| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-dwconv/unipass-avx2-mul16-vpmovsx.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2020 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
#include <assert.h> |
| 11 |
|
| 12 |
#include <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/dwconv.h" |
| 15 |
#include "xnnpack/unaligned.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qs8_dwconv_minmax_fp32_ukernel_25p16c__avx2_mul16_vpmovsx( |
| 19 |
size_t channels, |
| 20 |
size_t output_width, |
| 21 |
const int8_t** input, |
| 22 |
const void* weights, |
| 23 |
int8_t* output, |
| 24 |
intptr_t input_stride, |
| 25 |
size_t output_increment, |
| 26 |
size_t input_offset, |
| 27 |
const int8_t* zero, |
| 28 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 29 |
{ |
| 30 |
assert(channels != 0); |
| 31 |
assert(output_width != 0); |
| 32 |
|
| 33 |
const __m256 vscale = _mm256_set1_ps(params->fp32_scalar.scale); |
| 34 |
XNN_FORCE_REALIZATION(vscale); |
| 35 |
|
| 36 |
const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point); |
| 37 |
const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point); |
| 38 |
const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min); |
| 39 |
XNN_FORCE_REALIZATION(voutput_max_less_zero_point); |
| 40 |
XNN_FORCE_REALIZATION(voutput_zero_point); |
| 41 |
XNN_FORCE_REALIZATION(voutput_min); |
| 42 |
|
| 43 |
do { |
| 44 |
const int8_t* i0 = input[0]; |
| 45 |
assert(i0 != NULL); |
| 46 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 47 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 48 |
} |
| 49 |
const int8_t* i1 = input[1]; |
| 50 |
assert(i1 != NULL); |
| 51 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 52 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 53 |
} |
| 54 |
const int8_t* i2 = input[2]; |
| 55 |
assert(i2 != NULL); |
| 56 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 57 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 58 |
} |
| 59 |
const int8_t* i3 = input[3]; |
| 60 |
assert(i3 != NULL); |
| 61 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 62 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 63 |
} |
| 64 |
const int8_t* i4 = input[4]; |
| 65 |
assert(i4 != NULL); |
| 66 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 67 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 68 |
} |
| 69 |
const int8_t* i5 = input[5]; |
| 70 |
assert(i5 != NULL); |
| 71 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 72 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 73 |
} |
| 74 |
const int8_t* i6 = input[6]; |
| 75 |
assert(i6 != NULL); |
| 76 |
if XNN_UNPREDICTABLE(i6 != zero) { |
| 77 |
i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| 78 |
} |
| 79 |
const int8_t* i7 = input[7]; |
| 80 |
assert(i7 != NULL); |
| 81 |
if XNN_UNPREDICTABLE(i7 != zero) { |
| 82 |
i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); |
| 83 |
} |
| 84 |
const int8_t* i8 = input[8]; |
| 85 |
assert(i8 != NULL); |
| 86 |
if XNN_UNPREDICTABLE(i8 != zero) { |
| 87 |
i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); |
| 88 |
} |
| 89 |
const int8_t* i9 = input[9]; |
| 90 |
assert(i9 != NULL); |
| 91 |
if XNN_UNPREDICTABLE(i9 != zero) { |
| 92 |
i9 = (const int8_t*) ((uintptr_t) i9 + input_offset); |
| 93 |
} |
| 94 |
const int8_t* i10 = input[10]; |
| 95 |
assert(i10 != NULL); |
| 96 |
if XNN_UNPREDICTABLE(i10 != zero) { |
| 97 |
i10 = (const int8_t*) ((uintptr_t) i10 + input_offset); |
| 98 |
} |
| 99 |
const int8_t* i11 = input[11]; |
| 100 |
assert(i11 != NULL); |
| 101 |
if XNN_UNPREDICTABLE(i11 != zero) { |
| 102 |
i11 = (const int8_t*) ((uintptr_t) i11 + input_offset); |
| 103 |
} |
| 104 |
const int8_t* i12 = input[12]; |
| 105 |
assert(i12 != NULL); |
| 106 |
if XNN_UNPREDICTABLE(i12 != zero) { |
| 107 |
i12 = (const int8_t*) ((uintptr_t) i12 + input_offset); |
| 108 |
} |
| 109 |
const int8_t* i13 = input[13]; |
| 110 |
assert(i13 != NULL); |
| 111 |
if XNN_UNPREDICTABLE(i13 != zero) { |
| 112 |
i13 = (const int8_t*) ((uintptr_t) i13 + input_offset); |
| 113 |
} |
| 114 |
const int8_t* i14 = input[14]; |
| 115 |
assert(i14 != NULL); |
| 116 |
if XNN_UNPREDICTABLE(i14 != zero) { |
| 117 |
i14 = (const int8_t*) ((uintptr_t) i14 + input_offset); |
| 118 |
} |
| 119 |
const int8_t* i15 = input[15]; |
| 120 |
assert(i15 != NULL); |
| 121 |
if XNN_UNPREDICTABLE(i15 != zero) { |
| 122 |
i15 = (const int8_t*) ((uintptr_t) i15 + input_offset); |
| 123 |
} |
| 124 |
const int8_t* i16 = input[16]; |
| 125 |
assert(i16 != NULL); |
| 126 |
if XNN_UNPREDICTABLE(i16 != zero) { |
| 127 |
i16 = (const int8_t*) ((uintptr_t) i16 + input_offset); |
| 128 |
} |
