filters/overlay/intel/masked_merge_avx2_impl.hpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // AviSynth+ Copyright 2026- AviSynth+ Project | ||
| 2 | // SPDX-License-Identifier: GPL-2.0-or-later | ||
| 3 | // | ||
| 4 | // AVX2 masked-merge implementation templates. | ||
| 5 | // | ||
| 6 | // Implementation-include — NO include guards intentionally. | ||
| 7 | // Each including TU gets its own compilation with that TU's SIMD flags. | ||
| 8 | // All definitions are static so each TU has its own private copy; | ||
| 9 | // the linker never sees them as the same symbol. | ||
| 10 | // | ||
| 11 | // Requires the following to be visible in the including TU before this file: | ||
| 12 | // AVX2 intrinsics (<intrin.h> or <x86intrin.h>, <immintrin.h>) | ||
| 13 | // blend_common.h — MaskMode, MagicDiv, get_magic_div, magic_div_rt, | ||
| 14 | // prepare_effective_mask_for_row | ||
| 15 | // <vector> | ||
| 16 | |||
| 17 | // Provides simd_magic_div_32_avx2 + prepare_effective_mask_for_row_avx2. | ||
| 18 | #include "masked_rowprep_avx2_impl.h" | ||
| 19 | |||
| 20 | // --------------------------------------------------------------------------- | ||
| 21 | // 8-bit row — mask already has opacity baked in by rowprep. | ||
| 22 | // 32-wide 16-bit mulhi arithmetic (fast, overflow-safe). | ||
| 23 | // --------------------------------------------------------------------------- | ||
| 24 | #if defined(GCC) || defined(CLANG) | ||
| 25 | __attribute__((__target__("avx2"))) | ||
| 26 | #endif | ||
| 27 | 244 | static void blend8_masked_avx2_row( | |
| 28 | uint8_t* p1, const uint8_t* p2, const uint8_t* mask, | ||
| 29 | int width) | ||
| 30 | { | ||
| 31 | 244 | constexpr uint32_t half = 127u; | |
| 32 | 244 | constexpr uint32_t max_val = 255u; | |
| 33 | |||
| 34 | 244 | const __m256i v_half = _mm256_set1_epi16((short)half); | |
| 35 | 244 | const __m256i v_max = _mm256_set1_epi16((short)max_val); | |
| 36 | 244 | const __m256i v_magic = _mm256_set1_epi16((short)0x8081); | |
| 37 | |||
| 38 | 138 | auto magic_byte = [&](__m256i x) { | |
| 39 | 414 | return _mm256_srli_epi16(_mm256_mulhi_epu16(x, v_magic), 7); | |
| 40 | 244 | }; | |
| 41 | |||
| 42 | 244 | int x = 0; | |
| 43 |
2/2✓ Branch 45 → 15 taken 69 times.
✓ Branch 45 → 46 taken 244 times.
|
313 | for (; x <= width - 32; x += 32) { |
| 44 | 138 | __m256i mr = _mm256_loadu_si256((__m256i*)(mask + x)); | |
| 45 | |||
| 46 | // Unpack 32 bytes to 2x 16x16-bit halves | ||
| 47 | 69 | __m256i m_lo = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(mr)); | |
| 48 | 69 | __m256i m_hi = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(mr, 1)); | |
| 49 | |||
| 50 | 69 | __m256i ar = _mm256_loadu_si256((__m256i*)(p1 + x)); | |
| 51 | 69 | __m256i br = _mm256_loadu_si256((__m256i*)(p2 + x)); | |
| 52 | |||
| 53 | 138 | auto blend16 = [&](__m256i a, __m256i b, __m256i m) { | |
| 54 | 138 | __m256i invm = _mm256_sub_epi16(v_max, m); | |
| 55 | 690 | return magic_byte(_mm256_add_epi16( | |
| 56 | 138 | _mm256_add_epi16(_mm256_mullo_epi16(a, invm), _mm256_mullo_epi16(b, m)), v_half)); | |
| 57 | 69 | }; | |
| 58 | |||
| 59 | 276 | __m256i r_lo = blend16(_mm256_cvtepu8_epi16(_mm256_castsi256_si128(ar)), | |
| 60 | _mm256_cvtepu8_epi16(_mm256_castsi256_si128(br)), m_lo); | ||
| 61 | 207 | __m256i r_hi = blend16(_mm256_cvtepu8_epi16(_mm256_extracti128_si256(ar, 1)), | |
| 62 | _mm256_cvtepu8_epi16(_mm256_extracti128_si256(br, 1)), m_hi); | ||
| 63 | |||
| 64 | 69 | __m256i packed = _mm256_packus_epi16(r_lo, r_hi); | |
| 65 | 69 | packed = _mm256_permute4x64_epi64(packed, _MM_SHUFFLE(3, 1, 2, 0)); | |
| 66 | 69 | _mm256_storeu_si256((__m256i*)(p1 + x), packed); | |
| 67 | } | ||
| 68 |
2/2✓ Branch 48 → 47 taken 2572 times.
