convert/intel/convert_bits_avx2.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // Avisynth v2.5. Copyright 2002-2009 Ben Rudiak-Gould et al. | ||
| 2 | // http://avisynth.nl | ||
| 3 | |||
| 4 | // This program is free software; you can redistribute it and/or modify | ||
| 5 | // it under the terms of the GNU General Public License as published by | ||
| 6 | // the Free Software Foundation; either version 2 of the License, or | ||
| 7 | // (at your option) any later version. | ||
| 8 | // | ||
| 9 | // This program is distributed in the hope that it will be useful, | ||
| 10 | // but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 11 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 12 | // GNU General Public License for more details. | ||
| 13 | // | ||
| 14 | // You should have received a copy of the GNU General Public License | ||
| 15 | // along with this program; if not, write to the Free Software | ||
| 16 | // Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA, or visit | ||
| 17 | // http://www.gnu.org/copyleft/gpl.html . | ||
| 18 | // | ||
| 19 | // Linking Avisynth statically or dynamically with other modules is making a | ||
| 20 | // combined work based on Avisynth. Thus, the terms and conditions of the GNU | ||
| 21 | // General Public License cover the whole combination. | ||
| 22 | // | ||
| 23 | // As a special exception, the copyright holders of Avisynth give you | ||
| 24 | // permission to link Avisynth with independent modules that communicate with | ||
| 25 | // Avisynth solely through the interfaces defined in avisynth.h, regardless of the license | ||
| 26 | // terms of these independent modules, and to copy and distribute the | ||
| 27 | // resulting combined work under terms of your choice, provided that | ||
| 28 | // every copy of the combined work is accompanied by a complete copy of | ||
| 29 | // the source code of Avisynth (the version of Avisynth used to produce the | ||
| 30 | // combined work), being distributed under the terms of the GNU General | ||
| 31 | // Public License plus this exception. An independent module is a module | ||
| 32 | // which is not derived from or based on Avisynth, such as 3rd-party filters, | ||
| 33 | // import and export plugins, or graphical user interfaces. | ||
| 34 | |||
| 35 | |||
| 36 | #include <avs/alignment.h> | ||
| 37 | #include <avs/minmax.h> | ||
| 38 | |||
| 39 | #ifdef _MSC_VER | ||
| 40 | #include <intrin.h> | ||
| 41 | #else | ||
| 42 | #include <x86intrin.h> | ||
| 43 | #endif | ||
| 44 | #include <immintrin.h> | ||
| 45 | |||
| 46 | #ifndef _mm256_set_m128i | ||
| 47 | #define _mm256_set_m128i(v0, v1) _mm256_insertf128_si256(_mm256_castsi128_si256(v1), (v0), 1) | ||
| 48 | #endif | ||
| 49 | |||
| 50 | #ifndef _mm256_set_m128 | ||
| 51 | #define _mm256_set_m128(v0, v1) _mm256_insertf128_ps(_mm256_castps128_ps256(v1), (v0), 1) | ||
| 52 | #endif | ||
| 53 | |||
| 54 | #include "convert_bits_avx2.h" | ||
| 55 | #include "../convert_bits.h" | ||
| 56 | #include "../convert_helper.h" | ||
| 57 | |||
| 58 | #ifdef _MSC_VER | ||
| 59 | #pragma warning(push) | ||
| 60 | #pragma warning(disable: 4305) | ||
| 61 | #endif | ||
| 62 | |||
| 63 | template<typename pixel_t, bool chroma, bool fulls, bool fulld> | ||
| 64 | 5116 | void convert_32_to_uintN_avx2(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth) | |
| 65 | { | ||
| 66 | 5116 | const float* srcp0 = reinterpret_cast<const float*>(srcp); | |
| 67 | 5116 | pixel_t* dstp0 = reinterpret_cast<pixel_t*>(dstp); | |
| 68 | |||
| 69 | 5116 | src_pitch = src_pitch / sizeof(float); | |
| 70 | 5116 | dst_pitch = dst_pitch / sizeof(pixel_t); | |
| 71 | |||
| 72 | 5116 | const int src_width = src_rowsize / sizeof(float); | |
| 73 | |||
| 74 | //----------------------- | ||
| 75 |
7/32void convert_32_to_uintN_avx2<unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 1 time.
void convert_32_to_uintN_avx2<unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 1 time.
void convert_32_to_uintN_avx2<unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 1 time.
void convert_32_to_uintN_avx2<unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 1 time.
void convert_32_to_uintN_avx2<unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 5110 times.
void convert_32_to_uintN_avx2<unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 24 not taken.
void convert_32_to_uintN_avx2<unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
|
5116 | bits_conv_constants d; |
| 76 | get_bits_conv_constants(d, chroma, fulls, fulld, source_bitdepth, target_bitdepth); | ||
| 77 | |||
| 78 | 5116 | auto dst_offset_plus_round = d.dst_offset + 0.5f; | |
| 79 | 5116 | constexpr auto dst_pixel_min = 0; | |
| 80 | 5116 | const auto dst_pixel_max = (1 << target_bitdepth) - 1; | |
| 81 | |||
| 82 | 5116 | auto src_offset_ps = _mm256_set1_ps(d.src_offset); | |
| 83 | 10232 | auto factor_ps = _mm256_set1_ps(d.mul_factor); | |
| 84 | 5116 | auto dst_offset_plus_round_ps = _mm256_set1_ps(dst_offset_plus_round); | |
| 85 | 5116 | auto dst_pixel_min_ps = _mm256_set1_ps((float)dst_pixel_min); | |
| 86 | 5116 | auto dst_pixel_max_ps = _mm256_set1_ps((float)dst_pixel_max); | |
| 87 | |||
| 88 |
14/32void convert_32_to_uintN_avx2<unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 100 → 65 not taken.
✗ Branch 100 → 101 not taken.
void convert_32_to_uintN_avx2<unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 100 → 65 taken 4 times.
✓ Branch 100 → 101 taken 1 time.
void convert_32_to_uintN_avx2<unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 96 → 65 not taken.
✗ Branch 96 → 97 not taken.
void convert_32_to_uintN_avx2<unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 96 → 65 taken 5 times.
✓ Branch 96 → 97 taken 1 time.
void convert_32_to_uintN_avx2<unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 96 → 65 not taken.
✗ Branch 96 → 97 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 96 → 65 not taken.
✗ Branch 96 → 97 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 96 → 65 not taken.
✗ Branch 96 → 97 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 96 → 65 not taken.
✗ Branch 96 → 97 not taken.
void convert_32_to_uintN_avx2<unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 96 → 65 not taken.
✗ Branch 96 → 97 not taken.
void convert_32_to_uintN_avx2<unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 96 → 65 taken 5 times.
✓ Branch 96 → 97 taken 1 time.
void convert_32_to_uintN_avx2<unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 92 → 65 taken 4 times.
✓ Branch 92 → 93 taken 1 time.
void convert_32_to_uintN_avx2<unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 92 → 65 taken 5110 times.
✓ Branch 92 → 93 taken 5110 times.
void convert_32_to_uintN_avx2<unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 92 → 65 not taken.
✗ Branch 92 → 93 not taken.
void convert_32_to_uintN_avx2<unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 92 → 65 taken 5 times.
✓ Branch 92 → 93 taken 1 time.
void convert_32_to_uintN_avx2<unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 92 → 65 taken 5 times.
✓ Branch 92 → 93 taken 1 time.
void convert_32_to_uintN_avx2<unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 92 → 65 not taken.
✗ Branch 92 → 93 not taken.
|
10254 | for (int y = 0; y < src_height; y++) |
| 89 | { | ||
| 90 |
14/32void convert_32_to_uintN_avx2<unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 98 → 66 not taken.
✗ Branch 98 → 99 not taken.
void convert_32_to_uintN_avx2<unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 98 → 66 taken 8 times.
✓ Branch 98 → 99 taken 4 times.
void convert_32_to_uintN_avx2<unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 94 → 66 not taken.
✗ Branch 94 → 95 not taken.
void convert_32_to_uintN_avx2<unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 94 → 66 taken 5 times.
✓ Branch 94 → 95 taken 5 times.
void convert_32_to_uintN_avx2<unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 94 → 66 not taken.
✗ Branch 94 → 95 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 94 → 66 not taken.
✗ Branch 94 → 95 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 94 → 66 not taken.
✗ Branch 94 → 95 not taken.
void convert_32_to_uintN_avx2<unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 94 → 66 not taken.
✗ Branch 94 → 95 not taken.
void convert_32_to_uintN_avx2<unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 94 → 66 not taken.
✗ Branch 94 → 95 not taken.
void convert_32_to_uintN_avx2<unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 94 → 66 taken 5 times.
✓ Branch 94 → 95 taken 5 times.
void convert_32_to_uintN_avx2<unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 90 → 66 taken 8 times.
✓ Branch 90 → 91 taken 4 times.
void convert_32_to_uintN_avx2<unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 90 → 66 taken 5110 times.
✓ Branch 90 → 91 taken 5110 times.
void convert_32_to_uintN_avx2<unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 90 → 66 not taken.
✗ Branch 90 → 91 not taken.
void convert_32_to_uintN_avx2<unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 90 → 66 taken 5 times.
✓ Branch 90 → 91 taken 5 times.
void convert_32_to_uintN_avx2<unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 90 → 66 taken 5 times.
✓ Branch 90 → 91 taken 5 times.
void convert_32_to_uintN_avx2<unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 90 → 66 not taken.
✗ Branch 90 → 91 not taken.
