Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-dwconv/up-avx.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2019 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 | |
| 16 | |
| 17 | static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0}; |
| 18 | |
| 19 | void xnn_f32_dwconv_ukernel_up8x25__fma3_acc2( |
| 20 | size_t channels, |
| 21 | size_t output_width, |
| 22 | const float** input, |
| 23 | const float* weights, |
| 24 | float* output, |
| 25 | size_t input_stride, |
| 26 | size_t output_increment, |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 27 | const union xnn_f32_minmax_params params[restrict static 1]) |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 28 | { |
| 29 | assert(channels != 0); |
| 30 | assert(output_width != 0); |
| 31 | |
| 32 | const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max); |
| 33 | const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min); |
| 34 | do { |
| 35 | const float* i0 = input[0]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 36 | assert(i0 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 37 | const float* i1 = input[1]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 38 | assert(i1 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 39 | const float* i2 = input[2]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 40 | assert(i2 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 41 | const float* i3 = input[3]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 42 | assert(i3 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 43 | const float* i4 = input[4]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 44 | assert(i4 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 45 | const float* i5 = input[5]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 46 | assert(i5 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 47 | const float* i6 = input[6]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 48 | assert(i6 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 49 | const float* i7 = input[7]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 50 | assert(i7 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 51 | const float* i8 = input[8]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 52 | assert(i8 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 53 | const float* i9 = input[9]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 54 | assert(i9 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 55 | const float* i10 = input[10]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 56 | assert(i10 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 57 | const float* i11 = input[11]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 58 | assert(i11 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 59 | const float* i12 = input[12]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 60 | assert(i12 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 61 | const float* i13 = input[13]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 62 | assert(i13 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 63 | const float* i14 = input[14]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 64 | assert(i14 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 65 | const float* i15 = input[15]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 66 | assert(i15 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 67 | const float* i16 = input[16]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 68 | assert(i16 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 69 | const float* i17 = input[17]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 70 | assert(i17 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 71 | const float* i18 = input[18]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 72 | assert(i18 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 73 | const float* i19 = input[19]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 74 | assert(i19 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 75 | const float* i20 = input[20]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 76 | assert(i20 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 77 | const float* i21 = input[21]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 78 | assert(i21 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 79 | const float* i22 = input[22]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 80 | assert(i22 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 81 | const float* i23 = input[23]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 82 | assert(i23 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 83 | const float* i24 = input[24]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 84 | assert(i24 != NULL); |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 85 | input = (const float**) ((uintptr_t) input + input_stride); |
