third_party/libwebp/dsp/alpha_processing_sse2.c

   1 // Copyright 2014 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.
  11 //
  12 // Author: Skal (pascal.massimino@gmail.com)
  13
  14 #include "./dsp.h"
  15
  16 #if defined(WEBP_USE_SSE2)
  17 #include <emmintrin.h>
  18
  19 //------------------------------------------------------------------------------
  20
  21 static int ExtractAlpha(const uint8_t* argb, int argb_stride,
  22                         int width, int height,
  23                         uint8_t* alpha, int alpha_stride) {
  24   // alpha_and stores an 'and' operation of all the alpha[] values. The final
  25   // value is not 0xff if any of the alpha[] is not equal to 0xff.
  26   uint32_t alpha_and = 0xff;
  27   int i, j;
  28   const __m128i a_mask = _mm_set1_epi32(0xffu);  // to preserve alpha
  29   const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u);
  30   __m128i all_alphas = all_0xff;
  31
  32   // We must be able to access 3 extra bytes after the last written byte
  33   // 'src[4 * width - 4]', because we don't know if alpha is the first or the
  34   // last byte of the quadruplet.
  35   const int limit = (width - 1) & ~7;
  36
  37   for (j = 0; j < height; ++j) {
  38     const __m128i* src = (const __m128i*)argb;
  39     for (i = 0; i < limit; i += 8) {
  40       // load 32 argb bytes
  41       const __m128i a0 = _mm_loadu_si128(src + 0);
  42       const __m128i a1 = _mm_loadu_si128(src + 1);
  43       const __m128i b0 = _mm_and_si128(a0, a_mask);
  44       const __m128i b1 = _mm_and_si128(a1, a_mask);
  45       const __m128i c0 = _mm_packs_epi32(b0, b1);
  46       const __m128i d0 = _mm_packus_epi16(c0, c0);
  47       // store
  48       _mm_storel_epi64((__m128i*)&alpha[i], d0);
  49       // accumulate eight alpha 'and' in parallel
  50       all_alphas = _mm_and_si128(all_alphas, d0);
  51       src += 2;
  52     }
  53     for (; i < width; ++i) {
  54       const uint32_t alpha_value = argb[4 * i];
  55       alpha[i] = alpha_value;
  56       alpha_and &= alpha_value;
  57     }
  58     argb += argb_stride;
  59     alpha += alpha_stride;
  60   }
  61   // Combine the eight alpha 'and' into a 8-bit mask.
  62   alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff));
  63   return (alpha_and == 0xff);
  64 }
  65
  66 #endif   // WEBP_USE_SSE2
  67
  68 //------------------------------------------------------------------------------
  69 // Init function
  70
  71 extern void WebPInitAlphaProcessingSSE2(void);
  72
  73 void WebPInitAlphaProcessingSSE2(void) {
  74 #if defined(WEBP_USE_SSE2)
  75   WebPExtractAlpha = ExtractAlpha;
  76 #endif
  77 }