diff options
Diffstat (limited to 'thirdparty/libwebp/src/dsp')
26 files changed, 1303 insertions, 170 deletions
diff --git a/thirdparty/libwebp/src/dsp/alpha_processing.c b/thirdparty/libwebp/src/dsp/alpha_processing.c index 590e3bc312..819d1391f2 100644 --- a/thirdparty/libwebp/src/dsp/alpha_processing.c +++ b/thirdparty/libwebp/src/dsp/alpha_processing.c @@ -366,6 +366,16 @@ static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) { return (((uint32_t)a << 24) | (r << 16) | (g << 8) | b); } +#ifdef WORDS_BIGENDIAN +static void PackARGB_C(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]); + } +} +#endif + static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b, int len, int step, uint32_t* out) { int i, offset = 0; @@ -381,6 +391,10 @@ 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); void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size); +#ifdef WORDS_BIGENDIAN +void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, const uint8_t* g, + const uint8_t* b, int, uint32_t*); +#endif void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b, int len, int step, uint32_t* out); @@ -395,16 +409,14 @@ extern void WebPInitAlphaProcessingSSE2(void); extern void WebPInitAlphaProcessingSSE41(void); extern void WebPInitAlphaProcessingNEON(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; - +WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) { WebPMultARGBRow = WebPMultARGBRow_C; WebPMultRow = WebPMultRow_C; WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply_16b_C; +#ifdef WORDS_BIGENDIAN + WebPPackARGB = PackARGB_C; +#endif WebPPackRGB = PackRGB_C; #if !WEBP_NEON_OMIT_C_CODE WebPApplyAlphaMultiply = ApplyAlphaMultiply_C; @@ -451,9 +463,10 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessing(void) { assert(WebPDispatchAlphaToGreen != NULL); assert(WebPExtractAlpha != NULL); assert(WebPExtractGreen != NULL); +#ifdef WORDS_BIGENDIAN + assert(WebPPackARGB != NULL); +#endif assert(WebPPackRGB != NULL); assert(WebPHasAlpha8b != NULL); assert(WebPHasAlpha32b != NULL); - - alpha_processing_last_cpuinfo_used = VP8GetCPUInfo; } diff --git a/thirdparty/libwebp/src/dsp/alpha_processing_mips_dsp_r2.c b/thirdparty/libwebp/src/dsp/alpha_processing_mips_dsp_r2.c index e0dc91bab9..0090e87cd1 100644 --- a/thirdparty/libwebp/src/dsp/alpha_processing_mips_dsp_r2.c +++ b/thirdparty/libwebp/src/dsp/alpha_processing_mips_dsp_r2.c @@ -125,6 +125,49 @@ static void MultARGBRow_MIPSdspR2(uint32_t* const ptr, int width, } } +#ifdef WORDS_BIGENDIAN +static void PackARGB_MIPSdspR2(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" + ); +} +#endif // WORDS_BIGENDIAN + static void PackRGB_MIPSdspR2(const uint8_t* r, const uint8_t* g, const uint8_t* b, int len, int step, uint32_t* out) { @@ -172,6 +215,9 @@ extern void WebPInitAlphaProcessingMIPSdspR2(void); WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingMIPSdspR2(void) { WebPDispatchAlpha = DispatchAlpha_MIPSdspR2; WebPMultARGBRow = MultARGBRow_MIPSdspR2; +#ifdef WORDS_BIGENDIAN + WebPPackARGB = PackARGB_MIPSdspR2; +#endif WebPPackRGB = PackRGB_MIPSdspR2; } diff --git a/thirdparty/libwebp/src/dsp/common_sse2.h b/thirdparty/libwebp/src/dsp/common_sse2.h index 995d7cf4ea..e9f1ebff44 100644 --- a/thirdparty/libwebp/src/dsp/common_sse2.h +++ b/thirdparty/libwebp/src/dsp/common_sse2.h @@ -128,9 +128,9 @@ static WEBP_INLINE void VP8Transpose_2_4x4_16b( // Pack the planar buffers // rrrr... rrrr... gggg... gggg... bbbb... bbbb.... // triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ... -static WEBP_INLINE void VP8PlanarTo24b(__m128i* const in0, __m128i* const in1, - __m128i* const in2, __m128i* const in3, - __m128i* const in4, __m128i* const in5) { +static WEBP_INLINE void VP8PlanarTo24b_SSE2( + __m128i* const in0, __m128i* const in1, __m128i* const in2, + __m128i* const in3, __m128i* const in4, __m128i* const in5) { // The input is 6 registers of sixteen 8b but for the sake of explanation, // let's take 6 registers of four 8b values. // To pack, we will keep taking one every two 8b integer and move it @@ -159,10 +159,10 @@ static WEBP_INLINE void VP8PlanarTo24b(__m128i* const in0, __m128i* const in1, // Convert four packed four-channel buffers like argbargbargbargb... into the // split channels aaaaa ... rrrr ... gggg .... bbbbb ...... -static WEBP_INLINE void VP8L32bToPlanar(__m128i* const in0, - __m128i* const in1, - __m128i* const in2, - __m128i* const in3) { +static WEBP_INLINE void VP8L32bToPlanar_SSE2(__m128i* const in0, + __m128i* const in1, + __m128i* const in2, + __m128i* const in3) { // Column-wise transpose. const __m128i A0 = _mm_unpacklo_epi8(*in0, *in1); const __m128i A1 = _mm_unpackhi_epi8(*in0, *in1); diff --git a/thirdparty/libwebp/src/dsp/common_sse41.h b/thirdparty/libwebp/src/dsp/common_sse41.h new file mode 100644 index 0000000000..2f173c024a --- /dev/null +++ b/thirdparty/libwebp/src/dsp/common_sse41.h @@ -0,0 +1,132 @@ +// Copyright 2016 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 code common to several files. +// +// Author: Vincent Rabaud (vrabaud@google.com) + +#ifndef WEBP_DSP_COMMON_SSE41_H_ +#define WEBP_DSP_COMMON_SSE41_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#if defined(WEBP_USE_SSE41) +#include <smmintrin.h> + +//------------------------------------------------------------------------------ +// Channel mixing. +// Shuffles the input buffer as A0 0 0 A1 0 0 A2 ... +#define WEBP_SSE41_SHUFF(OUT, IN0, IN1) \ + OUT##0 = _mm_shuffle_epi8(*IN0, shuff0); \ + OUT##1 = _mm_shuffle_epi8(*IN0, shuff1); \ + OUT##2 = _mm_shuffle_epi8(*IN0, shuff2); \ + OUT##3 = _mm_shuffle_epi8(*IN1, shuff0); \ + OUT##4 = _mm_shuffle_epi8(*IN1, shuff1); \ + OUT##5 = _mm_shuffle_epi8(*IN1, shuff2); + +// Pack the planar buffers +// rrrr... rrrr... gggg... gggg... bbbb... bbbb.... +// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ... +static WEBP_INLINE void VP8PlanarTo24b_SSE41( + __m128i* const in0, __m128i* const in1, __m128i* const in2, + __m128i* const in3, __m128i* const in4, __m128i* const in5) { + __m128i R0, R1, R2, R3, R4, R5; + __m128i G0, G1, G2, G3, G4, G5; + __m128i B0, B1, B2, B3, B4, B5; + + // Process R. + { + const __m128i shuff0 = _mm_set_epi8( + 5, -1, -1, 4, -1, -1, 3, -1, -1, 2, -1, -1, 1, -1, -1, 0); + const __m128i shuff1 = _mm_set_epi8( + -1, 10, -1, -1, 9, -1, -1, 8, -1, -1, 7, -1, -1, 6, -1, -1); + const __m128i shuff2 = _mm_set_epi8( + -1, -1, 15, -1, -1, 14, -1, -1, 13, -1, -1, 12, -1, -1, 11, -1); + WEBP_SSE41_SHUFF(R, in0, in1) + } + + // Process G. + { + // Same as before, just shifted to the left by one and including the right + // padding. + const __m128i shuff0 = _mm_set_epi8( + -1, -1, 4, -1, -1, 3, -1, -1, 2, -1, -1, 1, -1, -1, 0, -1); + const __m128i shuff1 = _mm_set_epi8( + 10, -1, -1, 9, -1, -1, 8, -1, -1, 7, -1, -1, 6, -1, -1, 5); + const __m128i shuff2 = _mm_set_epi8( + -1, 15, -1, -1, 14, -1, -1, 13, -1, -1, 12, -1, -1, 11, -1, -1); + WEBP_SSE41_SHUFF(G, in2, in3) + } + + // Process B. + { + const __m128i shuff0 = _mm_set_epi8( + -1, 4, -1, -1, 3, -1, -1, 2, -1, -1, 1, -1, -1, 0, -1, -1); + const __m128i shuff1 = _mm_set_epi8( + -1, -1, 9, -1, -1, 8, -1, -1, 7, -1, -1, 6, -1, -1, 5, -1); + const __m128i shuff2 = _mm_set_epi8( + 15, -1, -1, 14, -1, -1, 13, -1, -1, 12, -1, -1, 11, -1, -1, 10); + WEBP_SSE41_SHUFF(B, in4, in5) + } + + // OR the different channels. + { + const __m128i RG0 = _mm_or_si128(R0, G0); + const __m128i RG1 = _mm_or_si128(R1, G1); + const __m128i RG2 = _mm_or_si128(R2, G2); + const __m128i RG3 = _mm_or_si128(R3, G3); + const __m128i RG4 = _mm_or_si128(R4, G4); + const __m128i RG5 = _mm_or_si128(R5, G5); + *in0 = _mm_or_si128(RG0, B0); + *in1 = _mm_or_si128(RG1, B1); + *in2 = _mm_or_si128(RG2, B2); + *in3 = _mm_or_si128(RG3, B3); + *in4 = _mm_or_si128(RG4, B4); + *in5 = _mm_or_si128(RG5, B5); + } +} + +#undef WEBP_SSE41_SHUFF + +// Convert four packed four-channel buffers like argbargbargbargb... into the +// split channels aaaaa ... rrrr ... gggg .... bbbbb ...... +static WEBP_INLINE void VP8L32bToPlanar_SSE41(__m128i* const in0, + __m128i* const in1, + __m128i* const in2, + __m128i* const in3) { + // aaaarrrrggggbbbb + const __m128i shuff0 = + _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0); + const __m128i A0 = _mm_shuffle_epi8(*in0, shuff0); + const __m128i A1 = _mm_shuffle_epi8(*in1, shuff0); + const __m128i A2 = _mm_shuffle_epi8(*in2, shuff0); + const __m128i A3 = _mm_shuffle_epi8(*in3, shuff0); + // A0A1R0R1 + // G0G1B0B1 + // A2A3R2R3 + // G0G1B0B1 + const __m128i B0 = _mm_unpacklo_epi32(A0, A1); + const __m128i B1 = _mm_unpackhi_epi32(A0, A1); + const __m128i B2 = _mm_unpacklo_epi32(A2, A3); + const __m128i B3 = _mm_unpackhi_epi32(A2, A3); + *in3 = _mm_unpacklo_epi64(B0, B2); + *in2 = _mm_unpackhi_epi64(B0, B2); + *in1 = _mm_unpacklo_epi64(B1, B3); + *in0 = _mm_unpackhi_epi64(B1, B3); +} + +#endif // WEBP_USE_SSE41 + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // WEBP_DSP_COMMON_SSE41_H_ diff --git a/thirdparty/libwebp/src/dsp/cost.c b/thirdparty/libwebp/src/dsp/cost.c index a732389d58..634ccc2085 100644 --- a/thirdparty/libwebp/src/dsp/cost.c +++ b/thirdparty/libwebp/src/dsp/cost.c @@ -378,12 +378,7 @@ 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; - +WEBP_DSP_INIT_FUNC(VP8EncDspCostInit) { VP8GetResidualCost = GetResidualCost_C; VP8SetResidualCoeffs = SetResidualCoeffs_C; @@ -405,8 +400,6 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInit(void) { } #endif } - - cost_last_cpuinfo_used = VP8GetCPUInfo; } //------------------------------------------------------------------------------ diff --git a/thirdparty/libwebp/src/dsp/dec.c b/thirdparty/libwebp/src/dsp/dec.c index 7e82407567..1119842dd3 100644 --- a/thirdparty/libwebp/src/dsp/dec.c +++ b/thirdparty/libwebp/src/dsp/dec.c @@ -741,12 +741,7 @@ extern void VP8DspInitMIPS32(void); extern void VP8DspInitMIPSdspR2(void); extern void VP8DspInitMSA(void); -static volatile VP8CPUInfo dec_last_cpuinfo_used = - (VP8CPUInfo)&dec_last_cpuinfo_used; - -WEBP_TSAN_IGNORE_FUNCTION void VP8DspInit(void) { - if (dec_last_cpuinfo_used == VP8GetCPUInfo) return; - +WEBP_DSP_INIT_FUNC(VP8DspInit) { VP8InitClipTables(); #if !WEBP_NEON_OMIT_C_CODE @@ -889,6 +884,4 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8DspInit(void) { assert(VP8PredChroma8[5] != NULL); assert(VP8PredChroma8[6] != NULL); assert(VP8DitherCombine8x8 != NULL); - - dec_last_cpuinfo_used = VP8GetCPUInfo; } diff --git a/thirdparty/libwebp/src/dsp/dsp.h b/thirdparty/libwebp/src/dsp/dsp.h index 99eefe092f..4ab77a5130 100644 --- a/thirdparty/libwebp/src/dsp/dsp.h +++ b/thirdparty/libwebp/src/dsp/dsp.h @@ -141,6 +141,42 @@ extern "C" { #endif #endif +#if defined(WEBP_USE_THREAD) && !defined(_WIN32) +#include <pthread.h> // NOLINT + +#define WEBP_DSP_INIT(func) do { \ + static volatile VP8CPUInfo func ## _last_cpuinfo_used = \ + (VP8CPUInfo)&func ## _last_cpuinfo_used; \ + static pthread_mutex_t func ## _lock = PTHREAD_MUTEX_INITIALIZER; \ + if (pthread_mutex_lock(&func ## _lock)) break; \ + if (func ## _last_cpuinfo_used != VP8GetCPUInfo) func(); \ + func ## _last_cpuinfo_used = VP8GetCPUInfo; \ + (void)pthread_mutex_unlock(&func ## _lock); \ +} while (0) +#else // !(defined(WEBP_USE_THREAD) && !defined(_WIN32)) +#define WEBP_DSP_INIT(func) do { \ + static volatile VP8CPUInfo func ## _last_cpuinfo_used = \ + (VP8CPUInfo)&func ## _last_cpuinfo_used; \ + if (func ## _last_cpuinfo_used == VP8GetCPUInfo) break; \ + func(); \ + func ## _last_cpuinfo_used = VP8GetCPUInfo; \ +} while (0) +#endif // defined(WEBP_USE_THREAD) && !defined(_WIN32) + +// Defines an Init + helper function that control multiple initialization of +// function pointers / tables. +/* Usage: + WEBP_DSP_INIT_FUNC(InitFunc) { + ...function body + } +*/ +#define WEBP_DSP_INIT_FUNC(name) \ + static WEBP_TSAN_IGNORE_FUNCTION void name ## _body(void); \ + WEBP_TSAN_IGNORE_FUNCTION void name(void) { \ + WEBP_DSP_INIT(name ## _body); \ + } \ + static WEBP_TSAN_IGNORE_FUNCTION void name ## _body(void) + #define WEBP_UBSAN_IGNORE_UNDEF #define WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW #if defined(__clang__) && defined(__has_attribute) @@ -166,6 +202,13 @@ extern "C" { #define WEBP_SWAP_16BIT_CSP 0 #endif +// some endian fix (e.g.: mips-gcc doesn't define __BIG_ENDIAN__) +#if !defined(WORDS_BIGENDIAN) && \ + (defined(__BIG_ENDIAN__) || defined(_M_PPC) || \ + (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__))) +#define WORDS_BIGENDIAN +#endif + typedef enum { kSSE2, kSSE3, @@ -189,7 +232,7 @@ WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo; // avoiding a compiler warning. #define WEBP_DSP_INIT_STUB(func) \ extern void func(void); \ - WEBP_TSAN_IGNORE_FUNCTION void func(void) {} + void func(void) {} //------------------------------------------------------------------------------ // Encoding @@ -578,6 +621,13 @@ void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha, int width, int inverse); void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse); +#ifdef WORDS_BIGENDIAN +// ARGB packing function: a/r/g/b input is rgba or bgra order. +extern void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, + const uint8_t* g, const uint8_t* b, int len, + uint32_t* out); +#endif + // RGB packing function. 'step' can be 3 or 4. r/g/b input is rgb or bgr order. extern void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b, int len, int step, uint32_t* out); diff --git a/thirdparty/libwebp/src/dsp/enc.c b/thirdparty/libwebp/src/dsp/enc.c index 1c807f1df7..fa23b40a30 100644 --- a/thirdparty/libwebp/src/dsp/enc.c +++ b/thirdparty/libwebp/src/dsp/enc.c @@ -740,12 +740,7 @@ extern void VP8EncDspInitMIPS32(void); extern void VP8EncDspInitMIPSdspR2(void); extern void VP8EncDspInitMSA(void); -static volatile VP8CPUInfo enc_last_cpuinfo_used = - (VP8CPUInfo)&enc_last_cpuinfo_used; - -WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInit(void) { - if (enc_last_cpuinfo_used == VP8GetCPUInfo) return; - +WEBP_DSP_INIT_FUNC(VP8EncDspInit) { VP8DspInit(); // common inverse transforms InitTables(); @@ -838,6 +833,4 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInit(void) { assert(VP8EncQuantizeBlockWHT != NULL); assert(VP8Copy4x4 != NULL); assert(VP8Copy16x8 != NULL); - - enc_last_cpuinfo_used = VP8GetCPUInfo; } diff --git a/thirdparty/libwebp/src/dsp/filters.c b/thirdparty/libwebp/src/dsp/filters.c index ca5f877da7..069a22eaef 100644 --- a/thirdparty/libwebp/src/dsp/filters.c +++ b/thirdparty/libwebp/src/dsp/filters.c @@ -238,12 +238,7 @@ extern void VP8FiltersInitMSA(void); extern void VP8FiltersInitNEON(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; - +WEBP_DSP_INIT_FUNC(VP8FiltersInit) { WebPUnfilters[WEBP_FILTER_NONE] = NULL; #if !WEBP_NEON_OMIT_C_CODE WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_C; @@ -289,6 +284,4 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInit(void) { assert(WebPFilters[WEBP_FILTER_HORIZONTAL] != NULL); assert(WebPFilters[WEBP_FILTER_VERTICAL] != NULL); assert(WebPFilters[WEBP_FILTER_GRADIENT] != NULL); - - filters_last_cpuinfo_used = VP8GetCPUInfo; } diff --git a/thirdparty/libwebp/src/dsp/lossless.c b/thirdparty/libwebp/src/dsp/lossless.c index 83f553d9ad..f9b3c182d3 100644 --- a/thirdparty/libwebp/src/dsp/lossless.c +++ b/thirdparty/libwebp/src/dsp/lossless.c @@ -577,9 +577,6 @@ extern void VP8LDspInitNEON(void); extern void VP8LDspInitMIPSdspR2(void); extern void VP8LDspInitMSA(void); -static volatile VP8CPUInfo lossless_last_cpuinfo_used = - (VP8CPUInfo)&lossless_last_cpuinfo_used; - #define COPY_PREDICTOR_ARRAY(IN, OUT) do { \ (OUT)[0] = IN##0_C; \ (OUT)[1] = IN##1_C; \ @@ -599,9 +596,7 @@ static volatile VP8CPUInfo lossless_last_cpuinfo_used = (OUT)[15] = IN##0_C; \ } while (0); -WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInit(void) { - if (lossless_last_cpuinfo_used == VP8GetCPUInfo) return; - +WEBP_DSP_INIT_FUNC(VP8LDspInit) { COPY_PREDICTOR_ARRAY(Predictor, VP8LPredictors) COPY_PREDICTOR_ARRAY(Predictor, VP8LPredictors_C) COPY_PREDICTOR_ARRAY(PredictorAdd, VP8LPredictorsAdd) @@ -658,8 +653,6 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInit(void) { assert(VP8LConvertBGRAToRGB565 != NULL); assert(VP8LMapColor32b != NULL); assert(VP8LMapColor8b != NULL); - - lossless_last_cpuinfo_used = VP8GetCPUInfo; } #undef COPY_PREDICTOR_ARRAY diff --git a/thirdparty/libwebp/src/dsp/lossless.h b/thirdparty/libwebp/src/dsp/lossless.h index a99dbda686..b2bbdfc93c 100644 --- a/thirdparty/libwebp/src/dsp/lossless.h +++ b/thirdparty/libwebp/src/dsp/lossless.h @@ -25,10 +25,6 @@ extern "C" { #endif -#ifdef WEBP_EXPERIMENTAL_FEATURES -#include "src/enc/delta_palettization_enc.h" -#endif // WEBP_EXPERIMENTAL_FEATURES - //------------------------------------------------------------------------------ // Decoding diff --git a/thirdparty/libwebp/src/dsp/lossless_enc.c b/thirdparty/libwebp/src/dsp/lossless_enc.c index 92ca3c0542..d608326fef 100644 --- a/thirdparty/libwebp/src/dsp/lossless_enc.c +++ b/thirdparty/libwebp/src/dsp/lossless_enc.c @@ -863,12 +863,7 @@ extern void VP8LEncDspInitMIPS32(void); extern void VP8LEncDspInitMIPSdspR2(void); extern void VP8LEncDspInitMSA(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; - +WEBP_DSP_INIT_FUNC(VP8LEncDspInit) { VP8LDspInit(); #if !WEBP_NEON_OMIT_C_CODE @@ -1011,8 +1006,6 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInit(void) { assert(VP8LPredictorsSub_C[13] != NULL); assert(VP8LPredictorsSub_C[14] != NULL); assert(VP8LPredictorsSub_C[15] != NULL); - - lossless_enc_last_cpuinfo_used = VP8GetCPUInfo; } //------------------------------------------------------------------------------ diff --git a/thirdparty/libwebp/src/dsp/lossless_enc_sse2.c b/thirdparty/libwebp/src/dsp/lossless_enc_sse2.c index 1eaf35ca8e..f84a9909e1 100644 --- a/thirdparty/libwebp/src/dsp/lossless_enc_sse2.c +++ b/thirdparty/libwebp/src/dsp/lossless_enc_sse2.c @@ -46,16 +46,14 @@ static void SubtractGreenFromBlueAndRed_SSE2(uint32_t* argb_data, //------------------------------------------------------------------------------ // Color Transform +#define MK_CST_16(HI, LO) \ + _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff))) + static void TransformColor_SSE2(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 mults_rb = MK_CST_16(CST_5b(m->green_to_red_), + CST_5b(m->green_to_blue_)); + const __m128i mults_b2 = MK_CST_16(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; @@ -85,12 +83,8 @@ static void CollectColorBlueTransforms_SSE2(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 mults_r = MK_CST_16(CST_5b(red_to_blue), 0); + const __m128i mults_g = MK_CST_16(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; @@ -135,9 +129,7 @@ static void CollectColorBlueTransforms_SSE2(const uint32_t* argb, int stride, static void CollectColorRedTransforms_SSE2(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 mults_g = MK_CST_16(0, CST_5b(green_to_red)); const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask const __m128i mask = _mm_set1_epi32(0xff); @@ -174,6 +166,7 @@ static void CollectColorRedTransforms_SSE2(const uint32_t* argb, int stride, } } #undef SPAN +#undef MK_CST_16 //------------------------------------------------------------------------------ diff --git a/thirdparty/libwebp/src/dsp/lossless_enc_sse41.c b/thirdparty/libwebp/src/dsp/lossless_enc_sse41.c index 3526a342d3..2e12a712eb 100644 --- a/thirdparty/libwebp/src/dsp/lossless_enc_sse41.c +++ b/thirdparty/libwebp/src/dsp/lossless_enc_sse41.c @@ -18,6 +18,9 @@ #include <smmintrin.h> #include "src/dsp/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 @@ -39,12 +42,103 @@ static void SubtractGreenFromBlueAndRed_SSE41(uint32_t* argb_data, } //------------------------------------------------------------------------------ +// Color Transform + +#define SPAN 8 +static void CollectColorBlueTransforms_SSE41(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_set1_epi16(CST_5b(red_to_blue)); + const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_blue)); + const __m128i mask_g = _mm_set1_epi16(0xff00); // green mask + const __m128i mask_gb = _mm_set1_epi32(0xffff); // green/blue mask + const __m128i mask_b = _mm_set1_epi16(0x00ff); // blue mask + const __m128i shuffler_lo = _mm_setr_epi8(-1, 2, -1, 6, -1, 10, -1, 14, -1, + -1, -1, -1, -1, -1, -1, -1); + const __m128i shuffler_hi = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, + 2, -1, 6, -1, 10, -1, 14); + 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 r0 = _mm_shuffle_epi8(in0, shuffler_lo); + const __m128i r1 = _mm_shuffle_epi8(in1, shuffler_hi); + const __m128i r = _mm_or_si128(r0, r1); // r 0 + const __m128i gb0 = _mm_and_si128(in0, mask_gb); + const __m128i gb1 = _mm_and_si128(in1, mask_gb); + const __m128i gb = _mm_packus_epi32(gb0, gb1); // g b + const __m128i g = _mm_and_si128(gb, mask_g); // g 0 + const __m128i A = _mm_mulhi_epi16(r, mults_r); // x dbr + const __m128i B = _mm_mulhi_epi16(g, mults_g); // x dbg + const __m128i C = _mm_sub_epi8(gb, B); // x b' + const __m128i D = _mm_sub_epi8(C, A); // x b'' + const __m128i E = _mm_and_si128(D, mask_b); // 0 b'' + _mm_storeu_si128((__m128i*)values, E); + 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_SSE41(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]) { + const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_red)); + const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask + const __m128i mask = _mm_set1_epi16(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 g0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0 + const __m128i g1 = _mm_and_si128(in1, mask_g); + const __m128i g = _mm_packus_epi32(g0, g1); // g 0 + const __m128i A0 = _mm_srli_epi32(in0, 16); // 0 0 | x r + const __m128i A1 = _mm_srli_epi32(in1, 16); + const __m128i A = _mm_packus_epi32(A0, A1); // x r + const __m128i B = _mm_mulhi_epi16(g, mults_g); // x dr + const __m128i C = _mm_sub_epi8(A, B); // x r' + const __m128i D = _mm_and_si128(C, mask); // 0 r' + _mm_storeu_si128((__m128i*)values, D); + 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); + } + } +} + +//------------------------------------------------------------------------------ // Entry point extern void VP8LEncDspInitSSE41(void); WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE41(void) { VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_SSE41; + VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE41; + VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE41; } #else // !WEBP_USE_SSE41 diff --git a/thirdparty/libwebp/src/dsp/lossless_sse2.c b/thirdparty/libwebp/src/dsp/lossless_sse2.c index 653b466cd6..17d7576419 100644 --- a/thirdparty/libwebp/src/dsp/lossless_sse2.c +++ b/thirdparty/libwebp/src/dsp/lossless_sse2.c @@ -453,14 +453,11 @@ static void TransformColorInverse_SSE2(const VP8LMultipliers* const m, int num_pixels, uint32_t* dst) { // 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); +#define MK_CST_16(HI, LO) \ + _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff))) + const __m128i mults_rb = MK_CST_16(CST(green_to_red_), CST(green_to_blue_)); + const __m128i mults_b2 = MK_CST_16(CST(red_to_blue_), 0); +#undef MK_CST_16 #undef CST const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks int i; @@ -503,11 +500,11 @@ static void ConvertBGRAToRGB_SSE2(const uint32_t* src, int num_pixels, __m128i in5 = _mm_loadu_si128(in + 5); __m128i in6 = _mm_loadu_si128(in + 6); __m128i in7 = _mm_loadu_si128(in + 7); - VP8L32bToPlanar(&in0, &in1, &in2, &in3); - VP8L32bToPlanar(&in4, &in5, &in6, &in7); + VP8L32bToPlanar_SSE2(&in0, &in1, &in2, &in3); + VP8L32bToPlanar_SSE2(&in4, &in5, &in6, &in7); // At this points, in1/in5 contains red only, in2/in6 green only ... // Pack the colors in 24b RGB. - VP8PlanarTo24b(&in1, &in5, &in2, &in6, &in3, &in7); + VP8PlanarTo24b_SSE2(&in1, &in5, &in2, &in6, &in3, &in7); _mm_storeu_si128(out + 0, in1); _mm_storeu_si128(out + 1, in5); _mm_storeu_si128(out + 2, in2); diff --git a/thirdparty/libwebp/src/dsp/rescaler.c b/thirdparty/libwebp/src/dsp/rescaler.c index 4b6b7834e5..f307d35056 100644 --- a/thirdparty/libwebp/src/dsp/rescaler.c +++ b/thirdparty/libwebp/src/dsp/rescaler.c @@ -204,11 +204,7 @@ extern void WebPRescalerDspInitMIPSdspR2(void); extern void WebPRescalerDspInitMSA(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; +WEBP_DSP_INIT_FUNC(WebPRescalerDspInit) { #if !defined(WEBP_REDUCE_SIZE) #if !WEBP_NEON_OMIT_C_CODE WebPRescalerExportRowExpand = WebPRescalerExportRowExpand_C; @@ -253,5 +249,4 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInit(void) { assert(WebPRescalerImportRowExpand != NULL); assert(WebPRescalerImportRowShrink != NULL); #endif // WEBP_REDUCE_SIZE - rescaler_last_cpuinfo_used = VP8GetCPUInfo; } diff --git a/thirdparty/libwebp/src/dsp/rescaler_sse2.c b/thirdparty/libwebp/src/dsp/rescaler_sse2.c index f93b204fe1..64c50deab5 100644 --- a/thirdparty/libwebp/src/dsp/rescaler_sse2.c +++ b/thirdparty/libwebp/src/dsp/rescaler_sse2.c @@ -36,7 +36,7 @@ static void LoadTwoPixels_SSE2(const uint8_t* const src, __m128i* out) { } // input: 8 bytes ABCDEFGH -> output: A0B0C0D0E0F0G0H0 -static void LoadHeightPixels_SSE2(const uint8_t* const src, __m128i* out) { +static void LoadEightPixels_SSE2(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); @@ -50,13 +50,15 @@ static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk, int accum = x_add; __m128i cur_pixels; + // SSE2 implementation only works with 16b signed arithmetic at max. + if (wrk->src_width < 8 || accum >= (1 << 15)) { + WebPRescalerImportRowExpand_C(wrk, src); + return; + } + assert(!WebPRescalerInputDone(wrk)); assert(wrk->x_expand); if (wrk->num_channels == 4) { - if (wrk->src_width < 2) { - WebPRescalerImportRowExpand_C(wrk, src); - return; - } LoadTwoPixels_SSE2(src, &cur_pixels); src += 4; while (1) { @@ -75,11 +77,7 @@ static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk, } else { int left; const uint8_t* const src_limit = src + wrk->src_width - 8; - if (wrk->src_width < 8) { - WebPRescalerImportRowExpand_C(wrk, src); - return; - } - LoadHeightPixels_SSE2(src, &cur_pixels); + LoadEightPixels_SSE2(src, &cur_pixels); src += 7; left = 7; while (1) { @@ -94,7 +92,7 @@ static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk, if (--left) { cur_pixels = _mm_srli_si128(cur_pixels, 2); } else if (src <= src_limit) { - LoadHeightPixels_SSE2(src, &cur_pixels); + LoadEightPixels_SSE2(src, &cur_pixels); src += 7; left = 7; } else { // tail diff --git a/thirdparty/libwebp/src/dsp/ssim.c b/thirdparty/libwebp/src/dsp/ssim.c index dc1b518a33..989ce8254c 100644 --- a/thirdparty/libwebp/src/dsp/ssim.c +++ b/thirdparty/libwebp/src/dsp/ssim.c @@ -139,12 +139,7 @@ VP8AccumulateSSEFunc VP8AccumulateSSE; extern void VP8SSIMDspInitSSE2(void); -static volatile VP8CPUInfo ssim_last_cpuinfo_used = - (VP8CPUInfo)&ssim_last_cpuinfo_used; - -WEBP_TSAN_IGNORE_FUNCTION void VP8SSIMDspInit(void) { - if (ssim_last_cpuinfo_used == VP8GetCPUInfo) return; - +WEBP_DSP_INIT_FUNC(VP8SSIMDspInit) { #if !defined(WEBP_REDUCE_SIZE) VP8SSIMGetClipped = SSIMGetClipped_C; VP8SSIMGet = SSIMGet_C; @@ -161,6 +156,4 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8SSIMDspInit(void) { } #endif } - - ssim_last_cpuinfo_used = VP8GetCPUInfo; } diff --git a/thirdparty/libwebp/src/dsp/upsampling.c b/thirdparty/libwebp/src/dsp/upsampling.c index e72626a82a..9b60da5bbb 100644 --- a/thirdparty/libwebp/src/dsp/upsampling.c +++ b/thirdparty/libwebp/src/dsp/upsampling.c @@ -217,13 +217,9 @@ WebPYUV444Converter WebPYUV444Converters[MODE_LAST]; extern void WebPInitYUV444ConvertersMIPSdspR2(void); extern void WebPInitYUV444ConvertersSSE2(void); +extern void WebPInitYUV444ConvertersSSE41(void); -static volatile VP8CPUInfo upsampling_last_cpuinfo_used1 = - (VP8CPUInfo)&upsampling_last_cpuinfo_used1; - -WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444Converters(void) { - if (upsampling_last_cpuinfo_used1 == VP8GetCPUInfo) return; - +WEBP_DSP_INIT_FUNC(WebPInitYUV444Converters) { WebPYUV444Converters[MODE_RGBA] = WebPYuv444ToRgba_C; WebPYUV444Converters[MODE_BGRA] = WebPYuv444ToBgra_C; WebPYUV444Converters[MODE_RGB] = WebPYuv444ToRgb_C; @@ -242,29 +238,29 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444Converters(void) { WebPInitYUV444ConvertersSSE2(); } #endif +#if defined(WEBP_USE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + WebPInitYUV444ConvertersSSE41(); + } +#endif #if defined(WEBP_USE_MIPS_DSP_R2) if (VP8GetCPUInfo(kMIPSdspR2)) { WebPInitYUV444ConvertersMIPSdspR2(); } #endif } - upsampling_last_cpuinfo_used1 = VP8GetCPUInfo; } //------------------------------------------------------------------------------ // Main calls extern void WebPInitUpsamplersSSE2(void); +extern void WebPInitUpsamplersSSE41(void); extern void WebPInitUpsamplersNEON(void); extern void WebPInitUpsamplersMIPSdspR2(void); extern void WebPInitUpsamplersMSA(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; - +WEBP_DSP_INIT_FUNC(WebPInitUpsamplers) { #ifdef FANCY_UPSAMPLING #if !WEBP_NEON_OMIT_C_CODE WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair_C; @@ -287,6 +283,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplers(void) { WebPInitUpsamplersSSE2(); } #endif +#if defined(WEBP_USE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + WebPInitUpsamplersSSE41(); + } +#endif #if defined(WEBP_USE_MIPS_DSP_R2) if (VP8GetCPUInfo(kMIPSdspR2)) { WebPInitUpsamplersMIPSdspR2(); @@ -310,6 +311,7 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplers(void) { assert(WebPUpsamplers[MODE_BGRA] != NULL); assert(WebPUpsamplers[MODE_rgbA] != NULL); assert(WebPUpsamplers[MODE_bgrA] != NULL); +#if !defined(WEBP_REDUCE_CSP) || !WEBP_NEON_OMIT_C_CODE assert(WebPUpsamplers[MODE_RGB] != NULL); assert(WebPUpsamplers[MODE_BGR] != NULL); assert(WebPUpsamplers[MODE_ARGB] != NULL); @@ -317,9 +319,9 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplers(void) { assert(WebPUpsamplers[MODE_RGB_565] != NULL); assert(WebPUpsamplers[MODE_Argb] != NULL); assert(WebPUpsamplers[MODE_rgbA_4444] != NULL); +#endif #endif // FANCY_UPSAMPLING - upsampling_last_cpuinfo_used2 = VP8GetCPUInfo; } //------------------------------------------------------------------------------ diff --git a/thirdparty/libwebp/src/dsp/upsampling_msa.c b/thirdparty/libwebp/src/dsp/upsampling_msa.c index 535ffb772c..99eea70e7d 100644 --- a/thirdparty/libwebp/src/dsp/upsampling_msa.c +++ b/thirdparty/libwebp/src/dsp/upsampling_msa.c @@ -264,6 +264,7 @@ static void YuvToBgr(int y, int u, int v, uint8_t* const bgr) { bgr[2] = Clip8(r1 >> 6); } +#if !defined(WEBP_REDUCE_CSP) static void YuvToRgb565(int y, int u, int v, uint8_t* const rgb) { const int y1 = MultHi(y, 19077); const int r1 = y1 + MultHi(v, 26149) - 14234; @@ -306,6 +307,7 @@ static void YuvToArgb(uint8_t y, uint8_t u, uint8_t v, uint8_t* const argb) { argb[0] = 0xff; YuvToRgb(y, u, v, argb + 1); } +#endif // WEBP_REDUCE_CSP static void YuvToBgra(uint8_t y, uint8_t u, uint8_t v, uint8_t* const bgra) { YuvToBgr(y, u, v, bgra); @@ -317,6 +319,7 @@ static void YuvToRgba(uint8_t y, uint8_t u, uint8_t v, uint8_t* const rgba) { rgba[3] = 0xff; } +#if !defined(WEBP_REDUCE_CSP) static void YuvToRgbLine(const uint8_t* y, const uint8_t* u, const uint8_t* v, uint8_t* dst, int length) { v16u8 R, G, B; @@ -370,6 +373,7 @@ static void YuvToBgrLine(const uint8_t* y, const uint8_t* u, memcpy(dst, temp, length * 3 * sizeof(*dst)); } } +#endif // WEBP_REDUCE_CSP static void YuvToRgbaLine(const uint8_t* y, const uint8_t* u, const uint8_t* v, uint8_t* dst, int length) { @@ -427,6 +431,7 @@ static void YuvToBgraLine(const uint8_t* y, const uint8_t* u, } } +#if !defined(WEBP_REDUCE_CSP) static void YuvToArgbLine(const uint8_t* y, const uint8_t* u, const uint8_t* v, uint8_t* dst, int length) { v16u8 R, G, B; @@ -526,6 +531,7 @@ static void YuvToRgb565Line(const uint8_t* y, const uint8_t* u, memcpy(dst, temp, length * 2 * sizeof(*dst)); } } +#endif // WEBP_REDUCE_CSP #define UPSAMPLE_32PIXELS(a, b, c, d) do { \ v16u8 s = __msa_aver_u_b(a, d); \ diff --git a/thirdparty/libwebp/src/dsp/upsampling_sse2.c b/thirdparty/libwebp/src/dsp/upsampling_sse2.c index fd5d303982..340f1e2ac2 100644 --- a/thirdparty/libwebp/src/dsp/upsampling_sse2.c +++ b/thirdparty/libwebp/src/dsp/upsampling_sse2.c @@ -104,21 +104,6 @@ static void Upsample32Pixels_SSE2(const uint8_t r1[], const uint8_t r2[], Upsample32Pixels_SSE2(r1, r2, out); \ } -#define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, \ - top_dst, bottom_dst, cur_x, num_pixels) { \ - int n; \ - 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) != NULL) { \ - for (n = 0; n < (num_pixels); ++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_SSE2((top_y) + (cur_x), r_u, r_v, (top_dst) + (cur_x) * (XSTEP)); \ @@ -135,7 +120,7 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ int uv_pos, pos; \ /* 16byte-aligned array to cache reconstructed u and v */ \ - uint8_t uv_buf[4 * 32 + 15]; \ + uint8_t uv_buf[14 * 32 + 15] = { 0 }; \ uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ uint8_t* const r_v = r_u + 32; \ \ @@ -160,11 +145,22 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ } \ if (len > 1) { \ const int left_over = ((len + 1) >> 1) - (pos >> 1); \ + uint8_t* const tmp_top_dst = r_u + 4 * 32; \ + uint8_t* const tmp_bottom_dst = tmp_top_dst + 4 * 32; \ + uint8_t* const tmp_top = tmp_bottom_dst + 4 * 32; \ + uint8_t* const tmp_bottom = (bottom_y == NULL) ? NULL : tmp_top + 32; \ 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); \ + memcpy(tmp_top, top_y + pos, len - pos); \ + if (bottom_y != NULL) memcpy(tmp_bottom, bottom_y + pos, len - pos); \ + CONVERT2RGB_32(FUNC, XSTEP, tmp_top, tmp_bottom, tmp_top_dst, \ + tmp_bottom_dst, 0); \ + memcpy(top_dst + pos * (XSTEP), tmp_top_dst, (len - pos) * (XSTEP)); \ + if (bottom_y != NULL) { \ + memcpy(bottom_dst + pos * (XSTEP), tmp_bottom_dst, \ + (len - pos) * (XSTEP)); \ + } \ } \ } diff --git a/thirdparty/libwebp/src/dsp/upsampling_sse41.c b/thirdparty/libwebp/src/dsp/upsampling_sse41.c new file mode 100644 index 0000000000..648d456027 --- /dev/null +++ b/thirdparty/libwebp/src/dsp/upsampling_sse41.c @@ -0,0 +1,239 @@ +// 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. +// ----------------------------------------------------------------------------- +// +// SSE41 version of YUV to RGB upsampling functions. +// +// Author: somnath@google.com (Somnath Banerjee) + +#include "src/dsp/dsp.h" + +#if defined(WEBP_USE_SSE41) + +#include <assert.h> +#include <smmintrin.h> +#include <string.h> +#include "src/dsp/yuv.h" + +#ifdef FANCY_UPSAMPLING + +#if !defined(WEBP_REDUCE_CSP) + +// We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows +// u = (9*a + 3*b + 3*c + d + 8) / 16 +// = (a + (a + 3*b + 3*c + d) / 8 + 1) / 2 +// = (a + m + 1) / 2 +// where m = (a + 3*b + 3*c + d) / 8 +// = ((a + b + c + d) / 2 + b + c) / 4 +// +// Let's say k = (a + b + c + d) / 4. +// We can compute k as +// k = (s + t + 1) / 2 - ((a^d) | (b^c) | (s^t)) & 1 +// where s = (a + d + 1) / 2 and t = (b + c + 1) / 2 +// +// Then m can be written as +// m = (k + t + 1) / 2 - (((b^c) & (s^t)) | (k^t)) & 1 + +// Computes out = (k + in + 1) / 2 - ((ij & (s^t)) | (k^in)) & 1 +#define GET_M(ij, in, out) do { \ + const __m128i tmp0 = _mm_avg_epu8(k, (in)); /* (k + in + 1) / 2 */ \ + const __m128i tmp1 = _mm_and_si128((ij), st); /* (ij) & (s^t) */ \ + const __m128i tmp2 = _mm_xor_si128(k, (in)); /* (k^in) */ \ + const __m128i tmp3 = _mm_or_si128(tmp1, tmp2); /* ((ij) & (s^t)) | (k^in) */\ + const __m128i tmp4 = _mm_and_si128(tmp3, one); /* & 1 -> lsb_correction */ \ + (out) = _mm_sub_epi8(tmp0, tmp4); /* (k + in + 1) / 2 - lsb_correction */ \ +} while (0) + +// pack and store two alternating pixel rows +#define PACK_AND_STORE(a, b, da, db, out) do { \ + 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((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 */ \ + const __m128i st = _mm_xor_si128(s, t); /* st = s^t */ \ + \ + const __m128i ad = _mm_xor_si128(a, d); /* ad = a^d */ \ + const __m128i bc = _mm_xor_si128(b, c); /* bc = b^c */ \ + \ + const __m128i t1 = _mm_or_si128(ad, bc); /* (a^d) | (b^c) */ \ + const __m128i t2 = _mm_or_si128(t1, st); /* (a^d) | (b^c) | (s^t) */ \ + const __m128i t3 = _mm_and_si128(t2, one); /* (a^d) | (b^c) | (s^t) & 1 */ \ + const __m128i t4 = _mm_avg_epu8(s, t); \ + const __m128i k = _mm_sub_epi8(t4, t3); /* k = (a + b + c + d) / 4 */ \ + __m128i diag1, diag2; \ + \ + GET_M(bc, t, diag1); /* diag1 = (a + 3b + 3c + d) / 8 */ \ + GET_M(ad, s, diag2); /* diag2 = (3a + b + c + 3d) / 8 */ \ + \ + /* pack the alternate pixels */ \ + 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 +static void Upsample32Pixels_SSE41(const uint8_t r1[], const uint8_t r2[], + uint8_t* const out) { + UPSAMPLE_32PIXELS(r1, r2, out); +} + +#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) { \ + uint8_t r1[17], r2[17]; \ + memcpy(r1, (tb), (num_pixels)); \ + memcpy(r2, (bb), (num_pixels)); \ + /* replicate last byte */ \ + memset(r1 + (num_pixels), r1[(num_pixels) - 1], 17 - (num_pixels)); \ + memset(r2 + (num_pixels), r2[(num_pixels) - 1], 17 - (num_pixels)); \ + /* using the shared function instead of the macro saves ~3k code size */ \ + Upsample32Pixels_SSE41(r1, r2, out); \ +} + +#define CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, \ + top_dst, bottom_dst, cur_x) do { \ + FUNC##32_SSE41((top_y) + (cur_x), r_u, r_v, (top_dst) + (cur_x) * (XSTEP)); \ + if ((bottom_y) != NULL) { \ + FUNC##32_SSE41((bottom_y) + (cur_x), r_u + 64, r_v + 64, \ + (bottom_dst) + (cur_x) * (XSTEP)); \ + } \ +} while (0) + +#define SSE4_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 uv_pos, pos; \ + /* 16byte-aligned array to cache reconstructed u and v */ \ + uint8_t uv_buf[14 * 32 + 15] = { 0 }; \ + uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ + uint8_t* const r_v = r_u + 32; \ + \ + 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); \ + } \ + } \ + /* 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); \ + } \ + if (len > 1) { \ + const int left_over = ((len + 1) >> 1) - (pos >> 1); \ + uint8_t* const tmp_top_dst = r_u + 4 * 32; \ + uint8_t* const tmp_bottom_dst = tmp_top_dst + 4 * 32; \ + uint8_t* const tmp_top = tmp_bottom_dst + 4 * 32; \ + uint8_t* const tmp_bottom = (bottom_y == NULL) ? NULL : tmp_top + 32; \ + 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); \ + memcpy(tmp_top, top_y + pos, len - pos); \ + if (bottom_y != NULL) memcpy(tmp_bottom, bottom_y + pos, len - pos); \ + CONVERT2RGB_32(FUNC, XSTEP, tmp_top, tmp_bottom, tmp_top_dst, \ + tmp_bottom_dst, 0); \ + memcpy(top_dst + pos * (XSTEP), tmp_top_dst, (len - pos) * (XSTEP)); \ + if (bottom_y != NULL) { \ + memcpy(bottom_dst + pos * (XSTEP), tmp_bottom_dst, \ + (len - pos) * (XSTEP)); \ + } \ + } \ +} + +// SSE4 variants of the fancy upsampler. +SSE4_UPSAMPLE_FUNC(UpsampleRgbLinePair_SSE41, VP8YuvToRgb, 