✓ Branch 48 → 49 taken 244 times.
|
2816 | for (; x < width; ++x) { |
| 69 | 2572 | const uint32_t ms = mask[x]; | |
| 70 | 2572 | const uint32_t a = p1[x], b_v = p2[x]; | |
| 71 | 2572 | const uint32_t tr = a * (max_val - ms) + b_v * ms + half; | |
| 72 | 2572 | p1[x] = (uint8_t)((tr * 0x8081u) >> 23); | |
| 73 | } | ||
| 74 | 244 | } | |
| 75 | |||
| 76 | // --------------------------------------------------------------------------- | ||
| 77 | // 16-bit row (10, 12, 14, 16 bits) — mask already has opacity baked in. | ||
| 78 | // 32-bit arithmetic, 8 pixels per step. | ||
| 79 | // --------------------------------------------------------------------------- | ||
| 80 | template<int bits_per_pixel> | ||
| 81 | #if defined(GCC) || defined(CLANG) | ||
| 82 | __attribute__((__target__("avx2"))) | ||
| 83 | #endif | ||
| 84 | 186 | static void blend16_masked_avx2_row( | |
| 85 | uint16_t* p1, const uint16_t* p2, const uint16_t* mask, | ||
| 86 | int width) | ||
| 87 | { | ||
| 88 | 186 | constexpr MagicDiv m_div = get_magic_div(bits_per_pixel); | |
| 89 | 186 | constexpr uint32_t max_val = (1u << bits_per_pixel) - 1; | |
| 90 | 186 | constexpr uint32_t half = max_val / 2; | |
| 91 | |||
| 92 | 186 | const __m256i v_half = _mm256_set1_epi32((int)half); | |
| 93 | 186 | const __m256i v_max = _mm256_set1_epi32((int)max_val); | |
| 94 | |||
| 95 | 186 | int x = 0; | |
| 96 |
8/8void blend16_masked_avx2_row<10>(unsigned short*, unsigned short const*, unsigned short const*, int):
✓ Branch 53 → 7 taken 20 times.
✓ Branch 53 → 54 taken 10 times.
void blend16_masked_avx2_row<12>(unsigned short*, unsigned short const*, unsigned short const*, int):
✓ Branch 53 → 7 taken 20 times.
✓ Branch 53 → 54 taken 10 times.
void blend16_masked_avx2_row<14>(unsigned short*, unsigned short const*, unsigned short const*, int):
✓ Branch 53 → 7 taken 20 times.
✓ Branch 53 → 54 taken 10 times.
void blend16_masked_avx2_row<16>(unsigned short*, unsigned short const*, unsigned short const*, int):
✓ Branch 53 → 7 taken 242 times.
✓ Branch 53 → 54 taken 156 times.
|
488 | for (; x <= width - 8; x += 8) { |
| 97 | 604 | __m256i m = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)(mask + x))); | |
| 98 | 604 | __m256i a = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)(p1 + x))); | |
| 99 | 906 | __m256i b = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)(p2 + x))); | |
| 100 | 302 | __m256i invm = _mm256_sub_epi32(v_max, m); | |
| 101 | 1510 | __m256i res = simd_magic_div_32_avx2( | |
| 102 | _mm256_add_epi32( | ||
| 103 | _mm256_add_epi32(_mm256_mullo_epi32(a, invm), _mm256_mullo_epi32(b, m)), | ||
| 104 | v_half), | ||
| 105 | 302 | m_div.div, m_div.shift); | |
| 106 | |||
| 107 | // Pack 8x32 → 8x16 with correct order | ||
| 108 | 604 | __m256i packed = _mm256_packus_epi32(res, _mm256_setzero_si256()); | |
| 109 | 302 | __m128i lo = _mm256_castsi256_si128(packed); | |
| 110 | 302 | __m128i hi = _mm256_extracti128_si256(packed, 1); | |
| 111 | 302 | _mm_storeu_si128((__m128i*)(p1 + x), _mm_unpacklo_epi64(lo, hi)); | |
| 112 | } | ||
| 113 |
8/8void blend16_masked_avx2_row<10>(unsigned short*, unsigned short const*, unsigned short const*, int):
✓ Branch 58 → 55 taken 30 times.
