filters/intel/resample_avx512.hpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // AviSynth+. Copyright 2026- AviSynth+ Project | ||
| 2 | // https://avs-plus.net | ||
| 3 | // http://avisynth.nl | ||
| 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 | |||
| 37 | This is a common source cpp include file (not header) for multi-arch AVX512 functions. | ||
| 38 | Functions here are static, they will be compiled into each translation unit including this file. | ||
| 39 | |||
| 40 | */ | ||
| 41 | |||
| 42 | // Original function needed avx512vbmi feature flag, but we want to support also base AVX512 without VBMI. | ||
| 43 | // We use _mm512_permutex2var_epi8_SIMUL<UseVBMI> and _mm512_maskz_permutex2var_epi8_SIMUL<UseVBMI> | ||
| 44 | // Thus both Base AVX512 and ICL level arch is supported. | ||
| 45 | // We are using two separated source modules and include this hpp file templated | ||
| 46 | // with UseVBMI/UseVNNI | ||
| 47 | |||
| 48 | // Notes: | ||
| 49 | // As of January 2026, Visual Studio 2026 ships with clang-cl (LLVM 20.1.8). | ||
| 50 | // - This version typically avoids using VNNI vpdpwssd instructions, opting instead for separate madd and add operations. | ||
| 51 | // - Masked permute operations are not optimized: for example, instead of using masked permutex2var_epi8, | ||
| 52 | // it performs a basic permutex2var_epi8 followed by an "and" with a pre-loaded zmm mask. | ||
| 53 | // These behaviors result in slower code compared to MSVC builds, which utilize these instructions more efficiently. | ||
| 54 | // These optimization issues are resolved in LLVM 21 (e.g., Intel C++ Compiler 2025.3). | ||
| 55 | |||
| 56 | // helper function for simulating _mm512_permutex2var_epi8 when VBMI is not available | ||
| 57 | // The MSB bit (128) zeroing effect is _not_ considered here, the indices must be all positive and within 0-127 range. | ||
| 58 | // Helper for _mm512_permutex2var_epi8_SIMUL: gather 32 bytes from [a,b] using precomputed word_idx and shift_amt. | ||
| 59 | // Extracted from lambda to ensure MSVC inlines it (lambdas are not reliably inlined by MSVC). | ||
| 60 | // Accepts precomputed word_idx (target_idx>>1) and shift_amt ((target_idx<<3)&8) so callers outside | ||
| 61 | // the y-loop can hoist these invariants, avoiding recomputation every row. | ||
| 62 | // Returns 32 16-bit words with gathered bytes in low 8 bits (high 8 bits cleared). | ||
| 63 | // Returning __m512i instead of __m256i lets callers feed unpacklo/hi_epi16 directly, | ||
| 64 | // avoiding the cvtepu8_epi16 expansion round-trip that MSVC emits as vpmovwb+vpmovzxbw. | ||
| 65 | AVS_FORCEINLINE static __m512i _permutex2var_epi8_sim_get32(__m512i word_idx, __m512i shift_amt, __m512i a, __m512i b) | ||
| 66 | { | ||
| 67 | 1520 | __m512i words = _mm512_permutex2var_epi16(a, word_idx, b); | |
| 68 | // vpsrlvw: avoids k-register RAW dependency; MSVC compiles test_epi16_mask+mask_srli as | ||
| 69 | // vptestmw->k1->vpsrlw{k1}, creating a read-after-write through the mask unit (higher latency) | ||
| 70 | 1520 | words = _mm512_srlv_epi16(words, shift_amt); | |
| 71 | // clear high byte so the caller can use unpacklo/hi_epi16 without cvtepu8_epi16 | ||
| 72 | 3040 | return _mm512_and_si512(words, _mm512_set1_epi16(0x00FF)); | |
| 73 | } | ||
| 74 | |||
| 75 | template<bool UseVBMI> | ||
| 76 | AVS_FORCEINLINE static __m512i _mm512_permutex2var_epi8_SIMUL(__m512i a, __m512i idx, __m512i b) { | ||
| 77 | if constexpr (UseVBMI) { | ||
| 78 | return _mm512_permutex2var_epi8(a, idx, b); | ||
| 79 | } | ||
| 80 | else { | ||
| 81 | const __m512i c_8 = _mm512_set1_epi16(8); | ||
| 82 | |||
| 83 | __m512i idx_lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(idx)); | ||
| 84 | __m512i idx_hi = _mm512_cvtepu8_epi16(_mm512_extracti64x4_epi64(idx, 1)); | ||
| 85 | |||
| 86 | __m512i wi_lo = _mm512_srli_epi16(idx_lo, 1); | ||
| 87 | __m512i sa_lo = _mm512_and_si512(_mm512_slli_epi16(idx_lo, 3), c_8); | ||
| 88 | __m512i wi_hi = _mm512_srli_epi16(idx_hi, 1); | ||
| 89 | __m512i sa_hi = _mm512_and_si512(_mm512_slli_epi16(idx_hi, 3), c_8); | ||
| 90 | |||
| 91 | __m256i res_0_31 = _mm512_cvtepi16_epi8(_permutex2var_epi8_sim_get32(wi_lo, sa_lo, a, b)); | ||
| 92 | __m256i res_32_63 = _mm512_cvtepi16_epi8(_permutex2var_epi8_sim_get32(wi_hi, sa_hi, a, b)); | ||
| 93 | |||
| 94 | return _mm512_inserti64x4(_mm512_castsi256_si512(res_0_31), res_32_63, 1); | ||
| 95 | } | ||
| 96 | } | ||
| 97 | |||
| 98 | template<bool UseVBMI> | ||
| 99 | AVS_FORCEINLINE static __m512i _mm512_maskz_permutex2var_epi8_SIMUL( | ||
| 100 | __mmask64 k, | ||
| 101 | __m512i a, | ||
| 102 | __m512i idx, | ||
| 103 | __m512i b) | ||
| 104 | { | ||
| 105 | if constexpr (UseVBMI) { | ||
| 106 | 600 | return _mm512_maskz_permutex2var_epi8(k, a, idx, b); | |
| 107 | } | ||
| 108 | else { | ||
| 109 | // 1. Run the base simulation to get the permuted bytes | ||
| 110 | // Note: Using your existing logic to get the full 512-bit permuted result | ||
| 111 | __m512i res = _mm512_permutex2var_epi8_SIMUL<false>(a, idx, b); | ||
| 112 | |||
| 113 | // 2. Apply the zero-mask using AVX-512BW | ||
| 114 | // _mm512_maskz_mov_epi8(k, src, src) returns 'src' where k[i]==1, and 0 where k[i]==0 | ||
| 115 | return _mm512_maskz_mov_epi8(k, res); | ||
| 116 | } | ||
| 117 | } | ||
| 118 | |||
| 119 | // H-Float-Resampler: 16 pixels, filter size 4, transpose 4x (4x_m128) to 4x_m512 | ||
| 120 | // Transposes a 4x4 matrix of 4-float vectors (16x16 float matrix effectively). | ||
| 121 | // Input/Output: Four 512-bit vectors (16 floats each) passed by reference. | ||
| 122 | AVS_FORCEINLINE static void _MM_TRANSPOSE16_LANE4_PS(__m512& row0, __m512& row1, __m512& row2, __m512& row3) { | ||
| 123 | // Stage 1: Interleave 32-bit (float) elements within 128-bit chunks (lanes) | ||
| 124 | // t0 = (r0_lo, r1_lo) | t1 = (r0_hi, r1_hi) | ||
| 125 | // t2 = (r2_lo, r3_lo) | t3 = (r2_hi, r3_hi) | ||
| 126 | 62 | __m512 t0 = _mm512_unpacklo_ps(row0, row1); | |
| 127 | 62 | __m512 t1 = _mm512_unpackhi_ps(row0, row1); | |
| 128 | 62 | __m512 t2 = _mm512_unpacklo_ps(row2, row3); | |
| 129 | 124 | __m512 t3 = _mm512_unpackhi_ps(row2, row3); | |
| 130 | |||
| 131 | // Stage 2: Shuffle 128-bit chunks (lanes) to complete the transpose | ||
| 132 | // We use _mm512_shuffle_ps which shuffles 64-bit blocks across the 512-bit register. | ||
| 133 | // _MM_SHUFFLE(w, z, y, x) applies to the 64-bit pairs (4 floats) within each 128-bit lane. | ||
| 134 | // Result: row0 = columns 0, 1, 2, 3 | ||
| 135 | 62 | row0 = _mm512_shuffle_ps(t0, t2, _MM_SHUFFLE(1, 0, 1, 0)); | |
| 136 | // Result: row1 = columns 4, 5, 6, 7 | ||
| 137 | 62 | row1 = _mm512_shuffle_ps(t0, t2, _MM_SHUFFLE(3, 2, 3, 2)); | |
| 138 | // Result: row2 = columns 8, 9, 10, 11 | ||
| 139 | 62 | row2 = _mm512_shuffle_ps(t1, t3, _MM_SHUFFLE(1, 0, 1, 0)); | |
| 140 | // Result: row3 = columns 12, 13, 14, 15 | ||
| 141 | 62 | row3 = _mm512_shuffle_ps(t1, t3, _MM_SHUFFLE(3, 2, 3, 2)); | |
| 142 | 62 | } | |
| 143 | |||
| 144 | // H-float-resampler: 16 pixels, filter size 8, transpose 8x (2x_m256) to 8x_m512 | ||
| 145 | // Transposes an 8x8 matrix of 2-float vectors (16x16 float matrix). | ||
| 146 | // Input/Output: Eight 512-bit vectors (16 floats each) passed by reference. | ||
| 147 | AVS_FORCEINLINE static void _MM_TRANSPOSE8x16_PS( | ||
| 148 | __m512& r0, __m512& r1, __m512& r2, __m512& r3, | ||
| 149 | __m512& r4, __m512& r5, __m512& r6, __m512& r7) | ||
| 150 | { | ||
| 151 | // --- Stage 1: Unpack 32-bit (Pairs of rows) --- | ||
| 152 | 24 | __m512 t0 = _mm512_unpacklo_ps(r0, r1); | |
| 153 | 24 | __m512 t1 = _mm512_unpackhi_ps(r0, r1); | |
| 154 | 24 | __m512 t2 = _mm512_unpacklo_ps(r2, r3); | |
| 155 | 24 | __m512 t3 = _mm512_unpackhi_ps(r2, r3); | |
| 156 | 24 | __m512 t4 = _mm512_unpacklo_ps(r4, r5); | |
| 157 | 24 | __m512 t5 = _mm512_unpackhi_ps(r4, r5); | |
| 158 | 24 | __m512 t6 = _mm512_unpacklo_ps(r6, r7); | |
| 159 | 48 | __m512 t7 = _mm512_unpackhi_ps(r6, r7); | |
| 160 | |||
| 161 | // --- Stage 2: Unpack 64-bit (Quads of rows) --- | ||
| 162 | // Uses _mm512_unpacklo/hi_pd for 64-bit (double) to interleave pairs of __m512 floats | ||
| 163 | 96 | __m512 u0 = _mm512_castpd_ps(_mm512_unpacklo_pd(_mm512_castps_pd(t0), _mm512_castps_pd(t2))); | |
| 164 | 96 | __m512 u1 = _mm512_castpd_ps(_mm512_unpackhi_pd(_mm512_castps_pd(t0), _mm512_castps_pd(t2))); | |
| 165 | 96 | __m512 u2 = _mm512_castpd_ps(_mm512_unpacklo_pd(_mm512_castps_pd(t1), _mm512_castps_pd(t3))); | |
| 166 | 96 | __m512 u3 = _mm512_castpd_ps(_mm512_unpackhi_pd(_mm512_castps_pd(t1), _mm512_castps_pd(t3))); | |
| 167 | 96 | __m512 u4 = _mm512_castpd_ps(_mm512_unpacklo_pd(_mm512_castps_pd(t4), _mm512_castps_pd(t6))); | |
| 168 | 96 | __m512 u5 = _mm512_castpd_ps(_mm512_unpackhi_pd(_mm512_castps_pd(t4), _mm512_castps_pd(t6))); | |
| 169 | 96 | __m512 u6 = _mm512_castpd_ps(_mm512_unpacklo_pd(_mm512_castps_pd(t5), _mm512_castps_pd(t7))); | |
| 170 | 96 | __m512 u7 = _mm512_castpd_ps(_mm512_unpackhi_pd(_mm512_castps_pd(t5), _mm512_castps_pd(t7))); | |
| 171 | |||
| 172 | // --- Stage 3: Shuffle 128-bit lanes (Octets of rows) --- | ||
| 173 | // _mm512_shuffle_f32x4 shuffles the 128-bit (f32x4) sub-vectors within and between two __m512 vectors. | ||
| 174 | // 0x88 = (10001000)_2: selects lane 0 from first input and lane 0 from second input for lo/hi 256 bits. | ||
| 175 | // 0xDD = (11011101)_2: selects lane 3 from first input and lane 3 from second input for lo/hi 256 bits. | ||
| 176 | 24 | __m512 v0 = _mm512_shuffle_f32x4(u0, u4, 0x88); // Col 0, 4 (interleaved) | |
| 177 | 24 | __m512 v1 = _mm512_shuffle_f32x4(u0, u4, 0xDD); // Col 1, 5 (interleaved) | |
| 178 | 24 | __m512 v2 = _mm512_shuffle_f32x4(u1, u5, 0x88); // Col 2, 6 (interleaved) | |
| 179 | 24 | __m512 v3 = _mm512_shuffle_f32x4(u1, u5, 0xDD); // Col 3, 7 (interleaved) | |
| 180 | 24 | __m512 v4 = _mm512_shuffle_f32x4(u2, u6, 0x88); // Col 8, 12 (interleaved) | |
| 181 | 24 | __m512 v5 = _mm512_shuffle_f32x4(u2, u6, 0xDD); // Col 9, 13 (interleaved) | |
| 182 | 24 | __m512 v6 = _mm512_shuffle_f32x4(u3, u7, 0x88); // Col 10, 14 (interleaved) | |
| 183 | 24 | __m512 v7 = _mm512_shuffle_f32x4(u3, u7, 0xDD); // Col 11, 15 (interleaved) | |
| 184 | |||
| 185 | // --- Stage 4: Permute to Linearize Indices --- | ||
| 186 | // Corrects the order of the 128-bit lanes to linearize the columns. | ||
| 187 | // The columns are currently: (0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15) | ||
| 188 | // The required order is: (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15) | ||
| 189 | 24 | __m512i idx = _mm512_setr_epi32( | |
| 190 | 0, 4, 1, 5, /* Lane 0: Rows 0, 1, 2, 3 */ | ||
| 191 | 2, 6, 3, 7, /* Lane 1: Rows 4, 5, 6, 7 */ | ||
| 192 | 8, 12, 9, 13, /* Lane 2: Rows 8, 9, 10, 11 */ | ||
| 193 | 10, 14, 11, 15 /* Lane 3: Rows 12, 13, 14, 15 */ | ||
| 194 | ); | ||
| 195 | |||
| 196 | // --- Final Assignment with Correct Mapping --- | ||
| 197 | // Maps the permuted vector components back to the original row variables (now columns). | ||
| 198 | 24 | r0 = _mm512_permutexvar_ps(idx, v0); // Col 0 | |
| 199 | 24 | r1 = _mm512_permutexvar_ps(idx, v2); // Col 1 | |
| 200 | 24 | r2 = _mm512_permutexvar_ps(idx, v4); // Col 2 | |
| 201 | 24 | r3 = _mm512_permutexvar_ps(idx, v6); // Col 3 | |
| 202 | 24 | r4 = _mm512_permutexvar_ps(idx, v1); // Col 4 | |
| 203 | 24 | r5 = _mm512_permutexvar_ps(idx, v3); // Col 5 | |
| 204 | 24 | r6 = _mm512_permutexvar_ps(idx, v5); // Col 6 | |
| 205 | 24 | r7 = _mm512_permutexvar_ps(idx, v7); // Col 7 | |
| 206 | 24 | } | |
| 207 | |||
| 208 | // Loads two 256-bit float vectors from registers (__m256) into a single 512-bit register. | ||
| 209 | // Equivalent to _mm512_insertf32x8(_mm512_castps256_ps512(lo), hi, 1) | ||
| 210 | AVS_FORCEINLINE static __m512 _mm512_insert_2_m256(__m256 lo, __m256 hi) { | ||
| 211 | 384 | return _mm512_insertf32x8(_mm512_castps256_ps512(lo), hi, 1); | |
| 212 | } | ||
| 213 | |||
| 214 | // Loads four 128-bit float vectors (unaligned) into a single 512-bit register. | ||
| 215 | AVS_FORCEINLINE static __m512 _mm512_loadu_4_m128( | ||
| 216 | /* __m128 const* */ const float* addr1, | ||
| 217 | /* __m128 const* */ const float* addr2, | ||
| 218 | /* __m128 const* */ const float* addr3, | ||
| 219 | /* __m128 const* */ const float* addr4) | ||
| 220 | { | ||
| 221 | // The cast is needed for the first insertion to make the target a 512-bit register | ||
| 222 | 160 | __m512 v = _mm512_castps128_ps512(_mm_loadu_ps(addr1)); | |
| 223 | 80 | v = _mm512_insertf32x4(v, _mm_loadu_ps(addr2), 1); | |
| 224 | 80 | v = _mm512_insertf32x4(v, _mm_loadu_ps(addr3), 2); | |
| 225 | 80 | v = _mm512_insertf32x4(v, _mm_loadu_ps(addr4), 3); | |
| 226 | 80 | return v; | |
| 227 | } | ||
| 228 | |||
| 229 | // Loads four 128-bit float vectors (aligned) into a single 512-bit register. | ||
| 230 | AVS_FORCEINLINE static __m512 _mm512_load_4_m128( | ||
| 231 | /* __m128 const* */ const float* addr1, | ||
| 232 | /* __m128 const* */ const float* addr2, | ||
| 233 | /* __m128 const* */ const float* addr3, | ||
| 234 | /* __m128 const* */ const float* addr4) | ||
| 235 | { | ||
| 236 | // The cast is needed for the first insertion to make the target a 512-bit register | ||
| 237 | 336 | __m512 v = _mm512_castps128_ps512(_mm_load_ps(addr1)); | |
| 238 | 168 | v = _mm512_insertf32x4(v, _mm_load_ps(addr2), 1); | |
| 239 | 168 | v = _mm512_insertf32x4(v, _mm_load_ps(addr3), 2); | |
| 240 | 168 | v = _mm512_insertf32x4(v, _mm_load_ps(addr4), 3); | |
| 241 | 168 | return v; | |
| 242 | } | ||
| 243 | |||
| 244 | // Loads two 256 - bit unaligned integer vectors from registers(__m256i) into a single 512i register. | ||
| 245 | AVS_FORCEINLINE static __m512i _mm512i_loadu_2_m256i( | ||
| 246 | const __m256i* addr1, | ||
| 247 | const __m256i* addr2) | ||
| 248 | { | ||
| 249 | return _mm512_inserti64x4(_mm512_castsi256_si512(_mm256_loadu_si256(addr1)), _mm256_loadu_si256(addr2), 1); | ||
| 250 | } | ||
| 251 | |||
| 252 | // Loads two 256 - bit aligned integer vectors from registers(__m256) into a single 512 - bit register. | ||
| 253 | AVS_FORCEINLINE static __m512i _mm512i_load_2_m256i( | ||
| 254 | const __m256i* addr1, | ||
| 255 | const __m256i* addr2) | ||
| 256 | { | ||
| 257 | return _mm512_inserti64x4(_mm512_castsi256_si512(_mm256_load_si256(addr1)), _mm256_load_si256(addr2), 1); | ||
| 258 | } | ||
| 259 | |||
| 260 | // Integers | ||
| 261 | // Loads four 128-bit integer vectors (unaligned) into a single 512-bit integer register. | ||
| 262 | AVS_FORCEINLINE static __m512i _mm512i_loadu_4_m128i( | ||
| 263 | const __m128i* addr1, | ||
| 264 | const __m128i* addr2, | ||
| 265 | const __m128i* addr3, | ||
| 266 | const __m128i* addr4) | ||
| 267 | { | ||
| 268 | // The cast is needed for the first insertion to make the target a 512-bit register | ||
| 269 | __m512i v = _mm512_castsi128_si512(_mm_loadu_si128(addr1)); | ||
| 270 | v = _mm512_inserti32x4(v, _mm_loadu_si128(addr2), 1); | ||
| 271 | v = _mm512_inserti32x4(v, _mm_loadu_si128(addr3), 2); | ||
| 272 | v = _mm512_inserti32x4(v, _mm_loadu_si128(addr4), 3); | ||
| 273 | return v; | ||
| 274 | } | ||
| 275 | |||
| 276 | // filter size up to 4 | ||
| 277 | // 64 target uint8_t pixels at a time | ||
| 278 | // 128-byte source loads (128 uint8_t pixels) | ||
| 279 | // maximum permute index is 128 for _mm512_permutex2var_epi8 (uint8_t) | ||
| 280 | |||
| 281 | // filter size up to 8 | ||
| 282 | // 64 target uint8_t pixels at a time | ||
| 283 | // 128-byte source loads (128 uint8_t pixels) | ||
| 284 | // maximum permute index is 128 for _mm512_permutex2var_epi8(uint8_t) | ||
| 285 | |||
| 286 | // filter size up to 8 | ||
| 287 | // 64 target uint8_t pixels at a time in 2 groups of 32 to support longer source loading to each group to support lower downsample ratios | ||
| 288 | // support /2 downsample ratios for resizers with no-resize kernel size of 4 (or support of 2 ?) (Bicubic, Bilinear, and others, also SinPowResize (?)) | ||
| 289 | // 2 groups of 128-byte source loads (128 uint8_t pixels) | ||
| 290 | // maximum permute index is 128 for _mm512_permutex2var_epi8 (uint8_t) | ||
| 291 | |||
| 292 | // filter size up to 16 | ||
| 293 | // 32 target uint8_t pixels at a time | ||
| 294 | // 128-byte source loads (128 uint8_t pixels) | ||
| 295 | // maximum permute index is 128 for _mm512_permutex2var_epi8 (uint8_t) | ||
| 296 | // expect to support all upsampling ratios up to filter support of 8 (or 7..6 ?) and some downsampling ratios with filter support up to 3 (with downsample ratios from 0.5 or a bit lower) | ||
| 297 | |||
| 298 | // uint8_t "mp" versions | ||
| 299 | |||
| 300 | // filter size up to 4 | ||
| 301 | // 64 target uint8_t pixels at a time | ||
| 302 | // 127-byte source loads (127 uint8_t pixels) | ||
| 303 | // maximum permute index is 127 for _mm512_maskz_permutex2var_epi8 (uint8_t) | ||
| 304 | // more premutex version to create 8->16bit converted and low-hi unpacked registers in single permutex instruction | ||
| 305 | |||
| 306 | // filter size up to 4 | ||
| 307 | // 64 target uint8_t pixels at a time | ||
| 308 | // 127-byte source loads (127 uint8_t pixels) | ||
| 309 | // maximum permute index is 127 for _mm512_maskz_permutex2var_epi8 (uint8_t) | ||
| 310 | // more premutex version to create 8->16bit converted and low-hi unpacked registers in single permutex instruction | ||
| 311 | template<bool UseVNNI> | ||
| 312 | 76 | void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 313 | { | ||
| 314 | 76 | constexpr int PIXELS_AT_A_TIME = 64; | |
| 315 | |||
| 316 | // 'source_overread_beyond_targetx' indicates if the filter kernel can read beyond the target width. | ||
| 317 | // we load 2x64 source bytes at a time, so ensure safe overread if needed. | ||
| 318 | // Our main loop processes calculates for 64 target pixels at a time. | ||
| 319 | // Inside that, we load 128 source bytes (2x64) to be able to permutex from that. | ||
| 320 | // This we have to check at each mod-PIXELS_AT_A_TIME boundary, the allowance of 128-byte source load. | ||
| 321 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 75 times.
✗ Branch 2 → 4 not taken.
|
76 | const int width_safe_mod = (program->safelimit_128_pixels_each64th_target.overread_possible ? program->safelimit_128_pixels_each64th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 322 | |||
| 323 | // Preconditions: | ||
| 324 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 75 times.
|
76 | assert(program->filter_size_real <= 4); // We preload all relevant coefficients (up to 4) before the height loop. |
| 325 | |||
| 326 | // 'target_size_alignment' ensures we can safely access coefficients using offsets like | ||
| 327 | // 'filter_size * 15' when processing 16 H pixels at a time | ||
| 328 | // 'filter_size * 63' when processing 64 H pixels at a time | ||
| 329 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 75 times.
|
76 | assert(program->target_size_alignment >= 64); // Adjusted for 64 pixels (is it enough for uint8 ?) |
| 330 | |||
| 331 | assert(FRAME_ALIGN >= 64); // Good for 64x8 bit pixels | ||
| 332 | |||
| 333 | // Ensure that coefficient loading beyond the valid target size is safe for 4x4 float loads. | ||
| 334 | // We load 4x16bit coeffs at a time | ||
| 335 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 75 times.
|
76 | assert(program->filter_size_alignment >= 4); |
| 336 | |||
| 337 | 76 | const int max_scanlines = program->max_scanlines; | |
| 338 | |||
| 339 | 76 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale8bits - 1)); | |
| 340 | |||
| 341 | // Vertical stripe loop for L2 cache optimization | ||
| 342 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 75 times.
✓ Branch 24 → 25 taken 75 times.
|
152 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 343 | { | ||
| 344 | 76 | int y_to = std::min(y_from + max_scanlines, height); | |
| 345 | |||
| 346 | // Reset current_coeff for the start of the stripe (points to start of row's coeffs) | ||
| 347 | // const short* AVS_RESTRICT current_coeff = program->pixel_coefficient; | ||
| 348 | // Pre-transposed coefficients for AVX512, stored in the program structure, ready for direct use in the main loop without needing to transpose on the fly. | ||
| 349 | 76 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 350 | |||
| 351 | 76 | int x = 0; | |
| 352 | |||
| 353 | // Lambda to handle both safe (fast) and unsafe (masked/partial) loading paths | ||
| 354 | 156 | auto do_h_integer_core = [&](auto partial_load) { | |
| 355 | |||
| 356 | // prepare coefs in transposed V-form | ||
| 357 | // TODO: make storage in transposed form, 64 x uint16 transposition looks too slow | ||
| 358 | 80 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 359 | |||
| 360 | // load coeffs from prepared | ||
| 361 | 80 | const __m512i coef_r0r1_0_31lo = _mm512_load_si512(current_coeff_SIMD + 0); | |
| 362 | 80 | const __m512i coef_r0r1_0_31hi = _mm512_load_si512(current_coeff_SIMD + 1); // in count of __m512i | |
| 363 | |||
| 364 | 80 | const __m512i coef_r0r1_32_63lo = _mm512_load_si512(current_coeff_SIMD + 2); | |
| 365 | 80 | const __m512i coef_r0r1_32_63hi = _mm512_load_si512(current_coeff_SIMD + 3); | |
| 366 | |||
| 367 | 80 | const __m512i coef_r2r3_0_31lo = _mm512_load_si512(current_coeff_SIMD + 4); | |
| 368 | 80 | const __m512i coef_r2r3_0_31hi = _mm512_load_si512(current_coeff_SIMD + 5); | |
| 369 | |||
| 370 | 80 | const __m512i coef_r2r3_32_63lo = _mm512_load_si512(current_coeff_SIMD + 6); | |
| 371 | 80 | const __m512i coef_r2r3_32_63hi = _mm512_load_si512(current_coeff_SIMD + 7); | |
| 372 | |||
| 373 | // convert resampling program in H-form into permuting indexes for src transposition in V-form | ||
| 374 | 80 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); // 16 offsets | |
| 375 | 80 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); // 16 offsets | |
| 376 | 80 | __m512i perm_0_32_47 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 32])); // 16 offsets | |
| 377 | 80 | __m512i perm_0_48_63 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 48])); // 16 offsets | |
| 378 | |||
| 379 | 80 | int iStart = program->pixel_offset[x]; | |
| 380 | 80 | __m512i m512i_Start = _mm512_set1_epi32(iStart); | |
| 381 | |||
| 382 | 80 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, m512i_Start); | |
| 383 | 80 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, m512i_Start); | |
| 384 | 80 | perm_0_32_47 = _mm512_sub_epi32(perm_0_32_47, m512i_Start); | |
| 385 | 80 | perm_0_48_63 = _mm512_sub_epi32(perm_0_48_63, m512i_Start); | |
| 386 | |||
| 387 | 80 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 388 | 80 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 389 | 80 | __m256i m256i_perm_0_32_47 = _mm512_cvtepi32_epi16(perm_0_32_47); | |
| 390 | 80 | __m256i m256i_perm_0_48_63 = _mm512_cvtepi32_epi16(perm_0_48_63); | |
| 391 | |||
| 392 | 160 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 393 | 160 | __m512i perm_0_32_63 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_32_47), m256i_perm_0_48_63, 1); | |
| 394 | |||
| 395 | // Taps are contiguous (0, 1, 2, 3), so we increment perm indexes by 1. | ||
| 396 | 80 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 397 | 80 | __m512i perm_2_0_31 = _mm512_add_epi16(perm_1_0_31, one_epi16); | |
| 398 | 80 | __m512i perm_3_0_31 = _mm512_add_epi16(perm_2_0_31, one_epi16); | |
| 399 | |||
| 400 | 80 | __m512i perm_1_32_63 = _mm512_add_epi16(perm_0_32_63, one_epi16); | |
| 401 | 80 | __m512i perm_2_32_63 = _mm512_add_epi16(perm_1_32_63, one_epi16); | |
| 402 | 80 | __m512i perm_3_32_63 = _mm512_add_epi16(perm_2_32_63, one_epi16); | |
| 403 | |||
| 404 | 80 | const __m512i perm_r0r1_0_31lo = _mm512_unpacklo_epi16(perm_0_0_31, perm_1_0_31); | |
| 405 | 80 | const __m512i perm_r0r1_0_31hi = _mm512_unpackhi_epi16(perm_0_0_31, perm_1_0_31); | |
| 406 | |||
| 407 | 80 | const __m512i perm_r0r1_32_63lo = _mm512_unpacklo_epi16(perm_0_32_63, perm_1_32_63); | |
| 408 | 80 | const __m512i perm_r0r1_32_63hi = _mm512_unpackhi_epi16(perm_0_32_63, perm_1_32_63); | |
| 409 | |||
| 410 | 80 | const __m512i perm_r2r3_0_31lo = _mm512_unpacklo_epi16(perm_2_0_31, perm_3_0_31); | |
| 411 | 80 | const __m512i perm_r2r3_0_31hi = _mm512_unpackhi_epi16(perm_2_0_31, perm_3_0_31); | |
| 412 | |||
| 413 | 80 | const __m512i perm_r2r3_32_63lo = _mm512_unpacklo_epi16(perm_2_32_63, perm_3_32_63); | |
| 414 | 80 | const __m512i perm_r2r3_32_63hi = _mm512_unpackhi_epi16(perm_2_32_63, perm_3_32_63); | |
| 415 | |||
| 416 | 80 | __m512i wi_r0r1_0_31lo={},sa_r0r1_0_31lo={},wi_r0r1_0_31hi={},sa_r0r1_0_31hi={}; | |
| 417 | 80 | __m512i wi_r0r1_32_63lo={},sa_r0r1_32_63lo={},wi_r0r1_32_63hi={},sa_r0r1_32_63hi={}; | |
| 418 | 80 | __m512i wi_r2r3_0_31lo={},sa_r2r3_0_31lo={},wi_r2r3_0_31hi={},sa_r2r3_0_31hi={}; | |
| 419 | 80 | __m512i wi_r2r3_32_63lo={},sa_r2r3_32_63lo={},wi_r2r3_32_63hi={},sa_r2r3_32_63hi={}; | |
| 420 | if constexpr (!UseVNNI) { | ||
| 421 | 3 | const __m512i c_8w = _mm512_set1_epi16(8); | |
| 422 | 27 | auto make_wi_sa = [&](__m512i p, __m512i& wi, __m512i& sa) { | |
| 423 | 24 | wi = _mm512_srli_epi16(p, 1); | |
| 424 | 48 | sa = _mm512_and_si512(_mm512_slli_epi16(p, 3), c_8w); | |
| 425 | }; | ||
| 426 | 3 | make_wi_sa(perm_r0r1_0_31lo, wi_r0r1_0_31lo, sa_r0r1_0_31lo); | |
| 427 | 3 | make_wi_sa(perm_r0r1_0_31hi, wi_r0r1_0_31hi, sa_r0r1_0_31hi); | |
| 428 | 3 | make_wi_sa(perm_r0r1_32_63lo, wi_r0r1_32_63lo, sa_r0r1_32_63lo); | |
| 429 | 3 | make_wi_sa(perm_r0r1_32_63hi, wi_r0r1_32_63hi, sa_r0r1_32_63hi); | |
| 430 | 3 | make_wi_sa(perm_r2r3_0_31lo, wi_r2r3_0_31lo, sa_r2r3_0_31lo); | |
| 431 | 3 | make_wi_sa(perm_r2r3_0_31hi, wi_r2r3_0_31hi, sa_r2r3_0_31hi); | |
| 432 | 3 | make_wi_sa(perm_r2r3_32_63lo, wi_r2r3_32_63lo, sa_r2r3_32_63lo); | |
| 433 | 3 | make_wi_sa(perm_r2r3_32_63hi, wi_r2r3_32_63hi, sa_r2r3_32_63hi); | |
| 434 | } | ||
| 435 | |||
| 436 | 80 | uint8_t* AVS_RESTRICT dst_ptr = dst8 + x + y_from * dst_pitch; | |
| 437 | 80 | const uint8_t* src_ptr = src8 + iStart + y_from * src_pitch; // all permute offsets relative to this start offset | |
| 438 | |||
| 439 | // Calculate remaining pixels for bounds checking in partial_load mode. 1..128 remaining pixels possible. | ||
| 440 | 80 | const int remaining = program->source_size - iStart; | |
| 441 | 80 | __mmask64 k1 = _bzhi_u64(~0ULL, remaining); // _bzhi_u64 creates a mask with the lower N bits set. If N >= 64, it returns all ones (~0ULL). | |
| 442 | 80 | __mmask64 k2 = _bzhi_u64(~0ULL, std::max(0, remaining - 64)); | |
| 443 | |||
| 444 | // mask is used to zero out every odd byte, so that the result of the permute is a | ||
| 445 | // vector of zero-extended 8-bit values in 16-bit lanes, preparing 8-bit data for 16-bit FMA operations in AVX512. | ||
| 446 | 80 | const __mmask64 k_zh8 = 0x5555555555555555ULL; | |
| 447 | |||
| 448 |
8/8void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 305 → 127 taken 5 times.