| 86 | |
| 87 | size_t c = channels; |
| 88 | const float* w = weights; |
| 89 | for (; c >= 8; c -= 8) { |
| 90 | __m256 vacc01234567p0 = _mm256_load_ps(w); |
| 91 | |
| 92 | |
| 93 | const __m256 vi0x01234567 = _mm256_loadu_ps(i0); |
| 94 | i0 += 8; |
| 95 | |
| 96 | const __m256 vk0x01234567 = _mm256_load_ps(w + 8); |
| 97 | vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); |
| 98 | |
| 99 | const __m256 vi1x01234567 = _mm256_loadu_ps(i1); |
| 100 | i1 += 8; |
| 101 | |
| 102 | const __m256 vk1x01234567 = _mm256_load_ps(w + 16); |
| 103 | __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567); |
| 104 | |
| 105 | const __m256 vi2x01234567 = _mm256_loadu_ps(i2); |
| 106 | i2 += 8; |
| 107 | |
| 108 | const __m256 vk2x01234567 = _mm256_load_ps(w + 24); |
| 109 | vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0); |
| 110 | |
| 111 | const __m256 vi3x01234567 = _mm256_loadu_ps(i3); |
| 112 | i3 += 8; |
| 113 | |
| 114 | const __m256 vk3x01234567 = _mm256_load_ps(w + 32); |
| 115 | vacc01234567p1 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1); |
| 116 | |
| 117 | const __m256 vi4x01234567 = _mm256_loadu_ps(i4); |
| 118 | i4 += 8; |
| 119 | |
| 120 | const __m256 vk4x01234567 = _mm256_load_ps(w + 40); |
| 121 | vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0); |
| 122 | |
| 123 | const __m256 vi5x01234567 = _mm256_loadu_ps(i5); |
| 124 | i5 += 8; |
| 125 | |
| 126 | const __m256 vk5x01234567 = _mm256_load_ps(w + 48); |
| 127 | vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); |
| 128 | |
| 129 | const __m256 vi6x01234567 = _mm256_loadu_ps(i6); |
| 130 | i6 += 8; |
| 131 | |
| 132 | const __m256 vk6x01234567 = _mm256_load_ps(w + 56); |
| 133 | vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); |
| 134 | |
| 135 | const __m256 vi7x01234567 = _mm256_loadu_ps(i7); |
| 136 | i7 += 8; |
| 137 | |
| 138 | const __m256 vk7x01234567 = _mm256_load_ps(w + 64); |
| 139 | vacc01234567p1 = _mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1); |
| 140 | |
| 141 | const __m256 vi8x01234567 = _mm256_loadu_ps(i8); |
| 142 | i8 += 8; |
| 143 | |
| 144 | const __m256 vk8x01234567 = _mm256_load_ps(w + 72); |
| 145 | vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0); |
| 146 | |
| 147 | const __m256 vi9x01234567 = _mm256_loadu_ps(i9); |
| 148 | i9 += 8; |
| 149 | |
| 150 | const __m256 vk9x01234567 = _mm256_load_ps(w + 80); |
| 151 | vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1); |
| 152 | |
| 153 | const __m256 vi10x01234567 = _mm256_loadu_ps(i10); |
| 154 | i10 += 8; |
| 155 | |
| 156 | const __m256 vk10x01234567 = _mm256_load_ps(w + 88); |
| 157 | vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); |
| 158 | |
| 159 | const __m256 vi11x01234567 = _mm256_loadu_ps(i11); |
| 160 | i11 += 8; |
| 161 | |
| 162 | const __m256 vk11x01234567 = _mm256_load_ps(w + 96); |
| 163 | vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1); |
| 164 | |
| 165 | const __m256 vi12x01234567 = _mm256_loadu_ps(i12); |
| 166 | i12 += 8; |
| 167 | |
| 168 | const __m256 vk12x01234567 = _mm256_load_ps(w + 104); |
| 169 | vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); |
| 170 | |
| 171 | const __m256 vi13x01234567 = _mm256_loadu_ps(i13); |
| 172 | i13 += 8; |
| 173 | |
| 174 | const __m256 vk13x01234567 = _mm256_load_ps(w + 112); |
| 175 | vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1); |
| 176 | |
| 177 | const __m256 vi14x01234567 = _mm256_loadu_ps(i14); |
| 178 | i14 += 8; |
| 179 | |
| 180 | const __m256 vk14x01234567 = _mm256_load_ps(w + 120); |
| 181 | vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); |
| 182 | |
| 183 | const __m256 vi15x01234567 = _mm256_loadu_ps(i15); |
| 184 | i15 += 8; |
| 185 | |
| 186 | const __m256 vk15x01234567 = _mm256_load_ps(w + 128); |
| 187 | vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1); |
| 188 | |
| 189 | const __m256 vi16x01234567 = _mm256_loadu_ps(i16); |
| 190 | i16 += 8; |
| 191 | |
| 192 | const __m256 vk16x01234567 = _mm256_load_ps(w + 136); |
| 193 | vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); |