|
10284 | for (int x = 0; x < src_width; x += 16) // 16 pixels at a time (64 byte - alignment is OK) |
| 91 | { | ||
| 92 | __m256i result; | ||
| 93 | __m256i result_0, result_1; | ||
| 94 | 5146 | __m256 src_0 = _mm256_load_ps(reinterpret_cast<const float*>(srcp0 + x)); | |
| 95 | 10292 | __m256 src_1 = _mm256_load_ps(reinterpret_cast<const float*>(srcp0 + x + 8)); | |
| 96 | if constexpr (!chroma && !fulls) { | ||
| 97 | // when offset is different from 0 | ||
| 98 | 13 | src_0 = _mm256_sub_ps(src_0, src_offset_ps); | |
| 99 | 13 | src_1 = _mm256_sub_ps(src_1, src_offset_ps); | |
| 100 | } | ||
| 101 | |||
| 102 | 5146 | src_0 = _mm256_fmadd_ps(src_0, factor_ps, dst_offset_plus_round_ps); | |
| 103 | 5146 | src_1 = _mm256_fmadd_ps(src_1, factor_ps, dst_offset_plus_round_ps); | |
| 104 | |||
| 105 | 10292 | src_0 = _mm256_max_ps(_mm256_min_ps(src_0, dst_pixel_max_ps), dst_pixel_min_ps); | |
| 106 | 10292 | src_1 = _mm256_max_ps(_mm256_min_ps(src_1, dst_pixel_max_ps), dst_pixel_min_ps); | |
| 107 | 5146 | result_0 = _mm256_cvttps_epi32(src_0); // truncate | |
| 108 | 5146 | result_1 = _mm256_cvttps_epi32(src_1); | |
| 109 | if constexpr (sizeof(pixel_t) == 2) { | ||
| 110 | 5133 | result = _mm256_packus_epi32(result_0, result_1); | |
| 111 | 5133 | result = _mm256_permute4x64_epi64(result, (0 << 0) | (2 << 2) | (1 << 4) | (3 << 6)); | |
| 112 | 5133 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x), result); | |
| 113 | } | ||
| 114 | else { | ||
| 115 | 13 | result = _mm256_packs_epi32(result_0, result_1); | |
| 116 | 13 | result = _mm256_permute4x64_epi64(result, (0 << 0) | (2 << 2) | (1 << 4) | (3 << 6)); | |
| 117 | 13 | __m128i result128_lo = _mm256_castsi256_si128(result); | |
| 118 | 13 | __m128i result128_hi = _mm256_extractf128_si256(result, 1); | |
| 119 | 13 | __m128i result128 = _mm_packus_epi16(result128_lo, result128_hi); | |
| 120 | 13 | _mm_store_si128(reinterpret_cast<__m128i*>(dstp0 + x), result128); | |
| 121 | } | ||
| 122 | |||
| 123 | // c: | ||
| 124 | //const int pixel = (int)((srcp0[x] - src_offset) * mul_factor + dst_offset_plus_round); | ||
| 125 | //dstp0[x] = pixel_t(clamp(pixel, dst_pixel_min, dst_pixel_max)); | ||
| 126 | } | ||
| 127 | 5138 | dstp0 += dst_pitch; | |
| 128 | 5138 | srcp0 += src_pitch; | |
| 129 | } | ||
| 130 | 5116 | } | |
| 131 | |||
| 132 | #ifdef _MSC_VER | ||
| 133 | #pragma warning(pop) | ||
| 134 | #endif | ||
| 135 | |||
| 136 | #define convert_32_to_uintN_avx2_functions(type) \ | ||
| 137 | template void convert_32_to_uintN_avx2<type, false, true, true>(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 138 | template void convert_32_to_uintN_avx2<type, true, true, true>(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 139 | template void convert_32_to_uintN_avx2<type, false, true, false>(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 140 | template void convert_32_to_uintN_avx2<type, true, true, false>(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 141 | template void convert_32_to_uintN_avx2<type, false, false, true>(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 142 | template void convert_32_to_uintN_avx2<type, true, false, true>(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 143 | template void convert_32_to_uintN_avx2<type, false, false, false>(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 144 | template void convert_32_to_uintN_avx2<type, true, false, false>(const BYTE *srcp, BYTE *dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 145 | |||
| 146 | convert_32_to_uintN_avx2_functions(uint8_t) | ||
| 147 | convert_32_to_uintN_avx2_functions(uint16_t) | ||
| 148 | |||
| 149 | #undef convert_32_to_uintN_avx2_functions | ||
| 150 | |||
| 151 | /* | ||
| 152 | 3.7.1 test26: stop using manual _mm256_zeroupper() | ||
| 153 | This warning is left in Avisynth at one single place. | ||
| 154 | 1.) compilers do it automatically | ||
| 155 | 2.) clang is getting mad and produces suboptimal prologue and epilogue, saving and loading all xmm registers! | ||
| 156 | ** Good (no manual zeroupper) ** | ||
| 157 | push rsi | ||
| 158 | .seh_pushreg rsi | ||
| 159 | push rdi | ||
| 160 | .seh_pushreg rdi | ||
| 161 | .seh_endprologue | ||
| 162 | [...] | ||
| 163 | pop rdi | ||
| 164 | pop rsi | ||
| 165 | vzeroupper | ||
| 166 | ret | ||
| 167 | |||
| 168 | ** Bad, manual _mm256_zeroupper() ** | ||
| 169 | |||
| 170 | vmovaps xmmword ptr [rsp + 144], xmm15 # 16-byte Spill | ||
| 171 | .seh_savexmm xmm15, 144 | ||
| 172 | vmovaps xmmword ptr [rsp + 128], xmm14 # 16-byte Spill | ||
| 173 | .seh_savexmm xmm14, 128 | ||
| 174 | vmovaps xmmword ptr [rsp + 112], xmm13 # 16-byte Spill | ||
| 175 | .seh_savexmm xmm13, 112 | ||
| 176 | vmovaps xmmword ptr [rsp + 96], xmm12 # 16-byte Spill | ||
| 177 | .seh_savexmm xmm12, 96 | ||
| 178 | vmovaps xmmword ptr [rsp + 80], xmm11 # 16-byte Spill | ||
| 179 | .seh_savexmm xmm11, 80 | ||
| 180 | vmovaps xmmword ptr [rsp + 64], xmm10 # 16-byte Spill | ||
| 181 | .seh_savexmm xmm10, 64 | ||
| 182 | vmovaps xmmword ptr [rsp + 48], xmm9 # 16-byte Spill | ||
| 183 | .seh_savexmm xmm9, 48 | ||
| 184 | vmovaps xmmword ptr [rsp + 32], xmm8 # 16-byte Spill | ||
| 185 | .seh_savexmm xmm8, 32 | ||
| 186 | vmovaps xmmword ptr [rsp + 16], xmm7 # 16-byte Spill | ||
| 187 | .seh_savexmm xmm7, 16 | ||
| 188 | vmovaps xmmword ptr [rsp], xmm6 # 16-byte Spill | ||
| 189 | .seh_savexmm xmm6, 0 | ||
| 190 | .seh_endprologue | ||
| 191 | [...] | ||
| 192 | pop rdi | ||
| 193 | pop rsi | ||
| 194 | vzeroupper | ||
| 195 | vmovaps xmm6, xmmword ptr [rsp] # 16-byte Reload | ||
| 196 | vmovaps xmm7, xmmword ptr [rsp + 16] # 16-byte Reload | ||
| 197 | vmovaps xmm8, xmmword ptr [rsp + 32] # 16-byte Reload | ||
| 198 | vmovaps xmm9, xmmword ptr [rsp + 48] # 16-byte Reload | ||
| 199 | vmovaps xmm10, xmmword ptr [rsp + 64] # 16-byte Reload | ||
| 200 | vmovaps xmm11, xmmword ptr [rsp + 80] # 16-byte Reload | ||
| 201 | vmovaps xmm12, xmmword ptr [rsp + 96] # 16-byte Reload | ||
| 202 | vmovaps xmm13, xmmword ptr [rsp + 112] # 16-byte Reload | ||
| 203 | vmovaps xmm14, xmmword ptr [rsp + 128] # 16-byte Reload | ||
| 204 | vmovaps xmm15, xmmword ptr [rsp + 144] # 16-byte Reload | ||
| 205 | ret | ||
| 206 | */ | ||
| 207 | |||
| 208 | // YUV: bit shift 8-16 <=> 8-16 bits | ||
| 209 | // shift right or left, depending on expandrange | ||
| 210 | template<typename pixel_t_s, typename pixel_t_d> | ||
| 211 | 2 | static void convert_uint_limited_avx2(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth) | |
| 212 | { | ||
| 213 | 2 | const pixel_t_s* srcp0 = reinterpret_cast<const pixel_t_s*>(srcp); | |
| 214 | 2 | pixel_t_d* dstp0 = reinterpret_cast<pixel_t_d*>(dstp); | |
| 215 | |||
| 216 | 2 | src_pitch = src_pitch / sizeof(pixel_t_s); | |
| 217 | 2 | dst_pitch = dst_pitch / sizeof(pixel_t_d); | |
| 218 | |||
| 219 | 2 | const int src_width = src_rowsize / sizeof(pixel_t_s); | |
| 220 | |||
| 221 |
2/8void convert_uint_limited_avx2<unsigned char, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 23 not taken.
void convert_uint_limited_avx2<unsigned char, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 25 not taken.
void convert_uint_limited_avx2<unsigned short, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 19 taken 1 time.
void convert_uint_limited_avx2<unsigned short, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 21 not taken.
|
2 | if (target_bitdepth > source_bitdepth) // expandrange. pixel_t_d is always uint16_t |
| 222 | { | ||
| 223 | 1 | const int shift_bits = target_bitdepth - source_bitdepth; | |
| 224 | 1 | __m128i shift = _mm_set_epi32(0, 0, 0, shift_bits); | |
| 225 |
2/8void convert_uint_limited_avx2<unsigned char, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 6 not taken.
✗ Branch 22 → 35 not taken.
void convert_uint_limited_avx2<unsigned char, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 24 → 6 taken 5 times.
✓ Branch 24 → 37 taken 1 time.
void convert_uint_limited_avx2<unsigned short, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 18 → 6 not taken.
✗ Branch 18 → 46 not taken.
void convert_uint_limited_avx2<unsigned short, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 6 not taken.
✗ Branch 20 → 47 not taken.
|
6 | for (int y = 0; y < src_height; y++) |
| 226 | { | ||
| 227 |
2/8void convert_uint_limited_avx2<unsigned char, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 7 not taken.
✗ Branch 20 → 21 not taken.
void convert_uint_limited_avx2<unsigned char, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 22 → 7 taken 5 times.
✓ Branch 22 → 23 taken 5 times.
void convert_uint_limited_avx2<unsigned short, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 7 not taken.
✗ Branch 16 → 17 not taken.
void convert_uint_limited_avx2<unsigned short, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 18 → 7 not taken.
✗ Branch 18 → 19 not taken.
|
10 | for (int x = 0; x < src_width; x += 32) |
| 228 | { | ||
| 229 | __m256i src_lo, src_hi; | ||
| 230 | if constexpr (sizeof(pixel_t_s) == 1) { | ||
| 231 | 5 | auto src_lo_128 = _mm_load_si128(reinterpret_cast<const __m128i*>(srcp0 + x)); // 16* uint8 | |
| 232 | 10 | auto src_hi_128 = _mm_load_si128(reinterpret_cast<const __m128i*>(srcp0 + x + 16)); // 16* uint8 | |
| 233 | 5 | src_lo = _mm256_cvtepu8_epi16(src_lo_128); // 16* uint16 | |
| 234 | 5 | src_hi = _mm256_cvtepu8_epi16(src_hi_128); // 16* uint16 | |
| 235 | } | ||
| 236 | else { | ||
| 237 | // 64 bytes per cycle. Ok in avs+ | ||
| 238 | ✗ | src_lo = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x)); // 16* uint_16 | |
| 239 | ✗ | src_hi = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x + 16)); // 16* uint_16 | |
| 240 | } | ||
| 241 | 5 | src_lo = _mm256_sll_epi16(src_lo, shift); | |
| 242 | 5 | src_hi = _mm256_sll_epi16(src_hi, shift); | |
| 243 | if constexpr (sizeof(pixel_t_d) == 1) { | ||
| 244 | // upconvert always to 2 bytes | ||
| 245 | ✗ | assert(0); | |
| 246 | } | ||
| 247 | else { | ||
| 248 | 5 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x), src_lo); | |
| 249 | 5 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x + 16), src_hi); | |
| 250 | } | ||
| 251 | } | ||
| 252 | 5 | dstp0 += dst_pitch; | |
| 253 | 5 | srcp0 += src_pitch; | |
| 254 | } | ||
| 255 | } | ||
| 256 | else | ||
| 257 | { | ||
| 258 | // reduce range | ||
| 259 | 1 | const int shift_bits = source_bitdepth - target_bitdepth; | |
| 260 | 1 | const int round = 1 << (shift_bits - 1); | |
| 261 | 1 | __m128i shift = _mm_set_epi32(0, 0, 0, shift_bits); | |
| 262 | 1 | const auto round_simd = _mm256_set1_epi16(round); | |
| 263 | |||
| 264 |
2/8void convert_uint_limited_avx2<unsigned char, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 34 → 30 not taken.
✗ Branch 34 → 35 not taken.
void convert_uint_limited_avx2<unsigned char, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 36 → 32 not taken.
✗ Branch 36 → 37 not taken.
void convert_uint_limited_avx2<unsigned short, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 45 → 26 taken 5 times.
✓ Branch 45 → 46 taken 1 time.
void convert_uint_limited_avx2<unsigned short, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 46 → 28 not taken.
✗ Branch 46 → 47 not taken.
|
6 | for (int y = 0; y < src_height; y++) |
| 265 | { | ||
| 266 |
2/8void convert_uint_limited_avx2<unsigned char, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 32 → 31 not taken.
