diff options
Diffstat (limited to 'src/3rdparty/libwebp/src/dsp')
21 files changed, 440 insertions, 239 deletions
diff --git a/src/3rdparty/libwebp/src/dsp/alpha_processing.c b/src/3rdparty/libwebp/src/dsp/alpha_processing.c index 3a27990..1892929 100644 --- a/src/3rdparty/libwebp/src/dsp/alpha_processing.c +++ b/src/3rdparty/libwebp/src/dsp/alpha_processing.c @@ -157,7 +157,8 @@ void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse) { } } -void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha, +void WebPMultRow_C(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse) { int x; for (x = 0; x < width; ++x) { @@ -178,7 +179,8 @@ void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha, #undef MFIX void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse); -void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha, +void (*WebPMultRow)(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse); //------------------------------------------------------------------------------ @@ -193,8 +195,8 @@ void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows, } } -void WebPMultRows(uint8_t* ptr, int stride, - const uint8_t* alpha, int alpha_stride, +void WebPMultRows(uint8_t* WEBP_RESTRICT ptr, int stride, + const uint8_t* WEBP_RESTRICT alpha, int alpha_stride, int width, int num_rows, int inverse) { int n; for (n = 0; n < num_rows; ++n) { @@ -290,9 +292,9 @@ static void ApplyAlphaMultiply_16b_C(uint8_t* rgba4444, } #if !WEBP_NEON_OMIT_C_CODE -static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride, +static int DispatchAlpha_C(const uint8_t* WEBP_RESTRICT alpha, int alpha_stride, int width, int height, - uint8_t* dst, int dst_stride) { + uint8_t* WEBP_RESTRICT dst, int dst_stride) { uint32_t alpha_mask = 0xff; int i, j; @@ -309,9 +311,10 @@ static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride, return (alpha_mask != 0xff); } -static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint32_t* dst, int dst_stride) { +static void DispatchAlphaToGreen_C(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint32_t* WEBP_RESTRICT dst, + int dst_stride) { int i, j; for (j = 0; j < height; ++j) { for (i = 0; i < width; ++i) { @@ -322,9 +325,9 @@ static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride, } } -static int ExtractAlpha_C(const uint8_t* argb, int argb_stride, +static int ExtractAlpha_C(const uint8_t* WEBP_RESTRICT argb, int argb_stride, int width, int height, - uint8_t* alpha, int alpha_stride) { + uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { uint8_t alpha_mask = 0xff; int i, j; @@ -340,7 +343,8 @@ static int ExtractAlpha_C(const uint8_t* argb, int argb_stride, return (alpha_mask == 0xff); } -static void ExtractGreen_C(const uint32_t* argb, uint8_t* alpha, int size) { +static void ExtractGreen_C(const uint32_t* WEBP_RESTRICT argb, + uint8_t* WEBP_RESTRICT alpha, int size) { int i; for (i = 0; i < size; ++i) alpha[i] = argb[i] >> 8; } @@ -372,8 +376,11 @@ static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) { } #ifdef WORDS_BIGENDIAN -static void PackARGB_C(const uint8_t* a, const uint8_t* r, const uint8_t* g, - const uint8_t* b, int len, uint32_t* out) { +static void PackARGB_C(const uint8_t* WEBP_RESTRICT a, + const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, uint32_t* WEBP_RESTRICT out) { int i; for (i = 0; i < len; ++i) { out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]); @@ -381,8 +388,10 @@ static void PackARGB_C(const uint8_t* a, const uint8_t* r, const uint8_t* g, } #endif -static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b, - int len, int step, uint32_t* out) { +static void PackRGB_C(const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, int step, uint32_t* WEBP_RESTRICT out) { int i, offset = 0; for (i = 0; i < len; ++i) { out[i] = MakeARGB32(0xff, r[offset], g[offset], b[offset]); @@ -392,16 +401,22 @@ static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b, void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int); void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int); -int (*WebPDispatchAlpha)(const uint8_t*, int, int, int, uint8_t*, int); -void (*WebPDispatchAlphaToGreen)(const uint8_t*, int, int, int, uint32_t*, int); -int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int); -void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size); +int (*WebPDispatchAlpha)(const uint8_t* WEBP_RESTRICT, int, int, int, + uint8_t* WEBP_RESTRICT, int); +void (*WebPDispatchAlphaToGreen)(const uint8_t* WEBP_RESTRICT, int, int, int, + uint32_t* WEBP_RESTRICT, int); +int (*WebPExtractAlpha)(const uint8_t* WEBP_RESTRICT, int, int, int, + uint8_t* WEBP_RESTRICT, int); +void (*WebPExtractGreen)(const uint32_t* WEBP_RESTRICT argb, + uint8_t* WEBP_RESTRICT alpha, int size); #ifdef WORDS_BIGENDIAN void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, const uint8_t* g, const uint8_t* b, int, uint32_t*); #endif -void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b, - int len, int step, uint32_t* out); +void (*WebPPackRGB)(const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, int step, uint32_t* WEBP_RESTRICT out); int (*WebPHasAlpha8b)(const uint8_t* src, int length); int (*WebPHasAlpha32b)(const uint8_t* src, int length); @@ -438,10 +453,10 