| 129 |
const int8_t* i17 = input[17]; |
| 130 |
assert(i17 != NULL); |
| 131 |
if XNN_UNPREDICTABLE(i17 != zero) { |
| 132 |
i17 = (const int8_t*) ((uintptr_t) i17 + input_offset); |
| 133 |
} |
| 134 |
const int8_t* i18 = input[18]; |
| 135 |
assert(i18 != NULL); |
| 136 |
if XNN_UNPREDICTABLE(i18 != zero) { |
| 137 |
i18 = (const int8_t*) ((uintptr_t) i18 + input_offset); |
| 138 |
} |
| 139 |
const int8_t* i19 = input[19]; |
| 140 |
assert(i19 != NULL); |
| 141 |
if XNN_UNPREDICTABLE(i19 != zero) { |
| 142 |
i19 = (const int8_t*) ((uintptr_t) i19 + input_offset); |
| 143 |
} |
| 144 |
const int8_t* i20 = input[20]; |
| 145 |
assert(i20 != NULL); |
| 146 |
if XNN_UNPREDICTABLE(i20 != zero) { |
| 147 |
i20 = (const int8_t*) ((uintptr_t) i20 + input_offset); |
| 148 |
} |
| 149 |
const int8_t* i21 = input[21]; |
| 150 |
assert(i21 != NULL); |
| 151 |
if XNN_UNPREDICTABLE(i21 != zero) { |
| 152 |
i21 = (const int8_t*) ((uintptr_t) i21 + input_offset); |
| 153 |
} |
| 154 |
const int8_t* i22 = input[22]; |
| 155 |
assert(i22 != NULL); |
| 156 |
if XNN_UNPREDICTABLE(i22 != zero) { |
| 157 |
i22 = (const int8_t*) ((uintptr_t) i22 + input_offset); |
| 158 |
} |
| 159 |
const int8_t* i23 = input[23]; |
| 160 |
assert(i23 != NULL); |
| 161 |
if XNN_UNPREDICTABLE(i23 != zero) { |
| 162 |
i23 = (const int8_t*) ((uintptr_t) i23 + input_offset); |
| 163 |
} |
| 164 |
const int8_t* i24 = input[24]; |
| 165 |
assert(i24 != NULL); |
| 166 |
if XNN_UNPREDICTABLE(i24 != zero) { |
| 167 |
i24 = (const int8_t*) ((uintptr_t) i24 + input_offset); |
| 168 |
} |
| 169 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 170 |
|
| 171 |
size_t c = channels; |
| 172 |
const void* w = weights; |
| 173 |
for (; c >= 16; c -= 16) { |
| 174 |
__m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w); |
| 175 |
__m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t))); |
| 176 |
|
| 177 |
|
| 178 |
const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0)); |
| 179 |
const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)))); |
| 180 |
i0 += 16; |
| 181 |
|
| 182 |
const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF); |
| 183 |
const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1); |
| 184 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF))); |
| 185 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); |
| 186 |
|
| 187 |
const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1)); |
| 188 |
const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)))); |
| 189 |
i1 += 16; |
| 190 |
|
| 191 |
const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF); |
| 192 |
const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1); |
| 193 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF))); |
| 194 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); |
| 195 |
|
| 196 |
const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2)); |
| 197 |
const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)))); |
| 198 |
i2 += 16; |
| 199 |
|
| 200 |
const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF); |
| 201 |
const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1); |
| 202 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF))); |
| 203 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); |
| 204 |
|
| 205 |
const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3)); |
| 206 |
const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)))); |
| 207 |
i3 += 16; |
| 208 |
|
| 209 |
const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF); |
| 210 |
const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1); |
| 211 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF))); |
| 212 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF)); |
| 213 |
|
| 214 |
const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4)); |
| 215 |
const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)))); |
| 216 |
i4 += 16; |
| 217 |
|
| 218 |
const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF); |
| 219 |
const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1); |
| 220 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF))); |
| 221 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF)); |
| 222 |
|
| 223 |
const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5)); |
| 224 |
const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)))); |
| 225 |
i5 += 16; |
| 226 |
|
| 227 |
const __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF); |
| 228 |
const __m128i vprod5x89ABCDEF = _mm256_extracti128_si256(vprod5x0123456789ABCDEF, 1); |
| 229 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x0123456789ABCDEF))); |