✗ Branch 32 → 33 not taken.
void convert_uint_limited_avx2<unsigned char, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 34 → 33 not taken.
✗ Branch 34 → 35 not taken.
void convert_uint_limited_avx2<unsigned short, unsigned char>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 43 → 27 taken 5 times.
✓ Branch 43 → 44 taken 5 times.
void convert_uint_limited_avx2<unsigned short, unsigned short>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 44 → 29 not taken.
✗ Branch 44 → 45 not taken.
|
10 | for (int x = 0; x < src_width; x += 32) |
| 267 | { | ||
| 268 | if constexpr (sizeof(pixel_t_s) == 1) | ||
| 269 | ✗ | assert(0); | |
| 270 | // downconvert always from 2 bytes | ||
| 271 | // 64 bytes per cycle. Ok in avs+ | ||
| 272 | 5 | auto src_lo = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x)); // 16* uint_16 | |
| 273 | 10 | auto src_hi = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x + 16)); // 16* uint_16 | |
| 274 | 10 | src_lo = _mm256_srl_epi16(_mm256_adds_epu16(src_lo, round_simd), shift); | |
| 275 | 10 | src_hi = _mm256_srl_epi16(_mm256_adds_epu16(src_hi, round_simd), shift); | |
| 276 | if constexpr (sizeof(pixel_t_d) == 1) { | ||
| 277 | // to 8 bits | ||
| 278 | 5 | auto dst = _mm256_packus_epi16(src_lo, src_hi); | |
| 279 | 5 | dst = _mm256_permute4x64_epi64(dst, (0 << 0) | (2 << 2) | (1 << 4) | (3 << 6)); | |
| 280 | 5 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x), dst); | |
| 281 | } | ||
| 282 | else { | ||
| 283 | ✗ | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x), src_lo); | |
| 284 | ✗ | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x + 16), src_hi); | |
| 285 | } | ||
| 286 | } | ||
| 287 | 5 | dstp0 += dst_pitch; | |
| 288 | 5 | srcp0 += src_pitch; | |
| 289 | } | ||
| 290 | } | ||
| 291 | 2 | } | |
| 292 | |||
| 293 | template<typename pixel_t_s, typename pixel_t_d, bool chroma, bool fulls, bool fulld> | ||
| 294 | 15 | void convert_uint_avx2(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth) | |
| 295 | { | ||
| 296 | // limited to limited is bitshift, see in other function | ||
| 297 | if constexpr (!fulls && !fulld) { | ||
| 298 | 2 | convert_uint_limited_avx2<pixel_t_s, pixel_t_d>(srcp, dstp, src_rowsize, src_height, src_pitch, dst_pitch, source_bitdepth, target_bitdepth, dither_target_bitdepth); | |
| 299 | 2 | return; | |
| 300 | } | ||
| 301 | |||
| 302 | 13 | const pixel_t_s* srcp0 = reinterpret_cast<const pixel_t_s*>(srcp); | |
| 303 | 13 | pixel_t_d* dstp0 = reinterpret_cast<pixel_t_d*>(dstp); | |
| 304 | |||
| 305 | 13 | src_pitch = src_pitch / sizeof(pixel_t_s); | |
| 306 | 13 | dst_pitch = dst_pitch / sizeof(pixel_t_d); | |
| 307 | |||
| 308 | 13 | const int src_width = src_rowsize / sizeof(pixel_t_s); | |
| 309 | 13 | const int wmod32 = (src_width / 32) * 32; | |
| 310 | |||
| 311 | if constexpr (sizeof(pixel_t_s) == 1 && sizeof(pixel_t_d) == 2) { | ||
| 312 |
2/4✓ Branch 2 → 3 taken 2 times.
✗ Branch 2 → 32 not taken.
✓ Branch 3 → 4 taken 2 times.
✗ Branch 3 → 32 not taken.
|
2 | if (fulls && fulld && !chroma && source_bitdepth == 8 && target_bitdepth == 16) { |
| 313 | // special case 8->16 bit full scale: * 65535 / 255 = *257 | ||
| 314 | 2 | __m256i multiplier = _mm256_set1_epi16(257); | |
| 315 | |||
| 316 |
2/2✓ Branch 30 → 9 taken 6 times.
✓ Branch 30 → 31 taken 2 times.
|
8 | for (int y = 0; y < src_height; y++) |
| 317 | { | ||
| 318 |
2/2✓ Branch 25 → 10 taken 6 times.
✓ Branch 25 → 26 taken 6 times.
|
12 | for (int x = 0; x < src_width; x += 32) |
| 319 | { | ||
| 320 | 6 | auto src_lo_128 = _mm_load_si128(reinterpret_cast<const __m128i*>(srcp0 + x)); // 16* uint8 | |
| 321 | 12 | auto src_hi_128 = _mm_load_si128(reinterpret_cast<const __m128i*>(srcp0 + x + 16)); // 16* uint8 | |
| 322 | 6 | auto src_lo = _mm256_cvtepu8_epi16(src_lo_128); // 16* uint16 | |
| 323 | 6 | auto src_hi = _mm256_cvtepu8_epi16(src_hi_128); // 16* uint16 | |
| 324 | |||
| 325 | // *257 mullo is faster than x*257 = (x<<8 + x) add/or solution (i7) | ||
| 326 | 6 | auto res_lo = _mm256_mullo_epi16(src_lo, multiplier); // lower 16 bit of multiplication is enough | |
| 327 | 6 | auto res_hi = _mm256_mullo_epi16(src_hi, multiplier); | |
| 328 | // dstp[x] = srcp0[x] * 257; // RGB: full range 0..255 <-> 0..65535 (257 = 65535 / 255) | ||
| 329 | 6 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x), res_lo); | |
| 330 | 6 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x + 16), res_hi); | |
| 331 | } // for x | ||
| 332 | // rest | ||
| 333 |
2/2✓ Branch 28 → 27 taken 16 times.
✓ Branch 28 → 29 taken 6 times.
|
22 | for (int x = wmod32; x < src_width; x++) |
| 334 | { | ||
| 335 | 16 | dstp0[x] = (pixel_t_d)(srcp0[x] * 257); | |
| 336 | } | ||
| 337 | 6 | dstp0 += dst_pitch; | |
| 338 | 6 | srcp0 += src_pitch; | |
| 339 | } // for y | ||
| 340 | 2 | return; | |
| 341 | } | ||
| 342 | } | ||
| 343 | |||
| 344 | 11 | const int target_max = (1 << target_bitdepth) - 1; | |
| 345 | |||
| 346 |
6/48void convert_uint_avx2<unsigned char, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned char, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned char, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned char, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned char, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned char, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned char, unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 25 not taken.
void convert_uint_avx2<unsigned char, unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✓ Branch 3 → 25 taken 4 times.
void convert_uint_avx2<unsigned char, unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 32 → 33 not taken.
✗ Branch 32 → 54 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 25 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 25 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 25 not taken.
void convert_uint_avx2<unsigned short, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 2 times.
void convert_uint_avx2<unsigned short, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 1 time.
void convert_uint_avx2<unsigned short, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 2 times.
✗ Branch 2 → 24 not taken.
void convert_uint_avx2<unsigned short, unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
|
11 | bits_conv_constants d; |
| 347 | get_bits_conv_constants(d, chroma, fulls, fulld, source_bitdepth, target_bitdepth); | ||
| 348 | |||
| 349 | 11 | auto vect_mul_factor = _mm256_set1_ps(d.mul_factor); | |
| 350 | 11 | auto vect_src_offset = _mm256_set1_epi32(d.src_offset_i); | |
| 351 | 11 | auto dst_offset_plus_round = d.dst_offset + 0.5f; | |
| 352 | 11 | auto vect_dst_offset_plus_round = _mm256_set1_ps(dst_offset_plus_round); | |
| 353 | |||
| 354 | 22 | auto vect_target_max = _mm256_set1_epi16(target_max); | |
| 355 | |||
| 356 | 11 | auto zero = _mm256_setzero_si256(); | |
| 357 | |||
| 358 |
12/48void convert_uint_avx2<unsigned char, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 138 → 67 not taken.
✗ Branch 138 → 139 not taken.
void convert_uint_avx2<unsigned char, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 130 → 67 not taken.
✗ Branch 130 → 131 not taken.
void convert_uint_avx2<unsigned char, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 130 → 67 not taken.
✗ Branch 130 → 131 not taken.
void convert_uint_avx2<unsigned char, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 138 → 67 not taken.
✗ Branch 138 → 139 not taken.
void convert_uint_avx2<unsigned char, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 138 → 67 taken 5 times.
✓ Branch 138 → 139 taken 1 time.
void convert_uint_avx2<unsigned char, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 138 → 67 not taken.
✗ Branch 138 → 139 not taken.
void convert_uint_avx2<unsigned char, unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 138 → 68 not taken.
✗ Branch 138 → 139 not taken.
void convert_uint_avx2<unsigned char, unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 130 → 68 taken 12 times.
✓ Branch 130 → 131 taken 4 times.
void convert_uint_avx2<unsigned char, unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 159 → 97 not taken.
✗ Branch 159 → 160 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 138 → 68 not taken.
✗ Branch 138 → 139 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 138 → 68 not taken.
✗ Branch 138 → 139 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 138 → 68 not taken.
✗ Branch 138 → 139 not taken.
void convert_uint_avx2<unsigned short, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 134 → 67 taken 10 times.
✓ Branch 134 → 135 taken 2 times.
void convert_uint_avx2<unsigned short, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 126 → 67 not taken.
✗ Branch 126 → 127 not taken.
void convert_uint_avx2<unsigned short, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 126 → 67 taken 1 time.
✓ Branch 126 → 127 taken 1 time.
void convert_uint_avx2<unsigned short, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 134 → 67 not taken.
✗ Branch 134 → 135 not taken.
void convert_uint_avx2<unsigned short, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 134 → 67 not taken.
✗ Branch 134 → 135 not taken.
void convert_uint_avx2<unsigned short, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 134 → 67 not taken.
✗ Branch 134 → 135 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 67 not taken.
✗ Branch 133 → 134 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 125 → 67 not taken.
✗ Branch 125 → 126 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 125 → 67 not taken.
✗ Branch 125 → 126 not taken.
void convert_uint_avx2<unsigned short, unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 133 → 67 taken 5 times.
✓ Branch 133 → 134 taken 1 time.
void convert_uint_avx2<unsigned short, unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 133 → 67 taken 10 times.
✓ Branch 133 → 134 taken 2 times.
void convert_uint_avx2<unsigned short, unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 67 not taken.
✗ Branch 133 → 134 not taken.
|
54 | for (int y = 0; y < src_height; y++) |
| 359 | { | ||
| 360 |
12/48void convert_uint_avx2<unsigned char, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 136 → 68 not taken.
✗ Branch 136 → 137 not taken.
void convert_uint_avx2<unsigned char, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 128 → 68 not taken.
✗ Branch 128 → 129 not taken.
void convert_uint_avx2<unsigned char, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 128 → 68 not taken.
✗ Branch 128 → 129 not taken.
void convert_uint_avx2<unsigned char, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 136 → 68 not taken.
✗ Branch 136 → 137 not taken.
void convert_uint_avx2<unsigned char, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 136 → 68 taken 5 times.
✓ Branch 136 → 137 taken 5 times.
void convert_uint_avx2<unsigned char, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 136 → 68 not taken.
✗ Branch 136 → 137 not taken.
void convert_uint_avx2<unsigned char, unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 136 → 69 not taken.
✗ Branch 136 → 137 not taken.
void convert_uint_avx2<unsigned char, unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 128 → 69 taken 12 times.
✓ Branch 128 → 129 taken 12 times.
void convert_uint_avx2<unsigned char, unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 157 → 98 not taken.