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitAlphaProcessingSSE2(); -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitAlphaProcessingSSE41(); } @@ -455,7 +470,7 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { WebPInitAlphaProcessingNEON(); diff --git a/src/3rdparty/libwebp/src/dsp/alpha_processing_neon.c b/src/3rdparty/libwebp/src/dsp/alpha_processing_neon.c index 9d55421..9e0ace9 100644 --- a/src/3rdparty/libwebp/src/dsp/alpha_processing_neon.c +++ b/src/3rdparty/libwebp/src/dsp/alpha_processing_neon.c @@ -80,9 +80,9 @@ static void ApplyAlphaMultiply_NEON(uint8_t* rgba, int alpha_first, //------------------------------------------------------------------------------ -static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint8_t* dst, int dst_stride) { +static int DispatchAlpha_NEON(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint8_t* WEBP_RESTRICT dst, int dst_stride) { uint32_t alpha_mask = 0xffffffffu; uint8x8_t mask8 = vdup_n_u8(0xff); uint32_t tmp[2]; @@ -112,9 +112,10 @@ static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride, return (alpha_mask != 0xffffffffu); } -static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint32_t* dst, int dst_stride) { +static void DispatchAlphaToGreen_NEON(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint32_t* WEBP_RESTRICT dst, + int dst_stride) { int i, j; uint8x8x4_t greens; // leave A/R/B channels zero'd. greens.val[0] = vdup_n_u8(0); @@ -131,9 +132,9 @@ static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride, } } -static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride, +static int ExtractAlpha_NEON(const uint8_t* WEBP_RESTRICT argb, int argb_stride, int width, int height, - uint8_t* alpha, int alpha_stride) { + uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { uint32_t alpha_mask = 0xffffffffu; uint8x8_t mask8 = vdup_n_u8(0xff); uint32_t tmp[2]; @@ -161,8 +162,8 @@ static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride, return (alpha_mask == 0xffffffffu); } -static void ExtractGreen_NEON(const uint32_t* argb, - uint8_t* alpha, int size) { +static void ExtractGreen_NEON(const uint32_t* WEBP_RESTRICT argb, + uint8_t* WEBP_RESTRICT alpha, int size) { int i; for (i = 0; i + 16 <= size; i += 16) { const uint8x16x4_t rgbX = vld4q_u8((const uint8_t*)(argb + i)); diff --git a/src/3rdparty/libwebp/src/dsp/alpha_processing_sse2.c b/src/3rdparty/libwebp/src/dsp/alpha_processing_sse2.c index f6c6e0f..a5f8c9f 100644 --- a/src/3rdparty/libwebp/src/dsp/alpha_processing_sse2.c +++ b/src/3rdparty/libwebp/src/dsp/alpha_processing_sse2.c @@ -18,9 +18,9 @@ //------------------------------------------------------------------------------ -static int DispatchAlpha_SSE2(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint8_t* dst, int dst_stride) { +static int DispatchAlpha_SSE2(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint8_t* WEBP_RESTRICT dst, int dst_stride) { // alpha_and stores an 'and' operation of all the alpha[] values. The final // value is not 0xff if any of the alpha[] is not equal to 0xff. uint32_t alpha_and = 0xff; @@ -72,9 +72,10 @@ static int DispatchAlpha_SSE2(const uint8_t* alpha, int alpha_stride, return (alpha_and != 0xff); } -static void DispatchAlphaToGreen_SSE2(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint32_t* dst, int dst_stride) { +static void DispatchAlphaToGreen_SSE2(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint32_t* WEBP_RESTRICT dst, + int dst_stride) { int i, j; const __m128i zero = _mm_setzero_si128(); const int limit = width & ~15; @@ -98,9 +99,9 @@ static void DispatchAlphaToGreen_SSE2(const uint8_t* alpha, int alpha_stride, } } -static int ExtractAlpha_SSE2(const uint8_t* argb, int argb_stride, +static int ExtractAlpha_SSE2(const uint8_t* WEBP_RESTRICT argb, int argb_stride, int width, int height, - uint8_t* alpha, int alpha_stride) { + uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { // alpha_and stores an 'and' operation of all the alpha[] values. The final // value is not 0xff if any of the alpha[] is not equal to 0xff. uint32_t alpha_and = 0xff; @@ -317,7 +318,8 @@ static void MultARGBRow_SSE2(uint32_t* const ptr, int width, int inverse) { if (width > 0) WebPMultARGBRow_C(ptr + x, width, inverse); } -static void MultRow_SSE2(uint8_t* const ptr, const uint8_t* const alpha, +static void MultRow_SSE2(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse) { int x = 0; if (!inverse) { diff --git a/src/3rdparty/libwebp/src/dsp/alpha_processing_sse41.c b/src/3rdparty/libwebp/src/dsp/alpha_processing_sse41.c index 56040f9..cdf877c 100644 --- a/src/3rdparty/libwebp/src/dsp/alpha_processing_sse41.c +++ b/src/3rdparty/libwebp/src/dsp/alpha_processing_sse41.c @@ -19,9 +19,9 @@ //------------------------------------------------------------------------------ -static int ExtractAlpha_SSE41(const uint8_t* argb, int argb_stride, - int width, int height, - uint8_t* alpha, int alpha_stride) { +static int ExtractAlpha_SSE41(const uint8_t* WEBP_RESTRICT argb, + int argb_stride, int width, int height, + uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { // alpha_and stores an 'and' operation of all the alpha[] values. The final // value is not 0xff if any of the alpha[] is not equal to 0xff. uint32_t alpha_and = 0xff; diff --git