| 194 | |
| 195 | const __m256 vi17x01234567 = _mm256_loadu_ps(i17); |
| 196 | i17 += 8; |
| 197 | |
| 198 | const __m256 vk17x01234567 = _mm256_load_ps(w + 144); |
| 199 | vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1); |
| 200 | |
| 201 | const __m256 vi18x01234567 = _mm256_loadu_ps(i18); |
| 202 | i18 += 8; |
| 203 | |
| 204 | const __m256 vk18x01234567 = _mm256_load_ps(w + 152); |
| 205 | vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); |
| 206 | |
| 207 | const __m256 vi19x01234567 = _mm256_loadu_ps(i19); |
| 208 | i19 += 8; |
| 209 | |
| 210 | const __m256 vk19x01234567 = _mm256_load_ps(w + 160); |
| 211 | vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1); |
| 212 | |
| 213 | const __m256 vi20x01234567 = _mm256_loadu_ps(i20); |
| 214 | i20 += 8; |
| 215 | |
| 216 | const __m256 vk20x01234567 = _mm256_load_ps(w + 168); |
| 217 | vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); |
| 218 | |
| 219 | const __m256 vi21x01234567 = _mm256_loadu_ps(i21); |
| 220 | i21 += 8; |
| 221 | |
| 222 | const __m256 vk21x01234567 = _mm256_load_ps(w + 176); |
| 223 | vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); |
| 224 | |
| 225 | const __m256 vi22x01234567 = _mm256_loadu_ps(i22); |
| 226 | i22 += 8; |
| 227 | |
| 228 | const __m256 vk22x01234567 = _mm256_load_ps(w + 184); |
| 229 | vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); |
| 230 | |
| 231 | const __m256 vi23x01234567 = _mm256_loadu_ps(i23); |
| 232 | i23 += 8; |
| 233 | |
| 234 | const __m256 vk23x01234567 = _mm256_load_ps(w + 192); |
| 235 | vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1); |
| 236 | |
| 237 | const __m256 vi24x01234567 = _mm256_loadu_ps(i24); |
| 238 | i24 += 8; |
| 239 | |
| 240 | const __m256 vk24x01234567 = _mm256_load_ps(w + 200); |
| 241 | vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); |
| 242 | |
| 243 | w += 208; |
| 244 | |
| 245 | // Add up all accumulators to vacc01234567p0 |
| 246 | vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1); |
| 247 | |
| 248 | __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin); |
| 249 | vacc01234567 = _mm256_min_ps(vacc01234567, vmax); |
| 250 | |
| 251 | _mm256_storeu_ps(output, vacc01234567); |
| 252 | output += 8; |
| 253 | } |
| 254 | if XNN_UNLIKELY(c != 0) { |
| 255 | assert(c >= 1); |
| 256 | assert(c <= 7); |
| 257 | __m256i vmask = _mm256_loadu_si256((const __m256i*) &mask_table[7 - c]); |
| 258 | |
| 259 | __m256 vacc01234567p0 = _mm256_load_ps(w); |
| 260 | |
| 261 | const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); |
| 262 | const __m256 vk0x01234567 = _mm256_load_ps(w + 8); |
| 263 | vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); |
| 264 | |
| 265 | const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); |
| 266 | const __m256 vk1x01234567 = _mm256_load_ps(w + 16); |
| 267 | __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567); |
| 268 | |
| 269 | const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask); |
| 270 | const __m256 vk2x01234567 = _mm256_load_ps(w + 24); |
| 271 | vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0); |
| 272 | |
| 273 | const __m256 vi3x01234567 = _mm256_maskload_ps(i3, vmask); |
| 274 | const __m256 vk3x01234567 = _mm256_load_ps(w + 32); |
| 275 | vacc01234567p1 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1); |
| 276 | |
| 277 | const __m256 vi4x01234567 = _mm256_maskload_ps(i4, vmask); |
| 278 | const __m256 vk4x01234567 = _mm256_load_ps(w + 40); |
| 279 | vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0); |
| 280 | |
| 281 | const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); |
| 282 | const __m256 vk5x01234567 = _mm256_load_ps(w + 48); |
| 283 | vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); |
| 284 | |
| 285 | const __m256 vi6x01234567 = _mm256_maskload_ps(i6, vmask); |
| 286 | const __m256 vk6x01234567 = _mm256_load_ps(w + 56); |
| 287 | vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); |
| 288 | |
| 289 | const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); |
| 290 | const __m256 vk7x01234567 = _mm256_load_ps(w + 64); |
| 291 | vacc01234567p1 = _mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1); |
| 292 | |
| 293 | const __m256 vi8x01234567 = _mm256_maskload_ps(i8, vmask); |
| 294 | const __m256 vk8x01234567 = _mm256_load_ps(w + 72); |
| 295 | vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0); |
| 296 | |
| 297 | const __m256 vi9x01234567 = _mm256_maskload_ps(i9, vmask); |
