From 0f4f721b129f438731416c4b63445a324fd49864 Mon Sep 17 00:00:00 2001 From: skal Date: Thu, 27 Mar 2014 21:43:21 +0100 Subject: [PATCH] separate SSE2 lossless functions into its own file expose the predictor array as function pointers instead of each individual sub-function + merged Average2() into ClampedAddSubtractHalf directly + unified the signature as "VP8LProcessBlueAndRedFunc" no speed diff observed Change-Id: Ic3c45dff11884a8330a9ad38c2c8e82491c6e044 --- Android.mk | 1 + Makefile.vc | 1 + makefile.unix | 1 + src/dsp/Makefile.am | 1 + src/dsp/lossless.c | 323 +++++----------------------------------- src/dsp/lossless.h | 36 +++-- src/dsp/lossless_sse2.c | 306 +++++++++++++++++++++++++++++++++++++ 7 files changed, 373 insertions(+), 296 deletions(-) create mode 100644 src/dsp/lossless_sse2.c diff --git a/Android.mk b/Android.mk index 8661fd76..a626b4b7 100644 --- a/Android.mk +++ b/Android.mk @@ -32,6 +32,7 @@ LOCAL_SRC_FILES := \ src/dsp/enc.c \ src/dsp/enc_sse2.c \ src/dsp/lossless.c \ + src/dsp/lossless_sse2.c \ src/dsp/upsampling.c \ src/dsp/upsampling_mips32.c \ src/dsp/upsampling_sse2.c \ diff --git a/Makefile.vc b/Makefile.vc index c790ba70..d18fa32d 100644 --- a/Makefile.vc +++ b/Makefile.vc @@ -174,6 +174,7 @@ DSP_DEC_OBJS = \ $(DIROBJ)\dsp\dec_neon.obj \ $(DIROBJ)\dsp\dec_sse2.obj \ $(DIROBJ)\dsp\lossless.obj \ + $(DIROBJ)\dsp\lossless_sse2.obj \ $(DIROBJ)\dsp\upsampling.obj \ $(DIROBJ)\dsp\upsampling_mips32.obj \ $(DIROBJ)\dsp\upsampling_neon.obj \ diff --git a/makefile.unix b/makefile.unix index e8bdbef5..630b0a43 100644 --- a/makefile.unix +++ b/makefile.unix @@ -105,6 +105,7 @@ DSP_DEC_OBJS = \ src/dsp/dec_neon.o \ src/dsp/dec_sse2.o \ src/dsp/lossless.o \ + src/dsp/lossless_sse2.o \ src/dsp/upsampling.o \ src/dsp/upsampling_mips32.o \ src/dsp/upsampling_neon.o \ diff --git a/src/dsp/Makefile.am b/src/dsp/Makefile.am index 028902b5..7b883dcd 100644 --- a/src/dsp/Makefile.am +++ b/src/dsp/Makefile.am @@ -17,6 +17,7 @@ COMMON_SOURCES += dec_neon.c COMMON_SOURCES += dec_sse2.c COMMON_SOURCES += dsp.h COMMON_SOURCES += lossless.c +COMMON_SOURCES += lossless_sse2.c COMMON_SOURCES += lossless.h COMMON_SOURCES += upsampling.c COMMON_SOURCES += upsampling_mips32.c diff --git a/src/dsp/lossless.c b/src/dsp/lossless.c index cdc2ac30..b5d01ece 100644 --- a/src/dsp/lossless.c +++ b/src/dsp/lossless.c @@ -15,10 +15,6 @@ #include "./dsp.h" -#if defined(WEBP_USE_SSE2) -#include -#endif - #include #include #include "./lossless.h" @@ -519,21 +515,19 @@ static uint32_t Predictor10(uint32_t left, const uint32_t* const top) { return pred; } static uint32_t Predictor11(uint32_t left, const uint32_t* const top) { - const uint32_t pred = VP8LSelect(top[0], left, top[-1]); + const uint32_t pred = Select(top[0], left, top[-1]); return pred; } static uint32_t Predictor12(uint32_t left, const uint32_t* const top) { - const uint32_t pred = VP8LClampedAddSubtractFull(left, top[0], top[-1]); + const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]); return pred; } static uint32_t Predictor13(uint32_t left, const uint32_t* const top) { - const uint32_t pred = VP8LClampedAddSubtractHalf(left, top[0], top[-1]); + const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]); return pred; } -// TODO(vikasa): Export the predictor array, to allow SSE2 variants. -typedef uint32_t (*PredictorFunc)(uint32_t left, const uint32_t* const top); -static const PredictorFunc kPredictors[16] = { +static const VP8LPredictorFunc kPredictorsC[16] = { Predictor0, Predictor1, Predictor2, Predictor3, Predictor4, Predictor5, Predictor6, Predictor7, Predictor8, Predictor9, Predictor10, Predictor11, @@ -610,7 +604,7 @@ static int GetBestPredictorForTile(int width, int height, int mode; for (mode = 0; mode < kNumPredModes; ++mode) { const uint32_t* current_row = argb_scratch; - const PredictorFunc pred_func = kPredictors[mode]; + const VP8LPredictorFunc pred_func = VP8LPredictors[mode]; float cur_diff; int y; int histo_argb[4][256]; @@ -653,7 +647,7 @@ static void CopyTileWithPrediction(int width, int height, const int tile_size = 1 << bits; const int max_y = GetMin(tile_size, height - row_start); const int max_x = GetMin(tile_size, width - col_start); - const PredictorFunc pred_func = kPredictors[mode]; + const VP8LPredictorFunc pred_func = VP8LPredictors[mode]; const uint32_t* current_row = argb_scratch; int y; @@ -758,14 +752,14 @@ static void PredictorInverseTransform(const VP8LTransform* const transform, while (y < y_end) { 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: while (x < safe_width) { - pred_func = kPredictors[((*pred_mode_src++) >> 8) & 0xf]; + 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); @@ -773,7 +767,7 @@ static void PredictorInverseTransform(const VP8LTransform* const transform, t = 0; } if (x < width) { - pred_func = kPredictors[((*pred_mode_src++) >> 8) & 0xf]; + 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); @@ -788,9 +782,9 @@ static void PredictorInverseTransform(const VP8LTransform* const transform, } } -static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs) { - int i = 0; - for (; i < num_pixs; ++i) { +void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels) { + int i; + for (i = 0; i < num_pixels; ++i) { const uint32_t argb = argb_data[i]; const uint32_t green = (argb >> 8) & 0xff; const uint32_t new_r = (((argb >> 16) & 0xff) - green) & 0xff; @@ -801,14 +795,15 @@ static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs) { // Add green to blue and red channels (i.e. perform the inverse transform of // 'subtract green'). -static void AddGreenToBlueAndRed(uint32_t* data, const uint32_t* data_end) { - while (data < data_end) { - const uint32_t argb = *data; +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; } } @@ -1248,7 +1243,7 @@ void VP8LInverseTransform(const VP8LTransform* const transform, assert(row_end <= transform->ysize_); switch (transform->type_) { case SUBTRACT_GREEN: - VP8LAddGreenToBlueAndRed(out, out + (row_end - row_start) * width); + VP8LAddGreenToBlueAndRed(out, (row_end - row_start) * width); break; case PREDICTOR_TRANSFORM: PredictorInverseTransform(transform, row_start, row_end, out); @@ -1293,8 +1288,8 @@ 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++; @@ -1304,8 +1299,8 @@ 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++; @@ -1316,8 +1311,8 @@ 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++; @@ -1333,8 +1328,8 @@ static void ConvertBGRAToRGBA4444(const uint32_t* src, } } -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++; @@ -1350,8 +1345,8 @@ static void ConvertBGRAToRGB565(const uint32_t* src, } } -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++; @@ -1469,246 +1464,9 @@ void VP8LBundleColorMap(const uint8_t* const row, int width, //------------------------------------------------------------------------------ -// TODO(vikasa): Move the SSE2 functions to lossless_dsp.c (new file), once -// color-space conversion methods (ConvertFromBGRA) are also updated for SSE2. -#if defined(WEBP_USE_SSE2) -static WEBP_INLINE uint32_t ClampedAddSubtractFullSSE2(uint32_t c0, uint32_t c1, - uint32_t c2) { - const __m128i zero = _mm_setzero_si128(); - const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); - const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); - const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); - const __m128i V1 = _mm_add_epi16(C0, C1); - const __m128i V2 = _mm_sub_epi16(V1, C2); - const __m128i b = _mm_packus_epi16(V2, V2); - const uint32_t output = _mm_cvtsi128_si32(b); - return output; -} - -static WEBP_INLINE uint32_t ClampedAddSubtractHalfSSE2(uint32_t c0, uint32_t c1, - uint32_t c2) { - const uint32_t ave = Average2(c0, c1); - const __m128i zero = _mm_setzero_si128(); - const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ave), zero); - const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); - const __m128i A1 = _mm_sub_epi16(A0, B0); - const __m128i BgtA = _mm_cmpgt_epi16(B0, A0); - const __m128i A2 = _mm_sub_epi16(A1, BgtA); - const __m128i A3 = _mm_srai_epi16(A2, 1); - const __m128i A4 = _mm_add_epi16(A0, A3); - const __m128i A5 = _mm_packus_epi16(A4, A4); - const uint32_t output = _mm_cvtsi128_si32(A5); - return output; -} - -static WEBP_INLINE uint32_t SelectSSE2(uint32_t a, uint32_t b, uint32_t c) { - int pa_minus_pb; - const __m128i zero = _mm_setzero_si128(); - const __m128i A0 = _mm_cvtsi32_si128(a); - const __m128i B0 = _mm_cvtsi32_si128(b); - const __m128i C0 = _mm_cvtsi32_si128(c); - const __m128i AC0 = _mm_subs_epu8(A0, C0); - const __m128i CA0 = _mm_subs_epu8(C0, A0); - const __m128i BC0 = _mm_subs_epu8(B0, C0); - const __m128i CB0 = _mm_subs_epu8(C0, B0); - const __m128i AC = _mm_or_si128(AC0, CA0); - const __m128i BC = _mm_or_si128(BC0, CB0); - const __m128i pa = _mm_unpacklo_epi8(AC, zero); // |a - c| - const __m128i pb = _mm_unpacklo_epi8(BC, zero); // |b - c| - const __m128i diff = _mm_sub_epi16(pb, pa); - { - int16_t out[8]; - _mm_storeu_si128((__m128i*)out, diff); - pa_minus_pb = out[0] + out[1] + out[2] + out[3]; - } - return (pa_minus_pb <= 0) ? a : b; -} - -static void SubtractGreenFromBlueAndRedSSE2(uint32_t* argb_data, int num_pixs) { - int i = 0; - const __m128i mask = _mm_set1_epi32(0x0000ff00); - for (; i + 4 < num_pixs; i += 4) { - const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); - const __m128i in_00g0 = _mm_and_si128(in, mask); // 00g0|00g0|... - const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8); // 0g00|0g00|... - const __m128i in_000g = _mm_srli_epi32(in_00g0, 8); // 000g|000g|... - const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g); - const __m128i out = _mm_sub_epi8(in, in_0g0g); - _mm_storeu_si128((__m128i*)&argb_data[i], out); - } - // fallthrough and finish off with plain-C - SubtractGreenFromBlueAndRed(argb_data + i, num_pixs - i); -} - -static void AddGreenToBlueAndRedSSE2(uint32_t* data, const uint32_t* data_end) { - const __m128i mask = _mm_set1_epi32(0x0000ff00); - for (; data + 4 < data_end; data += 4) { - const __m128i in = _mm_loadu_si128((__m128i*)data); - const __m128i in_00g0 = _mm_and_si128(in, mask); // 00g0|00g0|... - const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8); // 0g00|0g00|... - const __m128i in_000g = _mm_srli_epi32(in_00g0, 8); // 000g|000g|... - const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g); - const __m128i out = _mm_add_epi8(in, in_0g0g); - _mm_storeu_si128((__m128i*)data, out); - } - // fallthrough and finish off with plain-C - AddGreenToBlueAndRed(data, data_end); -} - -static void ConvertBGRAToRGBASSE2(const uint32_t* src, - int num_pixels, uint8_t* dst) { - const __m128i* in = (const __m128i*)src; - __m128i* out = (__m128i*)dst; - while (num_pixels >= 8) { - const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 - const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 - const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... - const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... - const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... - const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... - const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 - const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 - const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 - const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 - const __m128i rg0 = _mm_unpacklo_epi8(rb0, ga0); // r0g0r1g1 ... r6g6r7g7 - const __m128i ba0 = _mm_unpackhi_epi8(rb0, ga0); // b0a0b1a1 ... b6a6b7a7 - const __m128i rgba0 = _mm_unpacklo_epi16(rg0, ba0); // rgba0|rgba1... - const __m128i rgba4 = _mm_unpackhi_epi16(rg0, ba0); // rgba4|rgba5... - _mm_storeu_si128(out++, rgba0); - _mm_storeu_si128(out++, rgba4); - num_pixels -= 8; - } - // left-overs - ConvertBGRAToRGBA((const uint32_t*)in, num_pixels, (uint8_t*)out); -} - -static void ConvertBGRAToRGBA4444SSE2(const uint32_t* src, - int num_pixels, uint8_t* dst) { - const __m128i mask_0x0f = _mm_set1_epi8(0x0f); - const __m128i mask_0xf0 = _mm_set1_epi8(0xf0); - const __m128i* in = (const __m128i*)src; - __m128i* out = (__m128i*)dst; - while (num_pixels >= 8) { - const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 - const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 - const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... - const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... - const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... - const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... - const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 - const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 - const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 - const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 - const __m128i ga1 = _mm_srli_epi16(ga0, 4); // g0-|g1-|...