✗ Branch 157 → 158 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 136 → 69 not taken.
✗ Branch 136 → 137 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 136 → 69 not taken.
✗ Branch 136 → 137 not taken.
void convert_uint_avx2<unsigned char, unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 136 → 69 not taken.
✗ Branch 136 → 137 not taken.
void convert_uint_avx2<unsigned short, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 132 → 68 taken 10 times.
✓ Branch 132 → 133 taken 10 times.
void convert_uint_avx2<unsigned short, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 124 → 68 not taken.
✗ Branch 124 → 125 not taken.
void convert_uint_avx2<unsigned short, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 124 → 68 taken 1 time.
✓ Branch 124 → 125 taken 1 time.
void convert_uint_avx2<unsigned short, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 132 → 68 not taken.
✗ Branch 132 → 133 not taken.
void convert_uint_avx2<unsigned short, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 132 → 68 not taken.
✗ Branch 132 → 133 not taken.
void convert_uint_avx2<unsigned short, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 132 → 68 not taken.
✗ Branch 132 → 133 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 131 → 68 not taken.
✗ Branch 131 → 132 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 123 → 68 not taken.
✗ Branch 123 → 124 not taken.
void convert_uint_avx2<unsigned short, unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 123 → 68 not taken.
✗ Branch 123 → 124 not taken.
void convert_uint_avx2<unsigned short, unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 131 → 68 taken 5 times.
✓ Branch 131 → 132 taken 5 times.
void convert_uint_avx2<unsigned short, unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 131 → 68 taken 10 times.
✓ Branch 131 → 132 taken 10 times.
void convert_uint_avx2<unsigned short, unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 131 → 68 not taken.
✗ Branch 131 → 132 not taken.
|
86 | for (int x = 0; x < src_width; x += 32) |
| 361 | { | ||
| 362 | __m256i src_lo, src_hi; | ||
| 363 | if constexpr (sizeof(pixel_t_s) == 1) { | ||
| 364 | 17 | auto src_lo_128 = _mm_load_si128(reinterpret_cast<const __m128i*>(srcp0 + x)); // 16* uint8 | |
| 365 | 34 | auto src_hi_128 = _mm_load_si128(reinterpret_cast<const __m128i*>(srcp0 + x + 16)); // 16* uint8 | |
| 366 | 17 | src_lo = _mm256_cvtepu8_epi16(src_lo_128); // 16* uint16 | |
| 367 | 17 | src_hi = _mm256_cvtepu8_epi16(src_hi_128); // 16* uint16 | |
| 368 | } | ||
| 369 | else { | ||
| 370 | // 64 bytes per cycle. Ok in avs+ | ||
| 371 | 26 | src_lo = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x)); // 8* uint_16 | |
| 372 | 52 | src_hi = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x + 16)); // 8* uint_16 | |
| 373 | } | ||
| 374 | |||
| 375 | __m256i result1, result2; | ||
| 376 | { | ||
| 377 | // src: 8*uint16 | ||
| 378 | // convert to int32, bias, to float | ||
| 379 | 43 | auto res_lo_i = _mm256_unpacklo_epi16(src_lo, zero); | |
| 380 | 43 | auto res_hi_i = _mm256_unpackhi_epi16(src_lo, zero); | |
| 381 | if constexpr (chroma || !fulls) { | ||
| 382 | // src_offset is not zero | ||
| 383 | 30 | res_lo_i = _mm256_sub_epi32(res_lo_i, vect_src_offset); | |
| 384 | 30 | res_hi_i = _mm256_sub_epi32(res_hi_i, vect_src_offset); | |
| 385 | } | ||
| 386 | 43 | auto res_lo = _mm256_cvtepi32_ps(res_lo_i); | |
| 387 | 43 | auto res_hi = _mm256_cvtepi32_ps(res_hi_i); | |
| 388 | // multiply, bias back+round | ||
| 389 | // avx2 mode: smart fma instruction inserted even from msvc! | ||
| 390 | 86 | res_lo = _mm256_add_ps(_mm256_mul_ps(res_lo, vect_mul_factor), vect_dst_offset_plus_round); | |
| 391 | 86 | res_hi = _mm256_add_ps(_mm256_mul_ps(res_hi, vect_mul_factor), vect_dst_offset_plus_round); | |
| 392 | // convert back w/ truncate | ||
| 393 | 43 | auto result_l = _mm256_cvttps_epi32(res_lo); // no banker's rounding | |
| 394 | 43 | auto result_h = _mm256_cvttps_epi32(res_hi); | |
| 395 | // 0 min is ensured by packus | ||
| 396 | // back to 16 bit | ||
| 397 | 43 | result1 = _mm256_packus_epi32(result_l, result_h); // 8 * 16 bit pixels | |
| 398 | 43 | result1 = _mm256_min_epu16(result1, vect_target_max); | |
| 399 | } | ||
| 400 | |||
| 401 | // byte target: not yet | ||
| 402 | if constexpr (sizeof(pixel_t_d) == 2) | ||
| 403 | 27 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x), result1); | |
| 404 | |||
| 405 | { | ||
| 406 | // src: 8*uint16 | ||
| 407 | // convert to int32, bias, to float | ||
| 408 | 43 | auto res_lo_i = _mm256_unpacklo_epi16(src_hi, zero); | |
| 409 | 43 | auto res_hi_i = _mm256_unpackhi_epi16(src_hi, zero); | |
| 410 | if constexpr (chroma || !fulls) { | ||
| 411 | // src_offset is not zero | ||
| 412 | 30 | res_lo_i = _mm256_sub_epi32(res_lo_i, vect_src_offset); | |
| 413 | 30 | res_hi_i = _mm256_sub_epi32(res_hi_i, vect_src_offset); | |
| 414 | } | ||
| 415 | 43 | auto res_lo = _mm256_cvtepi32_ps(res_lo_i); | |
| 416 | 43 | auto res_hi = _mm256_cvtepi32_ps(res_hi_i); | |
| 417 | // multiply, bias back+round | ||
| 418 | 86 | res_lo = _mm256_add_ps(_mm256_mul_ps(res_lo, vect_mul_factor), vect_dst_offset_plus_round); | |
| 419 | 86 | res_hi = _mm256_add_ps(_mm256_mul_ps(res_hi, vect_mul_factor), vect_dst_offset_plus_round); | |
| 420 | // convert back w/ truncate | ||
| 421 | 43 | auto result_l = _mm256_cvttps_epi32(res_lo); | |
| 422 | 43 | auto result_h = _mm256_cvttps_epi32(res_hi); | |
| 423 | // 0 min is ensured by packus | ||
| 424 | // back to 16 bit | ||
| 425 | 43 | result2 = _mm256_packus_epi32(result_l, result_h); // 8 * 16 bit pixels | |
| 426 | 43 | result2 = _mm256_min_epu16(result2, vect_target_max); | |
| 427 | } | ||
| 428 | |||
| 429 | if constexpr (sizeof(pixel_t_d) == 2) | ||
| 430 | 27 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x + 16), result2); | |
| 431 | else { | ||
| 432 | // byte target: store both | ||
| 433 | 16 | auto result12 = _mm256_packus_epi16(result1, result2); | |
| 434 | 16 | result12 = _mm256_permute4x64_epi64(result12, (0 << 0) | (2 << 2) | (1 << 4) | (3 << 6)); | |
| 435 | 16 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp0 + x), result12); | |
| 436 | } | ||
| 437 | } // for x | ||
| 438 | |||
| 439 | 43 | dstp0 += dst_pitch; | |
| 440 | 43 | srcp0 += src_pitch; | |
| 441 | } | ||
| 442 | 7 | } | |
| 443 | |||
| 444 | // instantiate them all | ||
| 445 | // template<typename pixel_t_s, typename pixel_t_d, bool chroma, bool fulls, bool fulld> | ||
| 446 | // spec: fulld=false, fulls=false | ||
| 447 | #define convert_uint_avx2_functions(uint_X_t, uint_X_dest_t) \ | ||
| 448 | template void convert_uint_avx2<uint_X_t, uint_X_dest_t, false, false, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 449 | template void convert_uint_avx2<uint_X_t, uint_X_dest_t, false, false, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 450 | template void convert_uint_avx2<uint_X_t, uint_X_dest_t, false, true, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 451 | template void convert_uint_avx2<uint_X_t, uint_X_dest_t, false, true, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 452 | template void convert_uint_avx2<uint_X_t, uint_X_dest_t, true, false, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 453 | template void convert_uint_avx2<uint_X_t, uint_X_dest_t, true, false, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 454 | template void convert_uint_avx2<uint_X_t, uint_X_dest_t, true, true, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 455 | template void convert_uint_avx2<uint_X_t, uint_X_dest_t, true, true, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 456 | |||
| 457 | convert_uint_avx2_functions(uint8_t, uint8_t) | ||
| 458 | convert_uint_avx2_functions(uint8_t, uint16_t) | ||
| 459 | convert_uint_avx2_functions(uint16_t, uint8_t) | ||
| 460 | convert_uint_avx2_functions(uint16_t, uint16_t) | ||
| 461 | |||
| 462 | #undef convert_uint_avx2_functions | ||
| 463 | |||
| 464 | template<typename pixel_t_s, typename pixel_t_d, bool chroma, bool fulls, bool fulld, bool TEMPLATE_NEED_BACKSCALE, bool TEMPLATE_LOW_DITHER_BITDEPTH> | ||
| 465 | 5 | static void do_convert_ordered_dither_uint_avx2(const BYTE* srcp8, BYTE* dstp8, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth) | |
| 466 | { | ||
| 467 | 5 | const pixel_t_s* srcp = reinterpret_cast<const pixel_t_s*>(srcp8); | |
| 468 | 5 | pixel_t_d* dstp = reinterpret_cast<pixel_t_d*>(dstp8); | |
| 469 | 5 | dst_pitch = dst_pitch / sizeof(pixel_t_d); | |
| 470 | |||
| 471 | 5 | src_pitch = src_pitch / sizeof(pixel_t_s); | |
| 472 | // we can safely process whole 16 (8 bit src) / 32 byte (16 bit source) chunks, Avisynth ensures at least 32 byte alignment | ||
| 473 | 5 | const int src_width = (src_rowsize / sizeof(pixel_t_s) + 15) & ~15; | |
| 474 | |||
| 475 | // helps compiler optimization | ||
| 476 | if constexpr (sizeof(pixel_t_s) == 1) | ||
| 477 | 1 | source_bitdepth = 8; | |
| 478 | if constexpr (sizeof(pixel_t_d) == 1) { | ||
| 479 | 3 | target_bitdepth = 8; | |
| 480 | if (!TEMPLATE_NEED_BACKSCALE) { | ||
| 481 | 2 | dither_target_bitdepth = 8; | |
| 482 | } | ||
| 483 | } | ||
| 484 | |||
| 485 | 5 | const int max_pixel_value_target = (1 << target_bitdepth) - 1; | |
| 486 | 5 | const int max_pixel_value_dithered = (1 << dither_target_bitdepth) - 1; | |
| 487 | // precheck ensures: | ||
| 488 | // target_bitdepth >= dither_target_bitdepth | ||
| 489 | // source_bitdepth - dither_target_bitdepth <= 8 (max precalculated table is 16x16) | ||
| 490 | 5 | const bool odd_diff = (source_bitdepth - dither_target_bitdepth) & 1; | |
| 491 | 5 | const int dither_bit_diff = (source_bitdepth - dither_target_bitdepth); | |
| 492 | 5 | const int dither_order = (dither_bit_diff + 1) / 2; | |
| 493 | 5 | const int dither_mask = (1 << dither_order) - 1; // 9,10=2 11,12=4 13,14=8 15,16=16 | |
| 494 | // 10->8: 0x01 (2x2) | ||
| 495 | // 11->8: 0x03 (4x4) | ||
| 496 | // 12->8: 0x03 (4x4) | ||
| 497 | // 14->8: 0x07 (8x8) | ||
| 498 | // 16->8: 0x0F (16x16) | ||
| 499 | const BYTE* matrix; | ||
| 500 |
5/360void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✓ Branch 3 → 8 taken 1 time.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 7 taken 1 time.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 8 not taken.
