1 // Copyright 2015 Google Inc. All Rights Reserved. 2 // 3 // Use of this source code is governed by a BSD-style license 4 // that can be found in the COPYING file in the root of the source 5 // tree. An additional intellectual property rights grant can be found 6 // in the file PATENTS. All contributing project authors may 7 // be found in the AUTHORS file in the root of the source tree. 8 // ----------------------------------------------------------------------------- 9 // 10 // SSE4.1 variant of methods for lossless encoder 11 // 12 // Author: Skal (pascal.massimino (at) gmail.com) 13 14 #include "src/dsp/dsp.h" 15 16 #if defined(WEBP_USE_SSE41) 17 #include <assert.h> 18 #include <smmintrin.h> 19 #include "src/dsp/lossless.h" 20 21 // For sign-extended multiplying constants, pre-shifted by 5: 22 #define CST_5b(X) (((int16_t)((uint16_t)(X) << 8)) >> 5) 23 24 //------------------------------------------------------------------------------ 25 // Subtract-Green Transform 26 27 static void SubtractGreenFromBlueAndRed_SSE41(uint32_t* argb_data, 28 int num_pixels) { 29 int i; 30 const __m128i kCstShuffle = _mm_set_epi8(-1, 13, -1, 13, -1, 9, -1, 9, 31 -1, 5, -1, 5, -1, 1, -1, 1); 32 for (i = 0; i + 4 <= num_pixels; i += 4) { 33 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); 34 const __m128i in_0g0g = _mm_shuffle_epi8(in, kCstShuffle); 35 const __m128i out = _mm_sub_epi8(in, in_0g0g); 36 _mm_storeu_si128((__m128i*)&argb_data[i], out); 37 } 38 // fallthrough and finish off with plain-C 39 if (i != num_pixels) { 40 VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i); 41 } 42 } 43 44 //------------------------------------------------------------------------------ 45 // Color Transform 46 47 #define SPAN 8 48 static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride, 49 int tile_width, int tile_height, 50 int green_to_blue, int red_to_blue, 51 int histo[]) { 52 const __m128i mults_r = _mm_set1_epi16(CST_5b(red_to_blue)); 53 const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_blue)); 54 const __m128i mask_g = _mm_set1_epi16(0xff00); // green mask 55 const __m128i mask_gb = _mm_set1_epi32(0xffff); // green/blue mask 56 const __m128i mask_b = _mm_set1_epi16(0x00ff); // blue mask 57 const __m128i shuffler_lo = _mm_setr_epi8(-1, 2, -1, 6, -1, 10, -1, 14, -1, 58 -1, -1, -1, -1, -1, -1, -1); 59 const __m128i shuffler_hi = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, 60 2, -1, 6, -1, 10, -1, 14); 61 int y; 62 for (y = 0; y < tile_height; ++y) { 63 const uint32_t* const src = argb + y * stride; 64 int i, x; 65 for (x = 0; x + SPAN <= tile_width; x += SPAN) { 66 uint16_t values[SPAN]; 67 const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); 68 const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); 69 const __m128i r0 = _mm_shuffle_epi8(in0, shuffler_lo); 70 const __m128i r1 = _mm_shuffle_epi8(in1, shuffler_hi); 71 const __m128i r = _mm_or_si128(r0, r1); // r 0 72 const __m128i gb0 = _mm_and_si128(in0, mask_gb); 73 const __m128i gb1 = _mm_and_si128(in1, mask_gb); 74 const __m128i gb = _mm_packus_epi32(gb0, gb1); // g b 75 const __m128i g = _mm_and_si128(gb, mask_g); // g 0 76 const __m128i A = _mm_mulhi_epi16(r, mults_r); // x dbr 77 const __m128i B = _mm_mulhi_epi16(g, mults_g); // x dbg 78 const __m128i C = _mm_sub_epi8(gb, B); // x b' 79 const __m128i D = _mm_sub_epi8(C, A); // x b'' 80 const __m128i E = _mm_and_si128(D, mask_b); // 0 b'' 81 _mm_storeu_si128((__m128i*)values, E); 82 for (i = 0; i < SPAN; ++i) ++histo[values[i]]; 83 } 84 } 85 { 86 const int left_over = tile_width & (SPAN - 1); 87 if (left_over > 0) { 88 VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride, 89 left_over, tile_height, 90 green_to_blue, red_to_blue, histo); 91 } 92 } 93 } 94 95 static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride, 96 int tile_width, int tile_height, 97 int green_to_red, int histo[]) { 98 const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_red)); 99 const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask 100 const __m128i mask = _mm_set1_epi16(0xff); 101 102 int y; 103 for (y = 0; y < tile_height; ++y) { 104 const uint32_t* const src = argb + y * stride; 105 int i, x; 106 for (x = 0; x + SPAN <= tile_width; x += SPAN) { 107 uint16_t values[SPAN]; 108 const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); 109 const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); 110 const __m128i g0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0 111 const __m128i g1 = _mm_and_si128(in1, mask_g); 112 const __m128i g = _mm_packus_epi32(g0, g1); // g 0 113 const __m128i A0 = _mm_srli_epi32(in0, 16); // 0 0 | x r 114 const __m128i A1 = _mm_srli_epi32(in1, 16); 115 const __m128i A = _mm_packus_epi32(A0, A1); // x r 116 const __m128i B = _mm_mulhi_epi16(g, mults_g); // x dr 117 const __m128i C = _mm_sub_epi8(A, B); // x r' 118 const __m128i D = _mm_and_si128(C, mask); // 0 r' 119 _mm_storeu_si128((__m128i*)values, D); 120 for (i = 0; i < SPAN; ++i) ++histo[values[i]]; 121 } 122 } 123 { 124 const int left_over = tile_width & (SPAN - 1); 125 if (left_over > 0) { 126 VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride, 127 left_over, tile_height, green_to_red, 128 histo); 129 } 130 } 131 } 132 133 //------------------------------------------------------------------------------ 134 // Entry point 135 136 extern void VP8LEncDspInitSSE41(void); 137 138 WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE41(void) { 139 VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_SSE41; 140 VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE41; 141 VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE41; 142 } 143 144 #else // !WEBP_USE_SSE41 145 146 WEBP_DSP_INIT_STUB(VP8LEncDspInitSSE41) 147 148 #endif // WEBP_USE_SSE41 149