| 230 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod5x89ABCDEF)); |
| 231 |
|
| 232 |
const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6)); |
| 233 |
const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)))); |
| 234 |
i6 += 16; |
| 235 |
|
| 236 |
const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF); |
| 237 |
const __m128i vprod6x89ABCDEF = _mm256_extracti128_si256(vprod6x0123456789ABCDEF, 1); |
| 238 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod6x0123456789ABCDEF))); |
| 239 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod6x89ABCDEF)); |
| 240 |
|
| 241 |
const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7)); |
| 242 |
const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)))); |
| 243 |
i7 += 16; |
| 244 |
|
| 245 |
const __m256i vprod7x0123456789ABCDEF = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF); |
| 246 |
const __m128i vprod7x89ABCDEF = _mm256_extracti128_si256(vprod7x0123456789ABCDEF, 1); |
| 247 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod7x0123456789ABCDEF))); |
| 248 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod7x89ABCDEF)); |
| 249 |
|
| 250 |
const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); |
| 251 |
const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)))); |
| 252 |
i8 += 16; |
| 253 |
|
| 254 |
const __m256i vprod8x0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF); |
| 255 |
const __m128i vprod8x89ABCDEF = _mm256_extracti128_si256(vprod8x0123456789ABCDEF, 1); |
| 256 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod8x0123456789ABCDEF))); |
| 257 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod8x89ABCDEF)); |
| 258 |
|
| 259 |
const __m256i vi9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i9)); |
| 260 |
const __m256i vk9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)))); |
| 261 |
i9 += 16; |
| 262 |
|
| 263 |
const __m256i vprod9x0123456789ABCDEF = _mm256_mullo_epi16(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF); |
| 264 |
const __m128i vprod9x89ABCDEF = _mm256_extracti128_si256(vprod9x0123456789ABCDEF, 1); |
| 265 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod9x0123456789ABCDEF))); |
| 266 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod9x89ABCDEF)); |
| 267 |
|
| 268 |
const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); |
| 269 |
const __m256i vk10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)))); |
| 270 |
i10 += 16; |
| 271 |
|
| 272 |
const __m256i vprod10x0123456789ABCDEF = _mm256_mullo_epi16(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF); |
| 273 |
const __m128i vprod10x89ABCDEF = _mm256_extracti128_si256(vprod10x0123456789ABCDEF, 1); |
| 274 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod10x0123456789ABCDEF))); |
| 275 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod10x89ABCDEF)); |
| 276 |
|
| 277 |
const __m256i vi11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i11)); |
| 278 |
const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)))); |
| 279 |
i11 += 16; |
| 280 |
|
| 281 |
const __m256i vprod11x0123456789ABCDEF = _mm256_mullo_epi16(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF); |
| 282 |
const __m128i vprod11x89ABCDEF = _mm256_extracti128_si256(vprod11x0123456789ABCDEF, 1); |
| 283 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod11x0123456789ABCDEF))); |
| 284 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod11x89ABCDEF)); |
| 285 |
|
| 286 |
const __m256i vi12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i12)); |
| 287 |
const __m256i vk12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)))); |
| 288 |
i12 += 16; |
| 289 |
|
| 290 |
const __m256i vprod12x0123456789ABCDEF = _mm256_mullo_epi16(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF); |
| 291 |
const __m128i vprod12x89ABCDEF = _mm256_extracti128_si256(vprod12x0123456789ABCDEF, 1); |
| 292 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod12x0123456789ABCDEF))); |
| 293 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod12x89ABCDEF)); |
| 294 |
|
| 295 |
const __m256i vi13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i13)); |
| 296 |
const __m256i vk13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)))); |
| 297 |
i13 += 16; |
| 298 |
|
| 299 |
const __m256i vprod13x0123456789ABCDEF = _mm256_mullo_epi16(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF); |
| 300 |
const __m128i vprod13x89ABCDEF = _mm256_extracti128_si256(vprod13x0123456789ABCDEF, 1); |
| 301 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod13x0123456789ABCDEF))); |
| 302 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod13x89ABCDEF)); |
| 303 |
|
| 304 |