✗ Branch 3 → 12 not taken.
✗ Branch 3 → 16 not taken.
✗ Branch 3 → 20 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✓ Branch 2 → 15 taken 1 time.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✓ Branch 2 → 11 taken 1 time.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 7 not taken.
✗ Branch 2 → 11 not taken.
✗ Branch 2 → 15 not taken.
✗ Branch 2 → 19 not taken.
|
5 | switch (dither_order) { |
| 501 |
1/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✓ Branch 3 → 5 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 4 → 5 not taken.
✗ Branch 4 → 6 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
|
1 | case 1: matrix = reinterpret_cast<const BYTE*>(odd_diff ? dither2x2a_data_sse2 : dither2x2_data_sse2); break; |
| 502 |
2/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 8 → 9 taken 1 time.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 8 → 9 not taken.
✗ Branch 8 → 10 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
|
2 | case 2: matrix = reinterpret_cast<const BYTE*>(odd_diff ? dither4x4a_data_sse2 : dither4x4_data_sse2); break; |
| 503 |
1/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 12 → 13 not taken.
✗ Branch 12 → 14 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 11 → 12 taken 1 time.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 11 → 12 not taken.
✗ Branch 11 → 13 not taken.
|
1 | case 3: matrix = reinterpret_cast<const BYTE*>(odd_diff ? dither8x8a_data_sse2 : dither8x8_data_sse2); break; |
| 504 |
1/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 16 → 17 not taken.
✗ Branch 16 → 18 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✓ Branch 15 → 17 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 17 not taken.
|
1 | case 4: matrix = reinterpret_cast<const BYTE*>(odd_diff ? dither16x16a_data : dither16x16_data); break; // no spec sse, already 16 bytes long |
| 505 | ✗ | default: return; // n/a | |
| 506 | } | ||
| 507 | |||
| 508 | 5 | const int bitdiff_between_dither_and_target = target_bitdepth - dither_target_bitdepth; | |
| 509 |
5/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✓ Branch 21 → 23 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✓ Branch 20 → 22 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✓ Branch 20 → 22 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✓ Branch 20 → 22 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✓ Branch 20 → 22 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 20 → 21 not taken.
✗ Branch 20 → 22 not taken.
|
5 | assert(TEMPLATE_NEED_BACKSCALE == (target_bitdepth != dither_target_bitdepth)); // dither to x, target to y |
| 510 | |||
| 511 |
5/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✓ Branch 23 → 25 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✓ Branch 22 → 24 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✓ Branch 22 → 24 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 25 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✓ Branch 22 → 24 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✓ Branch 22 → 24 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 22 → 23 not taken.
✗ Branch 22 → 24 not taken.
|
5 | assert(TEMPLATE_LOW_DITHER_BITDEPTH == (dither_target_bitdepth < 8)); |
| 512 | // e.g. instead of 0,1 => -0.5,+0.5; 0,1,2,3 => -1.5,-0.5,0.5,1.5 | ||
| 513 | 5 | const float half_maxcorr_value = ((1 << dither_bit_diff) - 1) / 2.0f; | |
| 514 | |||
| 515 | 5 | const int source_max = (1 << source_bitdepth) - 1; | |
| 516 | //----------------------- | ||
| 517 | // When calculating src_pixel, src and dst are of the same bit depth | ||
| 518 |
5/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✓ Branch 25 → 47 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✓ Branch 24 → 46 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✓ Branch 24 → 46 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 25 → 26 not taken.
✗ Branch 25 → 47 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✓ Branch 24 → 46 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 24 → 25 taken 1 time.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
|
5 | bits_conv_constants d; |
| 519 | get_bits_conv_constants(d, chroma, fulls, fulld, source_bitdepth, source_bitdepth); // both is src_bitdepth | ||
| 520 | |||
| 521 | 5 | auto dst_offset_plus_round = d.dst_offset + 0.5f; | |
| 522 | //constexpr auto src_pixel_min = 0; | ||
| 523 | 5 | const auto src_pixel_max = source_max; | |
| 524 | 5 | const float mul_factor_backfromlowdither = (float)max_pixel_value_target / max_pixel_value_dithered; | |
| 525 | //----------------------- | ||
| 526 | |||
| 527 | 10 | const auto mul_factor_simd = _mm256_set1_ps(d.mul_factor); | |
| 528 | 5 | const auto mul_factor_backfromlowdither_simd = _mm256_set1_ps(mul_factor_backfromlowdither); | |
| 529 | 5 | const auto half_maxcorr_value_simd = _mm256_set1_ps(half_maxcorr_value); | |
| 530 | 5 | const auto zero = _mm256_setzero_si256(); | |
| 531 | 5 | const BYTE* matrix_simd16 = matrix; | |
| 532 | const BYTE* current_matrix_line; | ||
| 533 | |||
| 534 |
10/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 106 → 85 not taken.
✗ Branch 106 → 107 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 116 → 86 not taken.
✗ Branch 116 → 117 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 171 → 86 not taken.
✗ Branch 171 → 172 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 142 → 85 not taken.
✗ Branch 142 → 143 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 152 → 86 not taken.
✗ Branch 152 → 153 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 201 → 86 not taken.
✗ Branch 201 → 202 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 142 → 85 not taken.
✗ Branch 142 → 143 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 152 → 86 not taken.
✗ Branch 152 → 153 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 201 → 86 not taken.
✗ Branch 201 → 202 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 106 → 85 not taken.
✗ Branch 106 → 107 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 116 → 86 not taken.
✗ Branch 116 → 117 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 171 → 86 taken 5 times.
✓ Branch 171 → 172 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 106 → 85 not taken.
✗ Branch 106 → 107 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 116 → 86 not taken.
✗ Branch 116 → 117 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 171 → 86 not taken.
✗ Branch 171 → 172 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 142 → 85 not taken.
✗ Branch 142 → 143 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 152 → 86 not taken.
✗ Branch 152 → 153 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 201 → 86 not taken.
✗ Branch 201 → 202 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 142 → 85 not taken.
✗ Branch 142 → 143 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 152 → 86 not taken.
✗ Branch 152 → 153 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 201 → 86 not taken.
✗ Branch 201 → 202 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 106 → 85 not taken.
✗ Branch 106 → 107 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 116 → 86 not taken.
✗ Branch 116 → 117 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 171 → 86 not taken.
✗ Branch 171 → 172 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 119 → 85 taken 7 times.
✓ Branch 119 → 120 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 129 → 86 not taken.
✗ Branch 129 → 130 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 169 → 86 not taken.
✗ Branch 169 → 170 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 149 → 85 not taken.
✗ Branch 149 → 150 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 159 → 86 not taken.
✗ Branch 159 → 160 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 199 → 86 not taken.
✗ Branch 199 → 200 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 149 → 85 not taken.
✗ Branch 149 → 150 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 159 → 86 not taken.
✗ Branch 159 → 160 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 199 → 86 not taken.
✗ Branch 199 → 200 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 119 → 85 taken 5 times.
✓ Branch 119 → 120 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 129 → 86 not taken.
✗ Branch 129 → 130 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 169 → 86 not taken.
✗ Branch 169 → 170 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 119 → 85 not taken.
✗ Branch 119 → 120 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 129 → 86 not taken.
✗ Branch 129 → 130 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 169 → 86 not taken.
✗ Branch 169 → 170 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 149 → 85 not taken.
✗ Branch 149 → 150 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 159 → 86 not taken.
✗ Branch 159 → 160 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 199 → 86 not taken.
✗ Branch 199 → 200 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 149 → 85 not taken.
✗ Branch 149 → 150 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 159 → 86 not taken.
✗ Branch 159 → 160 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 199 → 86 not taken.
✗ Branch 199 → 200 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 119 → 85 not taken.
✗ Branch 119 → 120 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 129 → 86 not taken.
✗ Branch 129 → 130 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 169 → 86 not taken.
✗ Branch 169 → 170 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 121 → 85 not taken.
✗ Branch 121 → 122 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 130 → 85 not taken.
✗ Branch 130 → 131 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 170 → 85 not taken.
✗ Branch 170 → 171 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 151 → 85 not taken.
✗ Branch 151 → 152 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 160 → 85 not taken.
✗ Branch 160 → 161 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 200 → 85 not taken.
✗ Branch 200 → 201 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 151 → 85 not taken.
✗ Branch 151 → 152 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 160 → 85 not taken.
✗ Branch 160 → 161 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 200 → 85 not taken.
✗ Branch 200 → 201 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 121 → 85 not taken.
✗ Branch 121 → 122 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 130 → 85 taken 17 times.
✓ Branch 130 → 131 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 170 → 85 not taken.
✗ Branch 170 → 171 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 121 → 85 not taken.
✗ Branch 121 → 122 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 130 → 85 not taken.
✗ Branch 130 → 131 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 170 → 85 not taken.
✗ Branch 170 → 171 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 151 → 85 not taken.
✗ Branch 151 → 152 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 160 → 85 not taken.
✗ Branch 160 → 161 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 200 → 85 not taken.
✗ Branch 200 → 201 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 151 → 85 taken 9 times.
✓ Branch 151 → 152 taken 1 time.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 160 → 85 not taken.
✗ Branch 160 → 161 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 200 → 85 not taken.
✗ Branch 200 → 201 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 121 → 85 not taken.
✗ Branch 121 → 122 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 130 → 85 not taken.
✗ Branch 130 → 131 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 170 → 85 not taken.
✗ Branch 170 → 171 not taken.
|
48 | for (int y = 0; y < src_height; y++) |
| 535 | { | ||
| 536 | // int _y = (y & dither_mask) << dither_order; // for C version | ||
| 537 | 43 | current_matrix_line = matrix_simd16 + ((y & dither_mask) << 4); // always 16 byte boundary instead of dither order | |
| 538 | // load and convert corr to 16 bit | ||
| 539 | 43 | __m256i corr = _mm256_cvtepu8_epi16(_mm_load_si128(reinterpret_cast<const __m128i*>(current_matrix_line))); // int corr = matrix[_y | (x & dither_mask)]; | |
| 540 | |||
| 541 |
10/144void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 104 → 90 not taken.
✗ Branch 104 → 105 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 114 → 91 not taken.
✗ Branch 114 → 115 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 169 → 91 not taken.
✗ Branch 169 → 170 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 140 → 90 not taken.
✗ Branch 140 → 141 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 150 → 91 not taken.
✗ Branch 150 → 151 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 199 → 91 not taken.
✗ Branch 199 → 200 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 140 → 90 not taken.
✗ Branch 140 → 141 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 150 → 91 not taken.
✗ Branch 150 → 151 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 199 → 91 not taken.
✗ Branch 199 → 200 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 104 → 90 not taken.
✗ Branch 104 → 105 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 114 → 91 not taken.
✗ Branch 114 → 115 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 169 → 91 taken 10 times.
✓ Branch 169 → 170 taken 5 times.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 104 → 90 not taken.
✗ Branch 104 → 105 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 114 → 91 not taken.
✗ Branch 114 → 115 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 169 → 91 not taken.
✗ Branch 169 → 170 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 140 → 90 not taken.
✗ Branch 140 → 141 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 150 → 91 not taken.
✗ Branch 150 → 151 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 199 → 91 not taken.
✗ Branch 199 → 200 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 140 → 90 not taken.
✗ Branch 140 → 141 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 150 → 91 not taken.
