Path: blob/master/thirdparty/libwebp/src/dsp/alpha_processing_sse41.c
9913 views
// Copyright 2015 Google Inc. All Rights Reserved.1//2// Use of this source code is governed by a BSD-style license3// that can be found in the COPYING file in the root of the source4// tree. An additional intellectual property rights grant can be found5// in the file PATENTS. All contributing project authors may6// be found in the AUTHORS file in the root of the source tree.7// -----------------------------------------------------------------------------8//9// Utilities for processing transparent channel, SSE4.1 variant.10//11// Author: Skal ([email protected])1213#include "src/dsp/dsp.h"1415#if defined(WEBP_USE_SSE41)1617#include <smmintrin.h>1819//------------------------------------------------------------------------------2021static int ExtractAlpha_SSE41(const uint8_t* WEBP_RESTRICT argb,22int argb_stride, int width, int height,23uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {24// alpha_and stores an 'and' operation of all the alpha[] values. The final25// value is not 0xff if any of the alpha[] is not equal to 0xff.26uint32_t alpha_and = 0xff;27int i, j;28const __m128i all_0xff = _mm_set1_epi32(~0);29__m128i all_alphas = all_0xff;3031// We must be able to access 3 extra bytes after the last written byte32// 'src[4 * width - 4]', because we don't know if alpha is the first or the33// last byte of the quadruplet.34const int limit = (width - 1) & ~15;35const __m128i kCstAlpha0 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1,36-1, -1, -1, -1, 12, 8, 4, 0);37const __m128i kCstAlpha1 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1,3812, 8, 4, 0, -1, -1, -1, -1);39const __m128i kCstAlpha2 = _mm_set_epi8(-1, -1, -1, -1, 12, 8, 4, 0,40-1, -1, -1, -1, -1, -1, -1, -1);41const __m128i kCstAlpha3 = _mm_set_epi8(12, 8, 4, 0, -1, -1, -1, -1,42-1, -1, -1, -1, -1, -1, -1, -1);43for (j = 0; j < height; ++j) {44const __m128i* src = (const __m128i*)argb;45for (i = 0; i < limit; i += 16) {46// load 64 argb bytes47const __m128i a0 = _mm_loadu_si128(src + 0);48const __m128i a1 = _mm_loadu_si128(src + 1);49const __m128i a2 = _mm_loadu_si128(src + 2);50const __m128i a3 = _mm_loadu_si128(src + 3);51const __m128i b0 = _mm_shuffle_epi8(a0, kCstAlpha0);52const __m128i b1 = _mm_shuffle_epi8(a1, kCstAlpha1);53const __m128i b2 = _mm_shuffle_epi8(a2, kCstAlpha2);54const __m128i b3 = _mm_shuffle_epi8(a3, kCstAlpha3);55const __m128i c0 = _mm_or_si128(b0, b1);56const __m128i c1 = _mm_or_si128(b2, b3);57const __m128i d0 = _mm_or_si128(c0, c1);58// store59_mm_storeu_si128((__m128i*)&alpha[i], d0);60// accumulate sixteen alpha 'and' in parallel61all_alphas = _mm_and_si128(all_alphas, d0);62src += 4;63}64for (; i < width; ++i) {65const uint32_t alpha_value = argb[4 * i];66alpha[i] = alpha_value;67alpha_and &= alpha_value;68}69argb += argb_stride;70alpha += alpha_stride;71}72// Combine the sixteen alpha 'and' into an 8-bit mask.73alpha_and |= 0xff00u; // pretend the upper bits [8..15] were tested ok.74alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff));75return (alpha_and == 0xffffu);76}7778//------------------------------------------------------------------------------79// Entry point8081extern void WebPInitAlphaProcessingSSE41(void);8283WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE41(void) {84WebPExtractAlpha = ExtractAlpha_SSE41;85}8687#else // !WEBP_USE_SSE418889WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingSSE41)9091#endif // WEBP_USE_SSE41929394