const __m256i vi14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i14)); |
| 305 |
const __m256i vk14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)))); |
| 306 |
i14 += 16; |
| 307 |
|
| 308 |
const __m256i vprod14x0123456789ABCDEF = _mm256_mullo_epi16(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF); |
| 309 |
const __m128i vprod14x89ABCDEF = _mm256_extracti128_si256(vprod14x0123456789ABCDEF, 1); |
| 310 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod14x0123456789ABCDEF))); |
| 311 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod14x89ABCDEF)); |
| 312 |
|
| 313 |
const __m256i vi15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i15)); |
| 314 |
const __m256i vk15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)))); |
| 315 |
i15 += 16; |
| 316 |
|
| 317 |
const __m256i vprod15x0123456789ABCDEF = _mm256_mullo_epi16(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF); |
| 318 |
const __m128i vprod15x89ABCDEF = _mm256_extracti128_si256(vprod15x0123456789ABCDEF, 1); |
| 319 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod15x0123456789ABCDEF))); |
| 320 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod15x89ABCDEF)); |
| 321 |
|
| 322 |
const __m256i vi16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i16)); |
| 323 |
const __m256i vk16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)))); |
| 324 |
i16 += 16; |
| 325 |
|
| 326 |
const __m256i vprod16x0123456789ABCDEF = _mm256_mullo_epi16(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF); |
| 327 |
const __m128i vprod16x89ABCDEF = _mm256_extracti128_si256(vprod16x0123456789ABCDEF, 1); |
| 328 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod16x0123456789ABCDEF))); |
| 329 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod16x89ABCDEF)); |
| 330 |
|
| 331 |
const __m256i vi17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i17)); |
| 332 |
const __m256i vk17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)))); |
| 333 |
i17 += 16; |
| 334 |
|
| 335 |
const __m256i vprod17x0123456789ABCDEF = _mm256_mullo_epi16(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF); |
| 336 |
const __m128i vprod17x89ABCDEF = _mm256_extracti128_si256(vprod17x0123456789ABCDEF, 1); |
| 337 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod17x0123456789ABCDEF))); |
| 338 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod17x89ABCDEF)); |
| 339 |
|
| 340 |
const __m256i vi18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i18)); |
| 341 |
const __m256i vk18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)))); |
| 342 |
i18 += 16; |
| 343 |
|
| 344 |
const __m256i vprod18x0123456789ABCDEF = _mm256_mullo_epi16(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF); |
| 345 |
const __m128i vprod18x89ABCDEF = _mm256_extracti128_si256(vprod18x0123456789ABCDEF, 1); |
| 346 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod18x0123456789ABCDEF))); |
| 347 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod18x89ABCDEF)); |
| 348 |
|
| 349 |
const __m256i vi19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i19)); |
| 350 |
const __m256i vk19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)))); |
| 351 |
i19 += 16; |
| 352 |
|
| 353 |
const __m256i vprod19x0123456789ABCDEF = _mm256_mullo_epi16(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF); |
| 354 |
const __m128i vprod19x89ABCDEF = _mm256_extracti128_si256(vprod19x0123456789ABCDEF, 1); |
| 355 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod19x0123456789ABCDEF))); |
| 356 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod19x89ABCDEF)); |
| 357 |
|
| 358 |
const __m256i vi20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i20)); |
| 359 |
const __m256i vk20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)))); |
| 360 |
i20 += 16; |
| 361 |
|
| 362 |
const __m256i vprod20x0123456789ABCDEF = _mm256_mullo_epi16(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF); |
| 363 |
const __m128i vprod20x89ABCDEF = _mm256_extracti128_si256(vprod20x0123456789ABCDEF, 1); |
| 364 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod20x0123456789ABCDEF))); |
| 365 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod20x89ABCDEF)); |
| 366 |
|
| 367 |
const __m256i vi21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i21)); |
| 368 |
const __m256i vk21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)))); |
| 369 |
i21 += 16; |
| 370 |
|
| 371 |
const __m256i vprod21x0123456789ABCDEF = _mm256_mullo_epi16(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF); |
| 372 |
const __m128i vprod21x89ABCDEF = _mm256_extracti128_si256(vprod21x0123456789ABCDEF, 1); |