✓ Branch 58 → 59 taken 10 times.
void blend16_masked_avx2_row<12>(unsigned short*, unsigned short const*, unsigned short const*, int):
✓ Branch 58 → 55 taken 30 times.
✓ Branch 58 → 59 taken 10 times.
void blend16_masked_avx2_row<14>(unsigned short*, unsigned short const*, unsigned short const*, int):
✓ Branch 58 → 55 taken 30 times.
✓ Branch 58 → 59 taken 10 times.
void blend16_masked_avx2_row<16>(unsigned short*, unsigned short const*, unsigned short const*, int):
✓ Branch 58 → 55 taken 288 times.
✓ Branch 58 → 59 taken 156 times.
|
564 | for (; x < width; ++x) { |
| 114 | 378 | const uint32_t ms = mask[x]; | |
| 115 | 378 | const uint32_t a = p1[x]; | |
| 116 | 756 | p1[x] = (uint16_t)magic_div_rt<uint16_t>(a * (max_val - ms) + (uint32_t)p2[x] * ms + half, m_div); | |
| 117 | } | ||
| 118 | 186 | } | |
| 119 | |||
| 120 | // --------------------------------------------------------------------------- | ||
| 121 | // float row — mask already has opacity baked in (range 0.0 to 1.0). | ||
| 122 | // Linear interpolation: p1[x] = p1[x] * (1.0f - mask[x]) + p2[x] * mask[x] | ||
| 123 | // 8 pixels per step. | ||
| 124 | // --------------------------------------------------------------------------- | ||
| 125 | #if defined(GCC) || defined(CLANG) | ||
| 126 | __attribute__((__target__("avx2"))) | ||
| 127 | #endif | ||
| 128 | 30 | static void blend_masked_float_avx2_row( | |
| 129 | float* p1, const float* p2, const float* mask, int width) | ||
| 130 | { | ||
| 131 | 30 | int x = 0; | |
| 132 | // const __m256 v_one = _mm256_set1_ps(1.0f); in case of 1-x implementation | ||
| 133 | |||
| 134 |
2/2✓ Branch 17 → 3 taken 50 times.
✓ Branch 17 → 18 taken 30 times.
|
80 | for (; x <= width - 8; x += 8) { |
| 135 | 50 | __m256 m = _mm256_loadu_ps(mask + x); | |
| 136 | 50 | __m256 a = _mm256_loadu_ps(p1 + x); | |
| 137 | 100 | __m256 b = _mm256_loadu_ps(p2 + x); | |
| 138 | |||
| 139 | // Standard lerp: a + m * (b - a) | ||
| 140 | // This is generally more accurate and faster (1 mul, 2 adds) than | ||
| 141 | // a * (1-m) + b * m (2 muls, 2 adds). | ||
| 142 | 50 | __m256 diff = _mm256_sub_ps(b, a); | |
| 143 | 50 | __m256 res = _mm256_add_ps(a, _mm256_mul_ps(m, diff)); | |
| 144 | |||
| 145 | 50 | _mm256_storeu_ps(p1 + x, res); | |
| 146 | } | ||
| 147 | |||
| 148 | // Scalar tail | ||
| 149 |
2/2✓ Branch 20 → 19 taken 70 times.
