• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2016 Google Inc.
3  *
4  * Use of this source code is governed by a BSD-style license that can be
5  * found in the LICENSE file.
6  */
7 
8 // It is not safe to #include any header file here unless it has been vetted for ODR safety:
9 // all symbols used must be file-scoped static or in an anonymous namespace.  This applies
10 // to _all_ header files:  C standard library, C++ standard library, Skia... everything.
11 
12 #include <immintrin.h>   // ODR safe
13 #include <stdint.h>      // ODR safe
14 
15 #if defined(__AVX2__)
16 
17 namespace hsw {
18 
convolve_vertically(const int16_t * filter,int filterLen,uint8_t * const * srcRows,int width,uint8_t * out,bool hasAlpha)19     void convolve_vertically(const int16_t* filter, int filterLen,
20                              uint8_t* const* srcRows, int width,
21                              uint8_t* out, bool hasAlpha) {
22         // It's simpler to work with the output array in terms of 4-byte pixels.
23         auto dst = (int*)out;
24 
25         // Output up to eight pixels per iteration.
26         for (int x = 0; x < width; x += 8) {
27             // Accumulated result for 4 (non-adjacent) pairs of pixels,
28             // with each channel in signed 17.14 fixed point.
29             auto accum04 = _mm256_setzero_si256(),
30                  accum15 = _mm256_setzero_si256(),
31                  accum26 = _mm256_setzero_si256(),
32                  accum37 = _mm256_setzero_si256();
33 
34             // Convolve with the filter.  (This inner loop is where we spend ~all our time.)
35             // While we can, we consume 2 filter coefficients and 2 rows of 8 pixels each at a time.
36             auto convolve_16_pixels = [&](__m256i interlaced_coeffs,
37                                           __m256i pixels_01234567, __m256i pixels_89ABCDEF) {
38                 // Interlaced R0R8 G0G8 B0B8 A0A8 R1R9 G1G9... 32 8-bit values each.
39                 auto _08194C5D = _mm256_unpacklo_epi8(pixels_01234567, pixels_89ABCDEF),
40                      _2A3B6E7F = _mm256_unpackhi_epi8(pixels_01234567, pixels_89ABCDEF);
41 
42                 // Still interlaced R0R8 G0G8... as above, each channel expanded to 16-bit lanes.
43                 auto _084C = _mm256_unpacklo_epi8(_08194C5D, _mm256_setzero_si256()),
44                      _195D = _mm256_unpackhi_epi8(_08194C5D, _mm256_setzero_si256()),
45                      _2A6E = _mm256_unpacklo_epi8(_2A3B6E7F, _mm256_setzero_si256()),
46                      _3B7F = _mm256_unpackhi_epi8(_2A3B6E7F, _mm256_setzero_si256());
47 
48                 // accum0_R += R0*coeff0 + R8*coeff1, etc.
49                 accum04 = _mm256_add_epi32(accum04, _mm256_madd_epi16(_084C, interlaced_coeffs));
50                 accum15 = _mm256_add_epi32(accum15, _mm256_madd_epi16(_195D, interlaced_coeffs));
51                 accum26 = _mm256_add_epi32(accum26, _mm256_madd_epi16(_2A6E, interlaced_coeffs));
52                 accum37 = _mm256_add_epi32(accum37, _mm256_madd_epi16(_3B7F, interlaced_coeffs));
53             };
54 
55             int i = 0;
56             for (; i < filterLen/2*2; i += 2) {
57                 convolve_16_pixels(_mm256_set1_epi32(*(const int32_t*)(filter+i)),
58                                    _mm256_loadu_si256((const __m256i*)(srcRows[i+0] + x*4)),
59                                    _mm256_loadu_si256((const __m256i*)(srcRows[i+1] + x*4)));
60             }
61             if (i < filterLen) {
62                 convolve_16_pixels(_mm256_set1_epi32(*(const int16_t*)(filter+i)),
63                                    _mm256_loadu_si256((const __m256i*)(srcRows[i] + x*4)),
64                                    _mm256_setzero_si256());
65             }
66 
67             // Trim the fractional parts off the accumulators.
68             accum04 = _mm256_srai_epi32(accum04, 14);
69             accum15 = _mm256_srai_epi32(accum15, 14);
70             accum26 = _mm256_srai_epi32(accum26, 14);
71             accum37 = _mm256_srai_epi32(accum37, 14);
72 
73             // Pack back down to 8-bit channels.
74             auto pixels = _mm256_packus_epi16(_mm256_packs_epi32(accum04, accum15),
75                                               _mm256_packs_epi32(accum26, accum37));
76 
77             if (hasAlpha) {
78                 // Clamp alpha to the max of r,g,b to make sure we stay premultiplied.
79                 __m256i max_rg  = _mm256_max_epu8(pixels, _mm256_srli_epi32(pixels,  8)),
80                         max_rgb = _mm256_max_epu8(max_rg, _mm256_srli_epi32(pixels, 16));
81                 pixels = _mm256_max_epu8(pixels, _mm256_slli_epi32(max_rgb, 24));
82             } else {
83                 // Force opaque.
84                 pixels = _mm256_or_si256(pixels, _mm256_set1_epi32(0xff000000));
85             }
86 
87             // Normal path to store 8 pixels.
88             if (x + 8 <= width) {
89                 _mm256_storeu_si256((__m256i*)dst, pixels);
90                 dst += 8;
91                 continue;
92             }
93 
94             // Store one pixel at a time on the last iteration.
95             for (int i = x; i < width; i++) {
96                 *dst++ = _mm_cvtsi128_si32(_mm256_castsi256_si128(pixels));
97                 pixels = _mm256_permutevar8x32_epi32(pixels, _mm256_setr_epi32(1,2,3,4,5,6,7,0));
98             }
99         }
100     }
101 
102 }
103 
104 namespace SkOpts {
105     // See SkOpts.h, writing SkConvolutionFilter1D::ConvolutionFixed as the underlying type.
106     extern void (*convolve_vertically)(const int16_t* filter, int filterLen,
107                                        uint8_t* const* srcRows, int width,
108                                        uint8_t* out, bool hasAlpha);
Init_hsw()109     void Init_hsw() {
110         convolve_vertically = hsw::convolve_vertically;
111     }
112 }
113 
114 #else  // defined(__AVX2__) is not true...
115 
Init_hsw()116 namespace SkOpts { void Init_hsw() {} }
117 
118 #endif
119