| 373 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod21x0123456789ABCDEF))); |
| 374 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod21x89ABCDEF)); |
| 375 |
|
| 376 |
const __m256i vi22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i22)); |
| 377 |
const __m256i vk22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)))); |
| 378 |
i22 += 16; |
| 379 |
|
| 380 |
const __m256i vprod22x0123456789ABCDEF = _mm256_mullo_epi16(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF); |
| 381 |
const __m128i vprod22x89ABCDEF = _mm256_extracti128_si256(vprod22x0123456789ABCDEF, 1); |
| 382 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod22x0123456789ABCDEF))); |
| 383 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod22x89ABCDEF)); |
| 384 |
|
| 385 |
const __m256i vi23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i23)); |
| 386 |
const __m256i vk23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)))); |
| 387 |
i23 += 16; |
| 388 |
|
| 389 |
const __m256i vprod23x0123456789ABCDEF = _mm256_mullo_epi16(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF); |
| 390 |
const __m128i vprod23x89ABCDEF = _mm256_extracti128_si256(vprod23x0123456789ABCDEF, 1); |
| 391 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod23x0123456789ABCDEF))); |
| 392 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod23x89ABCDEF)); |
| 393 |
|
| 394 |
const __m256i vi24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i24)); |
| 395 |
const __m256i vk24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)))); |
| 396 |
i24 += 16; |
| 397 |
|
| 398 |
const __m256i vprod24x0123456789ABCDEF = _mm256_mullo_epi16(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF); |
| 399 |
const __m128i vprod24x89ABCDEF = _mm256_extracti128_si256(vprod24x0123456789ABCDEF, 1); |
| 400 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod24x0123456789ABCDEF))); |
| 401 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod24x89ABCDEF)); |
| 402 |
|
| 403 |
w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t)); |
| 404 |
|
| 405 |
__m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567); |
| 406 |
__m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF); |
| 407 |
|
| 408 |
vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale); |
| 409 |
vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale); |
| 410 |
|
| 411 |
vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point); |
| 412 |
vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point); |
| 413 |
|
| 414 |
vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567); |
| 415 |
vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF); |
| 416 |
|
| 417 |
const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point); |
| 418 |
|
| 419 |
__m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0)); |
| 420 |
|
| 421 |
vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min); |
| 422 |
|
| 423 |
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); |
| 424 |
output += 16; |
| 425 |
} |
| 426 |
if XNN_UNLIKELY(c != 0) { |
| 427 |
{ |
| 428 |
__m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w); |
| 429 |
__m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t))); |
| 430 |
|
| 431 |
|
| 432 |
const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0)); |
| 433 |
const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)))); |
| 434 |
|
| 435 |
const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF); |
| 436 |
const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1); |
| 437 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF))); |
| 438 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); |
| 439 |
|
| 440 |
const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1)); |
| 441 |
const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)))); |
| 442 |
|
| 443 |
const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF); |
| 444 |
const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1); |
| 445 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF))); |
| 446 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); |
| 447 |
|
| 448 |
const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2)); |
| 449 |
const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)))); |
| 450 |
|
| 451 |
const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF); |
| 452 |
const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1); |
| 453 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF))); |
| 454 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); |
| 455 |
|
| 456 |
const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3)); |
| 457 |