✓ Branch 305 → 306 taken 1 time.
✓ Branch 309 → 127 taken 10 times.
✓ Branch 309 → 310 taken 2 times.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 183 → 117 taken 5 times.
✓ Branch 183 → 184 taken 1 time.
✓ Branch 187 → 117 taken 291 times.
✓ Branch 187 → 188 taken 76 times.
|
391 | for (int y = y_from; y < y_to; y++) |
| 449 | { | ||
| 450 | // 8 coeffs + 8 permute_idx + 2 src + 4 temporal + 1 rounder ~= 23 regs (permute2var overwrite first source - really may be more needed) | ||
| 451 | __m512i data_src, data_src2; | ||
| 452 | |||
| 453 | if constexpr (partial_load) { | ||
| 454 | // Safe masked loads for the image edge | ||
| 455 | 301 | data_src = _mm512_maskz_loadu_epi8(k1, src_ptr); | |
| 456 | 592 | data_src2 = _mm512_maskz_loadu_epi8(k2, src_ptr + 64); | |
| 457 | } | ||
| 458 | else { | ||
| 459 | // Fast unaligned loads for the safe zone | ||
| 460 | 10 | data_src = _mm512_loadu_si512(src_ptr); | |
| 461 | 15 | data_src2 = _mm512_loadu_si512(src_ptr + 64); | |
| 462 | } | ||
| 463 | |||
| 464 | __m512i src_r0r1_0_31lo, src_r0r1_0_31hi, src_r0r1_32_63lo, src_r0r1_32_63hi; | ||
| 465 | __m512i src_r2r3_0_31lo, src_r2r3_0_31hi, src_r2r3_32_63lo, src_r2r3_32_63hi; | ||
| 466 | if constexpr (UseVNNI) { | ||
| 467 | 296 | src_r0r1_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_r0r1_0_31lo, data_src2); | |
| 468 | 296 | src_r0r1_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_r0r1_0_31hi, data_src2); | |
| 469 | 296 | src_r0r1_32_63lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_r0r1_32_63lo, data_src2); | |
| 470 | 296 | src_r0r1_32_63hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_r0r1_32_63hi, data_src2); | |
| 471 | 296 | src_r2r3_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_r2r3_0_31lo, data_src2); | |
| 472 | 296 | src_r2r3_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_r2r3_0_31hi, data_src2); | |
| 473 | 296 | src_r2r3_32_63lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_r2r3_32_63lo, data_src2); | |
| 474 | 296 | src_r2r3_32_63hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_r2r3_32_63hi, data_src2); | |
| 475 | } else { | ||
| 476 | 15 | src_r0r1_0_31lo = _permutex2var_epi8_sim_get32(wi_r0r1_0_31lo, sa_r0r1_0_31lo, data_src, data_src2); | |
| 477 | 15 | src_r0r1_0_31hi = _permutex2var_epi8_sim_get32(wi_r0r1_0_31hi, sa_r0r1_0_31hi, data_src, data_src2); | |
| 478 | 15 | src_r0r1_32_63lo = _permutex2var_epi8_sim_get32(wi_r0r1_32_63lo, sa_r0r1_32_63lo, data_src, data_src2); | |
| 479 | 15 | src_r0r1_32_63hi = _permutex2var_epi8_sim_get32(wi_r0r1_32_63hi, sa_r0r1_32_63hi, data_src, data_src2); | |
| 480 | 15 | src_r2r3_0_31lo = _permutex2var_epi8_sim_get32(wi_r2r3_0_31lo, sa_r2r3_0_31lo, data_src, data_src2); | |
| 481 | 15 | src_r2r3_0_31hi = _permutex2var_epi8_sim_get32(wi_r2r3_0_31hi, sa_r2r3_0_31hi, data_src, data_src2); | |
| 482 | 15 | src_r2r3_32_63lo = _permutex2var_epi8_sim_get32(wi_r2r3_32_63lo, sa_r2r3_32_63lo, data_src, data_src2); | |
| 483 | 30 | src_r2r3_32_63hi = _permutex2var_epi8_sim_get32(wi_r2r3_32_63hi, sa_r2r3_32_63hi, data_src, data_src2); | |
| 484 | } | ||
| 485 | |||
| 486 | __m512i result_0_31lo, result_0_31hi; | ||
| 487 | __m512i result_32_63lo, result_32_63hi; | ||
| 488 | |||
| 489 | if constexpr (UseVNNI) | ||
| 490 | { | ||
| 491 | 592 | result_0_31lo = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31lo, coef_r0r1_0_31lo); | |
| 492 | 296 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r2r3_0_31lo, coef_r2r3_0_31lo); | |
| 493 | |||
| 494 | 592 | result_0_31hi = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31hi, coef_r0r1_0_31hi); | |
| 495 | 296 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r2r3_0_31hi, coef_r2r3_0_31hi); | |
| 496 | |||
| 497 | 592 | result_32_63lo = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63lo, coef_r0r1_32_63lo); | |
| 498 | 296 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r2r3_32_63lo, coef_r2r3_32_63lo); | |
| 499 | |||
| 500 | 592 | result_32_63hi = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63hi, coef_r0r1_32_63hi); | |
| 501 | 296 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r2r3_32_63hi, coef_r2r3_32_63hi); | |
| 502 | } | ||
| 503 | else | ||
| 504 | { | ||
| 505 | // making FMA in 32bits accs as in AVX256 V-resize | ||
| 506 | 45 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, coef_r0r1_0_31lo), _mm512_madd_epi16(src_r2r3_0_31lo, coef_r2r3_0_31lo)); | |
| 507 | 45 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, coef_r0r1_0_31hi), _mm512_madd_epi16(src_r2r3_0_31hi, coef_r2r3_0_31hi)); | |
| 508 | |||
| 509 | 45 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63lo, coef_r0r1_32_63lo), _mm512_madd_epi16(src_r2r3_32_63lo, coef_r2r3_32_63lo)); | |
| 510 | 30 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63hi, coef_r0r1_32_63hi), _mm512_madd_epi16(src_r2r3_32_63hi, coef_r2r3_32_63hi)); | |
| 511 | |||
| 512 | // rounding | ||
| 513 | 15 | result_0_31lo = _mm512_add_epi32(result_0_31lo, rounder); | |
| 514 | 15 | result_0_31hi = _mm512_add_epi32(result_0_31hi, rounder); | |
| 515 | 15 | result_32_63lo = _mm512_add_epi32(result_32_63lo, rounder); | |
| 516 | 30 | result_32_63hi = _mm512_add_epi32(result_32_63hi, rounder); | |
| 517 | } | ||
| 518 | |||
| 519 | // scaling down | ||
| 520 | 311 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale8bits); | |
| 521 | 311 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale8bits); | |
| 522 | 311 | result_32_63lo = _mm512_srai_epi32(result_32_63lo, FPScale8bits); | |
| 523 | 311 | result_32_63hi = _mm512_srai_epi32(result_32_63hi, FPScale8bits); | |
| 524 | |||
| 525 | 311 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 526 | 311 | __m512i result_32_63_int16 = _mm512_packus_epi32(result_32_63lo, result_32_63hi); | |
| 527 | |||
| 528 | // cast is enough, no need to _mm512_zextsi256_si512 | ||
| 529 | 1244 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), | |
| 530 | _mm512_inserti64x4(_mm512_castsi256_si512(_mm512_cvtusepi16_epi8(result_0_31_int16)), _mm512_cvtusepi16_epi8(result_32_63_int16), 1)); | ||
| 531 | |||
| 532 | 311 | dst_ptr += dst_pitch; | |
| 533 | 311 | src_ptr += src_pitch; | |
| 534 | } | ||
| 535 | |||
| 536 | // current_coeff += filter_size * PIXELS_AT_A_TIME; | ||
| 537 | 80 | current_coeff_SIMD += 8; // in number of __m512i | |
| 538 | |||
| 539 | }; | ||
| 540 | |||
| 541 | // Process the 'safe zone' where direct full unaligned loads are acceptable. | ||
| 542 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 1 time.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 1 time.
✓ Branch 18 → 19 taken 75 times.
|
78 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 543 | { | ||
| 544 |
1/2✓ Branch 16 → 17 taken 1 time.
✗ Branch 16 → 26 not taken.
|
2 | do_h_integer_core(std::false_type{}); |
| 545 | } | ||
| 546 | |||
| 547 | // Process the potentially 'unsafe zone' near the image edge, using safe masked loading. | ||
| 548 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 2 times.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 76 times.
✓ Branch 22 → 23 taken 75 times.
|
154 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 549 | { | ||
| 550 |
1/2✓ Branch 20 → 21 taken 76 times.
✗ Branch 20 → 26 not taken.
|
78 | do_h_integer_core(std::true_type{}); |
| 551 | } | ||
| 552 | } | ||
| 553 | 76 | } | |
| 554 | |||
| 555 | // filter size up to 8 | ||
| 556 | // 64 target uint8_t pixels at a time | ||
| 557 | // 128-byte source loads (128 uint8_t pixels) | ||
| 558 | // maximum permute index is 128 for _mm512_maskz_permutex2var_epi8 (uint8_t) | ||
| 559 | // support VNNI and madd FMA | ||
| 560 | |||
| 561 | // filter size up to 16 | ||
| 562 | // 32 target uint8_t pixels at a time | ||
| 563 | // 128-byte source loads (128 uint8_t pixels) | ||
| 564 | // maximum permute index is 128 for _mm512_maskz_permutex2var_epi8 (uint8_t) | ||
| 565 | |||
| 566 | // uint16_t | ||
| 567 | |||
| 568 | // filter size up to 4 | ||
| 569 | // 32 target uint16_t pixels at a time | ||
| 570 | // 128-byte source loads (64 uint16_t pixels) | ||
| 571 | // maximum permute index is 64 for _mm512_permutex2var_epi16 (uint16_t) | ||
| 572 | // making lo-hi unpacking with single permutex2var operation and optional VNNI FMA | ||
| 573 | |||
| 574 | // filter size up to 8 | ||
| 575 | // 64 target uint16_t pixels at a time in 4 groups of 16 | ||
| 576 | // 128-byte source loads (64 uint16_t pixels), 2 groups of 128 byte source loads for each 16 output samples, | ||
| 577 | // expect to support /2 downsize (used in ConvertTo42x with default bicubic too) | ||
| 578 | // maximum permute index is 64 for _mm512_permutex2var_epi16 (uint16_t) | ||
| 579 | // support VNNI and madd FMA | ||
| 580 | // checker: (16/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, 8/*kernel_size*/)) | ||
| 581 | |||
| 582 | template <bool AdvancePerm, bool UseVNNI> | ||
| 583 | AVS_FORCEINLINE static void process_row_pair( | ||
| 584 | // 1. Accumulators (Modified across calls) | ||
| 585 | __m512i& res_lo, __m512i& res_hi, | ||
| 586 | __m512i& res_32_63lo, __m512i& res_32_63hi, | ||
| 587 | // 2. Permutation Vectors (Modified across calls) | ||
| 588 | __m512i& p_lo, __m512i& p_hi, | ||
| 589 | __m512i& p_32_lo, __m512i& p_32_hi, | ||
| 590 | // 3. Coefficients (Read-only) | ||
| 591 | const __m512i& c_lo, const __m512i& c_hi, | ||
| 592 | const __m512i& c_32_63lo, const __m512i& c_32_63hi, | ||
| 593 | // 4. Source Data and Constants | ||
| 594 | const __m512i& d0_15, const __m512i& d16_31, | ||
| 595 | const __m512i& d32_47, const __m512i& d48_63, | ||
| 596 | const __m512i& d2_0_15, const __m512i& d2_16_31, | ||
| 597 | const __m512i& d2_32_47, const __m512i& d2_48_63, | ||
| 598 | const __mmask32 &k_hi, const __m512i &v_two) | ||
| 599 | { | ||
| 600 | // Generate source vectors via permutations | ||
| 601 | auto src_lo = _mm512_mask_blend_epi16(k_hi, | ||
| 602 | _mm512_permutex2var_epi16(d0_15, p_lo, d2_0_15), | ||
| 603 | _mm512_permutex2var_epi16(d16_31, p_lo, d2_16_31)); | ||
| 604 | |||
| 605 | auto src_hi = _mm512_mask_blend_epi16(k_hi, | ||
| 606 | _mm512_permutex2var_epi16(d0_15, p_hi, d2_0_15), | ||
| 607 | _mm512_permutex2var_epi16(d16_31, p_hi, d2_16_31)); | ||
| 608 | |||
| 609 | auto src_32_63lo = _mm512_mask_blend_epi16(k_hi, | ||
| 610 | _mm512_permutex2var_epi16(d32_47, p_32_lo, d2_32_47), | ||
| 611 | _mm512_permutex2var_epi16(d48_63, p_32_lo, d2_48_63)); | ||
| 612 | |||
| 613 | auto src_32_63hi = _mm512_mask_blend_epi16(k_hi, | ||
| 614 | _mm512_permutex2var_epi16(d32_47, p_32_hi, d2_32_47), | ||
| 615 | _mm512_permutex2var_epi16(d48_63, p_32_hi, d2_48_63)); | ||
| 616 | |||
| 617 | // Accumulate results | ||
| 618 | if constexpr (UseVNNI) { | ||
| 619 | res_lo = _mm512_dpwssd_epi32(res_lo, src_lo, c_lo); | ||
| 620 | res_hi = _mm512_dpwssd_epi32(res_hi, src_hi, c_hi); | ||
| 621 | res_32_63lo = _mm512_dpwssd_epi32(res_32_63lo, src_32_63lo, c_32_63lo); | ||
| 622 | res_32_63hi = _mm512_dpwssd_epi32(res_32_63hi, src_32_63hi, c_32_63hi); | ||
| 623 | } | ||
| 624 | else { | ||
| 625 | res_lo = _mm512_add_epi32(res_lo, _mm512_madd_epi16(src_lo, c_lo)); | ||
| 626 | res_hi = _mm512_add_epi32(res_hi, _mm512_madd_epi16(src_hi, c_hi)); | ||
| 627 | res_32_63lo = _mm512_add_epi32(res_32_63lo, _mm512_madd_epi16(src_32_63lo, c_32_63lo)); | ||
| 628 | res_32_63hi = _mm512_add_epi32(res_32_63hi, _mm512_madd_epi16(src_32_63hi, c_32_63hi)); | ||
| 629 | } | ||
| 630 | |||
| 631 | if constexpr (AdvancePerm) { | ||
| 632 | p_lo = _mm512_add_epi16(p_lo, v_two); | ||
| 633 | p_hi = _mm512_add_epi16(p_hi, v_two); | ||
| 634 | p_32_lo = _mm512_add_epi16(p_32_lo, v_two); | ||
| 635 | p_32_hi = _mm512_add_epi16(p_32_hi, v_two); | ||
| 636 | } | ||
| 637 | } | ||
| 638 | |||
| 639 | // filter size up to 4 | ||
| 640 | // 64 target uint16_t pixels at a time in 2 groups of 32 | ||
| 641 | // 128-byte source loads (64 uint16_t pixels) | ||
| 642 | // maximum permute index is 64 for _mm512_permutex2var_epi16 (uint16_t) | ||
| 643 | // making lo-hi unpacking with single permutex2var operation and optional VNNI FMA | ||
| 644 | |||
| 645 | // filter size up to 8 | ||
| 646 | // 64 target uint16_t pixels at a time in 2 groups of 32 | ||
| 647 | // 128-byte source loads (64 uint16_t pixels) | ||
| 648 | // maximum permute index is 64 for _mm512_permutex2var_epi16 (uint16_t) | ||
| 649 | // support VNNI and madd FMA | ||
| 650 | |||
| 651 | // filter size up to 8 | ||
| 652 | // 32 target uint16_t pixels at a time | ||
| 653 | // 128-byte source loads (64 uint16_t pixels), 2 groups of 128 byte source loads for each 16 output samples, expect to support /2 downsize (used in ConvertTo42x with default bicubic too) | ||
| 654 | // maximum permute index is 64 for _mm512_permutex2var_epi16 (uint16_t) | ||
| 655 | // support VNNI and madd FMA | ||
| 656 | // checker function is program->resize_h_planar_gather_permutex_vstripe_check(16/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, 8/*kernel_size*/)) | ||
| 657 | |||
| 658 | // filter size up to 8 | ||
| 659 | // 32 target uint16_t pixels at a time | ||
| 660 | // 128-byte source loads (64 uint16_t pixels) | ||
| 661 | // maximum permute index is 64 for _mm512_permutex2var_epi16 (uint16_t) | ||
| 662 | // support VNNI and madd FMA | ||
| 663 | |||
| 664 | // filter size up to 16 | ||
| 665 | // 32 target uint16_t pixels at a time | ||
| 666 | // 128-byte source loads (64 uint16_t pixels) | ||
| 667 | // maximum permute index is 64 for _mm512_permutex2var_epi16 (uint16_t) | ||
| 668 | // support VNNI and madd FMA | ||
| 669 | |||
| 670 | // filter size up to 8, pretransposed coefficients | ||
| 671 | // 64 target uint8_t pixels at a time | ||
| 672 | // 128-byte source loads (2x64 uint8_t pixels) | ||
| 673 | template<bool UseVNNI> | ||
| 674 | 5 | void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 675 | { | ||
| 676 | 5 | constexpr int PIXELS_AT_A_TIME = 64; | |
| 677 | |||
| 678 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 4 times.
✗ Branch 2 → 4 not taken.
|
5 | const int width_safe_mod = (program->safelimit_128_pixels_each64th_target.overread_possible ? program->safelimit_128_pixels_each64th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 679 | |||
| 680 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 4 times.
|
5 | assert(program->filter_size_real <= 8); |
| 681 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 4 times.
|
5 | assert(program->target_size_alignment >= 64); |
| 682 | assert(FRAME_ALIGN >= 64); | ||
| 683 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 4 times.
|
5 | assert(program->filter_size_alignment >= 8); |
| 684 | |||
| 685 | 5 | const int max_scanlines = program->max_scanlines; | |
| 686 | |||
| 687 | 5 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale8bits - 1)); | |
| 688 | |||
| 689 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 4 times.
✓ Branch 24 → 25 taken 4 times.
|
10 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 690 | { | ||
| 691 | 5 | int y_to = std::min(y_from + max_scanlines, height); | |
| 692 | |||
| 693 | 5 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 694 | |||
| 695 | 5 | int x = 0; | |
| 696 | |||
| 697 | 14 | auto do_h_integer_core = [&](auto partial_load) { | |
| 698 | 9 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 699 | |||
| 700 | 9 | const __m512i coef_r0r1_0_31lo = _mm512_load_si512(current_coeff_SIMD + 0); | |
| 701 | 9 | const __m512i coef_r0r1_0_31hi = _mm512_load_si512(current_coeff_SIMD + 1); | |
| 702 | 9 | const __m512i coef_r0r1_32_63lo = _mm512_load_si512(current_coeff_SIMD + 2); | |
| 703 | 9 | const __m512i coef_r0r1_32_63hi = _mm512_load_si512(current_coeff_SIMD + 3); | |
| 704 | 9 | const __m512i coef_r2r3_0_31lo = _mm512_load_si512(current_coeff_SIMD + 4); | |
| 705 | 9 | const __m512i coef_r2r3_0_31hi = _mm512_load_si512(current_coeff_SIMD + 5); | |
| 706 | 9 | const __m512i coef_r2r3_32_63lo = _mm512_load_si512(current_coeff_SIMD + 6); | |
| 707 | 9 | const __m512i coef_r2r3_32_63hi = _mm512_load_si512(current_coeff_SIMD + 7); | |
| 708 | 9 | const __m512i coef_r4r5_0_31lo = _mm512_load_si512(current_coeff_SIMD + 8); | |
| 709 | 9 | const __m512i coef_r4r5_0_31hi = _mm512_load_si512(current_coeff_SIMD + 9); | |
| 710 | 9 | const __m512i coef_r4r5_32_63lo = _mm512_load_si512(current_coeff_SIMD + 10); | |
| 711 | 9 | const __m512i coef_r4r5_32_63hi = _mm512_load_si512(current_coeff_SIMD + 11); | |
| 712 | 9 | const __m512i coef_r6r7_0_31lo = _mm512_load_si512(current_coeff_SIMD + 12); | |
| 713 | 9 | const __m512i coef_r6r7_0_31hi = _mm512_load_si512(current_coeff_SIMD + 13); | |
| 714 | 9 | const __m512i coef_r6r7_32_63lo = _mm512_load_si512(current_coeff_SIMD + 14); | |
| 715 | 9 | const __m512i coef_r6r7_32_63hi = _mm512_load_si512(current_coeff_SIMD + 15); | |
| 716 | |||
| 717 | 9 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 718 | 9 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 719 | 9 | __m512i perm_0_32_47 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 32])); | |
| 720 | 9 | __m512i perm_0_48_63 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 48])); | |
| 721 | |||
| 722 | 9 | int iStart = program->pixel_offset[x]; | |
| 723 | 9 | __m512i m512i_Start = _mm512_set1_epi32(iStart); | |
| 724 | |||
| 725 | 9 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, m512i_Start); | |
| 726 | 9 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, m512i_Start); | |
| 727 | 9 | perm_0_32_47 = _mm512_sub_epi32(perm_0_32_47, m512i_Start); | |
| 728 | 9 | perm_0_48_63 = _mm512_sub_epi32(perm_0_48_63, m512i_Start); | |
| 729 | |||
| 730 | 9 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 731 | 9 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 732 | 9 | __m256i m256i_perm_0_32_47 = _mm512_cvtepi32_epi16(perm_0_32_47); | |
| 733 | 9 | __m256i m256i_perm_0_48_63 = _mm512_cvtepi32_epi16(perm_0_48_63); | |
| 734 | |||
| 735 | 18 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 736 | 18 | __m512i perm_0_32_63 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_32_47), m256i_perm_0_48_63, 1); | |
| 737 | |||
| 738 | 9 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 739 | 9 | __m512i perm_1_32_63 = _mm512_add_epi16(perm_0_32_63, one_epi16); | |
| 740 | |||
| 741 | 9 | const __m512i perm_r0r1_0_31lo = _mm512_unpacklo_epi16(perm_0_0_31, perm_1_0_31); | |
| 742 | 9 | const __m512i perm_r0r1_0_31hi = _mm512_unpackhi_epi16(perm_0_0_31, perm_1_0_31); | |
| 743 | 9 | const __m512i perm_r0r1_32_63lo = _mm512_unpacklo_epi16(perm_0_32_63, perm_1_32_63); | |
| 744 | 9 | const __m512i perm_r0r1_32_63hi = _mm512_unpackhi_epi16(perm_0_32_63, perm_1_32_63); | |
| 745 | |||
| 746 | 9 | const __m512i two_epi16 = _mm512_set1_epi16(2); | |
| 747 | |||
| 748 | 9 | __m512i wi_r0r1_0_31lo={},sa_r0r1_0_31lo={},wi_r0r1_0_31hi={},sa_r0r1_0_31hi={}; | |
| 749 | 9 | __m512i wi_r0r1_32_63lo={},sa_r0r1_32_63lo={},wi_r0r1_32_63hi={},sa_r0r1_32_63hi={}; | |
| 750 | 9 | __m512i wi_r2r3_0_31lo={},sa_r2r3_0_31lo={},wi_r2r3_0_31hi={},sa_r2r3_0_31hi={}; | |
| 751 | 9 | __m512i wi_r2r3_32_63lo={},sa_r2r3_32_63lo={},wi_r2r3_32_63hi={},sa_r2r3_32_63hi={}; | |
| 752 | 9 | __m512i wi_r4r5_0_31lo={},sa_r4r5_0_31lo={},wi_r4r5_0_31hi={},sa_r4r5_0_31hi={}; | |
| 753 | 9 | __m512i wi_r4r5_32_63lo={},sa_r4r5_32_63lo={},wi_r4r5_32_63hi={},sa_r4r5_32_63hi={}; | |
| 754 | 9 | __m512i wi_r6r7_0_31lo={},sa_r6r7_0_31lo={},wi_r6r7_0_31hi={},sa_r6r7_0_31hi={}; | |
| 755 | 9 | __m512i wi_r6r7_32_63lo={},sa_r6r7_32_63lo={},wi_r6r7_32_63hi={},sa_r6r7_32_63hi={}; | |
| 756 | if constexpr (!UseVNNI) { | ||
| 757 | 3 | const __m512i c_8w = _mm512_set1_epi16(8); | |
| 758 | 51 | auto make_wi_sa = [&](__m512i p, __m512i& wi, __m512i& sa) { | |
| 759 | 48 | wi = _mm512_srli_epi16(p, 1); | |
| 760 | 96 | sa = _mm512_and_si512(_mm512_slli_epi16(p, 3), c_8w); | |
| 761 | }; | ||
| 762 | 3 | __m512i p_lo = perm_r0r1_0_31lo, p_hi = perm_r0r1_0_31hi; | |
| 763 | 3 | __m512i p_lo2 = perm_r0r1_32_63lo, p_hi2 = perm_r0r1_32_63hi; | |
| 764 | 3 | make_wi_sa(p_lo, wi_r0r1_0_31lo, sa_r0r1_0_31lo); make_wi_sa(p_hi, wi_r0r1_0_31hi, sa_r0r1_0_31hi); | |
| 765 | 3 | make_wi_sa(p_lo2, wi_r0r1_32_63lo, sa_r0r1_32_63lo); make_wi_sa(p_hi2, wi_r0r1_32_63hi, sa_r0r1_32_63hi); | |
| 766 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 767 | 3 | p_lo2 = _mm512_add_epi16(p_lo2, two_epi16); p_hi2 = _mm512_add_epi16(p_hi2, two_epi16); | |
| 768 | 3 | make_wi_sa(p_lo, wi_r2r3_0_31lo, sa_r2r3_0_31lo); make_wi_sa(p_hi, wi_r2r3_0_31hi, sa_r2r3_0_31hi); | |
| 769 | 3 | make_wi_sa(p_lo2, wi_r2r3_32_63lo, sa_r2r3_32_63lo); make_wi_sa(p_hi2, wi_r2r3_32_63hi, sa_r2r3_32_63hi); | |
| 770 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 771 | 3 | p_lo2 = _mm512_add_epi16(p_lo2, two_epi16); p_hi2 = _mm512_add_epi16(p_hi2, two_epi16); | |
| 772 | 3 | make_wi_sa(p_lo, wi_r4r5_0_31lo, sa_r4r5_0_31lo); make_wi_sa(p_hi, wi_r4r5_0_31hi, sa_r4r5_0_31hi); | |
| 773 | 3 | make_wi_sa(p_lo2, wi_r4r5_32_63lo, sa_r4r5_32_63lo); make_wi_sa(p_hi2, wi_r4r5_32_63hi, sa_r4r5_32_63hi); | |
| 774 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 775 | 3 | p_lo2 = _mm512_add_epi16(p_lo2, two_epi16); p_hi2 = _mm512_add_epi16(p_hi2, two_epi16); | |
| 776 | 3 | make_wi_sa(p_lo, wi_r6r7_0_31lo, sa_r6r7_0_31lo); make_wi_sa(p_hi, wi_r6r7_0_31hi, sa_r6r7_0_31hi); | |
| 777 | 3 | make_wi_sa(p_lo2, wi_r6r7_32_63lo, sa_r6r7_32_63lo); make_wi_sa(p_hi2, wi_r6r7_32_63hi, sa_r6r7_32_63hi); | |
| 778 | } | ||
| 779 | |||
| 780 | 9 | uint8_t* AVS_RESTRICT dst_ptr = dst8 + x + y_from * dst_pitch; | |
| 781 | 9 | const uint8_t* src_ptr = src8 + iStart + y_from * src_pitch; | |
| 782 | |||
| 783 | 9 | const int remaining = program->source_size - iStart; | |
| 784 | 9 | __mmask64 k1 = _bzhi_u64(~0ULL, remaining); | |
| 785 | 9 | const __mmask64 k2 = _bzhi_u64(~0ULL, std::max(0, remaining - 64)); | |
| 786 | 9 | const __mmask64 k_zh8 = 0x5555555555555555ULL; | |
| 787 | |||
| 788 |
8/8void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 499 → 153 taken 5 times.
✓ Branch 499 → 500 taken 1 time.
✓ Branch 503 → 153 taken 10 times.
✓ Branch 503 → 504 taken 2 times.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 233 → 111 taken 5 times.
✓ Branch 233 → 234 taken 1 time.
✓ Branch 237 → 111 taken 37 times.