|a6-|a7- - const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0); // -r0|-r1|...|-b6|-a7 - const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f); // g0-|g1-|...|a6-|a7- - const __m128i rgba0 = _mm_or_si128(ga2, rb1); // rg0..rg7 | ba0..ba7 - const __m128i rgba1 = _mm_srli_si128(rgba0, 8); // ba0..ba7 | 0 -#ifdef WEBP_SWAP_16BIT_CSP - const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0); // barg0...barg7 -#else - const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1); // rgba0...rgba7 -#endif - _mm_storeu_si128(out++, rgba); - num_pixels -= 8; - } - // left-overs - ConvertBGRAToRGBA4444((const uint32_t*)in, num_pixels, (uint8_t*)out); -} - -static void ConvertBGRAToRGB565SSE2(const uint32_t* src, - int num_pixels, uint8_t* dst) { - const __m128i mask_0xe0 = _mm_set1_epi8(0xe0); - const __m128i mask_0xf8 = _mm_set1_epi8(0xf8); - const __m128i mask_0x07 = _mm_set1_epi8(0x07); - const __m128i* in = (const __m128i*)src; - __m128i* out = (__m128i*)dst; - while (num_pixels >= 8) { - const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 - const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 - const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... - const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... - const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... - const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... - const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 - const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 - const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 - const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 - const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8); // -r0..-r7|-b0..-b7 - const __m128i g_lo1 = _mm_srli_epi16(ga0, 5); - const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07); // g0-...g7-|xx (3b) - const __m128i g_hi1 = _mm_slli_epi16(ga0, 3); - const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0); // -g0...-g7|xx (3b) - const __m128i b0 = _mm_srli_si128(rb1, 8); // -b0...-b7|0 - const __m128i rg1 = _mm_or_si128(rb1, g_lo2); // gr0...gr7|xx - const __m128i b1 = _mm_srli_epi16(b0, 3); - const __m128i gb1 = _mm_or_si128(b1, g_hi2); // bg0...bg7|xx -#ifdef WEBP_SWAP_16BIT_CSP - const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1); // rggb0...rggb7 -#else - const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1); // bgrb0...bgrb7 -#endif - _mm_storeu_si128(out++, rgba); - num_pixels -= 8; - } - // left-overs - ConvertBGRAToRGB565((const uint32_t*)in, num_pixels, (uint8_t*)out); -} - -static void ConvertBGRAToBGRSSE2(const uint32_t* src, - int num_pixels, uint8_t* dst) { - const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff); - const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0); - const __m128i* in = (const __m128i*)src; - const uint8_t* const end = dst + num_pixels * 3; - // the last storel_epi64 below writes 8 bytes starting at offset 18 - while (dst + 26 <= end) { - const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 - const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 - const __m128i a0l = _mm_and_si128(bgra0, mask_l); // bgr0|0|bgr0|0 - const __m128i a4l = _mm_and_si128(bgra4, mask_l); // bgr0|0|bgr0|0 - const __m128i a0h = _mm_and_si128(bgra0, mask_h); // 0|bgr0|0|bgr0 - const __m128i a4h = _mm_and_si128(bgra4, mask_h); // 0|bgr0|0|bgr0 - const __m128i b0h = _mm_srli_epi64(a0h, 8); // 000b|gr00|000b|gr00 - const __m128i b4h = _mm_srli_epi64(a4h, 8); // 000b|gr00|000b|gr00 - const __m128i c0 = _mm_or_si128(a0l, b0h); // rgbrgb00|rgbrgb00 - const __m128i c4 = _mm_or_si128(a4l, b4h); // rgbrgb00|rgbrgb00 - const __m128i c2 = _mm_srli_si128(c0, 8); - const __m128i c6 = _mm_srli_si128(c4, 8); - _mm_storel_epi64((__m128i*)(dst + 0), c0); - _mm_storel_epi64((__m128i*)(dst + 6), c2); - _mm_storel_epi64((__m128i*)(dst + 12), c4); - _mm_storel_epi64((__m128i*)(dst + 