const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)))); |
| 458 |
|
| 459 |
const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF); |
| 460 |
const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1); |
| 461 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF))); |
| 462 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF)); |
| 463 |
|
| 464 |
const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4)); |
| 465 |
const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)))); |
| 466 |
|
| 467 |
const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF); |
| 468 |
const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1); |
| 469 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF))); |
| 470 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF)); |
| 471 |
|
| 472 |
const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5)); |
| 473 |
const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)))); |
| 474 |
|
| 475 |
const __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF); |
| 476 |
const __m128i vprod5x89ABCDEF = _mm256_extracti128_si256(vprod5x0123456789ABCDEF, 1); |
| 477 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x0123456789ABCDEF))); |
| 478 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod5x89ABCDEF)); |
| 479 |
|
| 480 |
const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6)); |
| 481 |
const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)))); |
| 482 |
|
| 483 |
const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF); |
| 484 |
const __m128i vprod6x89ABCDEF = _mm256_extracti128_si256(vprod6x0123456789ABCDEF, 1); |
| 485 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod6x0123456789ABCDEF))); |
| 486 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod6x89ABCDEF)); |
| 487 |
|
| 488 |
const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7)); |
| 489 |
const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)))); |
| 490 |
|
| 491 |
const __m256i vprod7x0123456789ABCDEF = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF); |
| 492 |
const __m128i vprod7x89ABCDEF = _mm256_extracti128_si256(vprod7x0123456789ABCDEF, 1); |
| 493 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod7x0123456789ABCDEF))); |
| 494 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod7x89ABCDEF)); |
| 495 |
|
| 496 |
const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); |
| 497 |
const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)))); |
| 498 |
|
| 499 |
const __m256i vprod8x0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF); |
| 500 |
const __m128i vprod8x89ABCDEF = _mm256_extracti128_si256(vprod8x0123456789ABCDEF, 1); |
| 501 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod8x0123456789ABCDEF))); |
| 502 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod8x89ABCDEF)); |
| 503 |
|
| 504 |
const __m256i vi9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i9)); |
| 505 |
const __m256i vk9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)))); |
| 506 |
|
| 507 |
const __m256i vprod9x0123456789ABCDEF = _mm256_mullo_epi16(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF); |
| 508 |
const __m128i vprod9x89ABCDEF = _mm256_extracti128_si256(vprod9x0123456789ABCDEF, 1); |
| 509 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod9x0123456789ABCDEF))); |
| 510 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod9x89ABCDEF)); |
| 511 |
|
| 512 |
const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); |
| 513 |
const __m256i vk10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)))); |
| 514 |
|
| 515 |
const __m256i vprod10x0123456789ABCDEF = _mm256_mullo_epi16(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF); |
| 516 |
const __m128i vprod10x89ABCDEF = _mm256_extracti128_si256(vprod10x0123456789ABCDEF, 1); |
| 517 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod10x0123456789ABCDEF))); |
| 518 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod10x89ABCDEF)); |
| 519 |
|
| 520 |
const __m256i vi11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i11)); |
| 521 |
const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)))); |
| 522 |
|
| 523 |
const __m256i vprod11x0123456789ABCDEF = _mm256_mullo_epi16(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF); |
| 524 |
const __m128i vprod11x89ABCDEF = _mm256_extracti128_si256(vprod11x0123456789ABCDEF, 1); |
| 525 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod11x0123456789ABCDEF))); |
| 526 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod11x89ABCDEF)); |
| 527 |
|
| 528 |
const __m256i vi12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i12)); |
| 529 |
