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 % 16 == 0 8$assert CHANNEL_TILE >= 16 9$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_mul16( 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, 16): 51 $if C == 0: 52 const __m256i vi${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i${K})); 53 $else: 54 const __m256i vi${K}x${ABC[C:C+16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i${K} + ${C}))); 55 const __m256i vk${K}x${ABC[C:C+16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(int8_t)))); 56 i${K} += ${CHANNEL_TILE}; 57 58 $for C in range(0, CHANNEL_TILE, 16): 59 const __m256i vprod${K}x${ABC[C:C+16]} = _mm256_mullo_epi16(vi${K}x${ABC[C:C+16]}, vk${K}x${ABC[C:C+16]}); 60 const __m128i vprod${K}x${ABC[C+8:C+16]} = _mm256_extracti128_si256(vprod${K}x${ABC[C:C+16]}, 1); 61 vacc${ABC[C:C+8]} = _mm256_add_epi32(vacc${ABC[C:C+8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod${K}x${ABC[C:C+16]}))); 62 vacc${ABC[C+8:C+16]} = _mm256_add_epi32(vacc${ABC[C+8:C+16]}, _mm256_cvtepi16_epi32(vprod${K}x${ABC[C+8:C+16]})); 63 64 w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(int8_t)); 65 66 const __m256i vmultiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.multiplier)); 67 const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding)); 68 69 $for C in range(0, CHANNEL_TILE, 8): 70 const __m256i vacc${ABC[C+1:C+8:2]} = _mm256_shuffle_epi32(vacc${ABC[C:C+8]}, _MM_SHUFFLE(3, 3, 1, 1)); 71 72 $for C in range(0, CHANNEL_TILE, 8): 73 const __m256i vprod${ABC[C:C+8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[C:C+8]}, vmultiplier), vrounding); 74 const __m256i vprod${ABC[C+1:C+8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[C+1:C+8:2]}, vmultiplier), vrounding); 75 76 $for C in range(0, CHANNEL_TILE, 8): 77 const __m256i vq31prod${ABC[C:C+8:2]} = _mm256_srli_epi64(vprod${ABC[C:C+8:2]}, 31); 78 const __m256i vq31prod${ABC[C+1:C+8:2]} = _mm256_add_epi64(vprod${ABC[C+1:C+8:2]}, vprod${ABC[C+1:C+8:2]}); 79 80 $for C in range(0, CHANNEL_TILE, 8): 81 const __m256i vq31prod${ABC[C:C+8]} = _mm256_blend_epi16(vq31prod${ABC[C:C+8:2]}, vq31prod${ABC[C+1:C+8:2]}, 0xCC); 82 83 const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask)); 84 $for C in range(0, CHANNEL_TILE, 8): 85 const __m256i vrem${ABC[C:C+8]} = 86 _mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[C:C+8]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[C:C+8]})); 87 88 const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold)); 89 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift); 90 $for C in range(0, CHANNEL_TILE, 8): 91 vacc${ABC[C:C+8]} = 92 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[C:C+8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[C:C+8]}, vremainder_threshold)); 93 94 const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point)); 95 $for C in range(0, CHANNEL_TILE, 16): 96 __m256i vout${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); 97 98 const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min)); 99 const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max)); 100 $for C in range(0, CHANNEL_TILE, 16): 101 vout${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); 102 103 $for C in range(0, CHANNEL_TILE, 16): 104 __m128i vout${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)); 105 106 _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]}); 107 $for C in range(16, CHANNEL_TILE, 16): 108 _mm_storeu_si128((__m128i*) (output + ${C}), vout${ABC[C:C+16]}); 109 output += ${CHANNEL_TILE}; 110 } 111 if XNN_UNLIKELY(c != 0) { 112 $if CHANNEL_TILE > 16: 113 const int8_t* k = (const int8_t*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t)); 114 ${"do " if CHANNEL_TILE > 16 else ""}{ 115 __m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w); 116 __m256i vacc${ABC[8:16]} = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t))); 117 118 $for K in range(KERNEL_TILE): 119 120 const __m256i vi${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i${K})); 121 $if CHANNEL_TILE > 16: 122 $if K == 0: 123 const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k)); 124 $else: 125 const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + ${K * CHANNEL_TILE}))); 126 $else: 127 const __m256i vk${K}x${ABC[0:16]} = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(int8_t)))); 128 $if CHANNEL_TILE > 16: 129 i${K} += 16; 130 131 const __m256i vprod${K}x${ABC[0:16]} = _mm256_mullo_epi16(vi${K}x${ABC[0:16]}, vk${K}x${ABC[0:16]}); 132 const __m128i vprod${K}x${ABC[8:16]} = _mm256_extracti128_si256(vprod${K}x${ABC[0:16]}, 1); 133 vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod${K}x${ABC[0:16]}))); 134 vacc${ABC[8:16]} = _mm256_add_epi32(vacc${ABC[8:16]}, _mm256_cvtepi16_epi32(vprod${K}x${ABC[8:16]})); 