a/src/3rdparty/libwebp/src/dsp/cost.c b/src/3rdparty/libwebp/src/dsp/cost.c index cc681cd..460ec4f 100644 --- a/src/3rdparty/libwebp/src/dsp/cost.c +++ b/src/3rdparty/libwebp/src/dsp/cost.c @@ -395,12 +395,12 @@ WEBP_DSP_INIT_FUNC(VP8EncDspCostInit) { VP8EncDspCostInitMIPSdspR2(); } #endif -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8EncDspCostInitSSE2(); } #endif -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (VP8GetCPUInfo(kNEON)) { VP8EncDspCostInitNEON(); } diff --git a/src/3rdparty/libwebp/src/dsp/cpu.c b/src/3rdparty/libwebp/src/dsp/cpu.c index 4ca90d8..3145e19 100644 --- a/src/3rdparty/libwebp/src/dsp/cpu.c +++ b/src/3rdparty/libwebp/src/dsp/cpu.c @@ -189,17 +189,17 @@ VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo; // Use compile flags as an indicator of SIMD support instead of a runtime check. static int wasmCPUInfo(CPUFeature feature) { switch (feature) { -#ifdef WEBP_USE_SSE2 +#ifdef WEBP_HAVE_SSE2 case kSSE2: return 1; #endif -#ifdef WEBP_USE_SSE41 +#ifdef WEBP_HAVE_SSE41 case kSSE3: case kSlowSSSE3: case kSSE4_1: return 1; #endif -#ifdef WEBP_USE_NEON +#ifdef WEBP_HAVE_NEON case kNEON: return 1; #endif @@ -209,9 +209,10 @@ static int wasmCPUInfo(CPUFeature feature) { return 0; } VP8CPUInfo VP8GetCPUInfo = wasmCPUInfo; -#elif defined(WEBP_USE_NEON) -// define a dummy function to enable turning off NEON at runtime by setting -// VP8DecGetCPUInfo = NULL +#elif defined(WEBP_HAVE_NEON) +// In most cases this function doesn't check for NEON support (it's assumed by +// the configuration), but enables turning off NEON at runtime, for testing +// purposes, by setting VP8DecGetCPUInfo = NULL. static int armCPUInfo(CPUFeature feature) { if (feature != kNEON) return 0; #if defined(__linux__) && defined(WEBP_HAVE_NEON_RTCD) diff --git a/src/3rdparty/libwebp/src/dsp/dec.c b/src/3rdparty/libwebp/src/dsp/dec.c index 1119842..537c701 100644 --- a/src/3rdparty/libwebp/src/dsp/dec.c +++ b/src/3rdparty/libwebp/src/dsp/dec.c @@ -807,10 +807,10 @@ WEBP_DSP_INIT_FUNC(VP8DspInit) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8DspInitSSE2(); -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { VP8DspInitSSE41(); } @@ -834,7 +834,7 @@ WEBP_DSP_INIT_FUNC(VP8DspInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { VP8DspInitNEON(); diff --git a/src/3rdparty/libwebp/src/dsp/dsp.h b/src/3rdparty/libwebp/src/dsp/dsp.h index 6df48cf..35085e0 100644 --- a/src/3rdparty/libwebp/src/dsp/dsp.h +++ b/src/3rdparty/libwebp/src/dsp/dsp.h @@ -27,6 +27,23 @@ extern "C" { #define BPS 32 // this is the common stride for enc/dec //------------------------------------------------------------------------------ +// WEBP_RESTRICT + +// Declares a pointer with the restrict type qualifier if available. +// This allows code to hint to the compiler that only this pointer references a +// particular object or memory region within the scope of the block in which it +// is declared. This may allow for improved optimizations due to the lack of +// pointer aliasing. See also: +// https://en.cppreference.com/w/c/language/restrict +#if defined(__GNUC__) +#define WEBP_RESTRICT __restrict__ +#elif defined(_MSC_VER) +#define WEBP_RESTRICT __restrict +#else +#define WEBP_RESTRICT +#endif + +//------------------------------------------------------------------------------ // CPU detection #if defined(__GNUC__) @@ -67,21 +84,31 @@ extern "C" { // 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) +#if (defined(__SSE2__) || defined(WEBP_MSC_SSE2)) && \ + (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE2)) #define WEBP_USE_SSE2 #endif -#if defined(__SSE4_1__) || defined(WEBP_MSC_SSE41) || defined(WEBP_HAVE_SSE41) +#if defined(WEBP_USE_SSE2) && !defined(WEBP_HAVE_SSE2) +#define WEBP_HAVE_SSE2 +#endif + +#if (defined(__SSE4_1__) || defined(WEBP_MSC_SSE41)) && \ + (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE41)) #define WEBP_USE_SSE41 #endif +#if defined(WEBP_USE_SSE41) && !defined(WEBP_HAVE_SSE41) +#define WEBP_HAVE_SSE41 +#endif + #undef WEBP_MSC_SSE41 #undef WEBP_MSC_SSE2 // 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(__aarch64__) || defined(WEBP_HAVE_NEON)) && \ +#if ((defined(__ARM_NEON__) || defined(__aarch64__)) && \ + (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_NEON))) && \ !defined(__native_client__) #define WEBP_USE_NEON #endif @@ -97,6 +124,10 @@ extern "C" { #define WEBP_USE_INTRINSICS #endif +#if defined(WEBP_USE_NEON) && !defined(WEBP_HAVE_NEON) +#define WEBP_HAVE_NEON +#endif + #if defined(__mips__) && !defined(__mips64) && \ defined(__mips_isa_rev) && (__mips_isa_rev >= 1) && (__mips_isa_rev < 6) #define WEBP_USE_MIPS32 @@ -116,7 +147,7 @@ extern "C" { #define WEBP_DSP_OMIT_C_CODE 1 #endif -#if (defined(__aarch64__) || defined(__ARM_NEON__)) && WEBP_DSP_OMIT_C_CODE +#if defined(WEBP_USE_NEON) && WEBP_DSP_OMIT_C_CODE #define WEBP_NEON_OMIT_C_CODE 1 #else #define WEBP_NEON_OMIT_C_CODE 0 @@ -578,26 +609,29 @@ extern void (*WebPApplyAlphaMultiply4444)( // 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); +extern int (*WebPDispatchAlpha)(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint8_t* WEBP_RESTRICT 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); +extern void (*WebPDispatchAlphaToGreen)(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint32_t* WEBP_RESTRICT 