); ++x) { + int temp0, temp1, temp2, temp3, temp4; + __asm__ volatile ( + "lw %[temp0], 0(%[p_argb]) \n\t" + "lw %[temp1], 4(%[p_argb]) \n\t" + "precrq.ph.w %[temp4], %[temp0], %[temp1] \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "shra.ph %[temp3], %[temp1], 8 \n\t" + "mul.ph %[temp2], %[temp3], %[gtr] \n\t" + "addiu %[p_argb], %[p_argb], 8 \n\t" + "shra.ph %[temp2], %[temp2], 5 \n\t" + "subu.qb %[temp2], %[temp4], %[temp2] \n\t" + : [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4) + : [gtr]"r"(gtr) + : "memory", "hi", "lo" + ); + ++histo[(uint8_t)(temp2 >> 16)]; + ++histo[(uint8_t)temp2]; + } + if (tile_width & 1) { + ++histo[TransformColorRed(green_to_red, *p_argb)]; + } + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPSdspR2(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed; + VP8LTransformColor = TransformColor; + VP8LCollectColorBlueTransforms = CollectColorBlueTransforms; + VP8LCollectColorRedTransforms = CollectColorRedTransforms; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8LEncDspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/lossless_enc_neon.c b/drivers/webp/dsp/lossless_enc_neon.c new file mode 100644 index 0000000000..4c56f2594b --- /dev/null +++ b/drivers/webp/dsp/lossless_enc_neon.c @@ -0,0 +1,143 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// NEON variant of methods for lossless encoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_NEON) + +#include <arm_neon.h> + +#include "./lossless.h" +#include "./neon.h" + +//------------------------------------------------------------------------------ +// Subtract-Green Transform + +// vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use +// non-standard versions there. +#if defined(__APPLE__) && defined(__aarch64__) && \ + defined(__apple_build_version__) && (__apple_build_version__< 6020037) +#define USE_VTBLQ +#endif + +#ifdef USE_VTBLQ +// 255 = byte will be zeroed +static const uint8_t kGreenShuffle[16] = { + 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255 +}; + +static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb, + const uint8x16_t shuffle) { + return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)), + vtbl1q_u8(argb, vget_high_u8(shuffle))); +} +#else // !USE_VTBLQ +// 255 = byte will be zeroed +static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255 }; + +static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb, + const uint8x8_t shuffle) { + return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle), + vtbl1_u8(vget_high_u8(argb), shuffle)); +} +#endif // USE_VTBLQ + +static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) { + const uint32_t* const end = argb_data + (num_pixels & ~3); +#ifdef USE_VTBLQ + const uint8x16_t shuffle = vld1q_u8(kGreenShuffle); +#else + const uint8x8_t shuffle = vld1_u8(kGreenShuffle); +#endif + for (; argb_data < end; argb_data += 4) { + const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data); + const uint8x16_t greens = DoGreenShuffle(argb, shuffle); + vst1q_u8((uint8_t*)argb_data, vsubq_u8(argb, greens)); + } + // fallthrough and finish off with plain-C + VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3); +} + +//------------------------------------------------------------------------------ +// Color Transform + +static void TransformColor(const VP8LMultipliers* const m, + uint32_t* argb_data, int num_pixels) { + // sign-extended multiplying constants, pre-shifted by 6. +#define CST(X) (((int16_t)(m->X << 8)) >> 6) + const int16_t rb[8] = { + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_) + }; + const int16x8_t mults_rb = vld1q_s16(rb); + const int16_t b2[8] = { + 0, CST(red_to_blue_), 0, CST(red_to_blue_), + 0, CST(red_to_blue_), 0, CST(red_to_blue_), + }; + const int16x8_t mults_b2 = vld1q_s16(b2); +#undef CST +#ifdef USE_VTBLQ + static const uint8_t kg0g0[16] = { + 255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13 + }; + const uint8x16_t shuffle = vld1q_u8(kg0g0); +#else + static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 }; + const uint8x8_t shuffle = vld1_u8(k0g0g); +#endif + const uint32x4_t mask_rb = vdupq_n_u32(0x00ff00ffu); // red-blue masks + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const uint8x16_t in = vld1q_u8((uint8_t*)(argb_data + i)); + // 0 g 0 g + const uint8x16_t greens = DoGreenShuffle(in, shuffle); + // x dr x db1 + const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb); + // r 0 b 0 + const int16x8_t B = vshlq_n_s16(vreinterpretq_s16_u8(in), 8); + // x db2 0 0 + const int16x8_t C = vqdmulhq_s16(B, mults_b2); + // 0 0 x db2 + const uint32x4_t D = vshrq_n_u32(vreinterpretq_u32_s16(C), 16); + // x dr x db + const int8x16_t E = vaddq_s8(vreinterpretq_s8_u32(D), + vreinterpretq_s8_s16(A)); + // 0 dr 0 db + const uint32x4_t F = vandq_u32(vreinterpretq_u32_s8(E), mask_rb); + const int8x16_t out = vsubq_s8(vreinterpretq_s8_u8(in), + vreinterpretq_s8_u32(F)); + vst1q_s8((int8_t*)(argb_data + i), out); + } + // fallthrough and finish off with plain-C + VP8LTransformColor_C(m, argb_data + i, num_pixels - i); +} + +#undef USE_VTBLQ + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitNEON(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed; + VP8LTransformColor = TransformColor; +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(VP8LEncDspInitNEON) + +#endif // WEBP_USE_NEON diff --git a/drivers/webp/dsp/lossless_enc_sse2.c b/drivers/webp/dsp/lossless_enc_sse2.c new file mode 100644 index 0000000000..1374b3ef64 --- /dev/null +++ b/drivers/webp/dsp/lossless_enc_sse2.c @@ -0,0 +1,270 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 variant of methods for lossless encoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE2) +#include <assert.h> +#include <emmintrin.h> +#include "./lossless.h" + +// For sign-extended multiplying constants, pre-shifted by 5: +#define CST_5b(X) (((int16_t)((uint16_t)X << 8)) >> 5) + +//------------------------------------------------------------------------------ +// Subtract-Green Transform + +static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb + const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g + const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); + const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g + const __m128i out = _mm_sub_epi8(in, C); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i); +} + +//------------------------------------------------------------------------------ +// Color Transform + +static void TransformColor(const VP8LMultipliers* const m, + uint32_t* argb_data, int num_pixels) { + const __m128i mults_rb = _mm_set_epi16( + CST_5b(m->green_to_red_), CST_5b(m->green_to_blue_), + CST_5b(m->green_to_red_), CST_5b(m->green_to_blue_), + CST_5b(m->green_to_red_), CST_5b(m->green_to_blue_), + CST_5b(m->green_to_red_), CST_5b(m->green_to_blue_)); + const __m128i mults_b2 = _mm_set_epi16( + CST_5b(m->red_to_blue_), 0, CST_5b(m->red_to_blue_), 0, + CST_5b(m->red_to_blue_), 0, CST_5b(m->red_to_blue_), 0); + const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks + const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff); // red-blue masks + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb + const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0 + const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); + const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0 + const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1 + const __m128i E = _mm_slli_epi16(in, 8); // r 0 b 0 + const __m128i F = _mm_mulhi_epi16(E, mults_b2); // x db2 0 0 + const __m128i G = _mm_srli_epi32(F, 16); // 0 0 x db2 + const __m128i H = _mm_add_epi8(G, D); // x dr x db + const __m128i I = _mm_and_si128(H, mask_rb); // 0 dr 0 db + const __m128i out = _mm_sub_epi8(in, I); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + VP8LTransformColor_C(m, argb_data + i, num_pixels - i); +} + +//------------------------------------------------------------------------------ +#define SPAN 8 +static void CollectColorBlueTransforms(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_blue, int red_to_blue, + int histo[]) { + const __m128i mults_r = _mm_set_epi16( + CST_5b(red_to_blue), 0, CST_5b(red_to_blue), 0, + CST_5b(red_to_blue), 0, CST_5b(red_to_blue), 0); + const __m128i mults_g = _mm_set_epi16( + 0, CST_5b(green_to_blue), 0, CST_5b(green_to_blue), + 0, CST_5b(green_to_blue), 0, CST_5b(green_to_blue)); + const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask + const __m128i mask_b = _mm_set1_epi32(0x0000ff); // blue mask + int y; + for (y = 0; y < tile_height; ++y) { + const uint32_t* const src = argb + y * stride; + int i, x; + for (x = 0; x + SPAN <= tile_width; x += SPAN) { + uint16_t values[SPAN]; + const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); + const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); + const __m128i A0 = _mm_slli_epi16(in0, 8); // r 0 | b 0 + const __m128i A1 = _mm_slli_epi16(in1, 8); + const __m128i B0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0 + const __m128i B1 = _mm_and_si128(in1, mask_g); + const __m128i C0 = _mm_mulhi_epi16(A0, mults_r); // x db | 0 0 + const __m128i C1 = _mm_mulhi_epi16(A1, mults_r); + const __m128i D0 = _mm_mulhi_epi16(B0, mults_g); // 0 0 | x db + const __m128i D1 = _mm_mulhi_epi16(B1, mults_g); + const __m128i E0 = _mm_sub_epi8(in0, D0); // x x | x b' + const __m128i E1 = _mm_sub_epi8(in1, D1); + const __m128i F0 = _mm_srli_epi32(C0, 16); // 0 0 | x db + const __m128i F1 = _mm_srli_epi32(C1, 16); + const __m128i G0 = _mm_sub_epi8(E0, F0); // 0 0 | x b' + const __m128i G1 = _mm_sub_epi8(E1, F1); + const __m128i H0 = _mm_and_si128(G0, mask_b); // 0 0 | 0 b + const __m128i H1 = _mm_and_si128(G1, mask_b); + const __m128i I = _mm_packs_epi32(H0, H1); // 0 b' | 0 b' + _mm_storeu_si128((__m128i*)values, I); + for (i = 0; i < SPAN; ++i) ++histo[values[i]]; + } + } + { + const int left_over = tile_width & (SPAN - 1); + if (left_over > 0) { + VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride, + left_over, tile_height, + green_to_blue, red_to_blue, histo); + } + } +} + +static void CollectColorRedTransforms(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]) { + const __m128i mults_g = _mm_set_epi16( + 0, CST_5b(green_to_red), 0, CST_5b(green_to_red), + 0, CST_5b(green_to_red), 0, CST_5b(green_to_red)); + const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask + const __m128i mask = _mm_set1_epi32(0xff); + + int y; + for (y = 0; y < tile_height; ++y) { + const uint32_t* const src = argb + y * stride; + int i, x; + for (x = 0; x + SPAN <= tile_width; x += SPAN) { + uint16_t values[SPAN]; + const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); + const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); + const __m128i A0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0 + const __m128i A1 = _mm_and_si128(in1, mask_g); + const __m128i B0 = _mm_srli_epi32(in0, 16); // 0 0 | x r + const __m128i B1 = _mm_srli_epi32(in1, 16); + const __m128i C0 = _mm_mulhi_epi16(A0, mults_g); // 0 0 | x dr + const __m128i C1 = _mm_mulhi_epi16(A1, mults_g); + const __m128i E0 = _mm_sub_epi8(B0, C0); // x x | x r' + const __m128i E1 = _mm_sub_epi8(B1, C1); + const __m128i F0 = _mm_and_si128(E0, mask); // 0 0 | 0 r' + const __m128i F1 = _mm_and_si128(E1, mask); + const __m128i I = _mm_packs_epi32(F0, F1); + _mm_storeu_si128((__m128i*)values, I); + for (i = 0; i < SPAN; ++i) ++histo[values[i]]; + } + } + { + const int left_over = tile_width & (SPAN - 1); + if (left_over > 0) { + VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride, + left_over, tile_height, + green_to_red, histo); + } + } +} +#undef SPAN + +//------------------------------------------------------------------------------ + +#define LINE_SIZE 16 // 8 or 16 +static void AddVector(const uint32_t* a, const uint32_t* b, uint32_t* out, + int size) { + int i; + assert(size % LINE_SIZE == 0); + for (i = 0; i < size; i += LINE_SIZE) { + const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]); + const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]); +#if (LINE_SIZE == 16) + const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]); + const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]); +#endif + const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]); + const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]); +#if (LINE_SIZE == 16) + const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i + 8]); + const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]); +#endif + _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0)); + _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1)); +#if (LINE_SIZE == 16) + _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2)); + _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3)); +#endif + } +} + +static void AddVectorEq(const uint32_t* a, uint32_t* out, int size) { + int i; + assert(size % LINE_SIZE == 0); + for (i = 0; i < size; i += LINE_SIZE) { + const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]); + const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]); +#if (LINE_SIZE == 16) + const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]); + const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]); +#endif + const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i + 0]); + const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i + 4]); +#if (LINE_SIZE == 16) + const __m128i b2 = _mm_loadu_si128((const __m128i*)&out[i + 8]); + const __m128i b3 = _mm_loadu_si128((const __m128i*)&out[i + 12]); +#endif + _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0)); + _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1)); +#if (LINE_SIZE == 16) + _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2)); + _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3)); +#endif + } +} +#undef LINE_SIZE + +// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But +// that's ok since the histogram values are less than 1<<28 (max picture size). +static void HistogramAdd(const VP8LHistogram* const a, + const VP8LHistogram* const b, + VP8LHistogram* const out) { + int i; + const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_); + assert(a->palette_code_bits_ == b->palette_code_bits_); + if (b != out) { + AddVector(a->literal_, b->literal_, out->literal_, NUM_LITERAL_CODES); + AddVector(a->red_, b->red_, out->red_, NUM_LITERAL_CODES); + AddVector(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES); + AddVector(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES); + } else { + AddVectorEq(a->literal_, out->literal_, NUM_LITERAL_CODES); + AddVectorEq(a->red_, out->red_, NUM_LITERAL_CODES); + AddVectorEq(a->blue_, out->blue_, NUM_LITERAL_CODES); + AddVectorEq(a->alpha_, out->alpha_, NUM_LITERAL_CODES); + } + for (i = NUM_LITERAL_CODES; i < literal_size; ++i) { + out->literal_[i] = a->literal_[i] + b->literal_[i]; + } + for (i = 0; i < NUM_DISTANCE_CODES; ++i) { + out->distance_[i] = a->distance_[i] + b->distance_[i]; + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed; + VP8LTransformColor = TransformColor; + VP8LCollectColorBlueTransforms = CollectColorBlueTransforms; + VP8LCollectColorRedTransforms = CollectColorRedTransforms; + VP8LHistogramAdd = HistogramAdd; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8LEncDspInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/lossless_enc_sse41.c b/drivers/webp/dsp/lossless_enc_sse41.c new file mode 100644 index 0000000000..3e493198db --- /dev/null +++ b/drivers/webp/dsp/lossless_enc_sse41.c @@ -0,0 +1,51 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE4.1 variant of methods for lossless encoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE41) +#include <assert.h> +#include <smmintrin.h> +#include "./lossless.h" + +//------------------------------------------------------------------------------ +// Subtract-Green Transform + +static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) { + int i; + const __m128i kCstShuffle = _mm_set_epi8(-1, 13, -1, 13, -1, 9, -1, 9, + -1, 5, -1, 5, -1, 1, -1, 1); + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); + const __m128i in_0g0g = _mm_shuffle_epi8(in, kCstShuffle); + const __m128i out = _mm_sub_epi8(in, in_0g0g); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE41(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed; +} + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(VP8LEncDspInitSSE41) + +#endif // WEBP_USE_SSE41 diff --git a/drivers/webp/dsp/lossless_mips_dsp_r2.c b/drivers/webp/dsp/lossless_mips_dsp_r2.c new file mode 100644 index 0000000000..90aed7f151 --- /dev/null +++ b/drivers/webp/dsp/lossless_mips_dsp_r2.c @@ -0,0 +1,680 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Image transforms and color space conversion methods for lossless decoder. +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "./lossless.h" + +#define MAP_COLOR_FUNCS(FUNC_NAME, TYPE, GET_INDEX, GET_VALUE) \ +static void FUNC_NAME(const TYPE* src, \ + const uint32_t* const color_map, \ + TYPE* dst, int y_start, int y_end, \ + int width) { \ + int y; \ + for (y = y_start; y < y_end; ++y) { \ + int x; \ + for (x = 0; x < (width >> 2); ++x) { \ + int tmp1, tmp2, tmp3, tmp4; \ + __asm__ volatile ( \ + ".ifc " #TYPE ", uint8_t \n\t" \ + "lbu %[tmp1], 0(%[src]) \n\t" \ + "lbu %[tmp2], 1(%[src]) \n\t" \ + "lbu %[tmp3], 2(%[src]) \n\t" \ + "lbu %[tmp4], 3(%[src]) \n\t" \ + "addiu %[src], %[src], 4 \n\t" \ + ".endif \n\t" \ + ".ifc " #TYPE ", uint32_t \n\t" \ + "lw %[tmp1], 0(%[src]) \n\t" \ + "lw %[tmp2], 4(%[src]) \n\t" \ + "lw %[tmp3], 8(%[src]) \n\t" \ + "lw %[tmp4], 12(%[src]) \n\t" \ + "ext %[tmp1], %[tmp1], 8, 8 \n\t" \ + "ext %[tmp2], %[tmp2], 8, 8 \n\t" \ + "ext %[tmp3], %[tmp3], 8, 8 \n\t" \ + "ext %[tmp4], %[tmp4], 8, 8 \n\t" \ + "addiu %[src], %[src], 16 \n\t" \ + ".endif \n\t" \ + "sll %[tmp1], %[tmp1], 2 \n\t" \ + "sll %[tmp2], %[tmp2], 2 \n\t" \ + "sll %[tmp3], %[tmp3], 2 \n\t" \ + "sll %[tmp4], %[tmp4], 2 \n\t" \ + "lwx %[tmp1], %[tmp1](%[color_map]) \n\t" \ + "lwx %[tmp2], %[tmp2](%[color_map]) \n\t" \ + "lwx %[tmp3], %[tmp3](%[color_map]) \n\t" \ + "lwx %[tmp4], %[tmp4](%[color_map]) \n\t" \ + ".ifc " #TYPE ", uint8_t \n\t" \ + "ext %[tmp1], %[tmp1], 8, 8 \n\t" \ + "ext %[tmp2], %[tmp2], 8, 8 \n\t" \ + "ext %[tmp3], %[tmp3], 8, 8 \n\t" \ + "ext %[tmp4], %[tmp4], 8, 8 \n\t" \ + "sb %[tmp1], 0(%[dst]) \n\t" \ + "sb %[tmp2], 1(%[dst]) \n\t" \ + "sb %[tmp3], 2(%[dst]) \n\t" \ + "sb %[tmp4], 3(%[dst]) \n\t" \ + "addiu %[dst], %[dst], 4 \n\t" \ + ".endif \n\t" \ + ".ifc " #TYPE ", uint32_t \n\t" \ + "sw %[tmp1], 0(%[dst]) \n\t" \ + "sw %[tmp2], 4(%[dst]) \n\t" \ + "sw %[tmp3], 8(%[dst]) \n\t" \ + "sw %[tmp4], 12(%[dst]) \n\t" \ + "addiu %[dst], %[dst], 16 \n\t" \ + ".endif \n\t" \ + : [tmp1]"=&r"(tmp1), [tmp2]"=&r"(tmp2), [tmp3]"=&r"(tmp3), \ + [tmp4]"=&r"(tmp4), [src]"+&r"(src), [dst]"+r"(dst) \ + : [color_map]"r"(color_map) \ + : "memory" \ + ); \ + } \ + for (x = 0; x < (width & 3); ++x) { \ + *dst++ = GET_VALUE(color_map[GET_INDEX(*src++)]); \ + } \ + } \ +} + +MAP_COLOR_FUNCS(MapARGB, uint32_t, VP8GetARGBIndex, VP8GetARGBValue) +MAP_COLOR_FUNCS(MapAlpha, uint8_t, VP8GetAlphaIndex, VP8GetAlphaValue) + +#undef MAP_COLOR_FUNCS + +static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, + uint32_t c2) { + int temp0, temp1, temp2, temp3, temp4, temp5; + __asm__ volatile ( + "preceu.ph.qbr %[temp1], %[c0] \n\t" + "preceu.ph.qbl %[temp2], %[c0] \n\t" + "preceu.ph.qbr %[temp3], %[c1] \n\t" + "preceu.ph.qbl %[temp4], %[c1] \n\t" + "preceu.ph.qbr %[temp5], %[c2] \n\t" + "preceu.ph.qbl %[temp0], %[c2] \n\t" + "subq.ph %[temp3], %[temp3], %[temp5] \n\t" + "subq.ph %[temp4], %[temp4], %[temp0] \n\t" + "addq.ph %[temp1], %[temp1], %[temp3] \n\t" + "addq.ph %[temp2], %[temp2], %[temp4] \n\t" + "shll_s.ph %[temp1], %[temp1], 7 \n\t" + "shll_s.ph %[temp2], %[temp2], 7 \n\t" + "precrqu_s.qb.ph %[temp2], %[temp2], %[temp1] \n\t" + : [temp0]"=r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5) + : [c0]"r"(c0), [c1]"r"(c1), [c2]"r"(c2) + : "memory" + ); + return temp2; +} + +static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, + uint32_t c2) { + int temp0, temp1, temp2, temp3, temp4, temp5; + __asm__ volatile ( + "adduh.qb %[temp5], %[c0], %[c1] \n\t" + "preceu.ph.qbr %[temp3], %[c2] \n\t" + "preceu.ph.qbr %[temp1], %[temp5] \n\t" + "preceu.ph.qbl %[temp2], %[temp5] \n\t" + "preceu.ph.qbl %[temp4], %[c2] \n\t" + "subq.ph %[temp3], %[temp1], %[temp3] \n\t" + "subq.ph %[temp4], %[temp2], %[temp4] \n\t" + "shrl.ph %[temp5], %[temp3], 15 \n\t" + "shrl.ph %[temp0], %[temp4], 15 \n\t" + "addq.ph %[temp3], %[temp3], %[temp5] \n\t" + "addq.ph %[temp4], %[temp0], %[temp4] \n\t" + "shra.ph %[temp3], %[temp3], 1 \n\t" + "shra.ph %[temp4], %[temp4], 1 \n\t" + "addq.ph %[temp1], %[temp1], %[temp3] \n\t" + "addq.ph %[temp2], %[temp2], %[temp4] \n\t" + "shll_s.ph %[temp1], %[temp1], 7 \n\t" + "shll_s.ph %[temp2], %[temp2], 7 \n\t" + "precrqu_s.qb.ph %[temp1], %[temp2], %[temp1] \n\t" + : [temp0]"=r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=r"(temp4), [temp5]"=&r"(temp5) + : [c0]"r"(c0), [c1]"r"(c1), [c2]"r"(c2) + : "memory" + ); + return temp1; +} + +static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { + int temp0, temp1, temp2, temp3, temp4, temp5; + __asm__ volatile ( + "cmpgdu.lt.qb %[temp1], %[c], %[b] \n\t" + "pick.qb %[temp1], %[b], %[c] \n\t" + "pick.qb %[temp2], %[c], %[b] \n\t" + "cmpgdu.lt.qb %[temp4], %[c], %[a] \n\t" + "pick.qb %[temp4], %[a], %[c] \n\t" + "pick.qb %[temp5], %[c], %[a] \n\t" + "subu.qb %[temp3], %[temp1], %[temp2] \n\t" + "subu.qb %[temp0], %[temp4], %[temp5] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "raddu.w.qb %[temp0], %[temp0] \n\t" + "subu %[temp3], %[temp3], %[temp0] \n\t" + "slti %[temp0], %[temp3], 0x1 \n\t" + "movz %[a], %[b], %[temp0] \n\t" + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp0]"=&r"(temp0), + [a]"+&r"(a) + : [b]"r"(b), [c]"r"(c) + ); + return a; +} + +static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) { + __asm__ volatile ( + "adduh.qb %[a0], %[a0], %[a1] \n\t" + : [a0]"+r"(a0) + : [a1]"r"(a1) + ); + return a0; +} + +static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) { + return Average2(Average2(a0, a2), a1); +} + +static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1, + uint32_t a2, uint32_t a3) { + return Average2(Average2(a0, a1), Average2(a2, a3)); +} + +static uint32_t Predictor5(uint32_t left, const uint32_t* const top) { + return Average3(left, top[0], top[1]); +} + +static uint32_t Predictor6(uint32_t left, const uint32_t* const top) { + return Average2(left, top[-1]); +} + +static uint32_t Predictor7(uint32_t left, const uint32_t* const top) { + return Average2(left, top[0]); +} + +static uint32_t Predictor8(uint32_t left, const uint32_t* const top) { + (void)left; + return Average2(top[-1], top[0]); +} + +static uint32_t Predictor9(uint32_t left, const uint32_t* const top) { + (void)left; + return Average2(top[0], top[1]); +} + +static uint32_t Predictor10(uint32_t left, const uint32_t* const top) { + return Average4(left, top[-1], top[0], top[1]); +} + +static uint32_t Predictor11(uint32_t left, const uint32_t* const top) { + return Select(top[0], left, top[-1]); +} + +static uint32_t Predictor12(uint32_t left, const uint32_t* const top) { + return ClampedAddSubtractFull(left, top[0], top[-1]); +} + +static uint32_t Predictor13(uint32_t left, const uint32_t* const top) { + return ClampedAddSubtractHalf(left, top[0], top[-1]); +} + +// Add green to blue and red channels (i.e. perform the inverse transform of +// 'subtract green'). +static void AddGreenToBlueAndRed(uint32_t* data, int num_pixels) { + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + uint32_t* const p_loop1_end = data + (num_pixels & ~3); + uint32_t* const p_loop2_end = data + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[data], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[data]) \n\t" + "lw %[temp1], 4(%[data]) \n\t" + "lw %[temp2], 8(%[data]) \n\t" + "lw %[temp3], 12(%[data]) \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "ext %[temp5], %[temp1], 8, 8 \n\t" + "ext %[temp6], %[temp2], 8, 8 \n\t" + "ext %[temp7], %[temp3], 8, 8 \n\t" + "addiu %[data], %[data], 16 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "replv.ph %[temp5], %[temp5] \n\t" + "replv.ph %[temp6], %[temp6] \n\t" + "replv.ph %[temp7], %[temp7] \n\t" + "addu.qb %[temp0], %[temp0], %[temp4] \n\t" + "addu.qb %[temp1], %[temp1], %[temp5] \n\t" + "addu.qb %[temp2], %[temp2], %[temp6] \n\t" + "addu.qb %[temp3], %[temp3], %[temp7] \n\t" + "sw %[temp0], -16(%[data]) \n\t" + "sw %[temp1], -12(%[data]) \n\t" + "sw %[temp2], -8(%[data]) \n\t" + "bne %[data], %[p_loop1_end], 0b \n\t" + " sw %[temp3], -4(%[data]) \n\t" + "3: \n\t" + "beq %[data], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[data]) \n\t" + "addiu %[data], %[data], 4 \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "addu.qb %[temp0], %[temp0], %[temp4] \n\t" + "bne %[data], %[p_loop2_end], 1b \n\t" + " sw %[temp0], -4(%[data]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [data]"+&r"(data), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), + [temp5]"=&r"(temp5), [temp6]"=&r"(temp6), [temp7]"=&r"(temp7) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void TransformColorInverse(const VP8LMultipliers* const m, + uint32_t* data, int num_pixels) { + int temp0, temp1, temp2, temp3, temp4, temp5; + uint32_t argb, argb1, new_red; + const uint32_t G_to_R = m->green_to_red_; + const uint32_t G_to_B = m->green_to_blue_; + const uint32_t R_to_B = m->red_to_blue_; + uint32_t* const p_loop_end = data + (num_pixels & ~1); + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[data], %[p_loop_end], 1f \n\t" + " nop \n\t" + "replv.ph %[temp0], %[G_to_R] \n\t" + "replv.ph %[temp1], %[G_to_B] \n\t" + "replv.ph %[temp2], %[R_to_B] \n\t" + "shll.ph %[temp0], %[temp0], 8 \n\t" + "shll.ph %[temp1], %[temp1], 8 \n\t" + "shll.ph %[temp2], %[temp2], 8 \n\t" + "shra.ph %[temp0], %[temp0], 8 \n\t" + "shra.ph %[temp1], %[temp1], 8 \n\t" + "shra.ph %[temp2], %[temp2], 8 \n\t" + "0: \n\t" + "lw %[argb], 0(%[data]) \n\t" + "lw %[argb1], 4(%[data]) \n\t" + "addiu %[data], %[data], 8 \n\t" + "precrq.qb.ph %[temp3], %[argb], %[argb1] \n\t" + "preceu.ph.qbra %[temp3], %[temp3] \n\t" + "shll.ph %[temp3], %[temp3], 8 \n\t" + "shra.ph %[temp3], %[temp3], 8 \n\t" + "mul.ph %[temp5], %[temp3], %[temp0] \n\t" + "mul.ph %[temp3], %[temp3], %[temp1] \n\t" + "precrq.ph.w %[new_red], %[argb], %[argb1] \n\t" + "ins %[argb1], %[argb], 16, 16 \n\t" + "shra.ph %[temp5], %[temp5], 5 \n\t" + "shra.ph %[temp3], %[temp3], 5 \n\t" + "addu.ph %[new_red], %[new_red], %[temp5] \n\t" + "addu.ph %[argb1], %[argb1], %[temp3] \n\t" + "preceu.ph.qbra %[temp5], %[new_red] \n\t" + "shll.ph %[temp4], %[temp5], 8 \n\t" + "shra.ph %[temp4], %[temp4], 8 \n\t" + "mul.ph %[temp4], %[temp4], %[temp2] \n\t" + "sb %[temp5], -2(%[data]) \n\t" + "sra %[temp5], %[temp5], 16 \n\t" + "shra.ph %[temp4], %[temp4], 5 \n\t" + "addu.ph %[argb1], %[argb1], %[temp4] \n\t" + "preceu.ph.qbra %[temp3], %[argb1] \n\t" + "sb %[temp5], -6(%[data]) \n\t" + "sb %[temp3], -4(%[data]) \n\t" + "sra %[temp3], %[temp3], 16 \n\t" + "bne %[data], %[p_loop_end], 0b \n\t" + " sb %[temp3], -8(%[data]) \n\t" + "1: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [new_red]"=&r"(new_red), [argb]"=&r"(argb), + [argb1]"=&r"(argb1), [data]"+&r"(data) + : [G_to_R]"r"(G_to_R), [R_to_B]"r"(R_to_B), + [G_to_B]"r"(G_to_B), [p_loop_end]"r"(p_loop_end) + : "memory", "hi", "lo" + ); + + // Fall-back to C-version for left-overs. + if (num_pixels & 1) VP8LTransformColorInverse_C(m, data, 1); +} + +static void ConvertBGRAToRGB(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "ins %[temp3], %[temp2], 24, 8 \n\t" + "sll %[temp2], %[temp2], 8 \n\t" + "rotr %[temp3], %[temp3], 16 \n\t" + "ins %[temp2], %[temp1], 0, 16 \n\t" + "sll %[temp1], %[temp1], 8 \n\t" + "wsbh %[temp3], %[temp3] \n\t" + "balign %[temp0], %[temp1], 1 \n\t" + "wsbh %[temp2], %[temp2] \n\t" + "wsbh %[temp0], %[temp0] \n\t" + "usw %[temp3], 8(%[dst]) \n\t" + "rotr %[temp0], %[temp0], 16 \n\t" + "usw %[temp2], 4(%[dst]) \n\t" + "addiu %[src], %[src], 16 \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 12 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "addiu %[src], %[src], 4 \n\t" + "wsbh %[temp1], %[temp0] \n\t" + "addiu %[dst], %[dst], 3 \n\t" + "ush %[temp1], -2(%[dst]) \n\t" + "sra %[temp0], %[temp0], 16 \n\t" + "bne %[src], %[p_loop2_end], 1b \n\t" + " sb %[temp0], -3(%[dst]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void ConvertBGRAToRGBA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "wsbh %[temp0], %[temp0] \n\t" + "wsbh %[temp1], %[temp1] \n\t" + "wsbh %[temp2], %[temp2] \n\t" + "wsbh %[temp3], %[temp3] \n\t" + "addiu %[src], %[src], 16 \n\t" + "balign %[temp0], %[temp0], 1 \n\t" + "balign %[temp1], %[temp1], 1 \n\t" + "balign %[temp2], %[temp2], 1 \n\t" + "balign %[temp3], %[temp3], 1 \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "usw %[temp1], 4(%[dst]) \n\t" + "usw %[temp2], 8(%[dst]) \n\t" + "usw %[temp3], 12(%[dst]) \n\t" + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 16 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "wsbh %[temp0], %[temp0] \n\t" + "addiu %[src], %[src], 4 \n\t" + "balign %[temp0], %[temp0], 1 \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "bne %[src], %[p_loop2_end], 1b \n\t" + " addiu %[dst], %[dst], 4 \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void ConvertBGRAToRGBA4444(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3, temp4, temp5; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "ext %[temp4], %[temp0], 28, 4 \n\t" + "ext %[temp5], %[temp0], 12, 4 \n\t" + "ins %[temp0], %[temp4], 0, 4 \n\t" + "ext %[temp4], %[temp1], 28, 4 \n\t" + "ins %[temp0], %[temp5], 16, 4 \n\t" + "ext %[temp5], %[temp1], 12, 4 \n\t" + "ins %[temp1], %[temp4], 0, 4 \n\t" + "ext %[temp4], %[temp2], 28, 4 \n\t" + "ins %[temp1], %[temp5], 16, 4 \n\t" + "ext %[temp5], %[temp2], 12, 4 \n\t" + "ins %[temp2], %[temp4], 0, 4 \n\t" + "ext %[temp4], %[temp3], 28, 4 \n\t" + "ins %[temp2], %[temp5], 16, 4 \n\t" + "ext %[temp5], %[temp3], 12, 4 \n\t" + "ins %[temp3], %[temp4], 0, 4 \n\t" + "precr.qb.ph %[temp1], %[temp1], %[temp0] \n\t" + "ins %[temp3], %[temp5], 16, 4 \n\t" + "addiu %[src], %[src], 16 \n\t" + "precr.qb.ph %[temp3], %[temp3], %[temp2] \n\t" +#ifdef WEBP_SWAP_16BIT_CSP + "usw %[temp1], 0(%[dst]) \n\t" + "usw %[temp3], 4(%[dst]) \n\t" +#else + "wsbh %[temp1], %[temp1] \n\t" + "wsbh %[temp3], %[temp3] \n\t" + "usw %[temp1], 0(%[dst]) \n\t" + "usw %[temp3], 4(%[dst]) \n\t" +#endif + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 8 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "ext %[temp4], %[temp0], 28, 4 \n\t" + "ext %[temp5], %[temp0], 12, 4 \n\t" + "ins %[temp0], %[temp4], 0, 4 \n\t" + "ins %[temp0], %[temp5], 16, 4 \n\t" + "addiu %[src], %[src], 4 \n\t" + "precr.qb.ph %[temp0], %[temp0], %[temp0] \n\t" +#ifdef WEBP_SWAP_16BIT_CSP + "ush %[temp0], 0(%[dst]) \n\t" +#else + "wsbh %[temp0], %[temp0] \n\t" + "ush %[temp0], 0(%[dst]) \n\t" +#endif + "bne %[src], %[p_loop2_end], 1b \n\t" + " addiu %[dst], %[dst], 2 \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void ConvertBGRAToRGB565(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3, temp4, temp5; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "ext %[temp4], %[temp0], 8, 16 \n\t" + "ext %[temp5], %[temp0], 5, 11 \n\t" + "ext %[temp0], %[temp0], 3, 5 \n\t" + "ins %[temp4], %[temp5], 0, 11 \n\t" + "ext %[temp5], %[temp1], 5, 11 \n\t" + "ins %[temp4], %[temp0], 0, 5 \n\t" + "ext %[temp0], %[temp1], 8, 16 \n\t" + "ext %[temp1], %[temp1], 3, 5 \n\t" + "ins %[temp0], %[temp5], 0, 11 \n\t" + "ext %[temp5], %[temp2], 5, 11 \n\t" + "ins %[temp0], %[temp1], 0, 5 \n\t" + "ext %[temp1], %[temp2], 8, 16 \n\t" + "ext %[temp2], %[temp2], 3, 5 \n\t" + "ins %[temp1], %[temp5], 0, 11 \n\t" + "ext %[temp5], %[temp3], 5, 11 \n\t" + "ins %[temp1], %[temp2], 0, 5 \n\t" + "ext %[temp2], %[temp3], 8, 16 \n\t" + "ext %[temp3], %[temp3], 3, 5 \n\t" + "ins %[temp2], %[temp5], 0, 11 \n\t" + "append %[temp0], %[temp4], 16 \n\t" + "ins %[temp2], %[temp3], 0, 5 \n\t" + "addiu %[src], %[src], 16 \n\t" + "append %[temp2], %[temp1], 16 \n\t" +#ifdef WEBP_SWAP_16BIT_CSP + "usw %[temp0], 0(%[dst]) \n\t" + "usw %[temp2], 4(%[dst]) \n\t" +#else + "wsbh %[temp0], %[temp0] \n\t" + "wsbh %[temp2], %[temp2] \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "usw %[temp2], 4(%[dst]) \n\t" +#endif + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 8 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "ext %[temp4], %[temp0], 8, 16 \n\t" + "ext %[temp5], %[temp0], 5, 11 \n\t" + "ext %[temp0], %[temp0], 3, 5 \n\t" + "ins %[temp4], %[temp5], 0, 11 \n\t" + "addiu %[src], %[src], 4 \n\t" + "ins %[temp4], %[temp0], 0, 5 \n\t" +#ifdef WEBP_SWAP_16BIT_CSP + "ush %[temp4], 0(%[dst]) \n\t" +#else + "wsbh %[temp4], %[temp4] \n\t" + "ush %[temp4], 0(%[dst]) \n\t" +#endif + "bne %[src], %[p_loop2_end], 1b \n\t" + " addiu %[dst], %[dst], 2 \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void ConvertBGRAToBGR(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "ins %[temp0], %[temp1], 24, 8 \n\t" + "sra %[temp1], %[temp1], 8 \n\t" + "ins %[temp1], %[temp2], 16, 16 \n\t" + "sll %[temp2], %[temp2], 8 \n\t" + "balign %[temp3], %[temp2], 1 \n\t" + "addiu %[src], %[src], 16 \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "usw %[temp1], 4(%[dst]) \n\t" + "usw %[temp3], 8(%[dst]) \n\t" + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 12 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "addiu %[src], %[src], 4 \n\t" + "addiu %[dst], %[dst], 3 \n\t" + "ush %[temp0], -3(%[dst]) \n\t" + "sra %[temp0], %[temp0], 16 \n\t" + "bne %[src], %[p_loop2_end], 1b \n\t" + " sb %[temp0], -1(%[dst]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LDspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMIPSdspR2(void) { + VP8LMapColor32b = MapARGB; + VP8LMapColor8b = MapAlpha; + VP8LPredictors[5] = Predictor5; + VP8LPredictors[6] = Predictor6; + VP8LPredictors[7] = Predictor7; + VP8LPredictors[8] = Predictor8; + VP8LPredictors[9] = Predictor9; + VP8LPredictors[10] = Predictor10; + VP8LPredictors[11] = Predictor11; + VP8LPredictors[12] = Predictor12; + VP8LPredictors[13] = Predictor13; + VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; + VP8LTransformColorInverse = TransformColorInverse; + VP8LConvertBGRAToRGB = ConvertBGRAToRGB; + VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; + VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444; + VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8LDspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/lossless_neon.c b/drivers/webp/dsp/lossless_neon.c new file mode 100644 index 0000000000..6faccb8f97 --- /dev/null +++ b/drivers/webp/dsp/lossless_neon.c @@ -0,0 +1,269 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// NEON variant of methods for lossless decoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_NEON) + +#include <arm_neon.h> + +#include "./lossless.h" +#include "./neon.h" + +//------------------------------------------------------------------------------ +// Colorspace conversion functions + +#if !defined(WORK_AROUND_GCC) +// gcc 4.6.0 had some trouble (NDK-r9) with this code. We only use it for +// gcc-4.8.x at least. +static void ConvertBGRAToRGBA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint32_t* const end = src + (num_pixels & ~15); + for (; src < end; src += 16) { + uint8x16x4_t pixel = vld4q_u8((uint8_t*)src); + // swap B and R. (VSWP d0,d2 has no intrinsics equivalent!) + const uint8x16_t tmp = pixel.val[0]; + pixel.val[0] = pixel.val[2]; + pixel.val[2] = tmp; + vst4q_u8(dst, pixel); + dst += 64; + } + VP8LConvertBGRAToRGBA_C(src, num_pixels & 15, dst); // left-overs +} + +static void ConvertBGRAToBGR(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint32_t* const end = src + (num_pixels & ~15); + for (; src < end; src += 16) { + const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src); + const uint8x16x3_t tmp = { { pixel.val[0], pixel.val[1], pixel.val[2] } }; + vst3q_u8(dst, tmp); + dst += 48; + } + VP8LConvertBGRAToBGR_C(src, num_pixels & 15, dst); // left-overs +} + +static void ConvertBGRAToRGB(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint32_t* const end = src + (num_pixels & ~15); + for (; src < end; src += 16) { + const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src); + const uint8x16x3_t tmp = { { pixel.val[2], pixel.val[1], pixel.val[0] } }; + vst3q_u8(dst, tmp); + dst += 48; + } + VP8LConvertBGRAToRGB_C(src, num_pixels & 15, dst); // left-overs +} + +#else // WORK_AROUND_GCC + +// gcc-4.6.0 fallback + +static const uint8_t kRGBAShuffle[8] = { 2, 1, 0, 3, 6, 5, 4, 7 }; + +static void ConvertBGRAToRGBA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint32_t* const end = src + (num_pixels & ~1); + const uint8x8_t shuffle = vld1_u8(kRGBAShuffle); + for (; src < end; src += 2) { + const uint8x8_t pixels = vld1_u8((uint8_t*)src); + vst1_u8(dst, vtbl1_u8(pixels, shuffle)); + dst += 8; + } + VP8LConvertBGRAToRGBA_C(src, num_pixels & 1, dst); // left-overs +} + +static const uint8_t kBGRShuffle[3][8] = { + { 0, 1, 2, 4, 5, 6, 8, 9 }, + { 10, 12, 13, 14, 16, 17, 18, 20 }, + { 21, 22, 24, 25, 26, 28, 29, 30 } +}; + +static void ConvertBGRAToBGR(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint32_t* const end = src + (num_pixels & ~7); + const uint8x8_t shuffle0 = vld1_u8(kBGRShuffle[0]); + const uint8x8_t shuffle1 = vld1_u8(kBGRShuffle[1]); + const uint8x8_t shuffle2 = vld1_u8(kBGRShuffle[2]); + for (; src < end; src += 8) { + uint8x8x4_t pixels; + INIT_VECTOR4(pixels, + vld1_u8((const uint8_t*)(src + 0)), + vld1_u8((const uint8_t*)(src + 2)), + vld1_u8((const uint8_t*)(src + 4)), + vld1_u8((const uint8_t*)(src + 6))); + vst1_u8(dst + 0, vtbl4_u8(pixels, shuffle0)); + vst1_u8(dst + 8, vtbl4_u8(pixels, shuffle1)); + vst1_u8(dst + 16, vtbl4_u8(pixels, shuffle2)); + dst += 8 * 3; + } + VP8LConvertBGRAToBGR_C(src, num_pixels & 7, dst); // left-overs +} + +static const uint8_t kRGBShuffle[3][8] = { + { 2, 1, 0, 6, 5, 4, 10, 9 }, + { 8, 14, 13, 12, 18, 17, 16, 22 }, + { 21, 20, 26, 25, 24, 30, 29, 28 } +}; + +static void ConvertBGRAToRGB(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint32_t* const end = src + (num_pixels & ~7); + const uint8x8_t shuffle0 = vld1_u8(kRGBShuffle[0]); + const uint8x8_t shuffle1 = vld1_u8(kRGBShuffle[1]); + const uint8x8_t shuffle2 = vld1_u8(kRGBShuffle[2]); + for (; src < end; src += 8) { + uint8x8x4_t pixels; + INIT_VECTOR4(pixels, + vld1_u8((const uint8_t*)(src + 0)), + vld1_u8((const uint8_t*)(src + 2)), + vld1_u8((const uint8_t*)(src + 4)), + vld1_u8((const uint8_t*)(src + 6))); + vst1_u8(dst + 0, vtbl4_u8(pixels, shuffle0)); + vst1_u8(dst + 8, vtbl4_u8(pixels, shuffle1)); + vst1_u8(dst + 16, vtbl4_u8(pixels, shuffle2)); + dst += 8 * 3; + } + VP8LConvertBGRAToRGB_C(src, num_pixels & 7, dst); // left-overs +} + +#endif // !WORK_AROUND_GCC + +//------------------------------------------------------------------------------ +// Subtract-Green Transform + +// vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use +// non-standard versions there. +#if defined(__APPLE__) && defined(__aarch64__) && \ + defined(__apple_build_version__) && (__apple_build_version__< 6020037) +#define USE_VTBLQ +#endif + +#ifdef USE_VTBLQ +// 255 = byte will be zeroed +static const uint8_t kGreenShuffle[16] = { + 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255 +}; + +static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb, + const uint8x16_t shuffle) { + return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)), + vtbl1q_u8(argb, vget_high_u8(shuffle))); +} +#else // !USE_VTBLQ +// 255 = byte will be zeroed +static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255 }; + +static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb, + const uint8x8_t shuffle) { + return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle), + vtbl1_u8(vget_high_u8(argb), shuffle)); +} +#endif // USE_VTBLQ + +static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) { + const uint32_t* const end = argb_data + (num_pixels & ~3); +#ifdef USE_VTBLQ + const uint8x16_t shuffle = vld1q_u8(kGreenShuffle); +#else + const uint8x8_t shuffle = vld1_u8(kGreenShuffle); +#endif + for (; argb_data < end; argb_data += 4) { + const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data); + const uint8x16_t greens = DoGreenShuffle(argb, shuffle); + vst1q_u8((uint8_t*)argb_data, vaddq_u8(argb, greens)); + } + // fallthrough and finish off with plain-C + VP8LAddGreenToBlueAndRed_C(argb_data, num_pixels & 3); +} + +//------------------------------------------------------------------------------ +// Color Transform + +static void TransformColorInverse(const VP8LMultipliers* const m, + uint32_t* argb_data, int num_pixels) { + // sign-extended multiplying constants, pre-shifted by 6. +#define CST(X) (((int16_t)(m->X << 8)) >> 6) + const int16_t rb[8] = { + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_) + }; + const int16x8_t mults_rb = vld1q_s16(rb); + const int16_t b2[8] = { + 0, CST(red_to_blue_), 0, CST(red_to_blue_), + 0, CST(red_to_blue_), 0, CST(red_to_blue_), + }; + const int16x8_t mults_b2 = vld1q_s16(b2); +#undef CST +#ifdef USE_VTBLQ + static const uint8_t kg0g0[16] = { + 255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13 + }; + const uint8x16_t shuffle = vld1q_u8(kg0g0); +#else + static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 }; + const uint8x8_t shuffle = vld1_u8(k0g0g); +#endif + const uint32x4_t mask_ag = vdupq_n_u32(0xff00ff00u); + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const uint8x16_t in = vld1q_u8((uint8_t*)(argb_data + i)); + const uint32x4_t a0g0 = vandq_u32(vreinterpretq_u32_u8(in), mask_ag); + // 0 g 0 g + const uint8x16_t greens = DoGreenShuffle(in, shuffle); + // x dr x db1 + const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb); + // x r' x b' + const int8x16_t B = vaddq_s8(vreinterpretq_s8_u8(in), + vreinterpretq_s8_s16(A)); + // r' 0 b' 0 + const int16x8_t C = vshlq_n_s16(vreinterpretq_s16_s8(B), 8); + // x db2 0 0 + const int16x8_t D = vqdmulhq_s16(C, mults_b2); + // 0 x db2 0 + const uint32x4_t E = vshrq_n_u32(vreinterpretq_u32_s16(D), 8); + // r' x b'' 0 + const int8x16_t F = vaddq_s8(vreinterpretq_s8_u32(E), + vreinterpretq_s8_s16(C)); + // 0 r' 0 b'' + const uint16x8_t G = vshrq_n_u16(vreinterpretq_u16_s8(F), 8); + const uint32x4_t out = vorrq_u32(vreinterpretq_u32_u16(G), a0g0); + vst1q_u32(argb_data + i, out); + } + // Fall-back to C-version for left-overs. + VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i); +} + +#undef USE_VTBLQ + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LDspInitNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitNEON(void) { + VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR; + VP8LConvertBGRAToRGB = ConvertBGRAToRGB; + + VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; + VP8LTransformColorInverse = TransformColorInverse; +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(VP8LDspInitNEON) + +#endif // WEBP_USE_NEON diff --git a/drivers/webp/dsp/lossless_sse2.c b/drivers/webp/dsp/lossless_sse2.c new file mode 100644 index 0000000000..2d016c2911 --- /dev/null +++ b/drivers/webp/dsp/lossless_sse2.c @@ -0,0 +1,372 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 variant of methods for lossless decoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE2) +#include <assert.h> +#include <emmintrin.h> +#include "./lossless.h" + +//------------------------------------------------------------------------------ +// Predictor Transform + +static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, + uint32_t c2) { + const __m128i zero = _mm_setzero_si128(); + const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); + const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); + const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); + const __m128i V1 = _mm_add_epi16(C0, C1); + const __m128i V2 = _mm_sub_epi16(V1, C2); + const __m128i b = _mm_packus_epi16(V2, V2); + const uint32_t output = _mm_cvtsi128_si32(b); + return output; +} + +static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, + uint32_t c2) { + const __m128i zero = _mm_setzero_si128(); + const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); + const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); + const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); + const __m128i avg = _mm_add_epi16(C1, C0); + const __m128i A0 = _mm_srli_epi16(avg, 1); + const __m128i A1 = _mm_sub_epi16(A0, B0); + const __m128i BgtA = _mm_cmpgt_epi16(B0, A0); + const __m128i A2 = _mm_sub_epi16(A1, BgtA); + const __m128i A3 = _mm_srai_epi16(A2, 1); + const __m128i A4 = _mm_add_epi16(A0, A3); + const __m128i A5 = _mm_packus_epi16(A4, A4); + const uint32_t output = _mm_cvtsi128_si32(A5); + return output; +} + +static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { + int pa_minus_pb; + const __m128i zero = _mm_setzero_si128(); + const __m128i A0 = _mm_cvtsi32_si128(a); + const __m128i B0 = _mm_cvtsi32_si128(b); + const __m128i C0 = _mm_cvtsi32_si128(c); + const __m128i AC0 = _mm_subs_epu8(A0, C0); + const __m128i CA0 = _mm_subs_epu8(C0, A0); + const __m128i BC0 = _mm_subs_epu8(B0, C0); + const __m128i CB0 = _mm_subs_epu8(C0, B0); + const __m128i AC = _mm_or_si128(AC0, CA0); + const __m128i BC = _mm_or_si128(BC0, CB0); + const __m128i pa = _mm_unpacklo_epi8(AC, zero); // |a - c| + const __m128i pb = _mm_unpacklo_epi8(BC, zero); // |b - c| + const __m128i diff = _mm_sub_epi16(pb, pa); + { + int16_t out[8]; + _mm_storeu_si128((__m128i*)out, diff); + pa_minus_pb = out[0] + out[1] + out[2] + out[3]; + } + return (pa_minus_pb <= 0) ? a : b; +} + +static WEBP_INLINE __m128i Average2_128i(uint32_t a0, uint32_t a1) { + const __m128i zero = _mm_setzero_si128(); + const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a0), zero); + const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero); + const __m128i sum = _mm_add_epi16(A1, A0); + const __m128i avg = _mm_srli_epi16(sum, 1); + return avg; +} + +static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) { + const __m128i avg = Average2_128i(a0, a1); + const __m128i A2 = _mm_packus_epi16(avg, avg); + const uint32_t output = _mm_cvtsi128_si32(A2); + return output; +} + +static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) { + const __m128i zero = _mm_setzero_si128(); + const __m128i avg1 = Average2_128i(a0, a2); + const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero); + const __m128i sum = _mm_add_epi16(avg1, A1); + const __m128i avg2 = _mm_srli_epi16(sum, 1); + const __m128i A2 = _mm_packus_epi16(avg2, avg2); + const uint32_t output = _mm_cvtsi128_si32(A2); + return output; +} + +static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1, + uint32_t a2, uint32_t a3) { + const __m128i avg1 = Average2_128i(a0, a1); + const __m128i avg2 = Average2_128i(a2, a3); + const __m128i sum = _mm_add_epi16(avg2, avg1); + const __m128i avg3 = _mm_srli_epi16(sum, 1); + const __m128i A0 = _mm_packus_epi16(avg3, avg3); + const uint32_t output = _mm_cvtsi128_si32(A0); + return output; +} + +static uint32_t Predictor5(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average3(left, top[0], top[1]); + return pred; +} +static uint32_t Predictor6(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(left, top[-1]); + return pred; +} +static uint32_t Predictor7(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(left, top[0]); + return pred; +} +static uint32_t Predictor8(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(top[-1], top[0]); + (void)left; + return pred; +} +static uint32_t Predictor9(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(top[0], top[1]); + (void)left; + return pred; +} +static uint32_t Predictor10(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average4(left, top[-1], top[0], top[1]); + return pred; +} +static uint32_t Predictor11(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Select(top[0], left, top[-1]); + return pred; +} +static uint32_t Predictor12(uint32_t left, const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]); + return pred; +} +static uint32_t Predictor13(uint32_t left, const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]); + return pred; +} + +//------------------------------------------------------------------------------ +// Subtract-Green Transform + +static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb + const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g + const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); + const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g + const __m128i out = _mm_add_epi8(in, C); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + VP8LAddGreenToBlueAndRed_C(argb_data + i, num_pixels - i); +} + +//------------------------------------------------------------------------------ +// Color Transform + +static void TransformColorInverse(const VP8LMultipliers* const m, + uint32_t* argb_data, int num_pixels) { + // sign-extended multiplying constants, pre-shifted by 5. +#define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend + const __m128i mults_rb = _mm_set_epi16( + CST(green_to_red_), CST(green_to_blue_), + CST(green_to_red_), CST(green_to_blue_), + CST(green_to_red_), CST(green_to_blue_), + CST(green_to_red_), CST(green_to_blue_)); + const __m128i mults_b2 = _mm_set_epi16( + CST(red_to_blue_), 0, CST(red_to_blue_), 0, + CST(red_to_blue_), 0, CST(red_to_blue_), 0); +#undef CST + const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb + const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0 + const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); + const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0 + const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1 + const __m128i E = _mm_add_epi8(in, D); // x r' x b' + const __m128i F = _mm_slli_epi16(E, 8); // r' 0 b' 0 + const __m128i G = _mm_mulhi_epi16(F, mults_b2); // x db2 0 0 + const __m128i H = _mm_srli_epi32(G, 8); // 0 x db2 0 + const __m128i I = _mm_add_epi8(H, F); // r' x b'' 0 + const __m128i J = _mm_srli_epi16(I, 8); // 0 r' 0 b'' + const __m128i out = _mm_or_si128(J, A); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // Fall-back to C-version for left-overs. + VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i); +} + +//------------------------------------------------------------------------------ +// Color-space conversion functions + +static void ConvertBGRAToRGBA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + while (num_pixels >= 8) { + const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 + const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 + const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... + const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... + const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... + const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... + const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 + const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 + const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 + const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 + const __m128i rg0 = _mm_unpacklo_epi8(rb0, ga0); // r0g0r1g1 ... r6g6r7g7 + const __m128i ba0 = _mm_unpackhi_epi8(rb0, ga0); // b0a0b1a1 ... b6a6b7a7 + const __m128i rgba0 = _mm_unpacklo_epi16(rg0, ba0); // rgba0|rgba1... + const __m128i rgba4 = _mm_unpackhi_epi16(rg0, ba0); // rgba4|rgba5... + _mm_storeu_si128(out++, rgba0); + _mm_storeu_si128(out++, rgba4); + num_pixels -= 8; + } + // left-overs + VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out); +} + +static void ConvertBGRAToRGBA4444(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i mask_0x0f = _mm_set1_epi8(0x0f); + const __m128i mask_0xf0 = _mm_set1_epi8(0xf0); + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + while (num_pixels >= 8) { + const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 + const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 + const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... + const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... + const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... + const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... + const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 + const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 + const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 + const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 + const __m128i ga1 = _mm_srli_epi16(ga0, 4); // g0-|g1-|...|a6-|a7- + const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0); // -r0|-r1|...|-b6|-a7 + const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f); // g0-|g1-|...|a6-|a7- + const __m128i rgba0 = _mm_or_si128(ga2, rb1); // rg0..rg7 | ba0..ba7 + const __m128i rgba1 = _mm_srli_si128(rgba0, 8); // ba0..ba7 | 0 +#ifdef WEBP_SWAP_16BIT_CSP + const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0); // barg0...barg7 +#else + const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1); // rgba0...rgba7 +#endif + _mm_storeu_si128(out++, rgba); + num_pixels -= 8; + } + // left-overs + VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out); +} + +static void ConvertBGRAToRGB565(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i mask_0xe0 = _mm_set1_epi8(0xe0); + const __m128i mask_0xf8 = _mm_set1_epi8(0xf8); + const __m128i mask_0x07 = _mm_set1_epi8(0x07); + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + while (num_pixels >= 8) { + const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 + const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 + const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... + const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... + const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... + const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... + const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 + const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 + const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 + const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 + const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8); // -r0..-r7|-b0..-b7 + const __m128i g_lo1 = _mm_srli_epi16(ga0, 5); + const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07); // g0-...g7-|xx (3b) + const __m128i g_hi1 = _mm_slli_epi16(ga0, 3); + const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0); // -g0...-g7|xx (3b) + const __m128i b0 = _mm_srli_si128(rb1, 8); // -b0...