✓ Branch 20 → 21 taken 30 times.
|
100 | for (; x < width; ++x) { |
| 150 | 70 | const float m = mask[x]; | |
| 151 | 70 | const float a = p1[x]; | |
| 152 | 70 | const float b = p2[x]; | |
| 153 | 70 | p1[x] = a + m * (b - a); | |
| 154 | } | ||
| 155 | 30 | } | |
| 156 | |||
| 157 | // --------------------------------------------------------------------------- | ||
| 158 | // Inner loop — full_opacity known at compile time. | ||
| 159 | // Rowprep bakes opacity when !full_opacity; blend row receives pre-scaled mask. | ||
| 160 | // MASK444 + full_opacity: returns mask ptr directly (no buffer, no copy). | ||
| 161 | // MASK444 + !full_opacity: copies row with opacity scaling into buffer. | ||
| 162 | // --------------------------------------------------------------------------- | ||
| 163 | template<MaskMode maskMode, bool full_opacity> | ||
| 164 | #if defined(GCC) || defined(CLANG) | ||
| 165 | __attribute__((__target__("avx2"))) | ||
| 166 | #endif | ||
| 167 | 117 | static void masked_merge_avx2_impl_inner( | |
| 168 | BYTE* p1, const BYTE* p2, const BYTE* mask, | ||
| 169 | int p1_pitch, int p2_pitch, int mask_pitch, | ||
| 170 | int width, int height, int opacity, int bits_per_pixel) | ||
| 171 | { | ||
| 172 | 117 | const MagicDiv mag = get_magic_div(bits_per_pixel); | |
| 173 | 117 | const int max_val = (1 << bits_per_pixel) - 1; | |
| 174 | 117 | const int half = max_val / 2; | |
| 175 | |||
| 176 |
28/32void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 13 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 13 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 5 times.
✓ Branch 3 → 13 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 2 times.
✓ Branch 3 → 13 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 2 times.
✓ Branch 3 → 13 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 1 time.
✓ Branch 3 → 13 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 1 time.
✓ Branch 3 → 13 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 1 time.
✓ Branch 3 → 13 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 2 times.
✓ Branch 3 → 13 taken 3 times.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 2 times.
✓ Branch 3 → 13 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 1 time.
✓ Branch 3 → 13 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 1 time.
✓ Branch 3 → 13 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 1 time.
✓ Branch 3 → 13 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 1 time.
✓ Branch 3 → 13 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 26 times.
✓ Branch 3 → 13 taken 15 times.
void masked_merge_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 3 → 4 taken 27 times.
✓ Branch 3 → 12 taken 12 times.
|
117 | if (bits_per_pixel == 8) { |
| 177 | 73 | const uint8_t* maskp = reinterpret_cast<const uint8_t*>(mask); | |
| 178 | 73 | const int mpx = mask_pitch; | |
| 179 | 73 | const int mask_adv = (maskMode == MASK420 || maskMode == MASK420_MPEG2 || maskMode == MASK420_TOPLEFT) ? mpx * 2 : mpx; | |
| 180 | |||
| 181 | 73 | std::vector<uint8_t> eff_buf; | |
| 182 |
13/30void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 5 times.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 2 times.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 2 times.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 1 time.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 1 time.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 1 time.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 2 times.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 2 times.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 1 time.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 1 time.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 1 time.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 1 time.
✗ Branch 5 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 5 → 6 taken 26 times.
✗ Branch 5 → 43 not taken.
|
46 | if constexpr (maskMode != MASK444 || !full_opacity) eff_buf.resize(width); |
| 183 | |||
| 184 |
28/32void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 10 → 7 not taken.
✗ Branch 10 → 11 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 10 → 7 not taken.
✗ Branch 10 → 11 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 19 times.
✓ Branch 10 → 11 taken 5 times.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 10 times.
✓ Branch 10 → 11 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 12 times.
✓ Branch 10 → 11 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 5 times.
✓ Branch 10 → 11 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 5 times.
✓ Branch 10 → 11 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 5 times.
✓ Branch 10 → 11 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 10 times.
✓ Branch 10 → 11 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 10 times.
✓ Branch 10 → 11 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 5 times.
✓ Branch 10 → 11 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 5 times.
✓ Branch 10 → 11 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 5 times.
✓ Branch 10 → 11 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 5 times.
✓ Branch 10 → 11 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 10 → 7 taken 99 times.
✓ Branch 10 → 11 taken 26 times.
void masked_merge_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 9 → 6 taken 49 times.
✓ Branch 9 → 10 taken 27 times.
|
317 | for (int y = 0; y < height; y++) { |
| 185 |
14/32void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 19 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 10 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 12 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 5 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 5 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 5 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 10 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 10 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 5 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 5 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 5 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 5 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 7 → 8 taken 99 times.