18), c6); - dst += 24; - num_pixels -= 8; - } - // left-overs - ConvertBGRAToBGR((const uint32_t*)in, num_pixels, dst); -} - -extern void VP8LDspInitSSE2(void); - -void VP8LDspInitSSE2(void) { - VP8LClampedAddSubtractFull = ClampedAddSubtractFullSSE2; - VP8LClampedAddSubtractHalf = ClampedAddSubtractHalfSSE2; - VP8LSelect = SelectSSE2; - VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRedSSE2; - VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRedSSE2; - VP8LConvertBGRAToRGBA = ConvertBGRAToRGBASSE2; - VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444SSE2; - VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565SSE2; - VP8LConvertBGRAToBGR = ConvertBGRAToBGRSSE2; -} - -#endif // WEBP_USE_SSE2 - -//------------------------------------------------------------------------------ - -VP8LPredClampedAddSubFunc VP8LClampedAddSubtractFull; -VP8LPredClampedAddSubFunc VP8LClampedAddSubtractHalf; -VP8LPredSelectFunc VP8LSelect; -VP8LSubtractGreenFromBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed; -VP8LAddGreenToBlueAndRedFunc VP8LAddGreenToBlueAndRed; +VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed; +VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed; +VP8LPredictorFunc VP8LPredictors[16]; VP8LConvertFunc VP8LConvertBGRAToRGB; VP8LConvertFunc VP8LConvertBGRAToRGBA; @@ -1717,17 +1475,16 @@ VP8LConvertFunc VP8LConvertBGRAToRGB565; VP8LConvertFunc VP8LConvertBGRAToBGR; void VP8LDspInit(void) { - VP8LClampedAddSubtractFull = ClampedAddSubtractFull; - VP8LClampedAddSubtractHalf = ClampedAddSubtractHalf; - VP8LSelect = Select; - VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed; - VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; + memcpy(VP8LPredictors, kPredictorsC, sizeof(VP8LPredictors)); - VP8LConvertBGRAToRGB = ConvertBGRAToRGB; - VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; - VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444; - VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565; - VP8LConvertBGRAToBGR = ConvertBGRAToBGR; + VP8LSubtractGreenFromBlueAndRed = VP8LSubtractGreenFromBlueAndRed_C; + VP8LAddGreenToBlueAndRed = VP8LAddGreenToBlueAndRed_C; + + VP8LConvertBGRAToRGB = VP8LConvertBGRAToRGB_C; + VP8LConvertBGRAToRGBA = VP8LConvertBGRAToRGBA_C; + VP8LConvertBGRAToRGBA4444 = VP8LConvertBGRAToRGBA4444_C; + VP8LConvertBGRAToRGB565 = VP8LConvertBGRAToRGB565_C; + VP8LConvertBGRAToBGR = VP8LConvertBGRAToBGR_C; // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { diff --git a/src/dsp/lossless.h b/src/dsp/lossless.h index 3671021c..9ff52309 100644 --- a/src/dsp/lossless.h +++ b/src/dsp/lossless.h @@ -23,21 +23,14 @@ extern "C" { #endif //------------------------------------------------------------------------------ -// +// Signatures and generic function-pointers -typedef uint32_t (*VP8LPredClampedAddSubFunc)(uint32_t c0, uint32_t c1, - uint32_t c2); -typedef uint32_t (*VP8LPredSelectFunc)(uint32_t c0, uint32_t c1, uint32_t c2); -typedef void (*VP8LSubtractGreenFromBlueAndRedFunc)(uint32_t* argb_data, - int num_pixs); -typedef void (*VP8LAddGreenToBlueAndRedFunc)(uint32_t* data_start, - const uint32_t* data_end); +typedef uint32_t (*VP8LPredictorFunc)(uint32_t left, const uint32_t* const top); +extern VP8LPredictorFunc VP8LPredictors[16]; -extern VP8LPredClampedAddSubFunc VP8LClampedAddSubtractFull; -extern VP8LPredClampedAddSubFunc VP8LClampedAddSubtractHalf; -extern VP8LPredSelectFunc VP8LSelect; -extern VP8LSubtractGreenFromBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed; -extern VP8LAddGreenToBlueAndRedFunc VP8LAddGreenToBlueAndRed; +typedef void (*VP8LProcessBlueAndRedFunc)(uint32_t* argb_data, int num_pixels); +extern VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed; +extern VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed; typedef void (*VP8LConvertFunc)(const uint32_t* src, int num_pixels, uint8_t* dst); @@ -47,6 +40,23 @@ extern VP8LConvertFunc VP8LConvertBGRAToRGBA4444; extern VP8LConvertFunc VP8LConvertBGRAToRGB565; extern VP8LConvertFunc VP8LConvertBGRAToBGR; +extern void VP8LDspInitSSE2(void); + +// Expose some C-only fallback functions +extern void VP8LConvertBGRAToRGB_C(const uint32_t* src, + int