-b7|0 + const __m128i rg1 = _mm_or_si128(rb1, g_lo2); // gr0...gr7|xx + const __m128i b1 = _mm_srli_epi16(b0, 3); + const __m128i gb1 = _mm_or_si128(b1, g_hi2); // bg0...bg7|xx +#ifdef WEBP_SWAP_16BIT_CSP + const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1); // rggb0...rggb7 +#else + const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1); // bgrb0...bgrb7 +#endif + _mm_storeu_si128(out++, rgba); + num_pixels -= 8; + } + // left-overs + VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out); +} + +static void ConvertBGRAToBGR(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff); + const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0); + const __m128i* in = (const __m128i*)src; + const uint8_t* const end = dst + num_pixels * 3; + // the last storel_epi64 below writes 8 bytes starting at offset 18 + while (dst + 26 <= end) { + const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 + const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 + const __m128i a0l = _mm_and_si128(bgra0, mask_l); // bgr0|0|bgr0|0 + const __m128i a4l = _mm_and_si128(bgra4, mask_l); // bgr0|0|bgr0|0 + const __m128i a0h = _mm_and_si128(bgra0, mask_h); // 0|bgr0|0|bgr0 + const __m128i a4h = _mm_and_si128(bgra4, mask_h); // 0|bgr0|0|bgr0 + const __m128i b0h = _mm_srli_epi64(a0h, 8); // 000b|gr00|000b|gr00 + const __m128i b4h = _mm_srli_epi64(a4h, 8); // 000b|gr00|000b|gr00 + const __m128i c0 = _mm_or_si128(a0l, b0h); // rgbrgb00|rgbrgb00 + const __m128i c4 = _mm_or_si128(a4l, b4h); // rgbrgb00|rgbrgb00 + const __m128i c2 = _mm_srli_si128(c0, 8); + const __m128i c6 = _mm_srli_si128(c4, 8); + _mm_storel_epi64((__m128i*)(dst + 0), c0); + _mm_storel_epi64((__m128i*)(dst + 6), c2); + _mm_storel_epi64((__m128i*)(dst + 12), c4); + _mm_storel_epi64((__m128i*)(dst + 18), c6); + dst += 24; + num_pixels -= 8; + } + // left-overs + VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LDspInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE2(void) { + VP8LPredictors[5] = Predictor5; + VP8LPredictors[6] = Predictor6; + VP8LPredictors[7] = Predictor7; + VP8LPredictors[8] = Predictor8; + VP8LPredictors[9] = Predictor9; + VP8LPredictors[10] = Predictor10; + VP8LPredictors[11] = Predictor11; + VP8LPredictors[12] = Predictor12; + VP8LPredictors[13] = Predictor13; + + VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; + VP8LTransformColorInverse = TransformColorInverse; + + VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; + VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444; + VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8LDspInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/mips_macro.h b/drivers/webp/dsp/mips_macro.h new file mode 100644 index 0000000000..44aba9b71d --- /dev/null +++ b/drivers/webp/dsp/mips_macro.h @@ -0,0 +1,200 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS common macros + +#ifndef WEBP_DSP_MIPS_MACRO_H_ +#define WEBP_DSP_MIPS_MACRO_H_ + +#if defined(__GNUC__) && defined(__ANDROID__) && LOCAL_GCC_VERSION == 0x409 +#define WORK_AROUND_GCC +#endif + +#define STR(s) #s +#define XSTR(s) STR(s) + +// O0[31..16 | 15..0] = I0[31..16 | 15..0] + I1[31..16 | 15..0] +// O1[31..16 | 15..0] = I0[31..16 | 15..0] - I1[31..16 | 15..0] +// O - output +// I - input (macro doesn't change it) +#define ADD_SUB_HALVES(O0, O1, \ + I0, I1) \ + "addq.ph %[" #O0 "], %[" #I0 "], %[" #I1 "] \n\t" \ + "subq.ph %[" #O1 "], %[" #I0 "], %[" #I1 "] \n\t" + +// O - output +// I - input (macro doesn't change it) +// I[0/1] - offset in bytes +#define LOAD_IN_X2(O0, O1, \ + I0, I1) \ + "lh %[" #O0 "], " #I0 "(%[in]) \n\t" \ + "lh %[" #O1 "], " #I1 "(%[in]) \n\t" + +// I0 - location +// I1..I9 - offsets in bytes +#define LOAD_WITH_OFFSET_X4(O0, O1, O2, O3, \ + I0, I1, I2, I3, I4, I5, I6, I7, I8, I9) \ + "ulw %[" #O0 "], " #I1 "+" XSTR(I9) "*" #I5 "(%[" #I0 "]) \n\t" \ + "ulw %[" #O1 "], " #I2 "+" XSTR(I9) "*" #I6 "(%[" #I0 "]) \n\t" \ + "ulw %[" #O2 "], " #I3 "+" XSTR(I9) "*" #I7 "(%[" #I0 "]) \n\t" \ + "ulw %[" #O3 "], " #I4 "+" XSTR(I9) "*" #I8 "(%[" #I0 "]) \n\t" + +// O - output +// IO - input/output +// I - input (macro doesn't change it) +#define MUL_SHIFT_SUM(O0, O1, O2, O3, O4, O5, O6, O7, \ + IO0, IO1, IO2, IO3, \ + I0, I1, I2, I3, I4, I5, I6, I7) \ + "mul %[" #O0 "], %[" #I0 "], %[kC2] \n\t" \ + "mul %[" #O1 "], %[" #I0 "], %[kC1] \n\t" \ + "mul %[" #O2 "], %[" #I1 "], %[kC2] \n\t" \ + "mul %[" #O3 "], %[" #I1 "], %[kC1] \n\t" \ + "mul %[" #O4 "], %[" #I2 "], %[kC2] \n\t" \ + "mul %[" #O5 "], %[" #I2 "], %[kC1] \n\t" \ + "mul %[" #O6 "], %[" #I3 "], %[kC2] \n\t" \ + "mul %[" #O7 "], %[" #I3 "], %[kC1] \n\t" \ + "sra %[" #O0 "], %[" #O0 "], 16 \n\t" \ + "sra %[" #O1 "], %[" #O1 "], 16 \n\t" \ + "sra %[" #O2 "], %[" #O2 "], 16 \n\t" \ + "sra %[" #O3 "], %[" #O3 "], 16 \n\t" \ + "sra %[" #O4 "], %[" #O4 "], 16 \n\t" \ + "sra %[" #O5 "], %[" #O5 "], 16 \n\t" \ + "sra %[" #O6 "], %[" #O6 "], 16 \n\t" \ + "sra %[" #O7 "], %[" #O7 "], 16 \n\t" \ + "addu %[" #IO0 "], %[" #IO0 "], %[" #I4 "] \n\t" \ + "addu %[" #IO1 "], %[" #IO1 "], %[" #I5 "] \n\t" \ + "subu %[" #IO2 "], %[" #IO2 "], %[" #I6 "] \n\t" \ + "subu %[" #IO3 "], %[" #IO3 "], %[" #I7 "] \n\t" + +// O - output +// I - input (macro doesn't change it) +#define INSERT_HALF_X2(O0, O1, \ + I0, I1) \ + "ins %[" #O0 "], %[" #I0 "], 16, 16 \n\t" \ + "ins %[" #O1 "], %[" #I1 "], 16, 16 \n\t" + +// O - output +// I - input (macro doesn't change it) +#define SRA_16(O0, O1, O2, O3, \ + I0, I1, I2, I3) \ + "sra %[" #O0 "], %[" #I0 "], 16 \n\t" \ + "sra %[" #O1 "], %[" #I1 "], 16 \n\t" \ + "sra %[" #O2 "], %[" #I2 "], 16 \n\t" \ + "sra %[" #O3 "], %[" #I3 "], 16 \n\t" + +// temp0[31..16 | 15..0] = temp8[31..16 | 15..0] + temp12[31..16 | 15..0] +// temp1[31..16 | 15..0] = temp8[31..16 | 15..0] - temp12[31..16 | 15..0] +// temp0[31..16 | 15..0] = temp0[31..16 >> 3 | 15..0 >> 3] +// temp1[31..16 | 15..0] = temp1[31..16 >> 3 | 15..0 >> 3] +// O - output +// I - input (macro doesn't change it) +#define SHIFT_R_SUM_X2(O0, O1, O2, O3, O4, O5, O6, O7, \ + I0, I1, I2, I3, I4, I5, I6, I7) \ + "addq.ph %[" #O0 "], %[" #I0 "], %[" #I4 "] \n\t" \ + "subq.ph %[" #O1 "], %[" #I0 "], %[" #I4 "] \n\t" \ + "addq.ph %[" #O2 "], %[" #I1 "], %[" #I5 "] \n\t" \ + "subq.ph %[" #O3 "], %[" #I1 "], %[" #I5 "] \n\t" \ + "addq.ph %[" #O4 "], %[" #I2 "], %[" #I6 "] \n\t" \ + "subq.ph %[" #O5 "], %[" #I2 "], %[" #I6 "] \n\t" \ + "addq.ph %[" #O6 "], %[" #I3 "], %[" #I7 "] \n\t" \ + "subq.ph %[" #O7 "], %[" #I3 "], %[" #I7 "] \n\t" \ + "shra.ph %[" #O0 "], %[" #O0 "], 3 \n\t" \ + "shra.ph %[" #O1 "], %[" #O1 "], 3 \n\t" \ + "shra.ph %[" #O2 "], %[" #O2 "], 3 \n\t" \ + "shra.ph %[" #O3 "], %[" #O3 "], 3 \n\t" \ + "shra.ph %[" #O4 "], %[" #O4 "], 3 \n\t" \ + "shra.ph %[" #O5 "], %[" #O5 "], 3 \n\t" \ + "shra.ph %[" #O6 "], %[" #O6 "], 3 \n\t" \ + "shra.ph %[" #O7 "], %[" #O7 "], 3 \n\t" + +// precrq.ph.w temp0, temp8, temp2 +// temp0 = temp8[31..16] | temp2[31..16] +// ins temp2, temp8, 16, 16 +// temp2 = temp8[31..16] | temp2[15..0] +// O - output +// IO - input/output +// I - input (macro doesn't change it) +#define PACK_2_HALVES_TO_WORD(O0, O1, O2, O3, \ + IO0, IO1, IO2, IO3, \ + I0, I1, I2, I3) \ + "precrq.ph.w %[" #O0 "], %[" #I0 "], %[" #IO0 "] \n\t" \ + "precrq.ph.w %[" #O1 "], %[" #I1 "], %[" #IO1 "] \n\t" \ + "ins %[" #IO0 "], %[" #I0 "], 16, 16 \n\t" \ + "ins %[" #IO1 "], %[" #I1 "], 16, 16 \n\t" \ + "precrq.ph.w %[" #O2 "], %[" #I2 "], %[" #IO2 "] \n\t" \ + "precrq.ph.w %[" #O3 "], %[" #I3 "], %[" #IO3 "] \n\t" \ + "ins %[" #IO2 "], %[" #I2 "], 16, 16 \n\t" \ + "ins %[" #IO3 "], %[" #I3 "], 16, 16 \n\t" + +// preceu.ph.qbr temp0, temp8 +// temp0 = 0 | 0 | temp8[23..16] | temp8[7..0] +// preceu.ph.qbl temp1, temp8 +// temp1 = temp8[23..16] | temp8[7..0] | 0 | 0 +// O - output +// I - input (macro doesn't change it) +#define CONVERT_2_BYTES_TO_HALF(O0, O1, O2, O3, O4, O5, O6, O7, \ + I0, I1, I2, I3) \ + "preceu.ph.qbr %[" #O0 "], %[" #I0 "] \n\t" \ + "preceu.ph.qbl %[" #O1 "], %[" #I0 "] \n\t" \ + "preceu.ph.qbr %[" #O2 "], %[" #I1 "] \n\t" \ + "preceu.ph.qbl %[" #O3 "], %[" #I1 "] \n\t" \ + "preceu.ph.qbr %[" #O4 "], %[" #I2 "] \n\t" \ + "preceu.ph.qbl %[" #O5 "], %[" #I2 "] \n\t" \ + "preceu.ph.qbr %[" #O6 "], %[" #I3 "] \n\t" \ + "preceu.ph.qbl %[" #O7 "], %[" #I3 "] \n\t" + +// temp0[31..16 | 15..0] = temp0[31..16 | 15..0] + temp8[31..16 | 15..0] +// temp0[31..16 | 15..0] = temp0[31..16 <<(s) 7 | 15..0 <<(s) 7] +// temp1..temp7 same as temp0 +// precrqu_s.qb.ph temp0, temp1, temp0: +// temp0 = temp1[31..24] | temp1[15..8] | temp0[31..24] | temp0[15..8] +// store temp0 to dst +// IO - input/output +// I - input (macro doesn't change it) +#define STORE_SAT_SUM_X2(IO0, IO1, IO2, IO3, IO4, IO5, IO6, IO7, \ + I0, I1, I2, I3, I4, I5, I6, I7, \ + I8, I9, I10, I11, I12, I13) \ + "addq.ph %[" #IO0 "], %[" #IO0 "], %[" #I0 "] \n\t" \ + "addq.ph %[" #IO1 "], %[" #IO1 "], %[" #I1 "] \n\t" \ + "addq.ph %[" #IO2 "], %[" #IO2 "], %[" #I2 "] \n\t" \ + "addq.ph %[" #IO3 "], %[" #IO3 "], %[" #I3 "] \n\t" \ + "addq.ph %[" #IO4 "], %[" #IO4 "], %[" #I4 "] \n\t" \ + "addq.ph %[" #IO5 "], %[" #IO5 "], %[" #I5 "] \n\t" \ + "addq.ph %[" #IO6 "], %[" #IO6 "], %[" #I6 "] \n\t" \ + "addq.ph %[" #IO7 "], %[" #IO7 "], %[" #I7 "] \n\t" \ + "shll_s.ph %[" #IO0 "], %[" #IO0 "], 7 \n\t" \ + "shll_s.ph %[" #IO1 "], %[" #IO1 "], 7 \n\t" \ + "shll_s.ph %[" #IO2 "], %[" #IO2 "], 7 \n\t" \ + "shll_s.ph %[" #IO3 "], %[" #IO3 "], 7 \n\t" \ + "shll_s.ph %[" #IO4 "], %[" #IO4 "], 7 \n\t" \ + "shll_s.ph %[" #IO5 "], %[" #IO5 "], 7 \n\t" \ + "shll_s.ph %[" #IO6 "], %[" #IO6 "], 7 \n\t" \ + "shll_s.ph %[" #IO7 "], %[" #IO7 "], 7 \n\t" \ + "precrqu_s.qb.ph %[" #IO0 "], %[" #IO1 "], %[" #IO0 "] \n\t" \ + "precrqu_s.qb.ph %[" #IO2 "], %[" #IO3 "], %[" #IO2 "] \n\t" \ + "precrqu_s.qb.ph %[" #IO4 "], %[" #IO5 "], %[" #IO4 "] \n\t" \ + "precrqu_s.qb.ph %[" #IO6 "], %[" #IO7 "], %[" #IO6 "] \n\t" \ + "usw %[" #IO0 "], " XSTR(I13) "*" #I9 "(%[" #I8 "]) \n\t" \ + "usw %[" #IO2 "], " XSTR(I13) "*" #I10 "(%[" #I8 "]) \n\t" \ + "usw %[" #IO4 "], " XSTR(I13) "*" #I11 "(%[" #I8 "]) \n\t" \ + "usw %[" #IO6 "], " XSTR(I13) "*" #I12 "(%[" #I8 "]) \n\t" + +#define OUTPUT_EARLY_CLOBBER_REGS_10() \ + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), \ + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6), \ + [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), [temp9]"=&r"(temp9), \ + [temp10]"=&r"(temp10) + +#define OUTPUT_EARLY_CLOBBER_REGS_18() \ + OUTPUT_EARLY_CLOBBER_REGS_10(), \ + [temp11]"=&r"(temp11), [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), \ + [temp14]"=&r"(temp14), [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), \ + [temp17]"=&r"(temp17), [temp18]"=&r"(temp18) + +#endif // WEBP_DSP_MIPS_MACRO_H_ diff --git a/drivers/webp/dsp/neon.h b/drivers/webp/dsp/neon.h new file mode 100644 index 0000000000..0a06266848 --- /dev/null +++ b/drivers/webp/dsp/neon.h @@ -0,0 +1,82 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// NEON common code. + +#ifndef WEBP_DSP_NEON_H_ +#define WEBP_DSP_NEON_H_ + +#include <arm_neon.h> + +#include "./dsp.h" + +// Right now, some intrinsics functions seem slower, so we disable them +// everywhere except aarch64 where the inline assembly is incompatible. +#if defined(__aarch64__) +#define WEBP_USE_INTRINSICS // use intrinsics when possible +#endif + +#define INIT_VECTOR2(v, a, b) do { \ + v.val[0] = a; \ + v.val[1] = b; \ +} while (0) + +#define INIT_VECTOR3(v, a, b, c) do { \ + v.val[0] = a; \ + v.val[1] = b; \ + v.val[2] = c; \ +} while (0) + +#define INIT_VECTOR4(v, a, b, c, d) do { \ + v.val[0] = a; \ + v.val[1] = b; \ + v.val[2] = c; \ + v.val[3] = d; \ +} while (0) + +// if using intrinsics, this flag avoids some functions that make gcc-4.6.3 +// crash ("internal compiler error: in immed_double_const, at emit-rtl."). +// (probably similar to gcc.gnu.org/bugzilla/show_bug.cgi?id=48183) +#if !(LOCAL_GCC_PREREQ(4,8) || defined(__aarch64__)) +#define WORK_AROUND_GCC +#endif + +static WEBP_INLINE int32x4x4_t Transpose4x4(const int32x4x4_t rows) { + uint64x2x2_t row01, row23; + + row01.val[0] = vreinterpretq_u64_s32(rows.val[0]); + row01.val[1] = vreinterpretq_u64_s32(rows.val[1]); + row23.val[0] = vreinterpretq_u64_s32(rows.val[2]); + row23.val[1] = vreinterpretq_u64_s32(rows.val[3]); + // Transpose 64-bit values (there's no vswp equivalent) + { + const uint64x1_t row0h = vget_high_u64(row01.val[0]); + const uint64x1_t row2l = vget_low_u64(row23.val[0]); + const uint64x1_t row1h = vget_high_u64(row01.val[1]); + const uint64x1_t row3l = vget_low_u64(row23.val[1]); + row01.val[0] = vcombine_u64(vget_low_u64(row01.val[0]), row2l); + row23.val[0] = vcombine_u64(row0h, vget_high_u64(row23.val[0])); + row01.val[1] = vcombine_u64(vget_low_u64(row01.val[1]), row3l); + row23.val[1] = vcombine_u64(row1h, vget_high_u64(row23.val[1])); + } + { + const int32x4x2_t out01 = vtrnq_s32(vreinterpretq_s32_u64(row01.val[0]), + vreinterpretq_s32_u64(row01.val[1])); + const int32x4x2_t out23 = vtrnq_s32(vreinterpretq_s32_u64(row23.val[0]), + vreinterpretq_s32_u64(row23.val[1])); + int32x4x4_t out; + out.val[0] = out01.val[0]; + out.val[1] = out01.val[1]; + out.val[2] = out23.val[0]; + out.val[3] = out23.val[1]; + return out; + } +} + +#endif // WEBP_DSP_NEON_H_ diff --git a/drivers/webp/dsp/rescaler.c b/drivers/webp/dsp/rescaler.c new file mode 100644 index 0000000000..bc743d5dc5 --- /dev/null +++ b/drivers/webp/dsp/rescaler.c @@ -0,0 +1,238 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Rescaling functions +// +// Author: Skal (pascal.massimino@gmail.com) + +#include <assert.h> + +#include "./dsp.h" +#include "../utils/rescaler.h" + +//------------------------------------------------------------------------------ +// Implementations of critical functions ImportRow / ExportRow + +#define ROUNDER (WEBP_RESCALER_ONE >> 1) +#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX) + +//------------------------------------------------------------------------------ +// Row import + +void WebPRescalerImportRowExpandC(WebPRescaler* const wrk, const uint8_t* src) { + const int x_stride = wrk->num_channels; + const int x_out_max = wrk->dst_width * wrk->num_channels; + int channel; + assert(!WebPRescalerInputDone(wrk)); + assert(wrk->x_expand); + for (channel = 0; channel < x_stride; ++channel) { + int x_in = channel; + int x_out = channel; + // simple bilinear interpolation + int accum = wrk->x_add; + int left = src[x_in]; + int right = (wrk->src_width > 1) ? src[x_in + x_stride] : left; + x_in += x_stride; + while (1) { + wrk->frow[x_out] = right * wrk->x_add + (left - right) * accum; + x_out += x_stride; + if (x_out >= x_out_max) break; + accum -= wrk->x_sub; + if (accum < 0) { + left = right; + x_in += x_stride; + assert(x_in < wrk->src_width * x_stride); + right = src[x_in]; + accum += wrk->x_add; + } + } + assert(wrk->x_sub == 0 /* <- special case for src_width=1 */ || accum == 0); + } +} + +void WebPRescalerImportRowShrinkC(WebPRescaler* const wrk, const uint8_t* src) { + const int x_stride = wrk->num_channels; + const int x_out_max = wrk->dst_width * wrk->num_channels; + int channel; + assert(!WebPRescalerInputDone(wrk)); + assert(!wrk->x_expand); + for (channel = 0; channel < x_stride; ++channel) { + int x_in = channel; + int x_out = channel; + uint32_t sum = 0; + int accum = 0; + while (x_out < x_out_max) { + uint32_t base = 0; + accum += wrk->x_add; + while (accum > 0) { + accum -= wrk->x_sub; + assert(x_in < wrk->src_width * x_stride); + base = src[x_in]; + sum += base; + x_in += x_stride; + } + { // Emit next horizontal pixel. + const rescaler_t frac = base * (-accum); + wrk->frow[x_out] = sum * wrk->x_sub - frac; + // fresh fractional start for next pixel + sum = (int)MULT_FIX(frac, wrk->fx_scale); + } + x_out += x_stride; + } + assert(accum == 0); + } +} + +//------------------------------------------------------------------------------ +// Row export + +void WebPRescalerExportRowExpandC(WebPRescaler* const wrk) { + int x_out; + uint8_t* const dst = wrk->dst; + rescaler_t* const irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* const frow = wrk->frow; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(wrk->y_expand); + assert(wrk->y_sub != 0); + if (wrk->y_accum == 0) { + for (x_out = 0; x_out < x_out_max; ++x_out) { + const uint32_t J = frow[x_out]; + const int v = (int)MULT_FIX(J, wrk->fy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + } + } else { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + for (x_out = 0; x_out < x_out_max; ++x_out) { + const uint64_t I = (uint64_t)A * frow[x_out] + + (uint64_t)B * irow[x_out]; + const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); + const int v = (int)MULT_FIX(J, wrk->fy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + } + } +} + +void WebPRescalerExportRowShrinkC(WebPRescaler* const wrk) { + int x_out; + uint8_t* const dst = wrk->dst; + rescaler_t* const irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* const frow = wrk->frow; + const uint32_t yscale = wrk->fy_scale * (-wrk->y_accum); + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + if (yscale) { + for (x_out = 0; x_out < x_out_max; ++x_out) { + const uint32_t frac = (uint32_t)MULT_FIX(frow[x_out], yscale); + const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + irow[x_out] = frac; // new fractional start + } + } else { + for (x_out = 0; x_out < x_out_max; ++x_out) { + const int v = (int)MULT_FIX(irow[x_out], wrk->fxy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + irow[x_out] = 0; + } + } +} + +#undef MULT_FIX +#undef ROUNDER + +//------------------------------------------------------------------------------ +// Main entry calls + +void WebPRescalerImportRow(WebPRescaler* const wrk, const uint8_t* src) { + assert(!WebPRescalerInputDone(wrk)); + if (!wrk->x_expand) { + WebPRescalerImportRowShrink(wrk, src); + } else { + WebPRescalerImportRowExpand(wrk, src); + } +} + +void WebPRescalerExportRow(WebPRescaler* const wrk) { + if (wrk->y_accum <= 0) { + assert(!WebPRescalerOutputDone(wrk)); + if (wrk->y_expand) { + WebPRescalerExportRowExpand(wrk); + } else if (wrk->fxy_scale) { + WebPRescalerExportRowShrink(wrk); + } else { // very special case for src = dst = 1x1 + int i; + assert(wrk->src_width == 1 && wrk->dst_width <= 2); + assert(wrk->src_height == 1 && wrk->dst_height == 1); + for (i = 0; i < wrk->num_channels * wrk->dst_width; ++i) { + wrk->dst[i] = wrk->irow[i]; + wrk->irow[i] = 0; + } + } + wrk->y_accum += wrk->y_add; + wrk->dst += wrk->dst_stride; + ++wrk->dst_y; + } +} + +//------------------------------------------------------------------------------ + +WebPRescalerImportRowFunc WebPRescalerImportRowExpand; +WebPRescalerImportRowFunc WebPRescalerImportRowShrink; + +WebPRescalerExportRowFunc WebPRescalerExportRowExpand; +WebPRescalerExportRowFunc WebPRescalerExportRowShrink; + +extern void WebPRescalerDspInitSSE2(void); +extern void WebPRescalerDspInitMIPS32(void); +extern void WebPRescalerDspInitMIPSdspR2(void); +extern void WebPRescalerDspInitNEON(void); + +static volatile VP8CPUInfo rescaler_last_cpuinfo_used = + (VP8CPUInfo)&rescaler_last_cpuinfo_used; + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInit(void) { + if (rescaler_last_cpuinfo_used == VP8GetCPUInfo) return; + + WebPRescalerImportRowExpand = WebPRescalerImportRowExpandC; + WebPRescalerImportRowShrink = WebPRescalerImportRowShrinkC; + WebPRescalerExportRowExpand = WebPRescalerExportRowExpandC; + WebPRescalerExportRowShrink = WebPRescalerExportRowShrinkC; + + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + WebPRescalerDspInitSSE2(); + } +#endif +#if defined(WEBP_USE_NEON) + if (VP8GetCPUInfo(kNEON)) { + WebPRescalerDspInitNEON(); + } +#endif +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + WebPRescalerDspInitMIPS32(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + WebPRescalerDspInitMIPSdspR2(); + } +#endif + } + rescaler_last_cpuinfo_used = VP8GetCPUInfo; +} diff --git a/drivers/webp/dsp/rescaler_mips32.c b/drivers/webp/dsp/rescaler_mips32.c new file mode 100644 index 0000000000..ddaa391336 --- /dev/null +++ b/drivers/webp/dsp/rescaler_mips32.c @@ -0,0 +1,291 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of rescaling functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include <assert.h> +#include "../utils/rescaler.h" + +//------------------------------------------------------------------------------ +// Row import + +static void ImportRowShrink(WebPRescaler* const wrk, const uint8_t* src) { + const int x_stride = wrk->num_channels; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const int fx_scale = wrk->fx_scale; + const int x_add = wrk->x_add; + const int x_sub = wrk->x_sub; + const int x_stride1 = x_stride << 2; + int channel; + assert(!wrk->x_expand); + assert(!WebPRescalerInputDone(wrk)); + + for (channel = 0; channel < x_stride; ++channel) { + const uint8_t* src1 = src + channel; + rescaler_t* frow = wrk->frow + channel; + int temp1, temp2, temp3; + int base, frac, sum; + int accum, accum1; + int loop_c = x_out_max - channel; + + __asm__ volatile ( + "li %[temp1], 0x8000 \n\t" + "li %[temp2], 0x10000 \n\t" + "li %[sum], 0 \n\t" + "li %[accum], 0 \n\t" + "1: \n\t" + "addu %[accum], %[accum], %[x_add] \n\t" + "li %[base], 0 \n\t" + "blez %[accum], 3f \n\t" + "2: \n\t" + "lbu %[base], 0(%[src1]) \n\t" + "subu %[accum], %[accum], %[x_sub] \n\t" + "addu %[src1], %[src1], %[x_stride] \n\t" + "addu %[sum], %[sum], %[base] \n\t" + "bgtz %[accum], 2b \n\t" + "3: \n\t" + "negu %[accum1], %[accum] \n\t" + "mul %[frac], %[base], %[accum1] \n\t" + "mul %[temp3], %[sum], %[x_sub] \n\t" + "subu %[loop_c], %[loop_c], %[x_stride] \n\t" + "mult %[temp1], %[temp2] \n\t" + "maddu %[frac], %[fx_scale] \n\t" + "mfhi %[sum] \n\t" + "subu %[temp3], %[temp3], %[frac] \n\t" + "sw %[temp3], 0(%[frow]) \n\t" + "addu %[frow], %[frow], %[x_stride1] \n\t" + "bgtz %[loop_c], 1b \n\t" + : [accum]"=&r"(accum), [src1]"+r"(src1), [temp3]"=&r"(temp3), + [sum]"=&r"(sum), [base]"=&r"(base), [frac]"=&r"(frac), + [frow]"+r"(frow), [accum1]"=&r"(accum1), + [temp2]"=&r"(temp2), [temp1]"=&r"(temp1) + : [x_stride]"r"(x_stride), [fx_scale]"r"(fx_scale), + [x_sub]"r"(x_sub), [x_add]"r"(x_add), + [loop_c]"r"(loop_c), [x_stride1]"r"(x_stride1) + : "memory", "hi", "lo" + ); + assert(accum == 0); + } +} + +static void ImportRowExpand(WebPRescaler* const wrk, const uint8_t* src) { + const int x_stride = wrk->num_channels; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const int x_add = wrk->x_add; + const int x_sub = wrk->x_sub; + const int src_width = wrk->src_width; + const int x_stride1 = x_stride << 2; + int channel; + assert(wrk->x_expand); + assert(!WebPRescalerInputDone(wrk)); + + for (channel = 0; channel < x_stride; ++channel) { + const uint8_t* src1 = src + channel; + rescaler_t* frow = wrk->frow + channel; + int temp1, temp2, temp3, temp4; + int frac; + int accum; + int x_out = channel; + + __asm__ volatile ( + "addiu %[temp3], %[src_width], -1 \n\t" + "lbu %[temp2], 0(%[src1]) \n\t" + "addu %[src1], %[src1], %[x_stride] \n\t" + "bgtz %[temp3], 0f \n\t" + "addiu %[temp1], %[temp2], 0 \n\t" + "b 3f \n\t" + "0: \n\t" + "lbu %[temp1], 0(%[src1]) \n\t" + "3: \n\t" + "addiu %[accum], %[x_add], 0 \n\t" + "1: \n\t" + "subu %[temp3], %[temp2], %[temp1] \n\t" + "mul %[temp3], %[temp3], %[accum] \n\t" + "mul %[temp4], %[temp1], %[x_add] \n\t" + "addu %[temp3], %[temp4], %[temp3] \n\t" + "sw %[temp3], 0(%[frow]) \n\t" + "addu %[frow], %[frow], %[x_stride1] \n\t" + "addu %[x_out], %[x_out], %[x_stride] \n\t" + "subu %[temp3], %[x_out], %[x_out_max] \n\t" + "bgez %[temp3], 2f \n\t" + "subu %[accum], %[accum], %[x_sub] \n\t" + "bgez %[accum], 4f \n\t" + "addiu %[temp2], %[temp1], 0 \n\t" + "addu %[src1], %[src1], %[x_stride] \n\t" + "lbu %[temp1], 0(%[src1]) \n\t" + "addu %[accum], %[accum], %[x_add] \n\t" + "4: \n\t" + "b 1b \n\t" + "2: \n\t" + : [src1]"+r"(src1), [accum]"=&r"(accum), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), + [x_out]"+r"(x_out), [frac]"=&r"(frac), [frow]"+r"(frow) + : [x_stride]"r"(x_stride), [x_add]"r"(x_add), [x_sub]"r"(x_sub), + [x_stride1]"r"(x_stride1), [src_width]"r"(src_width), + [x_out_max]"r"(x_out_max) + : "memory", "hi", "lo" + ); + assert(wrk->x_sub == 0 /* <- special case for src_width=1 */ || accum == 0); + } +} + +//------------------------------------------------------------------------------ +// Row export + +static void ExportRowExpand(WebPRescaler* const wrk) { + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* frow = wrk->frow; + int temp0, temp1, temp3, temp4, temp5, loop_end; + const int temp2 = (int)wrk->fy_scale; + const int temp6 = x_out_max << 2; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(wrk->y_expand); + assert(wrk->y_sub != 0); + if (wrk->y_accum == 0) { + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "addiu %[dst], %[dst], 1 \n\t" + "addiu %[frow], %[frow], 4 \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[temp0], %[temp2] \n\t" + "mfhi %[temp5] \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [dst]"+r"(dst), [loop_end]"=&r"(loop_end) + : [temp2]"r"(temp2), [temp6]"r"(temp6) + : "memory", "hi", "lo" + ); + } else { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "lw %[temp1], 0(%[irow]) \n\t" + "addiu %[dst], %[dst], 1 \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[A], %[temp0] \n\t" + "maddu %[B], %[temp1] \n\t" + "addiu %[frow], %[frow], 4 \n\t" + "addiu %[irow], %[irow], 4 \n\t" + "mfhi %[temp5] \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[temp5], %[temp2] \n\t" + "mfhi %[temp5] \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [irow]"+r"(irow), [dst]"+r"(dst), [loop_end]"=&r"(loop_end) + : [temp2]"r"(temp2), [temp6]"r"(temp6), [A]"r"(A), [B]"r"(B) + : "memory", "hi", "lo" + ); + } +} + +static void ExportRowShrink(WebPRescaler* const wrk) { + const int x_out_max = wrk->dst_width * wrk->num_channels; + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const rescaler_t* frow = wrk->frow; + const int yscale = wrk->fy_scale * (-wrk->y_accum); + int temp0, temp1, temp3, temp4, temp5, loop_end; + const int temp2 = (int)wrk->fxy_scale; + const int temp6 = x_out_max << 2; + + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + assert(wrk->fxy_scale != 0); + if (yscale) { + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "mult %[temp3], %[temp4] \n\t" + "addiu %[frow], %[frow], 4 \n\t" + "maddu %[temp0], %[yscale] \n\t" + "mfhi %[temp1] \n\t" + "lw %[temp0], 0(%[irow]) \n\t" + "addiu %[dst], %[dst], 1 \n\t" + "addiu %[irow], %[irow], 4 \n\t" + "subu %[temp0], %[temp0], %[temp1] \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[temp0], %[temp2] \n\t" + "mfhi %[temp5] \n\t" + "sw %[temp1], -4(%[irow]) \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [irow]"+r"(irow), [dst]"+r"(dst), [loop_end]"=&r"(loop_end) + : [temp2]"r"(temp2), [yscale]"r"(yscale), [temp6]"r"(temp6) + : "memory", "hi", "lo" + ); + } else { + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[irow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[irow]) \n\t" + "addiu %[dst], %[dst], 1 \n\t" + "addiu %[irow], %[irow], 4 \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[temp0], %[temp2] \n\t" + "mfhi %[temp5] \n\t" + "sw $zero, -4(%[irow]) \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[irow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [irow]"+r"(irow), + [dst]"+r"(dst), [loop_end]"=&r"(loop_end) + : [temp2]"r"(temp2), [temp6]"r"(temp6) + : "memory", "hi", "lo" + ); + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPRescalerDspInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitMIPS32(void) { + WebPRescalerImportRowExpand = ImportRowExpand; + WebPRescalerImportRowShrink = ImportRowShrink; + WebPRescalerExportRowExpand = ExportRowExpand; + WebPRescalerExportRowShrink = ExportRowShrink; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(WebPRescalerDspInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/drivers/webp/dsp/rescaler_mips_dsp_r2.c b/drivers/webp/dsp/rescaler_mips_dsp_r2.c new file mode 100644 index 0000000000..b457d0a30a --- /dev/null +++ b/drivers/webp/dsp/rescaler_mips_dsp_r2.c @@ -0,0 +1,314 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of rescaling functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include <assert.h> +#include "../utils/rescaler.h" + +#define ROUNDER (WEBP_RESCALER_ONE >> 1) +#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX) + +//------------------------------------------------------------------------------ +// Row export + +static void ExportRowShrink(WebPRescaler* const wrk) { + int i; + const int x_out_max = wrk->dst_width * wrk->num_channels; + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const rescaler_t* frow = wrk->frow; + const int yscale = wrk->fy_scale * (-wrk->y_accum); + int temp0, temp1, temp2, temp3, temp4, temp5, loop_end; + const int temp7 = (int)wrk->fxy_scale; + const int temp6 = (x_out_max & ~0x3) << 2; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + assert(wrk->fxy_scale != 0); + if (yscale) { + if (x_out_max >= 4) { + int temp8, temp9, temp10, temp11; + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "lw %[temp1], 4(%[frow]) \n\t" + "lw %[temp2], 8(%[frow]) \n\t" + "lw %[temp5], 12(%[frow]) \n\t" + "mult $ac0, %[temp3], %[temp4] \n\t" + "maddu $ac0, %[temp0], %[yscale] \n\t" + "mult $ac1, %[temp3], %[temp4] \n\t" + "maddu $ac1, %[temp1], %[yscale] \n\t" + "mult $ac2, %[temp3], %[temp4] \n\t" + "maddu $ac2, %[temp2], %[yscale] \n\t" + "mult $ac3, %[temp3], %[temp4] \n\t" + "maddu $ac3, %[temp5], %[yscale] \n\t" + "addiu %[frow], %[frow], 16 \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp5], $ac3 \n\t" + "lw %[temp8], 0(%[irow]) \n\t" + "lw %[temp9], 4(%[irow]) \n\t" + "lw %[temp10], 8(%[irow]) \n\t" + "lw %[temp11], 12(%[irow]) \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "addiu %[irow], %[irow], 16 \n\t" + "subu %[temp8], %[temp8], %[temp0] \n\t" + "subu %[temp9], %[temp9], %[temp1] \n\t" + "subu %[temp10], %[temp10], %[temp2] \n\t" + "subu %[temp11], %[temp11], %[temp5] \n\t" + "mult $ac0, %[temp3], %[temp4] \n\t" + "maddu $ac0, %[temp8], %[temp7] \n\t" + "mult $ac1, %[temp3], %[temp4] \n\t" + "maddu $ac1, %[temp9], %[temp7] \n\t" + "mult $ac2, %[temp3], %[temp4] \n\t" + "maddu $ac2, %[temp10], %[temp7] \n\t" + "mult $ac3, %[temp3], %[temp4] \n\t" + "maddu $ac3, %[temp11], %[temp7] \n\t" + "mfhi %[temp8], $ac0 \n\t" + "mfhi %[temp9], $ac1 \n\t" + "mfhi %[temp10], $ac2 \n\t" + "mfhi %[temp11], $ac3 \n\t" + "sw %[temp0], -16(%[irow]) \n\t" + "sw %[temp1], -12(%[irow]) \n\t" + "sw %[temp2], -8(%[irow]) \n\t" + "sw %[temp5], -4(%[irow]) \n\t" + "sb %[temp8], -4(%[dst]) \n\t" + "sb %[temp9], -3(%[dst]) \n\t" + "sb %[temp10], -2(%[dst]) \n\t" + "sb %[temp11], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [irow]"+r"(irow), [dst]"+r"(dst), [loop_end]"=&r"(loop_end), + [temp8]"=&r"(temp8), [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), + [temp11]"=&r"(temp11), [temp2]"=&r"(temp2) + : [temp7]"r"(temp7), [yscale]"r"(yscale), [temp6]"r"(temp6) + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + } + for (i = 0; i < (x_out_max & 0x3); ++i) { + const uint32_t frac = (uint32_t)MULT_FIX(*frow++, yscale); + const int v = (int)MULT_FIX(*irow - frac, wrk->fxy_scale); + assert(v >= 0 && v <= 255); + *dst++ = v; + *irow++ = frac; // new fractional start + } + } else { + if (x_out_max >= 4) { + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[irow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[irow]) \n\t" + "lw %[temp1], 4(%[irow]) \n\t" + "lw %[temp2], 8(%[irow]) \n\t" + "lw %[temp5], 12(%[irow]) \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "addiu %[irow], %[irow], 16 \n\t" + "mult $ac0, %[temp3], %[temp4] \n\t" + "maddu $ac0, %[temp0], %[temp7] \n\t" + "mult $ac1, %[temp3], %[temp4] \n\t" + "maddu $ac1, %[temp1], %[temp7] \n\t" + "mult $ac2, %[temp3], %[temp4] \n\t" + "maddu $ac2, %[temp2], %[temp7] \n\t" + "mult $ac3, %[temp3], %[temp4] \n\t" + "maddu $ac3, %[temp5], %[temp7] \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp5], $ac3 \n\t" + "sw $zero, -16(%[irow]) \n\t" + "sw $zero, -12(%[irow]) \n\t" + "sw $zero, -8(%[irow]) \n\t" + "sw $zero, -4(%[irow]) \n\t" + "sb %[temp0], -4(%[dst]) \n\t" + "sb %[temp1], -3(%[dst]) \n\t" + "sb %[temp2], -2(%[dst]) \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[irow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [irow]"+r"(irow), + [dst]"+r"(dst), [loop_end]"=&r"(loop_end), [temp2]"=&r"(temp2) + : [temp7]"r"(temp7), [temp6]"r"(temp6) + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + } + for (i = 0; i < (x_out_max & 0x3); ++i) { + const int v = (int)MULT_FIX(*irow, wrk->fxy_scale); + assert(v >= 0 && v <= 255); + *dst++ = v; + *irow++ = 0; + } + } +} + +static void ExportRowExpand(WebPRescaler* const wrk) { + int i; + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* frow = wrk->frow; + int temp0, temp1, temp2, temp3, temp4, temp5, loop_end; + const int temp6 = (x_out_max & ~0x3) << 2; + const int temp7 = (int)wrk->fy_scale; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(wrk->y_expand); + assert(wrk->y_sub != 0); + if (wrk->y_accum == 0) { + if (x_out_max >= 4) { + __asm__ volatile ( + "li %[temp4], 0x10000 \n\t" + "li %[temp5], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "lw %[temp1], 4(%[frow]) \n\t" + "lw %[temp2], 8(%[frow]) \n\t" + "lw %[temp3], 12(%[frow]) \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "addiu %[frow], %[frow], 16 \n\t" + "mult $ac0, %[temp4], %[temp5] \n\t" + "maddu $ac0, %[temp0], %[temp7] \n\t" + "mult $ac1, %[temp4], %[temp5] \n\t" + "maddu $ac1, %[temp1], %[temp7] \n\t" + "mult $ac2, %[temp4], %[temp5] \n\t" + "maddu $ac2, %[temp2], %[temp7] \n\t" + "mult $ac3, %[temp4], %[temp5] \n\t" + "maddu $ac3, %[temp3], %[temp7] \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp3], $ac3 \n\t" + "sb %[temp0], -4(%[dst]) \n\t" + "sb %[temp1], -3(%[dst]) \n\t" + "sb %[temp2], -2(%[dst]) \n\t" + "sb %[temp3], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [dst]"+r"(dst), [loop_end]"=&r"(loop_end), [temp2]"=&r"(temp2) + : [temp7]"r"(temp7), [temp6]"r"(temp6) + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + } + for (i = 0; i < (x_out_max & 0x3); ++i) { + const uint32_t J = *frow++; + const int v = (int)MULT_FIX(J, wrk->fy_scale); + assert(v >= 0 && v <= 255); + *dst++ = v; + } + } else { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + if (x_out_max >= 4) { + int temp8, temp9, temp10, temp11; + __asm__ volatile ( + "li %[temp8], 0x10000 \n\t" + "li %[temp9], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "lw %[temp1], 4(%[frow]) \n\t" + "lw %[temp2], 8(%[frow]) \n\t" + "lw %[temp3], 12(%[frow]) \n\t" + "lw %[temp4], 0(%[irow]) \n\t" + "lw %[temp5], 4(%[irow]) \n\t" + "lw %[temp10], 8(%[irow]) \n\t" + "lw %[temp11], 12(%[irow]) \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "mult $ac0, %[temp8], %[temp9] \n\t" + "maddu $ac0, %[A], %[temp0] \n\t" + "maddu $ac0, %[B], %[temp4] \n\t" + "mult $ac1, %[temp8], %[temp9] \n\t" + "maddu $ac1, %[A], %[temp1] \n\t" + "maddu $ac1, %[B], %[temp5] \n\t" + "mult $ac2, %[temp8], %[temp9] \n\t" + "maddu $ac2, %[A], %[temp2] \n\t" + "maddu $ac2, %[B], %[temp10] \n\t" + "mult $ac3, %[temp8], %[temp9] \n\t" + "maddu $ac3, %[A], %[temp3] \n\t" + "maddu $ac3, %[B], %[temp11] \n\t" + "addiu %[frow], %[frow], 16 \n\t" + "addiu %[irow], %[irow], 16 \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp3], $ac3 \n\t" + "mult $ac0, %[temp8], %[temp9] \n\t" + "maddu $ac0, %[temp0], %[temp7] \n\t" + "mult $ac1, %[temp8], %[temp9] \n\t" + "maddu $ac1, %[temp1], %[temp7] \n\t" + "mult $ac2, %[temp8], %[temp9] \n\t" + "maddu $ac2, %[temp2], %[temp7] \n\t" + "mult $ac3, %[temp8], %[temp9] \n\t" + "maddu $ac3, %[temp3], %[temp7] \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp3], $ac3 \n\t" + "sb %[temp0], -4(%[dst]) \n\t" + "sb %[temp1], -3(%[dst]) \n\t" + "sb %[temp2], -2(%[dst]) \n\t" + "sb %[temp3], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [irow]"+r"(irow), [dst]"+r"(dst), [loop_end]"=&r"(loop_end), + [temp8]"=&r"(temp8), [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), + [temp11]"=&r"(temp11), [temp2]"=&r"(temp2) + : [temp7]"r"(temp7), [temp6]"r"(temp6), [A]"r"(A), [B]"r"(B) + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + } + for (i = 0; i < (x_out_max & 0x3); ++i) { + const uint64_t I = (uint64_t)A * *frow++ + + (uint64_t)B * *irow++; + const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); + const int v = (int)MULT_FIX(J, wrk->fy_scale); + assert(v >= 0 && v <= 255); + *dst++ = v; + } + } +} + +#undef MULT_FIX +#undef ROUNDER + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPRescalerDspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitMIPSdspR2(void) { + WebPRescalerExportRowExpand = ExportRowExpand; + WebPRescalerExportRowShrink = ExportRowShrink; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(WebPRescalerDspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/rescaler_neon.c b/drivers/webp/dsp/rescaler_neon.c new file mode 100644 index 0000000000..16fd450ea3 --- /dev/null +++ b/drivers/webp/dsp/rescaler_neon.c @@ -0,0 +1,186 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// NEON version of rescaling functions +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_NEON) + +#include <arm_neon.h> +#include <assert.h> +#include "./neon.h" +#include "../utils/rescaler.h" + +#define ROUNDER (WEBP_RESCALER_ONE >> 1) +#define MULT_FIX_C(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX) + +#define LOAD_32x4(SRC, DST) const uint32x4_t DST = vld1q_u32((SRC)) +#define LOAD_32x8(SRC, DST0, DST1) \ + LOAD_32x4(SRC + 0, DST0); \ + LOAD_32x4(SRC + 4, DST1) + +#define STORE_32x8(SRC0, SRC1, DST) do { \ + vst1q_u32((DST) + 0, SRC0); \ + vst1q_u32((DST) + 4, SRC1); \ +} while (0); + +#if (WEBP_RESCALER_RFIX == 32) +#define MAKE_HALF_CST(C) vdupq_n_s32((int32_t)((C) >> 1)) +#define MULT_FIX(A, B) /* note: B is actualy scale>>1. See MAKE_HALF_CST */ \ + vreinterpretq_u32_s32(vqrdmulhq_s32(vreinterpretq_s32_u32((A)), (B))) +#else +#error "MULT_FIX/WEBP_RESCALER_RFIX need some more work" +#endif + +static uint32x4_t Interpolate(const rescaler_t* const frow, + const rescaler_t* const irow, + uint32_t A, uint32_t B) { + LOAD_32x4(frow, A0); + LOAD_32x4(irow, B0); + const uint64x2_t C0 = vmull_n_u32(vget_low_u32(A0), A); + const uint64x2_t C1 = vmull_n_u32(vget_high_u32(A0), A); + const uint64x2_t D0 = vmlal_n_u32(C0, vget_low_u32(B0), B); + const uint64x2_t D1 = vmlal_n_u32(C1, vget_high_u32(B0), B); + const uint32x4_t E = vcombine_u32( + vrshrn_n_u64(D0, WEBP_RESCALER_RFIX), + vrshrn_n_u64(D1, WEBP_RESCALER_RFIX)); + return E; +} + +static void RescalerExportRowExpand(WebPRescaler* const wrk) { + int x_out; + uint8_t* const dst = wrk->dst; + rescaler_t* const irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const int max_span = x_out_max & ~7; + const rescaler_t* const frow = wrk->frow; + const uint32_t fy_scale = wrk->fy_scale; + const int32x4_t fy_scale_half = MAKE_HALF_CST(fy_scale); + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(wrk->y_expand); + assert(wrk->y_sub != 0); + if (wrk->y_accum == 0) { + for (x_out = 0; x_out < max_span; x_out += 8) { + LOAD_32x4(frow + x_out + 0, A0); + LOAD_32x4(frow + x_out + 4, A1); + const uint32x4_t B0 = MULT_FIX(A0, fy_scale_half); + const uint32x4_t B1 = MULT_FIX(A1, fy_scale_half); + const uint16x4_t C0 = vmovn_u32(B0); + const uint16x4_t C1 = vmovn_u32(B1); + const uint8x8_t D = vmovn_u16(vcombine_u16(C0, C1)); + vst1_u8(dst + x_out, D); + } + for (; x_out < x_out_max; ++x_out) { + const uint32_t J = frow[x_out]; + const int v = (int)MULT_FIX_C(J, fy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + } + } else { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + for (x_out = 0; x_out < max_span; x_out += 8) { + const uint32x4_t C0 = + Interpolate(frow + x_out + 0, irow + x_out + 0, A, B); + const uint32x4_t C1 = + Interpolate(frow + x_out + 4, irow + x_out + 4, A, B); + const uint32x4_t D0 = MULT_FIX(C0, fy_scale_half); + const uint32x4_t D1 = MULT_FIX(C1, fy_scale_half); + const uint16x4_t E0 = vmovn_u32(D0); + const uint16x4_t E1 = vmovn_u32(D1); + const uint8x8_t F = vmovn_u16(vcombine_u16(E0, E1)); + vst1_u8(dst + x_out, F); + } + for (; x_out < x_out_max; ++x_out) { + const uint64_t I = (uint64_t)A * frow[x_out] + + (uint64_t)B * irow[x_out]; + const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); + const int v = (int)MULT_FIX_C(J, fy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + } + } +} + +static void RescalerExportRowShrink(WebPRescaler* const wrk) { + int x_out; + uint8_t* const dst = wrk->dst; + rescaler_t* const irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const int max_span = x_out_max & ~7; + const rescaler_t* const frow = wrk->frow; + const uint32_t yscale = wrk->fy_scale * (-wrk->y_accum); + const uint32_t fxy_scale = wrk->fxy_scale; + const uint32x4_t zero = vdupq_n_u32(0); + const int32x4_t yscale_half = MAKE_HALF_CST(yscale); + const int32x4_t fxy_scale_half = MAKE_HALF_CST(fxy_scale); + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + if (yscale) { + for (x_out = 0; x_out < max_span; x_out += 8) { + LOAD_32x8(frow + x_out, in0, in1); + LOAD_32x8(irow + x_out, in2, in3); + const uint32x4_t A0 = MULT_FIX(in0, yscale_half); + const uint32x4_t A1 = MULT_FIX(in1, yscale_half); + const uint32x4_t B0 = vqsubq_u32(in2, A0); + const uint32x4_t B1 = vqsubq_u32(in3, A1); + const uint32x4_t C0 = MULT_FIX(B0, fxy_scale_half); + const uint32x4_t C1 = MULT_FIX(B1, fxy_scale_half); + const uint16x4_t D0 = vmovn_u32(C0); + const uint16x4_t D1 = vmovn_u32(C1); + const uint8x8_t E = vmovn_u16(vcombine_u16(D0, D1)); + vst1_u8(dst + x_out, E); + STORE_32x8(A0, A1, irow + x_out); + } + for (; x_out < x_out_max; ++x_out) { + const uint32_t frac = (uint32_t)MULT_FIX_C(frow[x_out], yscale); + const int v = (int)MULT_FIX_C(irow[x_out] - frac, wrk->fxy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + irow[x_out] = frac; // new fractional start + } + } else { + for (x_out = 0; x_out < max_span; x_out += 8) { + LOAD_32x8(irow + x_out, in0, in1); + const uint32x4_t A0 = MULT_FIX(in0, fxy_scale_half); + const uint32x4_t A1 = MULT_FIX(in1, fxy_scale_half); + const uint16x4_t B0 = vmovn_u32(A0); + const uint16x4_t B1 = vmovn_u32(A1); + const uint8x8_t C = vmovn_u16(vcombine_u16(B0, B1)); + vst1_u8(dst + x_out, C); + STORE_32x8(zero, zero, irow + x_out); + } + for (; x_out < x_out_max; ++x_out) { + const int v = (int)MULT_FIX_C(irow[x_out], fxy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + irow[x_out] = 0; + } + } +} + +//------------------------------------------------------------------------------ + +extern void WebPRescalerDspInitNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitNEON(void) { + WebPRescalerExportRowExpand = RescalerExportRowExpand; + WebPRescalerExportRowShrink = RescalerExportRowShrink; +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(WebPRescalerDspInitNEON) + +#endif // WEBP_USE_NEON diff --git a/drivers/webp/dsp/rescaler_sse2.c b/drivers/webp/dsp/rescaler_sse2.c new file mode 100644 index 0000000000..186edb653e --- /dev/null +++ b/drivers/webp/dsp/rescaler_sse2.c @@ -0,0 +1,373 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 Rescaling functions +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE2) +#include <emmintrin.h> + +#include <assert.h> +#include "../utils/rescaler.h" + +//------------------------------------------------------------------------------ +// Implementations of critical functions ImportRow / ExportRow + +#define ROUNDER (WEBP_RESCALER_ONE >> 1) +#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX) + +// input: 8 bytes ABCDEFGH -> output: A0E0B0F0C0G0D0H0 +static void LoadTwoPixels(const uint8_t* const src, __m128i* out) { + const __m128i zero = _mm_setzero_si128(); + const __m128i A = _mm_loadl_epi64((const __m128i*)(src)); // ABCDEFGH + const __m128i B = _mm_unpacklo_epi8(A, zero); // A0B0C0D0E0F0G0H0 + const __m128i C = _mm_srli_si128(B, 8); // E0F0G0H0 + *out = _mm_unpacklo_epi16(B, C); +} + +// input: 8 bytes ABCDEFGH -> output: A0B0C0D0E0F0G0H0 +static void LoadHeightPixels(const uint8_t* const src, __m128i* out) { + const __m128i zero = _mm_setzero_si128(); + const __m128i A = _mm_loadl_epi64((const __m128i*)(src)); // ABCDEFGH + *out = _mm_unpacklo_epi8(A, zero); +} + +static void RescalerImportRowExpandSSE2(WebPRescaler* const wrk, + const uint8_t* src) { + rescaler_t* frow = wrk->frow; + const rescaler_t* const frow_end = frow + wrk->dst_width * wrk->num_channels; + const int x_add = wrk->x_add; + int accum = x_add; + __m128i cur_pixels; + + assert(!WebPRescalerInputDone(wrk)); + assert(wrk->x_expand); + if (wrk->num_channels == 4) { + if (wrk->src_width < 2) { + WebPRescalerImportRowExpandC(wrk, src); + return; + } + LoadTwoPixels(src, &cur_pixels); + src += 4; + while (1) { + const __m128i mult = _mm_set1_epi32(((x_add - accum) << 16) | accum); + const __m128i out = _mm_madd_epi16(cur_pixels, mult); + _mm_storeu_si128((__m128i*)frow, out); + frow += 4; + if (frow >= frow_end) break; + accum -= wrk->x_sub; + if (accum < 0) { + LoadTwoPixels(src, &cur_pixels); + src += 4; + accum += x_add; + } + } + } else { + int left; + const uint8_t* const src_limit = src + wrk->src_width - 8; + if (wrk->src_width < 8) { + WebPRescalerImportRowExpandC(wrk, src); + return; + } + LoadHeightPixels(src, &cur_pixels); + src += 7; + left = 7; + while (1) { + const __m128i mult = _mm_cvtsi32_si128(((x_add - accum) << 16) | accum); + const __m128i out = _mm_madd_epi16(cur_pixels, mult); + *(uint32_t*)frow = _mm_cvtsi128_si32(out); + frow += 1; + if (frow >= frow_end) break; + accum -= wrk->x_sub; + if (accum < 0) { + if (--left) { + cur_pixels = _mm_srli_si128(cur_pixels, 2); + } else if (src <= src_limit) { + LoadHeightPixels(src, &cur_pixels); + src += 7; + left = 7; + } else { // tail + cur_pixels = _mm_srli_si128(cur_pixels, 2); + cur_pixels = _mm_insert_epi16(cur_pixels, src[1], 1); + src += 1; + left = 1; + } + accum += x_add; + } + } + } + assert(accum == 0); +} + +static void RescalerImportRowShrinkSSE2(WebPRescaler* const wrk, + const uint8_t* src) { + const int x_sub = wrk->x_sub; + int accum = 0; + const __m128i zero = _mm_setzero_si128(); + const __m128i mult0 = _mm_set1_epi16(x_sub); + const __m128i mult1 = _mm_set1_epi32(wrk->fx_scale); + const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER); + __m128i sum = zero; + rescaler_t* frow = wrk->frow; + const rescaler_t* const frow_end = wrk->frow + 4 * wrk->dst_width; + + if (wrk->num_channels != 4 || wrk->x_add > (x_sub << 7)) { + WebPRescalerImportRowShrinkC(wrk, src); + return; + } + assert(!WebPRescalerInputDone(wrk)); + assert(!wrk->x_expand); + + for (; frow < frow_end; frow += 4) { + __m128i base = zero; + accum += wrk->x_add; + while (accum > 0) { + const __m128i A = _mm_cvtsi32_si128(*(int*)src); + src += 4; + base = _mm_unpacklo_epi8(A, zero); + // To avoid overflow, we need: base * x_add / x_sub < 32768 + // => x_add < x_sub << 7. That's a 1/128 reduction ratio limit. + sum = _mm_add_epi16(sum, base); + accum -= x_sub; + } + { // Emit next horizontal pixel. + const __m128i mult = _mm_set1_epi16(-accum); + const __m128i frac0 = _mm_mullo_epi16(base, mult); // 16b x 16b -> 32b + const __m128i frac1 = _mm_mulhi_epu16(base, mult); + const __m128i frac = _mm_unpacklo_epi16(frac0, frac1); // frac is 32b + const __m128i A0 = _mm_mullo_epi16(sum, mult0); + const __m128i A1 = _mm_mulhi_epu16(sum, mult0); + const __m128i B0 = _mm_unpacklo_epi16(A0, A1); // sum * x_sub + const __m128i frow_out = _mm_sub_epi32(B0, frac); // sum * x_sub - frac + const __m128i D0 = _mm_srli_epi64(frac, 32); + const __m128i D1 = _mm_mul_epu32(frac, mult1); // 32b x 16b -> 64b + const __m128i D2 = _mm_mul_epu32(D0, mult1); + const __m128i E1 = _mm_add_epi64(D1, rounder); + const __m128i E2 = _mm_add_epi64(D2, rounder); + const __m128i F1 = _mm_shuffle_epi32(E1, 1 | (3 << 2)); + const __m128i F2 = _mm_shuffle_epi32(E2, 1 | (3 << 2)); + const __m128i G = _mm_unpacklo_epi32(F1, F2); + sum = _mm_packs_epi32(G, zero); + _mm_storeu_si128((__m128i*)frow, frow_out); + } + } + assert(accum == 0); +} + +//------------------------------------------------------------------------------ +// Row export + +// load *src as epi64, multiply by mult and store result in [out0 ... out3] +static WEBP_INLINE void LoadDispatchAndMult(const rescaler_t* const src, + const __m128i* const mult, + __m128i* const out0, + __m128i* const out1, + __m128i* const out2, + __m128i* const out3) { + const __m128i A0 = _mm_loadu_si128((const __m128i*)(src + 0)); + const __m128i A1 = _mm_loadu_si128((const __m128i*)(src + 4)); + const __m128i A2 = _mm_srli_epi64(A0, 32); + const __m128i A3 = _mm_srli_epi64(A1, 32); + if (mult != NULL) { + *out0 = _mm_mul_epu32(A0, *mult); + *out1 = _mm_mul_epu32(A1, *mult); + *out2 = _mm_mul_epu32(A2, *mult); + *out3 = _mm_mul_epu32(A3, *mult); + } else { + *out0 = A0; + *out1 = A1; + *out2 = A2; + *out3 = A3; + } +} + +static WEBP_INLINE void ProcessRow(const __m128i* const A0, + const __m128i* const A1, + const __m128i* const A2, + const __m128i* const A3, + const __m128i* const mult, + uint8_t* const dst) { + const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER); + const __m128i mask = _mm_set_epi32(0xffffffffu, 0, 0xffffffffu, 0); + const __m128i B0 = _mm_mul_epu32(*A0, *mult); + const __m128i B1 = _mm_mul_epu32(*A1, *mult); + const __m128i B2 = _mm_mul_epu32(*A2, *mult); + const __m128i B3 = _mm_mul_epu32(*A3, *mult); + const __m128i C0 = _mm_add_epi64(B0, rounder); + const __m128i C1 = _mm_add_epi64(B1, rounder); + const __m128i C2 = _mm_add_epi64(B2, rounder); + const __m128i C3 = _mm_add_epi64(B3, rounder); + const __m128i D0 = _mm_srli_epi64(C0, WEBP_RESCALER_RFIX); + const __m128i D1 = _mm_srli_epi64(C1, WEBP_RESCALER_RFIX); +#if (WEBP_RESCALER_FIX < 32) + const __m128i D2 = + _mm_and_si128(_mm_slli_epi64(C2, 32 - WEBP_RESCALER_RFIX), mask); + const __m128i D3 = + _mm_and_si128(_mm_slli_epi64(C3, 32 - WEBP_RESCALER_RFIX), mask); +#else + const __m128i D2 = _mm_and_si128(C2, mask); + const __m128i D3 = _mm_and_si128(C3, mask); +#endif + const __m128i E0 = _mm_or_si128(D0, D2); + const __m128i E1 = _mm_or_si128(D1, D3); + const __m128i F = _mm_packs_epi32(E0, E1); + const __m128i G = _mm_packus_epi16(F, F); + _mm_storel_epi64((__m128i*)dst, G); +} + +static void RescalerExportRowExpandSSE2(WebPRescaler* const wrk) { + int x_out; + uint8_t* const dst = wrk->dst; + rescaler_t* const irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* const frow = wrk->frow; + const __m128i mult = _mm_set_epi32(0, wrk->fy_scale, 0, wrk->fy_scale); + + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0 && wrk->y_sub + wrk->y_accum >= 0); + assert(wrk->y_expand); + if (wrk->y_accum == 0) { + for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) { + __m128i A0, A1, A2, A3; + LoadDispatchAndMult(frow + x_out, NULL, &A0, &A1, &A2, &A3); + ProcessRow(&A0, &A1, &A2, &A3, &mult, dst + x_out); + } + for (; x_out < x_out_max; ++x_out) { + const uint32_t J = frow[x_out]; + const int v = (int)MULT_FIX(J, wrk->fy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + } + } else { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + const __m128i mA = _mm_set_epi32(0, A, 0, A); + const __m128i mB = _mm_set_epi32(0, B, 0, B); + const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER); + for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) { + __m128i A0, A1, A2, A3, B0, B1, B2, B3; + LoadDispatchAndMult(frow + x_out, &mA, &A0, &A1, &A2, &A3); + LoadDispatchAndMult(irow + x_out, &mB, &B0, &B1, &B2, &B3); + { + const __m128i C0 = _mm_add_epi64(A0, B0); + const __m128i C1 = _mm_add_epi64(A1, B1); + const __m128i C2 = _mm_add_epi64(A2, B2); + const __m128i C3 = _mm_add_epi64(A3, B3); + const __m128i D0 = _mm_add_epi64(C0, rounder); + const __m128i D1 = _mm_add_epi64(C1, rounder); + const __m128i D2 = _mm_add_epi64(C2, rounder); + const __m128i D3 = _mm_add_epi64(C3, rounder); + const __m128i E0 = _mm_srli_epi64(D0, WEBP_RESCALER_RFIX); + const __m128i E1 = _mm_srli_epi64(D1, WEBP_RESCALER_RFIX); + const __m128i E2 = _mm_srli_epi64(D2, WEBP_RESCALER_RFIX); + const __m128i E3 = _mm_srli_epi64(D3, WEBP_RESCALER_RFIX); + ProcessRow(&E0, &E1, &E2, &E3, &mult, dst + x_out); + } + } + for (; x_out < x_out_max; ++x_out) { + const uint64_t I = (uint64_t)A * frow[x_out] + + (uint64_t)B * irow[x_out]; + const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); + const int v = (int)MULT_FIX(J, wrk->fy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + } + } +} + +static void RescalerExportRowShrinkSSE2(WebPRescaler* const wrk) { + int x_out; + uint8_t* const dst = wrk->dst; + rescaler_t* const irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* const frow = wrk->frow; + const uint32_t yscale = wrk->fy_scale * (-wrk->y_accum); + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + if (yscale) { + const int scale_xy = wrk->fxy_scale; + const __m128i mult_xy = _mm_set_epi32(0, scale_xy, 0, scale_xy); + const __m128i mult_y = _mm_set_epi32(0, yscale, 0, yscale); + const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER); + for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) { + __m128i A0, A1, A2, A3, B0, B1, B2, B3; + LoadDispatchAndMult(irow + x_out, NULL, &A0, &A1, &A2, &A3); + LoadDispatchAndMult(frow + x_out, &mult_y, &B0, &B1, &B2, &B3); + { + const __m128i C0 = _mm_add_epi64(B0, rounder); + const __m128i C1 = _mm_add_epi64(B1, rounder); + const __m128i C2 = _mm_add_epi64(B2, rounder); + const __m128i C3 = _mm_add_epi64(B3, rounder); + const __m128i D0 = _mm_srli_epi64(C0, WEBP_RESCALER_RFIX); // = frac + const __m128i D1 = _mm_srli_epi64(C1, WEBP_RESCALER_RFIX); + const __m128i D2 = _mm_srli_epi64(C2, WEBP_RESCALER_RFIX); + const __m128i D3 = _mm_srli_epi64(C3, WEBP_RESCALER_RFIX); + const __m128i E0 = _mm_sub_epi64(A0, D0); // irow[x] - frac + const __m128i E1 = _mm_sub_epi64(A1, D1); + const __m128i E2 = _mm_sub_epi64(A2, D2); + const __m128i E3 = _mm_sub_epi64(A3, D3); + const __m128i F2 = _mm_slli_epi64(D2, 32); + const __m128i F3 = _mm_slli_epi64(D3, 32); + const __m128i G0 = _mm_or_si128(D0, F2); + const __m128i G1 = _mm_or_si128(D1, F3); + _mm_storeu_si128((__m128i*)(irow + x_out + 0), G0); + _mm_storeu_si128((__m128i*)(irow + x_out + 4), G1); + ProcessRow(&E0, &E1, &E2, &E3, &mult_xy, dst + x_out); + } + } + for (; x_out < x_out_max; ++x_out) { + const uint32_t frac = (int)MULT_FIX(frow[x_out], yscale); + const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + irow[x_out] = frac; // new fractional start + } + } else { + const uint32_t scale = wrk->fxy_scale; + const __m128i mult = _mm_set_epi32(0, scale, 0, scale); + const __m128i zero = _mm_setzero_si128(); + for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) { + __m128i A0, A1, A2, A3; + LoadDispatchAndMult(irow + x_out, NULL, &A0, &A1, &A2, &A3); + _mm_storeu_si128((__m128i*)(irow + x_out + 0), zero); + _mm_storeu_si128((__m128i*)(irow + x_out + 4), zero); + ProcessRow(&A0, &A1, &A2, &A3, &mult, dst + x_out); + } + for (; x_out < x_out_max; ++x_out) { + const int v = (int)MULT_FIX(irow[x_out], scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + irow[x_out] = 0; + } + } +} + +#undef MULT_FIX +#undef ROUNDER + +//------------------------------------------------------------------------------ + +extern void WebPRescalerDspInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitSSE2(void) { + WebPRescalerImportRowExpand = RescalerImportRowExpandSSE2; + WebPRescalerImportRowShrink = RescalerImportRowShrinkSSE2; + WebPRescalerExportRowExpand = RescalerExportRowExpandSSE2; + WebPRescalerExportRowShrink = RescalerExportRowShrinkSSE2; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(WebPRescalerDspInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/upsampling.c b/drivers/webp/dsp/upsampling.c index 4855eb1432..651274fcee 100644 --- a/drivers/webp/dsp/upsampling.c +++ b/drivers/webp/dsp/upsampling.c @@ -1,8 +1,10 @@ // Copyright 2011 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // YUV to RGB upsampling functions. @@ -12,9 +14,7 @@ #include "./dsp.h" #include "./yuv.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif +#include <assert.h> //------------------------------------------------------------------------------ // Fancy upsampler @@ -32,7 +32,7 @@ WebPUpsampleLinePairFunc WebPUpsamplers[MODE_LAST]; // ([3*a + b + 9*c + 3*d a + 3*b + 3*c + 9*d] [8 8]) / 16 // We process u and v together stashed into 32bit (16bit each). -#define LOAD_UV(u,v) ((u) | ((v) << 16)) +#define LOAD_UV(u, v) ((u) | ((v) << 16)) #define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ @@ -43,11 +43,12 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ const int last_pixel_pair = (len - 1) >> 1; \ uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]); /* top-left sample */ \ uint32_t l_uv = LOAD_UV(cur_u[0], cur_v[0]); /* left-sample */ \ - if (top_y) { \ + assert(top_y != NULL); \ + { \ const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \ } \ - if (bottom_y) { \ + if (bottom_y != NULL) { \ const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst); \ } \ @@ -58,7 +59,7 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ const uint32_t avg = tl_uv + t_uv + l_uv + uv + 0x00080008u; \ const uint32_t diag_12 = (avg + 2 * (t_uv + l_uv)) >> 3; \ const uint32_t diag_03 = (avg + 2 * (tl_uv + uv)) >> 3; \ - if (top_y) { \ + { \ const uint32_t uv0 = (diag_12 + tl_uv) >> 1; \ const uint32_t uv1 = (diag_03 + t_uv) >> 1; \ FUNC(top_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ @@ -66,7 +67,7 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16), \ top_dst + (2 * x - 0) * XSTEP); \ } \ - if (bottom_y) { \ + if (bottom_y != NULL) { \ const uint32_t uv0 = (diag_03 + l_uv) >> 1; \ const uint32_t uv1 = (diag_12 + uv) >> 1; \ FUNC(bottom_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ @@ -78,12 +79,12 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ l_uv = uv; \ } \ if (!(len & 1)) { \ - if (top_y) { \ + { \ const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ FUNC(top_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ top_dst + (len - 1) * XSTEP); \ } \ - if (bottom_y) { \ + if (bottom_y != NULL) { \ const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ FUNC(bottom_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ bottom_dst + (len - 1) * XSTEP); \ @@ -106,57 +107,6 @@ UPSAMPLE_FUNC(UpsampleRgb565LinePair, VP8YuvToRgb565, 2) #endif // FANCY_UPSAMPLING //------------------------------------------------------------------------------ -// simple point-sampling - -#define SAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ -static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ - const uint8_t* u, const uint8_t* v, \ - uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ - int i; \ - for (i = 0; i < len - 1; i += 2) { \ - FUNC(top_y[0], u[0], v[0], top_dst); \ - FUNC(top_y[1], u[0], v[0], top_dst + XSTEP); \ - FUNC(bottom_y[0], u[0], v[0], bottom_dst); \ - FUNC(bottom_y[1], u[0], v[0], bottom_dst + XSTEP); \ - top_y += 2; \ - bottom_y += 2; \ - u++; \ - v++; \ - top_dst += 2 * XSTEP; \ - bottom_dst += 2 * XSTEP; \ - } \ - if (i == len - 1) { /* last one */ \ - FUNC(top_y[0], u[0], v[0], top_dst); \ - FUNC(bottom_y[0], u[0], v[0], bottom_dst); \ - } \ -} - -// All variants implemented. -SAMPLE_FUNC(SampleRgbLinePair, VP8YuvToRgb, 3) -SAMPLE_FUNC(SampleBgrLinePair, VP8YuvToBgr, 3) -SAMPLE_FUNC(SampleRgbaLinePair, VP8YuvToRgba, 4) -SAMPLE_FUNC(SampleBgraLinePair, VP8YuvToBgra, 4) -SAMPLE_FUNC(SampleArgbLinePair, VP8YuvToArgb, 4) -SAMPLE_FUNC(SampleRgba4444LinePair, VP8YuvToRgba4444, 2) -SAMPLE_FUNC(SampleRgb565LinePair, VP8YuvToRgb565, 2) - -#undef SAMPLE_FUNC - -const WebPSampleLinePairFunc WebPSamplers[MODE_LAST] = { - SampleRgbLinePair, // MODE_RGB - SampleRgbaLinePair, // MODE_RGBA - SampleBgrLinePair, // MODE_BGR - SampleBgraLinePair, // MODE_BGRA - SampleArgbLinePair, // MODE_ARGB - SampleRgba4444LinePair, // MODE_RGBA_4444 - SampleRgb565LinePair, // MODE_RGB_565 - SampleRgbaLinePair, // MODE_rgbA - SampleBgraLinePair, // MODE_bgrA - SampleArgbLinePair, // MODE_Argb - SampleRgba4444LinePair // MODE_rgbA_4444 -}; - -//------------------------------------------------------------------------------ #if !defined(FANCY_UPSAMPLING) #define DUAL_SAMPLE_FUNC(FUNC_NAME, FUNC) \ @@ -166,7 +116,8 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y, \ uint8_t* top_dst, uint8_t* bot_dst, int len) { \ const int half_len = len >> 1; \ int x; \ - if (top_dst != NULL) { \ + assert(top_dst != NULL); \ + { \ for (x = 0; x < half_len; ++x) { \ FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x + 0); \ FUNC(top_y[2 * x + 1], top_u[x], top_v[x], top_dst + 8 * x + 4); \ @@ -202,116 +153,75 @@ WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last) { // YUV444 converter #define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP) \ -static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \ - uint8_t* dst, int len) { \ +extern void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len); \ +void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ int i; \ for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]); \ } -YUV444_FUNC(Yuv444ToRgb, VP8YuvToRgb, 3) -YUV444_FUNC(Yuv444ToBgr, VP8YuvToBgr, 3) -YUV444_FUNC(Yuv444ToRgba, VP8YuvToRgba, 4) -YUV444_FUNC(Yuv444ToBgra, VP8YuvToBgra, 4) -YUV444_FUNC(Yuv444ToArgb, VP8YuvToArgb, 4) -YUV444_FUNC(Yuv444ToRgba4444, VP8YuvToRgba4444, 2) -YUV444_FUNC(Yuv444ToRgb565, VP8YuvToRgb565, 2) +YUV444_FUNC(WebPYuv444ToRgbC, VP8YuvToRgb, 3) +YUV444_FUNC(WebPYuv444ToBgrC, VP8YuvToBgr, 3) +YUV444_FUNC(WebPYuv444ToRgbaC, VP8YuvToRgba, 4) +YUV444_FUNC(WebPYuv444ToBgraC, VP8YuvToBgra, 4) +YUV444_FUNC(WebPYuv444ToArgbC, VP8YuvToArgb, 4) +YUV444_FUNC(WebPYuv444ToRgba4444C, VP8YuvToRgba4444, 2) +YUV444_FUNC(WebPYuv444ToRgb565C, VP8YuvToRgb565, 2) #undef YUV444_FUNC -const WebPYUV444Converter WebPYUV444Converters[MODE_LAST] = { - Yuv444ToRgb, // MODE_RGB - Yuv444ToRgba, // MODE_RGBA - Yuv444ToBgr, // MODE_BGR - Yuv444ToBgra, // MODE_BGRA - Yuv444ToArgb, // MODE_ARGB - Yuv444ToRgba4444, // MODE_RGBA_4444 - Yuv444ToRgb565, // MODE_RGB_565 - Yuv444ToRgba, // MODE_rgbA - Yuv444ToBgra, // MODE_bgrA - Yuv444ToArgb, // MODE_Argb - Yuv444ToRgba4444 // MODE_rgbA_4444 -}; - -//------------------------------------------------------------------------------ -// Premultiplied modes - -// non dithered-modes - -// (x * a * 32897) >> 23 is bit-wise equivalent to (int)(x * a / 255.) -// for all 8bit x or a. For bit-wise equivalence to (int)(x * a / 255. + .5), -// one can use instead: (x * a * 65793 + (1 << 23)) >> 24 -#if 1 // (int)(x * a / 255.) -#define MULTIPLIER(a) ((a) * 32897UL) -#define PREMULTIPLY(x, m) (((x) * (m)) >> 23) -#else // (int)(x * a / 255. + .5) -#define MULTIPLIER(a) ((a) * 65793UL) -#define PREMULTIPLY(x, m) (((x) * (m) + (1UL << 23)) >> 24) -#endif - -static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_first, - int w, int h, int stride) { - while (h-- > 0) { - uint8_t* const rgb = rgba + (alpha_first ? 