✓ Branch 237 → 238 taken 5 times.
|
66 | for (int y = y_from; y < y_to; y++) |
| 789 | { | ||
| 790 | __m512i data_src, data_src2; | ||
| 791 | |||
| 792 | 57 | __m512i perm_rNrNp1_0_31lo_w = perm_r0r1_0_31lo; | |
| 793 | 57 | __m512i perm_rNrNp1_0_31hi_w = perm_r0r1_0_31hi; | |
| 794 | 57 | __m512i perm_rNrNp1_32_63lo_w = perm_r0r1_32_63lo; | |
| 795 | 57 | __m512i perm_rNrNp1_32_63hi_w = perm_r0r1_32_63hi; | |
| 796 | |||
| 797 | if constexpr (partial_load) { | ||
| 798 | 47 | data_src = _mm512_maskz_loadu_epi8(k1, src_ptr); | |
| 799 | 84 | data_src2 = _mm512_maskz_loadu_epi8(k2, src_ptr + 64); | |
| 800 | } | ||
| 801 | else { | ||
| 802 | 10 | data_src = _mm512_loadu_si512(src_ptr); | |
| 803 | 15 | data_src2 = _mm512_loadu_si512(src_ptr + 64); | |
| 804 | } | ||
| 805 | |||
| 806 | // rows 0..1 | ||
| 807 | __m512i src_r0r1_0_31lo, src_r0r1_0_31hi, src_r0r1_32_63lo, src_r0r1_32_63hi; | ||
| 808 | if constexpr (UseVNNI) { | ||
| 809 | 42 | src_r0r1_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 810 | 42 | src_r0r1_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 811 | 42 | src_r0r1_32_63lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_32_63lo_w, data_src2); | |
| 812 | 42 | src_r0r1_32_63hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_32_63hi_w, data_src2); | |
| 813 | } else { | ||
| 814 | 15 | src_r0r1_0_31lo = _permutex2var_epi8_sim_get32(wi_r0r1_0_31lo, sa_r0r1_0_31lo, data_src, data_src2); | |
| 815 | 15 | src_r0r1_0_31hi = _permutex2var_epi8_sim_get32(wi_r0r1_0_31hi, sa_r0r1_0_31hi, data_src, data_src2); | |
| 816 | 15 | src_r0r1_32_63lo = _permutex2var_epi8_sim_get32(wi_r0r1_32_63lo, sa_r0r1_32_63lo, data_src, data_src2); | |
| 817 | 30 | src_r0r1_32_63hi = _permutex2var_epi8_sim_get32(wi_r0r1_32_63hi, sa_r0r1_32_63hi, data_src, data_src2); | |
| 818 | } | ||
| 819 | |||
| 820 | // for r2r3 | ||
| 821 | 57 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 822 | 57 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 823 | 57 | perm_rNrNp1_32_63lo_w = _mm512_add_epi16(perm_rNrNp1_32_63lo_w, two_epi16); | |
| 824 | 57 | perm_rNrNp1_32_63hi_w = _mm512_add_epi16(perm_rNrNp1_32_63hi_w, two_epi16); | |
| 825 | |||
| 826 | __m512i src_r2r3_0_31lo, src_r2r3_0_31hi, src_r2r3_32_63lo, src_r2r3_32_63hi; | ||
| 827 | if constexpr (UseVNNI) { | ||
| 828 | 42 | src_r2r3_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 829 | 42 | src_r2r3_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 830 | 42 | src_r2r3_32_63lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_32_63lo_w, data_src2); | |
| 831 | 42 | src_r2r3_32_63hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_32_63hi_w, data_src2); | |
| 832 | } else { | ||
| 833 | 15 | src_r2r3_0_31lo = _permutex2var_epi8_sim_get32(wi_r2r3_0_31lo, sa_r2r3_0_31lo, data_src, data_src2); | |
| 834 | 15 | src_r2r3_0_31hi = _permutex2var_epi8_sim_get32(wi_r2r3_0_31hi, sa_r2r3_0_31hi, data_src, data_src2); | |
| 835 | 15 | src_r2r3_32_63lo = _permutex2var_epi8_sim_get32(wi_r2r3_32_63lo, sa_r2r3_32_63lo, data_src, data_src2); | |
| 836 | 30 | src_r2r3_32_63hi = _permutex2var_epi8_sim_get32(wi_r2r3_32_63hi, sa_r2r3_32_63hi, data_src, data_src2); | |
| 837 | } | ||
| 838 | |||
| 839 | // for r4r5 | ||
| 840 | 57 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 841 | 57 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 842 | 57 | perm_rNrNp1_32_63lo_w = _mm512_add_epi16(perm_rNrNp1_32_63lo_w, two_epi16); | |
| 843 | 57 | perm_rNrNp1_32_63hi_w = _mm512_add_epi16(perm_rNrNp1_32_63hi_w, two_epi16); | |
| 844 | |||
| 845 | __m512i result_0_31lo, result_0_31hi; | ||
| 846 | __m512i result_32_63lo, result_32_63hi; | ||
| 847 | |||
| 848 | if constexpr (UseVNNI) | ||
| 849 | { | ||
| 850 | 84 | result_0_31lo = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31lo, coef_r0r1_0_31lo); | |
| 851 | 42 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r2r3_0_31lo, coef_r2r3_0_31lo); | |
| 852 | 84 | result_0_31hi = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31hi, coef_r0r1_0_31hi); | |
| 853 | 42 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r2r3_0_31hi, coef_r2r3_0_31hi); | |
| 854 | 84 | result_32_63lo = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63lo, coef_r0r1_32_63lo); | |
| 855 | 42 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r2r3_32_63lo, coef_r2r3_32_63lo); | |
| 856 | 84 | result_32_63hi = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63hi, coef_r0r1_32_63hi); | |
| 857 | 42 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r2r3_32_63hi, coef_r2r3_32_63hi); | |
| 858 | } | ||
| 859 | else | ||
| 860 | { | ||
| 861 | 45 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, coef_r0r1_0_31lo), _mm512_madd_epi16(src_r2r3_0_31lo, coef_r2r3_0_31lo)); | |
| 862 | 45 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, coef_r0r1_0_31hi), _mm512_madd_epi16(src_r2r3_0_31hi, coef_r2r3_0_31hi)); | |
| 863 | 45 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63lo, coef_r0r1_32_63lo), _mm512_madd_epi16(src_r2r3_32_63lo, coef_r2r3_32_63lo)); | |
| 864 | 30 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63hi, coef_r0r1_32_63hi), _mm512_madd_epi16(src_r2r3_32_63hi, coef_r2r3_32_63hi)); | |
| 865 | } | ||
| 866 | |||
| 867 | // rows 4..5 | ||
| 868 | __m512i src_r4r5_0_31lo, src_r4r5_0_31hi, src_r4r5_32_63lo, src_r4r5_32_63hi; | ||
| 869 | if constexpr (UseVNNI) { | ||
| 870 | 42 | src_r4r5_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 871 | 42 | src_r4r5_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 872 | 42 | src_r4r5_32_63lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_32_63lo_w, data_src2); | |
| 873 | 42 | src_r4r5_32_63hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_32_63hi_w, data_src2); | |
| 874 | } else { | ||
| 875 | 15 | src_r4r5_0_31lo = _permutex2var_epi8_sim_get32(wi_r4r5_0_31lo, sa_r4r5_0_31lo, data_src, data_src2); | |
| 876 | 15 | src_r4r5_0_31hi = _permutex2var_epi8_sim_get32(wi_r4r5_0_31hi, sa_r4r5_0_31hi, data_src, data_src2); | |
| 877 | 15 | src_r4r5_32_63lo = _permutex2var_epi8_sim_get32(wi_r4r5_32_63lo, sa_r4r5_32_63lo, data_src, data_src2); | |
| 878 | 30 | src_r4r5_32_63hi = _permutex2var_epi8_sim_get32(wi_r4r5_32_63hi, sa_r4r5_32_63hi, data_src, data_src2); | |
| 879 | } | ||
| 880 | |||
| 881 | // for r6r7 | ||
| 882 | 57 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 883 | 57 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 884 | 57 | perm_rNrNp1_32_63lo_w = _mm512_add_epi16(perm_rNrNp1_32_63lo_w, two_epi16); | |
| 885 | 57 | perm_rNrNp1_32_63hi_w = _mm512_add_epi16(perm_rNrNp1_32_63hi_w, two_epi16); | |
| 886 | |||
| 887 | __m512i src_r6r7_0_31lo, src_r6r7_0_31hi, src_r6r7_32_63lo, src_r6r7_32_63hi; | ||
| 888 | if constexpr (UseVNNI) { | ||
| 889 | 42 | src_r6r7_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 890 | 42 | src_r6r7_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 891 | 42 | src_r6r7_32_63lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_32_63lo_w, data_src2); | |
| 892 | 42 | src_r6r7_32_63hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_32_63hi_w, data_src2); | |
| 893 | } else { | ||
| 894 | 15 | src_r6r7_0_31lo = _permutex2var_epi8_sim_get32(wi_r6r7_0_31lo, sa_r6r7_0_31lo, data_src, data_src2); | |
| 895 | 15 | src_r6r7_0_31hi = _permutex2var_epi8_sim_get32(wi_r6r7_0_31hi, sa_r6r7_0_31hi, data_src, data_src2); | |
| 896 | 15 | src_r6r7_32_63lo = _permutex2var_epi8_sim_get32(wi_r6r7_32_63lo, sa_r6r7_32_63lo, data_src, data_src2); | |
| 897 | 30 | src_r6r7_32_63hi = _permutex2var_epi8_sim_get32(wi_r6r7_32_63hi, sa_r6r7_32_63hi, data_src, data_src2); | |
| 898 | } | ||
| 899 | |||
| 900 | if constexpr (UseVNNI) | ||
| 901 | { | ||
| 902 | 42 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r4r5_0_31lo, coef_r4r5_0_31lo); | |
| 903 | 42 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r6r7_0_31lo, coef_r6r7_0_31lo); | |
| 904 | 42 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r4r5_0_31hi, coef_r4r5_0_31hi); | |
| 905 | 42 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r6r7_0_31hi, coef_r6r7_0_31hi); | |
| 906 | 42 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r4r5_32_63lo, coef_r4r5_32_63lo); | |
| 907 | 42 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r6r7_32_63lo, coef_r6r7_32_63lo); | |
| 908 | 42 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r4r5_32_63hi, coef_r4r5_32_63hi); | |
| 909 | 42 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r6r7_32_63hi, coef_r6r7_32_63hi); | |
| 910 | // rounding VNNI in first FMA already summed | ||
| 911 | } | ||
| 912 | else | ||
| 913 | { | ||
| 914 | 30 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_0_31lo, coef_r4r5_0_31lo), result_0_31lo); | |
| 915 | 30 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_0_31hi, coef_r4r5_0_31hi), result_0_31hi); | |
| 916 | 30 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_32_63lo, coef_r4r5_32_63lo), result_32_63lo); | |
| 917 | 30 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_32_63hi, coef_r4r5_32_63hi), result_32_63hi); | |
| 918 | |||
| 919 | 30 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_0_31lo, coef_r6r7_0_31lo), result_0_31lo); | |
| 920 | 30 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_0_31hi, coef_r6r7_0_31hi), result_0_31hi); | |
| 921 | 30 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_32_63lo, coef_r6r7_32_63lo), result_32_63lo); | |
| 922 | 15 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_32_63hi, coef_r6r7_32_63hi), result_32_63hi); | |
| 923 | |||
| 924 | 15 | result_0_31lo = _mm512_add_epi32(result_0_31lo, rounder); | |
| 925 | 15 | result_0_31hi = _mm512_add_epi32(result_0_31hi, rounder); | |
| 926 | 15 | result_32_63lo = _mm512_add_epi32(result_32_63lo, rounder); | |
| 927 | 30 | result_32_63hi = _mm512_add_epi32(result_32_63hi, rounder); | |
| 928 | } | ||
| 929 | |||
| 930 | 57 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale8bits); | |
| 931 | 57 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale8bits); | |
| 932 | 57 | result_32_63lo = _mm512_srai_epi32(result_32_63lo, FPScale8bits); | |
| 933 | 57 | result_32_63hi = _mm512_srai_epi32(result_32_63hi, FPScale8bits); | |
| 934 | |||
| 935 | 57 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 936 | 57 | __m512i result_32_63_int16 = _mm512_packus_epi32(result_32_63lo, result_32_63hi); | |
| 937 | |||
| 938 | 57 | __m256i result_0_31_u8 = _mm512_cvtusepi16_epi8(result_0_31_int16); | |
| 939 | 57 | __m256i result_32_63_u8 = _mm512_cvtusepi16_epi8(result_32_63_int16); | |
| 940 | |||
| 941 | 114 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), _mm512_inserti64x4(_mm512_castsi256_si512(result_0_31_u8), result_32_63_u8, 1)); | |
| 942 | |||
| 943 | 57 | dst_ptr += dst_pitch; | |
| 944 | 57 | src_ptr += src_pitch; | |
| 945 | } | ||
| 946 | |||
| 947 | 9 | current_coeff_SIMD += 16; | |
| 948 | }; | ||
| 949 | |||
| 950 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 1 time.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 1 time.
✓ Branch 18 → 19 taken 4 times.
|
7 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 951 | { | ||
| 952 |
1/2✓ Branch 16 → 17 taken 1 time.
✗ Branch 16 → 26 not taken.
|
2 | do_h_integer_core(std::false_type{}); |
| 953 | } | ||
| 954 | |||
| 955 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 2 times.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 5 times.
✓ Branch 22 → 23 taken 4 times.
|
12 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 956 | { | ||
| 957 |
1/2✓ Branch 20 → 21 taken 5 times.
✗ Branch 20 → 26 not taken.
|
7 | do_h_integer_core(std::true_type{}); |
| 958 | } | ||
| 959 | } | ||
| 960 | 5 | } | |
| 961 | |||
| 962 | // filter size up to 16, pretransposed coefficients | ||
| 963 | // 32 target uint8_t pixels at a time | ||
| 964 | // 128-byte source loads (2x64 uint8_t pixels) | ||
| 965 | template<bool UseVNNI> | ||
| 966 | 11 | void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 967 | { | ||
| 968 | 11 | constexpr int PIXELS_AT_A_TIME = 32; | |
| 969 | |||
| 970 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 10 times.
✗ Branch 2 → 4 not taken.
|
11 | const int width_safe_mod = (program->safelimit_128_pixels_each64th_target.overread_possible ? program->safelimit_128_pixels_each64th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 971 | |||
| 972 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 10 times.
|
11 | assert(program->filter_size_real <= 16); |
| 973 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 10 times.
|
11 | assert(program->target_size_alignment >= 32); |
| 974 | assert(FRAME_ALIGN >= 32); | ||
| 975 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 10 times.
|
11 | assert(program->filter_size_alignment >= 16); |
| 976 | |||
| 977 | 11 | const int max_scanlines = program->max_scanlines; | |
| 978 | |||
| 979 | 11 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale8bits - 1)); | |
| 980 | |||
| 981 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 10 times.
✓ Branch 24 → 25 taken 10 times.
|
22 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 982 | { | ||
| 983 | 11 | int y_to = std::min(y_from + max_scanlines, height); | |
| 984 | |||
| 985 | 11 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 986 | |||
| 987 | 11 | int x = 0; | |
| 988 | |||
| 989 | 32 | auto do_h_integer_core = [&](auto partial_load) { | |
| 990 | 21 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 991 | |||
| 992 | 21 | const __m512i coef_r0r1_0_31lo = _mm512_load_si512(current_coeff_SIMD + 0); | |
| 993 | 21 | const __m512i coef_r0r1_0_31hi = _mm512_load_si512(current_coeff_SIMD + 1); | |
| 994 | 21 | const __m512i coef_r2r3_0_31lo = _mm512_load_si512(current_coeff_SIMD + 2); | |
| 995 | 21 | const __m512i coef_r2r3_0_31hi = _mm512_load_si512(current_coeff_SIMD + 3); | |
| 996 | 21 | const __m512i coef_r4r5_0_31lo = _mm512_load_si512(current_coeff_SIMD + 4); | |
| 997 | 21 | const __m512i coef_r4r5_0_31hi = _mm512_load_si512(current_coeff_SIMD + 5); | |
| 998 | 21 | const __m512i coef_r6r7_0_31lo = _mm512_load_si512(current_coeff_SIMD + 6); | |
| 999 | 21 | const __m512i coef_r6r7_0_31hi = _mm512_load_si512(current_coeff_SIMD + 7); | |
| 1000 | 21 | const __m512i coef_r8r9_0_31lo = _mm512_load_si512(current_coeff_SIMD + 8); | |
| 1001 | 21 | const __m512i coef_r8r9_0_31hi = _mm512_load_si512(current_coeff_SIMD + 9); | |
| 1002 | 21 | const __m512i coef_r10r11_0_31lo = _mm512_load_si512(current_coeff_SIMD + 10); | |
| 1003 | 21 | const __m512i coef_r10r11_0_31hi = _mm512_load_si512(current_coeff_SIMD + 11); | |
| 1004 | 21 | const __m512i coef_r12r13_0_31lo = _mm512_load_si512(current_coeff_SIMD + 12); | |
| 1005 | 21 | const __m512i coef_r12r13_0_31hi = _mm512_load_si512(current_coeff_SIMD + 13); | |
| 1006 | 21 | const __m512i coef_r14r15_0_31lo = _mm512_load_si512(current_coeff_SIMD + 14); | |
| 1007 | 21 | const __m512i coef_r14r15_0_31hi = _mm512_load_si512(current_coeff_SIMD + 15); | |
| 1008 | |||
| 1009 | 21 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 1010 | 21 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 1011 | |||
| 1012 | 21 | int iStart = program->pixel_offset[x]; | |
| 1013 | 21 | __m512i m512i_Start = _mm512_set1_epi32(iStart); | |
| 1014 | |||
| 1015 | 21 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, m512i_Start); | |
| 1016 | 21 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, m512i_Start); | |
| 1017 | |||
| 1018 | 21 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 1019 | 21 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 1020 | |||
| 1021 | 42 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 1022 | 21 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 1023 | |||
| 1024 | 21 | const __m512i perm_r0r1_0_31lo = _mm512_unpacklo_epi16(perm_0_0_31, perm_1_0_31); | |
| 1025 | 21 | const __m512i perm_r0r1_0_31hi = _mm512_unpackhi_epi16(perm_0_0_31, perm_1_0_31); | |
| 1026 | |||
| 1027 | 21 | const __m512i two_epi16 = _mm512_set1_epi16(2); | |
| 1028 | |||
| 1029 | // BASE path: precompute word-index (wi) and shift-amount (sa) for each tap-pair perm. | ||
| 1030 | // perm_rNrNp1 holds 32 x 16-bit byte addresses; wi = addr>>1 selects the 16-bit word in | ||
| 1031 | // {data_src, data_src2}, sa = (addr<<3)&8 selects high (1) or low (0) byte within that word. | ||
| 1032 | // Precomputing here eliminates ~12 instructions/gather (cvtepu8+extract+slli+srli+maskz_mov) | ||
| 1033 | // from the hot y-loop, leaving only the 3-instruction sim_get32 core per tap-pair. | ||
| 1034 | 21 | __m512i wi_r0r1_lo={},sa_r0r1_lo={},wi_r0r1_hi={},sa_r0r1_hi={}; | |
| 1035 | 21 | __m512i wi_r2r3_lo={},sa_r2r3_lo={},wi_r2r3_hi={},sa_r2r3_hi={}; | |
| 1036 | 21 | __m512i wi_r4r5_lo={},sa_r4r5_lo={},wi_r4r5_hi={},sa_r4r5_hi={}; | |
| 1037 | 21 | __m512i wi_r6r7_lo={},sa_r6r7_lo={},wi_r6r7_hi={},sa_r6r7_hi={}; | |
| 1038 | 21 | __m512i wi_r8r9_lo={},sa_r8r9_lo={},wi_r8r9_hi={},sa_r8r9_hi={}; | |
| 1039 | 21 | __m512i wi_r10r11_lo={},sa_r10r11_lo={},wi_r10r11_hi={},sa_r10r11_hi={}; | |
| 1040 | 21 | __m512i wi_r12r13_lo={},sa_r12r13_lo={},wi_r12r13_hi={},sa_r12r13_hi={}; | |
| 1041 | 21 | __m512i wi_r14r15_lo={},sa_r14r15_lo={},wi_r14r15_hi={},sa_r14r15_hi={}; | |
| 1042 | if constexpr (!UseVNNI) { | ||
| 1043 | 6 | const __m512i c_8w = _mm512_set1_epi16(8); | |
| 1044 | 102 | auto make_wi_sa = [&](__m512i p, __m512i& wi, __m512i& sa) { | |
| 1045 | 96 | wi = _mm512_srli_epi16(p, 1); | |
| 1046 | 192 | sa = _mm512_and_si512(_mm512_slli_epi16(p, 3), c_8w); | |
| 1047 | }; | ||
| 1048 | 6 | __m512i p_lo = perm_r0r1_0_31lo, p_hi = perm_r0r1_0_31hi; | |
| 1049 | 6 | make_wi_sa(p_lo, wi_r0r1_lo, sa_r0r1_lo); make_wi_sa(p_hi, wi_r0r1_hi, sa_r0r1_hi); | |
| 1050 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 1051 | 6 | make_wi_sa(p_lo, wi_r2r3_lo, sa_r2r3_lo); make_wi_sa(p_hi, wi_r2r3_hi, sa_r2r3_hi); | |
| 1052 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 1053 | 6 | make_wi_sa(p_lo, wi_r4r5_lo, sa_r4r5_lo); make_wi_sa(p_hi, wi_r4r5_hi, sa_r4r5_hi); | |
| 1054 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 1055 | 6 | make_wi_sa(p_lo, wi_r6r7_lo, sa_r6r7_lo); make_wi_sa(p_hi, wi_r6r7_hi, sa_r6r7_hi); | |
| 1056 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 1057 | 6 | make_wi_sa(p_lo, wi_r8r9_lo, sa_r8r9_lo); make_wi_sa(p_hi, wi_r8r9_hi, sa_r8r9_hi); | |
| 1058 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 1059 | 6 | make_wi_sa(p_lo, wi_r10r11_lo, sa_r10r11_lo); make_wi_sa(p_hi, wi_r10r11_hi, sa_r10r11_hi); | |
| 1060 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 1061 | 6 | make_wi_sa(p_lo, wi_r12r13_lo, sa_r12r13_lo); make_wi_sa(p_hi, wi_r12r13_hi, sa_r12r13_hi); | |
| 1062 | 6 | p_lo = _mm512_add_epi16(p_lo, two_epi16); p_hi = _mm512_add_epi16(p_hi, two_epi16); | |
| 1063 | 6 | make_wi_sa(p_lo, wi_r14r15_lo, sa_r14r15_lo); make_wi_sa(p_hi, wi_r14r15_hi, sa_r14r15_hi); | |
| 1064 | } | ||
| 1065 | |||
| 1066 | 21 | uint8_t* AVS_RESTRICT dst_ptr = dst8 + x + y_from * dst_pitch; | |
| 1067 | 21 | const uint8_t* src_ptr = src8 + iStart + y_from * src_pitch; | |
| 1068 | |||
| 1069 | 21 | const int remaining = program->source_size - iStart; | |
| 1070 | 21 | __mmask64 k1 = _bzhi_u64(~0ULL, remaining); | |
| 1071 | 21 | const __mmask64 k2 = _bzhi_u64(~0ULL, std::max(0, remaining - 64)); | |
| 1072 | 21 | const __mmask64 k_zh8 = 0x5555555555555555ULL; | |
| 1073 | |||
| 1074 |
8/8void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 459 → 125 taken 10 times.
✓ Branch 459 → 460 taken 2 times.
✓ Branch 463 → 125 taken 20 times.
✓ Branch 463 → 464 taken 4 times.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 189 → 79 taken 10 times.
✓ Branch 189 → 190 taken 2 times.
✓ Branch 193 → 79 taken 101 times.
✓ Branch 193 → 194 taken 13 times.
|
162 | for (int y = y_from; y < y_to; y++) |
| 1075 | { | ||
| 1076 | __m512i data_src, data_src2; | ||
| 1077 | |||
| 1078 | 141 | __m512i perm_rNrNp1_0_31lo_w = perm_r0r1_0_31lo; | |
| 1079 | 141 | __m512i perm_rNrNp1_0_31hi_w = perm_r0r1_0_31hi; | |
| 1080 | |||
| 1081 | if constexpr (partial_load) { | ||
| 1082 | 121 | data_src = _mm512_maskz_loadu_epi8(k1, src_ptr); | |
| 1083 | 222 | data_src2 = _mm512_maskz_loadu_epi8(k2, src_ptr + 64); | |
| 1084 | } | ||
| 1085 | else { | ||
| 1086 | 20 | data_src = _mm512_loadu_si512(src_ptr); | |
| 1087 | 30 | data_src2 = _mm512_loadu_si512(src_ptr + 64); | |
| 1088 | } | ||
| 1089 | |||
| 1090 | // rows 0..1 | ||
| 1091 | __m512i src_r0r1_0_31lo, src_r0r1_0_31hi; | ||
| 1092 | if constexpr (UseVNNI) { | ||
| 1093 | 111 | src_r0r1_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 1094 | 111 | src_r0r1_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 1095 | } else { | ||
| 1096 | 30 | src_r0r1_0_31lo = _permutex2var_epi8_sim_get32(wi_r0r1_lo, sa_r0r1_lo, data_src, data_src2); | |
| 1097 | 60 | src_r0r1_0_31hi = _permutex2var_epi8_sim_get32(wi_r0r1_hi, sa_r0r1_hi, data_src, data_src2); | |
| 1098 | } | ||
| 1099 | |||
| 1100 | // for r2r3 | ||
| 1101 | 141 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 1102 | 141 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 1103 | |||
| 1104 | __m512i src_r2r3_0_31lo, src_r2r3_0_31hi; | ||
| 1105 | if constexpr (UseVNNI) { | ||
| 1106 | 111 | src_r2r3_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 1107 | 111 | src_r2r3_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 1108 | } else { | ||
| 1109 | 30 | src_r2r3_0_31lo = _permutex2var_epi8_sim_get32(wi_r2r3_lo, sa_r2r3_lo, data_src, data_src2); | |
| 1110 | 60 | src_r2r3_0_31hi = _permutex2var_epi8_sim_get32(wi_r2r3_hi, sa_r2r3_hi, data_src, data_src2); | |
| 1111 | } | ||
| 1112 | |||
| 1113 | // for r4r5 | ||
| 1114 | 141 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 1115 | 141 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 1116 | |||
| 1117 | __m512i result_0_31lo, result_0_31hi; | ||
| 1118 | |||
| 1119 | if constexpr (UseVNNI) | ||
| 1120 | { | ||
| 1121 | 222 | result_0_31lo = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31lo, coef_r0r1_0_31lo); | |
| 1122 | 111 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r2r3_0_31lo, coef_r2r3_0_31lo); | |
| 1123 | 222 | result_0_31hi = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31hi, coef_r0r1_0_31hi); | |
| 1124 | 111 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r2r3_0_31hi, coef_r2r3_0_31hi); | |
| 1125 | } | ||
| 1126 | else | ||
| 1127 | { | ||
| 1128 | 90 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, coef_r0r1_0_31lo), _mm512_madd_epi16(src_r2r3_0_31lo, coef_r2r3_0_31lo)); | |
| 1129 | 60 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, coef_r0r1_0_31hi), _mm512_madd_epi16(src_r2r3_0_31hi, coef_r2r3_0_31hi)); | |
| 1130 | } | ||
| 1131 | |||
| 1132 | // rows 4..5 | ||
| 1133 | __m512i src_r4r5_0_31lo, src_r4r5_0_31hi; | ||
| 1134 | if constexpr (UseVNNI) { | ||
| 1135 | 111 | src_r4r5_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 1136 | 111 | src_r4r5_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 1137 | } else { | ||
| 1138 | 30 | src_r4r5_0_31lo = _permutex2var_epi8_sim_get32(wi_r4r5_lo, sa_r4r5_lo, data_src, data_src2); | |
| 1139 | 60 | src_r4r5_0_31hi = _permutex2var_epi8_sim_get32(wi_r4r5_hi, sa_r4r5_hi, data_src, data_src2); | |
| 1140 | } | ||
| 1141 | |||
| 1142 | // for r6r7 | ||
| 1143 | 141 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 1144 | 141 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 1145 | |||
| 1146 | __m512i src_r6r7_0_31lo, src_r6r7_0_31hi; | ||
| 1147 | if constexpr (UseVNNI) { | ||
| 1148 | 111 | src_r6r7_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 1149 | 111 | src_r6r7_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 1150 | } else { | ||
| 1151 | 30 | src_r6r7_0_31lo = _permutex2var_epi8_sim_get32(wi_r6r7_lo, sa_r6r7_lo, data_src, data_src2); | |
| 1152 | 60 | src_r6r7_0_31hi = _permutex2var_epi8_sim_get32(wi_r6r7_hi, sa_r6r7_hi, data_src, data_src2); | |
| 1153 | } | ||
| 1154 | |||
| 1155 | // for r8r9 | ||
| 1156 | 141 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 1157 | 141 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 1158 | |||
| 1159 | if constexpr (UseVNNI) | ||
| 1160 | { | ||
| 1161 | 111 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r4r5_0_31lo, coef_r4r5_0_31lo); | |
| 1162 | 111 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r6r7_0_31lo, coef_r6r7_0_31lo); | |
| 1163 | 111 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r4r5_0_31hi, coef_r4r5_0_31hi); | |
| 1164 | 111 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r6r7_0_31hi, coef_r6r7_0_31hi); | |
| 1165 | } | ||
| 1166 | else | ||
| 1167 | { | ||
| 1168 | 60 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_0_31lo, coef_r4r5_0_31lo), result_0_31lo); | |
| 1169 | 60 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_0_31hi, coef_r4r5_0_31hi), result_0_31hi); | |
| 1170 | 60 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_0_31lo, coef_r6r7_0_31lo), result_0_31lo); | |
| 1171 | 30 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_0_31hi, coef_r6r7_0_31hi), result_0_31hi); | |
| 1172 | } | ||
| 1173 | |||
| 1174 | // rows 8..9 | ||
| 1175 | __m512i src_r8r9_0_31lo, src_r8r9_0_31hi; | ||
| 1176 | if constexpr (UseVNNI) { | ||
| 1177 | 111 | src_r8r9_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 1178 | 111 | src_r8r9_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 1179 | } else { | ||
| 1180 | 30 | src_r8r9_0_31lo = _permutex2var_epi8_sim_get32(wi_r8r9_lo, sa_r8r9_lo, data_src, data_src2); | |
| 1181 | 60 | src_r8r9_0_31hi = _permutex2var_epi8_sim_get32(wi_r8r9_hi, sa_r8r9_hi, data_src, data_src2); | |
| 1182 | } | ||
| 1183 | |||
| 1184 | // for r10r11 | ||
| 1185 | 141 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 1186 | 141 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 1187 | |||
| 1188 | __m512i src_r10r11_0_31lo, src_r10r11_0_31hi; | ||
| 1189 | if constexpr (UseVNNI) { | ||
| 1190 | 111 | src_r10r11_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 1191 | 111 | src_r10r11_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 1192 | } else { | ||
| 1193 | 30 | src_r10r11_0_31lo = _permutex2var_epi8_sim_get32(wi_r10r11_lo, sa_r10r11_lo, data_src, data_src2); | |
| 1194 | 60 | src_r10r11_0_31hi = _permutex2var_epi8_sim_get32(wi_r10r11_hi, sa_r10r11_hi, data_src, data_src2); | |
| 1195 | } | ||
| 1196 | |||
| 1197 | // for r12r13 | ||
| 1198 | 141 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 1199 | 141 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 1200 | |||
| 1201 | if constexpr (UseVNNI) | ||
| 1202 | { | ||
| 1203 | 111 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r8r9_0_31lo, coef_r8r9_0_31lo); | |
| 1204 | 111 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r10r11_0_31lo, coef_r10r11_0_31lo); | |
| 1205 | 111 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r8r9_0_31hi, coef_r8r9_0_31hi); | |
| 1206 | 111 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r10r11_0_31hi, coef_r10r11_0_31hi); | |
| 1207 | } | ||
| 1208 | else | ||
| 1209 | { | ||
| 1210 | 60 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r8r9_0_31lo, coef_r8r9_0_31lo), result_0_31lo); | |
| 1211 | 60 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r8r9_0_31hi, coef_r8r9_0_31hi), result_0_31hi); | |
| 1212 | 60 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r10r11_0_31lo, coef_r10r11_0_31lo), result_0_31lo); | |
| 1213 | 30 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r10r11_0_31hi, coef_r10r11_0_31hi), result_0_31hi); | |
| 1214 | } | ||
| 1215 | |||
| 1216 | // rows 12..13 | ||
| 1217 | __m512i src_r12r13_0_31lo, src_r12r13_0_31hi; | ||
| 1218 | if constexpr (UseVNNI) { | ||
| 1219 | 111 | src_r12r13_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 1220 | 111 | src_r12r13_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 1221 | } else { | ||
| 1222 | 30 | src_r12r13_0_31lo = _permutex2var_epi8_sim_get32(wi_r12r13_lo, sa_r12r13_lo, data_src, data_src2); | |
| 1223 | 60 | src_r12r13_0_31hi = _permutex2var_epi8_sim_get32(wi_r12r13_hi, sa_r12r13_hi, data_src, data_src2); | |
| 1224 | } | ||
| 1225 | |||
| 1226 | // for r14r15 | ||
| 1227 | 141 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 1228 | 141 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 1229 | |||
| 1230 | __m512i src_r14r15_0_31lo, src_r14r15_0_31hi; | ||
| 1231 | if constexpr (UseVNNI) { | ||
| 1232 | 111 | src_r14r15_0_31lo = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 1233 | 111 | src_r14r15_0_31hi = _mm512_maskz_permutex2var_epi8(k_zh8, data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 1234 | } else { | ||
| 1235 | 30 | src_r14r15_0_31lo = _permutex2var_epi8_sim_get32(wi_r14r15_lo, sa_r14r15_lo, data_src, data_src2); | |
| 1236 | 60 | src_r14r15_0_31hi = _permutex2var_epi8_sim_get32(wi_r14r15_hi, sa_r14r15_hi, data_src, data_src2); | |
| 1237 | } | ||
| 1238 | |||
| 1239 | if constexpr (UseVNNI) | ||
| 1240 | { | ||
| 1241 | 111 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r12r13_0_31lo, coef_r12r13_0_31lo); | |
| 1242 | 111 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r14r15_0_31lo, coef_r14r15_0_31lo); | |
| 1243 | 111 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r12r13_0_31hi, coef_r12r13_0_31hi); | |
| 1244 | 111 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r14r15_0_31hi, coef_r14r15_0_31hi); | |
| 1245 | // rounding VNNI in first FMA already summed | ||
| 1246 | } | ||
| 1247 | else | ||
| 1248 | { | ||
| 1249 | 60 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r12r13_0_31lo, coef_r12r13_0_31lo), result_0_31lo); | |
| 1250 | 60 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r12r13_0_31hi, coef_r12r13_0_31hi), result_0_31hi); | |
| 1251 | 60 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r14r15_0_31lo, coef_r14r15_0_31lo), result_0_31lo); | |
| 1252 | 30 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r14r15_0_31hi, coef_r14r15_0_31hi), result_0_31hi); | |
| 1253 | |||
| 1254 | 30 | result_0_31lo = _mm512_add_epi32(result_0_31lo, rounder); | |
| 1255 | 60 | result_0_31hi = _mm512_add_epi32(result_0_31hi, rounder); | |
| 1256 | } | ||
| 1257 | |||
| 1258 | 141 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale8bits); | |
| 1259 | 141 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale8bits); | |
| 1260 | |||
| 1261 | 141 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 1262 | |||
| 1263 | 141 | __m256i result_0_31_u8 = _mm512_cvtusepi16_epi8(result_0_31_int16); | |
| 1264 | |||
| 1265 | _mm256_stream_si256(reinterpret_cast<__m256i*>(dst_ptr), result_0_31_u8); | ||
| 1266 | |||
| 1267 | 141 | dst_ptr += dst_pitch; | |
| 1268 | 141 | src_ptr += src_pitch; | |
| 1269 | } | ||
| 1270 | |||
| 1271 | 21 | current_coeff_SIMD += 16; | |
| 1272 | }; | ||
| 1273 | |||
| 1274 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 2 times.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 2 times.