135 136 $if CHANNEL_TILE > 16: 137 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t)); 138 k += 16; 139 140 const __m256i vmultiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.multiplier)); 141 const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding)); 142 143 const __m256i vacc${ABC[1:8:2]} = _mm256_shuffle_epi32(vacc${ABC[0:8]}, _MM_SHUFFLE(3, 3, 1, 1)); 144 const __m256i vacc${ABC[9:16:2]} = _mm256_shuffle_epi32(vacc${ABC[8:16]}, _MM_SHUFFLE(3, 3, 1, 1)); 145 146 const __m256i vprod${ABC[0:8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[0:8]}, vmultiplier), vrounding); 147 const __m256i vprod${ABC[1:8:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[1:8:2]}, vmultiplier), vrounding); 148 const __m256i vprod${ABC[8:16:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[8:16]}, vmultiplier), vrounding); 149 const __m256i vprod${ABC[9:16:2]} = _mm256_add_epi64(_mm256_mul_epi32(vacc${ABC[9:16:2]}, vmultiplier), vrounding); 150 151 const __m256i vq31prod${ABC[0:8:2]} = _mm256_srli_epi64(vprod${ABC[0:8:2]}, 31); 152 const __m256i vq31prod${ABC[1:8:2]} = _mm256_add_epi64(vprod${ABC[1:8:2]}, vprod${ABC[1:8:2]}); 153 const __m256i vq31prod${ABC[8:16:2]} = _mm256_srli_epi64(vprod${ABC[8:16:2]}, 31); 154 const __m256i vq31prod${ABC[9:16:2]} = _mm256_add_epi64(vprod${ABC[9:16:2]}, vprod${ABC[9:16:2]}); 155 156 const __m256i vq31prod${ABC[0:8]} = _mm256_blend_epi16(vq31prod${ABC[0:8:2]}, vq31prod${ABC[1:8:2]}, 0xCC); 157 const __m256i vq31prod${ABC[8:16]} = _mm256_blend_epi16(vq31prod${ABC[8:16:2]}, vq31prod${ABC[9:16:2]}, 0xCC); 158 159 const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask)); 160 const __m256i vrem${ABC[0:8]} = 161 _mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[0:8]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[0:8]})); 162 const __m256i vrem${ABC[8:16]} = 163 _mm256_add_epi32(_mm256_and_si256(vq31prod${ABC[8:16]}, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${ABC[8:16]})); 164 165 const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold)); 166 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift); 167 vacc${ABC[0:8]} = 168 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[0:8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[0:8]}, vremainder_threshold)); 169 vacc${ABC[8:16]} = 170 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${ABC[8:16]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[8:16]}, vremainder_threshold)); 171 172 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); 173 __m128i vout${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); 174 __m128i vout${ABC[8:16]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[8:16]}), _mm256_extracti128_si256(vacc${ABC[8:16]}, 1)), voutput_zero_point); 175 176 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); 177 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max); 178 vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, voutput_min), voutput_max); 179 vout${ABC[8:16]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[8:16]}, voutput_min), voutput_max); 180 181 __m128i vout${ABC[0:16]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[8:16]}); 182 183 $if CHANNEL_TILE > 16: 184 if XNN_LIKELY(c >= 16) { 185 _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]}); 186 output += 16; 187 c -= 16; 188 } else { 189 if (c & 8) { 190 _mm_storel_epi64((__m128i*) output, vout${ABC[0:16]}); 191 vout${ABC[0:16]} = _mm_unpackhi_epi64(vout${ABC[0:16]}, vout${ABC[0:16]}); 192 output += 8; 193 } 194 if (c & 4) { 195 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:16]}); 196 vout${ABC[0:16]} = _mm_srli_epi64(vout${ABC[0:16]}, 32); 197 output += 4; 198 } 199 if (c & 2) { 200 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:16]}, 0); 201 vout${ABC[0:16]} = _mm_srli_epi32(vout${ABC[0:16]}, 16); 202 output += 2; 203 } 204 if (c & 1) { 205 *output = (int8_t) _mm_extract_epi8(vout${ABC[0:16]}, 0); 206 output += 1; 207 } 208 c = 0; 209 } 210 $else: 211 if (c & 8) { 212 _mm_storel_epi64((__m128i*) output, vout${ABC[0:16]}); 213 vout${ABC[0:16]} = _mm_unpackhi_epi64(vout${ABC[0:16]}, vout${ABC[0:16]}); 214 output += 8; 215 } 216 if (c & 4) { 217 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:16]}); 218 vout${ABC[0:16]} = _mm_srli_epi64(vout${ABC[0:16]}, 32); 219 output += 4; 220 } 221 if (c & 2) { 222 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:16]}, 0); 223 vout${ABC[0:16]} = _mm_srli_epi32(vout${ABC[0:16]}, 16); 224 output += 2; 225 } 226 if (c & 1) { 227 *output = (int8_t) _mm_extract_epi8(vout${ABC[0:16]}, 0); 228 output += 1; 229 } 230 }${" while (c != 0);" if CHANNEL_TILE > 16 else ""} 231 } 232 233 output = (int8_t*) ((uintptr_t) output + output_increment); 234 } while (--output_width != 0); 235} 236