✗ Branch 150 → 151 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 199 → 91 not taken.
✗ Branch 199 → 200 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 104 → 90 not taken.
✗ Branch 104 → 105 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 114 → 91 not taken.
✗ Branch 114 → 115 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 169 → 91 not taken.
✗ Branch 169 → 170 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 117 → 90 taken 14 times.
✓ Branch 117 → 118 taken 7 times.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 127 → 91 not taken.
✗ Branch 127 → 128 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 167 → 91 not taken.
✗ Branch 167 → 168 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 147 → 90 not taken.
✗ Branch 147 → 148 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 157 → 91 not taken.
✗ Branch 157 → 158 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 197 → 91 not taken.
✗ Branch 197 → 198 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 147 → 90 not taken.
✗ Branch 147 → 148 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 157 → 91 not taken.
✗ Branch 157 → 158 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 197 → 91 not taken.
✗ Branch 197 → 198 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 117 → 90 taken 10 times.
✓ Branch 117 → 118 taken 5 times.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 127 → 91 not taken.
✗ Branch 127 → 128 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 167 → 91 not taken.
✗ Branch 167 → 168 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 117 → 90 not taken.
✗ Branch 117 → 118 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 127 → 91 not taken.
✗ Branch 127 → 128 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 167 → 91 not taken.
✗ Branch 167 → 168 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 147 → 90 not taken.
✗ Branch 147 → 148 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 157 → 91 not taken.
✗ Branch 157 → 158 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 197 → 91 not taken.
✗ Branch 197 → 198 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 147 → 90 not taken.
✗ Branch 147 → 148 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 157 → 91 not taken.
✗ Branch 157 → 158 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 197 → 91 not taken.
✗ Branch 197 → 198 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 117 → 90 not taken.
✗ Branch 117 → 118 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 127 → 91 not taken.
✗ Branch 127 → 128 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 167 → 91 not taken.
✗ Branch 167 → 168 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 119 → 90 not taken.
✗ Branch 119 → 120 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 128 → 90 not taken.
✗ Branch 128 → 129 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 168 → 90 not taken.
✗ Branch 168 → 169 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 149 → 90 not taken.
✗ Branch 149 → 150 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 158 → 90 not taken.
✗ Branch 158 → 159 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 198 → 90 not taken.
✗ Branch 198 → 199 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 149 → 90 not taken.
✗ Branch 149 → 150 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 158 → 90 not taken.
✗ Branch 158 → 159 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 198 → 90 not taken.
✗ Branch 198 → 199 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 119 → 90 not taken.
✗ Branch 119 → 120 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 128 → 90 taken 34 times.
✓ Branch 128 → 129 taken 17 times.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 168 → 90 not taken.
✗ Branch 168 → 169 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 119 → 90 not taken.
✗ Branch 119 → 120 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 128 → 90 not taken.
✗ Branch 128 → 129 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 168 → 90 not taken.
✗ Branch 168 → 169 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 149 → 90 not taken.
✗ Branch 149 → 150 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 158 → 90 not taken.
✗ Branch 158 → 159 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 198 → 90 not taken.
✗ Branch 198 → 199 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 149 → 90 taken 18 times.
✓ Branch 149 → 150 taken 9 times.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 158 → 90 not taken.
✗ Branch 158 → 159 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 198 → 90 not taken.
✗ Branch 198 → 199 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 119 → 90 not taken.
✗ Branch 119 → 120 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 128 → 90 not taken.
✗ Branch 128 → 129 not taken.
void do_convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 168 → 90 not taken.
✗ Branch 168 → 169 not taken.
|
129 | for (int x = 0; x < src_width; x += 16) |
| 542 | { | ||
| 543 | //const int corr = matrix[_y | (x & dither_mask)]; | ||
| 544 | |||
| 545 | //int src_pixel = srcp[x]; | ||
| 546 | //__m128i src_lo, src_hi; | ||
| 547 | __m256i src; | ||
| 548 | if constexpr (sizeof(pixel_t_s) == 1) { | ||
| 549 | 30 | src = _mm256_cvtepu8_epi16(_mm_load_si128(reinterpret_cast<const __m128i*>(srcp + x))); // 16* uint8->uint16 | |
| 550 | } | ||
| 551 | else { | ||
| 552 | 134 | src= _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp + x)); // 16* uint16 | |
| 553 | } | ||
| 554 | |||
| 555 | [[maybe_unused]] __m256i src_lo; | ||
| 556 | [[maybe_unused]] __m256i src_hi; | ||
| 557 | if constexpr (fulls != fulld) { | ||
| 558 | // goint to 32 float | ||
| 559 | // const float val = (srcp[x] - src_offset) * mul_factor + dst_offset_plus_round; | ||
| 560 | 36 | const auto src_offset_simd = _mm256_set1_epi32(d.src_offset_i); | |
| 561 | 18 | const auto dst_offset_plus_round_simd = _mm256_set1_ps(dst_offset_plus_round); | |
| 562 | |||
| 563 | 54 | auto src_lo_ps = _mm256_cvtepi32_ps(_mm256_sub_epi32(_mm256_unpacklo_epi16(src, zero), src_offset_simd)); | |
| 564 | 54 | auto src_hi_ps = _mm256_cvtepi32_ps(_mm256_sub_epi32(_mm256_unpackhi_epi16(src, zero), src_offset_simd)); | |
| 565 | 36 | src_lo_ps = _mm256_add_ps(_mm256_mul_ps(src_lo_ps, mul_factor_simd), dst_offset_plus_round_simd); | |
| 566 | 36 | src_hi_ps = _mm256_add_ps(_mm256_mul_ps(src_hi_ps, mul_factor_simd), dst_offset_plus_round_simd); | |
| 567 | |||
| 568 | // src_pixel = clamp((int)val, src_pixel_min, src_pixel_max); | ||
| 569 | if constexpr (sizeof(pixel_t_s) == 1 && !TEMPLATE_LOW_DITHER_BITDEPTH) { | ||
| 570 | // Source is 8 bits. This branch is ineffective if we define TEMPLATE_LOW_DITHER_BITDEPTH for all <8 bits, as we do now | ||
| 571 | // We would reach this code for 5-7 bits when lo limit would be e.g. 4 bits | ||
| 572 | // 8 bit source: we can go back from 32 to 16 bit int | ||
| 573 | ✗ | src = _mm256_packus_epi32(_mm256_cvttps_epi32(src_lo_ps), _mm256_cvttps_epi32(src_hi_ps)); | |
| 574 | // min is 0, which was ensured by packus | ||
| 575 | ✗ | const auto src_pixel_max_simd = _mm256_set1_epi16(src_pixel_max); | |
| 576 | ✗ | src = _mm256_min_epu16(src, src_pixel_max_simd); | |
| 577 | // tmp result 16 bits in src | ||
| 578 | } | ||
| 579 | else { | ||
| 580 | // go back to int32 only | ||
| 581 | 18 | src_lo = _mm256_cvttps_epi32(src_lo_ps); | |
| 582 | 18 | src_hi = _mm256_cvttps_epi32(src_hi_ps); | |
| 583 | // min is 0, which was ensured by packus | ||
| 584 | 18 | const auto src_pixel_max_simd = _mm256_set1_epi32(src_pixel_max); | |
| 585 | 18 | src_lo = _mm256_min_epi32(src_lo, src_pixel_max_simd); | |
| 586 | 18 | src_hi = _mm256_min_epi32(src_hi, src_pixel_max_simd); | |
| 587 | // tmp result 32 bits in src_lo, src_hi | ||
| 588 | } | ||
| 589 | } | ||
| 590 | else { | ||
| 591 | // 10-16 bits or preshift: go into 32 bit int | ||
| 592 | if constexpr (sizeof(pixel_t_s) == 2 || TEMPLATE_LOW_DITHER_BITDEPTH) { | ||
| 593 | 68 | src_lo = _mm256_unpacklo_epi16(src, zero); | |
| 594 | 68 | src_hi = _mm256_unpackhi_epi16(src, zero); | |
| 595 | // tmp result 32 bits in src_lo, src_hi | ||
| 596 | } | ||
| 597 | } | ||
| 598 | |||
| 599 | __m256i new_pixel; | ||
| 600 | |||
| 601 | // scale down after adding dithering noise | ||
| 602 | // int new_pixel = ((src_pixel + corr) >> dither_bit_diff); | ||
| 603 | if constexpr (sizeof(pixel_t_s) == 1 && !TEMPLATE_LOW_DITHER_BITDEPTH) { | ||
| 604 | // when dither_target_bits is 1, then preshift==1 and dither_bit_diff==(7+1) would not fit in 16 bits | ||
| 605 | ✗ | new_pixel = _mm256_srai_epi16(_mm256_add_epi16(src, corr), dither_bit_diff); | |
| 606 | // new_pixel is ready with 16 bit data | ||
| 607 | } | ||
| 608 | else { | ||
| 609 | // source bits: >8. | ||
| 610 | // At 16 bit overflow can happen when 0xFFFF it dithered up. This is why we are at 32 bits int | ||
| 611 | // Theoretically fulls == fulld && <16 bit would allow 16 bit int workflow but we do not specialize for that | ||
| 612 | // source bits: >8 or 8 bit but PRESHIFT | ||
| 613 | // At 16 bit overflow can happen when 0xFFFF it dithered up. This is why we are at 32 bits int | ||
| 614 | // Theoretically fulls == fulld && <16 bit would allow 16 bit int workflow but we do not specialize for that | ||
| 615 | // low dither bitdepth is special as well | ||
| 616 | |||
| 617 | /* | ||
| 618 | int new_pixel; | ||
| 619 | if (TEMPLATE_LOW_DITHER_BITDEPTH) { | ||
| 620 | // accurate dither: +/- | ||
| 621 | const float corr_f = corr - half_maxcorr_value; | ||
| 622 | new_pixel = (int)(src_pixel + corr_f) >> dither_bit_diff; | ||
| 623 | } | ||
| 624 | else | ||
| 625 | new_pixel = ((src_pixel + corr) >> dither_bit_diff); | ||
| 626 | */ | ||
| 627 | __m256i new_pixel_lo, new_pixel_hi; | ||
| 628 | 86 | auto corr_lo = _mm256_unpacklo_epi16(corr, zero); | |
| 629 | 86 | auto corr_hi = _mm256_unpackhi_epi16(corr, zero); | |
| 630 | if (TEMPLATE_LOW_DITHER_BITDEPTH) { | ||
| 631 | // accurately positioned to the center | ||
| 632 | 20 | auto corr_lo_ps = _mm256_sub_ps(_mm256_cvtepi32_ps(corr_lo), half_maxcorr_value_simd); | |
| 633 | 20 | auto corr_hi_ps = _mm256_sub_ps(_mm256_cvtepi32_ps(corr_hi), half_maxcorr_value_simd); | |
| 634 | // accurate dither: +/- | ||
| 635 | 40 | new_pixel_lo = _mm256_srai_epi32(_mm256_cvttps_epi32(_mm256_add_ps(_mm256_cvtepi32_ps(src_lo), corr_lo_ps)), dither_bit_diff); | |