✓ Branch 18 → 19 taken 10 times.
|
15 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 1275 | { | ||
| 1276 |
1/2✓ Branch 16 → 17 taken 2 times.
✗ Branch 16 → 26 not taken.
|
4 | do_h_integer_core(std::false_type{}); |
| 1277 | } | ||
| 1278 | |||
| 1279 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 4 times.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 13 times.
✓ Branch 22 → 23 taken 10 times.
|
28 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 1280 | { | ||
| 1281 |
1/2✓ Branch 20 → 21 taken 13 times.
✗ Branch 20 → 26 not taken.
|
17 | do_h_integer_core(std::true_type{}); |
| 1282 | } | ||
| 1283 | } | ||
| 1284 | 11 | } | |
| 1285 | |||
| 1286 | // filter size up to 64, pretransposed coefficients | ||
| 1287 | // 64 target uint8_t pixels at a time in 2 groups of 32 | ||
| 1288 | // 2 groups of 128-byte source loads | ||
| 1289 | template<bool UseVNNI> | ||
| 1290 | 2 | void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 1291 | { | ||
| 1292 | 2 | int filter_size_real = program->filter_size_real; | |
| 1293 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
|
2 | if ((filter_size_real / 2 * 2) != filter_size_real) filter_size_real++; |
| 1294 | |||
| 1295 | 2 | constexpr int PIXELS_AT_A_TIME = 64; | |
| 1296 | |||
| 1297 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 4 → 5 taken 1 time.
✗ Branch 4 → 6 not taken.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 4 → 5 taken 1 time.
✗ Branch 4 → 6 not taken.
|
2 | const int width_safe_mod = (program->safelimit_128_pixels_each64th_target.overread_possible ? program->safelimit_128_pixels_each64th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 1298 | |||
| 1299 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
|
2 | assert(program->filter_size_real <= 64); |
| 1300 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1 time.
|
2 | assert(program->target_size_alignment >= 64); |
| 1301 | assert(FRAME_ALIGN >= 64); | ||
| 1302 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 11 → 12 not taken.
✓ Branch 11 → 13 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 11 → 12 not taken.
✓ Branch 11 → 13 taken 1 time.
|
2 | assert(program->filter_size_alignment >= 8); |
| 1303 | |||
| 1304 | 2 | const int max_scanlines = program->max_scanlines; | |
| 1305 | |||
| 1306 | 2 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale8bits - 1)); | |
| 1307 | |||
| 1308 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 26 → 16 taken 1 time.
✓ Branch 26 → 27 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 26 → 16 taken 1 time.
✓ Branch 26 → 27 taken 1 time.
|
4 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 1309 | { | ||
| 1310 | 2 | int y_to = std::min(y_from + max_scanlines, height); | |
| 1311 | |||
| 1312 | 2 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 1313 | |||
| 1314 | 2 | int x = 0; | |
| 1315 | |||
| 1316 | 14 | auto do_h_integer_core = [&](auto partial_load) { | |
| 1317 | |||
| 1318 | 6 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 1319 | 6 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 1320 | 6 | __m512i perm_0_32_47 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 32])); | |
| 1321 | 6 | __m512i perm_0_48_63 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 48])); | |
| 1322 | |||
| 1323 | 6 | int iStart = program->pixel_offset[x]; | |
| 1324 | 6 | int iStart_2 = program->pixel_offset[x + 32]; | |
| 1325 | 6 | __m512i m512i_Start = _mm512_set1_epi32(iStart); | |
| 1326 | 6 | __m512i m512i_Start_2 = _mm512_set1_epi32(iStart_2); | |
| 1327 | |||
| 1328 | 6 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, m512i_Start); | |
| 1329 | 6 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, m512i_Start); | |
| 1330 | 6 | perm_0_32_47 = _mm512_sub_epi32(perm_0_32_47, m512i_Start_2); | |
| 1331 | 6 | perm_0_48_63 = _mm512_sub_epi32(perm_0_48_63, m512i_Start_2); | |
| 1332 | |||
| 1333 | 6 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 1334 | 6 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 1335 | 6 | __m256i m256i_perm_0_32_47 = _mm512_cvtepi32_epi16(perm_0_32_47); | |
| 1336 | 6 | __m256i m256i_perm_0_48_63 = _mm512_cvtepi32_epi16(perm_0_48_63); | |
| 1337 | |||
| 1338 | 12 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 1339 | 12 | __m512i perm_0_32_63 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_32_47), m256i_perm_0_48_63, 1); | |
| 1340 | |||
| 1341 | 6 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 1342 | 6 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 1343 | 6 | __m512i perm_1_32_63 = _mm512_add_epi16(perm_0_32_63, one_epi16); | |
| 1344 | |||
| 1345 | 6 | const __m512i two_epi16 = _mm512_set1_epi16(2); | |
| 1346 | |||
| 1347 | // BASE: precompute wi/sa outside y-loop. perm vectors are epi16 (32 16-bit byte indices). | ||
| 1348 | // Each kr step the perm advances by +2 bytes = +1 word, so wi += 1 and sa is invariant. | ||
| 1349 | // perm_1 = perm_0 + 1: wi_p1[i] = wi_p0[i] + (perm_0[i] & 1), sa_p1 = sa_p0 ^ 8. | ||
| 1350 | 6 | __m512i wi_g1_p0_base = {}, wi_g2_p0_base = {}, wi_g1_p1_base = {}, wi_g2_p1_base = {}; | |
| 1351 | 6 | __m512i sa_g1_p0 = {}, sa_g2_p0 = {}, sa_g1_p1 = {}, sa_g2_p1 = {}; | |
| 1352 | if constexpr (!UseVNNI) { | ||
| 1353 | 3 | const __m512i c_8w = _mm512_set1_epi16(8); | |
| 1354 | 6 | sa_g1_p0 = _mm512_and_si512(_mm512_slli_epi16(perm_0_0_31, 3), c_8w); | |
| 1355 | 3 | wi_g1_p0_base = _mm512_srli_epi16(perm_0_0_31, 1); | |
| 1356 | 3 | sa_g1_p1 = _mm512_xor_si512(sa_g1_p0, c_8w); | |
| 1357 | 6 | wi_g1_p1_base = _mm512_add_epi16(wi_g1_p0_base, _mm512_srli_epi16(sa_g1_p0, 3)); | |
| 1358 | 6 | sa_g2_p0 = _mm512_and_si512(_mm512_slli_epi16(perm_0_32_63, 3), c_8w); | |
| 1359 | 3 | wi_g2_p0_base = _mm512_srli_epi16(perm_0_32_63, 1); | |
| 1360 | 3 | sa_g2_p1 = _mm512_xor_si512(sa_g2_p0, c_8w); | |
| 1361 | 3 | wi_g2_p1_base = _mm512_add_epi16(wi_g2_p0_base, _mm512_srli_epi16(sa_g2_p0, 3)); | |
| 1362 | } | ||
| 1363 | |||
| 1364 | 6 | uint8_t* AVS_RESTRICT dst_ptr = dst8 + x + y_from * dst_pitch; | |
| 1365 | 6 | const uint8_t* src_ptr = src8 + iStart + y_from * src_pitch; | |
| 1366 | 6 | const uint8_t* src_ptr_2 = src8 + iStart_2 + y_from * src_pitch; | |
| 1367 | |||
| 1368 | 6 | const int remaining = program->source_size - iStart; | |
| 1369 | 6 | const int remaining_2 = program->source_size - iStart_2; | |
| 1370 | 6 | const __mmask64 k1 = _bzhi_u64(~0ULL, remaining); | |
| 1371 | 6 | const __mmask64 k2 = _bzhi_u64(~0ULL, std::max(0, remaining - 64)); | |
| 1372 | 6 | const __mmask64 k1_2 = _bzhi_u64(~0ULL, remaining_2); | |
| 1373 | 6 | const __mmask64 k2_2 = _bzhi_u64(~0ULL, std::max(0, remaining_2 - 64)); | |
| 1374 | 6 | const __mmask64 k_zh8 = 0x5555555555555555ULL; | |
| 1375 | |||
| 1376 |
8/8void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 242 → 97 taken 5 times.
✓ Branch 242 → 243 taken 1 time.
✓ Branch 250 → 97 taken 10 times.
✓ Branch 250 → 251 taken 2 times.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 168 → 71 taken 5 times.
✓ Branch 168 → 169 taken 1 time.
✓ Branch 176 → 71 taken 10 times.
✓ Branch 176 → 177 taken 2 times.
|
36 | for (int y = y_from; y < y_to; y++) |
| 1377 | { | ||
| 1378 | __m512i data_src, data_src2, data_src_2, data_src2_2; | ||
| 1379 | |||
| 1380 | // VBMI: reset working perm copies per row; BASE: reset running wi counters per row | ||
| 1381 | 30 | __m512i perm_0_0_31w = {}, perm_0_32_63w = {}, perm_1_0_31w = {}, perm_1_32_63w = {}; | |
| 1382 | 30 | __m512i wi_g1_p0 = {}, wi_g2_p0 = {}, wi_g1_p1 = {}, wi_g2_p1 = {}; | |
| 1383 | if constexpr (UseVNNI) { | ||
| 1384 | 15 | perm_0_0_31w = perm_0_0_31; | |
| 1385 | 15 | perm_0_32_63w = perm_0_32_63; | |
| 1386 | 15 | perm_1_0_31w = perm_1_0_31; | |
| 1387 | 15 | perm_1_32_63w = perm_1_32_63; | |
| 1388 | } else { | ||
| 1389 | 15 | wi_g1_p0 = wi_g1_p0_base; | |
| 1390 | 15 | wi_g2_p0 = wi_g2_p0_base; | |
| 1391 | 15 | wi_g1_p1 = wi_g1_p1_base; | |
| 1392 | 15 | wi_g2_p1 = wi_g2_p1_base; | |
| 1393 | } | ||
| 1394 | |||
| 1395 | if constexpr (partial_load) { | ||
| 1396 | 20 | data_src = _mm512_maskz_loadu_epi8(k1, src_ptr); | |
| 1397 | 40 | data_src2 = _mm512_maskz_loadu_epi8(k2, src_ptr + 64); | |
| 1398 | 20 | data_src_2 = _mm512_maskz_loadu_epi8(k1_2, src_ptr_2); | |
| 1399 | 20 | data_src2_2 = _mm512_maskz_loadu_epi8(k2_2, src_ptr_2 + 64); | |
| 1400 | } | ||
| 1401 | else { | ||
| 1402 | 10 | data_src = _mm512_loadu_si512(src_ptr); | |
| 1403 | 20 | data_src2 = _mm512_loadu_si512(src_ptr + 64); | |
| 1404 | 10 | data_src_2 = _mm512_loadu_si512(src_ptr_2); | |
| 1405 | 10 | data_src2_2 = _mm512_loadu_si512(src_ptr_2 + 64); | |
| 1406 | } | ||
| 1407 | |||
| 1408 | 30 | __m512i result_0_31lo = rounder; | |
| 1409 | 30 | __m512i result_0_31hi = rounder; | |
| 1410 | 30 | __m512i result_32_63lo = rounder; | |
| 1411 | 30 | __m512i result_32_63hi = rounder; | |
| 1412 | |||
| 1413 | 30 | const __m512i* current_coeff_SIMDw = current_coeff_SIMD; | |
| 1414 | |||
| 1415 |
8/8void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 211 → 106 taken 50 times.
✓ Branch 211 → 212 taken 5 times.
✓ Branch 219 → 114 taken 100 times.
✓ Branch 219 → 220 taken 10 times.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 137 → 80 taken 50 times.
✓ Branch 137 → 138 taken 5 times.
✓ Branch 145 → 88 taken 100 times.
✓ Branch 145 → 146 taken 10 times.
|
330 | for (int kr = 0; kr < filter_size_real; kr += 2) |
| 1416 | { | ||
| 1417 | __m512i src_r0_0_31, src_r0_32_63, src_r1_0_31, src_r1_32_63; | ||
| 1418 | if constexpr (UseVNNI) { | ||
| 1419 | 150 | src_r0_0_31 = _mm512_maskz_permutex2var_epi8_SIMUL<true>(k_zh8, data_src, perm_0_0_31w, data_src2); | |
| 1420 | 150 | src_r0_32_63 = _mm512_maskz_permutex2var_epi8_SIMUL<true>(k_zh8, data_src_2, perm_0_32_63w, data_src2_2); | |
| 1421 | 150 | src_r1_0_31 = _mm512_maskz_permutex2var_epi8_SIMUL<true>(k_zh8, data_src, perm_1_0_31w, data_src2); | |
| 1422 | 150 | src_r1_32_63 = _mm512_maskz_permutex2var_epi8_SIMUL<true>(k_zh8, data_src_2, perm_1_32_63w, data_src2_2); | |
| 1423 | 150 | perm_0_0_31w = _mm512_add_epi16(perm_0_0_31w, two_epi16); | |
| 1424 | 150 | perm_0_32_63w = _mm512_add_epi16(perm_0_32_63w, two_epi16); | |
| 1425 | 150 | perm_1_0_31w = _mm512_add_epi16(perm_1_0_31w, two_epi16); | |
| 1426 | 150 | perm_1_32_63w = _mm512_add_epi16(perm_1_32_63w, two_epi16); | |
| 1427 | } else { | ||
| 1428 | // BASE: sim_get32 reads word wi (already indexed into [data_src | data_src2]) | ||
| 1429 | // and shifts by sa to extract the target byte; avoids full SIMUL recomputation | ||
| 1430 | 150 | src_r0_0_31 = _permutex2var_epi8_sim_get32(wi_g1_p0, sa_g1_p0, data_src, data_src2); | |
| 1431 | 150 | src_r0_32_63 = _permutex2var_epi8_sim_get32(wi_g2_p0, sa_g2_p0, data_src_2, data_src2_2); | |
| 1432 | 150 | src_r1_0_31 = _permutex2var_epi8_sim_get32(wi_g1_p1, sa_g1_p1, data_src, data_src2); | |
| 1433 | 150 | src_r1_32_63 = _permutex2var_epi8_sim_get32(wi_g2_p1, sa_g2_p1, data_src_2, data_src2_2); | |
| 1434 | 150 | wi_g1_p0 = _mm512_add_epi16(wi_g1_p0, one_epi16); | |
| 1435 | 150 | wi_g2_p0 = _mm512_add_epi16(wi_g2_p0, one_epi16); | |
| 1436 | 150 | wi_g1_p1 = _mm512_add_epi16(wi_g1_p1, one_epi16); | |
| 1437 | 150 | wi_g2_p1 = _mm512_add_epi16(wi_g2_p1, one_epi16); | |
| 1438 | } | ||
| 1439 | |||
| 1440 | 300 | __m512i src_r0r1_0_31lo = _mm512_unpacklo_epi16(src_r0_0_31, src_r1_0_31); | |
| 1441 | 300 | __m512i src_r0r1_0_31hi = _mm512_unpackhi_epi16(src_r0_0_31, src_r1_0_31); | |
| 1442 | 300 | __m512i src_r0r1_32_63lo = _mm512_unpacklo_epi16(src_r0_32_63, src_r1_32_63); | |
| 1443 | 300 | __m512i src_r0r1_32_63hi = _mm512_unpackhi_epi16(src_r0_32_63, src_r1_32_63); | |
| 1444 | |||
| 1445 | if constexpr (UseVNNI) | ||
| 1446 | { | ||
| 1447 | 150 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r0r1_0_31lo, _mm512_load_si512(current_coeff_SIMDw + 0)); | |
| 1448 | 300 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r0r1_0_31hi, _mm512_load_si512(current_coeff_SIMDw + 1)); | |
| 1449 | 300 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r0r1_32_63lo, _mm512_load_si512(current_coeff_SIMDw + 2)); | |
| 1450 | 300 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r0r1_32_63hi, _mm512_load_si512(current_coeff_SIMDw + 3)); | |
| 1451 | } | ||
| 1452 | else | ||
| 1453 | { | ||
| 1454 | 300 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, _mm512_load_si512(current_coeff_SIMDw + 0)), result_0_31lo); | |
| 1455 | 450 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, _mm512_load_si512(current_coeff_SIMDw + 1)), result_0_31hi); | |
| 1456 | 450 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63lo, _mm512_load_si512(current_coeff_SIMDw + 2)), result_32_63lo); | |
| 1457 | 450 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63hi, _mm512_load_si512(current_coeff_SIMDw + 3)), result_32_63hi); | |
| 1458 | } | ||
| 1459 | |||
| 1460 | 300 | current_coeff_SIMDw += 4; | |
| 1461 | } | ||
| 1462 | |||
| 1463 | 30 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale8bits); | |
| 1464 | 30 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale8bits); | |
| 1465 | 30 | result_32_63lo = _mm512_srai_epi32(result_32_63lo, FPScale8bits); | |
| 1466 | 30 | result_32_63hi = _mm512_srai_epi32(result_32_63hi, FPScale8bits); | |
| 1467 | |||
| 1468 | 30 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 1469 | 30 | __m512i result_32_63_int16 = _mm512_packus_epi32(result_32_63lo, result_32_63hi); | |
| 1470 | |||
| 1471 | 30 | __m256i result_0_31_u8 = _mm512_cvtusepi16_epi8(result_0_31_int16); | |
| 1472 | 30 | __m256i result_32_63_u8 = _mm512_cvtusepi16_epi8(result_32_63_int16); | |
| 1473 | |||
| 1474 | 60 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), _mm512_inserti64x4(_mm512_castsi256_si512(result_0_31_u8), result_32_63_u8, 1)); | |
| 1475 | |||
| 1476 | 30 | dst_ptr += dst_pitch; | |
| 1477 | 30 | src_ptr += src_pitch; | |
| 1478 | 30 | src_ptr_2 += src_pitch; | |
| 1479 | } | ||
| 1480 | |||
| 1481 | 6 | current_coeff_SIMD += filter_size_real * 2; | |
| 1482 | }; | ||
| 1483 | |||
| 1484 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 20 → 18 taken 1 time.
✓ Branch 20 → 21 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 20 → 18 taken 1 time.
✓ Branch 20 → 21 taken 1 time.
|
4 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 1485 | { | ||
| 1486 |
1/2✓ Branch 18 → 19 taken 1 time.
✗ Branch 18 → 28 not taken.
|
2 | do_h_integer_core(std::false_type{}); |
| 1487 | } | ||
| 1488 | |||
| 1489 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 2 times.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 2 times.
✓ Branch 24 → 25 taken 1 time.
|
6 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 1490 | { | ||
| 1491 |
1/2✓ Branch 22 → 23 taken 2 times.
✗ Branch 22 → 28 not taken.
|
4 | do_h_integer_core(std::true_type{}); |
| 1492 | } | ||
| 1493 | } | ||
| 1494 | 2 | } | |
| 1495 | |||
| 1496 | // filter size 8, pretransposed coefficients | ||
| 1497 | // 64 target uint8_t pixels at a time in 2 groups of 32 | ||
| 1498 | // 2 groups of 128-byte source loads | ||
| 1499 | template<bool UseVNNI> | ||
| 1500 | 2 | void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 1501 | { | ||
| 1502 | 2 | constexpr int PIXELS_AT_A_TIME = 64; | |
| 1503 | |||
| 1504 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
|
2 | const int width_safe_mod = (program->safelimit_128_pixels_each64th_target.overread_possible ? program->safelimit_128_pixels_each64th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 1505 | |||
| 1506 | 2 | const int max_scanlines = program->max_scanlines; | |
| 1507 | |||
| 1508 | 2 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale8bits - 1)); | |
| 1509 | |||
| 1510 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 8 taken 1 time.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 8 taken 1 time.
✓ Branch 18 → 19 taken 1 time.
|
4 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 1511 | { | ||
| 1512 | 2 | int y_to = std::min(y_from + max_scanlines, height); | |
| 1513 | |||
| 1514 | 2 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 1515 | |||
| 1516 | 2 | int x = 0; | |
| 1517 | |||
| 1518 | 6 | auto do_h_integer_core = [&](auto partial_load) { | |
| 1519 | 2 | const __m512i coef_r0r1_0_31lo = _mm512_load_si512(current_coeff_SIMD + 0); | |
| 1520 | 2 | const __m512i coef_r0r1_0_31hi = _mm512_load_si512(current_coeff_SIMD + 1); | |
| 1521 | 2 | const __m512i coef_r0r1_32_63lo = _mm512_load_si512(current_coeff_SIMD + 2); | |
| 1522 | 2 | const __m512i coef_r0r1_32_63hi = _mm512_load_si512(current_coeff_SIMD + 3); | |
| 1523 | 2 | const __m512i coef_r2r3_0_31lo = _mm512_load_si512(current_coeff_SIMD + 4); | |
| 1524 | 2 | const __m512i coef_r2r3_0_31hi = _mm512_load_si512(current_coeff_SIMD + 5); | |
| 1525 | 2 | const __m512i coef_r2r3_32_63lo = _mm512_load_si512(current_coeff_SIMD + 6); | |
| 1526 | 2 | const __m512i coef_r2r3_32_63hi = _mm512_load_si512(current_coeff_SIMD + 7); | |
| 1527 | 2 | const __m512i coef_r4r5_0_31lo = _mm512_load_si512(current_coeff_SIMD + 8); | |
| 1528 | 2 | const __m512i coef_r4r5_0_31hi = _mm512_load_si512(current_coeff_SIMD + 9); | |
| 1529 | 2 | const __m512i coef_r4r5_32_63lo = _mm512_load_si512(current_coeff_SIMD + 10); | |
| 1530 | 2 | const __m512i coef_r4r5_32_63hi = _mm512_load_si512(current_coeff_SIMD + 11); | |
| 1531 | 2 | const __m512i coef_r6r7_0_31lo = _mm512_load_si512(current_coeff_SIMD + 12); | |
| 1532 | 2 | const __m512i coef_r6r7_0_31hi = _mm512_load_si512(current_coeff_SIMD + 13); | |
| 1533 | 2 | const __m512i coef_r6r7_32_63lo = _mm512_load_si512(current_coeff_SIMD + 14); | |
| 1534 | 2 | const __m512i coef_r6r7_32_63hi = _mm512_load_si512(current_coeff_SIMD + 15); | |
| 1535 | |||
| 1536 | 2 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 1537 | 2 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 1538 | 2 | __m512i perm_0_32_47 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 32])); | |
| 1539 | 2 | __m512i perm_0_48_63 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 48])); | |
| 1540 | |||
| 1541 | 2 | int iStart = program->pixel_offset[x]; | |
| 1542 | 2 | int iStart_2 = program->pixel_offset[x + 32]; | |
| 1543 | 2 | __m512i m512i_Start = _mm512_set1_epi32(iStart); | |
| 1544 | 2 | __m512i m512i_Start_2 = _mm512_set1_epi32(iStart_2); | |
| 1545 | |||
| 1546 | 2 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, m512i_Start); | |
| 1547 | 2 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, m512i_Start); | |
| 1548 | 2 | perm_0_32_47 = _mm512_sub_epi32(perm_0_32_47, m512i_Start_2); | |
| 1549 | 2 | perm_0_48_63 = _mm512_sub_epi32(perm_0_48_63, m512i_Start_2); | |
| 1550 | |||
| 1551 | 2 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 1552 | 2 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 1553 | 2 | __m256i m256i_perm_0_32_47 = _mm512_cvtepi32_epi16(perm_0_32_47); | |
| 1554 | 2 | __m256i m256i_perm_0_48_63 = _mm512_cvtepi32_epi16(perm_0_48_63); | |
| 1555 | |||
| 1556 | 2 | __m128i mm128i_perm_0_0_15 = _mm256_cvtepi16_epi8(m256i_perm_0_0_15); | |
| 1557 | 2 | __m128i mm128i_perm_0_16_31 = _mm256_cvtepi16_epi8(m256i_perm_0_16_31); | |
| 1558 | 2 | __m128i mm128i_perm_0_32_47 = _mm256_cvtepi16_epi8(m256i_perm_0_32_47); | |
| 1559 | 2 | __m128i mm128i_perm_0_48_63 = _mm256_cvtepi16_epi8(m256i_perm_0_48_63); | |
| 1560 | |||
| 1561 | 2 | __m512i perm_0 = _mm512_inserti32x4(_mm512_castsi128_si512(mm128i_perm_0_0_15), mm128i_perm_0_16_31, 1); | |
| 1562 | 2 | perm_0 = _mm512_inserti32x4(perm_0, mm128i_perm_0_32_47, 2); | |
| 1563 | 2 | perm_0 = _mm512_inserti32x4(perm_0, mm128i_perm_0_48_63, 3); | |
| 1564 | |||
| 1565 | 2 | uint8_t* AVS_RESTRICT dst_ptr = dst8 + x + y_from * dst_pitch; | |
| 1566 | 2 | const uint8_t* src_ptr = src8 + iStart + y_from * src_pitch; | |
| 1567 | 2 | const uint8_t* src_ptr_2 = src8 + iStart_2 + y_from * src_pitch; | |
| 1568 | |||
| 1569 | 2 | const int remaining = program->source_size - iStart; | |
| 1570 | 2 | const int remaining_2 = program->source_size - iStart_2; | |
| 1571 | 2 | const __mmask64 k1 = _bzhi_u64(~0ULL, remaining); | |
| 1572 | 2 | const __mmask64 k2 = _bzhi_u64(~0ULL, std::max(0, remaining - 64)); | |
| 1573 | 2 | const __mmask64 k1_2 = _bzhi_u64(~0ULL, remaining_2); | |
| 1574 | 4 | const __mmask64 k2_2 = _bzhi_u64(~0ULL, std::max(0, remaining_2 - 64)); | |
| 1575 | |||
| 1576 | // perm_0+N for each kernel tap — loop-invariant, precomputed for both VBMI and BASE paths | ||
| 1577 | 4 | const __m512i pw_1 = _mm512_add_epi8(perm_0, _mm512_set1_epi8(1)); | |
| 1578 | 4 | const __m512i pw_2 = _mm512_add_epi8(perm_0, _mm512_set1_epi8(2)); | |
| 1579 | 4 | const __m512i pw_3 = _mm512_add_epi8(perm_0, _mm512_set1_epi8(3)); | |
| 1580 | 4 | const __m512i pw_4 = _mm512_add_epi8(perm_0, _mm512_set1_epi8(4)); | |
| 1581 | 4 | const __m512i pw_5 = _mm512_add_epi8(perm_0, _mm512_set1_epi8(5)); | |
| 1582 | 4 | const __m512i pw_6 = _mm512_add_epi8(perm_0, _mm512_set1_epi8(6)); | |
| 1583 | 2 | const __m512i pw_7 = _mm512_add_epi8(perm_0, _mm512_set1_epi8(7)); | |
| 1584 | |||
| 1585 | // BASE path only: precompute word_idx and shift_amt from each perm vector. | ||
| 1586 | // MSVC fails to hoist these from the y-loop even though perm_0 is invariant. | ||
| 1587 | 2 | __m512i wi_lo_0 = {}, sa_lo_0 = {}, wi_hi_0 = {}, sa_hi_0 = {}; | |
| 1588 | 2 | __m512i wi_lo_1 = {}, sa_lo_1 = {}, wi_hi_1 = {}, sa_hi_1 = {}; | |
| 1589 | 2 | __m512i wi_lo_2 = {}, sa_lo_2 = {}, wi_hi_2 = {}, sa_hi_2 = {}; | |
| 1590 | 2 | __m512i wi_lo_3 = {}, sa_lo_3 = {}, wi_hi_3 = {}, sa_hi_3 = {}; | |
| 1591 | 2 | __m512i wi_lo_4 = {}, sa_lo_4 = {}, wi_hi_4 = {}, sa_hi_4 = {}; | |
| 1592 | 2 | __m512i wi_lo_5 = {}, sa_lo_5 = {}, wi_hi_5 = {}, sa_hi_5 = {}; | |
| 1593 | 2 | __m512i wi_lo_6 = {}, sa_lo_6 = {}, wi_hi_6 = {}, sa_hi_6 = {}; | |
| 1594 | 2 | __m512i wi_lo_7 = {}, sa_lo_7 = {}, wi_hi_7 = {}, sa_hi_7 = {}; | |
| 1595 | if constexpr (!UseVNNI) { | ||
| 1596 | 1 | const __m512i c_8 = _mm512_set1_epi16(8); | |
| 1597 | 9 | auto make_wi_sa = [&](const __m512i pw, __m512i &wi_lo, __m512i &sa_lo, __m512i &wi_hi, __m512i &sa_hi) { | |
| 1598 | 16 | __m512i lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(pw)); | |
| 1599 | 16 | __m512i hi = _mm512_cvtepu8_epi16(_mm512_extracti64x4_epi64(pw, 1)); | |
| 1600 | 8 | wi_lo = _mm512_srli_epi16(lo, 1); | |
| 1601 | 24 | sa_lo = _mm512_and_si512(_mm512_slli_epi16(lo, 3), c_8); | |
| 1602 | 8 | wi_hi = _mm512_srli_epi16(hi, 1); | |
| 1603 | 16 | sa_hi = _mm512_and_si512(_mm512_slli_epi16(hi, 3), c_8); | |
| 1604 | }; | ||
| 1605 | 1 | make_wi_sa(perm_0, wi_lo_0, sa_lo_0, wi_hi_0, sa_hi_0); | |
| 1606 | 1 | make_wi_sa(pw_1, wi_lo_1, sa_lo_1, wi_hi_1, sa_hi_1); | |
| 1607 | 1 | make_wi_sa(pw_2, wi_lo_2, sa_lo_2, wi_hi_2, sa_hi_2); | |
| 1608 | 1 | make_wi_sa(pw_3, wi_lo_3, sa_lo_3, wi_hi_3, sa_hi_3); | |
| 1609 | 1 | make_wi_sa(pw_4, wi_lo_4, sa_lo_4, wi_hi_4, sa_hi_4); | |
| 1610 | 1 | make_wi_sa(pw_5, wi_lo_5, sa_lo_5, wi_hi_5, sa_hi_5); | |
| 1611 | 1 | make_wi_sa(pw_6, wi_lo_6, sa_lo_6, wi_hi_6, sa_hi_6); | |
| 1612 | 1 | make_wi_sa(pw_7, wi_lo_7, sa_lo_7, wi_hi_7, sa_hi_7); | |
| 1613 | } | ||
| 1614 | |||
| 1615 | // 512-bit vpermi2b + vextracti64x4 + vpunpcklw/hi saturated port 5 on Ice Lake | ||
| 1616 | // (56 port-5 uops/iter vs BASE's 32), costing ~8% vs the BASE simulation path. | ||
| 1617 | // Combined 256-bit indices eliminate extract+unpack, matching BASE at 32 port-5 uops. | ||
| 1618 | // Coefficient layout is unchanged; unpackbw+unpackqdq construction replicates the word | ||
| 1619 | // order of unpacklo/hi_epi16. Dual-AVX-512-pipe CPUs (Sapphire Rapids+) may not have | ||
| 1620 | // suffered this bottleneck due to higher port-5 throughput. | ||
| 1621 | 2 | __m256i comb_01_g1lo={}, comb_01_g1hi={}, comb_01_g2lo={}, comb_01_g2hi={}; | |
| 1622 | 2 | __m256i comb_23_g1lo={}, comb_23_g1hi={}, comb_23_g2lo={}, comb_23_g2hi={}; | |
| 1623 | 2 | __m256i comb_45_g1lo={}, comb_45_g1hi={}, comb_45_g2lo={}, comb_45_g2hi={}; | |
| 1624 | 2 | __m256i comb_67_g1lo={}, comb_67_g1hi={}, comb_67_g2lo={}, comb_67_g2hi={}; | |
| 1625 | if constexpr (UseVNNI) { | ||
| 1626 | 8 | auto make_combined = [&](__m256i pa, __m256i pb, __m256i& clo, __m256i& chi) { | |
| 1627 | 8 | __m256i t0 = _mm256_unpacklo_epi8(pa, pb); | |
| 1628 | 8 | __m256i t1 = _mm256_unpackhi_epi8(pa, pb); | |
| 1629 | 8 | clo = _mm256_unpacklo_epi64(t0, t1); | |
| 1630 | 8 | chi = _mm256_unpackhi_epi64(t0, t1); | |
| 1631 | }; | ||
| 1632 | 1 | __m256i p0g1 = _mm512_castsi512_si256(perm_0); | |
| 1633 | 1 | __m256i p0g2 = _mm512_extracti64x4_epi64(perm_0, 1); | |
| 1634 | // pw_N lower half is a free register alias; upper half = p0g2+N avoids 7 extra extracts | ||
| 1635 | 1 | make_combined(p0g1, _mm512_castsi512_si256(pw_1), comb_01_g1lo, comb_01_g1hi); | |
| 1636 | 2 | make_combined(p0g2, _mm256_add_epi8(p0g2, _mm256_set1_epi8(1)), comb_01_g2lo, comb_01_g2hi); | |
| 1637 | 2 | make_combined(_mm512_castsi512_si256(pw_2), _mm512_castsi512_si256(pw_3), comb_23_g1lo, comb_23_g1hi); | |
| 1638 | 4 | make_combined(_mm256_add_epi8(p0g2, _mm256_set1_epi8(2)), _mm256_add_epi8(p0g2, _mm256_set1_epi8(3)), comb_23_g2lo, comb_23_g2hi); | |
| 1639 | 2 | make_combined(_mm512_castsi512_si256(pw_4), _mm512_castsi512_si256(pw_5), comb_45_g1lo, comb_45_g1hi); | |
| 1640 | 4 | make_combined(_mm256_add_epi8(p0g2, _mm256_set1_epi8(4)), _mm256_add_epi8(p0g2, _mm256_set1_epi8(5)), comb_45_g2lo, comb_45_g2hi); | |
| 1641 | 2 | make_combined(_mm512_castsi512_si256(pw_6), _mm512_castsi512_si256(pw_7), comb_67_g1lo, comb_67_g1hi); | |
| 1642 | 4 | make_combined(_mm256_add_epi8(p0g2, _mm256_set1_epi8(6)), _mm256_add_epi8(p0g2, _mm256_set1_epi8(7)), comb_67_g2lo, comb_67_g2hi); | |
| 1643 | } | ||
| 1644 | |||
| 1645 |
4/8void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 535 → 145 taken 5 times.
