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/dsp.h"
15
16 #if defined(WEBP_USE_SSE41)
17
18 #include <smmintrin.h>
19
20 //------------------------------------------------------------------------------
21
ExtractAlpha_SSE41(const uint8_t * argb,int argb_stride,int width,int height,uint8_t * alpha,int alpha_stride)22 static int ExtractAlpha_SSE41(const uint8_t* argb, int argb_stride,
23 int width, int height,
24 uint8_t* alpha, int alpha_stride) {
25 // alpha_and stores an 'and' operation of all the alpha[] values. The final
26 // value is not 0xff if any of the alpha[] is not equal to 0xff.
27 uint32_t alpha_and = 0xff;
28 int i, j;
29 const __m128i all_0xff = _mm_set1_epi32(~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) & ~15;
36 const __m128i kCstAlpha0 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1,
37 -1, -1, -1, -1, 12, 8, 4, 0);
38 const __m128i kCstAlpha1 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1,
39 12, 8, 4, 0, -1, -1, -1, -1);
40 const __m128i kCstAlpha2 = _mm_set_epi8(-1, -1, -1, -1, 12, 8, 4, 0,
41 -1, -1, -1, -1, -1, -1, -1, -1);
42 const __m128i kCstAlpha3 = _mm_set_epi8(12, 8, 4, 0, -1, -1, -1, -1,
43 -1, -1, -1, -1, -1, -1, -1, -1);
44 for (j = 0; j < height; ++j) {
45 const __m128i* src = (const __m128i*)argb;
46 for (i = 0; i < limit; i += 16) {
47 // load 64 argb bytes
48 const __m128i a0 = _mm_loadu_si128(src + 0);
49 const __m128i a1 = _mm_loadu_si128(src + 1);
50 const __m128i a2 = _mm_loadu_si128(src + 2);
51 const __m128i a3 = _mm_loadu_si128(src + 3);
52 const __m128i b0 = _mm_shuffle_epi8(a0, kCstAlpha0);
53 const __m128i b1 = _mm_shuffle_epi8(a1, kCstAlpha1);
54 const __m128i b2 = _mm_shuffle_epi8(a2, kCstAlpha2);
55 const __m128i b3 = _mm_shuffle_epi8(a3, kCstAlpha3);
56 const __m128i c0 = _mm_or_si128(b0, b1);
57 const __m128i c1 = _mm_or_si128(b2, b3);
58 const __m128i d0 = _mm_or_si128(c0, c1);
59 // store
60 _mm_storeu_si128((__m128i*)&alpha[i], d0);
61 // accumulate sixteen alpha 'and' in parallel
62 all_alphas = _mm_and_si128(all_alphas, d0);
63 src += 4;
64 }
65 for (; i < width; ++i) {
66 const uint32_t alpha_value = argb[4 * i];
67 alpha[i] = alpha_value;
68 alpha_and &= alpha_value;
69 }
70 argb += argb_stride;
71 alpha += alpha_stride;
72 }
73 // Combine the sixteen alpha 'and' into an 8-bit mask.
74 alpha_and |= 0xff00u; // pretend the upper bits [8..15] were tested ok.
75 alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff));
76 return (alpha_and == 0xffffu);
77 }
78
79 //------------------------------------------------------------------------------
80 // Entry point
81
82 extern void WebPInitAlphaProcessingSSE41(void);
83
WebPInitAlphaProcessingSSE41(void)84 WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE41(void) {
85 WebPExtractAlpha = ExtractAlpha_SSE41;
86 }
87
88 #else // !WEBP_USE_SSE41
89
90 WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingSSE41)
91
92 #endif // WEBP_USE_SSE41
93