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); +extern int (*WebPExtractAlpha)(const uint8_t* WEBP_RESTRICT argb, + int argb_stride, int width, int height, + uint8_t* WEBP_RESTRICT alpha, + int alpha_stride); // Extract the green values from 32b values in argb[] and pack them into alpha[] // (this is the opposite of WebPDispatchAlphaToGreen). -extern void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size); +extern void (*WebPExtractGreen)(const uint32_t* WEBP_RESTRICT argb, + uint8_t* WEBP_RESTRICT alpha, int size); // 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. @@ -610,29 +644,35 @@ 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, +extern void (*WebPMultRow)(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT 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, +void WebPMultRows(uint8_t* WEBP_RESTRICT ptr, int stride, + const uint8_t* WEBP_RESTRICT alpha, int alpha_stride, int width, int num_rows, int inverse); // Plain-C versions, used as fallback by some implementations. -void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha, +void WebPMultRow_C(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse); void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse); #ifdef WORDS_BIGENDIAN // ARGB packing function: a/r/g/b input is rgba or bgra order. -extern void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, - const uint8_t* g, const uint8_t* b, int len, - uint32_t* out); +extern void (*WebPPackARGB)(const uint8_t* WEBP_RESTRICT a, + const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, uint32_t* WEBP_RESTRICT out); #endif // RGB packing function. 'step' can be 3 or 4. r/g/b input is rgb or bgr order. -extern void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b, - int len, int step, uint32_t* out); +extern void (*WebPPackRGB)(const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, int step, uint32_t* WEBP_RESTRICT out); // This function returns true if src[i] contains a value different from 0xff. extern int (*WebPHasAlpha8b)(const uint8_t* src, int length); diff --git a/src/3rdparty/libwebp/src/dsp/enc.c b/src/3rdparty/libwebp/src/dsp/enc.c index 2fddbc4..ea47a3f 100644 --- a/src/3rdparty/libwebp/src/dsp/enc.c +++ b/src/3rdparty/libwebp/src/dsp/enc.c @@ -773,10 +773,10 @@ WEBP_DSP_INIT_FUNC(VP8EncDspInit) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8EncDspInitSSE2(); -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { VP8EncDspInitSSE41(); } @@ -800,7 +800,7 @@ WEBP_DSP_INIT_FUNC(VP8EncDspInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { VP8EncDspInitNEON(); diff --git a/src/3rdparty/libwebp/src/dsp/filters.c b/src/3rdparty/libwebp/src/dsp/filters.c index 9e910d9..4506567 100644 --- a/src/3rdparty/libwebp/src/dsp/filters.c +++ b/src/3rdparty/libwebp/src/dsp/filters.c @@ -254,7 +254,7 @@ WEBP_DSP_INIT_FUNC(VP8FiltersInit) { #endif if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8FiltersInitSSE2(); } @@ -271,7 +271,7 @@ WEBP_DSP_INIT_FUNC(VP8FiltersInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { VP8FiltersInitNEON(); diff --git a/src/3rdparty/libwebp/src/dsp/filters_sse2.c b/src/3rdparty/libwebp/src/dsp/filters_sse2.c index 4b3f2d0..5c33ec1 100644 --- a/src/3rdparty/libwebp/src/dsp/filters_sse2.c +++ b/src/3rdparty/libwebp/src/dsp/filters_sse2.c @@ -320,7 +320,12 @@ extern void VP8FiltersInitSSE2(void); WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitSSE2(void) { WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_SSE2; +#if defined(CHROMIUM) + // TODO(crbug.com/654974) + (void)VerticalUnfilter_SSE2; +#else WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_SSE2; +#endif WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_SSE2; WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_SSE2; diff --git a/src/3rdparty/libwebp/src/dsp/lossless.c b/src/3rdparty/libwebp/src/dsp/lossless.c index 46b220e..d8bbb02 100644 --- a/src/3rdparty/libwebp/src/dsp/lossless.c +++ b/src/3rdparty/libwebp/src/dsp/lossless.c @@ -575,6 +575,7 @@ VP8LMapARGBFunc VP8LMapColor32b; VP8LMapAlphaFunc VP8LMapColor8b; extern void VP8LDspInitSSE2(void); +extern void VP8LDspInitSSE41(void); extern void VP8LDspInitNEON(void); extern void VP8LDspInitMIPSdspR2(void); extern void VP8LDspInitMSA(void); @@ -621,9 +622,14 @@ WEBP_DSP_INIT_FUNC(VP8LDspInit) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8LDspInitSSE2(); +#if defined(WEBP_HAVE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + VP8LDspInitSSE41(); + } +#endif } #endif #if defined(WEBP_USE_MIPS_DSP_R2) @@ -638,7 +644,7 @@ WEBP_DSP_INIT_FUNC(VP8LDspInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { VP8LDspInitNEON(); diff --git a/src/3rdparty/libwebp/src/dsp/lossless_enc.c b/src/3rdparty/libwebp/src/dsp/lossless_enc.c index a0c7ab9..c3e8537 100644 --- a/src/3rdparty/libwebp/src/dsp/lossless_enc.c +++ b/src/3rdparty/libwebp/src/dsp/lossless_enc.c @@ -329,6 +329,15 @@ const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX] = { static float FastSLog2Slow_C(uint32_t v) { assert(v >= LOG_LOOKUP_IDX_MAX); if (v < APPROX_LOG_WITH_CORRECTION_MAX) { +#if !defined(WEBP_HAVE_SLOW_CLZ_CTZ) + // use clz if available + const int log_cnt = BitsLog2Floor(v) - 7; + const uint32_t y = 1 << log_cnt; + int correction = 0; + const float