✓ Branch 535 → 536 taken 1 time.
✗ Branch 543 → 145 not taken.
✗ Branch 543 → 544 not taken.
void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 481 → 219 taken 5 times.
✓ Branch 481 → 482 taken 1 time.
✗ Branch 489 → 219 not taken.
✗ Branch 489 → 490 not taken.
|
12 | for (int y = y_from; y < y_to; y++) |
| 1646 | { | ||
| 1647 | __m512i data_src, data_src2, data_src_2, data_src2_2; | ||
| 1648 | |||
| 1649 | if constexpr (partial_load) { | ||
| 1650 | ✗ | data_src = _mm512_maskz_loadu_epi8(k1, src_ptr); | |
| 1651 | ✗ | data_src2 = _mm512_maskz_loadu_epi8(k2, src_ptr + 64); | |
| 1652 | ✗ | data_src_2 = _mm512_maskz_loadu_epi8(k1_2, src_ptr_2); | |
| 1653 | ✗ | data_src2_2 = _mm512_maskz_loadu_epi8(k2_2, src_ptr_2 + 64); | |
| 1654 | } | ||
| 1655 | else { | ||
| 1656 | 10 | data_src = _mm512_loadu_si512(src_ptr); | |
| 1657 | 20 | data_src2 = _mm512_loadu_si512(src_ptr + 64); | |
| 1658 | 10 | data_src_2 = _mm512_loadu_si512(src_ptr_2); | |
| 1659 | 10 | data_src2_2 = _mm512_loadu_si512(src_ptr_2 + 64); | |
| 1660 | } | ||
| 1661 | |||
| 1662 | __m512i src_r0r1_0_31lo, src_r0r1_0_31hi, src_r0r1_32_63lo, src_r0r1_32_63hi; | ||
| 1663 | __m512i src_r2r3_0_31lo, src_r2r3_0_31hi, src_r2r3_32_63lo, src_r2r3_32_63hi; | ||
| 1664 | |||
| 1665 | if constexpr (UseVNNI) { | ||
| 1666 | // 512-bit vpermi2b keeps the full 128-byte source window (256-bit vpermi2b only | ||
| 1667 | // addresses 64 bytes, corrupting pixels whose index falls in data_src[32..63]). | ||
| 1668 | // castsi256_si512 on index = free (upper 32 index bytes are undefined but only | ||
| 1669 | // affect output bytes 32-63 which are discarded by castsi512_si256 = also free). | ||
| 1670 | 20 | src_r0r1_0_31lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src, _mm512_castsi256_si512(comb_01_g1lo), data_src2))); | |
| 1671 | 20 | src_r0r1_0_31hi = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src, _mm512_castsi256_si512(comb_01_g1hi), data_src2))); | |
| 1672 | 20 | src_r0r1_32_63lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src_2, _mm512_castsi256_si512(comb_01_g2lo), data_src2_2))); | |
| 1673 | 20 | src_r0r1_32_63hi = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src_2, _mm512_castsi256_si512(comb_01_g2hi), data_src2_2))); | |
| 1674 | 20 | src_r2r3_0_31lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src, _mm512_castsi256_si512(comb_23_g1lo), data_src2))); | |
| 1675 | 20 | src_r2r3_0_31hi = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src, _mm512_castsi256_si512(comb_23_g1hi), data_src2))); | |
| 1676 | 20 | src_r2r3_32_63lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src_2, _mm512_castsi256_si512(comb_23_g2lo), data_src2_2))); | |
| 1677 | 20 | src_r2r3_32_63hi = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src_2, _mm512_castsi256_si512(comb_23_g2hi), data_src2_2))); | |
| 1678 | } else { | ||
| 1679 | // BASE: sim_get32 returns 32 16-bit words with gathered bytes in low 8 bits; | ||
| 1680 | // feed unpacklo/hi directly — avoids cvtepu8_epi16 round-trip | ||
| 1681 | 5 | __m512i d0l = _permutex2var_epi8_sim_get32(wi_lo_0, sa_lo_0, data_src, data_src2); | |
| 1682 | 5 | __m512i d0h = _permutex2var_epi8_sim_get32(wi_hi_0, sa_hi_0, data_src_2, data_src2_2); | |
| 1683 | 5 | __m512i d1l = _permutex2var_epi8_sim_get32(wi_lo_1, sa_lo_1, data_src, data_src2); | |
| 1684 | 5 | __m512i d1h = _permutex2var_epi8_sim_get32(wi_hi_1, sa_hi_1, data_src_2, data_src2_2); | |
| 1685 | 5 | __m512i d2l = _permutex2var_epi8_sim_get32(wi_lo_2, sa_lo_2, data_src, data_src2); | |
| 1686 | 5 | __m512i d2h = _permutex2var_epi8_sim_get32(wi_hi_2, sa_hi_2, data_src_2, data_src2_2); | |
| 1687 | 5 | __m512i d3l = _permutex2var_epi8_sim_get32(wi_lo_3, sa_lo_3, data_src, data_src2); | |
| 1688 | 10 | __m512i d3h = _permutex2var_epi8_sim_get32(wi_hi_3, sa_hi_3, data_src_2, data_src2_2); | |
| 1689 | 5 | src_r0r1_0_31lo = _mm512_unpacklo_epi16(d0l, d1l); | |
| 1690 | 5 | src_r0r1_0_31hi = _mm512_unpackhi_epi16(d0l, d1l); | |
| 1691 | 5 | src_r0r1_32_63lo = _mm512_unpacklo_epi16(d0h, d1h); | |
| 1692 | 5 | src_r0r1_32_63hi = _mm512_unpackhi_epi16(d0h, d1h); | |
| 1693 | 5 | src_r2r3_0_31lo = _mm512_unpacklo_epi16(d2l, d3l); | |
| 1694 | 5 | src_r2r3_0_31hi = _mm512_unpackhi_epi16(d2l, d3l); | |
| 1695 | 5 | src_r2r3_32_63lo = _mm512_unpacklo_epi16(d2h, d3h); | |
| 1696 | 5 | src_r2r3_32_63hi = _mm512_unpackhi_epi16(d2h, d3h); | |
| 1697 | } | ||
| 1698 | |||
| 1699 | __m512i result_0_31lo, result_0_31hi, result_32_63lo, result_32_63hi; | ||
| 1700 | |||
| 1701 | if constexpr (UseVNNI) | ||
| 1702 | { | ||
| 1703 | 5 | result_0_31lo = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31lo, coef_r0r1_0_31lo); | |
| 1704 | 5 | result_0_31hi = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31hi, coef_r0r1_0_31hi); | |
| 1705 | 5 | result_32_63lo = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63lo, coef_r0r1_32_63lo); | |
| 1706 | 10 | result_32_63hi = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63hi, coef_r0r1_32_63hi); | |
| 1707 | 5 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r2r3_0_31lo, coef_r2r3_0_31lo); | |
| 1708 | 5 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r2r3_0_31hi, coef_r2r3_0_31hi); | |
| 1709 | 5 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r2r3_32_63lo, coef_r2r3_32_63lo); | |
| 1710 | 5 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r2r3_32_63hi, coef_r2r3_32_63hi); | |
| 1711 | } | ||
| 1712 | else | ||
| 1713 | { | ||
| 1714 | 15 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, coef_r0r1_0_31lo), _mm512_madd_epi16(src_r2r3_0_31lo, coef_r2r3_0_31lo)); | |
| 1715 | 15 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, coef_r0r1_0_31hi), _mm512_madd_epi16(src_r2r3_0_31hi, coef_r2r3_0_31hi)); | |
| 1716 | 15 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63lo, coef_r0r1_32_63lo), _mm512_madd_epi16(src_r2r3_32_63lo, coef_r2r3_32_63lo)); | |
| 1717 | 10 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63hi, coef_r0r1_32_63hi), _mm512_madd_epi16(src_r2r3_32_63hi, coef_r2r3_32_63hi)); | |
| 1718 | } | ||
| 1719 | |||
| 1720 | __m512i src_r4r5_0_31lo, src_r4r5_0_31hi, src_r4r5_32_63lo, src_r4r5_32_63hi; | ||
| 1721 | __m512i src_r6r7_0_31lo, src_r6r7_0_31hi, src_r6r7_32_63lo, src_r6r7_32_63hi; | ||
| 1722 | |||
| 1723 | if constexpr (UseVNNI) { | ||
| 1724 | 20 | src_r4r5_0_31lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src, _mm512_castsi256_si512(comb_45_g1lo), data_src2))); | |
| 1725 | 20 | src_r4r5_0_31hi = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src, _mm512_castsi256_si512(comb_45_g1hi), data_src2))); | |
| 1726 | 20 | src_r4r5_32_63lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src_2, _mm512_castsi256_si512(comb_45_g2lo), data_src2_2))); | |
| 1727 | 20 | src_r4r5_32_63hi = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src_2, _mm512_castsi256_si512(comb_45_g2hi), data_src2_2))); | |
| 1728 | 20 | src_r6r7_0_31lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src, _mm512_castsi256_si512(comb_67_g1lo), data_src2))); | |
| 1729 | 20 | src_r6r7_0_31hi = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src, _mm512_castsi256_si512(comb_67_g1hi), data_src2))); | |
| 1730 | 20 | src_r6r7_32_63lo = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src_2, _mm512_castsi256_si512(comb_67_g2lo), data_src2_2))); | |
| 1731 | 25 | src_r6r7_32_63hi = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(_mm512_permutex2var_epi8(data_src_2, _mm512_castsi256_si512(comb_67_g2hi), data_src2_2))); | |
| 1732 | } else { | ||
| 1733 | 5 | __m512i d4l = _permutex2var_epi8_sim_get32(wi_lo_4, sa_lo_4, data_src, data_src2); | |
| 1734 | 5 | __m512i d4h = _permutex2var_epi8_sim_get32(wi_hi_4, sa_hi_4, data_src_2, data_src2_2); | |
| 1735 | 5 | __m512i d5l = _permutex2var_epi8_sim_get32(wi_lo_5, sa_lo_5, data_src, data_src2); | |
| 1736 | 5 | __m512i d5h = _permutex2var_epi8_sim_get32(wi_hi_5, sa_hi_5, data_src_2, data_src2_2); | |
| 1737 | 5 | __m512i d6l = _permutex2var_epi8_sim_get32(wi_lo_6, sa_lo_6, data_src, data_src2); | |
| 1738 | 5 | __m512i d6h = _permutex2var_epi8_sim_get32(wi_hi_6, sa_hi_6, data_src_2, data_src2_2); | |
| 1739 | 5 | __m512i d7l = _permutex2var_epi8_sim_get32(wi_lo_7, sa_lo_7, data_src, data_src2); | |
| 1740 | 10 | __m512i d7h = _permutex2var_epi8_sim_get32(wi_hi_7, sa_hi_7, data_src_2, data_src2_2); | |
| 1741 | 5 | src_r4r5_0_31lo = _mm512_unpacklo_epi16(d4l, d5l); | |
| 1742 | 5 | src_r4r5_0_31hi = _mm512_unpackhi_epi16(d4l, d5l); | |
| 1743 | 5 | src_r4r5_32_63lo = _mm512_unpacklo_epi16(d4h, d5h); | |
| 1744 | 5 | src_r4r5_32_63hi = _mm512_unpackhi_epi16(d4h, d5h); | |
| 1745 | 5 | src_r6r7_0_31lo = _mm512_unpacklo_epi16(d6l, d7l); | |
| 1746 | 5 | src_r6r7_0_31hi = _mm512_unpackhi_epi16(d6l, d7l); | |
| 1747 | 5 | src_r6r7_32_63lo = _mm512_unpacklo_epi16(d6h, d7h); | |
| 1748 | 5 | src_r6r7_32_63hi = _mm512_unpackhi_epi16(d6h, d7h); | |
| 1749 | } | ||
| 1750 | |||
| 1751 | if constexpr (UseVNNI) | ||
| 1752 | { | ||
| 1753 | 5 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r4r5_0_31lo, coef_r4r5_0_31lo); | |
| 1754 | 5 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r4r5_0_31hi, coef_r4r5_0_31hi); | |
| 1755 | 5 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r4r5_32_63lo, coef_r4r5_32_63lo); | |
| 1756 | 5 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r4r5_32_63hi, coef_r4r5_32_63hi); | |
| 1757 | 5 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r6r7_0_31lo, coef_r6r7_0_31lo); | |
| 1758 | 5 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r6r7_0_31hi, coef_r6r7_0_31hi); | |
| 1759 | 5 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r6r7_32_63lo, coef_r6r7_32_63lo); | |
| 1760 | 5 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r6r7_32_63hi, coef_r6r7_32_63hi); | |
| 1761 | } | ||
| 1762 | else | ||
| 1763 | { | ||
| 1764 | 10 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_0_31lo, coef_r4r5_0_31lo), result_0_31lo); | |
| 1765 | 10 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_0_31hi, coef_r4r5_0_31hi), result_0_31hi); | |
| 1766 | 10 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_32_63lo, coef_r4r5_32_63lo), result_32_63lo); | |
| 1767 | 10 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_32_63hi, coef_r4r5_32_63hi), result_32_63hi); | |
| 1768 | 10 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_0_31lo, coef_r6r7_0_31lo), result_0_31lo); | |
| 1769 | 10 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_0_31hi, coef_r6r7_0_31hi), result_0_31hi); | |
| 1770 | 10 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_32_63lo, coef_r6r7_32_63lo), result_32_63lo); | |
| 1771 | 5 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_32_63hi, coef_r6r7_32_63hi), result_32_63hi); | |
| 1772 | |||
| 1773 | 5 | result_0_31lo = _mm512_add_epi32(result_0_31lo, rounder); | |
| 1774 | 5 | result_0_31hi = _mm512_add_epi32(result_0_31hi, rounder); | |
| 1775 | 5 | result_32_63lo = _mm512_add_epi32(result_32_63lo, rounder); | |
| 1776 | 10 | result_32_63hi = _mm512_add_epi32(result_32_63hi, rounder); | |
| 1777 | } | ||
| 1778 | |||
| 1779 | 10 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale8bits); | |
| 1780 | 10 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale8bits); | |
| 1781 | 10 | result_32_63lo = _mm512_srai_epi32(result_32_63lo, FPScale8bits); | |
| 1782 | 10 | result_32_63hi = _mm512_srai_epi32(result_32_63hi, FPScale8bits); | |
| 1783 | |||
| 1784 | 10 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 1785 | 10 | __m512i result_32_63_int16 = _mm512_packus_epi32(result_32_63lo, result_32_63hi); | |
| 1786 | |||
| 1787 | 10 | __m256i result_0_31_u8 = _mm512_cvtusepi16_epi8(result_0_31_int16); | |
| 1788 | 10 | __m256i result_32_63_u8 = _mm512_cvtusepi16_epi8(result_32_63_int16); | |
| 1789 | |||
| 1790 | 20 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), _mm512_inserti64x4(_mm512_castsi256_si512(result_0_31_u8), result_32_63_u8, 1)); | |
| 1791 | |||
| 1792 | 10 | dst_ptr += dst_pitch; | |
| 1793 | 10 | src_ptr += src_pitch; | |
| 1794 | 10 | src_ptr_2 += src_pitch; | |
| 1795 | } | ||
| 1796 | |||
| 1797 | 2 | current_coeff_SIMD += 16; | |
| 1798 | }; | ||
| 1799 | |||
| 1800 |
4/4void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 12 → 10 taken 1 time.
✓ Branch 12 → 13 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 12 → 10 taken 1 time.
✓ Branch 12 → 13 taken 1 time.
|
4 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 1801 | { | ||
| 1802 |
1/2✓ Branch 10 → 11 taken 1 time.
✗ Branch 10 → 20 not taken.
|
2 | do_h_integer_core(std::false_type{}); |
| 1803 | } | ||
| 1804 | |||
| 1805 |
2/4void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 16 → 14 not taken.
✓ Branch 16 → 17 taken 1 time.
void resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_internal<true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 16 → 14 not taken.
✓ Branch 16 → 17 taken 1 time.
|
2 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 1806 | { | ||
| 1807 | ✗ | do_h_integer_core(std::true_type{}); | |
| 1808 | } | ||
| 1809 | } | ||
| 1810 | 2 | } | |
| 1811 | |||
| 1812 | // filter size up to 4, pretransposed coefficients | ||
| 1813 | // 64 target uint16_t pixels at a time in 2 groups of 32 | ||
| 1814 | template<bool lessthan16bit, bool UseVNNI> | ||
| 1815 | 9 | void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 1816 | { | ||
| 1817 | 9 | const uint16_t* src = (uint16_t*)src8; | |
| 1818 | 9 | uint16_t* AVS_RESTRICT dst = (uint16_t* AVS_RESTRICT)dst8; | |
| 1819 | 9 | dst_pitch = dst_pitch / sizeof(uint16_t); | |
| 1820 | 9 | src_pitch = src_pitch / sizeof(uint16_t); | |
| 1821 | |||
| 1822 | 9 | constexpr int PIXELS_AT_A_TIME = 64; | |
| 1823 | |||
| 1824 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 6 times.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
|
9 | const int width_safe_mod = (program->safelimit_64_pixels_each32th_target.overread_possible ? program->safelimit_64_pixels_each32th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 1825 | |||
| 1826 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 6 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 1 time.
|
9 | assert(program->filter_size_real <= 4); |
| 1827 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 6 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 1 time.
|
9 | assert(program->target_size_alignment >= 64); |
| 1828 | assert(FRAME_ALIGN >= 64); | ||
| 1829 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 6 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1 time.
|
9 | assert(program->filter_size_alignment >= 4); |
| 1830 | |||
| 1831 | 9 | const int max_scanlines = program->max_scanlines; | |
| 1832 | |||
| 1833 | 9 | const __m512i shifttosigned = _mm512_set1_epi16(-32768); | |
| 1834 | 9 | const __m512i shiftfromsigned = _mm512_set1_epi32(32768 << FPScale16bits); | |
| 1835 | |||
| 1836 | 9 | const int limit = (1 << bits_per_pixel) - 1; | |
| 1837 | 18 | __m512i clamp_limit = _mm512_set1_epi16((short)limit); | |
| 1838 | 9 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale16bits - 1)); | |
| 1839 | |||
| 1840 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 30 → 20 taken 1 time.
✓ Branch 30 → 31 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 30 → 20 taken 6 times.
✓ Branch 30 → 31 taken 6 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 30 → 20 taken 1 time.
✓ Branch 30 → 31 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 30 → 20 taken 1 time.
✓ Branch 30 → 31 taken 1 time.
|
18 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 1841 | { | ||
| 1842 | 9 | int y_to = std::min(y_from + max_scanlines, height); | |
| 1843 | |||
| 1844 | 9 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 1845 | |||
| 1846 | 9 | int x = 0; | |
| 1847 | |||
| 1848 | 26 | auto do_h_integer_core = [&](auto partial_load) { | |
| 1849 | 17 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 1850 | |||
| 1851 | 17 | const __m512i coef_r0r1_0_31lo = _mm512_load_si512(current_coeff_SIMD + 0); | |
| 1852 | 17 | const __m512i coef_r0r1_0_31hi = _mm512_load_si512(current_coeff_SIMD + 1); | |
| 1853 | 17 | const __m512i coef_r0r1_32_63lo = _mm512_load_si512(current_coeff_SIMD + 2); | |
| 1854 | 17 | const __m512i coef_r0r1_32_63hi = _mm512_load_si512(current_coeff_SIMD + 3); | |
| 1855 | 17 | const __m512i coef_r2r3_0_31lo = _mm512_load_si512(current_coeff_SIMD + 4); | |
| 1856 | 17 | const __m512i coef_r2r3_0_31hi = _mm512_load_si512(current_coeff_SIMD + 5); | |
| 1857 | 17 | const __m512i coef_r2r3_32_63lo = _mm512_load_si512(current_coeff_SIMD + 6); | |
| 1858 | 17 | const __m512i coef_r2r3_32_63hi = _mm512_load_si512(current_coeff_SIMD + 7); | |
| 1859 | |||
| 1860 | 17 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 1861 | 17 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 1862 | 17 | __m512i perm_0_32_47 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 32])); | |
| 1863 | 17 | __m512i perm_0_48_63 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 48])); | |
| 1864 | |||
| 1865 | 17 | int iStart_0_31 = program->pixel_offset[x]; | |
| 1866 | 17 | int iStart_32_63 = program->pixel_offset[x + 32]; | |
| 1867 | 17 | __m512i m512i_Start_0_31 = _mm512_set1_epi32(iStart_0_31); | |
| 1868 | 17 | __m512i m512i_Start_32_63 = _mm512_set1_epi32(iStart_32_63); | |
| 1869 | |||
| 1870 | 17 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, m512i_Start_0_31); | |
| 1871 | 17 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, m512i_Start_0_31); | |
| 1872 | 17 | perm_0_32_47 = _mm512_sub_epi32(perm_0_32_47, m512i_Start_32_63); | |
| 1873 | 17 | perm_0_48_63 = _mm512_sub_epi32(perm_0_48_63, m512i_Start_32_63); | |
| 1874 | |||
| 1875 | 17 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 1876 | 17 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 1877 | 17 | __m256i m256i_perm_0_32_47 = _mm512_cvtepi32_epi16(perm_0_32_47); | |
| 1878 | 17 | __m256i m256i_perm_0_48_63 = _mm512_cvtepi32_epi16(perm_0_48_63); | |
| 1879 | |||
| 1880 | 34 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 1881 | 34 | __m512i perm_0_32_63 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_32_47), m256i_perm_0_48_63, 1); | |
| 1882 | |||
| 1883 | 17 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 1884 | 17 | __m512i perm_1_32_63 = _mm512_add_epi16(perm_0_32_63, one_epi16); | |
| 1885 | |||
| 1886 | 17 | const __m512i perm_r0r1_0_31lo = _mm512_unpacklo_epi16(perm_0_0_31, perm_1_0_31); | |
| 1887 | 17 | const __m512i perm_r0r1_0_31hi = _mm512_unpackhi_epi16(perm_0_0_31, perm_1_0_31); | |
| 1888 | 17 | const __m512i perm_r0r1_32_63lo = _mm512_unpacklo_epi16(perm_0_32_63, perm_1_32_63); | |
| 1889 | 17 | const __m512i perm_r0r1_32_63hi = _mm512_unpackhi_epi16(perm_0_32_63, perm_1_32_63); | |
| 1890 | |||
| 1891 | 17 | const __m512i two_epi16 = _mm512_set1_epi16(2); | |
| 1892 | |||
| 1893 | 17 | uint16_t* AVS_RESTRICT dst_ptr = dst + x + y_from * dst_pitch; | |
| 1894 | 17 | const uint16_t* src_ptr_0_31 = src + iStart_0_31 + y_from * src_pitch; | |
| 1895 | 17 | const uint16_t* src_ptr_32_63 = src + iStart_32_63 + y_from * src_pitch; | |
| 1896 | |||
| 1897 | 17 | const int remaining_0_31 = program->source_size - iStart_0_31; | |
| 1898 | 17 | const __mmask32 k1_0_31 = _bzhi_u32(~0UL, remaining_0_31); | |
| 1899 | 17 | const __mmask32 k2_0_31 = _bzhi_u32(~0UL, remaining_0_31 - 32); | |
| 1900 | 17 | const int remaining_32_63 = program->source_size - iStart_32_63; | |
| 1901 | 17 | const __mmask32 k1_32_63 = _bzhi_u32(~0UL, remaining_32_63); | |
| 1902 | 17 | const __mmask32 k2_32_63 = _bzhi_u32(~0UL, remaining_32_63 - 32); | |
| 1903 | |||
| 1904 |
16/16✓ Branch 178 → 101 taken 10 times.
✓ Branch 178 → 179 taken 2 times.
✓ Branch 186 → 101 taken 5 times.
✓ Branch 186 → 187 taken 1 time.
✓ Branch 194 → 101 taken 10 times.
✓ Branch 194 → 195 taken 2 times.
✓ Branch 202 → 101 taken 26 times.
✓ Branch 202 → 203 taken 6 times.
✓ Branch 210 → 101 taken 10 times.
✓ Branch 210 → 211 taken 2 times.
✓ Branch 218 → 101 taken 5 times.
✓ Branch 218 → 219 taken 1 time.
✓ Branch 226 → 101 taken 10 times.
✓ Branch 226 → 227 taken 2 times.
✓ Branch 234 → 101 taken 5 times.
✓ Branch 234 → 235 taken 1 time.
|
98 | for (int y = y_from; y < y_to; y++) |
| 1905 | { | ||
| 1906 | __m512i data_src_0_31, data_src2_0_31, data_src_32_63, data_src2_32_63; | ||
| 1907 | |||
| 1908 | 81 | __m512i perm_rNrNp1_0_31lo = perm_r0r1_0_31lo; | |
| 1909 | 81 | __m512i perm_rNrNp1_0_31hi = perm_r0r1_0_31hi; | |
| 1910 | 81 | __m512i perm_rNrNp1_32_63lo = perm_r0r1_32_63lo; | |
| 1911 | 81 | __m512i perm_rNrNp1_32_63hi = perm_r0r1_32_63hi; | |
| 1912 | |||
| 1913 | if constexpr (partial_load) { | ||
| 1914 | 41 | data_src_0_31 = _mm512_maskz_loadu_epi16(k1_0_31, src_ptr_0_31); | |
| 1915 | 82 | data_src2_0_31 = _mm512_maskz_loadu_epi16(k2_0_31, src_ptr_0_31 + 32); | |
| 1916 | 41 | data_src_32_63 = _mm512_maskz_loadu_epi16(k1_32_63, src_ptr_32_63); | |
| 1917 | 82 | data_src2_32_63 = _mm512_maskz_loadu_epi16(k2_32_63, src_ptr_32_63 + 32); | |
| 1918 | } | ||
| 1919 | else { | ||
| 1920 | 40 | data_src_0_31 = _mm512_loadu_si512(src_ptr_0_31); | |
| 1921 | 80 | data_src2_0_31 = _mm512_loadu_si512(src_ptr_0_31 + 32); | |
| 1922 | 40 | data_src_32_63 = _mm512_loadu_si512(src_ptr_32_63); | |
| 1923 | 80 | data_src2_32_63 = _mm512_loadu_si512(src_ptr_32_63 + 32); | |
| 1924 | } | ||
| 1925 | |||
| 1926 | 81 | __m512i src_r0r1_0_31lo = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31lo, data_src2_0_31); | |
| 1927 | 81 | __m512i src_r0r1_0_31hi = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31hi, data_src2_0_31); | |
| 1928 | 81 | __m512i src_r0r1_32_63lo = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63lo, data_src2_32_63); | |
| 1929 | 81 | __m512i src_r0r1_32_63hi = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63hi, data_src2_32_63); | |
| 1930 | |||
| 1931 | // for r2r3 | ||
| 1932 | 81 | perm_rNrNp1_0_31lo = _mm512_add_epi16(perm_rNrNp1_0_31lo, two_epi16); | |
| 1933 | 81 | perm_rNrNp1_0_31hi = _mm512_add_epi16(perm_rNrNp1_0_31hi, two_epi16); | |
| 1934 | 81 | perm_rNrNp1_32_63lo = _mm512_add_epi16(perm_rNrNp1_32_63lo, two_epi16); | |
| 1935 | 81 | perm_rNrNp1_32_63hi = _mm512_add_epi16(perm_rNrNp1_32_63hi, two_epi16); | |
| 1936 | |||
| 1937 | if constexpr (!lessthan16bit) { | ||
| 1938 | 51 | src_r0r1_0_31lo = _mm512_add_epi16(src_r0r1_0_31lo, shifttosigned); | |
| 1939 | 51 | src_r0r1_0_31hi = _mm512_add_epi16(src_r0r1_0_31hi, shifttosigned); | |
| 1940 | 51 | src_r0r1_32_63lo = _mm512_add_epi16(src_r0r1_32_63lo, shifttosigned); | |
| 1941 | 51 | src_r0r1_32_63hi = _mm512_add_epi16(src_r0r1_32_63hi, shifttosigned); | |
| 1942 | } | ||
| 1943 | |||
| 1944 | __m512i result_0_31lo, result_0_31hi, result_32_63lo, result_32_63hi; | ||
| 1945 | |||
| 1946 | if constexpr (UseVNNI) | ||
| 1947 | { | ||
| 1948 | 51 | result_0_31lo = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31lo, coef_r0r1_0_31lo); | |
| 1949 | 51 | result_0_31hi = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31hi, coef_r0r1_0_31hi); | |
| 1950 | 51 | result_32_63lo = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63lo, coef_r0r1_32_63lo); | |
| 1951 | 102 | result_32_63hi = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63hi, coef_r0r1_32_63hi); | |
| 1952 | } | ||
| 1953 | else | ||
| 1954 | { | ||
| 1955 | 60 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, coef_r0r1_0_31lo), rounder); | |
| 1956 | 60 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, coef_r0r1_0_31hi), rounder); | |
| 1957 | 60 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63lo, coef_r0r1_32_63lo), rounder); | |
| 1958 | 90 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63hi, coef_r0r1_32_63hi), rounder); | |
| 1959 | } | ||
| 1960 | |||
| 1961 | 81 | __m512i src_r2r3_0_31lo = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31lo, data_src2_0_31); | |
| 1962 | 81 | __m512i src_r2r3_0_31hi = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31hi, data_src2_0_31); | |
| 1963 | 81 | __m512i src_r2r3_32_63lo = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63lo, data_src2_32_63); | |
| 1964 | 30 | __m512i src_r2r3_32_63hi = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63hi, data_src2_32_63); | |
| 1965 | |||
| 1966 | if constexpr (!lessthan16bit) { | ||
| 1967 | 51 | src_r2r3_0_31lo = _mm512_add_epi16(src_r2r3_0_31lo, shifttosigned); | |
| 1968 | 51 | src_r2r3_0_31hi = _mm512_add_epi16(src_r2r3_0_31hi, shifttosigned); | |
| 1969 | 51 | src_r2r3_32_63lo = _mm512_add_epi16(src_r2r3_32_63lo, shifttosigned); | |
| 1970 | 102 | src_r2r3_32_63hi = _mm512_add_epi16(src_r2r3_32_63hi, shifttosigned); | |
| 1971 | } | ||
| 1972 | |||
| 1973 | if constexpr (UseVNNI) | ||
| 1974 | { | ||
| 1975 | 51 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r2r3_0_31lo, coef_r2r3_0_31lo); | |
| 1976 | 51 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r2r3_0_31hi, coef_r2r3_0_31hi); | |
| 1977 | 51 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r2r3_32_63lo, coef_r2r3_32_63lo); | |
| 1978 | 15 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r2r3_32_63hi, coef_r2r3_32_63hi); | |
| 1979 | } | ||
| 1980 | else | ||
| 1981 | { | ||
| 1982 | 60 | result_0_31lo = _mm512_add_epi32(result_0_31lo, _mm512_madd_epi16(src_r2r3_0_31lo, coef_r2r3_0_31lo)); | |
| 1983 | 60 | result_0_31hi = _mm512_add_epi32(result_0_31hi, _mm512_madd_epi16(src_r2r3_0_31hi, coef_r2r3_0_31hi)); | |
| 1984 | 60 | result_32_63lo = _mm512_add_epi32(result_32_63lo, _mm512_madd_epi16(src_r2r3_32_63lo, coef_r2r3_32_63lo)); | |
| 1985 | 45 | result_32_63hi = _mm512_add_epi32(result_32_63hi, _mm512_madd_epi16(src_r2r3_32_63hi, coef_r2r3_32_63hi)); | |
| 1986 | } | ||
| 1987 | |||
| 1988 | if constexpr (!lessthan16bit) { | ||
| 1989 | 51 | result_0_31lo = _mm512_add_epi32(result_0_31lo, shiftfromsigned); | |
| 1990 | 51 | result_0_31hi = _mm512_add_epi32(result_0_31hi, shiftfromsigned); | |
| 1991 | 51 | result_32_63lo = _mm512_add_epi32(result_32_63lo, shiftfromsigned); | |
| 1992 | 102 | result_32_63hi = _mm512_add_epi32(result_32_63hi, shiftfromsigned); | |
| 1993 | } | ||
| 1994 | |||
| 1995 | 81 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale16bits); | |
| 1996 | 81 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale16bits); | |
| 1997 | 81 | result_32_63lo = _mm512_srai_epi32(result_32_63lo, FPScale16bits); | |
| 1998 | 81 | result_32_63hi = _mm512_srai_epi32(result_32_63hi, FPScale16bits); | |
| 1999 | |||
| 2000 | 81 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 2001 | 51 | __m512i result_32_63_int16 = _mm512_packus_epi32(result_32_63lo, result_32_63hi); | |
| 2002 | |||
| 2003 | if constexpr (lessthan16bit) { | ||
| 2004 | 30 | result_0_31_int16 = _mm512_min_epu16(result_0_31_int16, clamp_limit); | |
| 2005 | 60 | result_32_63_int16 = _mm512_min_epu16(result_32_63_int16, clamp_limit); | |
| 2006 | } | ||
| 2007 | |||
| 2008 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), result_0_31_int16); | ||
| 2009 | |||
| 2010 | 81 | const int w_mod32 = width / 32 * 32; | |
| 2011 |
9/16✓ Branch 174 → 175 taken 10 times.