const __m256i vk12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)))); |
| 530 |
|
| 531 |
const __m256i vprod12x0123456789ABCDEF = _mm256_mullo_epi16(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF); |
| 532 |
const __m128i vprod12x89ABCDEF = _mm256_extracti128_si256(vprod12x0123456789ABCDEF, 1); |
| 533 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod12x0123456789ABCDEF))); |
| 534 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod12x89ABCDEF)); |
| 535 |
|
| 536 |
const __m256i vi13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i13)); |
| 537 |
const __m256i vk13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)))); |
| 538 |
|
| 539 |
const __m256i vprod13x0123456789ABCDEF = _mm256_mullo_epi16(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF); |
| 540 |
const __m128i vprod13x89ABCDEF = _mm256_extracti128_si256(vprod13x0123456789ABCDEF, 1); |
| 541 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod13x0123456789ABCDEF))); |
| 542 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod13x89ABCDEF)); |
| 543 |
|
| 544 |
const __m256i vi14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i14)); |
| 545 |
const __m256i vk14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)))); |
| 546 |
|
| 547 |
const __m256i vprod14x0123456789ABCDEF = _mm256_mullo_epi16(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF); |
| 548 |
const __m128i vprod14x89ABCDEF = _mm256_extracti128_si256(vprod14x0123456789ABCDEF, 1); |
| 549 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod14x0123456789ABCDEF))); |
| 550 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod14x89ABCDEF)); |
| 551 |
|
| 552 |
const __m256i vi15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i15)); |
| 553 |
const __m256i vk15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)))); |
| 554 |
|
| 555 |
const __m256i vprod15x0123456789ABCDEF = _mm256_mullo_epi16(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF); |
| 556 |
const __m128i vprod15x89ABCDEF = _mm256_extracti128_si256(vprod15x0123456789ABCDEF, 1); |
| 557 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod15x0123456789ABCDEF))); |
| 558 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod15x89ABCDEF)); |
| 559 |
|
| 560 |
const __m256i vi16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i16)); |
| 561 |
const __m256i vk16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)))); |
| 562 |
|
| 563 |
const __m256i vprod16x0123456789ABCDEF = _mm256_mullo_epi16(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF); |
| 564 |
const __m128i vprod16x89ABCDEF = _mm256_extracti128_si256(vprod16x0123456789ABCDEF, 1); |
| 565 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod16x0123456789ABCDEF))); |
| 566 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod16x89ABCDEF)); |
| 567 |
|
| 568 |
const __m256i vi17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i17)); |
| 569 |
const __m256i vk17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)))); |
| 570 |
|
| 571 |
const __m256i vprod17x0123456789ABCDEF = _mm256_mullo_epi16(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF); |
| 572 |
const __m128i vprod17x89ABCDEF = _mm256_extracti128_si256(vprod17x0123456789ABCDEF, 1); |
| 573 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod17x0123456789ABCDEF))); |
| 574 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod17x89ABCDEF)); |
| 575 |
|
| 576 |
const __m256i vi18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i18)); |
| 577 |
const __m256i vk18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)))); |
| 578 |
|
| 579 |
const __m256i vprod18x0123456789ABCDEF = _mm256_mullo_epi16(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF); |
| 580 |
const __m128i vprod18x89ABCDEF = _mm256_extracti128_si256(vprod18x0123456789ABCDEF, 1); |
| 581 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod18x0123456789ABCDEF))); |
| 582 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod18x89ABCDEF)); |
| 583 |
|
| 584 |
const __m256i vi19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i19)); |
| 585 |
const __m256i vk19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)))); |
| 586 |
|
| 587 |
const __m256i vprod19x0123456789ABCDEF = _mm256_mullo_epi16(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF); |
| 588 |
const __m128i vprod19x89ABCDEF = _mm256_extracti128_si256(vprod19x0123456789ABCDEF, 1); |
| 589 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod19x0123456789ABCDEF))); |
| 590 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod19x89ABCDEF)); |
| 591 |
|
| 592 |
const __m256i vi20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i20)); |
| 593 |