1 : 0); - const uint8_t* const alpha = rgba + (alpha_first ? 0 : 3); - int i; - for (i = 0; i < w; ++i) { - const uint32_t a = alpha[4 * i]; - if (a != 0xff) { - const uint32_t mult = MULTIPLIER(a); - rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult); - rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult); - rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult); - } - } - rgba += stride; - } -} -#undef MULTIPLIER -#undef PREMULTIPLY - -// rgbA4444 +WebPYUV444Converter WebPYUV444Converters[MODE_LAST]; -#define MULTIPLIER(a) ((a) * 0x1111) // 0x1111 ~= (1 << 16) / 15 +extern void WebPInitYUV444ConvertersMIPSdspR2(void); +extern void WebPInitYUV444ConvertersSSE2(void); -static WEBP_INLINE uint8_t dither_hi(uint8_t x) { - return (x & 0xf0) | (x >> 4); -} +static volatile VP8CPUInfo upsampling_last_cpuinfo_used1 = + (VP8CPUInfo)&upsampling_last_cpuinfo_used1; -static WEBP_INLINE uint8_t dither_lo(uint8_t x) { - return (x & 0x0f) | (x << 4); -} +WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444Converters(void) { + if (upsampling_last_cpuinfo_used1 == VP8GetCPUInfo) return; -static WEBP_INLINE uint8_t multiply(uint8_t x, uint32_t m) { - return (x * m) >> 16; -} + WebPYUV444Converters[MODE_RGB] = WebPYuv444ToRgbC; + WebPYUV444Converters[MODE_RGBA] = WebPYuv444ToRgbaC; + WebPYUV444Converters[MODE_BGR] = WebPYuv444ToBgrC; + WebPYUV444Converters[MODE_BGRA] = WebPYuv444ToBgraC; + WebPYUV444Converters[MODE_ARGB] = WebPYuv444ToArgbC; + WebPYUV444Converters[MODE_RGBA_4444] = WebPYuv444ToRgba4444C; + WebPYUV444Converters[MODE_RGB_565] = WebPYuv444ToRgb565C; + WebPYUV444Converters[MODE_rgbA] = WebPYuv444ToRgbaC; + WebPYUV444Converters[MODE_bgrA] = WebPYuv444ToBgraC; + WebPYUV444Converters[MODE_Argb] = WebPYuv444ToArgbC; + WebPYUV444Converters[MODE_rgbA_4444] = WebPYuv444ToRgba4444C; -static void ApplyAlphaMultiply4444(uint8_t* rgba4444, - int w, int h, int stride) { - while (h-- > 0) { - int i; - for (i = 0; i < w; ++i) { - const uint8_t a = (rgba4444[2 * i + 1] & 0x0f); - const uint32_t mult = MULTIPLIER(a); - const uint8_t r = multiply(dither_hi(rgba4444[2 * i + 0]), mult); - const uint8_t g = multiply(dither_lo(rgba4444[2 * i + 0]), mult); - const uint8_t b = multiply(dither_hi(rgba4444[2 * i + 1]), mult); - rgba4444[2 * i + 0] = (r & 0xf0) | ((g >> 4) & 0x0f); - rgba4444[2 * i + 1] = (b & 0xf0) | a; + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + WebPInitYUV444ConvertersSSE2(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + WebPInitYUV444ConvertersMIPSdspR2(); } - rgba4444 += stride; +#endif } + upsampling_last_cpuinfo_used1 = VP8GetCPUInfo; } -#undef MULTIPLIER - -void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int) - = ApplyAlphaMultiply; -void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int) - = ApplyAlphaMultiply4444; //------------------------------------------------------------------------------ -// Main call +// Main calls + +extern void WebPInitUpsamplersSSE2(void); +extern void WebPInitUpsamplersNEON(void); +extern void WebPInitUpsamplersMIPSdspR2(void); + +static volatile VP8CPUInfo upsampling_last_cpuinfo_used2 = + (VP8CPUInfo)&upsampling_last_cpuinfo_used2; + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplers(void) { + if (upsampling_last_cpuinfo_used2 == VP8GetCPUInfo) return; -void WebPInitUpsamplers(void) { #ifdef FANCY_UPSAMPLING WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair; WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair; @@ -320,38 +230,31 @@ void WebPInitUpsamplers(void) { WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair; WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair; WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair; - - // If defined, use CPUInfo() to overwrite some pointers with faster versions. - if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) - if (VP8GetCPUInfo(kSSE2)) { - WebPInitUpsamplersSSE2(); - } -#endif - } -#endif // FANCY_UPSAMPLING -} - -void WebPInitPremultiply(void) { - WebPApplyAlphaMultiply = ApplyAlphaMultiply; - WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply4444; - -#ifdef FANCY_UPSAMPLING WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair; WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair; WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair; WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair; + // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { #if defined(WEBP_USE_SSE2) if (VP8GetCPUInfo(kSSE2)) { - WebPInitPremultiplySSE2(); + WebPInitUpsamplersSSE2(); + } +#endif +#if defined(WEBP_USE_NEON) + if (VP8GetCPUInfo(kNEON)) { + WebPInitUpsamplersNEON(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + WebPInitUpsamplersMIPSdspR2(); } #endif } #endif // FANCY_UPSAMPLING + upsampling_last_cpuinfo_used2 = VP8GetCPUInfo; } -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" -#endif +//------------------------------------------------------------------------------ diff --git a/drivers/webp/dsp/upsampling_mips_dsp_r2.c b/drivers/webp/dsp/upsampling_mips_dsp_r2.c new file mode 100644 index 0000000000..46f207b43e --- /dev/null +++ b/drivers/webp/dsp/upsampling_mips_dsp_r2.c @@ -0,0 +1,282 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// YUV to RGB upsampling functions. +// +// Author(s): Branimir Vasic (branimir.vasic@imgtec.com) +// Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include <assert.h> +#include "./yuv.h" + +#if !defined(WEBP_YUV_USE_TABLE) + +#define YUV_TO_RGB(Y, U, V, R, G, B) do { \ + const int t1 = kYScale * Y; \ + const int t2 = kVToG * V; \ + R = kVToR * V; \ + G = kUToG * U; \ + B = kUToB * U; \ + R = t1 + R; \ + G = t1 - G; \ + B = t1 + B; \ + R = R + kRCst; \ + G = G - t2 + kGCst; \ + B = B + kBCst; \ + __asm__ volatile ( \ + "shll_s.w %[" #R "], %[" #R "], 9 \n\t" \ + "shll_s.w %[" #G "], %[" #G "], 9 \n\t" \ + "shll_s.w %[" #B "], %[" #B "], 9 \n\t" \ + "precrqu_s.qb.ph %[" #R "], %[" #R "], $zero \n\t" \ + "precrqu_s.qb.ph %[" #G "], %[" #G "], $zero \n\t" \ + "precrqu_s.qb.ph %[" #B "], %[" #B "], $zero \n\t" \ + "srl %[" #R "], %[" #R "], 24 \n\t" \ + "srl %[" #G "], %[" #G "], 24 \n\t" \ + "srl %[" #B "], %[" #B "], 24 \n\t" \ + : [R]"+r"(R), [G]"+r"(G), [B]"+r"(B) \ + : \ + ); \ + } while (0) + +static WEBP_INLINE void YuvToRgb(int y, int u, int v, uint8_t* const rgb) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + rgb[0] = r; + rgb[1] = g; + rgb[2] = b; +} +static WEBP_INLINE void YuvToBgr(int y, int u, int v, uint8_t* const bgr) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + bgr[0] = b; + bgr[1] = g; + bgr[2] = r; +} +static WEBP_INLINE void YuvToRgb565(int y, int u, int v, uint8_t* const rgb) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + { + const int rg = (r & 0xf8) | (g >> 5); + const int gb = ((g << 3) & 0xe0) | (b >> 3); +#ifdef WEBP_SWAP_16BIT_CSP + rgb[0] = gb; + rgb[1] = rg; +#else + rgb[0] = rg; + rgb[1] = gb; +#endif + } +} +static WEBP_INLINE void YuvToRgba4444(int y, int u, int v, + uint8_t* const argb) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + { + const int rg = (r & 0xf0) | (g >> 4); + const int ba = (b & 0xf0) | 0x0f; // overwrite the lower 4 bits +#ifdef WEBP_SWAP_16BIT_CSP + argb[0] = ba; + argb[1] = rg; +#else + argb[0] = rg; + argb[1] = ba; +#endif + } +} +#endif // WEBP_YUV_USE_TABLE + +//----------------------------------------------------------------------------- +// Alpha handling variants + +static WEBP_INLINE void YuvToArgb(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const argb) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + argb[0] = 0xff; + argb[1] = r; + argb[2] = g; + argb[3] = b; +} +static WEBP_INLINE void YuvToBgra(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const bgra) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + bgra[0] = b; + bgra[1] = g; + bgra[2] = r; + bgra[3] = 0xff; +} +static WEBP_INLINE void YuvToRgba(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const rgba) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + rgba[0] = r; + rgba[1] = g; + rgba[2] = b; + rgba[3] = 0xff; +} + +//------------------------------------------------------------------------------ +// Fancy upsampler + +#ifdef FANCY_UPSAMPLING + +// Given samples laid out in a square as: +// [a b] +// [c d] +// we interpolate u/v as: +// ([9*a + 3*b + 3*c + d 3*a + 9*b + 3*c + d] + [8 8]) / 16 +// ([3*a + b + 9*c + 3*d a + 3*b + 3*c + 9*d] [8 8]) / 16 + +// We process u and v together stashed into 32bit (16bit each). +#define LOAD_UV(u, v) ((u) | ((v) << 16)) + +#define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ +static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ + const uint8_t* top_u, const uint8_t* top_v, \ + const uint8_t* cur_u, const uint8_t* cur_v, \ + uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ + int x; \ + const int last_pixel_pair = (len - 1) >> 1; \ + uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]); /* top-left sample */ \ + uint32_t l_uv = LOAD_UV(cur_u[0], cur_v[0]); /* left-sample */ \ + assert(top_y != NULL); \ + { \ + const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ + FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \ + } \ + if (bottom_y != NULL) { \ + const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ + FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst); \ + } \ + for (x = 1; x <= last_pixel_pair; ++x) { \ + const uint32_t t_uv = LOAD_UV(top_u[x], top_v[x]); /* top sample */ \ + const uint32_t uv = LOAD_UV(cur_u[x], cur_v[x]); /* sample */ \ + /* precompute invariant values associated with first and second diagonals*/\ + const uint32_t avg = tl_uv + t_uv + l_uv + uv + 0x00080008u; \ + const uint32_t diag_12 = (avg + 2 * (t_uv + l_uv)) >> 3; \ + const uint32_t diag_03 = (avg + 2 * (tl_uv + uv)) >> 3; \ + { \ + const uint32_t uv0 = (diag_12 + tl_uv) >> 1; \ + const uint32_t uv1 = (diag_03 + t_uv) >> 1; \ + FUNC(top_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ + top_dst + (2 * x - 1) * XSTEP); \ + FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16), \ + top_dst + (2 * x - 0) * XSTEP); \ + } \ + if (bottom_y != NULL) { \ + const uint32_t uv0 = (diag_03 + l_uv) >> 1; \ + const uint32_t uv1 = (diag_12 + uv) >> 1; \ + FUNC(bottom_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ + bottom_dst + (2 * x - 1) * XSTEP); \ + FUNC(bottom_y[2 * x + 0], uv1 & 0xff, (uv1 >> 16), \ + bottom_dst + (2 * x + 0) * XSTEP); \ + } \ + tl_uv = t_uv; \ + l_uv = uv; \ + } \ + if (!(len & 1)) { \ + { \ + const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ + FUNC(top_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ + top_dst + (len - 1) * XSTEP); \ + } \ + if (bottom_y != NULL) { \ + const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ + FUNC(bottom_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ + bottom_dst + (len - 1) * XSTEP); \ + } \ + } \ +} + +// All variants implemented. +UPSAMPLE_FUNC(UpsampleRgbLinePair, YuvToRgb, 3) +UPSAMPLE_FUNC(UpsampleBgrLinePair, YuvToBgr, 3) +UPSAMPLE_FUNC(UpsampleRgbaLinePair, YuvToRgba, 4) +UPSAMPLE_FUNC(UpsampleBgraLinePair, YuvToBgra, 4) +UPSAMPLE_FUNC(UpsampleArgbLinePair, YuvToArgb, 4) +UPSAMPLE_FUNC(UpsampleRgba4444LinePair, YuvToRgba4444, 2) +UPSAMPLE_FUNC(UpsampleRgb565LinePair, YuvToRgb565, 2) + +#undef LOAD_UV +#undef UPSAMPLE_FUNC + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitUpsamplersMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMIPSdspR2(void) { + WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair; + WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair; + WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair; + WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair; + WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair; + WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair; + WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair; + WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair; + WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair; +} + +#endif // FANCY_UPSAMPLING + +//------------------------------------------------------------------------------ +// YUV444 converter + +#define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP) \ +static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ + int i; \ + for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]); \ +} + +YUV444_FUNC(Yuv444ToRgb, YuvToRgb, 3) +YUV444_FUNC(Yuv444ToBgr, YuvToBgr, 3) +YUV444_FUNC(Yuv444ToRgba, YuvToRgba, 4) +YUV444_FUNC(Yuv444ToBgra, YuvToBgra, 4) +YUV444_FUNC(Yuv444ToArgb, YuvToArgb, 4) +YUV444_FUNC(Yuv444ToRgba4444, YuvToRgba4444, 2) +YUV444_FUNC(Yuv444ToRgb565, YuvToRgb565, 2) + +#undef YUV444_FUNC + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitYUV444ConvertersMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersMIPSdspR2(void) { + WebPYUV444Converters[MODE_RGB] = Yuv444ToRgb; + WebPYUV444Converters[MODE_RGBA] = Yuv444ToRgba; + WebPYUV444Converters[MODE_BGR] = Yuv444ToBgr; + WebPYUV444Converters[MODE_BGRA] = Yuv444ToBgra; + WebPYUV444Converters[MODE_ARGB] = Yuv444ToArgb; + WebPYUV444Converters[MODE_RGBA_4444] = Yuv444ToRgba4444; + WebPYUV444Converters[MODE_RGB_565] = Yuv444ToRgb565; + WebPYUV444Converters[MODE_rgbA] = Yuv444ToRgba; + WebPYUV444Converters[MODE_bgrA] = Yuv444ToBgra; + WebPYUV444Converters[MODE_Argb] = Yuv444ToArgb; + WebPYUV444Converters[MODE_rgbA_4444] = Yuv444ToRgba4444; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 + +#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_MIPS_DSP_R2)) +WEBP_DSP_INIT_STUB(WebPInitUpsamplersMIPSdspR2) +#endif diff --git a/drivers/webp/dsp/upsampling_neon.c b/drivers/webp/dsp/upsampling_neon.c new file mode 100644 index 0000000000..a8384c2149 --- /dev/null +++ b/drivers/webp/dsp/upsampling_neon.c @@ -0,0 +1,261 @@ +// Copyright 2011 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// NEON version of YUV to RGB upsampling functions. +// +// Author: mans@mansr.com (Mans Rullgard) +// Based on SSE code by: somnath@google.com (Somnath Banerjee) + +#include "./dsp.h" + +#if defined(WEBP_USE_NEON) + +#include <assert.h> +#include <arm_neon.h> +#include <string.h> +#include "./neon.h" +#include "./yuv.h" + +#ifdef FANCY_UPSAMPLING + +//----------------------------------------------------------------------------- +// U/V upsampling + +// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels. +#define UPSAMPLE_16PIXELS(r1, r2, out) { \ + uint8x8_t a = vld1_u8(r1); \ + uint8x8_t b = vld1_u8(r1 + 1); \ + uint8x8_t c = vld1_u8(r2); \ + uint8x8_t d = vld1_u8(r2 + 1); \ + \ + uint16x8_t al = vshll_n_u8(a, 1); \ + uint16x8_t bl = vshll_n_u8(b, 1); \ + uint16x8_t cl = vshll_n_u8(c, 1); \ + uint16x8_t dl = vshll_n_u8(d, 1); \ + \ + uint8x8_t diag1, diag2; \ + uint16x8_t sl; \ + \ + /* a + b + c + d */ \ + sl = vaddl_u8(a, b); \ + sl = vaddw_u8(sl, c); \ + sl = vaddw_u8(sl, d); \ + \ + al = vaddq_u16(sl, al); /* 3a + b + c + d */ \ + bl = vaddq_u16(sl, bl); /* a + 3b + c + d */ \ + \ + al = vaddq_u16(al, dl); /* 3a + b + c + 3d */ \ + bl = vaddq_u16(bl, cl); /* a + 3b + 3c + d */ \ + \ + diag2 = vshrn_n_u16(al, 3); \ + diag1 = vshrn_n_u16(bl, 3); \ + \ + a = vrhadd_u8(a, diag1); \ + b = vrhadd_u8(b, diag2); \ + c = vrhadd_u8(c, diag2); \ + d = vrhadd_u8(d, diag1); \ + \ + { \ + uint8x8x2_t a_b, c_d; \ + INIT_VECTOR2(a_b, a, b); \ + INIT_VECTOR2(c_d, c, d); \ + vst2_u8(out, a_b); \ + vst2_u8(out + 32, c_d); \ + } \ +} + +// Turn the macro into a function for reducing code-size when non-critical +static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2, + uint8_t *out) { + UPSAMPLE_16PIXELS(r1, r2, out); +} + +#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) { \ + uint8_t r1[9], r2[9]; \ + memcpy(r1, (tb), (num_pixels)); \ + memcpy(r2, (bb), (num_pixels)); \ + /* replicate last byte */ \ + memset(r1 + (num_pixels), r1[(num_pixels) - 1], 9 - (num_pixels)); \ + memset(r2 + (num_pixels), r2[(num_pixels) - 1], 9 - (num_pixels)); \ + Upsample16Pixels(r1, r2, out); \ +} + +//----------------------------------------------------------------------------- +// YUV->RGB conversion + +static const int16_t kCoeffs[4] = { kYScale, kVToR, kUToG, kVToG }; + +#define v255 vdup_n_u8(255) + +#define STORE_Rgb(out, r, g, b) do { \ + uint8x8x3_t r_g_b; \ + INIT_VECTOR3(r_g_b, r, g, b); \ + vst3_u8(out, r_g_b); \ +} while (0) + +#define STORE_Bgr(out, r, g, b) do { \ + uint8x8x3_t b_g_r; \ + INIT_VECTOR3(b_g_r, b, g, r); \ + vst3_u8(out, b_g_r); \ +} while (0) + +#define STORE_Rgba(out, r, g, b) do { \ + uint8x8x4_t r_g_b_v255; \ + INIT_VECTOR4(r_g_b_v255, r, g, b, v255); \ + vst4_u8(out, r_g_b_v255); \ +} while (0) + +#define STORE_Bgra(out, r, g, b) do { \ + uint8x8x4_t b_g_r_v255; \ + INIT_VECTOR4(b_g_r_v255, b, g, r, v255); \ + vst4_u8(out, b_g_r_v255); \ +} while (0) + +#define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) { \ + int i; \ + for (i = 0; i < N; i += 8) { \ + const int off = ((cur_x) + i) * XSTEP; \ + uint8x8_t y = vld1_u8((src_y) + (cur_x) + i); \ + uint8x8_t u = vld1_u8((src_uv) + i); \ + uint8x8_t v = vld1_u8((src_uv) + i + 16); \ + const int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16)); \ + const int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128)); \ + const int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128)); \ + int32x4_t yl = vmull_lane_s16(vget_low_s16(yy), cf16, 0); \ + int32x4_t yh = vmull_lane_s16(vget_high_s16(yy), cf16, 0); \ + const int32x4_t rl = vmlal_lane_s16(yl, vget_low_s16(vv), cf16, 1);\ + const int32x4_t rh = vmlal_lane_s16(yh, vget_high_s16(vv), cf16, 1);\ + int32x4_t gl = vmlsl_lane_s16(yl, vget_low_s16(uu), cf16, 2); \ + int32x4_t gh = vmlsl_lane_s16(yh, vget_high_s16(uu), cf16, 2); \ + const int32x4_t bl = vmovl_s16(vget_low_s16(uu)); \ + const int32x4_t bh = vmovl_s16(vget_high_s16(uu)); \ + gl = vmlsl_lane_s16(gl, vget_low_s16(vv), cf16, 3); \ + gh = vmlsl_lane_s16(gh, vget_high_s16(vv), cf16, 3); \ + yl = vmlaq_lane_s32(yl, bl, cf32, 0); \ + yh = vmlaq_lane_s32(yh, bh, cf32, 0); \ + /* vrshrn_n_s32() already incorporates the rounding constant */ \ + y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, YUV_FIX2), \ + vrshrn_n_s32(rh, YUV_FIX2))); \ + u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, YUV_FIX2), \ + vrshrn_n_s32(gh, YUV_FIX2))); \ + v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(yl, YUV_FIX2), \ + vrshrn_n_s32(yh, YUV_FIX2))); \ + STORE_ ## FMT(out + off, y, u, v); \ + } \ +} + +#define CONVERT1(FUNC, XSTEP, N, src_y, src_uv, rgb, cur_x) { \ + int i; \ + for (i = 0; i < N; i++) { \ + const int off = ((cur_x) + i) * XSTEP; \ + const int y = src_y[(cur_x) + i]; \ + const int u = (src_uv)[i]; \ + const int v = (src_uv)[i + 16]; \ + FUNC(y, u, v, rgb + off); \ + } \ +} + +#define CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, uv, \ + top_dst, bottom_dst, cur_x, len) { \ + CONVERT8(FMT, XSTEP, len, top_y, uv, top_dst, cur_x) \ + if (bottom_y != NULL) { \ + CONVERT8(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x) \ + } \ +} + +#define CONVERT2RGB_1(FUNC, XSTEP, top_y, bottom_y, uv, \ + top_dst, bottom_dst, cur_x, len) { \ + CONVERT1(FUNC, XSTEP, len, top_y, uv, top_dst, cur_x); \ + if (bottom_y != NULL) { \ + CONVERT1(FUNC, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x); \ + } \ +} + +#define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, XSTEP) \ +static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y, \ + const uint8_t *top_u, const uint8_t *top_v, \ + const uint8_t *cur_u, const uint8_t *cur_v, \ + uint8_t *top_dst, uint8_t *bottom_dst, int len) { \ + int block; \ + /* 16 byte aligned array to cache reconstructed u and v */ \ + uint8_t uv_buf[2 * 32 + 15]; \ + uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ + const int uv_len = (len + 1) >> 1; \ + /* 9 pixels must be read-able for each block */ \ + const int num_blocks = (uv_len - 1) >> 3; \ + const int leftover = uv_len - num_blocks * 8; \ + const int last_pos = 1 + 16 * num_blocks; \ + \ + const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \ + const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \ + \ + const int16x4_t cf16 = vld1_s16(kCoeffs); \ + const int32x2_t cf32 = vdup_n_s32(kUToB); \ + const uint8x8_t u16 = vdup_n_u8(16); \ + const uint8x8_t u128 = vdup_n_u8(128); \ + \ + /* Treat the first pixel in regular way */ \ + assert(top_y != NULL); \ + { \ + const int u0 = (top_u[0] + u_diag) >> 1; \ + const int v0 = (top_v[0] + v_diag) >> 1; \ + VP8YuvTo ## FMT(top_y[0], u0, v0, top_dst); \ + } \ + if (bottom_y != NULL) { \ + const int u0 = (cur_u[0] + u_diag) >> 1; \ + const int v0 = (cur_v[0] + v_diag) >> 1; \ + VP8YuvTo ## FMT(bottom_y[0], u0, v0, bottom_dst); \ + } \ + \ + for (block = 0; block < num_blocks; ++block) { \ + UPSAMPLE_16PIXELS(top_u, cur_u, r_uv); \ + UPSAMPLE_16PIXELS(top_v, cur_v, r_uv + 16); \ + CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, r_uv, \ + top_dst, bottom_dst, 16 * block + 1, 16); \ + top_u += 8; \ + cur_u += 8; \ + top_v += 8; \ + cur_v += 8; \ + } \ + \ + UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv); \ + UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 16); \ + CONVERT2RGB_1(VP8YuvTo ## FMT, XSTEP, top_y, bottom_y, r_uv, \ + top_dst, bottom_dst, last_pos, len - last_pos); \ +} + +// NEON variants of the fancy upsampler. +NEON_UPSAMPLE_FUNC(UpsampleRgbLinePair, Rgb, 3) +NEON_UPSAMPLE_FUNC(UpsampleBgrLinePair, Bgr, 3) +NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePair, Rgba, 4) +NEON_UPSAMPLE_FUNC(UpsampleBgraLinePair, Bgra, 4) + +//------------------------------------------------------------------------------ +// Entry point + +extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; + +extern void WebPInitUpsamplersNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersNEON(void) { + WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair; + WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair; + WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair; + WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair; +} + +#endif // FANCY_UPSAMPLING + +#endif // WEBP_USE_NEON + +#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_NEON)) +WEBP_DSP_INIT_STUB(WebPInitUpsamplersNEON) +#endif diff --git a/drivers/webp/dsp/upsampling_sse2.c b/drivers/webp/dsp/upsampling_sse2.c index 8cb275a02b..b85808e271 100644 --- a/drivers/webp/dsp/upsampling_sse2.c +++ b/drivers/webp/dsp/upsampling_sse2.c @@ -1,8 +1,10 @@ // Copyright 2011 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // SSE2 version of YUV to RGB upsampling functions. @@ -18,10 +20,6 @@ #include <string.h> #include "./yuv.