✗ Branch 7 → 43 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 6 → 7 taken 49 times.
✗ Branch 6 → 41 not taken.
|
244 | const uint8_t* eff = prepare_effective_mask_for_row_avx2<maskMode, uint8_t, full_opacity>( |
| 186 | maskp, mpx, width, eff_buf, opacity, half, mag); | ||
| 187 | 244 | blend8_masked_avx2_row( | |
| 188 | reinterpret_cast<uint8_t*>(p1), reinterpret_cast<const uint8_t*>(p2), eff, width); | ||
| 189 | 244 | p1 += p1_pitch; p2 += p2_pitch; maskp += mask_adv; | |
| 190 | } | ||
| 191 | 73 | return; | |
| 192 | 73 | } | |
| 193 | |||
| 194 | 44 | const uint16_t* maskp = reinterpret_cast<const uint16_t*>(mask); | |
| 195 | 44 | const int mpx = mask_pitch / 2; | |
| 196 | 44 | const int mask_adv = (maskMode == MASK420 || maskMode == MASK420_MPEG2 || maskMode == MASK420_TOPLEFT) ? mpx * 2 : mpx; | |
| 197 | |||
| 198 | 44 | std::vector<uint16_t> eff_buf; | |
| 199 |
13/30void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 14 → 15 not taken.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 14 → 15 not taken.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 2 times.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 2 times.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 1 time.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 1 time.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 1 time.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 1 time.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 3 times.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 2 times.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 1 time.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 1 time.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 1 time.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 1 time.
✗ Branch 14 → 46 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 14 → 15 taken 15 times.
✗ Branch 14 → 46 not taken.
|
32 | if constexpr (maskMode != MASK444 || !full_opacity) eff_buf.resize(width); |
| 200 | |||
| 201 | #define BLEND16_LOOP_AVX2(bpp) \ | ||
| 202 | for (int y = 0; y < height; y++) { \ | ||
| 203 | const uint16_t* eff = prepare_effective_mask_for_row_avx2<maskMode, uint16_t, full_opacity>( \ | ||
| 204 | maskp, mpx, width, eff_buf, opacity, half, mag); \ | ||
| 205 | blend16_masked_avx2_row<bpp>( \ | ||
| 206 | reinterpret_cast<uint16_t*>(p1), reinterpret_cast<const uint16_t*>(p2), eff, width); \ | ||
| 207 | p1 += p1_pitch; p2 += p2_pitch; maskp += mask_adv; \ | ||
| 208 | } break; | ||
| 209 | |||
| 210 |
20/80void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✗ Branch 15 → 34 not taken.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✗ Branch 15 → 34 not taken.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 2 times.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 2 times.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 1 time.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 1 time.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 1 time.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 1 time.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 3 times.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 2 times.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 1 time.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 1 time.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 1 time.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 15 → 16 not taken.
✗ Branch 15 → 22 not taken.
✗ Branch 15 → 28 not taken.
✓ Branch 15 → 34 taken 1 time.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 15 → 16 taken 1 time.
✓ Branch 15 → 22 taken 1 time.
✓ Branch 15 → 28 taken 1 time.
✓ Branch 15 → 34 taken 12 times.
✗ Branch 15 → 40 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 13 → 14 taken 1 time.
✓ Branch 13 → 20 taken 1 time.
✓ Branch 13 → 26 taken 1 time.
✓ Branch 13 → 32 taken 9 times.
✗ Branch 13 → 38 not taken.
|
44 | switch (bits_per_pixel) { |
| 211 |
8/96void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 17 → 18 not taken.
✗ Branch 17 → 46 not taken.
✗ Branch 18 → 19 not taken.
✗ Branch 18 → 46 not taken.
✗ Branch 20 → 17 not taken.
✗ Branch 20 → 21 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 17 → 18 taken 5 times.
✗ Branch 17 → 46 not taken.
✓ Branch 18 → 19 taken 5 times.
✗ Branch 18 → 46 not taken.
✓ Branch 20 → 17 taken 5 times.
✓ Branch 20 → 21 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 15 → 16 taken 5 times.
✗ Branch 15 → 44 not taken.
✓ Branch 16 → 17 taken 5 times.
✗ Branch 16 → 44 not taken.
✓ Branch 18 → 15 taken 5 times.