| 636 | 40 | new_pixel_hi = _mm256_srai_epi32(_mm256_cvttps_epi32(_mm256_add_ps(_mm256_cvtepi32_ps(src_hi), corr_hi_ps)), dither_bit_diff); | |
| 637 | } | ||
| 638 | else { | ||
| 639 | // Arrgh. If dither_bit_diff would be an immediate constant (e.g. template parameter) then we get 2390 fps vs 1378 even for sse4.1! | ||
| 640 | // for a simple Vide16bit.ConvertBits(8, dither=0, dither_bits=8) | ||
| 641 | // const auto dither_bit_diff = 8; | ||
| 642 | 152 | new_pixel_lo = _mm256_srai_epi32(_mm256_add_epi32(src_lo, corr_lo), dither_bit_diff); | |
| 643 | 152 | new_pixel_hi = _mm256_srai_epi32(_mm256_add_epi32(src_hi, corr_hi), dither_bit_diff); | |
| 644 | } | ||
| 645 | 86 | new_pixel = _mm256_packus_epi32(new_pixel_lo, new_pixel_hi); | |
| 646 | // new_pixel is ready with 16 bit data | ||
| 647 | } | ||
| 648 | |||
| 649 | // scale back to the required bit depth | ||
| 650 | // dither to x, target to y | ||
| 651 | if constexpr (TEMPLATE_NEED_BACKSCALE) { | ||
| 652 | /* | ||
| 653 | new_pixel = min(new_pixel, max_pixel_value_dithered); | ||
| 654 | if (TEMPLATE_LOW_DITHER_BITDEPTH) { | ||
| 655 | new_pixel = (int)(new_pixel * mul_factor_backfromlowdither + 0.5f); | ||
| 656 | } | ||
| 657 | else { | ||
| 658 | new_pixel = new_pixel << bitdiff_between_dither_and_target; | ||
| 659 | } | ||
| 660 | */ | ||
| 661 | 88 | const auto max_pixel_value_dithered_simd = _mm256_set1_epi16(max_pixel_value_dithered); | |
| 662 | 44 | new_pixel = _mm256_min_epu16(new_pixel, max_pixel_value_dithered_simd); | |
| 663 | // Interesting problem of dither_bits==1 (or in general at small dither_bits) | ||
| 664 | // After simple slli 0,1 becomes 0,128, we'd expect 0,255 instead. So we make cosmetics. | ||
| 665 | // dither_bits | ||
| 666 | // 1 0,1 => 0,128 => 0,255 | ||
| 667 | // 2 0,1,2,3 => 0,64,128,192 => 0,85,170,255 | ||
| 668 | // 3 0,...,7 => 0,32,...,224 => 0,....255 | ||
| 669 | // 4 0,..,15 => 0,16,...,240 => 0,....255 | ||
| 670 | // 5 0,..,31 => 0,8,....,248 => 0,....255 | ||
| 671 | // 6 0,..,63 => 0,4,....,252 => 0,....255 | ||
| 672 | // 7 0,.,127 => 0,2. ..,254 => 0,?,?,255 | ||
| 673 | if (TEMPLATE_LOW_DITHER_BITDEPTH) { | ||
| 674 | // new_pixel = (int)(new_pixel * mul_factor_backfromlowdither + 0.5f); | ||
| 675 | 10 | auto rounder_half_simd = _mm256_set1_ps(0.5f); | |
| 676 | |||
| 677 | 20 | auto new_pixel_lo_ps = _mm256_cvtepi32_ps(_mm256_unpacklo_epi16(new_pixel, zero)); | |
| 678 | 20 | auto new_pixel_hi_ps = _mm256_cvtepi32_ps(_mm256_unpackhi_epi16(new_pixel, zero)); | |
| 679 | 70 | new_pixel = _mm256_packus_epi32( | |
| 680 | _mm256_cvttps_epi32(_mm256_add_ps(_mm256_mul_ps(new_pixel_lo_ps, mul_factor_backfromlowdither_simd), rounder_half_simd)), | ||
| 681 | _mm256_cvttps_epi32(_mm256_add_ps(_mm256_mul_ps(new_pixel_hi_ps, mul_factor_backfromlowdither_simd), rounder_half_simd)) | ||
| 682 | ); | ||
| 683 | } | ||
| 684 | else { | ||
| 685 | 34 | new_pixel = _mm256_slli_epi16(new_pixel, bitdiff_between_dither_and_target); | |
| 686 | } | ||
| 687 | } | ||
| 688 | |||
| 689 | // dstp[x] = (pixel_t_d)(min((int)new_pixel, max_pixel_value_target)); | ||
| 690 | if constexpr (sizeof(pixel_t_d) == 1) | ||
| 691 | { | ||
| 692 | // byte target: 16x 2bytes -> 16x 1 bytes like cvtepu16->epu8 | ||
| 693 | 34 | auto result12 = _mm256_packus_epi16(new_pixel, new_pixel); | |
| 694 | 34 | result12 = _mm256_permute4x64_epi64(result12, (0 << 0) | (2 << 2) | (1 << 4) | (3 << 6)); | |
| 695 | 34 | _mm_store_si128(reinterpret_cast<__m128i*>(dstp + x), _mm256_castsi256_si128(result12)); | |
| 696 | } | ||
| 697 | else { | ||
| 698 | 104 | const auto max_pixel_value_target_simd = _mm256_set1_epi16(max_pixel_value_target); | |
| 699 | // max is unnecessary for exact 16 bit pixel type but we are not specialized for that case | ||
| 700 | 52 | new_pixel = _mm256_min_epu16(new_pixel, max_pixel_value_target_simd); | |
| 701 | 52 | _mm256_store_si256(reinterpret_cast<__m256i*>(dstp + x), new_pixel); | |
| 702 | } | ||
| 703 | } | ||
| 704 | 43 | dstp += dst_pitch; | |
| 705 | 43 | srcp += src_pitch; | |
| 706 | } | ||
| 707 | } | ||
| 708 | |||
| 709 | template<typename pixel_t_s, typename pixel_t_d, bool chroma, bool fulls, bool fulld> | ||
| 710 | 5 | void convert_ordered_dither_uint_avx2(const BYTE* srcp8, BYTE* dstp8, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth) | |
| 711 | { | ||
| 712 | 5 | const bool need_backscale = target_bitdepth != dither_target_bitdepth; // dither to x, target to y | |
| 713 | 5 | const bool low_dither_bitdepth = dither_target_bitdepth < 8; // 1-7 bits dither targets, need_backscale is always true, since 8 bit format is the minimum | |
| 714 |
5/48void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 6 taken 1 time.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 6 taken 1 time.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 6 taken 1 time.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 6 not taken.
|
5 | if (need_backscale) { |
| 715 |
2/48void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 1 time.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned char, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✓ Branch 3 → 5 taken 1 time.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
void convert_ordered_dither_uint_avx2<unsigned short, unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 5 not taken.
|
2 | if (low_dither_bitdepth) |
| 716 | 1 | do_convert_ordered_dither_uint_avx2<pixel_t_s, pixel_t_d, chroma, fulls, fulld, true, true>(srcp8, dstp8, src_rowsize, src_height, src_pitch, dst_pitch, source_bitdepth, target_bitdepth, dither_target_bitdepth); | |
| 717 | else | ||
| 718 | 1 | do_convert_ordered_dither_uint_avx2<pixel_t_s, pixel_t_d, chroma, fulls, fulld, true, false>(srcp8, dstp8, src_rowsize, src_height, src_pitch, dst_pitch, source_bitdepth, target_bitdepth, dither_target_bitdepth); | |
| 719 | } | ||
| 720 | else { | ||
| 721 | 3 | do_convert_ordered_dither_uint_avx2<pixel_t_s, pixel_t_d, chroma, fulls, fulld, false, false>(srcp8, dstp8, src_rowsize, src_height, src_pitch, dst_pitch, source_bitdepth, target_bitdepth, dither_target_bitdepth); | |
| 722 | } | ||
| 723 | 5 | } | |
| 724 | |||
| 725 | // instantiate them all | ||
| 726 | // template<typename pixel_t_s, typename pixel_t_d, bool chroma, bool fulls, bool fulld> | ||
| 727 | // spec: fulld=false, fulls=false | ||
| 728 | #define convert_ordered_dither_uint_sse4_functions(uint_X_t, uint_X_dest_t) \ | ||
| 729 | template void convert_ordered_dither_uint_avx2<uint_X_t, uint_X_dest_t, false, false, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 730 | template void convert_ordered_dither_uint_avx2<uint_X_t, uint_X_dest_t, false, false, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 731 | template void convert_ordered_dither_uint_avx2<uint_X_t, uint_X_dest_t, false, true, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 732 | template void convert_ordered_dither_uint_avx2<uint_X_t, uint_X_dest_t, false, true, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 733 | template void convert_ordered_dither_uint_avx2<uint_X_t, uint_X_dest_t, true, false, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 734 | template void convert_ordered_dither_uint_avx2<uint_X_t, uint_X_dest_t, true, false, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 735 | template void convert_ordered_dither_uint_avx2<uint_X_t, uint_X_dest_t, true, true, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); \ | ||
| 736 | template void convert_ordered_dither_uint_avx2<uint_X_t, uint_X_dest_t, true, true, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 737 | |||
| 738 | // dither only same of less bit depth | ||
| 739 | convert_ordered_dither_uint_sse4_functions(uint8_t, uint8_t) | ||
| 740 | convert_ordered_dither_uint_sse4_functions(uint16_t, uint8_t) | ||
| 741 | convert_ordered_dither_uint_sse4_functions(uint16_t, uint16_t) | ||
| 742 | |||
| 743 | #undef convert_ordered_dither_uint_sse4_functions | ||
| 744 | |||
| 745 | template<typename pixel_t, bool chroma, bool fulls, bool fulld> | ||
| 746 | 3 | void convert_uintN_to_float_avx2(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth) | |
| 747 | { | ||
| 748 | 3 | const pixel_t* srcp0 = reinterpret_cast<const pixel_t*>(srcp); | |
| 749 | 3 | float* dstp0 = reinterpret_cast<float*>(dstp); | |
| 750 | |||
| 751 | 3 | src_pitch = src_pitch / sizeof(pixel_t); | |
| 752 | 3 | dst_pitch = dst_pitch / sizeof(float); | |
| 753 | |||
| 754 | 3 | const int src_width = src_rowsize / sizeof(pixel_t); | |
| 755 | // write 32 floats (128 bytes) only if valid, scanline alignment is only 64 bytes (16 floats). | ||
| 756 | // Alignment, write-limited, 16 floats (64 bytes) is guaranteed by AviSynth | ||
| 757 | 3 | const int w16 = (src_width + 15) & ~15; | |
| 758 | //----------------------- | ||
| 759 |
3/32void convert_uintN_to_float_avx2<unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 1 time.
void convert_uintN_to_float_avx2<unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 24 taken 1 time.
void convert_uintN_to_float_avx2<unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 24 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 24 not taken.
|
3 | bits_conv_constants d; |
| 760 | get_bits_conv_constants(d, chroma, fulls, fulld, source_bitdepth, target_bitdepth); | ||
| 761 | |||
| 762 | 3 | const __m256i m256_src_offset_epi32 = _mm256_set1_epi32(d.src_offset_i); | |
| 763 | 3 | const __m256 m256_mul_factor = _mm256_set1_ps(d.mul_factor); | |
| 764 | 3 | const __m256 m256_dst_offset = _mm256_set1_ps(d.dst_offset); | |
| 765 | |||
| 766 |
6/32void convert_uintN_to_float_avx2<unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 133 → 61 taken 5 times.
✓ Branch 133 → 134 taken 1 time.
void convert_uintN_to_float_avx2<unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 133 → 61 taken 4 times.
✓ Branch 133 → 134 taken 1 time.
void convert_uintN_to_float_avx2<unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 133 → 61 taken 5 times.
✓ Branch 133 → 134 taken 1 time.
void convert_uintN_to_float_avx2<unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 133 → 61 not taken.
✗ Branch 133 → 134 not taken.
|
17 | for (int y = 0; y < src_height; y++) |
| 767 | { | ||
| 768 | // rows are 64 bytes - 16 float pixels - aligned we have a write constraint | ||
| 769 | int x; | ||
| 770 |
6/32void convert_uintN_to_float_avx2<unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 109 → 62 not taken.