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif - #ifdef FANCY_UPSAMPLING // We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows @@ -49,23 +47,23 @@ extern "C" { (out) = _mm_sub_epi8(tmp0, tmp4); /* (k + in + 1) / 2 - lsb_correction */ \ } while (0) -// pack and store two alterning pixel rows +// pack and store two alternating pixel rows #define PACK_AND_STORE(a, b, da, db, out) do { \ - const __m128i ta = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \ - const __m128i tb = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \ - const __m128i t1 = _mm_unpacklo_epi8(ta, tb); \ - const __m128i t2 = _mm_unpackhi_epi8(ta, tb); \ - _mm_store_si128(((__m128i*)(out)) + 0, t1); \ - _mm_store_si128(((__m128i*)(out)) + 1, t2); \ + const __m128i t_a = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \ + const __m128i t_b = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \ + const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b); \ + const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b); \ + _mm_store_si128(((__m128i*)(out)) + 0, t_1); \ + _mm_store_si128(((__m128i*)(out)) + 1, t_2); \ } while (0) // Loads 17 pixels each from rows r1 and r2 and generates 32 pixels. #define UPSAMPLE_32PIXELS(r1, r2, out) { \ const __m128i one = _mm_set1_epi8(1); \ - const __m128i a = _mm_loadu_si128((__m128i*)&(r1)[0]); \ - const __m128i b = _mm_loadu_si128((__m128i*)&(r1)[1]); \ - const __m128i c = _mm_loadu_si128((__m128i*)&(r2)[0]); \ - const __m128i d = _mm_loadu_si128((__m128i*)&(r2)[1]); \ + const __m128i a = _mm_loadu_si128((const __m128i*)&(r1)[0]); \ + const __m128i b = _mm_loadu_si128((const __m128i*)&(r1)[1]); \ + const __m128i c = _mm_loadu_si128((const __m128i*)&(r2)[0]); \ + const __m128i d = _mm_loadu_si128((const __m128i*)&(r2)[1]); \ \ const __m128i s = _mm_avg_epu8(a, d); /* s = (a + d + 1) / 2 */ \ const __m128i t = _mm_avg_epu8(b, c); /* t = (b + c + 1) / 2 */ \ @@ -85,8 +83,8 @@ extern "C" { GET_M(ad, s, diag2); /* diag2 = (3a + b + c + 3d) / 8 */ \ \ /* pack the alternate pixels */ \ - PACK_AND_STORE(a, b, diag1, diag2, &(out)[0 * 32]); \ - PACK_AND_STORE(c, d, diag2, diag1, &(out)[2 * 32]); \ + PACK_AND_STORE(a, b, diag1, diag2, out + 0); /* store top */ \ + PACK_AND_STORE(c, d, diag2, diag1, out + 2 * 32); /* store bottom */ \ } // Turn the macro into a function for reducing code-size when non-critical @@ -106,104 +104,140 @@ static void Upsample32Pixels(const uint8_t r1[], const uint8_t r2[], Upsample32Pixels(r1, r2, out); \ } -#define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, uv, \ +#define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, \ top_dst, bottom_dst, cur_x, num_pixels) { \ int n; \ - if (top_y) { \ - for (n = 0; n < (num_pixels); ++n) { \ - FUNC(top_y[(cur_x) + n], (uv)[n], (uv)[32 + n], \ - top_dst + ((cur_x) + n) * XSTEP); \ - } \ + for (n = 0; n < (num_pixels); ++n) { \ + FUNC(top_y[(cur_x) + n], r_u[n], r_v[n], \ + top_dst + ((cur_x) + n) * XSTEP); \ } \ - if (bottom_y) { \ + if (bottom_y != NULL) { \ for (n = 0; n < (num_pixels); ++n) { \ - FUNC(bottom_y[(cur_x) + n], (uv)[64 + n], (uv)[64 + 32 + n], \ + FUNC(bottom_y[(cur_x) + n], r_u[64 + n], r_v[64 + n], \ bottom_dst + ((cur_x) + n) * XSTEP); \ } \ } \ } +#define CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, \ + top_dst, bottom_dst, cur_x) do { \ + FUNC##32(top_y + (cur_x), r_u, r_v, top_dst + (cur_x) * XSTEP); \ + if (bottom_y != NULL) { \ + FUNC##32(bottom_y + (cur_x), r_u + 64, r_v + 64, \ + bottom_dst + (cur_x) * XSTEP); \ + } \ +} while (0) + #define SSE2_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ const uint8_t* top_u, const uint8_t* top_v, \ const uint8_t* cur_u, const uint8_t* cur_v, \ uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ - int b; \ - /* 16 byte aligned array to cache reconstructed u and v */ \ + int uv_pos, pos; \ + /* 16byte-aligned array to cache reconstructed u and v */ \ uint8_t uv_buf[4 * 32 + 15]; \ - uint8_t* const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ - const int uv_len = (len + 1) >> 1; \ - /* 17 pixels must be read-able for each block */ \ - const int num_blocks = (uv_len - 1) >> 4; \ - const int leftover = uv_len - num_blocks * 16; \ - const int last_pos = 1 + 32 * num_blocks; \ - \ - const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \ - const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \ + uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ + uint8_t* const r_v = r_u + 32; \ \ - assert(len > 0); \ - /* Treat the first pixel in regular way */ \ - if (top_y) { \ - const int u0 = (top_u[0] + u_diag) >> 1; \ - const int v0 = (top_v[0] + v_diag) >> 1; \ - FUNC(top_y[0], u0, v0, top_dst); \ + assert(top_y != NULL); \ + { /* Treat the first pixel in regular way */ \ + const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \ + const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \ + const int u0_t = (top_u[0] + u_diag) >> 1; \ + const int v0_t = (top_v[0] + v_diag) >> 1; \ + FUNC(top_y[0], u0_t, v0_t, top_dst); \ + if (bottom_y != NULL) { \ + const int u0_b = (cur_u[0] + u_diag) >> 1; \ + const int v0_b = (cur_v[0] + v_diag) >> 1; \ + FUNC(bottom_y[0], u0_b, v0_b, bottom_dst); \ + } \ } \ - if (bottom_y) { \ - const int u0 = (cur_u[0] + u_diag) >> 1; \ - const int v0 = (cur_v[0] + v_diag) >> 1; \ - FUNC(bottom_y[0], u0, v0, bottom_dst); \ + /* For UPSAMPLE_32PIXELS, 17 u/v values must be read-able for each block */ \ + for (pos = 1, uv_pos = 0; pos + 32 + 1 <= len; pos += 32, uv_pos += 16) { \ + UPSAMPLE_32PIXELS(top_u + uv_pos, cur_u + uv_pos, r_u); \ + UPSAMPLE_32PIXELS(top_v + uv_pos, cur_v + uv_pos, r_v); \ + CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst, pos); \ } \ - \ - for (b = 0; b < num_blocks; ++b) { \ - UPSAMPLE_32PIXELS(top_u, cur_u, r_uv + 0 * 32); \ - UPSAMPLE_32PIXELS(top_v, cur_v, r_uv + 1 * 32); \ - CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst, \ - 32 * b + 1, 32) \ - top_u += 16; \ - cur_u += 16; \ - top_v += 16; \ - cur_v += 16; \ + if (len > 1) { \ + const int left_over = ((len + 1) >> 1) - (pos >> 1); \ + assert(left_over > 0); \ + UPSAMPLE_LAST_BLOCK(top_u + uv_pos, cur_u + uv_pos, left_over, r_u); \ + UPSAMPLE_LAST_BLOCK(top_v + uv_pos, cur_v + uv_pos, left_over, r_v); \ + CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst, \ + pos, len - pos); \ } \ - \ - UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv + 0 * 32); \ - UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 1 * 32); \ - CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst, \ - last_pos, len - last_pos); \ } // SSE2 variants of the fancy upsampler. -SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePairSSE2, VP8YuvToRgb, 3) -SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePairSSE2, VP8YuvToBgr, 3) -SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePairSSE2, VP8YuvToRgba, 4) -SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4) +SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePair, VP8YuvToRgb, 3) +SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePair, VP8YuvToBgr, 3) +SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePair, VP8YuvToRgba, 4) +SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePair, VP8YuvToBgra, 4) #undef GET_M #undef PACK_AND_STORE #undef UPSAMPLE_32PIXELS #undef UPSAMPLE_LAST_BLOCK #undef CONVERT2RGB +#undef CONVERT2RGB_32 #undef SSE2_UPSAMPLE_FUNC //------------------------------------------------------------------------------ +// Entry point extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; -void WebPInitUpsamplersSSE2(void) { - WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePairSSE2; - WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2; - WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePairSSE2; - WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2; -} +extern void WebPInitUpsamplersSSE2(void); -void WebPInitPremultiplySSE2(void) { - WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2; - WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2; +WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersSSE2(void) { + VP8YUVInitSSE2(); + WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair; + WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair; + WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair; + WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair; } #endif // FANCY_UPSAMPLING -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" -#endif +//------------------------------------------------------------------------------ + +extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */]; +extern void WebPInitYUV444ConvertersSSE2(void); + +#define YUV444_FUNC(FUNC_NAME, CALL, XSTEP) \ +extern void WebP##FUNC_NAME##C(const uint8_t* y, const uint8_t* u, \ + const uint8_t* v, uint8_t* dst, int len); \ +static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ + int i; \ + const int max_len = len & ~31; \ + for (i = 0; i < max_len; i += 32) CALL(y + i, u + i, v + i, dst + i * XSTEP);\ + if (i < len) { /* C-fallback */ \ + WebP##FUNC_NAME##C(y + i, u + i, v + i, dst + i * XSTEP, len - i); \ + } \ +} -#endif // WEBP_USE_SSE2 +YUV444_FUNC(Yuv444ToRgba, VP8YuvToRgba32, 4); +YUV444_FUNC(Yuv444ToBgra, VP8YuvToBgra32, 4); +YUV444_FUNC(Yuv444ToRgb, VP8YuvToRgb32, 3); +YUV444_FUNC(Yuv444ToBgr, VP8YuvToBgr32, 3); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersSSE2(void) { + VP8YUVInitSSE2(); + WebPYUV444Converters[MODE_RGBA] = Yuv444ToRgba; + WebPYUV444Converters[MODE_BGRA] = Yuv444ToBgra; + WebPYUV444Converters[MODE_RGB] = Yuv444ToRgb; + WebPYUV444Converters[MODE_BGR] = Yuv444ToBgr; +} + +#else + +WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersSSE2) + +#endif // WEBP_USE_SSE2 + +#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_SSE2)) +WEBP_DSP_INIT_STUB(WebPInitUpsamplersSSE2) +#endif diff --git a/drivers/webp/dsp/yuv.c b/drivers/webp/dsp/yuv.c index 7f05f9a3aa..f50a253168 100644 --- a/drivers/webp/dsp/yuv.c +++ b/drivers/webp/dsp/yuv.c @@ -1,26 +1,19 @@ // Copyright 2010 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // -// YUV->RGB conversion function +// YUV->RGB conversion functions // // Author: Skal (pascal.massimino@gmail.com) #include "./yuv.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif - -enum { YUV_HALF = 1 << (YUV_FIX - 1) }; - -int16_t VP8kVToR[256], VP8kUToB[256]; -int32_t VP8kVToG[256], VP8kUToG[256]; -uint8_t VP8kClip[YUV_RANGE_MAX - YUV_RANGE_MIN]; -uint8_t VP8kClip4Bits[YUV_RANGE_MAX - YUV_RANGE_MIN]; +#if defined(WEBP_YUV_USE_TABLE) static int done = 0; @@ -28,11 +21,17 @@ static WEBP_INLINE uint8_t clip(int v, int max_value) { return v < 0 ? 0 : v > max_value ? max_value : v; } -void VP8YUVInit(void) { +int16_t VP8kVToR[256], VP8kUToB[256]; +int32_t VP8kVToG[256], VP8kUToG[256]; +uint8_t VP8kClip[YUV_RANGE_MAX - YUV_RANGE_MIN]; +uint8_t VP8kClip4Bits[YUV_RANGE_MAX - YUV_RANGE_MIN]; + +WEBP_TSAN_IGNORE_FUNCTION void VP8YUVInit(void) { int i; if (done) { return; } +#ifndef USE_YUVj for (i = 0; i < 256; ++i) { VP8kVToR[i] = (89858 * (i - 128) + YUV_HALF) >> YUV_FIX; VP8kUToG[i] = -22014 * (i - 128) + YUV_HALF; @@ -44,9 +43,238 @@ void VP8YUVInit(void) { VP8kClip[i - YUV_RANGE_MIN] = clip(k, 255); VP8kClip4Bits[i - YUV_RANGE_MIN] = clip((k + 8) >> 4, 15); } +#else + for (i = 0; i < 256; ++i) { + VP8kVToR[i] = (91881 * (i - 128) + YUV_HALF) >> YUV_FIX; + VP8kUToG[i] = -22554 * (i - 128) + YUV_HALF; + VP8kVToG[i] = -46802 * (i - 128); + VP8kUToB[i] = (116130 * (i - 128) + YUV_HALF) >> YUV_FIX; + } + for (i = YUV_RANGE_MIN; i < YUV_RANGE_MAX; ++i) { + const int k = i; + VP8kClip[i - YUV_RANGE_MIN] = clip(k, 255); + VP8kClip4Bits[i - YUV_RANGE_MIN] = clip((k + 8) >> 4, 15); + } +#endif + done = 1; } -#if defined(__cplusplus) || defined(c_plusplus) -} // extern "C" -#endif +#else + +WEBP_TSAN_IGNORE_FUNCTION void VP8YUVInit(void) {} + +#endif // WEBP_YUV_USE_TABLE + +//----------------------------------------------------------------------------- +// Plain-C version + +#define ROW_FUNC(FUNC_NAME, FUNC, XSTEP) \ +static void FUNC_NAME(const uint8_t* y, \ + const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ + const uint8_t* const end = dst + (len & ~1) * XSTEP; \ + while (dst != end) { \ + FUNC(y[0], u[0], v[0], dst); \ + FUNC(y[1], u[0], v[0], dst + XSTEP); \ + y += 2; \ + ++u; \ + ++v; \ + dst += 2 * XSTEP; \ + } \ + if (len & 1) { \ + FUNC(y[0], u[0], v[0], dst); \ + } \ +} \ + +// All variants implemented. +ROW_FUNC(YuvToRgbRow, VP8YuvToRgb, 3) +ROW_FUNC(YuvToBgrRow, VP8YuvToBgr, 3) +ROW_FUNC(YuvToRgbaRow, VP8YuvToRgba, 4) +ROW_FUNC(YuvToBgraRow, VP8YuvToBgra, 4) +ROW_FUNC(YuvToArgbRow, VP8YuvToArgb, 4) +ROW_FUNC(YuvToRgba4444Row, VP8YuvToRgba4444, 2) +ROW_FUNC(YuvToRgb565Row, VP8YuvToRgb565, 2) + +#undef ROW_FUNC + +// Main call for processing a plane with a WebPSamplerRowFunc function: +void WebPSamplerProcessPlane(const uint8_t* y, int y_stride, + const uint8_t* u, const uint8_t* v, int uv_stride, + uint8_t* dst, int dst_stride, + int width, int height, WebPSamplerRowFunc func) { + int j; + for (j = 0; j < height; ++j) { + func(y, u, v, dst, width); + y += y_stride; + if (j & 1) { + u += uv_stride; + v += uv_stride; + } + dst += dst_stride; + } +} + +//----------------------------------------------------------------------------- +// Main call + +WebPSamplerRowFunc WebPSamplers[MODE_LAST]; + +extern void WebPInitSamplersSSE2(void); +extern void WebPInitSamplersMIPS32(void); +extern void WebPInitSamplersMIPSdspR2(void); + +static volatile VP8CPUInfo yuv_last_cpuinfo_used = + (VP8CPUInfo)&yuv_last_cpuinfo_used; + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplers(void) { + if (yuv_last_cpuinfo_used == VP8GetCPUInfo) return; + + WebPSamplers[MODE_RGB] = YuvToRgbRow; + WebPSamplers[MODE_RGBA] = YuvToRgbaRow; + WebPSamplers[MODE_BGR] = YuvToBgrRow; + WebPSamplers[MODE_BGRA] = YuvToBgraRow; + WebPSamplers[MODE_ARGB] = YuvToArgbRow; + WebPSamplers[MODE_RGBA_4444] = YuvToRgba4444Row; + WebPSamplers[MODE_RGB_565] = YuvToRgb565Row; + WebPSamplers[MODE_rgbA] = YuvToRgbaRow; + WebPSamplers[MODE_bgrA] = YuvToBgraRow; + WebPSamplers[MODE_Argb] = YuvToArgbRow; + WebPSamplers[MODE_rgbA_4444] = YuvToRgba4444Row; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + WebPInitSamplersSSE2(); + } +#endif // WEBP_USE_SSE2 +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + WebPInitSamplersMIPS32(); + } +#endif // WEBP_USE_MIPS32 +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + WebPInitSamplersMIPSdspR2(); + } +#endif // WEBP_USE_MIPS_DSP_R2 + } + yuv_last_cpuinfo_used = VP8GetCPUInfo; +} + +//----------------------------------------------------------------------------- +// ARGB -> YUV converters + +static void ConvertARGBToY(const uint32_t* argb, uint8_t* y, int width) { + int i; + for (i = 0; i < width; ++i) { + const uint32_t p = argb[i]; + y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff, + YUV_HALF); + } +} + +void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v, + int src_width, int do_store) { + // No rounding. Last pixel is dealt with separately. + const int uv_width = src_width >> 1; + int i; + for (i = 0; i < uv_width; ++i) { + const uint32_t v0 = argb[2 * i + 0]; + const uint32_t v1 = argb[2 * i + 1]; + // VP8RGBToU/V expects four accumulated pixels. Hence we need to + // scale r/g/b value by a factor 2. We just shift v0/v1 one bit less. + const int r = ((v0 >> 15) & 0x1fe) + ((v1 >> 15) & 0x1fe); + const int g = ((v0 >> 7) & 0x1fe) + ((v1 >> 7) & 0x1fe); + const int b = ((v0 << 1) & 0x1fe) + ((v1 << 1) & 0x1fe); + const int tmp_u = VP8RGBToU(r, g, b, YUV_HALF << 2); + const int tmp_v = VP8RGBToV(r, g, b, YUV_HALF << 2); + if (do_store) { + u[i] = tmp_u; + v[i] = tmp_v; + } else { + // Approximated average-of-four. But it's an acceptable diff. + u[i] = (u[i] + tmp_u + 1) >> 1; + v[i] = (v[i] + tmp_v + 1) >> 1; + } + } + if (src_width & 1) { // last pixel + const uint32_t v0 = argb[2 * i + 0]; + const int r = (v0 >> 14) & 0x3fc; + const int g = (v0 >> 6) & 0x3fc; + const int b = (v0 << 2) & 0x3fc; + const int tmp_u = VP8RGBToU(r, g, b, YUV_HALF << 2); + const int tmp_v = VP8RGBToV(r, g, b, YUV_HALF << 2); + if (do_store) { + u[i] = tmp_u; + v[i] = tmp_v; + } else { + u[i] = (u[i] + tmp_u + 1) >> 1; + v[i] = (v[i] + tmp_v + 1) >> 1; + } + } +} + +//----------------------------------------------------------------------------- + +static void ConvertRGB24ToY(const uint8_t* rgb, uint8_t* y, int width) { + int i; + for (i = 0; i < width; ++i, rgb += 3) { + y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF); + } +} + +static void ConvertBGR24ToY(const uint8_t* bgr, uint8_t* y, int width) { + int i; + for (i = 0; i < width; ++i, bgr += 3) { + y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF); + } +} + +void WebPConvertRGBA32ToUV_C(const uint16_t* rgb, + uint8_t* u, uint8_t* v, int width) { + int i; + for (i = 0; i < width; i += 1, rgb += 4) { + const int r = rgb[0], g = rgb[1], b = rgb[2]; + u[i] = VP8RGBToU(r, g, b, YUV_HALF << 2); + v[i] = VP8RGBToV(r, g, b, YUV_HALF << 2); + } +} + +//----------------------------------------------------------------------------- + +void (*WebPConvertRGB24ToY)(const uint8_t* rgb, uint8_t* y, int width); +void (*WebPConvertBGR24ToY)(const uint8_t* bgr, uint8_t* y, int width); +void (*WebPConvertRGBA32ToUV)(const uint16_t* rgb, + uint8_t* u, uint8_t* v, int width); + +void (*WebPConvertARGBToY)(const uint32_t* argb, uint8_t* y, int width); +void (*WebPConvertARGBToUV)(const uint32_t* argb, uint8_t* u, uint8_t* v, + int src_width, int do_store); + +static volatile VP8CPUInfo rgba_to_yuv_last_cpuinfo_used = + (VP8CPUInfo)&rgba_to_yuv_last_cpuinfo_used; + +extern void WebPInitConvertARGBToYUVSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) { + if (rgba_to_yuv_last_cpuinfo_used == VP8GetCPUInfo) return; + + WebPConvertARGBToY = ConvertARGBToY; + WebPConvertARGBToUV = WebPConvertARGBToUV_C; + + WebPConvertRGB24ToY = ConvertRGB24ToY; + WebPConvertBGR24ToY = ConvertBGR24ToY; + + WebPConvertRGBA32ToUV = WebPConvertRGBA32ToUV_C; + + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + WebPInitConvertARGBToYUVSSE2(); + } +#endif // WEBP_USE_SSE2 + } + rgba_to_yuv_last_cpuinfo_used = VP8GetCPUInfo; +} diff --git a/drivers/webp/dsp/yuv.h b/drivers/webp/dsp/yuv.h index a569109c54..af435a5b3e 100644 --- a/drivers/webp/dsp/yuv.h +++ b/drivers/webp/dsp/yuv.h @@ -1,36 +1,165 @@ // Copyright 2010 Google Inc. All Rights Reserved. // -// This code is licensed under the same terms as WebM: -// Software License Agreement: http://www.webmproject.org/license/software/ -// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. // ----------------------------------------------------------------------------- // // inline YUV<->RGB conversion function // +// The exact naming is Y'CbCr, following the ITU-R BT.601 standard. +// More information at: http://en.wikipedia.org/wiki/YCbCr +// Y = 0.2569 * R + 0.5044 * G + 0.0979 * B + 16 +// U = -0.1483 * R - 0.2911 * G + 0.4394 * B + 128 +// V = 0.4394 * R - 0.3679 * G - 0.0715 * B + 128 +// We use 16bit fixed point operations for RGB->YUV conversion (YUV_FIX). +// +// For the Y'CbCr to RGB conversion, the BT.601 specification reads: +// R = 1.164 * (Y-16) + 1.596 * (V-128) +// G = 1.164 * (Y-16) - 0.813 * (V-128) - 0.391 * (U-128) +// B = 1.164 * (Y-16) + 2.018 * (U-128) +// where Y is in the [16,235] range, and U/V in the [16,240] range. +// In the table-lookup version (WEBP_YUV_USE_TABLE), the common factor +// "1.164 * (Y-16)" can be handled as an offset in the VP8kClip[] table. +// So in this case the formulae should read: +// R = 1.164 * [Y + 1.371 * (V-128) ] - 18.624 +// G = 1.164 * [Y - 0.698 * (V-128) - 0.336 * (U-128)] - 18.624 +// B = 1.164 * [Y + 1.733 * (U-128)] - 18.624 +// once factorized. +// For YUV->RGB conversion, only 14bit fixed precision is used (YUV_FIX2). +// That's the maximum possible for a convenient ARM implementation. +// // Author: Skal (pascal.massimino@gmail.com) #ifndef WEBP_DSP_YUV_H_ #define WEBP_DSP_YUV_H_ +#include "./dsp.h" #include "../dec/decode_vp8.h" +// Define the following to use the LUT-based code: +// #define WEBP_YUV_USE_TABLE + +#if defined(WEBP_EXPERIMENTAL_FEATURES) +// Do NOT activate this feature for real compression. This is only experimental! +// This flag is for comparison purpose against JPEG's "YUVj" natural colorspace. +// This colorspace is close to Rec.601's Y'CbCr model with the notable +// difference of allowing larger range for luma/chroma. +// See http://en.wikipedia.org/wiki/YCbCr#JPEG_conversion paragraph, and its +// difference with http://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion +// #define USE_YUVj +#endif + //------------------------------------------------------------------------------ // YUV -> RGB conversion -#if defined(__cplusplus) || defined(c_plusplus) +#ifdef __cplusplus extern "C" { #endif -enum { YUV_FIX = 16, // fixed-point precision - YUV_RANGE_MIN = -227, // min value of r/g/b output - YUV_RANGE_MAX = 256 + 226 // max value of r/g/b output +enum { + YUV_FIX = 16, // fixed-point precision for RGB->YUV + YUV_HALF = 1 << (YUV_FIX - 1), + YUV_MASK = (256 << YUV_FIX) - 1, + YUV_RANGE_MIN = -227, // min value of r/g/b output + YUV_RANGE_MAX = 256 + 226, // max value of r/g/b output + + YUV_FIX2 = 14, // fixed-point precision for YUV->RGB + YUV_HALF2 = 1 << (YUV_FIX2 - 1), + YUV_MASK2 = (256 << YUV_FIX2) - 1 }; + +// These constants are 14b fixed-point version of ITU-R BT.601 constants. +#define kYScale 19077 // 1.164 = 255 / 219 +#define kVToR 26149 // 1.596 = 255 / 112 * 0.701 +#define kUToG 6419 // 0.391 = 255 / 112 * 0.886 * 0.114 / 0.587 +#define kVToG 13320 // 0.813 = 255 / 112 * 0.701 * 0.299 / 0.587 +#define kUToB 33050 // 2.018 = 255 / 112 * 0.886 +#define kRCst (-kYScale * 16 - kVToR * 128 + YUV_HALF2) +#define kGCst (-kYScale * 16 + kUToG * 128 + kVToG * 128 + YUV_HALF2) +#define kBCst (-kYScale * 16 - kUToB * 128 + YUV_HALF2) + +//------------------------------------------------------------------------------ + +#if !defined(WEBP_YUV_USE_TABLE) + +// slower on x86 by ~7-8%, but bit-exact with the SSE2 version + +static WEBP_INLINE int VP8Clip8(int v) { + return ((v & ~YUV_MASK2) == 0) ? (v >> YUV_FIX2) : (v < 0) ? 0 : 255; +} + +static WEBP_INLINE int VP8YUVToR(int y, int v) { + return VP8Clip8(kYScale * y + kVToR * v + kRCst); +} + +static WEBP_INLINE int VP8YUVToG(int y, int u, int v) { + return VP8Clip8(kYScale * y - kUToG * u - kVToG * v + kGCst); +} + +static WEBP_INLINE int VP8YUVToB(int y, int u) { + return VP8Clip8(kYScale * y + kUToB * u + kBCst); +} + +static WEBP_INLINE void VP8YuvToRgb(int y, int u, int v, + uint8_t* const rgb) { + rgb[0] = VP8YUVToR(y, v); + rgb[1] = VP8YUVToG(y, u, v); + rgb[2] = VP8YUVToB(y, u); +} + +static WEBP_INLINE void VP8YuvToBgr(int y, int u, int v, + uint8_t* const bgr) { + bgr[0] = VP8YUVToB(y, u); + bgr[1] = VP8YUVToG(y, u, v); + bgr[2] = VP8YUVToR(y, v); +} + +static WEBP_INLINE void VP8YuvToRgb565(int y, int u, int v, + uint8_t* const rgb) { + const int r = VP8YUVToR(y, v); // 5 usable bits + const int g = VP8YUVToG(y, u, v); // 6 usable bits + const int b = VP8YUVToB(y, u); // 5 usable bits + const int rg = (r & 0xf8) | (g >> 5); + const int gb = ((g << 3) & 0xe0) | (b >> 3); +#ifdef WEBP_SWAP_16BIT_CSP + rgb[0] = gb; + rgb[1] = rg; +#else + rgb[0] = rg; + rgb[1] = gb; +#endif +} + +static WEBP_INLINE void VP8YuvToRgba4444(int y, int u, int v, + uint8_t* const argb) { + const int r = VP8YUVToR(y, v); // 4 usable bits + const int g = VP8YUVToG(y, u, v); // 4 usable bits + const int b = VP8YUVToB(y, u); // 4 usable bits + const int rg = (r & 0xf0) | (g >> 4); + const int ba = (b & 0xf0) | 0x0f; // overwrite the lower 4 bits +#ifdef WEBP_SWAP_16BIT_CSP + argb[0] = ba; + argb[1] = rg; +#else + argb[0] = rg; + argb[1] = ba; +#endif +} + +#else + +// Table-based version, not totally equivalent to the SSE2 version. +// Rounding diff is only +/-1 though. + extern int16_t VP8kVToR[256], VP8kUToB[256]; extern int32_t VP8kVToG[256], VP8kUToG[256]; extern uint8_t VP8kClip[YUV_RANGE_MAX - YUV_RANGE_MIN]; extern uint8_t VP8kClip4Bits[YUV_RANGE_MAX - YUV_RANGE_MIN]; -static WEBP_INLINE void VP8YuvToRgb(uint8_t y, uint8_t u, uint8_t v, +static WEBP_INLINE void VP8YuvToRgb(int y, int u, int v, uint8_t* const rgb) { const int r_off = VP8kVToR[v]; const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX; @@ -40,42 +169,60 @@ static WEBP_INLINE void VP8YuvToRgb(uint8_t y, uint8_t u, uint8_t v, rgb[2] = VP8kClip[y + b_off - YUV_RANGE_MIN]; } -static WEBP_INLINE void VP8YuvToRgb565(uint8_t y, uint8_t u, uint8_t v, - uint8_t* const rgb) { +static WEBP_INLINE void VP8YuvToBgr(int y, int u, int v, + uint8_t* const bgr) { const int r_off = VP8kVToR[v]; const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX; const int b_off = VP8kUToB[u]; - rgb[0] = ((VP8kClip[y + r_off - YUV_RANGE_MIN] & 0xf8) | - (VP8kClip[y + g_off - YUV_RANGE_MIN] >> 5)); - rgb[1] = (((VP8kClip[y + g_off - YUV_RANGE_MIN] << 3) & 0xe0) | - (VP8kClip[y + b_off - YUV_RANGE_MIN] >> 3)); -} - -static WEBP_INLINE void VP8YuvToArgb(uint8_t y, uint8_t u, uint8_t v, - uint8_t* const argb) { - argb[0] = 0xff; - VP8YuvToRgb(y, u, v, argb + 1); + bgr[0] = VP8kClip[y + b_off - YUV_RANGE_MIN]; + bgr[1] = VP8kClip[y + g_off - YUV_RANGE_MIN]; + bgr[2] = VP8kClip[y + r_off - YUV_RANGE_MIN]; } -static WEBP_INLINE void VP8YuvToRgba4444(uint8_t y, uint8_t u, uint8_t v, - uint8_t* const argb) { +static WEBP_INLINE void VP8YuvToRgb565(int y, int u, int v, + uint8_t* const rgb) { const int r_off = VP8kVToR[v]; const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX; const int b_off = VP8kUToB[u]; - // Don't update alpha (last 4 bits of argb[1]) - argb[0] = ((VP8kClip4Bits[y + r_off - YUV_RANGE_MIN] << 4) | - VP8kClip4Bits[y + g_off - YUV_RANGE_MIN]); - argb[1] = 0x0f | (VP8kClip4Bits[y + b_off - YUV_RANGE_MIN] << 4); + const int rg = ((VP8kClip[y + r_off - YUV_RANGE_MIN] & 0xf8) | + (VP8kClip[y + g_off - YUV_RANGE_MIN] >> 5)); + const int gb = (((VP8kClip[y + g_off - YUV_RANGE_MIN] << 3) & 0xe0) | + (VP8kClip[y + b_off - YUV_RANGE_MIN] >> 3)); +#ifdef WEBP_SWAP_16BIT_CSP + rgb[0] = gb; + rgb[1] = rg; +#else + rgb[0] = rg; + rgb[1] = gb; +#endif } -static WEBP_INLINE void VP8YuvToBgr(uint8_t y, uint8_t u, uint8_t v, - uint8_t* const bgr) { +static WEBP_INLINE void VP8YuvToRgba4444(int y, int u, int v, + uint8_t* const argb) { const int r_off = VP8kVToR[v]; const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX; const int b_off = VP8kUToB[u]; - bgr[0] = VP8kClip[y + b_off - YUV_RANGE_MIN]; - bgr[1] = VP8kClip[y + g_off - YUV_RANGE_MIN]; - bgr[2] = VP8kClip[y + r_off - YUV_RANGE_MIN]; + const int rg = ((VP8kClip4Bits[y + r_off - YUV_RANGE_MIN] << 4) | + VP8kClip4Bits[y + g_off - YUV_RANGE_MIN]); + const int ba = (VP8kClip4Bits[y + b_off - YUV_RANGE_MIN] << 4) | 0x0f; +#ifdef WEBP_SWAP_16BIT_CSP + argb[0] = ba; + argb[1] = rg; +#else + argb[0] = rg; + argb[1] = ba; +#endif +} + +#endif // WEBP_YUV_USE_TABLE + +//----------------------------------------------------------------------------- +// Alpha handling variants + +static WEBP_INLINE void VP8YuvToArgb(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const argb) { + argb[0] = 0xff; + VP8YuvToRgb(y, u, v, argb + 1); } static WEBP_INLINE void VP8YuvToBgra(uint8_t y, uint8_t u, uint8_t v, @@ -93,35 +240,79 @@ static WEBP_INLINE void VP8YuvToRgba(uint8_t y, uint8_t u, uint8_t v, // Must be called before everything, to initialize the tables. void VP8YUVInit(void); +//----------------------------------------------------------------------------- +// SSE2 extra functions (mostly for upsampling_sse2.c) + +#if defined(WEBP_USE_SSE2) + +// When the following is defined, tables are initialized statically, adding ~12k +// to the binary size. Otherwise, they are initialized at run-time (small cost). +#define WEBP_YUV_USE_SSE2_TABLES + +// Process 32 pixels and store the result (24b or 32b per pixel) in *dst. +void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst); +void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst); +void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst); +void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst); + +// Must be called to initialize tables before using the functions. +void VP8YUVInitSSE2(void); + +#endif // WEBP_USE_SSE2 + //------------------------------------------------------------------------------ // RGB -> YUV conversion -// The exact naming is Y'CbCr, following the ITU-R BT.601 standard. -// More information at: http://en.wikipedia.org/wiki/YCbCr -// Y = 0.2569 * R + 0.5044 * G + 0.0979 * B + 16 -// U = -0.1483 * R - 0.2911 * G + 0.4394 * B + 128 -// V = 0.4394 * R - 0.3679 * G - 0.0715 * B + 128 -// We use 16bit fixed point operations. -static WEBP_INLINE int VP8ClipUV(int v) { - v = (v + (257 << (YUV_FIX + 2 - 1))) >> (YUV_FIX + 2); - return ((v & ~0xff) == 0) ? v : (v < 0) ? 0 : 255; +// Stub functions that can be called with various rounding values: +static WEBP_INLINE int VP8ClipUV(int uv, int rounding) { + uv = (uv + rounding + (128 << (YUV_FIX + 2))) >> (YUV_FIX + 2); + return ((uv & ~0xff) == 0) ? uv : (uv < 0) ? 