v_f = (float)v; + const uint32_t orig_v = v; + v >>= log_cnt; +#else int log_cnt = 0; uint32_t y = 1; int correction = 0; @@ -339,6 +348,7 @@ static float FastSLog2Slow_C(uint32_t v) { v = v >> 1; y = y << 1; } while (v >= LOG_LOOKUP_IDX_MAX); +#endif // vf = (2^log_cnt) * Xf; where y = 2^log_cnt and Xf < 256 // Xf = floor(Xf) * (1 + (v % y) / v) // log2(Xf) = log2(floor(Xf)) + log2(1 + (v % y) / v) @@ -355,6 +365,14 @@ static float FastSLog2Slow_C(uint32_t v) { static float FastLog2Slow_C(uint32_t v) { assert(v >= LOG_LOOKUP_IDX_MAX); if (v < APPROX_LOG_WITH_CORRECTION_MAX) { +#if !defined(WEBP_HAVE_SLOW_CLZ_CTZ) + // use clz if available + const int log_cnt = BitsLog2Floor(v) - 7; + const uint32_t y = 1 << log_cnt; + const uint32_t orig_v = v; + double log_2; + v >>= log_cnt; +#else int log_cnt = 0; uint32_t y = 1; const uint32_t orig_v = v; @@ -364,6 +382,7 @@ static float FastLog2Slow_C(uint32_t v) { v = v >> 1; y = y << 1; } while (v >= LOG_LOOKUP_IDX_MAX); +#endif log_2 = kLog2Table[v] + log_cnt; if (orig_v >= APPROX_LOG_MAX) { // Since the division is still expensive, add this correction factor only @@ -843,10 +862,10 @@ WEBP_DSP_INIT_FUNC(VP8LEncDspInit) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8LEncDspInitSSE2(); -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { VP8LEncDspInitSSE41(); } @@ -870,7 +889,7 @@ WEBP_DSP_INIT_FUNC(VP8LEncDspInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { VP8LEncDspInitNEON(); diff --git a/src/3rdparty/libwebp/src/dsp/lossless_enc_sse2.c b/src/3rdparty/libwebp/src/dsp/lossless_enc_sse2.c index 90c2637..b2f83b8 100644 --- a/src/3rdparty/libwebp/src/dsp/lossless_enc_sse2.c +++ b/src/3rdparty/libwebp/src/dsp/lossless_enc_sse2.c @@ -232,79 +232,55 @@ static void AddVectorEq_SSE2(const uint32_t* a, uint32_t* out, int size) { //------------------------------------------------------------------------------ // Entropy -// Checks whether the X or Y contribution is worth computing and adding. -// Used in loop unrolling. -#define ANALYZE_X_OR_Y(x_or_y, j) \ - do { \ - if ((x_or_y)[i + (j)] != 0) retval -= VP8LFastSLog2((x_or_y)[i + (j)]); \ - } while (0) - -// Checks whether the X + Y contribution is worth computing and adding. -// Used in loop unrolling. -#define ANALYZE_XY(j) \ - do { \ - if (tmp[j] != 0) { \ - retval -= VP8LFastSLog2(tmp[j]); \ - ANALYZE_X_OR_Y(X, j); \ - } \ - } while (0) - -#if !(defined(__i386__) || defined(_M_IX86)) +// TODO(https://crbug.com/webp/499): this function produces different results +// from the C code due to use of double/float resulting in output differences +// when compared to -noasm. +#if !(defined(WEBP_HAVE_SLOW_CLZ_CTZ) || defined(__i386__) || defined(_M_IX86)) + static float CombinedShannonEntropy_SSE2(const int X[256], const int Y[256]) { int i; double retval = 0.; - int sumX, sumXY; - int32_t tmp[4]; - __m128i zero = _mm_setzero_si128(); - // Sums up X + Y, 4 ints at a time (and will merge it at the end for sumXY). - __m128i sumXY_128 = zero; - __m128i sumX_128 = zero; - - for (i = 0; i < 256; i += 4) { - const __m128i x = _mm_loadu_si128((const __m128i*)(X + i)); - const __m128i y = _mm_loadu_si128((const __m128i*)(Y + i)); - - // Check if any X is non-zero: this actually provides a speedup as X is - // usually sparse. - if (_mm_movemask_epi8(_mm_cmpeq_epi32(x, zero)) != 0xFFFF) { - const __m128i xy_128 = _mm_add_epi32(x, y); - sumXY_128 = _mm_add_epi32(sumXY_128, xy_128); - - sumX_128 = _mm_add_epi32(sumX_128, x); - - // Analyze the different X + Y. - _mm_storeu_si128((__m128i*)tmp, xy_128); - - ANALYZE_XY(0); - ANALYZE_XY(1); - ANALYZE_XY(2); - ANALYZE_XY(3); - } else { - // X is fully 0, so only deal with Y. - sumXY_128 = _mm_add_epi32(sumXY_128, y); - - ANALYZE_X_OR_Y(Y, 0); - ANALYZE_X_OR_Y(Y, 1); - ANALYZE_X_OR_Y(Y, 2); - ANALYZE_X_OR_Y(Y, 3); + int sumX = 0, sumXY = 0; + const __m128i zero = _mm_setzero_si128(); + + for (i = 0; i < 256; i += 16) { + const __m128i x0 = _mm_loadu_si128((const __m128i*)(X + i + 0)); + const __m128i y0 = _mm_loadu_si128((const __m128i*)(Y + i + 0)); + const __m128i x1 = _mm_loadu_si128((const __m128i*)(X + i + 4)); + const __m128i y1 = _mm_loadu_si128((const __m128i*)(Y + i + 4)); + const __m128i x2 = _mm_loadu_si128((const __m128i*)(X + i + 8)); + const __m128i y2 = _mm_loadu_si128((const __m128i*)(Y + i + 8)); + const __m128i x3 = _mm_loadu_si128((const __m128i*)(X + i + 12)); + const __m128i y3 = _mm_loadu_si128((const __m128i*)(Y + i + 12)); + const __m128i x4 = _mm_packs_epi16(_mm_packs_epi32(x0, x1), + _mm_packs_epi32(x2, x3)); + const __m128i y4 = _mm_packs_epi16(_mm_packs_epi32(y0, y1), + _mm_packs_epi32(y2, y3)); + const int32_t mx = _mm_movemask_epi8(_mm_cmpgt_epi8(x4, zero)); + int32_t my = _mm_movemask_epi8(_mm_cmpgt_epi8(y4, zero)) | mx; + while (my) { + const int32_t j = BitsCtz(my); + int xy; + if ((mx >> j) & 1) { + const int x = X[i + j]; + sumXY += x; + retval -= VP8LFastSLog2(x); + } + xy = X[i + j] + Y[i + j]; + sumX += xy; + retval -= VP8LFastSLog2(xy); + my &= my - 1; } } - - // Sum up sumX_128 to get sumX. - _mm_storeu_si128((__m128i*)tmp, sumX_128); - sumX = tmp[3] + tmp[2] + tmp[1] + tmp[0]; - - // Sum up sumXY_128 to get sumXY. - _mm_storeu_si128((__m128i*)tmp, sumXY_128); - sumXY = tmp[3] + tmp[2] + tmp[1] + tmp[0]; - retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY); return (float)retval; } -#endif // !