✗ Branch 174 → 177 not taken.
✓ Branch 182 → 183 taken 5 times.
✗ Branch 182 → 185 not taken.
✓ Branch 190 → 191 taken 10 times.
✗ Branch 190 → 193 not taken.
✓ Branch 198 → 199 taken 5 times.
✓ Branch 198 → 201 taken 21 times.
✓ Branch 206 → 207 taken 10 times.
✗ Branch 206 → 209 not taken.
✓ Branch 214 → 215 taken 5 times.
✗ Branch 214 → 217 not taken.
✓ Branch 222 → 223 taken 10 times.
✗ Branch 222 → 225 not taken.
✓ Branch 230 → 231 taken 5 times.
✗ Branch 230 → 233 not taken.
|
81 | if (x < (w_mod32 - 32)) |
| 2012 | 60 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr + 32), result_32_63_int16); | |
| 2013 | |||
| 2014 | 81 | dst_ptr += dst_pitch; | |
| 2015 | 81 | src_ptr_0_31 += src_pitch; | |
| 2016 | 81 | src_ptr_32_63 += src_pitch; | |
| 2017 | } | ||
| 2018 | |||
| 2019 | 17 | current_coeff_SIMD += 8; | |
| 2020 | }; | ||
| 2021 | |||
| 2022 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 2 times.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 2 times.
✓ Branch 24 → 25 taken 6 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 2 times.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 2 times.
✓ Branch 24 → 25 taken 1 time.
|
17 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 2023 | { | ||
| 2024 | 8 | do_h_integer_core(std::false_type{}); | |
| 2025 | } | ||
| 2026 | |||
| 2027 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 28 → 26 taken 1 time.
✓ Branch 28 → 29 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 28 → 26 taken 6 times.
✓ Branch 28 → 29 taken 6 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 28 → 26 taken 1 time.
✓ Branch 28 → 29 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 28 → 26 taken 1 time.
✓ Branch 28 → 29 taken 1 time.
|
18 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 2028 | { | ||
| 2029 | 9 | do_h_integer_core(std::true_type{}); | |
| 2030 | } | ||
| 2031 | } | ||
| 2032 | 9 | } | |
| 2033 | |||
| 2034 | template<bool lessthan16bit, bool UseVNNI> | ||
| 2035 | 7 | void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 2036 | { | ||
| 2037 | 7 | const uint16_t* src = (uint16_t*)src8; | |
| 2038 | 7 | uint16_t* AVS_RESTRICT dst = (uint16_t* AVS_RESTRICT)dst8; | |
| 2039 | 7 | dst_pitch = dst_pitch / sizeof(uint16_t); | |
| 2040 | 7 | src_pitch = src_pitch / sizeof(uint16_t); | |
| 2041 | |||
| 2042 | 7 | constexpr int PIXELS_AT_A_TIME = 64; | |
| 2043 | |||
| 2044 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 4 times.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
|
7 | const int width_safe_mod = (program->safelimit_64_pixels_each32th_target.overread_possible ? program->safelimit_64_pixels_each32th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 2045 | |||
| 2046 | 7 | const int max_scanlines = program->max_scanlines; | |
| 2047 | |||
| 2048 | 7 | const __m512i shifttosigned = _mm512_set1_epi16(-32768); | |
| 2049 | 7 | const __m512i shiftfromsigned = _mm512_set1_epi32(32768 << FPScale16bits); | |
| 2050 | |||
| 2051 | 7 | const int limit = (1 << bits_per_pixel) - 1; | |
| 2052 | 14 | __m512i clamp_limit = _mm512_set1_epi16((short)limit); | |
| 2053 | 7 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale16bits - 1)); | |
| 2054 | |||
| 2055 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 4 times.
✓ Branch 24 → 25 taken 4 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
|
14 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 2056 | { | ||
| 2057 | 7 | int y_to = std::min(y_from + max_scanlines, height); | |
| 2058 | |||
| 2059 | 7 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 2060 | |||
| 2061 | 7 | int x = 0; | |
| 2062 | |||
| 2063 | 22 | auto do_h_integer_core = [&](auto partial_load) { | |
| 2064 | 15 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 2065 | |||
| 2066 | 15 | const __m512i coef_r0r1_0_31lo = _mm512_load_si512(current_coeff_SIMD + 0); | |
| 2067 | 15 | const __m512i coef_r0r1_0_31hi = _mm512_load_si512(current_coeff_SIMD + 1); | |
| 2068 | 15 | const __m512i coef_r0r1_32_63lo = _mm512_load_si512(current_coeff_SIMD + 2); | |
| 2069 | 15 | const __m512i coef_r0r1_32_63hi = _mm512_load_si512(current_coeff_SIMD + 3); | |
| 2070 | 15 | const __m512i coef_r2r3_0_31lo = _mm512_load_si512(current_coeff_SIMD + 4); | |
| 2071 | 15 | const __m512i coef_r2r3_0_31hi = _mm512_load_si512(current_coeff_SIMD + 5); | |
| 2072 | 15 | const __m512i coef_r2r3_32_63lo = _mm512_load_si512(current_coeff_SIMD + 6); | |
| 2073 | 15 | const __m512i coef_r2r3_32_63hi = _mm512_load_si512(current_coeff_SIMD + 7); | |
| 2074 | 15 | const __m512i coef_r4r5_0_31lo = _mm512_load_si512(current_coeff_SIMD + 8); | |
| 2075 | 15 | const __m512i coef_r4r5_0_31hi = _mm512_load_si512(current_coeff_SIMD + 9); | |
| 2076 | 15 | const __m512i coef_r4r5_32_63lo = _mm512_load_si512(current_coeff_SIMD + 10); | |
| 2077 | 15 | const __m512i coef_r4r5_32_63hi = _mm512_load_si512(current_coeff_SIMD + 11); | |
| 2078 | 15 | const __m512i coef_r6r7_0_31lo = _mm512_load_si512(current_coeff_SIMD + 12); | |
| 2079 | 15 | const __m512i coef_r6r7_0_31hi = _mm512_load_si512(current_coeff_SIMD + 13); | |
| 2080 | 15 | const __m512i coef_r6r7_32_63lo = _mm512_load_si512(current_coeff_SIMD + 14); | |
| 2081 | 15 | const __m512i coef_r6r7_32_63hi = _mm512_load_si512(current_coeff_SIMD + 15); | |
| 2082 | |||
| 2083 | 15 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 2084 | 15 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 2085 | 15 | __m512i perm_0_32_47 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 32])); | |
| 2086 | 15 | __m512i perm_0_48_63 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 48])); | |
| 2087 | |||
| 2088 | 15 | int iStart_0_31 = program->pixel_offset[x]; | |
| 2089 | 15 | int iStart_32_63 = program->pixel_offset[x + 32]; | |
| 2090 | 15 | __m512i m512i_Start_0_31 = _mm512_set1_epi32(iStart_0_31); | |
| 2091 | 15 | __m512i m512i_Start_32_63 = _mm512_set1_epi32(iStart_32_63); | |
| 2092 | |||
| 2093 | 15 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, m512i_Start_0_31); | |
| 2094 | 15 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, m512i_Start_0_31); | |
| 2095 | 15 | perm_0_32_47 = _mm512_sub_epi32(perm_0_32_47, m512i_Start_32_63); | |
| 2096 | 15 | perm_0_48_63 = _mm512_sub_epi32(perm_0_48_63, m512i_Start_32_63); | |
| 2097 | |||
| 2098 | 15 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 2099 | 15 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 2100 | 15 | __m256i m256i_perm_0_32_47 = _mm512_cvtepi32_epi16(perm_0_32_47); | |
| 2101 | 15 | __m256i m256i_perm_0_48_63 = _mm512_cvtepi32_epi16(perm_0_48_63); | |
| 2102 | |||
| 2103 | 30 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 2104 | 30 | __m512i perm_0_32_63 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_32_47), m256i_perm_0_48_63, 1); | |
| 2105 | |||
| 2106 | 15 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 2107 | 15 | __m512i perm_1_32_63 = _mm512_add_epi16(perm_0_32_63, one_epi16); | |
| 2108 | |||
| 2109 | 15 | const __m512i perm_r0r1_0_31lo = _mm512_unpacklo_epi16(perm_0_0_31, perm_1_0_31); | |
| 2110 | 15 | const __m512i perm_r0r1_0_31hi = _mm512_unpackhi_epi16(perm_0_0_31, perm_1_0_31); | |
| 2111 | 15 | const __m512i perm_r0r1_32_63lo = _mm512_unpacklo_epi16(perm_0_32_63, perm_1_32_63); | |
| 2112 | 15 | const __m512i perm_r0r1_32_63hi = _mm512_unpackhi_epi16(perm_0_32_63, perm_1_32_63); | |
| 2113 | |||
| 2114 | 15 | const __m512i two_epi16 = _mm512_set1_epi16(2); | |
| 2115 | |||
| 2116 | 15 | uint16_t* AVS_RESTRICT dst_ptr = dst + x + y_from * dst_pitch; | |
| 2117 | 15 | const uint16_t* src_ptr_0_31 = src + iStart_0_31 + y_from * src_pitch; | |
| 2118 | 15 | const uint16_t* src_ptr_32_63 = src + iStart_32_63 + y_from * src_pitch; | |
| 2119 | |||
| 2120 | 15 | const int remaining_0_31 = program->source_size - iStart_0_31; | |
| 2121 | 15 | const __mmask32 k1_0_31 = _bzhi_u32(~0UL, remaining_0_31); | |
| 2122 | 15 | const __mmask32 k2_0_31 = _bzhi_u32(~0UL, remaining_0_31 - 32); | |
| 2123 | 15 | const int remaining_32_63 = program->source_size - iStart_32_63; | |
| 2124 | 15 | const __mmask32 k1_32_63 = _bzhi_u32(~0UL, remaining_32_63); | |
| 2125 | 15 | const __mmask32 k2_32_63 = _bzhi_u32(~0UL, remaining_32_63 - 32); | |
| 2126 | |||
| 2127 |
16/16✓ Branch 242 → 117 taken 10 times.
✓ Branch 242 → 243 taken 2 times.
✓ Branch 250 → 117 taken 5 times.
✓ Branch 250 → 251 taken 1 time.
✓ Branch 274 → 117 taken 10 times.
✓ Branch 274 → 275 taken 2 times.
✓ Branch 282 → 117 taken 20 times.
✓ Branch 282 → 283 taken 4 times.
✓ Branch 306 → 117 taken 10 times.
✓ Branch 306 → 307 taken 2 times.
✓ Branch 314 → 117 taken 5 times.
✓ Branch 314 → 315 taken 1 time.
✓ Branch 338 → 117 taken 10 times.
✓ Branch 338 → 339 taken 2 times.
✓ Branch 346 → 117 taken 5 times.
✓ Branch 346 → 347 taken 1 time.
|
90 | for (int y = y_from; y < y_to; y++) |
| 2128 | { | ||
| 2129 | __m512i data_src_0_31, data_src2_0_31, data_src_32_63, data_src2_32_63; | ||
| 2130 | |||
| 2131 | 75 | __m512i perm_rNrNp1_0_31lo = perm_r0r1_0_31lo; | |
| 2132 | 75 | __m512i perm_rNrNp1_0_31hi = perm_r0r1_0_31hi; | |
| 2133 | 75 | __m512i perm_rNrNp1_32_63lo = perm_r0r1_32_63lo; | |
| 2134 | 75 | __m512i perm_rNrNp1_32_63hi = perm_r0r1_32_63hi; | |
| 2135 | |||
| 2136 | if constexpr (partial_load) { | ||
| 2137 | 35 | data_src_0_31 = _mm512_maskz_loadu_epi16(k1_0_31, src_ptr_0_31); | |
| 2138 | 70 | data_src2_0_31 = _mm512_maskz_loadu_epi16(k2_0_31, src_ptr_0_31 + 32); | |
| 2139 | 35 | data_src_32_63 = _mm512_maskz_loadu_epi16(k1_32_63, src_ptr_32_63); | |
| 2140 | 70 | data_src2_32_63 = _mm512_maskz_loadu_epi16(k2_32_63, src_ptr_32_63 + 32); | |
| 2141 | } | ||
| 2142 | else { | ||
| 2143 | 40 | data_src_0_31 = _mm512_loadu_si512(src_ptr_0_31); | |
| 2144 | 80 | data_src2_0_31 = _mm512_loadu_si512(src_ptr_0_31 + 32); | |
| 2145 | 40 | data_src_32_63 = _mm512_loadu_si512(src_ptr_32_63); | |
| 2146 | 80 | data_src2_32_63 = _mm512_loadu_si512(src_ptr_32_63 + 32); | |
| 2147 | } | ||
| 2148 | |||
| 2149 | __m512i result_0_31lo, result_0_31hi, result_32_63lo, result_32_63hi; | ||
| 2150 | |||
| 2151 | 75 | __m512i src_r0r1_0_31lo = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31lo, data_src2_0_31); | |
| 2152 | 75 | __m512i src_r0r1_0_31hi = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31hi, data_src2_0_31); | |
| 2153 | 75 | __m512i src_r0r1_32_63lo = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63lo, data_src2_32_63); | |
| 2154 | 75 | __m512i src_r0r1_32_63hi = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63hi, data_src2_32_63); | |
| 2155 | |||
| 2156 | 75 | perm_rNrNp1_0_31lo = _mm512_add_epi16(perm_rNrNp1_0_31lo, two_epi16); | |
| 2157 | 75 | perm_rNrNp1_0_31hi = _mm512_add_epi16(perm_rNrNp1_0_31hi, two_epi16); | |
| 2158 | 75 | perm_rNrNp1_32_63lo = _mm512_add_epi16(perm_rNrNp1_32_63lo, two_epi16); | |
| 2159 | 75 | perm_rNrNp1_32_63hi = _mm512_add_epi16(perm_rNrNp1_32_63hi, two_epi16); | |
| 2160 | |||
| 2161 | if constexpr (!lessthan16bit) { | ||
| 2162 | 45 | src_r0r1_0_31lo = _mm512_add_epi16(src_r0r1_0_31lo, shifttosigned); | |
| 2163 | 45 | src_r0r1_0_31hi = _mm512_add_epi16(src_r0r1_0_31hi, shifttosigned); | |
| 2164 | 45 | src_r0r1_32_63lo = _mm512_add_epi16(src_r0r1_32_63lo, shifttosigned); | |
| 2165 | 45 | src_r0r1_32_63hi = _mm512_add_epi16(src_r0r1_32_63hi, shifttosigned); | |
| 2166 | } | ||
| 2167 | |||
| 2168 | if constexpr (UseVNNI) | ||
| 2169 | { | ||
| 2170 | 45 | result_0_31lo = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31lo, coef_r0r1_0_31lo); | |
| 2171 | 45 | result_0_31hi = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31hi, coef_r0r1_0_31hi); | |
| 2172 | 45 | result_32_63lo = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63lo, coef_r0r1_32_63lo); | |
| 2173 | 90 | result_32_63hi = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63hi, coef_r0r1_32_63hi); | |
| 2174 | } | ||
| 2175 | else | ||
| 2176 | { | ||
| 2177 | 60 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, coef_r0r1_0_31lo), rounder); | |
| 2178 | 60 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, coef_r0r1_0_31hi), rounder); | |
| 2179 | 60 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63lo, coef_r0r1_32_63lo), rounder); | |
| 2180 | 90 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63hi, coef_r0r1_32_63hi), rounder); | |
| 2181 | } | ||
| 2182 | |||
| 2183 | 75 | __m512i src_r2r3_0_31lo = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31lo, data_src2_0_31); | |
| 2184 | 75 | __m512i src_r2r3_0_31hi = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31hi, data_src2_0_31); | |
| 2185 | 75 | __m512i src_r2r3_32_63lo = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63lo, data_src2_32_63); | |
| 2186 | 75 | __m512i src_r2r3_32_63hi = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63hi, data_src2_32_63); | |
| 2187 | |||
| 2188 | 75 | perm_rNrNp1_0_31lo = _mm512_add_epi16(perm_rNrNp1_0_31lo, two_epi16); | |
| 2189 | 75 | perm_rNrNp1_0_31hi = _mm512_add_epi16(perm_rNrNp1_0_31hi, two_epi16); | |
| 2190 | 75 | perm_rNrNp1_32_63lo = _mm512_add_epi16(perm_rNrNp1_32_63lo, two_epi16); | |
| 2191 | 30 | perm_rNrNp1_32_63hi = _mm512_add_epi16(perm_rNrNp1_32_63hi, two_epi16); | |
| 2192 | |||
| 2193 | if constexpr (!lessthan16bit) { | ||
| 2194 | 45 | src_r2r3_0_31lo = _mm512_add_epi16(src_r2r3_0_31lo, shifttosigned); | |
| 2195 | 45 | src_r2r3_0_31hi = _mm512_add_epi16(src_r2r3_0_31hi, shifttosigned); | |
| 2196 | 45 | src_r2r3_32_63lo = _mm512_add_epi16(src_r2r3_32_63lo, shifttosigned); | |
| 2197 | 90 | src_r2r3_32_63hi = _mm512_add_epi16(src_r2r3_32_63hi, shifttosigned); | |
| 2198 | } | ||
| 2199 | |||
| 2200 | if constexpr (UseVNNI) | ||
| 2201 | { | ||
| 2202 | 45 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r2r3_0_31lo, coef_r2r3_0_31lo); | |
| 2203 | 45 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r2r3_0_31hi, coef_r2r3_0_31hi); | |
| 2204 | 45 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r2r3_32_63lo, coef_r2r3_32_63lo); | |
| 2205 | 45 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r2r3_32_63hi, coef_r2r3_32_63hi); | |
| 2206 | } | ||
| 2207 | else | ||
| 2208 | { | ||
| 2209 | 60 | result_0_31lo = _mm512_add_epi32(result_0_31lo, _mm512_madd_epi16(src_r2r3_0_31lo, coef_r2r3_0_31lo)); | |
| 2210 | 60 | result_0_31hi = _mm512_add_epi32(result_0_31hi, _mm512_madd_epi16(src_r2r3_0_31hi, coef_r2r3_0_31hi)); | |
| 2211 | 60 | result_32_63lo = _mm512_add_epi32(result_32_63lo, _mm512_madd_epi16(src_r2r3_32_63lo, coef_r2r3_32_63lo)); | |
| 2212 | 60 | result_32_63hi = _mm512_add_epi32(result_32_63hi, _mm512_madd_epi16(src_r2r3_32_63hi, coef_r2r3_32_63hi)); | |
| 2213 | } | ||
| 2214 | |||
| 2215 | 75 | __m512i src_r4r5_0_31lo = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31lo, data_src2_0_31); | |
| 2216 | 75 | __m512i src_r4r5_0_31hi = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31hi, data_src2_0_31); | |
| 2217 | 75 | __m512i src_r4r5_32_63lo = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63lo, data_src2_32_63); | |
| 2218 | 75 | __m512i src_r4r5_32_63hi = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63hi, data_src2_32_63); | |
| 2219 | |||
| 2220 | 75 | perm_rNrNp1_0_31lo = _mm512_add_epi16(perm_rNrNp1_0_31lo, two_epi16); | |
| 2221 | 75 | perm_rNrNp1_0_31hi = _mm512_add_epi16(perm_rNrNp1_0_31hi, two_epi16); | |
| 2222 | 75 | perm_rNrNp1_32_63lo = _mm512_add_epi16(perm_rNrNp1_32_63lo, two_epi16); | |
| 2223 | 30 | perm_rNrNp1_32_63hi = _mm512_add_epi16(perm_rNrNp1_32_63hi, two_epi16); | |
| 2224 | |||
| 2225 | if constexpr (!lessthan16bit) { | ||
| 2226 | 45 | src_r4r5_0_31lo = _mm512_add_epi16(src_r4r5_0_31lo, shifttosigned); | |
| 2227 | 45 | src_r4r5_0_31hi = _mm512_add_epi16(src_r4r5_0_31hi, shifttosigned); | |
| 2228 | 45 | src_r4r5_32_63lo = _mm512_add_epi16(src_r4r5_32_63lo, shifttosigned); | |
| 2229 | 90 | src_r4r5_32_63hi = _mm512_add_epi16(src_r4r5_32_63hi, shifttosigned); | |
| 2230 | } | ||
| 2231 | |||
| 2232 | if constexpr (UseVNNI) | ||
| 2233 | { | ||
| 2234 | 45 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r4r5_0_31lo, coef_r4r5_0_31lo); | |
| 2235 | 45 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r4r5_0_31hi, coef_r4r5_0_31hi); | |
| 2236 | 45 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r4r5_32_63lo, coef_r4r5_32_63lo); | |
| 2237 | 45 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r4r5_32_63hi, coef_r4r5_32_63hi); | |
| 2238 | } | ||
| 2239 | else | ||
| 2240 | { | ||
| 2241 | 60 | result_0_31lo = _mm512_add_epi32(result_0_31lo, _mm512_madd_epi16(src_r4r5_0_31lo, coef_r4r5_0_31lo)); | |
| 2242 | 60 | result_0_31hi = _mm512_add_epi32(result_0_31hi, _mm512_madd_epi16(src_r4r5_0_31hi, coef_r4r5_0_31hi)); | |
| 2243 | 60 | result_32_63lo = _mm512_add_epi32(result_32_63lo, _mm512_madd_epi16(src_r4r5_32_63lo, coef_r4r5_32_63lo)); | |
| 2244 | 60 | result_32_63hi = _mm512_add_epi32(result_32_63hi, _mm512_madd_epi16(src_r4r5_32_63hi, coef_r4r5_32_63hi)); | |
| 2245 | } | ||
| 2246 | |||
| 2247 | 75 | __m512i src_r6r7_0_31lo = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31lo, data_src2_0_31); | |
| 2248 | 75 | __m512i src_r6r7_0_31hi = _mm512_permutex2var_epi16(data_src_0_31, perm_rNrNp1_0_31hi, data_src2_0_31); | |
| 2249 | 75 | __m512i src_r6r7_32_63lo = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63lo, data_src2_32_63); | |
| 2250 | 30 | __m512i src_r6r7_32_63hi = _mm512_permutex2var_epi16(data_src_32_63, perm_rNrNp1_32_63hi, data_src2_32_63); | |
| 2251 | |||
| 2252 | if constexpr (!lessthan16bit) { | ||
| 2253 | 45 | src_r6r7_0_31lo = _mm512_add_epi16(src_r6r7_0_31lo, shifttosigned); | |
| 2254 | 45 | src_r6r7_0_31hi = _mm512_add_epi16(src_r6r7_0_31hi, shifttosigned); | |
| 2255 | 45 | src_r6r7_32_63lo = _mm512_add_epi16(src_r6r7_32_63lo, shifttosigned); | |
| 2256 | 90 | src_r6r7_32_63hi = _mm512_add_epi16(src_r6r7_32_63hi, shifttosigned); | |
| 2257 | } | ||
| 2258 | |||
| 2259 | if constexpr (UseVNNI) | ||
| 2260 | { | ||
| 2261 | 45 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r6r7_0_31lo, coef_r6r7_0_31lo); | |
| 2262 | 45 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r6r7_0_31hi, coef_r6r7_0_31hi); | |
| 2263 | 45 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r6r7_32_63lo, coef_r6r7_32_63lo); | |
| 2264 | 15 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r6r7_32_63hi, coef_r6r7_32_63hi); | |
| 2265 | } | ||
| 2266 | else | ||
| 2267 | { | ||
| 2268 | 60 | result_0_31lo = _mm512_add_epi32(result_0_31lo, _mm512_madd_epi16(src_r6r7_0_31lo, coef_r6r7_0_31lo)); | |
| 2269 | 60 | result_0_31hi = _mm512_add_epi32(result_0_31hi, _mm512_madd_epi16(src_r6r7_0_31hi, coef_r6r7_0_31hi)); | |
| 2270 | 60 | result_32_63lo = _mm512_add_epi32(result_32_63lo, _mm512_madd_epi16(src_r6r7_32_63lo, coef_r6r7_32_63lo)); | |
| 2271 | 45 | result_32_63hi = _mm512_add_epi32(result_32_63hi, _mm512_madd_epi16(src_r6r7_32_63hi, coef_r6r7_32_63hi)); | |
| 2272 | } | ||
| 2273 | |||
| 2274 | if constexpr (!lessthan16bit) { | ||
| 2275 | 45 | result_0_31lo = _mm512_add_epi32(result_0_31lo, shiftfromsigned); | |
| 2276 | 45 | result_0_31hi = _mm512_add_epi32(result_0_31hi, shiftfromsigned); | |
| 2277 | 45 | result_32_63lo = _mm512_add_epi32(result_32_63lo, shiftfromsigned); | |
| 2278 | 90 | result_32_63hi = _mm512_add_epi32(result_32_63hi, shiftfromsigned); | |
| 2279 | } | ||
| 2280 | |||
| 2281 | 75 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale16bits); | |
| 2282 | 75 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale16bits); | |
| 2283 | 75 | result_32_63lo = _mm512_srai_epi32(result_32_63lo, FPScale16bits); | |
| 2284 | 75 | result_32_63hi = _mm512_srai_epi32(result_32_63hi, FPScale16bits); | |
| 2285 | |||
| 2286 | 75 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 2287 | 45 | __m512i result_32_63_int16 = _mm512_packus_epi32(result_32_63lo, result_32_63hi); | |
| 2288 | |||
| 2289 | if constexpr (lessthan16bit) { | ||
| 2290 | 30 | result_0_31_int16 = _mm512_min_epu16(result_0_31_int16, clamp_limit); | |
| 2291 | 60 | result_32_63_int16 = _mm512_min_epu16(result_32_63_int16, clamp_limit); | |
| 2292 | } | ||
| 2293 | |||
| 2294 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), result_0_31_int16); | ||
| 2295 | |||
| 2296 | 75 | const int w_mod32 = width / 32 * 32; | |
| 2297 |
9/16✓ Branch 238 → 239 taken 10 times.
✗ Branch 238 → 241 not taken.
✓ Branch 246 → 247 taken 5 times.
✗ Branch 246 → 249 not taken.
✓ Branch 270 → 271 taken 10 times.
✗ Branch 270 → 273 not taken.
✓ Branch 278 → 279 taken 5 times.
✓ Branch 278 → 281 taken 15 times.
✓ Branch 302 → 303 taken 10 times.
✗ Branch 302 → 305 not taken.
✓ Branch 310 → 311 taken 5 times.
✗ Branch 310 → 313 not taken.
✓ Branch 334 → 335 taken 10 times.
✗ Branch 334 → 337 not taken.
✓ Branch 342 → 343 taken 5 times.
✗ Branch 342 → 345 not taken.
|
75 | if (x < (w_mod32 - 32)) |
| 2298 | 60 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr + 32), result_32_63_int16); | |
| 2299 | |||
| 2300 | 75 | dst_ptr += dst_pitch; | |
| 2301 | 75 | src_ptr_0_31 += src_pitch; | |
| 2302 | 75 | src_ptr_32_63 += src_pitch; | |
| 2303 | } | ||
| 2304 | |||
| 2305 | 15 | current_coeff_SIMD += 16; | |
| 2306 | }; | ||
| 2307 | |||
| 2308 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 2 times.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 2 times.
✓ Branch 18 → 19 taken 4 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 2 times.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 2 times.
✓ Branch 18 → 19 taken 1 time.
|
15 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 2309 | { | ||
| 2310 | 8 | do_h_integer_core(std::false_type{}); | |
| 2311 | } | ||
| 2312 | |||
| 2313 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 1 time.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 4 times.
✓ Branch 22 → 23 taken 4 times.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 1 time.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 1 time.