0 : 255; } -static WEBP_INLINE int VP8RGBToY(int r, int g, int b) { - const int kRound = (1 << (YUV_FIX - 1)) + (16 << YUV_FIX); +#ifndef USE_YUVj + +static WEBP_INLINE int VP8RGBToY(int r, int g, int b, int rounding) { const int luma = 16839 * r + 33059 * g + 6420 * b; - return (luma + kRound) >> YUV_FIX; // no need to clip + return (luma + rounding + (16 << YUV_FIX)) >> YUV_FIX; // no need to clip } -static WEBP_INLINE int VP8RGBToU(int r, int g, int b) { - return VP8ClipUV(-9719 * r - 19081 * g + 28800 * b); +static WEBP_INLINE int VP8RGBToU(int r, int g, int b, int rounding) { + const int u = -9719 * r - 19081 * g + 28800 * b; + return VP8ClipUV(u, rounding); } -static WEBP_INLINE int VP8RGBToV(int r, int g, int b) { - return VP8ClipUV(+28800 * r - 24116 * g - 4684 * b); +static WEBP_INLINE int VP8RGBToV(int r, int g, int b, int rounding) { + const int v = +28800 * r - 24116 * g - 4684 * b; + return VP8ClipUV(v, rounding); +} + +#else + +// This JPEG-YUV colorspace, only for comparison! +// These are also 16bit precision coefficients from Rec.601, but with full +// [0..255] output range. +static WEBP_INLINE int VP8RGBToY(int r, int g, int b, int rounding) { + const int luma = 19595 * r + 38470 * g + 7471 * b; + return (luma + rounding) >> YUV_FIX; // no need to clip } -#if defined(__cplusplus) || defined(c_plusplus) +static WEBP_INLINE int VP8RGBToU(int r, int g, int b, int rounding) { + const int u = -11058 * r - 21710 * g + 32768 * b; + return VP8ClipUV(u, rounding); +} + +static WEBP_INLINE int VP8RGBToV(int r, int g, int b, int rounding) { + const int v = 32768 * r - 27439 * g - 5329 * b; + return VP8ClipUV(v, rounding); +} + +#endif // USE_YUVj + +#ifdef __cplusplus } // extern "C" #endif diff --git a/drivers/webp/dsp/yuv_mips32.c b/drivers/webp/dsp/yuv_mips32.c new file mode 100644 index 0000000000..018f8ab774 --- /dev/null +++ b/drivers/webp/dsp/yuv_mips32.c @@ -0,0 +1,103 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of YUV to RGB upsampling functions. +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include "./yuv.h" + +//------------------------------------------------------------------------------ +// simple point-sampling + +#define ROW_FUNC(FUNC_NAME, XSTEP, R, G, B, A) \ +static void FUNC_NAME(const uint8_t* y, \ + const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ + int i, r, g, b; \ + int temp0, temp1, temp2, temp3, temp4; \ + for (i = 0; i < (len >> 1); i++) { \ + temp1 = kVToR * v[0]; \ + temp3 = kVToG * v[0]; \ + temp2 = kUToG * u[0]; \ + temp4 = kUToB * u[0]; \ + temp0 = kYScale * y[0]; \ + temp1 += kRCst; \ + temp3 -= kGCst; \ + temp2 += temp3; \ + temp4 += kBCst; \ + r = VP8Clip8(temp0 + temp1); \ + g = VP8Clip8(temp0 - temp2); \ + b = VP8Clip8(temp0 + temp4); \ + temp0 = kYScale * y[1]; \ + dst[R] = r; \ + dst[G] = g; \ + dst[B] = b; \ + if (A) dst[A] = 0xff; \ + r = VP8Clip8(temp0 + temp1); \ + g = VP8Clip8(temp0 - temp2); \ + b = VP8Clip8(temp0 + temp4); \ + dst[R + XSTEP] = r; \ + dst[G + XSTEP] = g; \ + dst[B + XSTEP] = b; \ + if (A) dst[A + XSTEP] = 0xff; \ + y += 2; \ + ++u; \ + ++v; \ + dst += 2 * XSTEP; \ + } \ + if (len & 1) { \ + temp1 = kVToR * v[0]; \ + temp3 = kVToG * v[0]; \ + temp2 = kUToG * u[0]; \ + temp4 = kUToB * u[0]; \ + temp0 = kYScale * y[0]; \ + temp1 += kRCst; \ + temp3 -= kGCst; \ + temp2 += temp3; \ + temp4 += kBCst; \ + r = VP8Clip8(temp0 + temp1); \ + g = VP8Clip8(temp0 - temp2); \ + b = VP8Clip8(temp0 + temp4); \ + dst[R] = r; \ + dst[G] = g; \ + dst[B] = b; \ + if (A) dst[A] = 0xff; \ + } \ +} + +ROW_FUNC(YuvToRgbRow, 3, 0, 1, 2, 0) +ROW_FUNC(YuvToRgbaRow, 4, 0, 1, 2, 3) +ROW_FUNC(YuvToBgrRow, 3, 2, 1, 0, 0) +ROW_FUNC(YuvToBgraRow, 4, 2, 1, 0, 3) + +#undef ROW_FUNC + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitSamplersMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersMIPS32(void) { + WebPSamplers[MODE_RGB] = YuvToRgbRow; + WebPSamplers[MODE_RGBA] = YuvToRgbaRow; + WebPSamplers[MODE_BGR] = YuvToBgrRow; + WebPSamplers[MODE_BGRA] = YuvToBgraRow; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(WebPInitSamplersMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/drivers/webp/dsp/yuv_mips_dsp_r2.c b/drivers/webp/dsp/yuv_mips_dsp_r2.c new file mode 100644 index 0000000000..45a2200352 --- /dev/null +++ b/drivers/webp/dsp/yuv_mips_dsp_r2.c @@ -0,0 +1,134 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS DSPr2 version of YUV to RGB upsampling functions. +// +// Author(s): Branimir Vasic (branimir.vasic@imgtec.com) +// Djordje Pesut (djordje.pesut@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "./yuv.h" + +//------------------------------------------------------------------------------ +// simple point-sampling + +#define ROW_FUNC_PART_1() \ + "lbu %[temp3], 0(%[v]) \n\t" \ + "lbu %[temp4], 0(%[u]) \n\t" \ + "lbu %[temp0], 0(%[y]) \n\t" \ + "mul %[temp1], %[t_con_1], %[temp3] \n\t" \ + "mul %[temp3], %[t_con_2], %[temp3] \n\t" \ + "mul %[temp2], %[t_con_3], %[temp4] \n\t" \ + "mul %[temp4], %[t_con_4], %[temp4] \n\t" \ + "mul %[temp0], %[t_con_5], %[temp0] \n\t" \ + "addu %[temp1], %[temp1], %[t_con_6] \n\t" \ + "subu %[temp3], %[temp3], %[t_con_7] \n\t" \ + "addu %[temp2], %[temp2], %[temp3] \n\t" \ + "addu %[temp4], %[temp4], %[t_con_8] \n\t" \ + +#define ROW_FUNC_PART_2(R, G, B, K) \ + "addu %[temp5], %[temp0], %[temp1] \n\t" \ + "subu %[temp6], %[temp0], %[temp2] \n\t" \ + "addu %[temp7], %[temp0], %[temp4] \n\t" \ +".if " #K " \n\t" \ + "lbu %[temp0], 1(%[y]) \n\t" \ +".endif \n\t" \ + "shll_s.w %[temp5], %[temp5], 9 \n\t" \ + "shll_s.w %[temp6], %[temp6], 9 \n\t" \ +".if " #K " \n\t" \ + "mul %[temp0], %[t_con_5], %[temp0] \n\t" \ +".endif \n\t" \ + "shll_s.w %[temp7], %[temp7], 9 \n\t" \ + "precrqu_s.qb.ph %[temp5], %[temp5], $zero \n\t" \ + "precrqu_s.qb.ph %[temp6], %[temp6], $zero \n\t" \ + "precrqu_s.qb.ph %[temp7], %[temp7], $zero \n\t" \ + "srl %[temp5], %[temp5], 24 \n\t" \ + "srl %[temp6], %[temp6], 24 \n\t" \ + "srl %[temp7], %[temp7], 24 \n\t" \ + "sb %[temp5], " #R "(%[dst]) \n\t" \ + "sb %[temp6], " #G "(%[dst]) \n\t" \ + "sb %[temp7], " #B "(%[dst]) \n\t" \ + +#define ASM_CLOBBER_LIST() \ + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), \ + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), \ + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7) \ + : [t_con_1]"r"(t_con_1), [t_con_2]"r"(t_con_2), [t_con_3]"r"(t_con_3), \ + [t_con_4]"r"(t_con_4), [t_con_5]"r"(t_con_5), [t_con_6]"r"(t_con_6), \ + [u]"r"(u), [v]"r"(v), [y]"r"(y), [dst]"r"(dst), \ + [t_con_7]"r"(t_con_7), [t_con_8]"r"(t_con_8) \ + : "memory", "hi", "lo" \ + +#define ROW_FUNC(FUNC_NAME, XSTEP, R, G, B, A) \ +static void FUNC_NAME(const uint8_t* y, \ + const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ + int i; \ + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; \ + const int t_con_1 = kVToR; \ + const int t_con_2 = kVToG; \ + const int t_con_3 = kUToG; \ + const int t_con_4 = kUToB; \ + const int t_con_5 = kYScale; \ + const int t_con_6 = kRCst; \ + const int t_con_7 = kGCst; \ + const int t_con_8 = kBCst; \ + for (i = 0; i < (len >> 1); i++) { \ + __asm__ volatile ( \ + ROW_FUNC_PART_1() \ + ROW_FUNC_PART_2(R, G, B, 1) \ + ROW_FUNC_PART_2(R + XSTEP, G + XSTEP, B + XSTEP, 0) \ + ASM_CLOBBER_LIST() \ + ); \ + if (A) dst[A] = dst[A + XSTEP] = 0xff; \ + y += 2; \ + ++u; \ + ++v; \ + dst += 2 * XSTEP; \ + } \ + if (len & 1) { \ + __asm__ volatile ( \ + ROW_FUNC_PART_1() \ + ROW_FUNC_PART_2(R, G, B, 0) \ + ASM_CLOBBER_LIST() \ + ); \ + if (A) dst[A] = 0xff; \ + } \ +} + +ROW_FUNC(YuvToRgbRow, 3, 0, 1, 2, 0) +ROW_FUNC(YuvToRgbaRow, 4, 0, 1, 2, 3) +ROW_FUNC(YuvToBgrRow, 3, 2, 1, 0, 0) +ROW_FUNC(YuvToBgraRow, 4, 2, 1, 0, 3) + +#undef ROW_FUNC +#undef ASM_CLOBBER_LIST +#undef ROW_FUNC_PART_2 +#undef ROW_FUNC_PART_1 + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitSamplersMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersMIPSdspR2(void) { + WebPSamplers[MODE_RGB] = YuvToRgbRow; + WebPSamplers[MODE_RGBA] = YuvToRgbaRow; + WebPSamplers[MODE_BGR] = YuvToBgrRow; + WebPSamplers[MODE_BGRA] = YuvToBgraRow; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(WebPInitSamplersMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/drivers/webp/dsp/yuv_sse2.c b/drivers/webp/dsp/yuv_sse2.c new file mode 100644 index 0000000000..283b3af228 --- /dev/null +++ b/drivers/webp/dsp/yuv_sse2.c @@ -0,0 +1,606 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// YUV->RGB conversion functions +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./yuv.h" + +#if defined(WEBP_USE_SSE2) + +#include <emmintrin.h> +#include <string.h> // for memcpy + +typedef union { // handy struct for converting SSE2 registers + int32_t i32[4]; + uint8_t u8[16]; + __m128i m; +} VP8kCstSSE2; + +#if defined(WEBP_YUV_USE_SSE2_TABLES) + +#include "./yuv_tables_sse2.h" + +WEBP_TSAN_IGNORE_FUNCTION void VP8YUVInitSSE2(void) {} + +#else + +static int done_sse2 = 0; +static VP8kCstSSE2 VP8kUtoRGBA[256], VP8kVtoRGBA[256], VP8kYtoRGBA[256]; + +WEBP_TSAN_IGNORE_FUNCTION void VP8YUVInitSSE2(void) { + if (!done_sse2) { + int i; + for (i = 0; i < 256; ++i) { + VP8kYtoRGBA[i].i32[0] = + VP8kYtoRGBA[i].i32[1] = + VP8kYtoRGBA[i].i32[2] = (i - 16) * kYScale + YUV_HALF2; + VP8kYtoRGBA[i].i32[3] = 0xff << YUV_FIX2; + + VP8kUtoRGBA[i].i32[0] = 0; + VP8kUtoRGBA[i].i32[1] = -kUToG * (i - 128); + VP8kUtoRGBA[i].i32[2] = kUToB * (i - 128); + VP8kUtoRGBA[i].i32[3] = 0; + + VP8kVtoRGBA[i].i32[0] = kVToR * (i - 128); + VP8kVtoRGBA[i].i32[1] = -kVToG * (i - 128); + VP8kVtoRGBA[i].i32[2] = 0; + VP8kVtoRGBA[i].i32[3] = 0; + } + done_sse2 = 1; + +#if 0 // code used to generate 'yuv_tables_sse2.h' + printf("static const VP8kCstSSE2 VP8kYtoRGBA[256] = {\n"); + for (i = 0; i < 256; ++i) { + printf(" {{0x%.8x, 0x%.8x, 0x%.8x, 0x%.8x}},\n", + VP8kYtoRGBA[i].i32[0], VP8kYtoRGBA[i].i32[1], + VP8kYtoRGBA[i].i32[2], VP8kYtoRGBA[i].i32[3]); + } + printf("};\n\n"); + printf("static const VP8kCstSSE2 VP8kUtoRGBA[256] = {\n"); + for (i = 0; i < 256; ++i) { + printf(" {{0, 0x%.8x, 0x%.8x, 0}},\n", + VP8kUtoRGBA[i].i32[1], VP8kUtoRGBA[i].i32[2]); + } + printf("};\n\n"); + printf("static VP8kCstSSE2 VP8kVtoRGBA[256] = {\n"); + for (i = 0; i < 256; ++i) { + printf(" {{0x%.8x, 0x%.8x, 0, 0}},\n", + VP8kVtoRGBA[i].i32[0], VP8kVtoRGBA[i].i32[1]); + } + printf("};\n\n"); +#endif + } +} + +#endif // WEBP_YUV_USE_SSE2_TABLES + +//----------------------------------------------------------------------------- + +static WEBP_INLINE __m128i LoadUVPart(int u, int v) { + const __m128i u_part = _mm_loadu_si128(&VP8kUtoRGBA[u].m); + const __m128i v_part = _mm_loadu_si128(&VP8kVtoRGBA[v].m); + const __m128i uv_part = _mm_add_epi32(u_part, v_part); + return uv_part; +} + +static WEBP_INLINE __m128i GetRGBA32bWithUV(int y, const __m128i uv_part) { + const __m128i y_part = _mm_loadu_si128(&VP8kYtoRGBA[y].m); + const __m128i rgba1 = _mm_add_epi32(y_part, uv_part); + const __m128i rgba2 = _mm_srai_epi32(rgba1, YUV_FIX2); + return rgba2; +} + +static WEBP_INLINE __m128i GetRGBA32b(int y, int u, int v) { + const __m128i uv_part = LoadUVPart(u, v); + return GetRGBA32bWithUV(y, uv_part); +} + +static WEBP_INLINE void YuvToRgbSSE2(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const rgb) { + const __m128i tmp0 = GetRGBA32b(y, u, v); + const __m128i tmp1 = _mm_packs_epi32(tmp0, tmp0); + const __m128i tmp2 = _mm_packus_epi16(tmp1, tmp1); + // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp + _mm_storel_epi64((__m128i*)rgb, tmp2); +} + +static WEBP_INLINE void YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const bgr) { + const __m128i tmp0 = GetRGBA32b(y, u, v); + const __m128i tmp1 = _mm_shuffle_epi32(tmp0, _MM_SHUFFLE(3, 0, 1, 2)); + const __m128i tmp2 = _mm_packs_epi32(tmp1, tmp1); + const __m128i tmp3 = _mm_packus_epi16(tmp2, tmp2); + // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp + _mm_storel_epi64((__m128i*)bgr, tmp3); +} + +//----------------------------------------------------------------------------- +// Convert spans of 32 pixels to various RGB formats for the fancy upsampler. + +void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst) { + int n; + for (n = 0; n < 32; n += 4) { + const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]); + const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]); + const __m128i tmp0_3 = GetRGBA32b(y[n + 2], u[n + 2], v[n + 2]); + const __m128i tmp0_4 = GetRGBA32b(y[n + 3], u[n + 3], v[n + 3]); + const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2); + const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4); + const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2); + _mm_storeu_si128((__m128i*)dst, tmp2); + dst += 4 * 4; + } +} + +void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst) { + int n; + for (n = 0; n < 32; n += 2) { + const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]); + const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]); + const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2)); + const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2)); + const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2); + const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1); + _mm_storel_epi64((__m128i*)dst, tmp3); + dst += 4 * 2; + } +} + +void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst) { + int n; + uint8_t tmp0[2 * 3 + 5 + 15]; + uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15); // align + for (n = 0; n < 30; ++n) { // we directly stomp the *dst memory + YuvToRgbSSE2(y[n], u[n], v[n], dst + n * 3); + } + // Last two pixels are special: we write in a tmp buffer before sending + // to dst. + YuvToRgbSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0); + YuvToRgbSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3); + memcpy(dst + n * 3, tmp, 2 * 3); +} + +void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst) { + int n; + uint8_t tmp0[2 * 3 + 5 + 15]; + uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15); // align + for (n = 0; n < 30; ++n) { + YuvToBgrSSE2(y[n], u[n], v[n], dst + n * 3); + } + YuvToBgrSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0); + YuvToBgrSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3); + memcpy(dst + n * 3, tmp, 2 * 3); +} + +//----------------------------------------------------------------------------- +// Arbitrary-length row conversion functions + +static void YuvToRgbaRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst, int len) { + int n; + for (n = 0; n + 4 <= len; n += 4) { + const __m128i uv_0 = LoadUVPart(u[0], v[0]); + const __m128i uv_1 = LoadUVPart(u[1], v[1]); + const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0); + const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0); + const __m128i tmp0_3 = GetRGBA32bWithUV(y[2], uv_1); + const __m128i tmp0_4 = GetRGBA32bWithUV(y[3], uv_1); + const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2); + const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4); + const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2); + _mm_storeu_si128((__m128i*)dst, tmp2); + dst += 4 * 4; + y += 4; + u += 2; + v += 2; + } + // Finish off + while (n < len) { + VP8YuvToRgba(y[0], u[0], v[0], dst); + dst += 4; + ++y; + u += (n & 1); + v += (n & 1); + ++n; + } +} + +static void YuvToBgraRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst, int len) { + int n; + for (n = 0; n + 2 <= len; n += 2) { + const __m128i uv_0 = LoadUVPart(u[0], v[0]); + const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0); + const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0); + const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2)); + const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2)); + const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2); + const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1); + _mm_storel_epi64((__m128i*)dst, tmp3); + dst += 4 * 2; + y += 2; + ++u; + ++v; + } + // Finish off + if (len & 1) { + VP8YuvToBgra(y[0], u[0], v[0], dst); + } +} + +static void YuvToArgbRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst, int len) { + int n; + for (n = 0; n + 2 <= len; n += 2) { + const __m128i uv_0 = LoadUVPart(u[0], v[0]); + const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0); + const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0); + const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(2, 1, 0, 3)); + const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(2, 1, 0, 3)); + const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2); + const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1); + _mm_storel_epi64((__m128i*)dst, tmp3); + dst += 4 * 2; + y += 2; + ++u; + ++v; + } + // Finish off + if (len & 1) { + VP8YuvToArgb(y[0], u[0], v[0], dst); + } +} + +static void YuvToRgbRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst, int len) { + int n; + for (n = 0; n + 2 < len; ++n) { // we directly stomp the *dst memory + YuvToRgbSSE2(y[0], u[0], v[0], dst); // stomps 8 bytes + dst += 3; + ++y; + u += (n & 1); + v += (n & 1); + } + VP8YuvToRgb(y[0], u[0], v[0], dst); + if (len > 1) { + VP8YuvToRgb(y[1], u[n & 1], v[n & 1], dst + 3); + } +} + +static void YuvToBgrRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst, int len) { + int n; + for (n = 0; n + 2 < len; ++n) { // we directly stomp the *dst memory + YuvToBgrSSE2(y[0], u[0], v[0], dst); // stomps 8 bytes + dst += 3; + ++y; + u += (n & 1); + v += (n & 1); + } + VP8YuvToBgr(y[0], u[0], v[0], dst + 0); + if (len > 1) { + VP8YuvToBgr(y[1], u[n & 1], v[n & 1], dst + 3); + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitSamplersSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersSSE2(void) { + WebPSamplers[MODE_RGB] = YuvToRgbRow; + WebPSamplers[MODE_RGBA] = YuvToRgbaRow; + WebPSamplers[MODE_BGR] = YuvToBgrRow; + WebPSamplers[MODE_BGRA] = YuvToBgraRow; + WebPSamplers[MODE_ARGB] = YuvToArgbRow; +} + +//------------------------------------------------------------------------------ +// RGB24/32 -> YUV converters + +// Load eight 16b-words from *src. +#define LOAD_16(src) _mm_loadu_si128((const __m128i*)(src)) +// Store either 16b-words into *dst +#define STORE_16(V, dst) _mm_storeu_si128((__m128i*)(dst), (V)) + +// Convert 8 packed RGB or BGR samples to r[], g[], b[] +static WEBP_INLINE void RGB24PackedToPlanar(const uint8_t* const rgb, + __m128i* const r, + __m128i* const g, + __m128i* const b, + int input_is_bgr) { + const __m128i zero = _mm_setzero_si128(); + // in0: r0 g0 b0 r1 | g1 b1 r2 g2 | b2 r3 g3 b3 | r4 g4 b4 r5 + // in1: b2 r3 g3 b3 | r4 g4 b4 r5 | g5 b5 r6 g6 | b6 r7 g7 b7 + const __m128i in0 = LOAD_16(rgb + 0); + const __m128i in1 = LOAD_16(rgb + 8); + // A0: | r2 g2 b2 r3 | g3 b3 r4 g4 | b4 r5 ... + // A1: ... b2 r3 | g3 b3 r4 g4 | b4 r5 g5 b5 | + const __m128i A0 = _mm_srli_si128(in0, 6); + const __m128i A1 = _mm_slli_si128(in1, 6); + // B0: r0 r2 g0 g2 | b0 b2 r1 r3 | g1 g3 b1 b3 | r2 r4 b2 b4 + // B1: g3 g5 b3 b5 | r4 r6 g4 g6 | b4 b6 r5 r7 | g5 g7 b5 b7 + const __m128i B0 = _mm_unpacklo_epi8(in0, A0); + const __m128i B1 = _mm_unpackhi_epi8(A1, in1); + // C0: r1 r3 g1 g3 | b1 b3 r2 r4 | b2 b4 ... + // C1: ... g3 g5 | b3 b5 r4 r6 | g4 g6 b4 b6 + const __m128i C0 = _mm_srli_si128(B0, 6); + const __m128i C1 = _mm_slli_si128(B1, 6); + // D0: r0 r1 r2 r3 | g0 g1 g2 g3 | b0 b1 b2 b3 | r1 r2 r3 r4 + // D1: b3 b4 b5 b6 | r4 r5 r6 r7 | g4 g5 g6 g7 | b4 b5 b6 b7 | + const __m128i D0 = _mm_unpacklo_epi8(B0, C0); + const __m128i D1 = _mm_unpackhi_epi8(C1, B1); + // r4 r5 r6 r7 | g4 g5 g6 g7 | b4 b5 b6 b7 | 0 + const __m128i D2 = _mm_srli_si128(D1, 4); + // r0 r1 r2 r3 | r4 r5 r6 r7 | g0 g1 g2 g3 | g4 g5 g6 g7 + const __m128i E0 = _mm_unpacklo_epi32(D0, D2); + // b0 b1 b2 b3 | b4 b5 b6 b7 | r1 r2 r3 r4 | 0 + const __m128i E1 = _mm_unpackhi_epi32(D0, D2); + // g0 g1 g2 g3 | g4 g5 g6 g7 | 0 + const __m128i E2 = _mm_srli_si128(E0, 8); + const __m128i F0 = _mm_unpacklo_epi8(E0, zero); // -> R + const __m128i F1 = _mm_unpacklo_epi8(E1, zero); // -> B + const __m128i F2 = _mm_unpacklo_epi8(E2, zero); // -> G + *g = F2; + if (input_is_bgr) { + *r = F1; + *b = F0; + } else { + *r = F0; + *b = F1; + } +} + +// Convert 8 packed ARGB to r[], g[], b[] +static WEBP_INLINE void RGB32PackedToPlanar(const uint32_t* const argb, + __m128i* const r, + __m128i* const g, + __m128i* const b) { + const __m128i zero = _mm_setzero_si128(); + const __m128i in0 = LOAD_16(argb + 0); // argb3 | argb2 | argb1 | argb0 + const __m128i in1 = LOAD_16(argb + 4); // argb7 | argb6 | argb5 | argb4 + // column-wise transpose + const __m128i A0 = _mm_unpacklo_epi8(in0, in1); + const __m128i A1 = _mm_unpackhi_epi8(in0, in1); + const __m128i B0 = _mm_unpacklo_epi8(A0, A1); + const __m128i B1 = _mm_unpackhi_epi8(A0, A1); + // C0 = g7 g6 ... g1 g0 | b7 b6 ... b1 b0 + // C1 = a7 a6 ... a1 a0 | r7 r6 ... r1 r0 + const __m128i C0 = _mm_unpacklo_epi8(B0, B1); + const __m128i C1 = _mm_unpackhi_epi8(B0, B1); + // store 16b + *r = _mm_unpacklo_epi8(C1, zero); + *g = _mm_unpackhi_epi8(C0, zero); + *b = _mm_unpacklo_epi8(C0, zero); +} + +// This macro computes (RG * MULT_RG + GB * MULT_GB + ROUNDER) >> DESCALE_FIX +// It's a macro and not a function because we need to use immediate values with +// srai_epi32, e.g. +#define TRANSFORM(RG_LO, RG_HI, GB_LO, GB_HI, MULT_RG, MULT_GB, \ + ROUNDER, DESCALE_FIX, OUT) do { \ + const __m128i V0_lo = _mm_madd_epi16(RG_LO, MULT_RG); \ + const __m128i V0_hi = _mm_madd_epi16(RG_HI, MULT_RG); \ + const __m128i V1_lo = _mm_madd_epi16(GB_LO, MULT_GB); \ + const __m128i V1_hi = _mm_madd_epi16(GB_HI, MULT_GB); \ + const __m128i V2_lo = _mm_add_epi32(V0_lo, V1_lo); \ + const __m128i V2_hi = _mm_add_epi32(V0_hi, V1_hi); \ + const __m128i V3_lo = _mm_add_epi32(V2_lo, ROUNDER); \ + const __m128i V3_hi = _mm_add_epi32(V2_hi, ROUNDER); \ + const __m128i V5_lo = _mm_srai_epi32(V3_lo, DESCALE_FIX); \ + const __m128i V5_hi = _mm_srai_epi32(V3_hi, DESCALE_FIX); \ + (OUT) = _mm_packs_epi32(V5_lo, V5_hi); \ +} while (0) + +#define MK_CST_16(A, B) _mm_set_epi16((B), (A), (B), (A), (B), (A), (B), (A)) +static WEBP_INLINE void ConvertRGBToY(const __m128i* const R, + const __m128i* const G, + const __m128i* const B, + __m128i* const Y) { + const __m128i kRG_y = MK_CST_16(16839, 33059 - 16384); + const __m128i kGB_y = MK_CST_16(16384, 6420); + const __m128i kHALF_Y = _mm_set1_epi32((16 << YUV_FIX) + YUV_HALF); + + const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G); + const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G); + const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B); + const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B); + TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_y, kGB_y, kHALF_Y, YUV_FIX, *Y); +} + +static WEBP_INLINE void ConvertRGBToUV(const __m128i* const R, + const __m128i* const G, + const __m128i* const B, + __m128i* const U, __m128i* const V) { + const __m128i kRG_u = MK_CST_16(-9719, -19081); + const __m128i kGB_u = MK_CST_16(0, 28800); + const __m128i kRG_v = MK_CST_16(28800, 0); + const __m128i kGB_v = MK_CST_16(-24116, -4684); + const __m128i kHALF_UV = _mm_set1_epi32(((128 << YUV_FIX) + YUV_HALF) << 2); + + const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G); + const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G); + const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B); + const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B); + TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_u, kGB_u, + kHALF_UV, YUV_FIX + 2, *U); + TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_v, kGB_v, + kHALF_UV, YUV_FIX + 2, *V); +} + +#undef MK_CST_16 +#undef TRANSFORM + +static void ConvertRGB24ToY(const uint8_t* rgb, uint8_t* y, int width) { + const int max_width = width & ~15; + int i; + for (i = 0; i < max_width; i += 16, rgb += 3 * 16) { + __m128i r, g, b, Y0, Y1; + RGB24PackedToPlanar(rgb + 0 * 8, &r, &g, &b, 0); + ConvertRGBToY(&r, &g, &b, &Y0); + RGB24PackedToPlanar(rgb + 3 * 8, &r, &g, &b, 0); + ConvertRGBToY(&r, &g, &b, &Y1); + STORE_16(_mm_packus_epi16(Y0, Y1), y + i); + } + for (; i < width; ++i, rgb += 3) { // left-over + y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF); + } +} + +static void ConvertBGR24ToY(const uint8_t* bgr, uint8_t* y, int width) { + int i; + const int max_width = width & ~15; + for (i = 0; i < max_width; i += 16, bgr += 3 * 16) { + __m128i r, g, b, Y0, Y1; + RGB24PackedToPlanar(bgr + 0 * 8, &r, &g, &b, 1); + ConvertRGBToY(&r, &g, &b, &Y0); + RGB24PackedToPlanar(bgr + 3 * 8, &r, &g, &b, 1); + ConvertRGBToY(&r, &g, &b, &Y1); + STORE_16(_mm_packus_epi16(Y0, Y1), y + i); + } + for (; i < width; ++i, bgr += 3) { // left-over + y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF); + } +} + +static void ConvertARGBToY(const uint32_t* argb, uint8_t* y, int width) { + const int max_width = width & ~15; + int i; + for (i = 0; i < max_width; i += 16) { + __m128i r, g, b, Y0, Y1; + RGB32PackedToPlanar(&argb[i + 0], &r, &g, &b); + ConvertRGBToY(&r, &g, &b, &Y0); + RGB32PackedToPlanar(&argb[i + 8], &r, &g, &b); + ConvertRGBToY(&r, &g, &b, &Y1); + STORE_16(_mm_packus_epi16(Y0, Y1), y + i); + } + for (; i < width; ++i) { // left-over + const uint32_t p = argb[i]; + y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff, + YUV_HALF); + } +} + +// Horizontal add (doubled) of two 16b values, result is 16b. +// in: A | B | C | D | ... -> out: 2*(A+B) | 2*(C+D) | ... +static void HorizontalAddPack(const __m128i* const A, const __m128i* const B, + __m128i* const out) { + const __m128i k2 = _mm_set1_epi16(2); + const __m128i C = _mm_madd_epi16(*A, k2); + const __m128i D = _mm_madd_epi16(*B, k2); + *out = _mm_packs_epi32(C, D); +} + +static void ConvertARGBToUV(const uint32_t* argb, uint8_t* u, uint8_t* v, + int src_width, int do_store) { + const int max_width = src_width & ~31; + int i; + for (i = 0; i < max_width; i += 32, u += 16, v += 16) { + __m128i r0, g0, b0, r1, g1, b1, U0, V0, U1, V1; + RGB32PackedToPlanar(&argb[i + 0], &r0, &g0, &b0); + RGB32PackedToPlanar(&argb[i + 8], &r1, &g1, &b1); + HorizontalAddPack(&r0, &r1, &r0); + HorizontalAddPack(&g0, &g1, &g0); + HorizontalAddPack(&b0, &b1, &b0); + ConvertRGBToUV(&r0, &g0, &b0, &U0, &V0); + + RGB32PackedToPlanar(&argb[i + 16], &r0, &g0, &b0); + RGB32PackedToPlanar(&argb[i + 24], &r1, &g1, &b1); + HorizontalAddPack(&r0, &r1, &r0); + HorizontalAddPack(&g0, &g1, &g0); + HorizontalAddPack(&b0, &b1, &b0); + ConvertRGBToUV(&r0, &g0, &b0, &U1, &V1); + + U0 = _mm_packus_epi16(U0, U1); + V0 = _mm_packus_epi16(V0, V1); + if (!do_store) { + const __m128i prev_u = LOAD_16(u); + const __m128i prev_v = LOAD_16(v); + U0 = _mm_avg_epu8(U0, prev_u); + V0 = _mm_avg_epu8(V0, prev_v); + } + STORE_16(U0, u); + STORE_16(V0, v); + } + if (i < src_width) { // left-over + WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store); + } +} + +// Convert 16 packed ARGB 16b-values to r[], g[], b[] +static WEBP_INLINE void RGBA32PackedToPlanar_16b(const uint16_t* const rgbx, + __m128i* const r, + __m128i* const g, + __m128i* const b) { + const __m128i in0 = LOAD_16(rgbx + 0); // r0 | g0 | b0 |x| r1 | g1 | b1 |x + const __m128i in1 = LOAD_16(rgbx + 8); // r2 | g2 | b2 |x| r3 | g3 | b3 |x + const __m128i in2 = LOAD_16(rgbx + 16); // r4 | ... + const __m128i in3 = LOAD_16(rgbx + 24); // r6 | ... + // column-wise transpose + const __m128i A0 = _mm_unpacklo_epi16(in0, in1); + const __m128i A1 = _mm_unpackhi_epi16(in0, in1); + const __m128i A2 = _mm_unpacklo_epi16(in2, in3); + const __m128i A3 = _mm_unpackhi_epi16(in2, in3); + const __m128i B0 = _mm_unpacklo_epi16(A0, A1); // r0 r1 r2 r3 | g0 g1 .. + const __m128i B1 = _mm_unpackhi_epi16(A0, A1); // b0 b1 b2 b3 | x x x x + const __m128i B2 = _mm_unpacklo_epi16(A2, A3); // r4 r5 r6 r7 | g4 g5 .. + const __m128i B3 = _mm_unpackhi_epi16(A2, A3); // b4 b5 b6 b7 | x x x x + *r = _mm_unpacklo_epi64(B0, B2); + *g = _mm_unpackhi_epi64(B0, B2); + *b = _mm_unpacklo_epi64(B1, B3); +} + +static void ConvertRGBA32ToUV(const uint16_t* rgb, + uint8_t* u, uint8_t* v, int width) { + const int max_width = width & ~15; + const uint16_t* const last_rgb = rgb + 4 * max_width; + while (rgb < last_rgb) { + __m128i r, g, b, U0, V0, U1, V1; + RGBA32PackedToPlanar_16b(rgb + 0, &r, &g, &b); + ConvertRGBToUV(&r, &g, &b, &U0, &V0); + RGBA32PackedToPlanar_16b(rgb + 32, &r, &g, &b); + ConvertRGBToUV(&r, &g, &b, &U1, &V1); + STORE_16(_mm_packus_epi16(U0, U1), u); + STORE_16(_mm_packus_epi16(V0, V1), v); + u += 16; + v += 16; + rgb += 2 * 32; + } + if (max_width < width) { // left-over + WebPConvertRGBA32ToUV_C(rgb, u, v, width - max_width); + } +} + +//------------------------------------------------------------------------------ + +extern void WebPInitConvertARGBToYUVSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVSSE2(void) { + WebPConvertARGBToY = ConvertARGBToY; + WebPConvertARGBToUV = ConvertARGBToUV; + + WebPConvertRGB24ToY = ConvertRGB24ToY; + WebPConvertBGR24ToY = ConvertBGR24ToY; + + WebPConvertRGBA32ToUV = ConvertRGBA32ToUV; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(WebPInitSamplersSSE2) +WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/drivers/webp/dsp/yuv_tables_sse2.h b/drivers/webp/dsp/yuv_tables_sse2.h new file mode 100644 index 0000000000..2b0f057518 --- /dev/null +++ b/drivers/webp/dsp/yuv_tables_sse2.h @@ -0,0 +1,536 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 tables for YUV->RGB conversion (12kB overall) +// +// Author: Skal (pascal.massimino@gmail.com) + +// This file is not compiled, but #include'd directly from yuv.c +// Only used if WEBP_YUV_USE_SSE2_TABLES is defined. + +static const VP8kCstSSE2 VP8kYtoRGBA[256] = { + {{0xfffb77b0, 0xfffb77b0, 0xfffb77b0, 0x003fc000}}, + {{0xfffbc235, 0xfffbc235, 0xfffbc235, 0x003fc000}}, + {{0xfffc0cba, 0xfffc0cba, 0xfffc0cba, 0x003fc000}}, + {{0xfffc573f, 0xfffc573f, 0xfffc573f, 0x003fc000}}, + {{0xfffca1c4, 0xfffca1c4, 0xfffca1c4, 0x003fc000}}, + {{0xfffcec49, 0xfffcec49, 0xfffcec49, 0x003fc000}}, + {{0xfffd36ce, 0xfffd36ce, 0xfffd36ce, 0x003fc000}}, + {{0xfffd8153, 0xfffd8153, 0xfffd8153, 0x003fc000}}, + {{0xfffdcbd8, 0xfffdcbd8, 0xfffdcbd8, 0x003fc000}}, + {{0xfffe165d, 0xfffe165d, 0xfffe165d, 0x003fc000}}, + {{0xfffe60e2, 0xfffe60e2, 0xfffe60e2, 0x003fc000}}, + {{0xfffeab67, 0xfffeab67, 0xfffeab67, 0x003fc000}}, + {{0xfffef5ec, 0xfffef5ec, 0xfffef5ec, 0x003fc000}}, + {{0xffff4071, 0xffff4071, 0xffff4071, 0x003fc000}}, + {{0xffff8af6, 0xffff8af6, 0xffff8af6, 0x003fc000}}, + {{0xffffd57b, 0xffffd57b, 0xffffd57b, 0x003fc000}}, + {{0x00002000, 0x00002000, 0x00002000, 0x003fc000}}, + {{0x00006a85, 0x00006a85, 0x00006a85, 0x003fc000}}, + {{0x0000b50a, 0x0000b50a, 0x0000b50a, 0x003fc000}}, + {{0x0000ff8f, 0x0000ff8f, 0x0000ff8f, 0x003fc000}}, + {{0x00014a14, 0x00014a14, 0x00014a14, 0x003fc000}}, + {{0x00019499, 0x00019499, 0x00019499, 0x003fc000}}, + {{0x0001df1e, 0x0001df1e, 0x0001df1e, 0x003fc000}}, + {{0x000229a3, 0x000229a3, 0x000229a3, 0x003fc000}}, + {{0x00027428, 0x00027428, 0x00027428, 0x003fc000}}, + {{0x0002bead, 0x0002bead, 0x0002bead, 0x003fc000}}, + {{0x00030932, 0x00030932, 0x00030932, 0x003fc000}}, + {{0x000353b7, 0x000353b7, 0x000353b7, 0x003fc000}}, + {{0x00039e3c, 0x00039e3c, 0x00039e3c, 0x003fc000}}, + {{0x0003e8c1, 0x0003e8c1, 0x0003e8c1, 0x003fc000}}, + {{0x00043346, 0x00043346, 0x00043346, 0x003fc000}}, + {{0x00047dcb, 0x00047dcb, 0x00047dcb, 0x003fc000}}, + {{0x0004c850, 0x0004c850, 0x0004c850, 0x003fc000}}, + {{0x000512d5, 0x000512d5, 0x000512d5, 0x003fc000}}, + {{0x00055d5a, 0x00055d5a, 0x00055d5a, 0x003fc000}}, + {{0x0005a7df, 0x0005a7df, 0x0005a7df, 0x003fc000}}, + {{0x0005f264, 0x0005f264, 0x0005f264, 0x003fc000}}, + {{0x00063ce9, 0x00063ce9, 0x00063ce9, 0x003fc000}}, + {{0x0006876e, 0x0006876e, 0x0006876e, 0x003fc000}}, + {{0x0006d1f3, 0x0006d1f3, 0x0006d1f3, 0x003fc000}}, + {{0x00071c78, 0x00071c78, 0x00071c78, 0x003fc000}}, + {{0x000766fd, 0x000766fd, 0x000766fd, 0x003fc000}}, + {{0x0007b182, 0x0007b182, 0x0007b182, 0x003fc000}}, + {{0x0007fc07, 0x0007fc07, 0x0007fc07, 0x003fc000}}, + {{0x0008468c, 0x0008468c, 0x0008468c, 0x003fc000}}, + {{0x00089111, 0x00089111, 0x00089111, 0x003fc000}}, + {{0x0008db96, 0x0008db96, 0x0008db96, 0x003fc000}}, + {{0x0009261b, 0x0009261b, 0x0009261b, 0x003fc000}}, + {{0x000970a0, 0x000970a0, 0x000970a0, 0x003fc000}}, + {{0x0009bb25, 0x0009bb25, 0x0009bb25, 0x003fc000}}, + {{0x000a05aa, 0x000a05aa, 0x000a05aa, 0x003fc000}}, + {{0x000a502f, 0x000a502f, 0x000a502f, 0x003fc000}}, + {{0x000a9ab4, 0x000a9ab4, 0x000a9ab4, 0x003fc000}}, + {{0x000ae539, 0x000ae539, 0x000ae539, 0x003fc000}}, + {{0x000b2fbe, 0x000b2fbe, 