(defined(__i386__) || defined(_M_IX86)) -#undef ANALYZE_X_OR_Y -#undef ANALYZE_XY +#else + +#define DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC // won't be faster + +#endif //------------------------------------------------------------------------------ @@ -662,10 +638,7 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) { VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE2; VP8LAddVector = AddVector_SSE2; VP8LAddVectorEq = AddVectorEq_SSE2; - // TODO(https://crbug.com/webp/499): this function produces different results - // from the C code due to use of double/float resulting in output differences - // when compared to -noasm. -#if !(defined(__i386__) || defined(_M_IX86)) +#if !defined(DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC) VP8LCombinedShannonEntropy = CombinedShannonEntropy_SSE2; #endif VP8LVectorMismatch = VectorMismatch_SSE2; diff --git a/src/3rdparty/libwebp/src/dsp/lossless_enc_sse41.c b/src/3rdparty/libwebp/src/dsp/lossless_enc_sse41.c index 719d8ed..ad358a6 100644 --- a/src/3rdparty/libwebp/src/dsp/lossless_enc_sse41.c +++ b/src/3rdparty/libwebp/src/dsp/lossless_enc_sse41.c @@ -44,46 +44,47 @@ static void SubtractGreenFromBlueAndRed_SSE41(uint32_t* argb_data, //------------------------------------------------------------------------------ // Color Transform -#define SPAN 8 +#define MK_CST_16(HI, LO) \ + _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff))) + static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride, int tile_width, int tile_height, int green_to_blue, int red_to_blue, int histo[]) { - const __m128i mults_r = _mm_set1_epi16(CST_5b(red_to_blue)); - const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_blue)); - const __m128i mask_g = _mm_set1_epi16((short)0xff00); // green mask - const __m128i mask_gb = _mm_set1_epi32(0xffff); // green/blue mask - const __m128i mask_b = _mm_set1_epi16(0x00ff); // blue mask - const __m128i shuffler_lo = _mm_setr_epi8(-1, 2, -1, 6, -1, 10, -1, 14, -1, - -1, -1, -1, -1, -1, -1, -1); - const __m128i shuffler_hi = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, - 2, -1, 6, -1, 10, -1, 14); - int y; - for (y = 0; y < tile_height; ++y) { - const uint32_t* const src = argb + y * stride; - int i, x; - for (x = 0; x + SPAN <= tile_width; x += SPAN) { - uint16_t values[SPAN]; - const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); - const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); - const __m128i r0 = _mm_shuffle_epi8(in0, shuffler_lo); - const __m128i r1 = _mm_shuffle_epi8(in1, shuffler_hi); - const __m128i r = _mm_or_si128(r0, r1); // r 0 - const __m128i gb0 = _mm_and_si128(in0, mask_gb); - const __m128i gb1 = _mm_and_si128(in1, mask_gb); - const __m128i gb = _mm_packus_epi32(gb0, gb1); // g b - const __m128i g = _mm_and_si128(gb, mask_g); // g 0 - const __m128i A = _mm_mulhi_epi16(r, mults_r); // x dbr - const __m128i B = _mm_mulhi_epi16(g, mults_g); // x dbg - const __m128i C = _mm_sub_epi8(gb, B); // x b' - const __m128i D = _mm_sub_epi8(C, A); // x b'' - const __m128i E = _mm_and_si128(D, mask_b); // 0 b'' - _mm_storeu_si128((__m128i*)values, E); - for (i = 0; i < SPAN; ++i) ++histo[values[i]]; + const __m128i mult = + MK_CST_16(CST_5b(red_to_blue) + 256,CST_5b(green_to_blue)); + const __m128i perm = + _mm_setr_epi8(-1, 1, -1, 2, -1, 5, -1, 6, -1, 9, -1, 10, -1, 13, -1, 14); + if (tile_width >= 4) { + int y; + for (y = 0; y < tile_height; ++y) { + const uint32_t* const src = argb + y * stride; + const __m128i A1 = _mm_loadu_si128((const __m128i*)src); + const __m128i B1 = _mm_shuffle_epi8(A1, perm); + const __m128i C1 = _mm_mulhi_epi16(B1, mult); + const __m128i D1 = _mm_sub_epi16(A1, C1); + __m128i E = _mm_add_epi16(_mm_srli_epi32(D1, 16), D1); + int x; + for (x = 4; x + 4 <= tile_width; x += 4) { + const __m128i A2 = _mm_loadu_si128((const __m128i*)(src + x)); + __m128i B2, C2, D2; + ++histo[_mm_extract_epi8(E, 0)]; + B2 = _mm_shuffle_epi8(A2, perm); + ++histo[_mm_extract_epi8(E, 4)]; + C2 = _mm_mulhi_epi16(B2, mult); + ++histo[_mm_extract_epi8(E, 8)]; + D2 = _mm_sub_epi16(A2, C2); + ++histo[_mm_extract_epi8(E, 12)]; + E = _mm_add_epi16(_mm_srli_epi32(D2, 16), D2); + } + ++histo[_mm_extract_epi8(E, 0)]; + ++histo[_mm_extract_epi8(E, 4)]; + ++histo[_mm_extract_epi8(E, 8)]; + ++histo[_mm_extract_epi8(E, 12)]; } } { - const int left_over = tile_width & (SPAN - 1); + const int left_over = tile_width & 3; if (left_over > 0) { VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride, left_over, tile_height, @@ -95,33 +96,37 @@ static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride, static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride, int tile_width, int tile_height, int green_to_red, int histo[]) { - const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_red)); - const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask - const __m128i mask = _mm_set1_epi16(0xff); - - int y; - for (y = 0; y < tile_height; ++y) { - const uint32_t* const src = argb + y * stride; - int i, x; - for (x = 0; x + SPAN <= tile_width; x += SPAN) { - uint16_t values[SPAN]; - const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); - const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); - const __m128i g0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0 - const __m128i g1 = _mm_and_si128(in1, mask_g); - const __m128i g = _mm_packus_epi32(g0, g1); // g 0 - const __m128i A0 = _mm_srli_epi32(in0, 16); // 0 0 | x r - const __m128i A1 = _mm_srli_epi32(in1, 16); - const __m128i A = _mm_packus_epi32(A0, A1); // x r - const __m128i B = _mm_mulhi_epi16(g, mults_g); // x dr - const __m128i C = _mm_sub_epi8(A, B); // x r' - const __m128i D = _mm_and_si128(C, mask); // 0 r' - _mm_storeu_si128((__m128i*)values, D); - for (i = 0; i < SPAN; ++i) ++histo[values[i]]; + + const __m128i mult = MK_CST_16(0, CST_5b(green_to_red)); + const __m128i mask_g = _mm_set1_epi32(0x0000ff00); + if (tile_width >= 4) { + int y; + for (y = 0; y < tile_height; ++y) { + const uint32_t* const src = argb + y * stride; + const __m128i A1 = _mm_loadu_si128((const __m128i*)src); + const __m128i B1 = _mm_and_si128(A1, mask_g); + const __m128i C1 = _mm_madd_epi16(B1, mult); + __m128i D = _mm_sub_epi16(A1, C1); + int x; + for (x = 4; x + 4 <= tile_width; x += 4) { + const __m128i A2 = _mm_loadu_si128((const __m128i*)(src + x)); + __m128i B2, C2; + ++histo[_mm_extract_epi8(D, 2)]; + B2 = _mm_and_si128(A2, mask_g); + ++histo[_mm_extract_epi8(D, 6)]; + C2 = _mm_madd_epi16(B2, mult); + ++histo[_mm_extract_epi8(D, 10)]; + ++histo[_mm_extract_epi8(D, 14)]; + D = _mm_sub_epi16(A2, C2); + } + ++histo[_mm_extract_epi8(D, 2)]; + ++histo[_mm_extract_epi8(D, 6)]; + ++histo[_mm_extract_epi8(D, 10)]; + ++histo[_mm_extract_epi8(D, 14)]; } } { - const int left_over = tile_width & (SPAN - 1); + const int left_over = tile_width & 3; if (left_over > 0) { VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride, left_over, tile_height, green_to_red, @@ -130,6 +135,8 @@ static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride, } } +#undef MK_CST_16 + //------------------------------------------------------------------------------ // Entry point diff --git a/src/3rdparty/libwebp/src/dsp/lossless_sse2.c b/src/3rdparty/libwebp/src/dsp/lossless_sse2.c index aef0cee..3a0eb44 100644 --- a/src/3rdparty/libwebp/src/dsp/lossless_sse2.c +++ b/src/3rdparty/libwebp/src/dsp/lossless_sse2.c @@ -18,7 +18,6 @@ #include "src/dsp/common_sse2.h" #include "src/dsp/lossless.h" #include "src/dsp/lossless_common.h" -#include <assert.h> #include <emmintrin.h> //------------------------------------------------------------------------------ diff --git a/src/3rdparty/libwebp/src/dsp/lossless_sse41.c b/src/3rdparty/libwebp/src/dsp/lossless_sse41.c new file mode 100644 index 0000000..b0d6daa --- /dev/null +++ b/src/3rdparty/libwebp/src/dsp/lossless_sse41.c @@ -0,0 +1,132 @@ +// Copyright 2021 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE41 variant of methods for lossless decoder + +#include "src/dsp/dsp.h" + +#if defined(WEBP_USE_SSE41) + +#include "src/dsp/common_sse41.h" +#include "src/dsp/lossless.h" +#include "src/dsp/lossless_common.h" + +//------------------------------------------------------------------------------ +// Color-space conversion functions + +static void TransformColorInverse_SSE41(const VP8LMultipliers* const m, + const uint32_t* const src, + int num_pixels, uint32_t* dst) { +// sign-extended multiplying constants, pre-shifted by 5. +#define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend + const __m128i mults_rb = _mm_set1_epi32((uint32_t)CST(green_to_red_) << 16 | + (CST(green_to_blue_) & 0xffff)); + const __m128i mults_b2 = _mm_set1_epi32(CST(red_to_blue_)); +#undef CST + const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); + const __m128i perm1 = _mm_setr_epi8(-1, 1, -1, 1, -1, 5, -1, 5, + -1, 9, -1, 9, -1, 13, -1, 13); + const __m128i perm2 = _mm_setr_epi8(-1, 2, -1, -1, -1, 6, -1, -1, + -1, 10, -1, -1, -1, 14, -1, -1); + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i A = _mm_loadu_si128((const __m128i*)(src + i)); + const __m128i B = _mm_shuffle_epi8(A, perm1); // argb -> g0g0 + const __m128i C = _mm_mulhi_epi16(B, mults_rb); + const __m128i D = _mm_add_epi8(A, C); + const __m128i E = _mm_shuffle_epi8(D, perm2); + const __m128i F = _mm_mulhi_epi16(E, mults_b2); + const __m128i G = _mm_add_epi8(D, F); + const __m128i out = _mm_blendv_epi8(G, A, mask_ag); + _mm_storeu_si128((__m128i*)&dst[i], out); + } + // Fall-back to C-version for left-overs. + if (i != num_pixels) { + VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i); + } +} + +//------------------------------------------------------------------------------ + +#define ARGB_TO_RGB_SSE41 do { \ + while (num_pixels >= 16) { \ + const __m128i in0 = _mm_loadu_si128(in + 0); \ + const __m128i in1 = _mm_loadu_si128(in + 1); \ + const __m128i in2 = _mm_loadu_si128(in + 2); \ + const __m128i in3 = _mm_loadu_si128(in + 3); \ + const __m128i a0 = _mm_shuffle_epi8(in0, perm0); \ + const __m128i a1 = _mm_shuffle_epi8(in1, perm1); \ + const __m128i a2 = _mm_shuffle_epi8(in2, perm2); \ + const __m128i a3 = _mm_shuffle_epi8(in3, perm3); \ + const __m128i b0 = _mm_blend_epi16(a0, a1, 0xc0); \ + const __m128i b1 = _mm_blend_epi16(a1, a2, 0xf0); \ + const __m128i b2 = _mm_blend_epi16(a2, a3, 0xfc); \ + _mm_storeu_si128(out + 0, b0); \ + _mm_storeu_si128(out + 1, b1); \ + _mm_storeu_si128(out + 2, b2); \ + in += 4; \ + out += 3; \ + num_pixels -= 16; \ + } \ +} while (0) + +static void ConvertBGRAToRGB_SSE41(const uint32_t* src, int num_pixels, + uint8_t* dst) { + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + const __m128i perm0 = _mm_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9, + 8, 14, 13, 12, -1, -1, -1, -1); + const __m128i perm1 = _mm_shuffle_epi32(perm0, 0x39); + const __m128i perm2 = _mm_shuffle_epi32(perm0, 0x4e); + const __m128i perm3 = _mm_shuffle_epi32(perm0, 0x93); + + ARGB_TO_RGB_SSE41; + + // left-overs + if (num_pixels > 0) { + VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out); + } +} + +static void ConvertBGRAToBGR_SSE41(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + const __m128i perm0 = _mm_setr_epi8(0, 1, 2, 4, 5, 6, 8, 9, 10, + 12, 13, 14, -1, -1, -1, -1); + const __m128i perm1 = _mm_shuffle_epi32(perm0, 0x39); + const __m128i perm2 = _mm_shuffle_epi32(perm0, 0x4e); + const __m128i perm3 = _mm_shuffle_epi32(perm0, 0x93); + + ARGB_TO_RGB_SSE41; + + // left-overs + if (num_pixels > 0) { + VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, (uint8_t*)out); + } +} + +#undef ARGB_TO_RGB_SSE41 + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LDspInitSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE41(void) { + VP8LTransformColorInverse = TransformColorInverse_SSE41; + VP8LConvertBGRAToRGB = ConvertBGRAToRGB_SSE41; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR_SSE41; +} + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(VP8LDspInitSSE41) + +#endif // WEBP_USE_SSE41 diff --git a/src/3rdparty/libwebp/src/dsp/rescaler.c b/src/3rdparty/libwebp/src/dsp/rescaler.c index c5a01e8..14620ce 100644 --- a/src/3rdparty/libwebp/src/dsp/rescaler.c +++ b/src/3rdparty/libwebp/src/dsp/rescaler.c @@ -38,8 +38,9 @@ void WebPRescalerImportRowExpand_C(WebPRescaler* const wrk, int x_out = channel; // simple bilinear interpolation int accum = wrk->x_add; - int left = src[x_in]; - int right = (wrk->src_width > 1) ? src[x_in + x_stride] : left; + rescaler_t left = (rescaler_t)src[x_in]; + rescaler_t right = + (wrk->src_width > 1) ? (rescaler_t)src[x_in + x_stride] : left; x_in += x_stride; while (1) { wrk->frow[x_out] = right * wrk->x_add + (left - right) * accum; @@ -50,7 +51,7 @@ void WebPRescalerImportRowExpand_C(WebPRescaler* const wrk, left = right; x_in += x_stride; assert(x_in < wrk->src_width * x_stride); - right = src[x_in]; + right = (rescaler_t)src[x_in]; accum += wrk->x_add; } } @@ -213,7 +214,7 @@ WEBP_DSP_INIT_FUNC(WebPRescalerDspInit) { WebPRescalerImportRowShrink = WebPRescalerImportRowShrink_C; if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPRescalerDspInitSSE2(); } @@ -235,7 +236,7 @@ WEBP_DSP_INIT_FUNC(WebPRescalerDspInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { WebPRescalerDspInitNEON(); diff --git a/src/3rdparty/libwebp/src/dsp/ssim.c b/src/3rdparty/libwebp/src/dsp/ssim.c index 989ce82..f85c2e6 100644 --- a/src/3rdparty/libwebp/src/dsp/ssim.c +++ b/src/3rdparty/libwebp/src/dsp/ssim.c @@ -150,7 +150,7 @@ WEBP_DSP_INIT_FUNC(VP8SSIMDspInit) { #endif if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8SSIMDspInitSSE2(); } diff --git a/src/3rdparty/libwebp/src/dsp/upsampling.c b/src/3rdparty/libwebp/src/dsp/upsampling.c index 9b60da5..87f771f 100644 --- a/src/3rdparty/libwebp/src/dsp/upsampling.c +++ b/src/3rdparty/libwebp/src/dsp/upsampling.c @@ -233,12 +233,12 @@ WEBP_DSP_INIT_FUNC(WebPInitYUV444Converters) { WebPYUV444Converters[MODE_rgbA_4444] = WebPYuv444ToRgba4444_C; if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitYUV444ConvertersSSE2(); } #endif -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitYUV444ConvertersSSE41(); } @@ -278,12 +278,12 @@ WEBP_DSP_INIT_FUNC(WebPInitUpsamplers) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitUpsamplersSSE2(); } #endif -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitUpsamplersSSE41(); } @@ -300,7 +300,7 @@ WEBP_DSP_INIT_FUNC(WebPInitUpsamplers) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { WebPInitUpsamplersNEON(); diff --git a/src/3rdparty/libwebp/src/dsp/yuv.c b/src/3rdparty/libwebp/src/dsp/yuv.c index 14e67fc..48466f8 100644 --- a/src/3rdparty/libwebp/src/dsp/yuv.c +++ b/src/3rdparty/libwebp/src/dsp/yuv.c @@ -90,16 +90,16 @@ WEBP_DSP_INIT_FUNC(WebPInitSamplers) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitSamplersSSE2(); } -#endif // WEBP_USE_SSE2 -#if defined(WEBP_USE_SSE41) +#endif // WEBP_HAVE_SSE2 +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitSamplersSSE41(); } -#endif // WEBP_USE_SSE41 +#endif // WEBP_HAVE_SSE41 #if defined(WEBP_USE_MIPS32) if (VP8GetCPUInfo(kMIPS32)) { WebPInitSamplersMIPS32(); @@ -276,26 +276,26 @@ WEBP_DSP_INIT_FUNC(WebPInitConvertARGBToYUV) { #endif if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitConvertARGBToYUVSSE2(); WebPInitSharpYUVSSE2(); } -#endif // WEBP_USE_SSE2 -#if defined(WEBP_USE_SSE41) +#endif // WEBP_HAVE_SSE2 +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitConvertARGBToYUVSSE41(); } -#endif // WEBP_USE_SSE41 +#endif // WEBP_HAVE_SSE41 } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { WebPInitConvertARGBToYUVNEON(); WebPInitSharpYUVNEON(); } -#endif // WEBP_USE_NEON +#endif // WEBP_HAVE_NEON assert(WebPConvertARGBToY != NULL); assert(WebPConvertARGBToUV != NULL); |