const __m256i vk20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)))); |
| 594 |
|
| 595 |
const __m256i vprod20x0123456789ABCDEF = _mm256_mullo_epi16(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF); |
| 596 |
const __m128i vprod20x89ABCDEF = _mm256_extracti128_si256(vprod20x0123456789ABCDEF, 1); |
| 597 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod20x0123456789ABCDEF))); |
| 598 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod20x89ABCDEF)); |
| 599 |
|
| 600 |
const __m256i vi21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i21)); |
| 601 |
const __m256i vk21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)))); |
| 602 |
|
| 603 |
const __m256i vprod21x0123456789ABCDEF = _mm256_mullo_epi16(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF); |
| 604 |
const __m128i vprod21x89ABCDEF = _mm256_extracti128_si256(vprod21x0123456789ABCDEF, 1); |
| 605 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod21x0123456789ABCDEF))); |
| 606 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod21x89ABCDEF)); |
| 607 |
|
| 608 |
const __m256i vi22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i22)); |
| 609 |
const __m256i vk22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)))); |
| 610 |
|
| 611 |
const __m256i vprod22x0123456789ABCDEF = _mm256_mullo_epi16(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF); |
| 612 |
const __m128i vprod22x89ABCDEF = _mm256_extracti128_si256(vprod22x0123456789ABCDEF, 1); |
| 613 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod22x0123456789ABCDEF))); |
| 614 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod22x89ABCDEF)); |
| 615 |
|
| 616 |
const __m256i vi23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i23)); |
| 617 |
const __m256i vk23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)))); |
| 618 |
|
| 619 |
const __m256i vprod23x0123456789ABCDEF = _mm256_mullo_epi16(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF); |
| 620 |
const __m128i vprod23x89ABCDEF = _mm256_extracti128_si256(vprod23x0123456789ABCDEF, 1); |
| 621 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod23x0123456789ABCDEF))); |
| 622 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod23x89ABCDEF)); |
| 623 |
|
| 624 |
const __m256i vi24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i24)); |
| 625 |
const __m256i vk24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)))); |
| 626 |
|
| 627 |
const __m256i vprod24x0123456789ABCDEF = _mm256_mullo_epi16(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF); |
| 628 |
const __m128i vprod24x89ABCDEF = _mm256_extracti128_si256(vprod24x0123456789ABCDEF, 1); |
| 629 |
vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod24x0123456789ABCDEF))); |
| 630 |
vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod24x89ABCDEF)); |
| 631 |
|
| 632 |
|
| 633 |
__m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567); |
| 634 |
__m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF); |
| 635 |
|
| 636 |
vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale); |
| 637 |
vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale); |
| 638 |
|
| 639 |
vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point); |
| 640 |
vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point); |
| 641 |
|
| 642 |
vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567); |
| 643 |
vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF); |
| 644 |
|
| 645 |
|
| 646 |
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point)); |
| 647 |
__m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), _mm256_castsi256_si128(voutput_zero_point)); |
| 648 |
|
| 649 |
__m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF); |
| 650 |
vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min); |
| 651 |
|
| 652 |
if (c & 8) { |
| 653 |
_mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF); |
| 654 |
vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF); |
| 655 |
output += 8; |
| 656 |
} |
| 657 |
if (c & 4) { |
| 658 |
unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF)); |
| 659 |
vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32); |
| 660 |
output += 4; |
| 661 |
} |
| 662 |
if (c & 2) { |
| 663 |
unaligned_store_u16(output, _mm_extract_epi16(vout0123456789ABCDEF, 0)); |
| 664 |
vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16); |
| 665 |
output += 2; |
| 666 |
} |
| 667 |
if (c & 1) { |
| 668 |
*output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0); |
| 669 |
output += 1; |
| 670 |
} |
| 671 |
} |
| 672 |
} |
| 673 |
|
| 674 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 675 |
} while (--output_width != 0); |
| 676 |
} |