✗ Branch 109 → 110 not taken.
void convert_uintN_to_float_avx2<unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 109 → 62 taken 5 times.
✓ Branch 109 → 110 taken 5 times.
void convert_uintN_to_float_avx2<unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 109 → 62 not taken.
✗ Branch 109 → 110 not taken.
void convert_uintN_to_float_avx2<unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 109 → 62 not taken.
✗ Branch 109 → 110 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 109 → 62 not taken.
✗ Branch 109 → 110 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 109 → 62 not taken.
✗ Branch 109 → 110 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 109 → 62 not taken.
✗ Branch 109 → 110 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 109 → 62 not taken.
✗ Branch 109 → 110 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 107 → 62 not taken.
✗ Branch 107 → 108 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 107 → 62 taken 4 times.
✓ Branch 107 → 108 taken 4 times.
void convert_uintN_to_float_avx2<unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 107 → 62 not taken.
✗ Branch 107 → 108 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 107 → 62 not taken.
✗ Branch 107 → 108 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 107 → 62 not taken.
✗ Branch 107 → 108 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 107 → 62 not taken.
✗ Branch 107 → 108 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✓ Branch 107 → 62 taken 5 times.
✓ Branch 107 → 108 taken 5 times.
void convert_uintN_to_float_avx2<unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 107 → 62 not taken.
✗ Branch 107 → 108 not taken.
|
28 | for (x = 0; x + 32 <= w16; x += 32) |
| 771 | { | ||
| 772 | __m256i src_0_32, src_1_32, src_2_32, src_3_32; | ||
| 773 | |||
| 774 | if constexpr (sizeof(pixel_t) == 1) // uint8_t | ||
| 775 | { | ||
| 776 | 10 | __m256i src_32 = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x)); | |
| 777 | // unpack to 4x64bits | ||
| 778 | 5 | src_0_32 = _mm256_cvtepu8_epi32(_mm256_castsi256_si128(src_32)); // bytes 0-7 | |
| 779 | |||
| 780 | 5 | __m256i src_1 = _mm256_permute4x64_epi64(src_32, 1); | |
| 781 | 5 | __m256i src_2 = _mm256_permute4x64_epi64(src_32, 2); | |
| 782 | 5 | __m256i src_3 = _mm256_permute4x64_epi64(src_32, 3); | |
| 783 | |||
| 784 | 10 | src_1_32 = _mm256_cvtepu8_epi32(_mm256_castsi256_si128(src_1)); | |
| 785 | 10 | src_2_32 = _mm256_cvtepu8_epi32(_mm256_castsi256_si128(src_2)); | |
| 786 | 10 | src_3_32 = _mm256_cvtepu8_epi32(_mm256_castsi256_si128(src_3)); | |
| 787 | } | ||
| 788 | |||
| 789 | if constexpr (sizeof(pixel_t) == 2) // uint16_t | ||
| 790 | { | ||
| 791 | 9 | __m256i src_16 = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x)); | |
| 792 | 18 | __m256i src_16_2 = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x + 16)); | |
| 793 | // unpack to 2x128bits 2x8 uint16 | ||
| 794 | 9 | src_0_32 = _mm256_cvtepu16_epi32(_mm256_castsi256_si128(src_16)); | |
| 795 | 18 | src_1_32 = _mm256_cvtepu16_epi32(_mm256_extracti128_si256(src_16, 1)); | |
| 796 | 9 | src_2_32 = _mm256_cvtepu16_epi32(_mm256_castsi256_si128(src_16_2)); | |
| 797 | 18 | src_3_32 = _mm256_cvtepu16_epi32(_mm256_extracti128_si256(src_16_2, 1)); | |
| 798 | } | ||
| 799 | |||
| 800 | 14 | src_0_32 = _mm256_sub_epi32(src_0_32, m256_src_offset_epi32); | |
| 801 | 14 | src_1_32 = _mm256_sub_epi32(src_1_32, m256_src_offset_epi32); | |
| 802 | 14 | src_2_32 = _mm256_sub_epi32(src_2_32, m256_src_offset_epi32); | |
| 803 | 14 | src_3_32 = _mm256_sub_epi32(src_3_32, m256_src_offset_epi32); | |
| 804 | |||
| 805 | 14 | __m256 src_0_ps = _mm256_cvtepi32_ps(src_0_32); | |
| 806 | 14 | __m256 src_1_ps = _mm256_cvtepi32_ps(src_1_32); | |
| 807 | 14 | __m256 src_2_ps = _mm256_cvtepi32_ps(src_2_32); | |
| 808 | 14 | __m256 src_3_ps = _mm256_cvtepi32_ps(src_3_32); | |
| 809 | |||
| 810 | 14 | __m256 out_0_ps = _mm256_fmadd_ps(src_0_ps, m256_mul_factor, m256_dst_offset); | |
| 811 | 14 | __m256 out_1_ps = _mm256_fmadd_ps(src_1_ps, m256_mul_factor, m256_dst_offset); | |
| 812 | 14 | __m256 out_2_ps = _mm256_fmadd_ps(src_2_ps, m256_mul_factor, m256_dst_offset); | |
| 813 | 14 | __m256 out_3_ps = _mm256_fmadd_ps(src_3_ps, m256_mul_factor, m256_dst_offset); | |
| 814 | |||
| 815 | // process 32 pixels, write 128 bytes | ||
| 816 | 14 | _mm256_store_ps((dstp0 + x + 0), out_0_ps); | |
| 817 | 14 | _mm256_store_ps((dstp0 + x + 8), out_1_ps); | |
| 818 | 14 | _mm256_store_ps((dstp0 + x + 16), out_2_ps); | |
| 819 | 14 | _mm256_store_ps((dstp0 + x + 24), out_3_ps); | |
| 820 | |||
| 821 | // const float pixel = (srcp0[x] - d.src_offset_i) * d.mul_factor + d.dst_offset; | ||
| 822 | // dstp0[x] = pixel; // no clamp | ||
| 823 | } | ||
| 824 | |||
| 825 | // second loop: process remaining 1-15 pixels, 16 at a time | ||
| 826 |
3/32void convert_uintN_to_float_avx2<unsigned char, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 110 → 111 not taken.
✗ Branch 110 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned char, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 110 → 111 not taken.
✓ Branch 110 → 132 taken 5 times.
void convert_uintN_to_float_avx2<unsigned char, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 110 → 111 not taken.
✗ Branch 110 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned char, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 110 → 111 not taken.
✗ Branch 110 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 110 → 111 not taken.
✗ Branch 110 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 110 → 111 not taken.
✗ Branch 110 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 110 → 111 not taken.
✗ Branch 110 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned char, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 110 → 111 not taken.
✗ Branch 110 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 108 → 109 not taken.
✗ Branch 108 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 108 → 109 not taken.
✓ Branch 108 → 132 taken 4 times.
void convert_uintN_to_float_avx2<unsigned short, false, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 108 → 109 not taken.
✗ Branch 108 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned short, false, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 108 → 109 not taken.
✗ Branch 108 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, false, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 108 → 109 not taken.
✗ Branch 108 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, false, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 108 → 109 not taken.
✗ Branch 108 → 132 not taken.
void convert_uintN_to_float_avx2<unsigned short, true, true, false>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 108 → 109 not taken.
✓ Branch 108 → 132 taken 5 times.
void convert_uintN_to_float_avx2<unsigned short, true, true, true>(unsigned char const*, unsigned char*, int, int, int, int, int, int, int):
✗ Branch 108 → 109 not taken.
✗ Branch 108 → 132 not taken.
|
14 | if (x < w16) { |
| 827 | __m256i src_0_32, src_1_32; | ||
| 828 | |||
| 829 | if constexpr (sizeof(pixel_t) == 1) // uint8_t | ||
| 830 | { | ||
| 831 | // _m128 enough, 16 pixels | ||
| 832 | ✗ | __m128i src_16_bytes = _mm_load_si128(reinterpret_cast<const __m128i*>(srcp0 + x)); | |
| 833 | ✗ | src_0_32 = _mm256_cvtepu8_epi32(src_16_bytes); // 0-7 | |
| 834 | ✗ | src_1_32 = _mm256_cvtepu8_epi32(_mm_srli_si128(src_16_bytes, 8)); // 8-15, quick permuteless | |
| 835 | } | ||
| 836 | else // uint16_t | ||
| 837 | { | ||
| 838 | ✗ | __m256i src_16 = _mm256_load_si256(reinterpret_cast<const __m256i*>(srcp0 + x)); | |
| 839 | ✗ | src_0_32 = _mm256_cvtepu16_epi32(_mm256_castsi256_si128(src_16)); // 0-7 | |
| 840 | ✗ | src_1_32 = _mm256_cvtepu16_epi32(_mm256_extracti128_si256(src_16, 1)); // 8-15 | |
| 841 | } | ||
| 842 | |||
| 843 | ✗ | src_0_32 = _mm256_sub_epi32(src_0_32, m256_src_offset_epi32); | |
| 844 | ✗ | src_1_32 = _mm256_sub_epi32(src_1_32, m256_src_offset_epi32); | |
| 845 | |||
| 846 | ✗ | __m256 out_0_ps = _mm256_fmadd_ps(_mm256_cvtepi32_ps(src_0_32), m256_mul_factor, m256_dst_offset); | |
| 847 | ✗ | __m256 out_1_ps = _mm256_fmadd_ps(_mm256_cvtepi32_ps(src_1_32), m256_mul_factor, m256_dst_offset); | |
| 848 | |||
| 849 | // 16 floats - 64 bytes always safe | ||
| 850 | ✗ | _mm256_store_ps((dstp0 + x + 0), out_0_ps); | |
| 851 | ✗ | _mm256_store_ps((dstp0 + x + 8), out_1_ps); | |
| 852 | |||
| 853 | // const float pixel = (srcp0[x] - d.src_offset_i) * d.mul_factor + d.dst_offset; | ||
| 854 | // dstp0[x] = pixel; // no clamp | ||
| 855 | } | ||
| 856 | 14 | dstp0 += dst_pitch; | |
| 857 | 14 | srcp0 += src_pitch; | |
| 858 | } | ||
| 859 | |||
| 860 | 3 | } | |
| 861 | |||
| 862 | // instantinate | ||
| 863 | template void convert_uintN_to_float_avx2<uint8_t, false, false, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 864 | template void convert_uintN_to_float_avx2<uint8_t, false, false, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 865 | template void convert_uintN_to_float_avx2<uint8_t, false, true, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 866 | template void convert_uintN_to_float_avx2<uint8_t, false, true, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 867 | template void convert_uintN_to_float_avx2<uint8_t, true, false, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 868 | template void convert_uintN_to_float_avx2<uint8_t, true, false, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 869 | template void convert_uintN_to_float_avx2<uint8_t, true, true, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 870 | template void convert_uintN_to_float_avx2<uint8_t, true, true, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 871 | |||
| 872 | template void convert_uintN_to_float_avx2<uint16_t, false, false, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 873 | template void convert_uintN_to_float_avx2<uint16_t, false, false, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 874 | template void convert_uintN_to_float_avx2<uint16_t, false, true, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 875 | template void convert_uintN_to_float_avx2<uint16_t, false, true, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 876 | template void convert_uintN_to_float_avx2<uint16_t, true, false, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 877 | template void convert_uintN_to_float_avx2<uint16_t, true, false, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 878 | template void convert_uintN_to_float_avx2<uint16_t, true, true, false>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 879 | template void convert_uintN_to_float_avx2<uint16_t, true, true, true>(const BYTE* srcp, BYTE* dstp, int src_rowsize, int src_height, int src_pitch, int dst_pitch, int source_bitdepth, int target_bitdepth, int dither_target_bitdepth); | ||
| 880 | |||
| 881 |