✓ Branch 22 → 23 taken 1 time.
|
14 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 2314 | { | ||
| 2315 | 7 | do_h_integer_core(std::true_type{}); | |
| 2316 | } | ||
| 2317 | } | ||
| 2318 | 7 | } | |
| 2319 | |||
| 2320 | template<bool lessthan16bit, bool UseVNNI> | ||
| 2321 | 4 | void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 2322 | { | ||
| 2323 | 4 | const uint16_t* src = (uint16_t*)src8; | |
| 2324 | 4 | uint16_t* AVS_RESTRICT dst = (uint16_t* AVS_RESTRICT)dst8; | |
| 2325 | 4 | dst_pitch = dst_pitch / sizeof(uint16_t); | |
| 2326 | 4 | src_pitch = src_pitch / sizeof(uint16_t); | |
| 2327 | |||
| 2328 | 4 | constexpr int PIXELS_AT_A_TIME = 64; | |
| 2329 | |||
| 2330 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
|
4 | const int width_safe_mod = (program->safelimit_64_pixels_each32th_target.overread_possible ? program->safelimit_64_pixels_each32th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 2331 | |||
| 2332 | 4 | const int max_scanlines = program->max_scanlines; | |
| 2333 | |||
| 2334 | 4 | const __m512i shifttosigned = _mm512_set1_epi16(-32768); | |
| 2335 | 4 | const __m512i shiftfromsigned = _mm512_set1_epi32(32768 << FPScale16bits); | |
| 2336 | |||
| 2337 | 4 | const int limit = (1 << bits_per_pixel) - 1; | |
| 2338 | 8 | __m512i clamp_limit = _mm512_set1_epi16((short)limit); | |
| 2339 | 4 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale16bits - 1)); | |
| 2340 | |||
| 2341 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
|
8 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 2342 | { | ||
| 2343 | 4 | int y_to = std::min(y_from + max_scanlines, height); | |
| 2344 | |||
| 2345 | 4 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 2346 | |||
| 2347 | 4 | int x = 0; | |
| 2348 | |||
| 2349 | 8 | auto do_h_integer_core = [&](auto partial_load) { | |
| 2350 | 4 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 2351 | |||
| 2352 | 4 | const __m512i coef_r0r1_0_31lo = _mm512_load_si512(current_coeff_SIMD + 0); | |
| 2353 | 4 | const __m512i coef_r0r1_0_31hi = _mm512_load_si512(current_coeff_SIMD + 1); | |
| 2354 | 4 | const __m512i coef_r0r1_32_63lo = _mm512_load_si512(current_coeff_SIMD + 2); | |
| 2355 | 4 | const __m512i coef_r0r1_32_63hi = _mm512_load_si512(current_coeff_SIMD + 3); | |
| 2356 | 4 | const __m512i coef_r2r3_0_31lo = _mm512_load_si512(current_coeff_SIMD + 4); | |
| 2357 | 4 | const __m512i coef_r2r3_0_31hi = _mm512_load_si512(current_coeff_SIMD + 5); | |
| 2358 | 4 | const __m512i coef_r2r3_32_63lo = _mm512_load_si512(current_coeff_SIMD + 6); | |
| 2359 | 4 | const __m512i coef_r2r3_32_63hi = _mm512_load_si512(current_coeff_SIMD + 7); | |
| 2360 | 4 | const __m512i coef_r4r5_0_31lo = _mm512_load_si512(current_coeff_SIMD + 8); | |
| 2361 | 4 | const __m512i coef_r4r5_0_31hi = _mm512_load_si512(current_coeff_SIMD + 9); | |
| 2362 | 4 | const __m512i coef_r4r5_32_63lo = _mm512_load_si512(current_coeff_SIMD + 10); | |
| 2363 | 4 | const __m512i coef_r4r5_32_63hi = _mm512_load_si512(current_coeff_SIMD + 11); | |
| 2364 | 4 | const __m512i coef_r6r7_0_31lo = _mm512_load_si512(current_coeff_SIMD + 12); | |
| 2365 | 4 | const __m512i coef_r6r7_0_31hi = _mm512_load_si512(current_coeff_SIMD + 13); | |
| 2366 | 4 | const __m512i coef_r6r7_32_63lo = _mm512_load_si512(current_coeff_SIMD + 14); | |
| 2367 | 4 | const __m512i coef_r6r7_32_63hi = _mm512_load_si512(current_coeff_SIMD + 15); | |
| 2368 | |||
| 2369 | 4 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 2370 | 4 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 2371 | 4 | __m512i perm_0_32_47 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 32])); | |
| 2372 | 4 | __m512i perm_0_48_63 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 48])); | |
| 2373 | |||
| 2374 | 4 | int iStart_0_15 = program->pixel_offset[x]; | |
| 2375 | 4 | int iStart_16_31 = program->pixel_offset[x + 16]; | |
| 2376 | 4 | int iStart_32_47 = program->pixel_offset[x + 32]; | |
| 2377 | 4 | int iStart_48_63 = program->pixel_offset[x + 48]; | |
| 2378 | |||
| 2379 | 8 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, _mm512_set1_epi32(iStart_0_15)); | |
| 2380 | 8 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, _mm512_set1_epi32(iStart_16_31)); | |
| 2381 | 8 | perm_0_32_47 = _mm512_sub_epi32(perm_0_32_47, _mm512_set1_epi32(iStart_32_47)); | |
| 2382 | 8 | perm_0_48_63 = _mm512_sub_epi32(perm_0_48_63, _mm512_set1_epi32(iStart_48_63)); | |
| 2383 | |||
| 2384 | 4 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 2385 | 4 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 2386 | 4 | __m256i m256i_perm_0_32_47 = _mm512_cvtepi32_epi16(perm_0_32_47); | |
| 2387 | 4 | __m256i m256i_perm_0_48_63 = _mm512_cvtepi32_epi16(perm_0_48_63); | |
| 2388 | |||
| 2389 | 8 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 2390 | 8 | __m512i perm_0_32_63 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_32_47), m256i_perm_0_48_63, 1); | |
| 2391 | |||
| 2392 | 4 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 2393 | 4 | __m512i perm_1_32_63 = _mm512_add_epi16(perm_0_32_63, one_epi16); | |
| 2394 | |||
| 2395 | 4 | const __m512i perm_r0r1_0_31lo = _mm512_unpacklo_epi16(perm_0_0_31, perm_1_0_31); | |
| 2396 | 4 | const __m512i perm_r0r1_0_31hi = _mm512_unpackhi_epi16(perm_0_0_31, perm_1_0_31); | |
| 2397 | 4 | const __m512i perm_r0r1_32_63lo = _mm512_unpacklo_epi16(perm_0_32_63, perm_1_32_63); | |
| 2398 | 4 | const __m512i perm_r0r1_32_63hi = _mm512_unpackhi_epi16(perm_0_32_63, perm_1_32_63); | |
| 2399 | |||
| 2400 | 4 | const __m512i two_epi16 = _mm512_set1_epi16(2); | |
| 2401 | 4 | const __mmask32 k_high = 0xFFFF0000; | |
| 2402 | |||
| 2403 | 4 | uint16_t* AVS_RESTRICT dst_ptr = dst + x + y_from * dst_pitch; | |
| 2404 | 4 | const uint16_t* src_ptr_0_15 = src + iStart_0_15 + y_from * src_pitch; | |
| 2405 | 4 | const uint16_t* src_ptr_16_31 = src + iStart_16_31 + y_from * src_pitch; | |
| 2406 | 4 | const uint16_t* src_ptr_32_47 = src + iStart_32_47 + y_from * src_pitch; | |
| 2407 | 4 | const uint16_t* src_ptr_48_63 = src + iStart_48_63 + y_from * src_pitch; | |
| 2408 | |||
| 2409 | 4 | const int remaining_0_15 = program->source_size - iStart_0_15; | |
| 2410 | 4 | const __mmask32 k1_0_15 = _bzhi_u32(~0UL, remaining_0_15); | |
| 2411 | 4 | const __mmask32 k2_0_15 = _bzhi_u32(~0UL, remaining_0_15 - 32); | |
| 2412 | 4 | const int remaining_16_31 = program->source_size - iStart_16_31; | |
| 2413 | 4 | const __mmask32 k1_16_31 = _bzhi_u32(~0UL, remaining_16_31); | |
| 2414 | 4 | const __mmask32 k2_16_31 = _bzhi_u32(~0UL, remaining_16_31 - 32); | |
| 2415 | 4 | const int remaining_32_47 = program->source_size - iStart_32_47; | |
| 2416 | 4 | const __mmask32 k1_32_47 = _bzhi_u32(~0UL, remaining_32_47); | |
| 2417 | 4 | const __mmask32 k2_32_47 = _bzhi_u32(~0UL, remaining_32_47 - 32); | |
| 2418 | 4 | const int remaining_48_63 = program->source_size - iStart_48_63; | |
| 2419 | 4 | const __mmask32 k1_48_63 = _bzhi_u32(~0UL, remaining_48_63); | |
| 2420 | 4 | const __mmask32 k2_48_63 = _bzhi_u32(~0UL, remaining_48_63 - 32); | |
| 2421 | |||
| 2422 |
8/16✓ Branch 296 → 131 taken 5 times.
✓ Branch 296 → 297 taken 1 time.
✗ Branch 312 → 131 not taken.
✗ Branch 312 → 313 not taken.
✓ Branch 328 → 131 taken 5 times.
✓ Branch 328 → 329 taken 1 time.
✗ Branch 344 → 131 not taken.
✗ Branch 344 → 345 not taken.
✓ Branch 360 → 131 taken 5 times.
✓ Branch 360 → 361 taken 1 time.
✗ Branch 376 → 131 not taken.
✗ Branch 376 → 377 not taken.
✓ Branch 392 → 131 taken 5 times.
✓ Branch 392 → 393 taken 1 time.
✗ Branch 408 → 131 not taken.
✗ Branch 408 → 409 not taken.
|
24 | for (int y = y_from; y < y_to; y++) |
| 2423 | { | ||
| 2424 | __m512i data_src_0_15, data_src2_0_15; | ||
| 2425 | __m512i data_src_16_31, data_src2_16_31; | ||
| 2426 | __m512i data_src_32_47, data_src2_32_47; | ||
| 2427 | __m512i data_src_48_63, data_src2_48_63; | ||
| 2428 | |||
| 2429 | 20 | __m512i perm_rNrNp1_0_31lo = perm_r0r1_0_31lo; | |
| 2430 | 20 | __m512i perm_rNrNp1_0_31hi = perm_r0r1_0_31hi; | |
| 2431 | 20 | __m512i perm_rNrNp1_32_63lo = perm_r0r1_32_63lo; | |
| 2432 | 20 | __m512i perm_rNrNp1_32_63hi = perm_r0r1_32_63hi; | |
| 2433 | |||
| 2434 | if constexpr (partial_load) { | ||
| 2435 | ✗ | data_src_0_15 = _mm512_maskz_loadu_epi16(k1_0_15, src_ptr_0_15); | |
| 2436 | ✗ | data_src_16_31 = _mm512_maskz_loadu_epi16(k1_16_31, src_ptr_16_31); | |
| 2437 | ✗ | data_src_32_47 = _mm512_maskz_loadu_epi16(k1_32_47, src_ptr_32_47); | |
| 2438 | ✗ | data_src_48_63 = _mm512_maskz_loadu_epi16(k1_48_63, src_ptr_48_63); | |
| 2439 | ✗ | data_src2_0_15 = _mm512_maskz_loadu_epi16(k2_0_15, src_ptr_0_15 + 32); | |
| 2440 | ✗ | data_src2_16_31 = _mm512_maskz_loadu_epi16(k2_16_31, src_ptr_16_31 + 32); | |
| 2441 | ✗ | data_src2_32_47 = _mm512_maskz_loadu_epi16(k2_32_47, src_ptr_32_47 + 32); | |
| 2442 | ✗ | data_src2_48_63 = _mm512_maskz_loadu_epi16(k2_48_63, src_ptr_48_63 + 32); | |
| 2443 | } | ||
| 2444 | else { | ||
| 2445 | 20 | data_src_0_15 = _mm512_loadu_si512(src_ptr_0_15); | |
| 2446 | 20 | data_src_16_31 = _mm512_loadu_si512(src_ptr_16_31); | |
| 2447 | 20 | data_src_32_47 = _mm512_loadu_si512(src_ptr_32_47); | |
| 2448 | 20 | data_src_48_63 = _mm512_loadu_si512(src_ptr_48_63); | |
| 2449 | 20 | data_src2_0_15 = _mm512_loadu_si512(src_ptr_0_15 + 32); | |
| 2450 | 20 | data_src2_16_31 = _mm512_loadu_si512(src_ptr_16_31 + 32); | |
| 2451 | 20 | data_src2_32_47 = _mm512_loadu_si512(src_ptr_32_47 + 32); | |
| 2452 | 40 | data_src2_48_63 = _mm512_loadu_si512(src_ptr_48_63 + 32); | |
| 2453 | } | ||
| 2454 | |||
| 2455 | __m512i result_0_31lo, result_0_31hi, result_32_63lo, result_32_63hi; | ||
| 2456 | |||
| 2457 | 40 | __m512i src_r0r1_0_31lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31lo, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31lo, data_src2_16_31)); | |
| 2458 | 40 | __m512i src_r0r1_0_31hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31hi, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31hi, data_src2_16_31)); | |
| 2459 | 40 | __m512i src_r0r1_32_63lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63lo, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63lo, data_src2_48_63)); | |
| 2460 | 40 | __m512i src_r0r1_32_63hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63hi, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63hi, data_src2_48_63)); | |
| 2461 | |||
| 2462 | 20 | perm_rNrNp1_0_31lo = _mm512_add_epi16(perm_rNrNp1_0_31lo, two_epi16); | |
| 2463 | 20 | perm_rNrNp1_0_31hi = _mm512_add_epi16(perm_rNrNp1_0_31hi, two_epi16); | |
| 2464 | 20 | perm_rNrNp1_32_63lo = _mm512_add_epi16(perm_rNrNp1_32_63lo, two_epi16); | |
| 2465 | 20 | perm_rNrNp1_32_63hi = _mm512_add_epi16(perm_rNrNp1_32_63hi, two_epi16); | |
| 2466 | |||
| 2467 | if constexpr (!lessthan16bit) { | ||
| 2468 | 10 | src_r0r1_0_31lo = _mm512_add_epi16(src_r0r1_0_31lo, shifttosigned); | |
| 2469 | 10 | src_r0r1_0_31hi = _mm512_add_epi16(src_r0r1_0_31hi, shifttosigned); | |
| 2470 | 10 | src_r0r1_32_63lo = _mm512_add_epi16(src_r0r1_32_63lo, shifttosigned); | |
| 2471 | 10 | src_r0r1_32_63hi = _mm512_add_epi16(src_r0r1_32_63hi, shifttosigned); | |
| 2472 | } | ||
| 2473 | |||
| 2474 | if constexpr (UseVNNI) | ||
| 2475 | { | ||
| 2476 | 10 | result_0_31lo = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31lo, coef_r0r1_0_31lo); | |
| 2477 | 10 | result_0_31hi = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31hi, coef_r0r1_0_31hi); | |
| 2478 | 10 | result_32_63lo = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63lo, coef_r0r1_32_63lo); | |
| 2479 | 20 | result_32_63hi = _mm512_dpwssd_epi32(rounder, src_r0r1_32_63hi, coef_r0r1_32_63hi); | |
| 2480 | } | ||
| 2481 | else | ||
| 2482 | { | ||
| 2483 | 20 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, coef_r0r1_0_31lo), rounder); | |
| 2484 | 20 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, coef_r0r1_0_31hi), rounder); | |
| 2485 | 20 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63lo, coef_r0r1_32_63lo), rounder); | |
| 2486 | 30 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63hi, coef_r0r1_32_63hi), rounder); | |
| 2487 | } | ||
| 2488 | |||
| 2489 | 40 | __m512i src_r2r3_0_31lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31lo, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31lo, data_src2_16_31)); | |
| 2490 | 40 | __m512i src_r2r3_0_31hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31hi, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31hi, data_src2_16_31)); | |
| 2491 | 40 | __m512i src_r2r3_32_63lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63lo, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63lo, data_src2_48_63)); | |
| 2492 | 40 | __m512i src_r2r3_32_63hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63hi, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63hi, data_src2_48_63)); | |
| 2493 | |||
| 2494 | 20 | perm_rNrNp1_0_31lo = _mm512_add_epi16(perm_rNrNp1_0_31lo, two_epi16); | |
| 2495 | 20 | perm_rNrNp1_0_31hi = _mm512_add_epi16(perm_rNrNp1_0_31hi, two_epi16); | |
| 2496 | 20 | perm_rNrNp1_32_63lo = _mm512_add_epi16(perm_rNrNp1_32_63lo, two_epi16); | |
| 2497 | 10 | perm_rNrNp1_32_63hi = _mm512_add_epi16(perm_rNrNp1_32_63hi, two_epi16); | |
| 2498 | |||
| 2499 | if constexpr (!lessthan16bit) { | ||
| 2500 | 10 | src_r2r3_0_31lo = _mm512_add_epi16(src_r2r3_0_31lo, shifttosigned); | |
| 2501 | 10 | src_r2r3_0_31hi = _mm512_add_epi16(src_r2r3_0_31hi, shifttosigned); | |
| 2502 | 10 | src_r2r3_32_63lo = _mm512_add_epi16(src_r2r3_32_63lo, shifttosigned); | |
| 2503 | 20 | src_r2r3_32_63hi = _mm512_add_epi16(src_r2r3_32_63hi, shifttosigned); | |
| 2504 | } | ||
| 2505 | |||
| 2506 | if constexpr (UseVNNI) | ||
| 2507 | { | ||
| 2508 | 10 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r2r3_0_31lo, coef_r2r3_0_31lo); | |
| 2509 | 10 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r2r3_0_31hi, coef_r2r3_0_31hi); | |
| 2510 | 10 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r2r3_32_63lo, coef_r2r3_32_63lo); | |
| 2511 | 10 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r2r3_32_63hi, coef_r2r3_32_63hi); | |
| 2512 | } | ||
| 2513 | else | ||
| 2514 | { | ||
| 2515 | 20 | result_0_31lo = _mm512_add_epi32(result_0_31lo, _mm512_madd_epi16(src_r2r3_0_31lo, coef_r2r3_0_31lo)); | |
| 2516 | 20 | result_0_31hi = _mm512_add_epi32(result_0_31hi, _mm512_madd_epi16(src_r2r3_0_31hi, coef_r2r3_0_31hi)); | |
| 2517 | 20 | result_32_63lo = _mm512_add_epi32(result_32_63lo, _mm512_madd_epi16(src_r2r3_32_63lo, coef_r2r3_32_63lo)); | |
| 2518 | 20 | result_32_63hi = _mm512_add_epi32(result_32_63hi, _mm512_madd_epi16(src_r2r3_32_63hi, coef_r2r3_32_63hi)); | |
| 2519 | } | ||
| 2520 | |||
| 2521 | 40 | __m512i src_r4r5_0_31lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31lo, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31lo, data_src2_16_31)); | |
| 2522 | 40 | __m512i src_r4r5_0_31hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31hi, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31hi, data_src2_16_31)); | |
| 2523 | 40 | __m512i src_r4r5_32_63lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63lo, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63lo, data_src2_48_63)); | |
| 2524 | 40 | __m512i src_r4r5_32_63hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63hi, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63hi, data_src2_48_63)); | |
| 2525 | |||
| 2526 | 20 | perm_rNrNp1_0_31lo = _mm512_add_epi16(perm_rNrNp1_0_31lo, two_epi16); | |
| 2527 | 20 | perm_rNrNp1_0_31hi = _mm512_add_epi16(perm_rNrNp1_0_31hi, two_epi16); | |
| 2528 | 20 | perm_rNrNp1_32_63lo = _mm512_add_epi16(perm_rNrNp1_32_63lo, two_epi16); | |
| 2529 | 10 | perm_rNrNp1_32_63hi = _mm512_add_epi16(perm_rNrNp1_32_63hi, two_epi16); | |
| 2530 | |||
| 2531 | if constexpr (!lessthan16bit) { | ||
| 2532 | 10 | src_r4r5_0_31lo = _mm512_add_epi16(src_r4r5_0_31lo, shifttosigned); | |
| 2533 | 10 | src_r4r5_0_31hi = _mm512_add_epi16(src_r4r5_0_31hi, shifttosigned); | |
| 2534 | 10 | src_r4r5_32_63lo = _mm512_add_epi16(src_r4r5_32_63lo, shifttosigned); | |
| 2535 | 20 | src_r4r5_32_63hi = _mm512_add_epi16(src_r4r5_32_63hi, shifttosigned); | |
| 2536 | } | ||
| 2537 | |||
| 2538 | if constexpr (UseVNNI) | ||
| 2539 | { | ||
| 2540 | 10 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r4r5_0_31lo, coef_r4r5_0_31lo); | |
| 2541 | 10 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r4r5_0_31hi, coef_r4r5_0_31hi); | |
| 2542 | 10 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r4r5_32_63lo, coef_r4r5_32_63lo); | |
| 2543 | 10 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r4r5_32_63hi, coef_r4r5_32_63hi); | |
| 2544 | } | ||
| 2545 | else | ||
| 2546 | { | ||
| 2547 | 20 | result_0_31lo = _mm512_add_epi32(result_0_31lo, _mm512_madd_epi16(src_r4r5_0_31lo, coef_r4r5_0_31lo)); | |
| 2548 | 20 | result_0_31hi = _mm512_add_epi32(result_0_31hi, _mm512_madd_epi16(src_r4r5_0_31hi, coef_r4r5_0_31hi)); | |
| 2549 | 20 | result_32_63lo = _mm512_add_epi32(result_32_63lo, _mm512_madd_epi16(src_r4r5_32_63lo, coef_r4r5_32_63lo)); | |
| 2550 | 20 | result_32_63hi = _mm512_add_epi32(result_32_63hi, _mm512_madd_epi16(src_r4r5_32_63hi, coef_r4r5_32_63hi)); | |
| 2551 | } | ||
| 2552 | |||
| 2553 | 40 | __m512i src_r6r7_0_31lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31lo, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31lo, data_src2_16_31)); | |
| 2554 | 40 | __m512i src_r6r7_0_31hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31hi, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31hi, data_src2_16_31)); | |
| 2555 | 40 | __m512i src_r6r7_32_63lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63lo, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63lo, data_src2_48_63)); | |
| 2556 | 30 | __m512i src_r6r7_32_63hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63hi, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63hi, data_src2_48_63)); | |
| 2557 | |||
| 2558 | if constexpr (!lessthan16bit) { | ||
| 2559 | 10 | src_r6r7_0_31lo = _mm512_add_epi16(src_r6r7_0_31lo, shifttosigned); | |
| 2560 | 10 | src_r6r7_0_31hi = _mm512_add_epi16(src_r6r7_0_31hi, shifttosigned); | |
| 2561 | 10 | src_r6r7_32_63lo = _mm512_add_epi16(src_r6r7_32_63lo, shifttosigned); | |
| 2562 | 20 | src_r6r7_32_63hi = _mm512_add_epi16(src_r6r7_32_63hi, shifttosigned); | |
| 2563 | } | ||
| 2564 | |||
| 2565 | if constexpr (UseVNNI) | ||
| 2566 | { | ||
| 2567 | 10 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r6r7_0_31lo, coef_r6r7_0_31lo); | |
| 2568 | 10 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r6r7_0_31hi, coef_r6r7_0_31hi); | |
| 2569 | 10 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r6r7_32_63lo, coef_r6r7_32_63lo); | |
| 2570 | 5 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r6r7_32_63hi, coef_r6r7_32_63hi); | |
| 2571 | } | ||
| 2572 | else | ||
| 2573 | { | ||
| 2574 | 20 | result_0_31lo = _mm512_add_epi32(result_0_31lo, _mm512_madd_epi16(src_r6r7_0_31lo, coef_r6r7_0_31lo)); | |
| 2575 | 20 | result_0_31hi = _mm512_add_epi32(result_0_31hi, _mm512_madd_epi16(src_r6r7_0_31hi, coef_r6r7_0_31hi)); | |
| 2576 | 20 | result_32_63lo = _mm512_add_epi32(result_32_63lo, _mm512_madd_epi16(src_r6r7_32_63lo, coef_r6r7_32_63lo)); | |
| 2577 | 15 | result_32_63hi = _mm512_add_epi32(result_32_63hi, _mm512_madd_epi16(src_r6r7_32_63hi, coef_r6r7_32_63hi)); | |
| 2578 | } | ||
| 2579 | |||
| 2580 | if constexpr (!lessthan16bit) { | ||
| 2581 | 10 | result_0_31lo = _mm512_add_epi32(result_0_31lo, shiftfromsigned); | |
| 2582 | 10 | result_0_31hi = _mm512_add_epi32(result_0_31hi, shiftfromsigned); | |
| 2583 | 10 | result_32_63lo = _mm512_add_epi32(result_32_63lo, shiftfromsigned); | |
| 2584 | 20 | result_32_63hi = _mm512_add_epi32(result_32_63hi, shiftfromsigned); | |
| 2585 | } | ||
| 2586 | |||
| 2587 | 20 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale16bits); | |
| 2588 | 20 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale16bits); | |
| 2589 | 20 | result_32_63lo = _mm512_srai_epi32(result_32_63lo, FPScale16bits); | |
| 2590 | 20 | result_32_63hi = _mm512_srai_epi32(result_32_63hi, FPScale16bits); | |
| 2591 | |||
| 2592 | 20 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 2593 | 10 | __m512i result_32_63_int16 = _mm512_packus_epi32(result_32_63lo, result_32_63hi); | |
| 2594 | |||
| 2595 | if constexpr (lessthan16bit) { | ||
| 2596 | 10 | result_0_31_int16 = _mm512_min_epu16(result_0_31_int16, clamp_limit); | |
| 2597 | 20 | result_32_63_int16 = _mm512_min_epu16(result_32_63_int16, clamp_limit); | |
| 2598 | } | ||
| 2599 | |||
| 2600 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), result_0_31_int16); | ||
| 2601 | |||
| 2602 | 20 | const int w_mod32 = width / 32 * 32; | |
| 2603 |
4/16✓ Branch 292 → 293 taken 5 times.
✗ Branch 292 → 295 not taken.
✗ Branch 308 → 309 not taken.
✗ Branch 308 → 311 not taken.
✓ Branch 324 → 325 taken 5 times.
✗ Branch 324 → 327 not taken.
✗ Branch 340 → 341 not taken.
✗ Branch 340 → 343 not taken.
✓ Branch 356 → 357 taken 5 times.
✗ Branch 356 → 359 not taken.
✗ Branch 372 → 373 not taken.
✗ Branch 372 → 375 not taken.
✓ Branch 388 → 389 taken 5 times.
✗ Branch 388 → 391 not taken.
✗ Branch 404 → 405 not taken.
✗ Branch 404 → 407 not taken.
|
20 | if (x < (w_mod32 - 32)) |
| 2604 | 20 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr + 32), result_32_63_int16); | |
| 2605 | |||
| 2606 | 20 | dst_ptr += dst_pitch; | |
| 2607 | 20 | src_ptr_0_15 += src_pitch; | |
| 2608 | 20 | src_ptr_16_31 += src_pitch; | |
| 2609 | 20 | src_ptr_32_47 += src_pitch; | |
| 2610 | 20 | src_ptr_48_63 += src_pitch; | |
| 2611 | } | ||
| 2612 | |||
| 2613 | 4 | current_coeff_SIMD += 16; | |
| 2614 | }; | ||
| 2615 | |||
| 2616 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 1 time.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 1 time.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 1 time.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 1 time.
✓ Branch 18 → 19 taken 1 time.
|
8 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 2617 | { | ||
| 2618 | 4 | do_h_integer_core(std::false_type{}); | |
| 2619 | } | ||
| 2620 | |||
| 2621 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 22 → 20 not taken.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 22 → 20 not taken.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 22 → 20 not taken.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 22 → 20 not taken.
✓ Branch 22 → 23 taken 1 time.
|
4 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 2622 | { | ||
| 2623 | ✗ | do_h_integer_core(std::true_type{}); | |
| 2624 | } | ||
| 2625 | } | ||
| 2626 | 4 | } | |
| 2627 | |||
| 2628 | template<bool lessthan16bit, bool UseVNNI> | ||
| 2629 | 4 | void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 2630 | { | ||
| 2631 | 4 | const uint16_t* src = (uint16_t*)src8; | |
| 2632 | 4 | uint16_t* AVS_RESTRICT dst = (uint16_t* AVS_RESTRICT)dst8; | |
| 2633 | 4 | dst_pitch = dst_pitch / sizeof(uint16_t); | |
| 2634 | 4 | src_pitch = src_pitch / sizeof(uint16_t); | |
| 2635 | |||
| 2636 | 4 | constexpr int PIXELS_AT_A_TIME = 32; | |
| 2637 | |||
| 2638 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 2 → 3 taken 1 time.
✗ Branch 2 → 4 not taken.
|
4 | const int width_safe_mod = (program->safelimit_64_pixels_each32th_target.overread_possible ? program->safelimit_64_pixels_each32th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 2639 | |||
| 2640 | 4 | const int max_scanlines = program->max_scanlines; | |
| 2641 | |||
| 2642 | 4 | const __m512i shifttosigned = _mm512_set1_epi16(-32768); | |
| 2643 | 4 | const __m512i shiftfromsigned = _mm512_set1_epi32(32768 << FPScale16bits); | |
| 2644 | |||
| 2645 | 4 | const int limit = (1 << bits_per_pixel) - 1; | |
| 2646 | 8 | __m512i clamp_limit = _mm512_set1_epi16((short)limit); | |
| 2647 | 4 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale16bits - 1)); | |
| 2648 | |||
| 2649 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 14 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
|
8 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 2650 | { | ||
| 2651 | 4 | int y_to = std::min(y_from + max_scanlines, height); | |
| 2652 | |||
| 2653 | 4 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 2654 | |||
| 2655 | 4 | int x = 0; | |
| 2656 | |||
| 2657 | 28 | auto do_h_integer_core = [&](auto partial_load) { | |
| 2658 | 24 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 2659 | |||
| 2660 | 24 | const __m512i coef_r0r1_0_31lo = _mm512_load_si512(current_coeff_SIMD + 0); | |
| 2661 | 24 | const __m512i coef_r0r1_0_31hi = _mm512_load_si512(current_coeff_SIMD + 1); | |
| 2662 | 24 | const __m512i coef_r2r3_0_31lo = _mm512_load_si512(current_coeff_SIMD + 2); | |
| 2663 | 24 | const __m512i coef_r2r3_0_31hi = _mm512_load_si512(current_coeff_SIMD + 3); | |
| 2664 | 24 | const __m512i coef_r4r5_0_31lo = _mm512_load_si512(current_coeff_SIMD + 4); | |
| 2665 | 24 | const __m512i coef_r4r5_0_31hi = _mm512_load_si512(current_coeff_SIMD + 5); | |
| 2666 | 24 | const __m512i coef_r6r7_0_31lo = _mm512_load_si512(current_coeff_SIMD + 6); | |
| 2667 | 24 | const __m512i coef_r6r7_0_31hi = _mm512_load_si512(current_coeff_SIMD + 7); | |
| 2668 | 24 | const __m512i coef_r8r9_0_31lo = _mm512_load_si512(current_coeff_SIMD + 8); | |
| 2669 | 24 | const __m512i coef_r8r9_0_31hi = _mm512_load_si512(current_coeff_SIMD + 9); | |
| 2670 | 24 | const __m512i coef_r10r11_0_31lo = _mm512_load_si512(current_coeff_SIMD + 10); | |
| 2671 | 24 | const __m512i coef_r10r11_0_31hi = _mm512_load_si512(current_coeff_SIMD + 11); | |
| 2672 | 24 | const __m512i coef_r12r13_0_31lo = _mm512_load_si512(current_coeff_SIMD + 12); | |
| 2673 | 24 | const __m512i coef_r12r13_0_31hi = _mm512_load_si512(current_coeff_SIMD + 13); | |
| 2674 | 24 | const __m512i coef_r14r15_0_31lo = _mm512_load_si512(current_coeff_SIMD + 14); | |
| 2675 | 24 | const __m512i coef_r14r15_0_31hi = _mm512_load_si512(current_coeff_SIMD + 15); | |
| 2676 | |||
| 2677 | 24 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 2678 | 24 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 2679 | |||
| 2680 | 24 | int iStart = program->pixel_offset[x]; | |
| 2681 | 24 | __m512i m512i_Start = _mm512_set1_epi32(iStart); | |
| 2682 | |||
| 2683 | 24 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, m512i_Start); | |
| 2684 | 24 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, m512i_Start); | |
| 2685 | |||
| 2686 | 24 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 2687 | 24 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 2688 | |||
| 2689 | 48 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 2690 | 24 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 2691 | |||
| 2692 | 24 | const __m512i perm_r0r1_0_31lo = _mm512_unpacklo_epi16(perm_0_0_31, perm_1_0_31); | |
| 2693 | 24 | const __m512i perm_r0r1_0_31hi = _mm512_unpackhi_epi16(perm_0_0_31, perm_1_0_31); | |
| 2694 | |||
| 2695 | 24 | const __m512i two_epi16 = _mm512_set1_epi16(2); | |
| 2696 | |||
| 2697 | 24 | uint16_t* AVS_RESTRICT dst_ptr = dst + x + y_from * dst_pitch; | |
| 2698 | 24 | const uint16_t* src_ptr = src + iStart + y_from * src_pitch; | |
| 2699 | |||
| 2700 | 24 | const int remaining = program->source_size - iStart; | |
| 2701 | 24 | const __mmask32 k1 = _bzhi_u32(~0UL, remaining); | |
| 2702 | 24 | const __mmask32 k2 = _bzhi_u32(~0UL, remaining - 32); | |
| 2703 | |||
| 2704 |
16/16✓ Branch 188 → 78 taken 20 times.
✓ Branch 188 → 189 taken 4 times.
✓ Branch 192 → 78 taken 10 times.
✓ Branch 192 → 193 taken 2 times.
✓ Branch 220 → 78 taken 20 times.
✓ Branch 220 → 221 taken 4 times.
✓ Branch 224 → 78 taken 10 times.
✓ Branch 224 → 225 taken 2 times.
✓ Branch 252 → 78 taken 20 times.
✓ Branch 252 → 253 taken 4 times.
✓ Branch 256 → 78 taken 10 times.
✓ Branch 256 → 257 taken 2 times.
✓ Branch 284 → 78 taken 20 times.
✓ Branch 284 → 285 taken 4 times.
✓ Branch 288 → 78 taken 10 times.