num_pixels, uint8_t* dst); +extern void VP8LConvertBGRAToRGBA_C(const uint32_t* src, + int num_pixels, uint8_t* dst); +extern void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src, + int num_pixels, uint8_t* dst); +extern void VP8LConvertBGRAToRGB565_C(const uint32_t* src, + int num_pixels, uint8_t* dst); +extern void VP8LConvertBGRAToBGR_C(const uint32_t* src, + int num_pixels, uint8_t* dst); +extern void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, + int num_pixels); +extern void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels); + // Must be called before calling any of the above methods. void VP8LDspInit(void); diff --git a/src/dsp/lossless_sse2.c b/src/dsp/lossless_sse2.c new file mode 100644 index 00000000..25b250bd --- /dev/null +++ b/src/dsp/lossless_sse2.c @@ -0,0 +1,306 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 variant of methods for lossless decoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_SSE2) +#include +#include "./lossless.h" + +//------------------------------------------------------------------------------ +// Predictors + +static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, + uint32_t c2) { + const __m128i zero = _mm_setzero_si128(); + const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); + const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); + const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); + const __m128i V1 = _mm_add_epi16(C0, C1); + const __m128i V2 = _mm_sub_epi16(V1, C2); + const __m128i b = _mm_packus_epi16(V2, V2); + const uint32_t output = _mm_cvtsi128_si32(b); + return output; +} + +static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, + uint32_t c2) { + const __m128i zero = _mm_setzero_si128(); + const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); + const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); + const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); + const __m128i avg = _mm_add_epi16(C1, C0); + const __m128i A0 = _mm_srli_epi16(avg, 1); + const __m128i A1 = _mm_sub_epi16(A0, B0); + const __m128i BgtA = _mm_cmpgt_epi16(B0, A0); + const __m128i A2 = _mm_sub_epi16(A1, BgtA); + const __m128i A3 = _mm_srai_epi16(A2, 1); + const __m128i A4 = _mm_add_epi16(A0, A3); + const __m128i A5 = _mm_packus_epi16(A4, A4); + const uint32_t output = _mm_cvtsi128_si32(A5); + return output; +} + +static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { + int pa_minus_pb; + const __m128i zero = _mm_setzero_si128(); + const __m128i A0 = _mm_cvtsi32_si128(a); + const __m128i B0 = _mm_cvtsi32_si128(b); + const __m128i C0 = _mm_cvtsi32_si128(c); + const __m128i AC0 = _mm_subs_epu8(A0, C0); + const __m128i CA0 = _mm_subs_epu8(C0, A0); + const __m128i BC0 = _mm_subs_epu8(B0, C0); + const __m128i CB0 = _mm_subs_epu8(C0, B0); + const __m128i AC = _mm_or_si128(AC0, CA0); + const __m128i BC = _mm_or_si128(BC0, CB0); + const __m128i pa = _mm_unpacklo_epi8(AC, zero); // |a - c| + const __m128i pb = _mm_unpacklo_epi8(BC, zero); // |b - c| + const __m128i diff = _mm_sub_epi16(pb, pa); + { + int16_t out[8]; + _mm_storeu_si128((__m128i*)out, diff); + pa_minus_pb = out[0] + out[1] + out[2] + out[3]; + } + return (pa_minus_pb <= 0) ? a : b; +} + +#if 0 +// TODO(skal): optimize these to SSE2 too. +static uint32_t Predictor5(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average3(left, top[0], top[1]); + return pred; +} +static uint32_t Predictor6(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(left, top[-1]); + return pred; +} +static uint32_t Predictor7(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(left, top[0]); + return pred; +} +static uint32_t Predictor8(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(top[-1], top[0]); + (void)left; + return pred; +} +static uint32_t Predictor9(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(top[0], top[1]); + (void)left; + return pred; +} +static uint32_t Predictor10(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average4(left, top[-1], top[0], top[1]); + return pred; +} +#endif + +static uint32_t Predictor11(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Select(top[0], left, top[-1]); + return pred; +} +static