diff options
author | Juan Linietsky <reduzio@gmail.com> | 2015-12-04 10:18:28 -0300 |
---|---|---|
committer | Juan Linietsky <reduzio@gmail.com> | 2015-12-04 10:18:28 -0300 |
commit | da113fe40d0a9410859912473d53e43903dc6c8e (patch) | |
tree | 23c6019a28a11d67241789721d1feecdd19410e6 /drivers/webp/dsp | |
parent | 064fd762fae75371658e773a3acf39616e813b08 (diff) |
-Upgraded webp to a MUCH newer version. Hoping it fixes some bugs in the process. Keeping old version just in case for now.
-Added ability to convert xml and tscn scenes to binary on export, makes loading of larger scenes faster
Diffstat (limited to 'drivers/webp/dsp')
-rw-r--r-- | drivers/webp/dsp/cpu.c | 95 | ||||
-rw-r--r-- | drivers/webp/dsp/dec.c | 338 | ||||
-rw-r--r-- | drivers/webp/dsp/dec_neon.c | 1478 | ||||
-rw-r--r-- | drivers/webp/dsp/dec_sse2.c | 1056 | ||||
-rw-r--r-- | drivers/webp/dsp/dsp.h | 368 | ||||
-rw-r--r-- | drivers/webp/dsp/enc.c | 347 | ||||
-rw-r--r-- | drivers/webp/dsp/enc_sse2.c | 1364 | ||||
-rw-r--r-- | drivers/webp/dsp/lossless.c | 1013 | ||||
-rw-r--r-- | drivers/webp/dsp/lossless.h | 292 | ||||
-rw-r--r-- | drivers/webp/dsp/upsampling.c | 259 | ||||
-rw-r--r-- | drivers/webp/dsp/upsampling_sse2.c | 196 | ||||
-rw-r--r-- | drivers/webp/dsp/yuv.c | 264 | ||||
-rw-r--r-- | drivers/webp/dsp/yuv.h | 287 |
13 files changed, 5123 insertions, 2234 deletions
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_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/dsp.h b/drivers/webp/dsp/dsp.h index fd686a8532..8395df40e4 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_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/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..ee6771333f 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/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_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 |