✓ Branch 18 → 19 taken 1 time.
|
12 | case 10: BLEND16_LOOP_AVX2(10) |
| 212 |
8/96void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 23 → 24 not taken.
✗ Branch 23 → 46 not taken.
✗ Branch 24 → 25 not taken.
✗ Branch 24 → 46 not taken.
✗ Branch 26 → 23 not taken.
✗ Branch 26 → 27 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 23 → 24 taken 5 times.
✗ Branch 23 → 46 not taken.
✓ Branch 24 → 25 taken 5 times.
✗ Branch 24 → 46 not taken.
✓ Branch 26 → 23 taken 5 times.
✓ Branch 26 → 27 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 21 → 22 taken 5 times.
✗ Branch 21 → 44 not taken.
✓ Branch 22 → 23 taken 5 times.
✗ Branch 22 → 44 not taken.
✓ Branch 24 → 21 taken 5 times.
✓ Branch 24 → 25 taken 1 time.
|
12 | case 12: BLEND16_LOOP_AVX2(12) |
| 213 |
8/96void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 29 → 30 not taken.
✗ Branch 29 → 46 not taken.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 46 not taken.
✗ Branch 32 → 29 not taken.
✗ Branch 32 → 33 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 29 → 30 taken 5 times.
✗ Branch 29 → 46 not taken.
✓ Branch 30 → 31 taken 5 times.
✗ Branch 30 → 46 not taken.
✓ Branch 32 → 29 taken 5 times.
✓ Branch 32 → 33 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 27 → 28 taken 5 times.
✗ Branch 27 → 44 not taken.
✓ Branch 28 → 29 taken 5 times.
✗ Branch 28 → 44 not taken.
✓ Branch 30 → 27 taken 5 times.
✓ Branch 30 → 31 taken 1 time.
|
12 | case 14: BLEND16_LOOP_AVX2(14) |
| 214 |
56/96void masked_merge_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 35 → 36 not taken.
✗ Branch 35 → 46 not taken.
✗ Branch 36 → 37 not taken.
✗ Branch 36 → 46 not taken.
✗ Branch 38 → 35 not taken.
✗ Branch 38 → 39 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 35 → 36 not taken.
✗ Branch 35 → 46 not taken.
✗ Branch 36 → 37 not taken.
✗ Branch 36 → 46 not taken.
✗ Branch 38 → 35 not taken.
✗ Branch 38 → 39 not taken.
void masked_merge_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 10 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 10 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 10 times.
✓ Branch 38 → 39 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 10 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 10 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 10 times.
✓ Branch 38 → 39 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 5 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 5 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 5 times.
✓ Branch 38 → 39 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 5 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 5 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 5 times.
✓ Branch 38 → 39 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 5 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 5 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 5 times.
✓ Branch 38 → 39 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 5 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 5 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 5 times.
✓ Branch 38 → 39 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 15 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 15 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 15 times.
✓ Branch 38 → 39 taken 3 times.
void masked_merge_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 10 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 10 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 10 times.
✓ Branch 38 → 39 taken 2 times.
void masked_merge_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 5 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 5 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 5 times.
✓ Branch 38 → 39 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 5 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 5 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 5 times.
✓ Branch 38 → 39 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 5 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 5 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 5 times.
✓ Branch 38 → 39 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 5 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 5 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 5 times.
✓ Branch 38 → 39 taken 1 time.
void masked_merge_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 35 → 36 taken 40 times.
✗ Branch 35 → 46 not taken.
✓ Branch 36 → 37 taken 10 times.
✗ Branch 36 → 46 not taken.
✓ Branch 38 → 35 taken 40 times.
✓ Branch 38 → 39 taken 12 times.
void masked_merge_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 33 → 34 taken 31 times.
✗ Branch 33 → 44 not taken.
✓ Branch 34 → 35 taken 10 times.
✗ Branch 34 → 44 not taken.
✓ Branch 36 → 33 taken 31 times.