✓ Branch 288 → 289 taken 2 times.
|
144 | for (int y = y_from; y < y_to; y++) |
| 2705 | { | ||
| 2706 | __m512i data_src, data_src2; | ||
| 2707 | |||
| 2708 | 120 | __m512i perm_rNrNp1_0_31lo_w = perm_r0r1_0_31lo; | |
| 2709 | 120 | __m512i perm_rNrNp1_0_31hi_w = perm_r0r1_0_31hi; | |
| 2710 | |||
| 2711 | if constexpr (partial_load) { | ||
| 2712 | 40 | data_src = _mm512_maskz_loadu_epi16(k1, src_ptr); | |
| 2713 | 80 | data_src2 = _mm512_maskz_loadu_epi16(k2, src_ptr + 32); | |
| 2714 | } | ||
| 2715 | else { | ||
| 2716 | 80 | data_src = _mm512_loadu_si512(src_ptr); | |
| 2717 | 160 | data_src2 = _mm512_loadu_si512(src_ptr + 32); | |
| 2718 | } | ||
| 2719 | |||
| 2720 | 120 | __m512i src_r0r1_0_31lo = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 2721 | 120 | __m512i src_r0r1_0_31hi = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 2722 | |||
| 2723 | 120 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 2724 | 120 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 2725 | |||
| 2726 | 120 | __m512i src_r2r3_0_31lo = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 2727 | 120 | __m512i src_r2r3_0_31hi = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 2728 | |||
| 2729 | 120 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 2730 | 90 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 2731 | |||
| 2732 | __m512i result_0_31lo, result_0_31hi; | ||
| 2733 | |||
| 2734 | if constexpr (!lessthan16bit) { | ||
| 2735 | 60 | src_r0r1_0_31lo = _mm512_add_epi16(src_r0r1_0_31lo, shifttosigned); | |
| 2736 | 60 | src_r0r1_0_31hi = _mm512_add_epi16(src_r0r1_0_31hi, shifttosigned); | |
| 2737 | 60 | src_r2r3_0_31lo = _mm512_add_epi16(src_r2r3_0_31lo, shifttosigned); | |
| 2738 | 90 | src_r2r3_0_31hi = _mm512_add_epi16(src_r2r3_0_31hi, shifttosigned); | |
| 2739 | } | ||
| 2740 | |||
| 2741 | if constexpr (UseVNNI) | ||
| 2742 | { | ||
| 2743 | 120 | result_0_31lo = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31lo, coef_r0r1_0_31lo); | |
| 2744 | 60 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r2r3_0_31lo, coef_r2r3_0_31lo); | |
| 2745 | 120 | result_0_31hi = _mm512_dpwssd_epi32(rounder, src_r0r1_0_31hi, coef_r0r1_0_31hi); | |
| 2746 | 60 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r2r3_0_31hi, coef_r2r3_0_31hi); | |
| 2747 | } | ||
| 2748 | else | ||
| 2749 | { | ||
| 2750 | 180 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, coef_r0r1_0_31lo), _mm512_madd_epi16(src_r2r3_0_31lo, coef_r2r3_0_31lo)); | |
| 2751 | 180 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, coef_r0r1_0_31hi), _mm512_madd_epi16(src_r2r3_0_31hi, coef_r2r3_0_31hi)); | |
| 2752 | } | ||
| 2753 | |||
| 2754 | 120 | __m512i src_r4r5_0_31lo = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 2755 | 120 | __m512i src_r4r5_0_31hi = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 2756 | |||
| 2757 | 120 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 2758 | 120 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 2759 | |||
| 2760 | 120 | __m512i src_r6r7_0_31lo = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 2761 | 120 | __m512i src_r6r7_0_31hi = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 2762 | |||
| 2763 | 120 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 2764 | 60 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 2765 | |||
| 2766 | if constexpr (!lessthan16bit) { | ||
| 2767 | 60 | src_r4r5_0_31lo = _mm512_add_epi16(src_r4r5_0_31lo, shifttosigned); | |
| 2768 | 60 | src_r4r5_0_31hi = _mm512_add_epi16(src_r4r5_0_31hi, shifttosigned); | |
| 2769 | 60 | src_r6r7_0_31lo = _mm512_add_epi16(src_r6r7_0_31lo, shifttosigned); | |
| 2770 | 120 | src_r6r7_0_31hi = _mm512_add_epi16(src_r6r7_0_31hi, shifttosigned); | |
| 2771 | } | ||
| 2772 | |||
| 2773 | if constexpr (UseVNNI) | ||
| 2774 | { | ||
| 2775 | 60 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r4r5_0_31lo, coef_r4r5_0_31lo); | |
| 2776 | 60 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r6r7_0_31lo, coef_r6r7_0_31lo); | |
| 2777 | 60 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r4r5_0_31hi, coef_r4r5_0_31hi); | |
| 2778 | 60 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r6r7_0_31hi, coef_r6r7_0_31hi); | |
| 2779 | } | ||
| 2780 | else | ||
| 2781 | { | ||
| 2782 | 120 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_0_31lo, coef_r4r5_0_31lo), result_0_31lo); | |
| 2783 | 120 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r4r5_0_31hi, coef_r4r5_0_31hi), result_0_31hi); | |
| 2784 | 120 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_0_31lo, coef_r6r7_0_31lo), result_0_31lo); | |
| 2785 | 120 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r6r7_0_31hi, coef_r6r7_0_31hi), result_0_31hi); | |
| 2786 | } | ||
| 2787 | |||
| 2788 | 120 | __m512i src_r8r9_0_31lo = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 2789 | 120 | __m512i src_r8r9_0_31hi = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 2790 | |||
| 2791 | 120 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 2792 | 120 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 2793 | |||
| 2794 | 120 | __m512i src_r10r11_0_31lo = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 2795 | 120 | __m512i src_r10r11_0_31hi = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 2796 | |||
| 2797 | 120 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 2798 | 60 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 2799 | |||
| 2800 | if constexpr (!lessthan16bit) { | ||
| 2801 | 60 | src_r8r9_0_31lo = _mm512_add_epi16(src_r8r9_0_31lo, shifttosigned); | |
| 2802 | 60 | src_r8r9_0_31hi = _mm512_add_epi16(src_r8r9_0_31hi, shifttosigned); | |
| 2803 | 60 | src_r10r11_0_31lo = _mm512_add_epi16(src_r10r11_0_31lo, shifttosigned); | |
| 2804 | 120 | src_r10r11_0_31hi = _mm512_add_epi16(src_r10r11_0_31hi, shifttosigned); | |
| 2805 | } | ||
| 2806 | |||
| 2807 | if constexpr (UseVNNI) | ||
| 2808 | { | ||
| 2809 | 60 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r8r9_0_31lo, coef_r8r9_0_31lo); | |
| 2810 | 60 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r10r11_0_31lo, coef_r10r11_0_31lo); | |
| 2811 | 60 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r8r9_0_31hi, coef_r8r9_0_31hi); | |
| 2812 | 60 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r10r11_0_31hi, coef_r10r11_0_31hi); | |
| 2813 | } | ||
| 2814 | else | ||
| 2815 | { | ||
| 2816 | 120 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r8r9_0_31lo, coef_r8r9_0_31lo), result_0_31lo); | |
| 2817 | 120 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r8r9_0_31hi, coef_r8r9_0_31hi), result_0_31hi); | |
| 2818 | 120 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r10r11_0_31lo, coef_r10r11_0_31lo), result_0_31lo); | |
| 2819 | 120 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r10r11_0_31hi, coef_r10r11_0_31hi), result_0_31hi); | |
| 2820 | } | ||
| 2821 | |||
| 2822 | 120 | __m512i src_r12r13_0_31lo = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 2823 | 120 | __m512i src_r12r13_0_31hi = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 2824 | |||
| 2825 | 120 | perm_rNrNp1_0_31lo_w = _mm512_add_epi16(perm_rNrNp1_0_31lo_w, two_epi16); | |
| 2826 | 120 | perm_rNrNp1_0_31hi_w = _mm512_add_epi16(perm_rNrNp1_0_31hi_w, two_epi16); | |
| 2827 | |||
| 2828 | 120 | __m512i src_r14r15_0_31lo = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31lo_w, data_src2); | |
| 2829 | 60 | __m512i src_r14r15_0_31hi = _mm512_permutex2var_epi16(data_src, perm_rNrNp1_0_31hi_w, data_src2); | |
| 2830 | |||
| 2831 | if constexpr (!lessthan16bit) { | ||
| 2832 | 60 | src_r12r13_0_31lo = _mm512_add_epi16(src_r12r13_0_31lo, shifttosigned); | |
| 2833 | 60 | src_r12r13_0_31hi = _mm512_add_epi16(src_r12r13_0_31hi, shifttosigned); | |
| 2834 | 60 | src_r14r15_0_31lo = _mm512_add_epi16(src_r14r15_0_31lo, shifttosigned); | |
| 2835 | 120 | src_r14r15_0_31hi = _mm512_add_epi16(src_r14r15_0_31hi, shifttosigned); | |
| 2836 | } | ||
| 2837 | |||
| 2838 | if constexpr (UseVNNI) | ||
| 2839 | { | ||
| 2840 | 60 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r12r13_0_31lo, coef_r12r13_0_31lo); | |
| 2841 | 60 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r14r15_0_31lo, coef_r14r15_0_31lo); | |
| 2842 | 60 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r12r13_0_31hi, coef_r12r13_0_31hi); | |
| 2843 | 30 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r14r15_0_31hi, coef_r14r15_0_31hi); | |
| 2844 | } | ||
| 2845 | else | ||
| 2846 | { | ||
| 2847 | 120 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r12r13_0_31lo, coef_r12r13_0_31lo), result_0_31lo); | |
| 2848 | 120 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r12r13_0_31hi, coef_r12r13_0_31hi), result_0_31hi); | |
| 2849 | 120 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r14r15_0_31lo, coef_r14r15_0_31lo), result_0_31lo); | |
| 2850 | 60 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r14r15_0_31hi, coef_r14r15_0_31hi), result_0_31hi); | |
| 2851 | |||
| 2852 | 60 | result_0_31lo = _mm512_add_epi32(result_0_31lo, rounder); | |
| 2853 | 90 | result_0_31hi = _mm512_add_epi32(result_0_31hi, rounder); | |
| 2854 | } | ||
| 2855 | |||
| 2856 | if constexpr (!lessthan16bit) { | ||
| 2857 | 60 | result_0_31lo = _mm512_add_epi32(result_0_31lo, shiftfromsigned); | |
| 2858 | 120 | result_0_31hi = _mm512_add_epi32(result_0_31hi, shiftfromsigned); | |
| 2859 | } | ||
| 2860 | |||
| 2861 | 120 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale16bits); | |
| 2862 | 120 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale16bits); | |
| 2863 | |||
| 2864 | 60 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 2865 | |||
| 2866 | if constexpr (lessthan16bit) { | ||
| 2867 | 120 | result_0_31_int16 = _mm512_min_epu16(result_0_31_int16, clamp_limit); | |
| 2868 | } | ||
| 2869 | |||
| 2870 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), result_0_31_int16); | ||
| 2871 | |||
| 2872 | 120 | dst_ptr += dst_pitch; | |
| 2873 | 120 | src_ptr += src_pitch; | |
| 2874 | } | ||
| 2875 | |||
| 2876 | 24 | current_coeff_SIMD += 16; | |
| 2877 | }; | ||
| 2878 | |||
| 2879 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 4 times.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 4 times.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 4 times.
✓ Branch 18 → 19 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 18 → 16 taken 4 times.
✓ Branch 18 → 19 taken 1 time.
|
20 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 2880 | { | ||
| 2881 | 16 | do_h_integer_core(std::false_type{}); | |
| 2882 | } | ||
| 2883 | |||
| 2884 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 2 times.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 2 times.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 2 times.
✓ Branch 22 → 23 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 22 → 20 taken 2 times.
✓ Branch 22 → 23 taken 1 time.
|
12 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 2885 | { | ||
| 2886 | 8 | do_h_integer_core(std::true_type{}); | |
| 2887 | } | ||
| 2888 | } | ||
| 2889 | 4 | } | |
| 2890 | |||
| 2891 | template<bool lessthan16bit, bool UseVNNI> | ||
| 2892 | 4 | void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) | |
| 2893 | { | ||
| 2894 | 4 | int filter_size_real = program->filter_size_real; | |
| 2895 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 1 time.
|
4 | if ((filter_size_real / 2 * 2) != filter_size_real) filter_size_real++; |
| 2896 | |||
| 2897 | 4 | const uint16_t* src = (uint16_t*)src8; | |
| 2898 | 4 | uint16_t* AVS_RESTRICT dst = (uint16_t* AVS_RESTRICT)dst8; | |
| 2899 | 4 | dst_pitch = dst_pitch / sizeof(uint16_t); | |
| 2900 | 4 | src_pitch = src_pitch / sizeof(uint16_t); | |
| 2901 | |||
| 2902 | 4 | constexpr int PIXELS_AT_A_TIME = 64; | |
| 2903 | |||
| 2904 |
4/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 4 → 5 taken 1 time.
✗ Branch 4 → 6 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 4 → 5 taken 1 time.
✗ Branch 4 → 6 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 4 → 5 taken 1 time.
✗ Branch 4 → 6 not taken.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 4 → 5 taken 1 time.
✗ Branch 4 → 6 not taken.
|
4 | const int width_safe_mod = (program->safelimit_64_pixels_each32th_target.overread_possible ? program->safelimit_64_pixels_each32th_target.source_overread_beyond_targetx : width) / PIXELS_AT_A_TIME * PIXELS_AT_A_TIME; |
| 2905 | |||
| 2906 | 4 | const int max_scanlines = program->max_scanlines; | |
| 2907 | |||
| 2908 | 4 | const __m512i shifttosigned = _mm512_set1_epi16(-32768); | |
| 2909 | 4 | const __m512i shiftfromsigned = _mm512_set1_epi32(32768 << FPScale16bits); | |
| 2910 | |||
| 2911 | 4 | const int limit = (1 << bits_per_pixel) - 1; | |
| 2912 | 8 | __m512i clamp_limit = _mm512_set1_epi16((short)limit); | |
| 2913 | 4 | __m512i rounder = _mm512_set1_epi32(1 << (FPScale16bits - 1)); | |
| 2914 | |||
| 2915 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 26 → 16 taken 1 time.
✓ Branch 26 → 27 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 26 → 16 taken 1 time.
✓ Branch 26 → 27 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 26 → 16 taken 1 time.
✓ Branch 26 → 27 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 26 → 16 taken 1 time.
✓ Branch 26 → 27 taken 1 time.
|
8 | for (int y_from = 0; y_from < height; y_from += max_scanlines) |
| 2916 | { | ||
| 2917 | 4 | int y_to = std::min(y_from + max_scanlines, height); | |
| 2918 | |||
| 2919 | 4 | const __m512i* AVS_RESTRICT current_coeff_SIMD = (__m512i*)program->pixel_coefficient_AVX512_H; | |
| 2920 | |||
| 2921 | 4 | int x = 0; | |
| 2922 | |||
| 2923 | 16 | auto do_h_integer_core = [&](auto partial_load) { | |
| 2924 | 12 | __m512i one_epi16 = _mm512_set1_epi16(1); | |
| 2925 | |||
| 2926 | 12 | __m512i perm_0_0_15 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x])); | |
| 2927 | 12 | __m512i perm_0_16_31 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 16])); | |
| 2928 | 12 | __m512i perm_0_32_47 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 32])); | |
| 2929 | 12 | __m512i perm_0_48_63 = _mm512_loadu_si512((__m512i*)(&program->pixel_offset[x + 48])); | |
| 2930 | |||
| 2931 | 12 | int iStart_0_15 = program->pixel_offset[x]; | |
| 2932 | 12 | int iStart_16_31 = program->pixel_offset[x + 16]; | |
| 2933 | 12 | int iStart_32_47 = program->pixel_offset[x + 32]; | |
| 2934 | 12 | int iStart_48_63 = program->pixel_offset[x + 48]; | |
| 2935 | |||
| 2936 | 24 | perm_0_0_15 = _mm512_sub_epi32(perm_0_0_15, _mm512_set1_epi32(iStart_0_15)); | |
| 2937 | 24 | perm_0_16_31 = _mm512_sub_epi32(perm_0_16_31, _mm512_set1_epi32(iStart_16_31)); | |
| 2938 | 24 | perm_0_32_47 = _mm512_sub_epi32(perm_0_32_47, _mm512_set1_epi32(iStart_32_47)); | |
| 2939 | 24 | perm_0_48_63 = _mm512_sub_epi32(perm_0_48_63, _mm512_set1_epi32(iStart_48_63)); | |
| 2940 | |||
| 2941 | 12 | __m256i m256i_perm_0_0_15 = _mm512_cvtepi32_epi16(perm_0_0_15); | |
| 2942 | 12 | __m256i m256i_perm_0_16_31 = _mm512_cvtepi32_epi16(perm_0_16_31); | |
| 2943 | 12 | __m256i m256i_perm_0_32_47 = _mm512_cvtepi32_epi16(perm_0_32_47); | |
| 2944 | 12 | __m256i m256i_perm_0_48_63 = _mm512_cvtepi32_epi16(perm_0_48_63); | |
| 2945 | |||
| 2946 | 24 | __m512i perm_0_0_31 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_0_15), m256i_perm_0_16_31, 1); | |
| 2947 | 24 | __m512i perm_0_32_63 = _mm512_inserti64x4(_mm512_castsi256_si512(m256i_perm_0_32_47), m256i_perm_0_48_63, 1); | |
| 2948 | |||
| 2949 | 12 | __m512i perm_1_0_31 = _mm512_add_epi16(perm_0_0_31, one_epi16); | |
| 2950 | 12 | __m512i perm_1_32_63 = _mm512_add_epi16(perm_0_32_63, one_epi16); | |
| 2951 | |||
| 2952 | 12 | const __m512i perm_r0r1_0_31lo = _mm512_unpacklo_epi16(perm_0_0_31, perm_1_0_31); | |
| 2953 | 12 | const __m512i perm_r0r1_0_31hi = _mm512_unpackhi_epi16(perm_0_0_31, perm_1_0_31); | |
| 2954 | 12 | const __m512i perm_r0r1_32_63lo = _mm512_unpacklo_epi16(perm_0_32_63, perm_1_32_63); | |
| 2955 | 12 | const __m512i perm_r0r1_32_63hi = _mm512_unpackhi_epi16(perm_0_32_63, perm_1_32_63); | |
| 2956 | |||
| 2957 | 12 | const __m512i two_epi16 = _mm512_set1_epi16(2); | |
| 2958 | 12 | const __mmask32 k_high = 0xFFFF0000; | |
| 2959 | |||
| 2960 | 12 | uint16_t* AVS_RESTRICT dst_ptr = dst + x + y_from * dst_pitch; | |
| 2961 | 12 | const uint16_t* src_ptr_0_15 = src + iStart_0_15 + y_from * src_pitch; | |
| 2962 | 12 | const uint16_t* src_ptr_16_31 = src + iStart_16_31 + y_from * src_pitch; | |
| 2963 | 12 | const uint16_t* src_ptr_32_47 = src + iStart_32_47 + y_from * src_pitch; | |
| 2964 | 12 | const uint16_t* src_ptr_48_63 = src + iStart_48_63 + y_from * src_pitch; | |
| 2965 | |||
| 2966 | 12 | const int remaining_0_15 = program->source_size - iStart_0_15; | |
| 2967 | 12 | const __mmask32 k1_0_15 = _bzhi_u32(~0UL, remaining_0_15); | |
| 2968 | 12 | const __mmask32 k2_0_15 = _bzhi_u32(~0UL, remaining_0_15 - 32); | |
| 2969 | 12 | const int remaining_16_31 = program->source_size - iStart_16_31; | |
| 2970 | 12 | const __mmask32 k1_16_31 = _bzhi_u32(~0UL, remaining_16_31); | |
| 2971 | 12 | const __mmask32 k2_16_31 = _bzhi_u32(~0UL, remaining_16_31 - 32); | |
| 2972 | 12 | const int remaining_32_47 = program->source_size - iStart_32_47; | |
| 2973 | 12 | const __mmask32 k1_32_47 = _bzhi_u32(~0UL, remaining_32_47); | |
| 2974 | 12 | const __mmask32 k2_32_47 = _bzhi_u32(~0UL, remaining_32_47 - 32); | |
| 2975 | 12 | const int remaining_48_63 = program->source_size - iStart_48_63; | |
| 2976 | 12 | const __mmask32 k1_48_63 = _bzhi_u32(~0UL, remaining_48_63); | |
| 2977 | 12 | const __mmask32 k2_48_63 = _bzhi_u32(~0UL, remaining_48_63 - 32); | |
| 2978 | |||
| 2979 |
12/12✓ Branch 187 → 99 taken 10 times.
✓ Branch 187 → 188 taken 2 times.
✓ Branch 195 → 99 taken 10 times.
✓ Branch 195 → 196 taken 2 times.
✓ Branch 203 → 99 taken 15 times.
✓ Branch 203 → 204 taken 3 times.
✓ Branch 211 → 99 taken 15 times.
✓ Branch 211 → 212 taken 3 times.
✓ Branch 219 → 99 taken 5 times.
✓ Branch 219 → 220 taken 1 time.
✓ Branch 227 → 99 taken 5 times.
✓ Branch 227 → 228 taken 1 time.
|
72 | for (int y = y_from; y < y_to; y++) |
| 2980 | { | ||
| 2981 | __m512i data_src_0_15, data_src2_0_15; | ||
| 2982 | __m512i data_src_16_31, data_src2_16_31; | ||
| 2983 | __m512i data_src_32_47, data_src2_32_47; | ||
| 2984 | __m512i data_src_48_63, data_src2_48_63; | ||
| 2985 | |||
| 2986 | 60 | __m512i perm_rNrNp1_0_31lo = perm_r0r1_0_31lo; | |
| 2987 | 60 | __m512i perm_rNrNp1_0_31hi = perm_r0r1_0_31hi; | |
| 2988 | 60 | __m512i perm_rNrNp1_32_63lo = perm_r0r1_32_63lo; | |
| 2989 | 60 | __m512i perm_rNrNp1_32_63hi = perm_r0r1_32_63hi; | |
| 2990 | |||
| 2991 | if constexpr (partial_load) { | ||
| 2992 | 20 | data_src_0_15 = _mm512_maskz_loadu_epi16(k1_0_15, src_ptr_0_15); | |
| 2993 | 20 | data_src_16_31 = _mm512_maskz_loadu_epi16(k1_16_31, src_ptr_16_31); | |
| 2994 | 20 | data_src_32_47 = _mm512_maskz_loadu_epi16(k1_32_47, src_ptr_32_47); | |
| 2995 | 20 | data_src_48_63 = _mm512_maskz_loadu_epi16(k1_48_63, src_ptr_48_63); | |
| 2996 | 20 | data_src2_0_15 = _mm512_maskz_loadu_epi16(k2_0_15, src_ptr_0_15 + 32); | |
| 2997 | 20 | data_src2_16_31 = _mm512_maskz_loadu_epi16(k2_16_31, src_ptr_16_31 + 32); | |
| 2998 | 20 | data_src2_32_47 = _mm512_maskz_loadu_epi16(k2_32_47, src_ptr_32_47 + 32); | |
| 2999 | 20 | data_src2_48_63 = _mm512_maskz_loadu_epi16(k2_48_63, src_ptr_48_63 + 32); | |
| 3000 | } | ||
| 3001 | else { | ||
| 3002 | 40 | data_src_0_15 = _mm512_loadu_si512(src_ptr_0_15); | |
| 3003 | 40 | data_src_16_31 = _mm512_loadu_si512(src_ptr_16_31); | |
| 3004 | 40 | data_src_32_47 = _mm512_loadu_si512(src_ptr_32_47); | |
| 3005 | 40 | data_src_48_63 = _mm512_loadu_si512(src_ptr_48_63); | |
| 3006 | 40 | data_src2_0_15 = _mm512_loadu_si512(src_ptr_0_15 + 32); | |
| 3007 | 40 | data_src2_16_31 = _mm512_loadu_si512(src_ptr_16_31 + 32); | |
| 3008 | 40 | data_src2_32_47 = _mm512_loadu_si512(src_ptr_32_47 + 32); | |
| 3009 | 40 | data_src2_48_63 = _mm512_loadu_si512(src_ptr_48_63 + 32); | |
| 3010 | } | ||
| 3011 | |||
| 3012 | 60 | __m512i result_0_31lo = rounder; | |
| 3013 | 60 | __m512i result_0_31hi = rounder; | |
| 3014 | 60 | __m512i result_32_63lo = rounder; | |
| 3015 | 60 | __m512i result_32_63hi = rounder; | |
| 3016 | |||
| 3017 | 60 | const __m512i* current_coeff_SIMDw = current_coeff_SIMD; | |
| 3018 | |||
| 3019 |
14/14✓ Branch 157 → 116 taken 100 times.
✓ Branch 157 → 158 taken 10 times.
✓ Branch 165 → 116 taken 100 times.
✓ Branch 165 → 166 taken 10 times.
✓ Branch 173 → 116 taken 100 times.
✓ Branch 173 → 132 taken 50 times.
✓ Branch 173 → 174 taken 15 times.
✓ Branch 181 → 116 taken 100 times.
✓ Branch 181 → 132 taken 50 times.
✓ Branch 181 → 182 taken 15 times.
✓ Branch 189 → 132 taken 50 times.
✓ Branch 189 → 190 taken 5 times.
✓ Branch 197 → 132 taken 50 times.
✓ Branch 197 → 198 taken 5 times.
|
660 | for (int kr = 0; kr < filter_size_real; kr += 2) |
| 3020 | { | ||
| 3021 | 1200 | __m512i src_r0r1_0_31lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31lo, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31lo, data_src2_16_31)); | |
| 3022 | 1200 | __m512i src_r0r1_0_31hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_0_15, perm_rNrNp1_0_31hi, data_src2_0_15), _mm512_permutex2var_epi16(data_src_16_31, perm_rNrNp1_0_31hi, data_src2_16_31)); | |
| 3023 | 1200 | __m512i src_r0r1_32_63lo = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63lo, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63lo, data_src2_48_63)); | |
| 3024 | 1200 | __m512i src_r0r1_32_63hi = _mm512_mask_blend_epi16(k_high, _mm512_permutex2var_epi16(data_src_32_47, perm_rNrNp1_32_63hi, data_src2_32_47), _mm512_permutex2var_epi16(data_src_48_63, perm_rNrNp1_32_63hi, data_src2_48_63)); | |
| 3025 | |||
| 3026 | 600 | perm_rNrNp1_0_31lo = _mm512_add_epi16(perm_rNrNp1_0_31lo, two_epi16); | |
| 3027 | 600 | perm_rNrNp1_0_31hi = _mm512_add_epi16(perm_rNrNp1_0_31hi, two_epi16); | |
| 3028 | 600 | perm_rNrNp1_32_63lo = _mm512_add_epi16(perm_rNrNp1_32_63lo, two_epi16); | |
| 3029 | 300 | perm_rNrNp1_32_63hi = _mm512_add_epi16(perm_rNrNp1_32_63hi, two_epi16); | |
| 3030 | |||
| 3031 | if constexpr (!lessthan16bit) { | ||
| 3032 | 300 | src_r0r1_0_31lo = _mm512_add_epi16(src_r0r1_0_31lo, shifttosigned); | |
| 3033 | 300 | src_r0r1_0_31hi = _mm512_add_epi16(src_r0r1_0_31hi, shifttosigned); | |
| 3034 | 300 | src_r0r1_32_63lo = _mm512_add_epi16(src_r0r1_32_63lo, shifttosigned); | |
| 3035 | 600 | src_r0r1_32_63hi = _mm512_add_epi16(src_r0r1_32_63hi, shifttosigned); | |
| 3036 | } | ||
| 3037 | |||
| 3038 | if constexpr (UseVNNI) | ||
| 3039 | { | ||
| 3040 | 300 | result_0_31lo = _mm512_dpwssd_epi32(result_0_31lo, src_r0r1_0_31lo, _mm512_load_si512(current_coeff_SIMDw + 0)); | |
| 3041 | 600 | result_0_31hi = _mm512_dpwssd_epi32(result_0_31hi, src_r0r1_0_31hi, _mm512_load_si512(current_coeff_SIMDw + 1)); | |
| 3042 | 600 | result_32_63lo = _mm512_dpwssd_epi32(result_32_63lo, src_r0r1_32_63lo, _mm512_load_si512(current_coeff_SIMDw + 2)); | |
| 3043 | 600 | result_32_63hi = _mm512_dpwssd_epi32(result_32_63hi, src_r0r1_32_63hi, _mm512_load_si512(current_coeff_SIMDw + 3)); | |
| 3044 | } | ||
| 3045 | else | ||
| 3046 | { | ||
| 3047 | 600 | result_0_31lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31lo, _mm512_load_si512(current_coeff_SIMDw + 0)), result_0_31lo); | |
| 3048 | 900 | result_0_31hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_0_31hi, _mm512_load_si512(current_coeff_SIMDw + 1)), result_0_31hi); | |
| 3049 | 900 | result_32_63lo = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63lo, _mm512_load_si512(current_coeff_SIMDw + 2)), result_32_63lo); | |
| 3050 | 900 | result_32_63hi = _mm512_add_epi32(_mm512_madd_epi16(src_r0r1_32_63hi, _mm512_load_si512(current_coeff_SIMDw + 3)), result_32_63hi); | |
| 3051 | } | ||
| 3052 | |||
| 3053 | 600 | current_coeff_SIMDw += 4; | |
| 3054 | } | ||
| 3055 | |||
| 3056 | if constexpr (!lessthan16bit) { | ||
| 3057 | 30 | result_0_31lo = _mm512_add_epi32(result_0_31lo, shiftfromsigned); | |
| 3058 | 30 | result_0_31hi = _mm512_add_epi32(result_0_31hi, shiftfromsigned); | |
| 3059 | 30 | result_32_63lo = _mm512_add_epi32(result_32_63lo, shiftfromsigned); | |
| 3060 | 60 | result_32_63hi = _mm512_add_epi32(result_32_63hi, shiftfromsigned); | |
| 3061 | } | ||
| 3062 | |||
| 3063 | 60 | result_0_31lo = _mm512_srai_epi32(result_0_31lo, FPScale16bits); | |
| 3064 | 60 | result_0_31hi = _mm512_srai_epi32(result_0_31hi, FPScale16bits); | |
| 3065 | 60 | result_32_63lo = _mm512_srai_epi32(result_32_63lo, FPScale16bits); | |
| 3066 | 60 | result_32_63hi = _mm512_srai_epi32(result_32_63hi, FPScale16bits); | |
| 3067 | |||
| 3068 | 60 | __m512i result_0_31_int16 = _mm512_packus_epi32(result_0_31lo, result_0_31hi); | |
| 3069 | 30 | __m512i result_32_63_int16 = _mm512_packus_epi32(result_32_63lo, result_32_63hi); | |
| 3070 | |||
| 3071 | if constexpr (lessthan16bit) { | ||
| 3072 | 30 | result_0_31_int16 = _mm512_min_epu16(result_0_31_int16, clamp_limit); | |
| 3073 | 60 | result_32_63_int16 = _mm512_min_epu16(result_32_63_int16, clamp_limit); | |
| 3074 | } | ||
| 3075 | |||
| 3076 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr), result_0_31_int16); | ||
| 3077 | |||
| 3078 | 60 | const int w_mod32 = width / 32 * 32; | |
| 3079 |
6/12✓ Branch 183 → 184 taken 10 times.
✗ Branch 183 → 186 not taken.
✓ Branch 191 → 192 taken 10 times.
✗ Branch 191 → 194 not taken.
✓ Branch 199 → 200 taken 15 times.
✗ Branch 199 → 202 not taken.
✓ Branch 207 → 208 taken 15 times.
✗ Branch 207 → 210 not taken.
✓ Branch 215 → 216 taken 5 times.
✗ Branch 215 → 218 not taken.
✓ Branch 223 → 224 taken 5 times.
✗ Branch 223 → 226 not taken.
|
60 | if (x < (w_mod32 - 32)) |
| 3080 | 60 | _mm512_stream_si512(reinterpret_cast<__m512i*>(dst_ptr + 32), result_32_63_int16); | |
| 3081 | |||
| 3082 | 60 | dst_ptr += dst_pitch; | |
| 3083 | 60 | src_ptr_0_15 += src_pitch; | |
| 3084 | 60 | src_ptr_16_31 += src_pitch; | |
| 3085 | 60 | src_ptr_32_47 += src_pitch; | |
| 3086 | 60 | src_ptr_48_63 += src_pitch; | |
| 3087 | } | ||
| 3088 | |||
| 3089 | 12 | current_coeff_SIMD += filter_size_real * 2; | |
| 3090 | }; | ||
| 3091 | |||
| 3092 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 20 → 18 taken 2 times.
✓ Branch 20 → 21 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 20 → 18 taken 2 times.
✓ Branch 20 → 21 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 20 → 18 taken 2 times.
✓ Branch 20 → 21 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 20 → 18 taken 2 times.
✓ Branch 20 → 21 taken 1 time.
|
12 | for (; x < width_safe_mod; x += PIXELS_AT_A_TIME) |
| 3093 | { | ||
| 3094 | 8 | do_h_integer_core(std::false_type{}); | |
| 3095 | } | ||
| 3096 | |||
| 3097 |
8/8void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<false, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, false>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
void resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_internal<true, true>(unsigned char*, unsigned char const*, int, int, ResamplingProgram*, int, int, int):
✓ Branch 24 → 22 taken 1 time.
✓ Branch 24 → 25 taken 1 time.
|
8 | for (; x < width; x += PIXELS_AT_A_TIME) |
| 3098 | { | ||
| 3099 | 4 | do_h_integer_core(std::true_type{}); | |
| 3100 | } | ||
| 3101 | } | ||
| 3102 | 4 | } | |
| 3103 |