3) +SSE4_UPSAMPLE_FUNC(UpsampleBgrLinePair_SSE41, VP8YuvToBgr, 3) + +#undef GET_M +#undef PACK_AND_STORE +#undef UPSAMPLE_32PIXELS +#undef UPSAMPLE_LAST_BLOCK +#undef CONVERT2RGB +#undef CONVERT2RGB_32 +#undef SSE4_UPSAMPLE_FUNC + +#endif // WEBP_REDUCE_CSP + +//------------------------------------------------------------------------------ +// Entry point + +extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; + +extern void WebPInitUpsamplersSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersSSE41(void) { +#if !defined(WEBP_REDUCE_CSP) + WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair_SSE41; + WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair_SSE41; +#endif // WEBP_REDUCE_CSP +} + +#endif // FANCY_UPSAMPLING + +//------------------------------------------------------------------------------ + +extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */]; +extern void WebPInitYUV444ConvertersSSE41(void); + +#define YUV444_FUNC(FUNC_NAME, CALL, CALL_C, XSTEP) \ +extern void CALL_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 */ \ + CALL_C(y + i, u + i, v + i, dst + i * (XSTEP), len - i); \ + } \ +} + +#if !defined(WEBP_REDUCE_CSP) +YUV444_FUNC(Yuv444ToRgb_SSE41, VP8YuvToRgb32_SSE41, WebPYuv444ToRgb_C, 3); +YUV444_FUNC(Yuv444ToBgr_SSE41, VP8YuvToBgr32_SSE41, WebPYuv444ToBgr_C, 3); +#endif // WEBP_REDUCE_CSP + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersSSE41(void) { +#if !defined(WEBP_REDUCE_CSP) + WebPYUV444Converters[MODE_RGB] = Yuv444ToRgb_SSE41; + WebPYUV444Converters[MODE_BGR] = Yuv444ToBgr_SSE41; +#endif // WEBP_REDUCE_CSP +} + +#else + +WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersSSE41) + +#endif // WEBP_USE_SSE41 + +#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_SSE41)) +WEBP_DSP_INIT_STUB(WebPInitUpsamplersSSE41) +#endif diff --git a/thirdparty/libwebp/src/dsp/yuv.c b/thirdparty/libwebp/src/dsp/yuv.c index bddf81fe09..14e67fc28e 100644 --- a/thirdparty/libwebp/src/dsp/yuv.c +++ b/thirdparty/libwebp/src/dsp/yuv.c @@ -71,15 +71,11 @@ void WebPSamplerProcessPlane(const uint8_t* y, int y_stride, WebPSamplerRowFunc WebPSamplers[MODE_LAST]; extern void WebPInitSamplersSSE2(void); +extern void WebPInitSamplersSSE41(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; - +WEBP_DSP_INIT_FUNC(WebPInitSamplers) { WebPSamplers[MODE_RGB] = YuvToRgbRow; WebPSamplers[MODE_RGBA] = YuvToRgbaRow; WebPSamplers[MODE_BGR] = YuvToBgrRow; @@ -99,6 +95,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplers(void) { WebPInitSamplersSSE2(); } #endif // WEBP_USE_SSE2 +#if defined(WEBP_USE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + WebPInitSamplersSSE41(); + } +#endif // WEBP_USE_SSE41 #if defined(WEBP_USE_MIPS32) if (VP8GetCPUInfo(kMIPS32)) { WebPInitSamplersMIPS32(); @@ -110,7 +111,6 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplers(void) { } #endif // WEBP_USE_MIPS_DSP_R2 } - yuv_last_cpuinfo_used = VP8GetCPUInfo; } //----------------------------------------------------------------------------- @@ -254,17 +254,13 @@ void (*WebPSharpYUVUpdateRGB)(const int16_t* ref, const int16_t* src, void (*WebPSharpYUVFilterRow)(const int16_t* A, const int16_t* B, int len, const uint16_t* best_y, uint16_t* out); -static volatile VP8CPUInfo rgba_to_yuv_last_cpuinfo_used = - (VP8CPUInfo)&rgba_to_yuv_last_cpuinfo_used; - extern void WebPInitConvertARGBToYUVSSE2(void); +extern void WebPInitConvertARGBToYUVSSE41(void); extern void WebPInitConvertARGBToYUVNEON(void); extern void WebPInitSharpYUVSSE2(void); extern void WebPInitSharpYUVNEON(void); -WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) { - if (rgba_to_yuv_last_cpuinfo_used == VP8GetCPUInfo) return; - +WEBP_DSP_INIT_FUNC(WebPInitConvertARGBToYUV) { WebPConvertARGBToY = ConvertARGBToY_C; WebPConvertARGBToUV = WebPConvertARGBToUV_C; @@ -286,6 +282,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) { WebPInitSharpYUVSSE2(); } #endif // WEBP_USE_SSE2 +#if defined(WEBP_USE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + WebPInitConvertARGBToYUVSSE41(); + } +#endif // WEBP_USE_SSE41 } #if defined(WEBP_USE_NEON) @@ -304,6 +305,4 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) { assert(WebPSharpYUVUpdateY != NULL); assert(WebPSharpYUVUpdateRGB != NULL); assert(WebPSharpYUVFilterRow != NULL); - - rgba_to_yuv_last_cpuinfo_used = VP8GetCPUInfo; } diff --git a/thirdparty/libwebp/src/dsp/yuv.h b/thirdparty/libwebp/src/dsp/yuv.h index c8a55832d4..eb787270d2 100644 --- a/thirdparty/libwebp/src/dsp/yuv.h +++ b/thirdparty/libwebp/src/dsp/yuv.h @@ -166,6 +166,19 @@ void VP8YuvToRgb56532_SSE2(const uint8_t* y, const uint8_t* u, const uint8_t* v, #endif // WEBP_USE_SSE2 +//----------------------------------------------------------------------------- +// SSE41 extra functions (mostly for upsampling_sse41.c) + +#if defined(WEBP_USE_SSE41) + +// Process 32 pixels and store the result (16b, 24b or 32b per pixel) in *dst. +void VP8YuvToRgb32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst); +void VP8YuvToBgr32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst); + +#endif // WEBP_USE_SSE41 + //------------------------------------------------------------------------------ // RGB -> YUV conversion diff --git a/thirdparty/libwebp/src/dsp/yuv_sse2.c b/thirdparty/libwebp/src/dsp/yuv_sse2.c index 6810bf8d15..baa48d5371 100644 --- a/thirdparty/libwebp/src/dsp/yuv_sse2.c +++ b/thirdparty/libwebp/src/dsp/yuv_sse2.c @@ -180,7 +180,7 @@ static WEBP_INLINE void PlanarTo24b_SSE2(__m128i* const in0, __m128i* const in1, // Repeat the same permutations twice more: // r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7 // r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7 - VP8PlanarTo24b(in0, in1, in2, in3, in4, in5); + VP8PlanarTo24b_SSE2(in0, in1, in2, in3, in4, in5); _mm_storeu_si128((__m128i*)(rgb + 0), *in0); _mm_storeu_si128((__m128i*)(rgb + 16), *in1); @@ -492,7 +492,7 @@ static WEBP_INLINE void RGB32PackedToPlanar_SSE2(const uint32_t* const argb, __m128i a1 = LOAD_16(argb + 4); __m128i a2 = LOAD_16(argb + 8); __m128i a3 = LOAD_16(argb + 12); - VP8L32bToPlanar(&a0, &a1, &a2, &a3); + VP8L32bToPlanar_SSE2(&a0, &a1, &a2, &a3); rgb[0] = _mm_unpacklo_epi8(a1, zero); rgb[1] = _mm_unpackhi_epi8(a1, zero); rgb[2] = _mm_unpacklo_epi8(a2, zero); diff --git a/thirdparty/libwebp/src/dsp/yuv_sse41.c b/thirdparty/libwebp/src/dsp/yuv_sse41.c new file mode 100644 index 0000000000..579d1f7402 --- /dev/null +++ b/thirdparty/libwebp/src/dsp/yuv_sse41.c @@ -0,0 +1,613 @@ +// 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 "src/dsp/yuv.h" + +#if defined(WEBP_USE_SSE41) + +#include "src/dsp/common_sse41.h" +#include <stdlib.h> +#include <smmintrin.h> + +//----------------------------------------------------------------------------- +// Convert spans of 32 pixels to various RGB formats for the fancy upsampler. + +// These constants are 14b fixed-point version of ITU-R BT.601 constants. +// R = (19077 * y + 26149 * v - 14234) >> 6 +// G = (19077 * y - 6419 * u - 13320 * v + 8708) >> 6 +// B = (19077 * y + 33050 * u - 17685) >> 6 +static void ConvertYUV444ToRGB_SSE41(const __m128i* const Y0, + const __m128i* const U0, + const __m128i* const V0, + __m128i* const R, + __m128i* const G, + __m128i* const B) { + const __m128i k19077 = _mm_set1_epi16(19077); + const __m128i k26149 = _mm_set1_epi16(26149); + const __m128i k14234 = _mm_set1_epi16(14234); + // 33050 doesn't fit in a signed short: only use this with unsigned arithmetic + const __m128i k33050 = _mm_set1_epi16((short)33050); + const __m128i k17685 = _mm_set1_epi16(17685); + const __m128i k6419 = _mm_set1_epi16(6419); + const __m128i k13320 = _mm_set1_epi16(13320); + const __m128i k8708 = _mm_set1_epi16(8708); + + const __m128i Y1 = _mm_mulhi_epu16(*Y0, k19077); + + const __m128i R0 = _mm_mulhi_epu16(*V0, k26149); + const __m128i R1 = _mm_sub_epi16(Y1, k14234); + const __m128i R2 = _mm_add_epi16(R1, R0); + + const __m128i G0 = _mm_mulhi_epu16(*U0, k6419); + const __m128i G1 = _mm_mulhi_epu16(*V0, k13320); + const __m128i G2 = _mm_add_epi16(Y1, k8708); + const __m128i G3 = _mm_add_epi16(G0, G1); + const __m128i G4 = _mm_sub_epi16(G2, G3); + + // be careful with the saturated *unsigned* arithmetic here! + const __m128i B0 = _mm_mulhi_epu16(*U0, k33050); + const __m128i B1 = _mm_adds_epu16(B0, Y1); + const __m128i B2 = _mm_subs_epu16(B1, k17685); + + // use logical shift for B2, which can be larger than 32767 + *R = _mm_srai_epi16(R2, 6); // range: [-14234, 30815] + *G = _mm_srai_epi16(G4, 6); // range: [-10953, 27710] + *B = _mm_srli_epi16(B2, 6); // range: [0, 34238] +} + +// Load the bytes into the *upper* part of 16b words. That's "<< 8", basically. +static WEBP_INLINE __m128i Load_HI_16_SSE41(const uint8_t* src) { + const __m128i zero = _mm_setzero_si128(); + return _mm_unpacklo_epi8(zero, _mm_loadl_epi64((const __m128i*)src)); +} + +// Load and replicate the U/V samples +static WEBP_INLINE __m128i Load_UV_HI_8_SSE41(const uint8_t* src) { + const __m128i zero = _mm_setzero_si128(); + const __m128i tmp0 = _mm_cvtsi32_si128(*(const uint32_t*)src); + const __m128i tmp1 = _mm_unpacklo_epi8(zero, tmp0); + return _mm_unpacklo_epi16(tmp1, tmp1); // replicate samples +} + +// Convert 32 samples of YUV444 to R/G/B +static void YUV444ToRGB_SSE41(const uint8_t* const y, + const uint8_t* const u, + const uint8_t* const v, + __m128i* const R, __m128i* const G, + __m128i* const B) { + const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_HI_16_SSE41(u), + V0 = Load_HI_16_SSE41(v); + ConvertYUV444ToRGB_SSE41(&Y0, &U0, &V0, R, G, B); +} + +// Convert 32 samples of YUV420 to R/G/B +static void YUV420ToRGB_SSE41(const uint8_t* const y, + const uint8_t* const u, + const uint8_t* const v, + __m128i* const R, __m128i* const G, + __m128i* const B) { + const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_UV_HI_8_SSE41(u), + V0 = Load_UV_HI_8_SSE41(v); + ConvertYUV444ToRGB_SSE41(&Y0, &U0, &V0, R, G, B); +} + +// Pack the planar buffers +// rrrr... rrrr... gggg... gggg... bbbb... bbbb.... +// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ... +static WEBP_INLINE void PlanarTo24b_SSE41( + __m128i* const in0, __m128i* const in1, __m128i* const in2, + __m128i* const in3, __m128i* const in4, __m128i* const in5, + uint8_t* const rgb) { + // The input is 6 registers of sixteen 8b but for the sake of explanation, + // let's take 6 registers of four 8b values. + // To pack, we will keep taking one every two 8b integer and move it + // around as follows: + // Input: + // r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7 + // Split the 6 registers in two sets of 3 registers: the first set as the even + // 8b bytes, the second the odd ones: + // r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7 + // Repeat the same permutations twice more: + // r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7 + // r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7 + VP8PlanarTo24b_SSE41(in0, in1, in2, in3, in4, in5); + + _mm_storeu_si128((__m128i*)(rgb + 0), *in0); + _mm_storeu_si128((__m128i*)(rgb + 16), *in1); + _mm_storeu_si128((__m128i*)(rgb + 32), *in2); + _mm_storeu_si128((__m128i*)(rgb + 48), *in3); + _mm_storeu_si128((__m128i*)(rgb + 64), *in4); + _mm_storeu_si128((__m128i*)(rgb + 80), *in5); +} + +void VP8YuvToRgb32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst) { + __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; + __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5; + + YUV444ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0); + YUV444ToRGB_SSE41(y + 8, u + 8, v + 8, &R1, &G1, &B1); + YUV444ToRGB_SSE41(y + 16, u + 16, v + 16, &R2, &G2, &B2); + YUV444ToRGB_SSE41(y + 24, u + 24, v + 24, &R3, &G3, &B3); + + // Cast to 8b and store as RRRRGGGGBBBB. + rgb0 = _mm_packus_epi16(R0, R1); + rgb1 = _mm_packus_epi16(R2, R3); + rgb2 = _mm_packus_epi16(G0, G1); + rgb3 = _mm_packus_epi16(G2, G3); + rgb4 = _mm_packus_epi16(B0, B1); + rgb5 = _mm_packus_epi16(B2, B3); + + // Pack as RGBRGBRGBRGB. + PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst); +} + +void VP8YuvToBgr32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v, + uint8_t* dst) { + __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; + __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5; + + YUV444ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0); + YUV444ToRGB_SSE41(y + 8, u + 8, v + 8, &R1, &G1, &B1); + YUV444ToRGB_SSE41(y + 16, u + 16, v + 16, &R2, &G2, &B2); + YUV444ToRGB_SSE41(y + 24, u + 24, v + 24, &R3, &G3, &B3); + + // Cast to 8b and store as BBBBGGGGRRRR. + bgr0 = _mm_packus_epi16(B0, B1); + bgr1 = _mm_packus_epi16(B2, B3); + bgr2 = _mm_packus_epi16(G0, G1); + bgr3 = _mm_packus_epi16(G2, G3); + bgr4 = _mm_packus_epi16(R0, R1); + bgr5= _mm_packus_epi16(R2, R3); + + // Pack as BGRBGRBGRBGR. + PlanarTo24b_SSE41(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst); +} + +//----------------------------------------------------------------------------- +// Arbitrary-length row conversion functions + +static void YuvToRgbRow_SSE41(const uint8_t* y, + const uint8_t* u, const uint8_t* v, + uint8_t* dst, int len) { + int n; + for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) { + __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; + __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5; + + YUV420ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0); + YUV420ToRGB_SSE41(y + 8, u + 4, v + 4, &R1, &G1, &B1); + YUV420ToRGB_SSE41(y + 16, u + 8, v + 8, &R2, &G2, &B2); + YUV420ToRGB_SSE41(y + 24, u + 12, v + 12, &R3, &G3, &B3); + + // Cast to 8b and store as RRRRGGGGBBBB. + rgb0 = _mm_packus_epi16(R0, R1); + rgb1 = _mm_packus_epi16(R2, R3); + rgb2 = _mm_packus_epi16(G0, G1); + rgb3 = _mm_packus_epi16(G2, G3); + rgb4 = _mm_packus_epi16(B0, B1); + rgb5 = _mm_packus_epi16(B2, B3); + + // Pack as RGBRGBRGBRGB. + PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst); + + y += 32; + u += 16; + v += 16; + } + for (; n < len; ++n) { // Finish off + VP8YuvToRgb(y[0], u[0], v[0], dst); + dst += 3; + y += 1; + u += (n & 1); + v += (n & 1); + } +} + +static void YuvToBgrRow_SSE41(const uint8_t* y, + const uint8_t* u, const uint8_t* v, + uint8_t* dst, int len) { + int n; + for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) { + __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; + __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5; + + YUV420ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0); + YUV420ToRGB_SSE41(y + 8, u + 4, v + 4, &R1, &G1, &B1); + YUV420ToRGB_SSE41(y + 16, u + 8, v + 8, &R2, &G2, &B2); + YUV420ToRGB_SSE41(y + 24, u + 12, v + 12, &R3, &G3, &B3); + + // Cast to 8b and store as BBBBGGGGRRRR. + bgr0 = _mm_packus_epi16(B0, B1); + bgr1 = _mm_packus_epi16(B2, B3); + bgr2 = _mm_packus_epi16(G0, G1); + bgr3 = _mm_packus_epi16(G2, G3); + bgr4 = _mm_packus_epi16(R0, R1); + bgr5 = _mm_packus_epi16(R2, R3); + + // Pack as BGRBGRBGRBGR. + PlanarTo24b_SSE41(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst); + + y += 32; + u += 16; + v += 16; + } + for (; n < len; ++n) { // Finish off + VP8YuvToBgr(y[0], u[0], v[0], dst); + dst += 3; + y += 1; + u += (n & 1); + v += (n & 1); + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitSamplersSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersSSE41(void) { + WebPSamplers[MODE_RGB] = YuvToRgbRow_SSE41; + WebPSamplers[MODE_BGR] = YuvToBgrRow_SSE41; +} + +//------------------------------------------------------------------------------ +// 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)) + +#define WEBP_SSE41_SHUFF(OUT) do { \ + const __m128i tmp0 = _mm_shuffle_epi8(A0, shuff0); \ + const __m128i tmp1 = _mm_shuffle_epi8(A1, shuff1); \ + const __m128i tmp2 = _mm_shuffle_epi8(A2, shuff2); \ + const __m128i tmp3 = _mm_shuffle_epi8(A3, shuff0); \ + const __m128i tmp4 = _mm_shuffle_epi8(A4, shuff1); \ + const __m128i tmp5 = _mm_shuffle_epi8(A5, shuff2); \ + \ + /* OR everything to get one channel */ \ + const __m128i tmp6 = _mm_or_si128(tmp0, tmp1); \ + const __m128i tmp7 = _mm_or_si128(tmp3, tmp4); \ + out[OUT + 0] = _mm_or_si128(tmp6, tmp2); \ + out[OUT + 1] = _mm_or_si128(tmp7, tmp5); \ +} while (0); + +// Unpack the 8b input rgbrgbrgbrgb ... as contiguous registers: +// rrrr... rrrr... gggg... gggg... bbbb... bbbb.... +// Similar to PlanarTo24bHelper(), but in reverse order. +static WEBP_INLINE void RGB24PackedToPlanar_SSE41( + const uint8_t* const rgb, __m128i* const out /*out[6]*/) { + const __m128i A0 = _mm_loadu_si128((const __m128i*)(rgb + 0)); + const __m128i A1 = _mm_loadu_si128((const __m128i*)(rgb + 16)); + const __m128i A2 = _mm_loadu_si128((const __m128i*)(rgb + 32)); + const __m128i A3 = _mm_loadu_si128((const __m128i*)(rgb + 48)); + const __m128i A4 = _mm_loadu_si128((const __m128i*)(rgb + 64)); + const __m128i A5 = _mm_loadu_si128((const __m128i*)(rgb + 80)); + + // Compute RR. + { + const __m128i shuff0 = _mm_set_epi8( + -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 15, 12, 9, 6, 3, 0); + const __m128i shuff1 = _mm_set_epi8( + -1, -1, -1, -1, -1, 14, 11, 8, 5, 2, -1, -1, -1, -1, -1, -1); + const __m128i shuff2 = _mm_set_epi8( + 13, 10, 7, 4, 1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1); + WEBP_SSE41_SHUFF(0) + } + // Compute GG. + { + const __m128i shuff0 = _mm_set_epi8( + -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 13, 10, 7, 4, 1); + const __m128i shuff1 = _mm_set_epi8( + -1, -1, -1, -1, -1, 15, 12, 9, 6, 3, 0, -1, -1, -1, -1, -1); + const __m128i shuff2 = _mm_set_epi8( + 14, 11, 8, 5, 2, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1); + WEBP_SSE41_SHUFF(2) + } + // Compute BB. + { + const __m128i shuff0 = _mm_set_epi8( + -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 14, 11, 8, 5, 2); + const __m128i shuff1 = _mm_set_epi8( + -1, -1, -1, -1, -1, -1, 13, 10, 7, 4, 1, -1, -1, -1, -1, -1); + const __m128i shuff2 = _mm_set_epi8( + 15, 12, 9, 6, 3, 0, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1); + WEBP_SSE41_SHUFF(4) + } +} + +#undef WEBP_SSE41_SHUFF + +// Convert 8 packed ARGB to r[], g[], b[] +static WEBP_INLINE void RGB32PackedToPlanar_SSE41( + const uint32_t* const argb, __m128i* const rgb /*in[6]*/) { + const __m128i zero = _mm_setzero_si128(); + __m128i a0 = LOAD_16(argb + 0); + __m128i a1 = LOAD_16(argb + 4); + __m128i a2 = LOAD_16(argb + 8); + __m128i a3 = LOAD_16(argb + 12); + VP8L32bToPlanar_SSE41(&a0, &a1, &a2, &a3); + rgb[0] = _mm_unpacklo_epi8(a1, zero); + rgb[1] = _mm_unpackhi_epi8(a1, zero); + rgb[2] = _mm_unpacklo_epi8(a2, zero); + rgb[3] = _mm_unpackhi_epi8(a2, zero); + rgb[4] = _mm_unpacklo_epi8(a3, zero); + rgb[5] = _mm_unpackhi_epi8(a3, 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_SSE41(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_SSE41(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_SSE41(const uint8_t* rgb, uint8_t* y, int width) { + const int max_width = width & ~31; + int i; + for (i = 0; i < max_width; rgb += 3 * 16 * 2) { + __m128i rgb_plane[6]; + int j; + + RGB24PackedToPlanar_SSE41(rgb, rgb_plane); + + for (j = 0; j < 2; ++j, i += 16) { + const __m128i zero = _mm_setzero_si128(); + __m128i r, g, b, Y0, Y1; + + // Convert to 16-bit Y. + r = _mm_unpacklo_epi8(rgb_plane[0 + j], zero); + g = _mm_unpacklo_epi8(rgb_plane[2 + j], zero); + b = _mm_unpacklo_epi8(rgb_plane[4 + j], zero); + ConvertRGBToY_SSE41(&r, &g, &b, &Y0); + + // Convert to 16-bit Y. + r = _mm_unpackhi_epi8(rgb_plane[0 + j], zero); + g = _mm_unpackhi_epi8(rgb_plane[2 + j], zero); + b = _mm_unpackhi_epi8(rgb_plane[4 + j], zero); + ConvertRGBToY_SSE41(&r, &g, &b, &Y1); + + // Cast to 8-bit and store. + 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_SSE41(const uint8_t* bgr, uint8_t* y, int width) { + const int max_width = width & ~31; + int i; + for (i = 0; i < max_width; bgr += 3 * 16 * 2) { + __m128i bgr_plane[6]; + int j; + + RGB24PackedToPlanar_SSE41(bgr, bgr_plane); + + for (j = 0; j < 2; ++j, i += 16) { + const __m128i zero = _mm_setzero_si128(); + __m128i r, g, b, Y0, Y1; + + // Convert to 16-bit Y. + b = _mm_unpacklo_epi8(bgr_plane[0 + j], zero); + g = _mm_unpacklo_epi8(bgr_plane[2 + j], zero); + r = _mm_unpacklo_epi8(bgr_plane[4 + j], zero); + ConvertRGBToY_SSE41(&r, &g, &b, &Y0); + + // Convert to 16-bit Y. + b = _mm_unpackhi_epi8(bgr_plane[0 + j], zero); + g = _mm_unpackhi_epi8(bgr_plane[2 + j], zero); + r = _mm_unpackhi_epi8(bgr_plane[4 + j], zero); + ConvertRGBToY_SSE41(&r, &g, &b, &Y1); + + // Cast to 8-bit and store. + 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_SSE41(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 Y0, Y1, rgb[6]; + RGB32PackedToPlanar_SSE41(&argb[i], rgb); + ConvertRGBToY_SSE41(&rgb[0], &rgb[2], &rgb[4], &Y0); + ConvertRGBToY_SSE41(&rgb[1], &rgb[3], &rgb[5], &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_SSE41(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_SSE41(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 rgb[6], U0, V0, U1, V1; + RGB32PackedToPlanar_SSE41(&argb[i], rgb); + HorizontalAddPack_SSE41(&rgb[0], &rgb[1], &rgb[0]); + HorizontalAddPack_SSE41(&rgb[2], &rgb[3], &rgb[2]); + HorizontalAddPack_SSE41(&rgb[4], &rgb[5], &rgb[4]); + ConvertRGBToUV_SSE41(&rgb[0], &rgb[2], &rgb[4], &U0, &V0); + + RGB32PackedToPlanar_SSE41(&argb[i + 16], rgb); + HorizontalAddPack_SSE41(&rgb[0], &rgb[1], &rgb[0]); + HorizontalAddPack_SSE41(&rgb[2], &rgb[3], &rgb[2]); + HorizontalAddPack_SSE41(&rgb[4], &rgb[5], &rgb[4]); + ConvertRGBToUV_SSE41(&rgb[0], &rgb[2], &rgb[4], &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_SSE41( + 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 | ... + // aarrggbb as 16-bit. + const __m128i shuff0 = + _mm_set_epi8(-1, -1, -1, -1, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0); + const __m128i shuff1 = + _mm_set_epi8(13, 12, 5, 4, -1, -1, -1, -1, 11, 10, 3, 2, 9, 8, 1, 0); + const __m128i A0 = _mm_shuffle_epi8(in0, shuff0); + const __m128i A1 = _mm_shuffle_epi8(in1, shuff1); + const __m128i A2 = _mm_shuffle_epi8(in2, shuff0); + const __m128i A3 = _mm_shuffle_epi8(in3, shuff1); + // R0R1G0G1 + // B0B1**** + // R2R3G2G3 + // B2B3**** + // (OR is used to free port 5 for the unpack) + const __m128i B0 = _mm_unpacklo_epi32(A0, A1); + const __m128i B1 = _mm_or_si128(A0, A1); + const __m128i B2 = _mm_unpacklo_epi32(A2, A3); + const __m128i B3 = _mm_or_si128(A2, A3); + // Gather the channels. + *r = _mm_unpacklo_epi64(B0, B2); + *g = _mm_unpackhi_epi64(B0, B2); + *b = _mm_unpackhi_epi64(B1, B3); +} + +static void ConvertRGBA32ToUV_SSE41(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_SSE41(rgb + 0, &r, &g, &b); + ConvertRGBToUV_SSE41(&r, &g, &b, &U0, &V0); + RGBA32PackedToPlanar_16b_SSE41(rgb + 32, &r, &g, &b); + ConvertRGBToUV_SSE41(&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 WebPInitConvertARGBToYUVSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVSSE41(void) { + WebPConvertARGBToY = ConvertARGBToY_SSE41; + WebPConvertARGBToUV = ConvertARGBToUV_SSE41; + + WebPConvertRGB24ToY = ConvertRGB24ToY_SSE41; + WebPConvertBGR24ToY = ConvertBGR24ToY_SSE41; + + WebPConvertRGBA32ToUV = ConvertRGBA32ToUV_SSE41; +} + +//------------------------------------------------------------------------------ + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(WebPInitSamplersSSE41) +WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVSSE41) + +#endif // WEBP_USE_SSE41 |