✓ Branch 36 → 37 taken 9 times.
|
194 | case 16: BLEND16_LOOP_AVX2(16) |
| 215 | } | ||
| 216 | #undef BLEND16_LOOP_AVX2 | ||
| 217 | 44 | } | |
| 218 | |||
| 219 | // --------------------------------------------------------------------------- | ||
| 220 | // Inner loop — full_opacity known at compile time. | ||
| 221 | // Rowprep bakes opacity when !full_opacity; blend row receives pre-scaled mask. | ||
| 222 | // MASK444 + full_opacity: returns mask ptr directly (no buffer, no copy). | ||
| 223 | // MASK444 + !full_opacity: copies row with opacity scaling into buffer. | ||
| 224 | // --------------------------------------------------------------------------- | ||
| 225 | template<MaskMode maskMode, bool full_opacity> | ||
| 226 | #if defined(GCC) || defined(CLANG) | ||
| 227 | __attribute__((__target__("avx2"))) | ||
| 228 | #endif | ||
| 229 | 6 | static void masked_merge_float_avx2_impl_inner( | |
| 230 | BYTE* p1, const BYTE* p2, const BYTE* mask, | ||
| 231 | int p1_pitch, int p2_pitch, int mask_pitch, | ||
| 232 | int width, int height, float opacity) | ||
| 233 | { | ||
| 234 | 6 | const float* maskp = reinterpret_cast<const float*>(mask); | |
| 235 | 6 | const int mpx = mask_pitch / sizeof(float); | |
| 236 | 6 | const int mask_adv = (maskMode == MASK420 || maskMode == MASK420_MPEG2 || maskMode == MASK420_TOPLEFT) ? mpx * 2 : mpx; | |
| 237 | |||
| 238 | 6 | std::vector<float> eff_buf; | |
| 239 |
5/30void masked_merge_float_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 3 → 4 taken 1 time.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 3 → 4 taken 1 time.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 3 → 4 taken 1 time.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 3 → 4 taken 1 time.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 3 → 4 not taken.
✗ Branch 3 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 3 → 4 taken 1 time.
✗ Branch 3 → 11 not taken.
|
5 | if constexpr (maskMode != MASK444 || !full_opacity) eff_buf.resize(width); |
| 240 | |||
| 241 |
12/32void masked_merge_float_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 8 → 5 taken 5 times.
✓ Branch 8 → 9 taken 1 time.
void masked_merge_float_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 8 → 5 taken 5 times.
✓ Branch 8 → 9 taken 1 time.
void masked_merge_float_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 8 → 5 taken 5 times.
✓ Branch 8 → 9 taken 1 time.
void masked_merge_float_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 8 → 5 taken 5 times.
✓ Branch 8 → 9 taken 1 time.
void masked_merge_float_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 8 → 5 not taken.
✗ Branch 8 → 9 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 8 → 5 taken 5 times.
✓ Branch 8 → 9 taken 1 time.
void masked_merge_float_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 7 → 4 taken 5 times.
✓ Branch 7 → 8 taken 1 time.
|
36 | for (int y = 0; y < height; y++) { |
| 242 |
6/32void masked_merge_float_avx2_impl_inner<(MaskMode)0, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)0, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)1, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 5 → 6 taken 5 times.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)1, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 5 → 6 taken 5 times.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)2, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)2, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)3, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)3, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)4, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 5 → 6 taken 5 times.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)4, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 5 → 6 taken 5 times.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)5, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)5, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)6, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)6, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 5 → 6 not taken.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)7, false>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 5 → 6 taken 5 times.
✗ Branch 5 → 11 not taken.
void masked_merge_float_avx2_impl_inner<(MaskMode)7, true>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 4 → 5 taken 5 times.
✗ Branch 4 → 10 not taken.
|
30 | const float* eff = prepare_effective_mask_for_row_float_avx2<maskMode, full_opacity>( |
| 243 | maskp, mpx, width, eff_buf, opacity); | ||
| 244 | 30 | blend_masked_float_avx2_row( | |
| 245 | reinterpret_cast<float*>(p1), reinterpret_cast<const float*>(p2), eff, width); | ||
| 246 | 30 | p1 += p1_pitch; p2 += p2_pitch; maskp += mask_adv; | |
| 247 | } | ||
| 248 | 6 | } | |
| 249 | |||
| 250 | // --------------------------------------------------------------------------- | ||
| 251 | // Outer: dispatch on full_opacity (opacity == max_pixel_value) at the call site. | ||
| 252 | // --------------------------------------------------------------------------- | ||
| 253 | template<MaskMode maskMode> | ||
| 254 | #if defined(GCC) || defined(CLANG) | ||
| 255 | __attribute__((__target__("avx2"))) | ||
| 256 | #endif | ||
| 257 | 117 | static void masked_merge_avx2_impl( | |
| 258 | BYTE* p1, const BYTE* p2, const BYTE* mask, | ||
| 259 | int p1_pitch, int p2_pitch, int mask_pitch, | ||
| 260 | int width, int height, int opacity, int bits_per_pixel) | ||
| 261 | { | ||
| 262 | 117 | const int max_val = (1 << bits_per_pixel) - 1; | |
| 263 |
14/16void masked_merge_avx2_impl<(MaskMode)0>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 4 not taken.