uint32_t Predictor12(uint32_t left, const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]); + return pred; +} +static uint32_t Predictor13(uint32_t left, const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]); + return pred; +} + +//------------------------------------------------------------------------------ +// Colorspace conversion functions + +static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) { + const __m128i mask = _mm_set1_epi32(0x0000ff00); + int i; + for (i = 0; i + 4 < num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); + const __m128i in_00g0 = _mm_and_si128(in, mask); // 00g0|00g0|... + const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8); // 0g00|0g00|... + const __m128i in_000g = _mm_srli_epi32(in_00g0, 8); // 000g|000g|... + const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g); + const __m128i out = _mm_sub_epi8(in, in_0g0g); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i); +} + +static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) { + const __m128i mask = _mm_set1_epi32(0x0000ff00); + int i; + for (i = 0; i + 4 < num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); + const __m128i in_00g0 = _mm_and_si128(in, mask); // 00g0|00g0|... + const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8); // 0g00|0g00|... + const __m128i in_000g = _mm_srli_epi32(in_00g0, 8); // 000g|000g|... + const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g); + const __m128i out = _mm_add_epi8(in, in_0g0g); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + VP8LAddGreenToBlueAndRed_C(argb_data + i, num_pixels - i); +} + +static void ConvertBGRAToRGBA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + while (num_pixels >= 8) { + const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 + const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 + const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... + const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... + const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... + const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... + const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 + const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 + const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 + const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 + const __m128i rg0 = _mm_unpacklo_epi8(rb0, ga0); // r0g0r1g1 ... r6g6r7g7 + const __m128i ba0 = _mm_unpackhi_epi8(rb0, ga0); // b0a0b1a1 ... b6a6b7a7 + const __m128i rgba0 = _mm_unpacklo_epi16(rg0, ba0); // rgba0|rgba1... + const __m128i rgba4 = _mm_unpackhi_epi16(rg0, ba0); // rgba4|rgba5... + _mm_storeu_si128(out++, rgba0); + _mm_storeu_si128(out++, rgba4); + num_pixels -= 8; + } + // left-overs + VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out); +} + +static void ConvertBGRAToRGBA4444(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i mask_0x0f = _mm_set1_epi8(0x0f); + const __m128i mask_0xf0 = _mm_set1_epi8(0xf0); + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + while (num_pixels >= 8) { + const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 + const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 + const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... + const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... + const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... + const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... + const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 + const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 + const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 + const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 + const __m128i ga1 = _mm_srli_epi16(ga0, 4); // g0-|g1-|...|a6-|a7- + const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0); // -r0|-r1|...|-b6|-a7 + const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f); // g0-|g1-|...