1// Copyright 2020 Google LLC 2// 3// This source code is licensed under the BSD-style license found in the 4// LICENSE file in the root directory of this source tree. 5 6$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 7$assert CHANNEL_TILE % 8 == 0 8$assert CHANNEL_TILE >= 8 9$assert KERNEL_TILE >= 2 10#include <assert.h> 11 12#include <immintrin.h> 13 14#include <xnnpack/dwconv.h> 15 16 17void xnn_qs8_dwconv_minmax_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__avx2_mul32( 18 size_t channels, 19 size_t output_width, 20 const int8_t** input, 21 const void* weights, 22 int8_t* output, 23 size_t input_stride, 24 size_t output_increment, 25 size_t input_offset, 26 const int8_t* zero, 27 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN 28{ 29 assert(channels != 0); 30 assert(output_width != 0); 31 32 do { 33 $for K in range(KERNEL_TILE): 34 const int8_t* i${K} = input[${K}]; 35 assert(i${K} != NULL); 36 if XNN_UNPREDICTABLE(i${K} != zero) { 37 i${K} = (const int8_t*) ((uintptr_t) i${K} + input_offset); 38 } 39 input = (const int8_t**) ((uintptr_t) input + input_stride); 40 41 size_t c = channels; 42 const void* w = weights; 43 for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) { 44 __m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w); 45 $for C in range(8, CHANNEL_TILE, 8): 46 __m256i vacc${ABC[C:C+8]} = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + ${C} * sizeof(int32_t))); 47 48 $for K in range(KERNEL_TILE): 49 50 $for C in range(0, CHANNEL_TILE, 8): 51 $if C == 0: 52 const __m256i vi${K}x${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i${K})); 53 $else: 54 const __m256i vi${K}x${ABC[C:C+8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i${K} + ${C}))); 55 const __m256i vk${K}x${ABC[C:C+8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(int8_t)))); 56 i${K} += ${CHANNEL_TILE}; 57 58 $for C in range(0, CHANNEL_TILE, 8): 59 vacc${ABC[C:C+8]} = _mm256_add_epi32(vacc${ABC[C:C+8]}, _mm256_mullo_epi32(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]})); 60 61 w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(int8_t)); 62 63 const __m256i vmultiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.multiplier)); 64 const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding)); 65 66 $for C in range(0, CHANNEL_TILE, 8): 67 const __m256i vacc${ABC[C+1:C+8:2]} = _mm256_shuffle_epi32(vacc${ABC[C:C+8]}, _MM_SHUFFLE(3, 3, 1, 1)); 68 69 $for C in range(0, CHANNEL_TILE, 8): 70 const __m256i vprod${ABC[C:C+8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[C:C+8]}, vmultiplier), vrounding); 71 const __m256i vprod${ABC[C+1:C+8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[C+1:C+8:2]}, vmultiplier), vrounding); 72 73 $for C in range(0, CHANNEL_TILE, 8): 74 const __m256i vq31prod${ABC[C:C+8:2]} = _mm256_srli_epi64(vprod${ABC[C:C+8:2]}, 31); 75 const __m256i vq31prod${ABC[C+1:C+8:2]} = _mm256_add_epi64(vprod${ABC[C+1:C+8:2]}, vprod${ABC[C+1:C+8:2]}); 76 77 $for C in range(0, CHANNEL_TILE, 8): 78 const __m256i vq31prod${ABC[C:C+8]} = _mm256_blend_epi16(vq31prod${ABC[C:C+8:2]}, vq31prod${ABC[C+1:C+8:2]}, 0xCC); 79 80 const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask)); 81 $for C in range(0, CHANNEL_TILE, 8): 82 const __m256i vrem${ABC[C:C+8]} = 83 _mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[C:C+8]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[C:C+8]})); 84 85 const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold)); 86 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift); 87 $for C in range(0, CHANNEL_TILE, 8): 88 vacc${ABC[C:C+8]} = 89 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[C:C+8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[C:C+8]}, vremainder_threshold)); 90 91 $if CHANNEL_TILE > 8: 92 const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point)); 93 $else: 94 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); 95 $for C in range(0, CHANNEL_TILE, 16): 96 $if C + 8 < CHANNEL_TILE: 97 __m256i vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_adds_epi16(_mm256_packs_epi32(vacc${ABC[C:C+8]}, vacc${ABC[C+8:C+16]}), voutput_zero_point); 98 $elif CHANNEL_TILE > 8: 99 __m128i vout${ABC[C:C+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[C:C+8]}), _mm256_extracti128_si256(vacc${ABC[C:C+8]}, 1)), _mm256_castsi256_si128(voutput_zero_point)); 100 $else: 101 __m128i vout${ABC[C:C+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[C:C+8]}), _mm256_extracti128_si256(vacc${ABC[C:C+8]}, 1)), voutput_zero_point); 102 103 $if CHANNEL_TILE > 8: 104 const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min)); 105 const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max)); 106 $else: 107 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); 108 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max); 109 $for C in range(0, CHANNEL_TILE, 16): 110 $if C + 8 < CHANNEL_TILE: 111 vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_min_epi16(_mm256_max_epi16(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}, voutput_min), voutput_max); 112 $elif CHANNEL_TILE > 8: 113 vout${ABC[C:C+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[C:C+8]}, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max)); 114 $else: 115 vout${ABC[C:C+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[C:C+8]}, voutput_min), voutput_max); 116 117 $for C in range(0, CHANNEL_TILE, 16): 118 $if C + 8 < CHANNEL_TILE: 119 __m128i vout${ABC[C:C+16]} = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}), _mm256_extracti128_si256(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0)); 120 $else: 121 __m128i vout${ABC[C:C+8]}${ABC[C:C+8]} = _mm_packs_epi16(vout${ABC[C:C+8]}, vout${ABC[C:C+8]}); 122 123 $if CHANNEL_TILE > 8: 124 _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]}); 125 $else: 126 _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[C:C+8]}); 127 $for C in range(16, CHANNEL_TILE, 16): 128 $if C + 8 < CHANNEL_TILE: 129 _mm_storeu_si128((__m128i*) (output + ${C}), vout${ABC[C:C+16]}); 130 $else: 131 _mm_storel_epi64((__m128i*) (output + ${C}), vout${ABC[C:C+8]}${ABC[C:C+8]}); 132 output += ${CHANNEL_TILE}; 133 } 134 if XNN_UNLIKELY(c != 0) { 135 $if CHANNEL_TILE > 8: 136 const int8_t* k = (const int8_t*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t)); 137 ${"do " if CHANNEL_TILE > 8 else ""}{ 138 __m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w); 139 140 $for K in range(KERNEL_TILE): 141 142 const __m256i vi${K}x${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i${K})); 143 $if CHANNEL_TILE > 8: 144 $if K == 0: 145 const __m256i vk${K}x${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) k)); 146 $else: 147 const __m256i vk${K}x${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + ${K * CHANNEL_TILE}))); 148 $else: 149 const __m256i vk${K}x${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(int8_t)))); 150 $if CHANNEL_TILE > 8: 151 i${K} += 8; 152 153 vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_mullo_epi32(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]})); 154 155 $if CHANNEL_TILE > 8: 156 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t)); 157 k += 8; 158 159 const __m256i vmultiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.multiplier)); 160 const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding)); 161 162 const __m256i vacc${ABC[1:8:2]} = _mm256_shuffle_epi32(vacc${ABC[0:8]}, _MM_SHUFFLE(3, 3, 1, 1)); 163 164 const __m256i vprod${ABC[0:8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[0:8]}, vmultiplier), vrounding); 165 const __m256i vprod${ABC[1:8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[1:8:2]}, vmultiplier), vrounding); 166 167 const __m256i vq31prod${ABC[0:8:2]} = _mm256_srli_epi64(vprod${ABC[0:8:2]}, 31); 168 const __m256i vq31prod${ABC[1:8:2]} = _mm256_add_epi64(vprod${ABC[1:8:2]}, vprod${ABC[1:8:2]}); 169 170 const __m256i vq31prod${ABC[0:8]} = _mm256_blend_epi16(vq31prod${ABC[0:8:2]}, vq31prod${ABC[1:8:2]}, 0xCC); 171 172 const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask)); 173 const __m256i vrem${ABC[0:8]} = 174 _mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[0:8]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[0:8]})); 175 176 const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold)); 177 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift); 178 vacc${ABC[0:8]} = 179 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[0:8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[0:8]}, vremainder_threshold)); 180 181 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); 182 __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point); 183 184 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); 185 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max); 186 vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, voutput_min), voutput_max); 187 188 __m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]}); 189 190 $if CHANNEL_TILE > 8: 191 if XNN_LIKELY(c >= 8) { 192 _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]}); 193 output += 8; 194 c -= 8; 195 } else { 196 if (c & 4) { 197 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]}); 198 vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32); 199 output += 4; 200 } 201 if (c & 2) { 202 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0); 203 vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16); 204 output += 2; 205 } 206 if (c & 1) { 207 *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0); 208 output += 1; 209 } 210 c = 0; 211 } 212 $else: 213 if (c & 4) { 214 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]}); 215 vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32); 216 output += 4; 217 } 218 if (c & 2) { 219 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0); 220 vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16); 221 output += 2; 222 } 223 if (c & 1) { 224 *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0); 225 output += 1; 226 } 227 }${" while (c != 0);" if CHANNEL_TILE > 8 else ""} 228 } 229 230 output = (int8_t*) ((uintptr_t) output + output_increment); 231 } while (--output_width != 0); 232} 233