0x000b2fbe, 0x003fc000}}, + {{0x000b7a43, 0x000b7a43, 0x000b7a43, 0x003fc000}}, + {{0x000bc4c8, 0x000bc4c8, 0x000bc4c8, 0x003fc000}}, + {{0x000c0f4d, 0x000c0f4d, 0x000c0f4d, 0x003fc000}}, + {{0x000c59d2, 0x000c59d2, 0x000c59d2, 0x003fc000}}, + {{0x000ca457, 0x000ca457, 0x000ca457, 0x003fc000}}, + {{0x000ceedc, 0x000ceedc, 0x000ceedc, 0x003fc000}}, + {{0x000d3961, 0x000d3961, 0x000d3961, 0x003fc000}}, + {{0x000d83e6, 0x000d83e6, 0x000d83e6, 0x003fc000}}, + {{0x000dce6b, 0x000dce6b, 0x000dce6b, 0x003fc000}}, + {{0x000e18f0, 0x000e18f0, 0x000e18f0, 0x003fc000}}, + {{0x000e6375, 0x000e6375, 0x000e6375, 0x003fc000}}, + {{0x000eadfa, 0x000eadfa, 0x000eadfa, 0x003fc000}}, + {{0x000ef87f, 0x000ef87f, 0x000ef87f, 0x003fc000}}, + {{0x000f4304, 0x000f4304, 0x000f4304, 0x003fc000}}, + {{0x000f8d89, 0x000f8d89, 0x000f8d89, 0x003fc000}}, + {{0x000fd80e, 0x000fd80e, 0x000fd80e, 0x003fc000}}, + {{0x00102293, 0x00102293, 0x00102293, 0x003fc000}}, + {{0x00106d18, 0x00106d18, 0x00106d18, 0x003fc000}}, + {{0x0010b79d, 0x0010b79d, 0x0010b79d, 0x003fc000}}, + {{0x00110222, 0x00110222, 0x00110222, 0x003fc000}}, + {{0x00114ca7, 0x00114ca7, 0x00114ca7, 0x003fc000}}, + {{0x0011972c, 0x0011972c, 0x0011972c, 0x003fc000}}, + {{0x0011e1b1, 0x0011e1b1, 0x0011e1b1, 0x003fc000}}, + {{0x00122c36, 0x00122c36, 0x00122c36, 0x003fc000}}, + {{0x001276bb, 0x001276bb, 0x001276bb, 0x003fc000}}, + {{0x0012c140, 0x0012c140, 0x0012c140, 0x003fc000}}, + {{0x00130bc5, 0x00130bc5, 0x00130bc5, 0x003fc000}}, + {{0x0013564a, 0x0013564a, 0x0013564a, 0x003fc000}}, + {{0x0013a0cf, 0x0013a0cf, 0x0013a0cf, 0x003fc000}}, + {{0x0013eb54, 0x0013eb54, 0x0013eb54, 0x003fc000}}, + {{0x001435d9, 0x001435d9, 0x001435d9, 0x003fc000}}, + {{0x0014805e, 0x0014805e, 0x0014805e, 0x003fc000}}, + {{0x0014cae3, 0x0014cae3, 0x0014cae3, 0x003fc000}}, + {{0x00151568, 0x00151568, 0x00151568, 0x003fc000}}, + {{0x00155fed, 0x00155fed, 0x00155fed, 0x003fc000}}, + {{0x0015aa72, 0x0015aa72, 0x0015aa72, 0x003fc000}}, + {{0x0015f4f7, 0x0015f4f7, 0x0015f4f7, 0x003fc000}}, + {{0x00163f7c, 0x00163f7c, 0x00163f7c, 0x003fc000}}, + {{0x00168a01, 0x00168a01, 0x00168a01, 0x003fc000}}, + {{0x0016d486, 0x0016d486, 0x0016d486, 0x003fc000}}, + {{0x00171f0b, 0x00171f0b, 0x00171f0b, 0x003fc000}}, + {{0x00176990, 0x00176990, 0x00176990, 0x003fc000}}, + {{0x0017b415, 0x0017b415, 0x0017b415, 0x003fc000}}, + {{0x0017fe9a, 0x0017fe9a, 0x0017fe9a, 0x003fc000}}, + {{0x0018491f, 0x0018491f, 0x0018491f, 0x003fc000}}, + {{0x001893a4, 0x001893a4, 0x001893a4, 0x003fc000}}, + {{0x0018de29, 0x0018de29, 0x0018de29, 0x003fc000}}, + {{0x001928ae, 0x001928ae, 0x001928ae, 0x003fc000}}, + {{0x00197333, 0x00197333, 0x00197333, 0x003fc000}}, + {{0x0019bdb8, 0x0019bdb8, 0x0019bdb8, 0x003fc000}}, + {{0x001a083d, 0x001a083d, 0x001a083d, 0x003fc000}}, + {{0x001a52c2, 0x001a52c2, 0x001a52c2, 0x003fc000}}, + {{0x001a9d47, 0x001a9d47, 0x001a9d47, 0x003fc000}}, + {{0x001ae7cc, 0x001ae7cc, 0x001ae7cc, 0x003fc000}}, + {{0x001b3251, 0x001b3251, 0x001b3251, 0x003fc000}}, + {{0x001b7cd6, 0x001b7cd6, 0x001b7cd6, 0x003fc000}}, + {{0x001bc75b, 0x001bc75b, 0x001bc75b, 0x003fc000}}, + {{0x001c11e0, 0x001c11e0, 0x001c11e0, 0x003fc000}}, + {{0x001c5c65, 0x001c5c65, 0x001c5c65, 0x003fc000}}, + {{0x001ca6ea, 0x001ca6ea, 0x001ca6ea, 0x003fc000}}, + {{0x001cf16f, 0x001cf16f, 0x001cf16f, 0x003fc000}}, + {{0x001d3bf4, 0x001d3bf4, 0x001d3bf4, 0x003fc000}}, + {{0x001d8679, 0x001d8679, 0x001d8679, 0x003fc000}}, + {{0x001dd0fe, 0x001dd0fe, 0x001dd0fe, 0x003fc000}}, + {{0x001e1b83, 0x001e1b83, 0x001e1b83, 0x003fc000}}, + {{0x001e6608, 0x001e6608, 0x001e6608, 0x003fc000}}, + {{0x001eb08d, 0x001eb08d, 0x001eb08d, 0x003fc000}}, + {{0x001efb12, 0x001efb12, 0x001efb12, 0x003fc000}}, + {{0x001f4597, 0x001f4597, 0x001f4597, 0x003fc000}}, + {{0x001f901c, 0x001f901c, 0x001f901c, 0x003fc000}}, + {{0x001fdaa1, 0x001fdaa1, 0x001fdaa1, 0x003fc000}}, + {{0x00202526, 0x00202526, 0x00202526, 0x003fc000}}, + {{0x00206fab, 0x00206fab, 0x00206fab, 0x003fc000}}, + {{0x0020ba30, 0x0020ba30, 0x0020ba30, 0x003fc000}}, + {{0x002104b5, 0x002104b5, 0x002104b5, 0x003fc000}}, + {{0x00214f3a, 0x00214f3a, 0x00214f3a, 0x003fc000}}, + {{0x002199bf, 0x002199bf, 0x002199bf, 0x003fc000}}, + {{0x0021e444, 0x0021e444, 0x0021e444, 0x003fc000}}, + {{0x00222ec9, 0x00222ec9, 0x00222ec9, 0x003fc000}}, + {{0x0022794e, 0x0022794e, 0x0022794e, 0x003fc000}}, + {{0x0022c3d3, 0x0022c3d3, 0x0022c3d3, 0x003fc000}}, + {{0x00230e58, 0x00230e58, 0x00230e58, 0x003fc000}}, + {{0x002358dd, 0x002358dd, 0x002358dd, 0x003fc000}}, + {{0x0023a362, 0x0023a362, 0x0023a362, 0x003fc000}}, + {{0x0023ede7, 0x0023ede7, 0x0023ede7, 0x003fc000}}, + {{0x0024386c, 0x0024386c, 0x0024386c, 0x003fc000}}, + {{0x002482f1, 0x002482f1, 0x002482f1, 0x003fc000}}, + {{0x0024cd76, 0x0024cd76, 0x0024cd76, 0x003fc000}}, + {{0x002517fb, 0x002517fb, 0x002517fb, 0x003fc000}}, + {{0x00256280, 0x00256280, 0x00256280, 0x003fc000}}, + {{0x0025ad05, 0x0025ad05, 0x0025ad05, 0x003fc000}}, + {{0x0025f78a, 0x0025f78a, 0x0025f78a, 0x003fc000}}, + {{0x0026420f, 0x0026420f, 0x0026420f, 0x003fc000}}, + {{0x00268c94, 0x00268c94, 0x00268c94, 0x003fc000}}, + {{0x0026d719, 0x0026d719, 0x0026d719, 0x003fc000}}, + {{0x0027219e, 0x0027219e, 0x0027219e, 0x003fc000}}, + {{0x00276c23, 0x00276c23, 0x00276c23, 0x003fc000}}, + {{0x0027b6a8, 0x0027b6a8, 0x0027b6a8, 0x003fc000}}, + {{0x0028012d, 0x0028012d, 0x0028012d, 0x003fc000}}, + {{0x00284bb2, 0x00284bb2, 0x00284bb2, 0x003fc000}}, + {{0x00289637, 0x00289637, 0x00289637, 0x003fc000}}, + {{0x0028e0bc, 0x0028e0bc, 0x0028e0bc, 0x003fc000}}, + {{0x00292b41, 0x00292b41, 0x00292b41, 0x003fc000}}, + {{0x002975c6, 0x002975c6, 0x002975c6, 0x003fc000}}, + {{0x0029c04b, 0x0029c04b, 0x0029c04b, 0x003fc000}}, + {{0x002a0ad0, 0x002a0ad0, 0x002a0ad0, 0x003fc000}}, + {{0x002a5555, 0x002a5555, 0x002a5555, 0x003fc000}}, + {{0x002a9fda, 0x002a9fda, 0x002a9fda, 0x003fc000}}, + {{0x002aea5f, 0x002aea5f, 0x002aea5f, 0x003fc000}}, + {{0x002b34e4, 0x002b34e4, 0x002b34e4, 0x003fc000}}, + {{0x002b7f69, 0x002b7f69, 0x002b7f69, 0x003fc000}}, + {{0x002bc9ee, 0x002bc9ee, 0x002bc9ee, 0x003fc000}}, + {{0x002c1473, 0x002c1473, 0x002c1473, 0x003fc000}}, + {{0x002c5ef8, 0x002c5ef8, 0x002c5ef8, 0x003fc000}}, + {{0x002ca97d, 0x002ca97d, 0x002ca97d, 0x003fc000}}, + {{0x002cf402, 0x002cf402, 0x002cf402, 0x003fc000}}, + {{0x002d3e87, 0x002d3e87, 0x002d3e87, 0x003fc000}}, + {{0x002d890c, 0x002d890c, 0x002d890c, 0x003fc000}}, + {{0x002dd391, 0x002dd391, 0x002dd391, 0x003fc000}}, + {{0x002e1e16, 0x002e1e16, 0x002e1e16, 0x003fc000}}, + {{0x002e689b, 0x002e689b, 0x002e689b, 0x003fc000}}, + {{0x002eb320, 0x002eb320, 0x002eb320, 0x003fc000}}, + {{0x002efda5, 0x002efda5, 0x002efda5, 0x003fc000}}, + {{0x002f482a, 0x002f482a, 0x002f482a, 0x003fc000}}, + {{0x002f92af, 0x002f92af, 0x002f92af, 0x003fc000}}, + {{0x002fdd34, 0x002fdd34, 0x002fdd34, 0x003fc000}}, + {{0x003027b9, 0x003027b9, 0x003027b9, 0x003fc000}}, + {{0x0030723e, 0x0030723e, 0x0030723e, 0x003fc000}}, + {{0x0030bcc3, 0x0030bcc3, 0x0030bcc3, 0x003fc000}}, + {{0x00310748, 0x00310748, 0x00310748, 0x003fc000}}, + {{0x003151cd, 0x003151cd, 0x003151cd, 0x003fc000}}, + {{0x00319c52, 0x00319c52, 0x00319c52, 0x003fc000}}, + {{0x0031e6d7, 0x0031e6d7, 0x0031e6d7, 0x003fc000}}, + {{0x0032315c, 0x0032315c, 0x0032315c, 0x003fc000}}, + {{0x00327be1, 0x00327be1, 0x00327be1, 0x003fc000}}, + {{0x0032c666, 0x0032c666, 0x0032c666, 0x003fc000}}, + {{0x003310eb, 0x003310eb, 0x003310eb, 0x003fc000}}, + {{0x00335b70, 0x00335b70, 0x00335b70, 0x003fc000}}, + {{0x0033a5f5, 0x0033a5f5, 0x0033a5f5, 0x003fc000}}, + {{0x0033f07a, 0x0033f07a, 0x0033f07a, 0x003fc000}}, + {{0x00343aff, 0x00343aff, 0x00343aff, 0x003fc000}}, + {{0x00348584, 0x00348584, 0x00348584, 0x003fc000}}, + {{0x0034d009, 0x0034d009, 0x0034d009, 0x003fc000}}, + {{0x00351a8e, 0x00351a8e, 0x00351a8e, 0x003fc000}}, + {{0x00356513, 0x00356513, 0x00356513, 0x003fc000}}, + {{0x0035af98, 0x0035af98, 0x0035af98, 0x003fc000}}, + {{0x0035fa1d, 0x0035fa1d, 0x0035fa1d, 0x003fc000}}, + {{0x003644a2, 0x003644a2, 0x003644a2, 0x003fc000}}, + {{0x00368f27, 0x00368f27, 0x00368f27, 0x003fc000}}, + {{0x0036d9ac, 0x0036d9ac, 0x0036d9ac, 0x003fc000}}, + {{0x00372431, 0x00372431, 0x00372431, 0x003fc000}}, + {{0x00376eb6, 0x00376eb6, 0x00376eb6, 0x003fc000}}, + {{0x0037b93b, 0x0037b93b, 0x0037b93b, 0x003fc000}}, + {{0x003803c0, 0x003803c0, 0x003803c0, 0x003fc000}}, + {{0x00384e45, 0x00384e45, 0x00384e45, 0x003fc000}}, + {{0x003898ca, 0x003898ca, 0x003898ca, 0x003fc000}}, + {{0x0038e34f, 0x0038e34f, 0x0038e34f, 0x003fc000}}, + {{0x00392dd4, 0x00392dd4, 0x00392dd4, 0x003fc000}}, + {{0x00397859, 0x00397859, 0x00397859, 0x003fc000}}, + {{0x0039c2de, 0x0039c2de, 0x0039c2de, 0x003fc000}}, + {{0x003a0d63, 0x003a0d63, 0x003a0d63, 0x003fc000}}, + {{0x003a57e8, 0x003a57e8, 0x003a57e8, 0x003fc000}}, + {{0x003aa26d, 0x003aa26d, 0x003aa26d, 0x003fc000}}, + {{0x003aecf2, 0x003aecf2, 0x003aecf2, 0x003fc000}}, + {{0x003b3777, 0x003b3777, 0x003b3777, 0x003fc000}}, + {{0x003b81fc, 0x003b81fc, 0x003b81fc, 0x003fc000}}, + {{0x003bcc81, 0x003bcc81, 0x003bcc81, 0x003fc000}}, + {{0x003c1706, 0x003c1706, 0x003c1706, 0x003fc000}}, + {{0x003c618b, 0x003c618b, 0x003c618b, 0x003fc000}}, + {{0x003cac10, 0x003cac10, 0x003cac10, 0x003fc000}}, + {{0x003cf695, 0x003cf695, 0x003cf695, 0x003fc000}}, + {{0x003d411a, 0x003d411a, 0x003d411a, 0x003fc000}}, + {{0x003d8b9f, 0x003d8b9f, 0x003d8b9f, 0x003fc000}}, + {{0x003dd624, 0x003dd624, 0x003dd624, 0x003fc000}}, + {{0x003e20a9, 0x003e20a9, 0x003e20a9, 0x003fc000}}, + {{0x003e6b2e, 0x003e6b2e, 0x003e6b2e, 0x003fc000}}, + {{0x003eb5b3, 0x003eb5b3, 0x003eb5b3, 0x003fc000}}, + {{0x003f0038, 0x003f0038, 0x003f0038, 0x003fc000}}, + {{0x003f4abd, 0x003f4abd, 0x003f4abd, 0x003fc000}}, + {{0x003f9542, 0x003f9542, 0x003f9542, 0x003fc000}}, + {{0x003fdfc7, 0x003fdfc7, 0x003fdfc7, 0x003fc000}}, + {{0x00402a4c, 0x00402a4c, 0x00402a4c, 0x003fc000}}, + {{0x004074d1, 0x004074d1, 0x004074d1, 0x003fc000}}, + {{0x0040bf56, 0x0040bf56, 0x0040bf56, 0x003fc000}}, + {{0x004109db, 0x004109db, 0x004109db, 0x003fc000}}, + {{0x00415460, 0x00415460, 0x00415460, 0x003fc000}}, + {{0x00419ee5, 0x00419ee5, 0x00419ee5, 0x003fc000}}, + {{0x0041e96a, 0x0041e96a, 0x0041e96a, 0x003fc000}}, + {{0x004233ef, 0x004233ef, 0x004233ef, 0x003fc000}}, + {{0x00427e74, 0x00427e74, 0x00427e74, 0x003fc000}}, + {{0x0042c8f9, 0x0042c8f9, 0x0042c8f9, 0x003fc000}}, + {{0x0043137e, 0x0043137e, 0x0043137e, 0x003fc000}}, + {{0x00435e03, 0x00435e03, 0x00435e03, 0x003fc000}}, + {{0x0043a888, 0x0043a888, 0x0043a888, 0x003fc000}}, + {{0x0043f30d, 0x0043f30d, 0x0043f30d, 0x003fc000}}, + {{0x00443d92, 0x00443d92, 0x00443d92, 0x003fc000}}, + {{0x00448817, 0x00448817, 0x00448817, 0x003fc000}}, + {{0x0044d29c, 0x0044d29c, 0x0044d29c, 0x003fc000}}, + {{0x00451d21, 0x00451d21, 0x00451d21, 0x003fc000}}, + {{0x004567a6, 0x004567a6, 0x004567a6, 0x003fc000}}, + {{0x0045b22b, 0x0045b22b, 0x0045b22b, 0x003fc000}} +}; + +static const VP8kCstSSE2 VP8kUtoRGBA[256] = { + {{0, 0x000c8980, 0xffbf7300, 0}}, {{0, 0x000c706d, 0xffbff41a, 0}}, + {{0, 0x000c575a, 0xffc07534, 0}}, {{0, 0x000c3e47, 0xffc0f64e, 0}}, + {{0, 0x000c2534, 0xffc17768, 0}}, {{0, 0x000c0c21, 0xffc1f882, 0}}, + {{0, 0x000bf30e, 0xffc2799c, 0}}, {{0, 0x000bd9fb, 0xffc2fab6, 0}}, + {{0, 0x000bc0e8, 0xffc37bd0, 0}}, {{0, 0x000ba7d5, 0xffc3fcea, 0}}, + {{0, 0x000b8ec2, 0xffc47e04, 0}}, {{0, 0x000b75af, 0xffc4ff1e, 0}}, + {{0, 0x000b5c9c, 0xffc58038, 0}}, {{0, 0x000b4389, 0xffc60152, 0}}, + {{0, 0x000b2a76, 0xffc6826c, 0}}, {{0, 0x000b1163, 0xffc70386, 0}}, + {{0, 0x000af850, 0xffc784a0, 0}}, {{0, 0x000adf3d, 0xffc805ba, 0}}, + {{0, 0x000ac62a, 0xffc886d4, 0}}, {{0, 0x000aad17, 0xffc907ee, 0}}, + {{0, 0x000a9404, 0xffc98908, 0}}, {{0, 0x000a7af1, 0xffca0a22, 0}}, + {{0, 0x000a61de, 0xffca8b3c, 0}}, {{0, 0x000a48cb, 0xffcb0c56, 0}}, + {{0, 0x000a2fb8, 0xffcb8d70, 0}}, {{0, 0x000a16a5, 0xffcc0e8a, 0}}, + {{0, 0x0009fd92, 0xffcc8fa4, 0}}, {{0, 0x0009e47f, 0xffcd10be, 0}}, + {{0, 0x0009cb6c, 0xffcd91d8, 0}}, {{0, 0x0009b259, 0xffce12f2, 0}}, + {{0, 0x00099946, 0xffce940c, 0}}, {{0, 0x00098033, 0xffcf1526, 0}}, + {{0, 0x00096720, 0xffcf9640, 0}}, {{0, 0x00094e0d, 0xffd0175a, 0}}, + {{0, 0x000934fa, 0xffd09874, 0}}, {{0, 0x00091be7, 0xffd1198e, 0}}, + {{0, 0x000902d4, 0xffd19aa8, 0}}, {{0, 0x0008e9c1, 0xffd21bc2, 0}}, + {{0, 0x0008d0ae, 0xffd29cdc, 0}}, {{0, 0x0008b79b, 0xffd31df6, 0}}, + {{0, 0x00089e88, 0xffd39f10, 0}}, {{0, 0x00088575, 0xffd4202a, 0}}, + {{0, 0x00086c62, 0xffd4a144, 0}}, {{0, 0x0008534f, 0xffd5225e, 0}}, + {{0, 0x00083a3c, 0xffd5a378, 0}}, {{0, 0x00082129, 0xffd62492, 0}}, + {{0, 0x00080816, 0xffd6a5ac, 0}}, {{0, 0x0007ef03, 0xffd726c6, 0}}, + {{0, 0x0007d5f0, 0xffd7a7e0, 0}}, {{0, 0x0007bcdd, 0xffd828fa, 0}}, + {{0, 0x0007a3ca, 0xffd8aa14, 0}}, {{0, 0x00078ab7, 0xffd92b2e, 0}}, + {{0, 0x000771a4, 0xffd9ac48, 0}}, {{0, 0x00075891, 0xffda2d62, 0}}, + {{0, 0x00073f7e, 0xffdaae7c, 0}}, {{0, 0x0007266b, 0xffdb2f96, 0}}, + {{0, 0x00070d58, 0xffdbb0b0, 0}}, {{0, 0x0006f445, 0xffdc31ca, 0}}, + {{0, 0x0006db32, 0xffdcb2e4, 0}}, {{0, 0x0006c21f, 0xffdd33fe, 0}}, + {{0, 0x0006a90c, 0xffddb518, 0}}, {{0, 0x00068ff9, 0xffde3632, 0}}, + {{0, 0x000676e6, 0xffdeb74c, 0}}, {{0, 0x00065dd3, 0xffdf3866, 0}}, + {{0, 0x000644c0, 0xffdfb980, 0}}, {{0, 0x00062bad, 0xffe03a9a, 0}}, + {{0, 0x0006129a, 0xffe0bbb4, 0}}, {{0, 0x0005f987, 0xffe13cce, 0}}, + {{0, 0x0005e074, 0xffe1bde8, 0}}, {{0, 0x0005c761, 0xffe23f02, 0}}, + {{0, 0x0005ae4e, 0xffe2c01c, 0}}, {{0, 0x0005953b, 0xffe34136, 0}}, + {{0, 0x00057c28, 0xffe3c250, 0}}, {{0, 0x00056315, 0xffe4436a, 0}}, + {{0, 0x00054a02, 0xffe4c484, 0}}, {{0, 0x000530ef, 0xffe5459e, 0}}, + {{0, 0x000517dc, 0xffe5c6b8, 0}}, {{0, 0x0004fec9, 0xffe647d2, 0}}, + {{0, 0x0004e5b6, 0xffe6c8ec, 0}}, {{0, 0x0004cca3, 0xffe74a06, 0}}, + {{0, 0x0004b390, 0xffe7cb20, 0}}, {{0, 0x00049a7d, 0xffe84c3a, 0}}, + {{0, 0x0004816a, 0xffe8cd54, 0}}, {{0, 0x00046857, 0xffe94e6e, 0}}, + {{0, 0x00044f44, 0xffe9cf88, 0}}, {{0, 0x00043631, 0xffea50a2, 0}}, + {{0, 0x00041d1e, 0xffead1bc, 0}}, {{0, 0x0004040b, 0xffeb52d6, 0}}, + {{0, 0x0003eaf8, 0xffebd3f0, 0}}, {{0, 0x0003d1e5, 0xffec550a, 0}}, + {{0, 0x0003b8d2, 0xffecd624, 0}}, {{0, 0x00039fbf, 0xffed573e, 0}}, + {{0, 0x000386ac, 0xffedd858, 0}}, {{0, 0x00036d99, 0xffee5972, 0}}, + {{0, 0x00035486, 0xffeeda8c, 0}}, {{0, 0x00033b73, 0xffef5ba6, 0}}, + {{0, 0x00032260, 0xffefdcc0, 0}}, {{0, 0x0003094d, 0xfff05dda, 0}}, + {{0, 0x0002f03a, 0xfff0def4, 0}}, {{0, 0x0002d727, 0xfff1600e, 0}}, + {{0, 0x0002be14, 0xfff1e128, 0}}, {{0, 0x0002a501, 0xfff26242, 0}}, + {{0, 0x00028bee, 0xfff2e35c, 0}}, {{0, 0x000272db, 0xfff36476, 0}}, + {{0, 0x000259c8, 0xfff3e590, 0}}, {{0, 0x000240b5, 0xfff466aa, 0}}, + {{0, 0x000227a2, 0xfff4e7c4, 0}}, {{0, 0x00020e8f, 0xfff568de, 0}}, + {{0, 0x0001f57c, 0xfff5e9f8, 0}}, {{0, 0x0001dc69, 0xfff66b12, 0}}, + {{0, 0x0001c356, 0xfff6ec2c, 0}}, {{0, 0x0001aa43, 0xfff76d46, 0}}, + {{0, 0x00019130, 0xfff7ee60, 0}}, {{0, 0x0001781d, 0xfff86f7a, 0}}, + {{0, 0x00015f0a, 0xfff8f094, 0}}, {{0, 0x000145f7, 0xfff971ae, 0}}, + {{0, 0x00012ce4, 0xfff9f2c8, 0}}, {{0, 0x000113d1, 0xfffa73e2, 0}}, + {{0, 0x0000fabe, 0xfffaf4fc, 0}}, {{0, 0x0000e1ab, 0xfffb7616, 0}}, + {{0, 0x0000c898, 0xfffbf730, 0}}, {{0, 0x0000af85, 0xfffc784a, 0}}, + {{0, 0x00009672, 0xfffcf964, 0}}, {{0, 0x00007d5f, 0xfffd7a7e, 0}}, + {{0, 0x0000644c, 0xfffdfb98, 0}}, {{0, 0x00004b39, 0xfffe7cb2, 0}}, + {{0, 0x00003226, 0xfffefdcc, 0}}, {{0, 0x00001913, 0xffff7ee6, 0}}, + {{0, 0x00000000, 0x00000000, 0}}, {{0, 0xffffe6ed, 0x0000811a, 0}}, + {{0, 0xffffcdda, 0x00010234, 0}}, {{0, 0xffffb4c7, 0x0001834e, 0}}, + {{0, 0xffff9bb4, 0x00020468, 0}}, {{0, 0xffff82a1, 0x00028582, 0}}, + {{0, 0xffff698e, 0x0003069c, 0}}, {{0, 0xffff507b, 0x000387b6, 0}}, + {{0, 0xffff3768, 0x000408d0, 0}}, {{0, 0xffff1e55, 0x000489ea, 0}}, + {{0, 0xffff0542, 0x00050b04, 0}}, {{0, 0xfffeec2f, 0x00058c1e, 0}}, + {{0, 0xfffed31c, 0x00060d38, 0}}, {{0, 0xfffeba09, 0x00068e52, 0}}, + {{0, 0xfffea0f6, 0x00070f6c, 0}}, {{0, 0xfffe87e3, 0x00079086, 0}}, + {{0, 0xfffe6ed0, 0x000811a0, 0}}, {{0, 0xfffe55bd, 0x000892ba, 0}}, + {{0, 0xfffe3caa, 0x000913d4, 0}}, {{0, 0xfffe2397, 0x000994ee, 0}}, + {{0, 0xfffe0a84, 0x000a1608, 0}}, {{0, 0xfffdf171, 0x000a9722, 0}}, + {{0, 0xfffdd85e, 0x000b183c, 0}}, {{0, 0xfffdbf4b, 0x000b9956, 0}}, + {{0, 0xfffda638, 0x000c1a70, 0}}, {{0, 0xfffd8d25, 0x000c9b8a, 0}}, + {{0, 0xfffd7412, 0x000d1ca4, 0}}, {{0, 0xfffd5aff, 0x000d9dbe, 0}}, + {{0, 0xfffd41ec, 0x000e1ed8, 0}}, {{0, 0xfffd28d9, 0x000e9ff2, 0}}, + {{0, 0xfffd0fc6, 0x000f210c, 0}}, {{0, 0xfffcf6b3, 0x000fa226, 0}}, + {{0, 0xfffcdda0, 0x00102340, 0}}, {{0, 0xfffcc48d, 0x0010a45a, 0}}, + {{0, 0xfffcab7a, 0x00112574, 0}}, {{0, 0xfffc9267, 0x0011a68e, 0}}, + {{0, 0xfffc7954, 0x001227a8, 0}}, {{0, 0xfffc6041, 0x0012a8c2, 0}}, + {{0, 0xfffc472e, 0x001329dc, 0}}, {{0, 0xfffc2e1b, 0x0013aaf6, 0}}, + {{0, 0xfffc1508, 0x00142c10, 0}}, {{0, 0xfffbfbf5, 0x0014ad2a, 0}}, + {{0, 0xfffbe2e2, 0x00152e44, 0}}, {{0, 0xfffbc9cf, 0x0015af5e, 0}}, + {{0, 0xfffbb0bc, 0x00163078, 0}}, {{0, 0xfffb97a9, 0x0016b192, 0}}, + {{0, 0xfffb7e96, 0x001732ac, 0}}, {{0, 0xfffb6583, 0x0017b3c6, 0}}, + {{0, 0xfffb4c70, 0x001834e0, 0}}, {{0, 0xfffb335d, 0x0018b5fa, 0}}, + {{0, 0xfffb1a4a, 0x00193714, 0}}, {{0, 0xfffb0137, 0x0019b82e, 0}}, + {{0, 0xfffae824, 0x001a3948, 0}}, {{0, 0xfffacf11, 0x001aba62, 0}}, + {{0, 0xfffab5fe, 0x001b3b7c, 0}}, {{0, 0xfffa9ceb, 0x001bbc96, 0}}, + {{0, 0xfffa83d8, 0x001c3db0, 0}}, {{0, 0xfffa6ac5, 0x001cbeca, 0}}, + {{0, 0xfffa51b2, 0x001d3fe4, 0}}, {{0, 0xfffa389f, 0x001dc0fe, 0}}, + {{0, 0xfffa1f8c, 0x001e4218, 0}}, {{0, 0xfffa0679, 0x001ec332, 0}}, + {{0, 0xfff9ed66, 0x001f444c, 0}}, {{0, 0xfff9d453, 0x001fc566, 0}}, + {{0, 0xfff9bb40, 0x00204680, 0}}, {{0, 0xfff9a22d, 0x0020c79a, 0}}, + {{0, 0xfff9891a, 0x002148b4, 0}}, {{0, 0xfff97007, 0x0021c9ce, 0}}, + {{0, 0xfff956f4, 0x00224ae8, 0}}, {{0, 0xfff93de1, 0x0022cc02, 0}}, + {{0, 0xfff924ce, 0x00234d1c, 0}}, {{0, 0xfff90bbb, 0x0023ce36, 0}}, + {{0, 0xfff8f2a8, 0x00244f50, 0}}, {{0, 0xfff8d995, 0x0024d06a, 0}}, + {{0, 0xfff8c082, 0x00255184, 0}}, {{0, 0xfff8a76f, 0x0025d29e, 0}}, + {{0, 0xfff88e5c, 0x002653b8, 0}}, {{0, 0xfff87549, 0x0026d4d2, 0}}, + {{0, 0xfff85c36, 0x002755ec, 0}}, {{0, 0xfff84323, 0x0027d706, 0}}, + {{0, 0xfff82a10, 0x00285820, 0}}, {{0, 0xfff810fd, 0x0028d93a, 0}}, + {{0, 0xfff7f7ea, 0x00295a54, 0}}, {{0, 0xfff7ded7, 0x0029db6e, 0}}, + {{0, 0xfff7c5c4, 0x002a5c88, 0}}, {{0, 0xfff7acb1, 0x002adda2, 0}}, + {{0, 0xfff7939e, 0x002b5ebc, 0}}, {{0, 0xfff77a8b, 0x002bdfd6, 0}}, + {{0, 0xfff76178, 0x002c60f0, 0}}, {{0, 0xfff74865, 0x002ce20a, 0}}, + {{0, 0xfff72f52, 0x002d6324, 0}}, {{0, 0xfff7163f, 0x002de43e, 0}}, + {{0, 0xfff6fd2c, 0x002e6558, 0}}, {{0, 0xfff6e419, 0x002ee672, 0}}, + {{0, 0xfff6cb06, 0x002f678c, 0}}, {{0, 0xfff6b1f3, 0x002fe8a6, 0}}, + {{0, 0xfff698e0, 0x003069c0, 0}}, {{0, 0xfff67fcd, 0x0030eada, 0}}, + {{0, 0xfff666ba, 0x00316bf4, 0}}, {{0, 0xfff64da7, 0x0031ed0e, 0}}, + {{0, 0xfff63494, 0x00326e28, 0}}, {{0, 0xfff61b81, 0x0032ef42, 0}}, + {{0, 0xfff6026e, 0x0033705c, 0}}, {{0, 0xfff5e95b, 0x0033f176, 0}}, + {{0, 0xfff5d048, 0x00347290, 0}}, {{0, 0xfff5b735, 0x0034f3aa, 0}}, + {{0, 0xfff59e22, 0x003574c4, 0}}, {{0, 0xfff5850f, 0x0035f5de, 0}}, + {{0, 0xfff56bfc, 0x003676f8, 0}}, {{0, 0xfff552e9, 0x0036f812, 0}}, + {{0, 0xfff539d6, 0x0037792c, 0}}, {{0, 0xfff520c3, 0x0037fa46, 0}}, + {{0, 0xfff507b0, 0x00387b60, 0}}, {{0, 0xfff4ee9d, 0x0038fc7a, 0}}, + {{0, 0xfff4d58a, 0x00397d94, 0}}, {{0, 0xfff4bc77, 0x0039feae, 0}}, + {{0, 0xfff4a364, 0x003a7fc8, 0}}, {{0, 0xfff48a51, 0x003b00e2, 0}}, + {{0, 0xfff4713e, 0x003b81fc, 0}}, {{0, 0xfff4582b, 0x003c0316, 0}}, + {{0, 0xfff43f18, 0x003c8430, 0}}, {{0, 0xfff42605, 0x003d054a, 0}}, + {{0, 0xfff40cf2, 0x003d8664, 0}}, {{0, 0xfff3f3df, 0x003e077e, 0}}, + {{0, 0xfff3dacc, 0x003e8898, 0}}, {{0, 0xfff3c1b9, 0x003f09b2, 0}}, + {{0, 0xfff3a8a6, 0x003f8acc, 0}}, {{0, 0xfff38f93, 0x00400be6, 0}} +}; + +static VP8kCstSSE2 VP8kVtoRGBA[256] = { + {{0xffcced80, 0x001a0400, 0, 0}}, {{0xffcd53a5, 0x0019cff8, 0, 0}}, + {{0xffcdb9ca, 0x00199bf0, 0, 0}}, {{0xffce1fef, 0x001967e8, 0, 0}}, + {{0xffce8614, 0x001933e0, 0, 0}}, {{0xffceec39, 0x0018ffd8, 0, 0}}, + {{0xffcf525e, 0x0018cbd0, 0, 0}}, {{0xffcfb883, 0x001897c8, 0, 0}}, + {{0xffd01ea8, 0x001863c0, 0, 0}}, {{0xffd084cd, 0x00182fb8, 0, 0}}, + {{0xffd0eaf2, 0x0017fbb0, 0, 0}}, {{0xffd15117, 0x0017c7a8, 0, 0}}, + {{0xffd1b73c, 0x001793a0, 0, 0}}, {{0xffd21d61, 0x00175f98, 0, 0}}, + {{0xffd28386, 0x00172b90, 0, 0}}, {{0xffd2e9ab, 0x0016f788, 0, 0}}, + {{0xffd34fd0, 0x0016c380, 0, 0}}, {{0xffd3b5f5, 0x00168f78, 0, 0}}, + {{0xffd41c1a, 0x00165b70, 0, 0}}, {{0xffd4823f, 0x00162768, 0, 0}}, + {{0xffd4e864, 0x0015f360, 0, 0}}, {{0xffd54e89, 0x0015bf58, 0, 0}}, + {{0xffd5b4ae, 0x00158b50, 0, 0}}, {{0xffd61ad3, 0x00155748, 0, 0}}, + {{0xffd680f8, 0x00152340, 0, 0}}, {{0xffd6e71d, 0x0014ef38, 0, 0}}, + {{0xffd74d42, 0x0014bb30, 0, 0}}, {{0xffd7b367, 0x00148728, 0, 0}}, + {{0xffd8198c, 0x00145320, 0, 0}}, {{0xffd87fb1, 0x00141f18, 0, 0}}, + {{0xffd8e5d6, 0x0013eb10, 0, 0}}, {{0xffd94bfb, 0x0013b708, 0, 0}}, + {{0xffd9b220, 0x00138300, 0, 0}}, {{0xffda1845, 0x00134ef8, 0, 0}}, + {{0xffda7e6a, 0x00131af0, 0, 0}}, {{0xffdae48f, 0x0012e6e8, 0, 0}}, + {{0xffdb4ab4, 0x0012b2e0, 0, 0}}, {{0xffdbb0d9, 0x00127ed8, 0, 0}}, + {{0xffdc16fe, 0x00124ad0, 0, 0}}, {{0xffdc7d23, 0x001216c8, 0, 0}}, + {{0xffdce348, 0x0011e2c0, 0, 0}}, {{0xffdd496d, 0x0011aeb8, 0, 0}}, + {{0xffddaf92, 0x00117ab0, 0, 0}}, {{0xffde15b7, 0x001146a8, 0, 0}}, + {{0xffde7bdc, 0x001112a0, 0, 0}}, {{0xffdee201, 0x0010de98, 0, 0}}, + {{0xffdf4826, 0x0010aa90, 0, 0}}, {{0xffdfae4b, 0x00107688, 0, 0}}, + {{0xffe01470, 0x00104280, 0, 0}}, {{0xffe07a95, 0x00100e78, 0, 0}}, + {{0xffe0e0ba, 0x000fda70, 0, 0}}, {{0xffe146df, 0x000fa668, 0, 0}}, + {{0xffe1ad04, 0x000f7260, 0, 0}}, {{0xffe21329, 0x000f3e58, 0, 0}}, + {{0xffe2794e, 0x000f0a50, 0, 0}}, {{0xffe2df73, 0x000ed648, 0, 0}}, + {{0xffe34598, 0x000ea240, 0, 0}}, {{0xffe3abbd, 0x000e6e38, 0, 0}}, + {{0xffe411e2, 0x000e3a30, 0, 0}}, {{0xffe47807, 0x000e0628, 0, 0}}, + {{0xffe4de2c, 0x000dd220, 0, 0}}, {{0xffe54451, 0x000d9e18, 0, 0}}, + {{0xffe5aa76, 0x000d6a10, 0, 0}}, {{0xffe6109b, 0x000d3608, 0, 0}}, + {{0xffe676c0, 0x000d0200, 0, 0}}, {{0xffe6dce5, 0x000ccdf8, 0, 0}}, + {{0xffe7430a, 0x000c99f0, 0, 0}}, {{0xffe7a92f, 0x000c65e8, 0, 0}}, + {{0xffe80f54, 0x000c31e0, 0, 0}}, {{0xffe87579, 0x000bfdd8, 0, 0}}, + {{0xffe8db9e, 0x000bc9d0, 0, 0}}, {{0xffe941c3, 0x000b95c8, 0, 0}}, + {{0xffe9a7e8, 0x000b61c0, 0, 0}}, {{0xffea0e0d, 0x000b2db8, 0, 0}}, + {{0xffea7432, 0x000af9b0, 0, 0}}, {{0xffeada57, 0x000ac5a8, 0, 0}}, + {{0xffeb407c, 0x000a91a0, 0, 0}}, {{0xffeba6a1, 0x000a5d98, 0, 0}}, + {{0xffec0cc6, 0x000a2990, 0, 0}}, {{0xffec72eb, 0x0009f588, 0, 0}}, + {{0xffecd910, 0x0009c180, 0, 0}}, {{0xffed3f35, 0x00098d78, 0, 0}}, + {{0xffeda55a, 0x00095970, 0, 0}}, {{0xffee0b7f, 0x00092568, 0, 0}}, + {{0xffee71a4, 0x0008f160, 0, 0}}, {{0xffeed7c9, 0x0008bd58, 0, 0}}, + {{0xffef3dee, 0x00088950, 0, 0}}, {{0xffefa413, 0x00085548, 0, 0}}, + {{0xfff00a38, 0x00082140, 0, 0}}, {{0xfff0705d, 0x0007ed38, 0, 0}}, + {{0xfff0d682, 0x0007b930, 0, 0}}, {{0xfff13ca7, 0x00078528, 0, 0}}, + {{0xfff1a2cc, 0x00075120, 0, 0}}, {{0xfff208f1, 0x00071d18, 0, 0}}, + {{0xfff26f16, 0x0006e910, 0, 0}}, {{0xfff2d53b, 0x0006b508, 0, 0}}, + {{0xfff33b60, 0x00068100, 0, 0}}, {{0xfff3a185, 0x00064cf8, 0, 0}}, + {{0xfff407aa, 0x000618f0, 0, 0}}, {{0xfff46dcf, 0x0005e4e8, 0, 0}}, + {{0xfff4d3f4, 0x0005b0e0, 0, 0}}, {{0xfff53a19, 0x00057cd8, 0, 0}}, + {{0xfff5a03e, 0x000548d0, 0, 0}}, {{0xfff60663, 0x000514c8, 0, 0}}, + {{0xfff66c88, 0x0004e0c0, 0, 0}}, {{0xfff6d2ad, 0x0004acb8, 0, 0}}, + {{0xfff738d2, 0x000478b0, 0, 0}}, {{0xfff79ef7, 0x000444a8, 0, 0}}, + {{0xfff8051c, 0x000410a0, 0, 0}}, {{0xfff86b41, 0x0003dc98, 0, 0}}, + {{0xfff8d166, 0x0003a890, 0, 0}}, {{0xfff9378b, 0x00037488, 0, 0}}, + {{0xfff99db0, 0x00034080, 0, 0}}, {{0xfffa03d5, 0x00030c78, 0, 0}}, + {{0xfffa69fa, 0x0002d870, 0, 0}}, {{0xfffad01f, 0x0002a468, 0, 0}}, + {{0xfffb3644, 0x00027060, 0, 0}}, {{0xfffb9c69, 0x00023c58, 0, 0}}, + {{0xfffc028e, 0x00020850, 0, 0}}, {{0xfffc68b3, 0x0001d448, 0, 0}}, + {{0xfffcced8, 0x0001a040, 0, 0}}, {{0xfffd34fd, 0x00016c38, 0, 0}}, + {{0xfffd9b22, 0x00013830, 0, 0}}, {{0xfffe0147, 0x00010428, 0, 0}}, + {{0xfffe676c, 0x0000d020, 0, 0}}, {{0xfffecd91, 0x00009c18, 0, 0}}, + {{0xffff33b6, 0x00006810, 0, 0}}, {{0xffff99db, 0x00003408, 0, 0}}, + {{0x00000000, 0x00000000, 0, 0}}, {{0x00006625, 0xffffcbf8, 0, 0}}, + {{0x0000cc4a, 0xffff97f0, 0, 0}}, {{0x0001326f, 0xffff63e8, 0, 0}}, + {{0x00019894, 0xffff2fe0, 0, 0}}, {{0x0001feb9, 0xfffefbd8, 0, 0}}, + {{0x000264de, 0xfffec7d0, 0, 0}}, {{0x0002cb03, 0xfffe93c8, 0, 0}}, + {{0x00033128, 0xfffe5fc0, 0, 0}}, {{0x0003974d, 0xfffe2bb8, 0, 0}}, + {{0x0003fd72, 0xfffdf7b0, 0, 0}}, {{0x00046397, 0xfffdc3a8, 0, 0}}, + {{0x0004c9bc, 0xfffd8fa0, 0, 0}}, {{0x00052fe1, 0xfffd5b98, 0, 0}}, + {{0x00059606, 0xfffd2790, 0, 0}}, {{0x0005fc2b, 0xfffcf388, 0, 0}}, + {{0x00066250, 0xfffcbf80, 0, 0}}, {{0x0006c875, 0xfffc8b78, 0, 0}}, + {{0x00072e9a, 0xfffc5770, 0, 0}}, {{0x000794bf, 0xfffc2368, 0, 0}}, + {{0x0007fae4, 0xfffbef60, 0, 0}}, {{0x00086109, 0xfffbbb58, 0, 0}}, + {{0x0008c72e, 0xfffb8750, 0, 0}}, {{0x00092d53, 0xfffb5348, 0, 0}}, + {{0x00099378, 0xfffb1f40, 0, 0}}, {{0x0009f99d, 0xfffaeb38, 0, 0}}, + {{0x000a5fc2, 0xfffab730, 0, 0}}, {{0x000ac5e7, 0xfffa8328, 0, 0}}, + {{0x000b2c0c, 0xfffa4f20, 0, 0}}, {{0x000b9231, 0xfffa1b18, 0, 0}}, + {{0x000bf856, 0xfff9e710, 0, 0}}, {{0x000c5e7b, 0xfff9b308, 0, 0}}, + {{0x000cc4a0, 0xfff97f00, 0, 0}}, {{0x000d2ac5, 0xfff94af8, 0, 0}}, + {{0x000d90ea, 0xfff916f0, 0, 0}}, {{0x000df70f, 0xfff8e2e8, 0, 0}}, + {{0x000e5d34, 0xfff8aee0, 0, 0}}, {{0x000ec359, 0xfff87ad8, 0, 0}}, + {{0x000f297e, 0xfff846d0, 0, 0}}, {{0x000f8fa3, 0xfff812c8, 0, 0}}, + {{0x000ff5c8, 0xfff7dec0, 0, 0}}, {{0x00105bed, 0xfff7aab8, 0, 0}}, + {{0x0010c212, 0xfff776b0, 0, 0}}, {{0x00112837, 0xfff742a8, 0, 0}}, + {{0x00118e5c, 0xfff70ea0, 0, 0}}, {{0x0011f481, 0xfff6da98, 0, 0}}, + {{0x00125aa6, 0xfff6a690, 0, 0}}, {{0x0012c0cb, 0xfff67288, 0, 0}}, + {{0x001326f0, 0xfff63e80, 0, 0}}, {{0x00138d15, 0xfff60a78, 0, 0}}, + {{0x0013f33a, 0xfff5d670, 0, 0}}, {{0x0014595f, 0xfff5a268, 0, 0}}, + {{0x0014bf84, 0xfff56e60, 0, 0}}, {{0x001525a9, 0xfff53a58, 0, 0}}, + {{0x00158bce, 0xfff50650, 0, 0}}, {{0x0015f1f3, 0xfff4d248, 0, 0}}, + {{0x00165818, 0xfff49e40, 0, 0}}, {{0x0016be3d, 0xfff46a38, 0, 0}}, + {{0x00172462, 0xfff43630, 0, 0}}, {{0x00178a87, 0xfff40228, 0, 0}}, + {{0x0017f0ac, 0xfff3ce20, 0, 0}}, {{0x001856d1, 0xfff39a18, 0, 0}}, + {{0x0018bcf6, 0xfff36610, 0, 0}}, {{0x0019231b, 0xfff33208, 0, 0}}, + {{0x00198940, 0xfff2fe00, 0, 0}}, {{0x0019ef65, 0xfff2c9f8, 0, 0}}, + {{0x001a558a, 0xfff295f0, 0, 0}}, {{0x001abbaf, 0xfff261e8, 0, 0}}, + {{0x001b21d4, 0xfff22de0, 0, 0}}, {{0x001b87f9, 0xfff1f9d8, 0, 0}}, + {{0x001bee1e, 0xfff1c5d0, 0, 0}}, {{0x001c5443, 0xfff191c8, 0, 0}}, + {{0x001cba68, 0xfff15dc0, 0, 0}}, {{0x001d208d, 0xfff129b8, 0, 0}}, + {{0x001d86b2, 0xfff0f5b0, 0, 0}}, {{0x001decd7, 0xfff0c1a8, 0, 0}}, + {{0x001e52fc, 0xfff08da0, 0, 0}}, {{0x001eb921, 0xfff05998, 0, 0}}, + {{0x001f1f46, 0xfff02590, 0, 0}}, {{0x001f856b, 0xffeff188, 0, 0}}, + {{0x001feb90, 0xffefbd80, 0, 0}}, {{0x002051b5, 0xffef8978, 0, 0}}, + {{0x0020b7da, 0xffef5570, 0, 0}}, {{0x00211dff, 0xffef2168, 0, 0}}, + {{0x00218424, 0xffeeed60, 0, 0}}, {{0x0021ea49, 0xffeeb958, 0, 0}}, + {{0x0022506e, 0xffee8550, 0, 0}}, {{0x0022b693, 0xffee5148, 0, 0}}, + {{0x00231cb8, 0xffee1d40, 0, 0}}, {{0x002382dd, 0xffede938, 0, 0}}, + {{0x0023e902, 0xffedb530, 0, 0}}, {{0x00244f27, 0xffed8128, 0, 0}}, + {{0x0024b54c, 0xffed4d20, 0, 0}}, {{0x00251b71, 0xffed1918, 0, 0}}, + {{0x00258196, 0xffece510, 0, 0}}, {{0x0025e7bb, 0xffecb108, 0, 0}}, + {{0x00264de0, 0xffec7d00, 0, 0}}, {{0x0026b405, 0xffec48f8, 0, 0}}, + {{0x00271a2a, 0xffec14f0, 0, 0}}, {{0x0027804f, 0xffebe0e8, 0, 0}}, + {{0x0027e674, 0xffebace0, 0, 0}}, {{0x00284c99, 0xffeb78d8, 0, 0}}, + {{0x0028b2be, 0xffeb44d0, 0, 0}}, {{0x002918e3, 0xffeb10c8, 0, 0}}, + {{0x00297f08, 0xffeadcc0, 0, 0}}, {{0x0029e52d, 0xffeaa8b8, 0, 0}}, + {{0x002a4b52, 0xffea74b0, 0, 0}}, {{0x002ab177, 0xffea40a8, 0, 0}}, + {{0x002b179c, 0xffea0ca0, 0, 0}}, {{0x002b7dc1, 0xffe9d898, 0, 0}}, + {{0x002be3e6, 0xffe9a490, 0, 0}}, {{0x002c4a0b, 0xffe97088, 0, 0}}, + {{0x002cb030, 0xffe93c80, 0, 0}}, {{0x002d1655, 0xffe90878, 0, 0}}, + {{0x002d7c7a, 0xffe8d470, 0, 0}}, {{0x002de29f, 0xffe8a068, 0, 0}}, + {{0x002e48c4, 0xffe86c60, 0, 0}}, {{0x002eaee9, 0xffe83858, 0, 0}}, + {{0x002f150e, 0xffe80450, 0, 0}}, {{0x002f7b33, 0xffe7d048, 0, 0}}, + {{0x002fe158, 0xffe79c40, 0, 0}}, {{0x0030477d, 0xffe76838, 0, 0}}, + {{0x0030ada2, 0xffe73430, 0, 0}}, {{0x003113c7, 0xffe70028, 0, 0}}, + {{0x003179ec, 0xffe6cc20, 0, 0}}, {{0x0031e011, 0xffe69818, 0, 0}}, + {{0x00324636, 0xffe66410, 0, 0}}, {{0x0032ac5b, 0xffe63008, 0, 0}} +}; |