alpha_processing_sse41.c (3820B)
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 // Utilities for processing transparent channel, SSE4.1 variant. 11 // 12 // Author: Skal (pascal.massimino@gmail.com) 13 14 #include "src/dsp/cpu.h" 15 #include "src/webp/types.h" 16 #include "src/dsp/dsp.h" 17 18 #if defined(WEBP_USE_SSE41) 19 #include <emmintrin.h> 20 #include <smmintrin.h> 21 22 //------------------------------------------------------------------------------ 23 24 static int ExtractAlpha_SSE41(const uint8_t* WEBP_RESTRICT argb, 25 int argb_stride, int width, int height, 26 uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { 27 // alpha_and stores an 'and' operation of all the alpha[] values. The final 28 // value is not 0xff if any of the alpha[] is not equal to 0xff. 29 uint32_t alpha_and = 0xff; 30 int i, j; 31 const __m128i all_0xff = _mm_set1_epi32(~0); 32 __m128i all_alphas = all_0xff; 33 34 // We must be able to access 3 extra bytes after the last written byte 35 // 'src[4 * width - 4]', because we don't know if alpha is the first or the 36 // last byte of the quadruplet. 37 const int limit = (width - 1) & ~15; 38 const __m128i kCstAlpha0 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, 39 -1, -1, -1, -1, 12, 8, 4, 0); 40 const __m128i kCstAlpha1 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, 41 12, 8, 4, 0, -1, -1, -1, -1); 42 const __m128i kCstAlpha2 = _mm_set_epi8(-1, -1, -1, -1, 12, 8, 4, 0, 43 -1, -1, -1, -1, -1, -1, -1, -1); 44 const __m128i kCstAlpha3 = _mm_set_epi8(12, 8, 4, 0, -1, -1, -1, -1, 45 -1, -1, -1, -1, -1, -1, -1, -1); 46 for (j = 0; j < height; ++j) { 47 const __m128i* src = (const __m128i*)argb; 48 for (i = 0; i < limit; i += 16) { 49 // load 64 argb bytes 50 const __m128i a0 = _mm_loadu_si128(src + 0); 51 const __m128i a1 = _mm_loadu_si128(src + 1); 52 const __m128i a2 = _mm_loadu_si128(src + 2); 53 const __m128i a3 = _mm_loadu_si128(src + 3); 54 const __m128i b0 = _mm_shuffle_epi8(a0, kCstAlpha0); 55 const __m128i b1 = _mm_shuffle_epi8(a1, kCstAlpha1); 56 const __m128i b2 = _mm_shuffle_epi8(a2, kCstAlpha2); 57 const __m128i b3 = _mm_shuffle_epi8(a3, kCstAlpha3); 58 const __m128i c0 = _mm_or_si128(b0, b1); 59 const __m128i c1 = _mm_or_si128(b2, b3); 60 const __m128i d0 = _mm_or_si128(c0, c1); 61 // store 62 _mm_storeu_si128((__m128i*)&alpha[i], d0); 63 // accumulate sixteen alpha 'and' in parallel 64 all_alphas = _mm_and_si128(all_alphas, d0); 65 src += 4; 66 } 67 for (; i < width; ++i) { 68 const uint32_t alpha_value = argb[4 * i]; 69 alpha[i] = alpha_value; 70 alpha_and &= alpha_value; 71 } 72 argb += argb_stride; 73 alpha += alpha_stride; 74 } 75 // Combine the sixteen alpha 'and' into an 8-bit mask. 76 alpha_and |= 0xff00u; // pretend the upper bits [8..15] were tested ok. 77 alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff)); 78 return (alpha_and == 0xffffu); 79 } 80 81 //------------------------------------------------------------------------------ 82 // Entry point 83 84 extern void WebPInitAlphaProcessingSSE41(void); 85 86 WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE41(void) { 87 WebPExtractAlpha = ExtractAlpha_SSE41; 88 } 89 90 #else // !WEBP_USE_SSE41 91 92 WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingSSE41) 93 94 #endif // WEBP_USE_SSE41