void masked_merge_avx2_impl<(MaskMode)1>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 4 times.
✓ Branch 2 → 4 taken 7 times.
void masked_merge_avx2_impl<(MaskMode)2>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 2 times.
✓ Branch 2 → 4 taken 3 times.
void masked_merge_avx2_impl<(MaskMode)3>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 2 times.
✓ Branch 2 → 4 taken 2 times.
void masked_merge_avx2_impl<(MaskMode)4>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 4 times.
✓ Branch 2 → 4 taken 5 times.
void masked_merge_avx2_impl<(MaskMode)5>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 2 times.
✓ Branch 2 → 4 taken 2 times.
void masked_merge_avx2_impl<(MaskMode)6>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 2 times.
✓ Branch 2 → 4 taken 2 times.
void masked_merge_avx2_impl<(MaskMode)7>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, int, int):
✓ Branch 2 → 3 taken 39 times.
✓ Branch 2 → 4 taken 41 times.
|
117 | if (opacity == max_val) |
| 264 | 55 | masked_merge_avx2_impl_inner<maskMode, true>( | |
| 265 | p1, p2, mask, p1_pitch, p2_pitch, mask_pitch, width, height, opacity, bits_per_pixel); | ||
| 266 | else | ||
| 267 | 62 | masked_merge_avx2_impl_inner<maskMode, false>( | |
| 268 | p1, p2, mask, p1_pitch, p2_pitch, mask_pitch, width, height, opacity, bits_per_pixel); | ||
| 269 | 117 | } | |
| 270 | |||
| 271 | template<MaskMode maskMode> | ||
| 272 | #if defined(GCC) || defined(CLANG) | ||
| 273 | __attribute__((__target__("avx2"))) | ||
| 274 | #endif | ||
| 275 | 6 | static void masked_merge_float_avx2_impl( | |
| 276 | BYTE* p1, const BYTE* p2, const BYTE* mask, | ||
| 277 | int p1_pitch, int p2_pitch, int mask_pitch, | ||
| 278 | int width, int height, float opacity) | ||
| 279 | { | ||
| 280 |
6/16void masked_merge_float_avx2_impl<(MaskMode)0>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 4 not taken.
void masked_merge_float_avx2_impl<(MaskMode)1>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 2 → 3 taken 1 time.
✓ Branch 2 → 4 taken 1 time.
void masked_merge_float_avx2_impl<(MaskMode)2>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 4 not taken.
void masked_merge_float_avx2_impl<(MaskMode)3>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 4 not taken.
void masked_merge_float_avx2_impl<(MaskMode)4>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 2 → 3 taken 1 time.
✓ Branch 2 → 4 taken 1 time.
void masked_merge_float_avx2_impl<(MaskMode)5>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 4 not taken.
void masked_merge_float_avx2_impl<(MaskMode)6>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✗ Branch 2 → 3 not taken.
✗ Branch 2 → 4 not taken.
void masked_merge_float_avx2_impl<(MaskMode)7>(unsigned char*, unsigned char const*, unsigned char const*, int, int, int, int, int, float):
✓ Branch 2 → 3 taken 1 time.
✓ Branch 2 → 4 taken 1 time.
|
6 | if (opacity >= 1.0f) |
| 281 | 3 | masked_merge_float_avx2_impl_inner<maskMode, true>( | |
| 282 | p1, p2, mask, p1_pitch, p2_pitch, mask_pitch, width, height, opacity); | ||
| 283 | else | ||
| 284 | 3 | masked_merge_float_avx2_impl_inner<maskMode, false>( | |
| 285 | p1, p2, mask, p1_pitch, p2_pitch, mask_pitch, width, height, opacity); | ||
| 286 | 6 | } | |
| 287 | |||
| 288 |