|a6-|a7- + const __m128i rgba0 = _mm_or_si128(ga2, rb1); // rg0..rg7 | ba0..ba7 + const __m128i rgba1 = _mm_srli_si128(rgba0, 8); // ba0..ba7 | 0 +#ifdef WEBP_SWAP_16BIT_CSP + const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0); // barg0...barg7 +#else + const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1); // rgba0...rgba7 +#endif + _mm_storeu_si128(out++, rgba); + num_pixels -= 8; + } + // left-overs + VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out); +} + +static void ConvertBGRAToRGB565(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i mask_0xe0 = _mm_set1_epi8(0xe0); + const __m128i mask_0xf8 = _mm_set1_epi8(0xf8); + const __m128i mask_0x07 = _mm_set1_epi8(0x07); + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + while (num_pixels >= 8) { + const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 + const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 + const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... + const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... + const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... + const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... + const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 + const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 + const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 + const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 + const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8); // -r0..-r7|-b0..-b7 + const __m128i g_lo1 = _mm_srli_epi16(ga0, 5); + const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07); // g0-...g7-|xx (3b) + const __m128i g_hi1 = _mm_slli_epi16(ga0, 3); + const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0); // -g0...-g7|xx (3b) + const __m128i b0 = _mm_srli_si128(rb1, 8); // -b0...-b7|0 + const __m128i rg1 = _mm_or_si128(rb1, g_lo2); // gr0...gr7|xx + const __m128i b1 = _mm_srli_epi16(b0, 3); + const __m128i gb1 = _mm_or_si128(b1, g_hi2); // bg0...bg7|xx +#ifdef WEBP_SWAP_16BIT_CSP + const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1); // rggb0...rggb7 +#else + const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1); // bgrb0...bgrb7 +#endif + _mm_storeu_si128(out++, rgba); + num_pixels -= 8; + } + // left-overs + VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out); +} + +static void ConvertBGRAToBGR(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff); + const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0); + const __m128i* in = (const __m128i*)src; + const uint8_t* const end = dst + num_pixels * 3; + // the last storel_epi64 below writes 8 bytes starting at offset 18 + while (dst + 26 <= end) { + const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 + const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 + const __m128i a0l = _mm_and_si128(bgra0, mask_l); // bgr0|0|bgr0|0 + const __m128i a4l = _mm_and_si128(bgra4, mask_l); // bgr0|0|bgr0|0 + const __m128i a0h = _mm_and_si128(bgra0, mask_h); // 0|bgr0|0|bgr0 + const __m128i a4h = _mm_and_si128(bgra4, mask_h); // 0|bgr0|0|bgr0 + const __m128i b0h = _mm_srli_epi64(a0h, 8); // 000b|gr00|000b|gr00 + const __m128i b4h = _mm_srli_epi64(a4h, 8); // 000b|gr00|000b|gr00 + const __m128i c0 = _mm_or_si128(a0l, b0h); // rgbrgb00|rgbrgb00 + const __m128i c4 = _mm_or_si128(a4l, b4h); // rgbrgb00|rgbrgb00 + const __m128i c2 = _mm_srli_si128(c0, 8); + const __m128i c6 = _mm_srli_si128(c4, 8); + _mm_storel_epi64((__m128i*)(dst + 0), c0); + _mm_storel_epi64((__m128i*)(dst + 6), c2); + _mm_storel_epi64((__m128i*)(dst + 12), c4); + _mm_storel_epi64((__m128i*)(dst + 18), c6); + dst += 24; + num_pixels -= 8; + } + // left-overs + VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst); +} + +#endif // WEBP_USE_SSE2 + +//------------------------------------------------------------------------------ + +void VP8LDspInitSSE2(void) { +#if defined(WEBP_USE_SSE2) + VP8LPredictors[11] = Predictor11; + VP8LPredictors[12] = Predictor12; + VP8LPredictors[13] = Predictor13; + + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed; + VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; + + VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; + VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444; + VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR; +#endif // WEBP_USE_SSE2 +} + +//------------------------------------------------------------------------------