| 298 | const __m256 vk9x01234567 = _mm256_load_ps(w + 80); |
| 299 | vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1); |
| 300 | |
| 301 | const __m256 vi10x01234567 = _mm256_maskload_ps(i10, vmask); |
| 302 | const __m256 vk10x01234567 = _mm256_load_ps(w + 88); |
| 303 | vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); |
| 304 | |
| 305 | const __m256 vi11x01234567 = _mm256_maskload_ps(i11, vmask); |
| 306 | const __m256 vk11x01234567 = _mm256_load_ps(w + 96); |
| 307 | vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1); |
| 308 | |
| 309 | const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); |
| 310 | const __m256 vk12x01234567 = _mm256_load_ps(w + 104); |
| 311 | vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); |
| 312 | |
| 313 | const __m256 vi13x01234567 = _mm256_maskload_ps(i13, vmask); |
| 314 | const __m256 vk13x01234567 = _mm256_load_ps(w + 112); |
| 315 | vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1); |
| 316 | |
| 317 | const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); |
| 318 | const __m256 vk14x01234567 = _mm256_load_ps(w + 120); |
| 319 | vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); |
| 320 | |
| 321 | const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); |
| 322 | const __m256 vk15x01234567 = _mm256_load_ps(w + 128); |
| 323 | vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1); |
| 324 | |
| 325 | const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); |
| 326 | const __m256 vk16x01234567 = _mm256_load_ps(w + 136); |
| 327 | vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); |
| 328 | |
| 329 | const __m256 vi17x01234567 = _mm256_maskload_ps(i17, vmask); |
| 330 | const __m256 vk17x01234567 = _mm256_load_ps(w + 144); |
| 331 | vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1); |
| 332 | |
| 333 | const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); |
| 334 | const __m256 vk18x01234567 = _mm256_load_ps(w + 152); |
| 335 | vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); |
| 336 | |
| 337 | const __m256 vi19x01234567 = _mm256_maskload_ps(i19, vmask); |
| 338 | const __m256 vk19x01234567 = _mm256_load_ps(w + 160); |
| 339 | vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1); |
| 340 | |
| 341 | const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); |
| 342 | const __m256 vk20x01234567 = _mm256_load_ps(w + 168); |
| 343 | vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); |
| 344 | |
| 345 | const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); |
| 346 | const __m256 vk21x01234567 = _mm256_load_ps(w + 176); |
| 347 | vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); |
| 348 | |
| 349 | const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); |
| 350 | const __m256 vk22x01234567 = _mm256_load_ps(w + 184); |
| 351 | vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); |
| 352 | |
| 353 | const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); |
| 354 | const __m256 vk23x01234567 = _mm256_load_ps(w + 192); |
| 355 | vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1); |
| 356 | |
| 357 | const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); |
| 358 | const __m256 vk24x01234567 = _mm256_load_ps(w + 200); |
| 359 | vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); |
| 360 | |
| 361 | // Add up all accumulators to vacc01234567p0 |
| 362 | vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1); |
| 363 | |
| 364 | __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin); |
| 365 | vacc01234567 = _mm256_min_ps(vacc01234567, vmax); |
| 366 | |
Marat Dukhan | d16d0f0 | 2020-02-05 21:33:05 -0800 | [diff] [blame] | 367 | // _mm256_maskstore_ps(output, vmask, vacc01234567); output += c; could be used here, but triggers msan failures (probably an msan bug). |
| 368 | __m128 vacc0123 = _mm256_castps256_ps128(vacc01234567); |
| 369 | if (c & 4) { |
| 370 | _mm_storeu_ps(output, vacc0123); |
| 371 | vacc0123 = _mm256_extractf128_ps(vacc01234567, 1); |
| 372 | output += 4; |
| 373 | } |
| 374 | if (c & 2) { |
| 375 | _mm_storel_pi((__m64*) output, vacc0123); |
| 376 | vacc0123 = _mm_movehl_ps(vacc0123, vacc0123); |
| 377 | output += 2; |
| 378 | } |
| 379 | if (c & 1) { |
| 380 | _mm_store_ss(output, vacc0123); |
| 381 | output += 1; |
| 382 | } |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 383 | } |
| 384 | |
| 385 | output = (float*) ((uintptr_t) output + output_increment); |
| 386 | } while (--output_width != 0); |
| 387 | } |