Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-igemm/avx-broadcast.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/igemm.h> |
| 15 | |
| 16 | |
| 17 | void xnn_f32_igemm_ukernel_7x8__fma3_broadcast( |
| 18 | size_t mr, |
| 19 | size_t nc, |
| 20 | size_t kc, |
| 21 | size_t ks, |
| 22 | const float**restrict a, |
| 23 | const float*restrict w, |
| 24 | float*restrict c, |
| 25 | size_t cm_stride, |
| 26 | size_t cn_stride, |
| 27 | size_t a_offset, |
| 28 | const float* zero, |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 29 | const union xnn_f32_minmax_params params[restrict static 1]) |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 30 | { |
| 31 | assert(mr != 0); |
| 32 | assert(mr <= 7); |
| 33 | assert(nc != 0); |
| 34 | assert(kc != 0); |
| 35 | assert(kc % sizeof(float) == 0); |
| 36 | assert(ks != 0); |
| 37 | assert(ks % (7 * sizeof(void*)) == 0); |
| 38 | assert(a_offset % sizeof(float) == 0); |
| 39 | assert(a != NULL); |
| 40 | assert(w != NULL); |
| 41 | assert(c != NULL); |
| 42 | |
| 43 | float* c0 = c; |
| 44 | float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 45 | if XNN_UNPREDICTABLE(mr < 2) { |
| 46 | c1 = c0; |
| 47 | } |
| 48 | float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 49 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 50 | c2 = c1; |
| 51 | } |
| 52 | float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 53 | if XNN_UNPREDICTABLE(mr < 4) { |
| 54 | c3 = c2; |
| 55 | } |
| 56 | float* c4 = (float*) ((uintptr_t) c3 + cm_stride); |
| 57 | if XNN_UNPREDICTABLE(mr <= 4) { |
| 58 | c4 = c3; |
| 59 | } |
| 60 | float* c5 = (float*) ((uintptr_t) c4 + cm_stride); |
| 61 | if XNN_UNPREDICTABLE(mr < 6) { |
| 62 | c5 = c4; |
| 63 | } |
| 64 | float* c6 = (float*) ((uintptr_t) c5 + cm_stride); |
| 65 | if XNN_UNPREDICTABLE(mr <= 6) { |
| 66 | c6 = c5; |
| 67 | } |
| 68 | |
| 69 | do { |
| 70 | __m256 vacc0x01234567 = _mm256_load_ps(w); |
| 71 | __m256 vacc1x01234567 = vacc0x01234567; |
| 72 | __m256 vacc2x01234567 = vacc0x01234567; |
| 73 | __m256 vacc3x01234567 = vacc0x01234567; |
| 74 | __m256 vacc4x01234567 = vacc0x01234567; |
| 75 | __m256 vacc5x01234567 = vacc0x01234567; |
| 76 | __m256 vacc6x01234567 = vacc0x01234567; |
| 77 | w += 8; |
| 78 | |
| 79 | size_t p = ks; |
| 80 | do { |
| 81 | const float* restrict a0 = a[0]; |
| 82 | assert(a0 != NULL); |
| 83 | if XNN_UNPREDICTABLE(a0 != zero) { |
| 84 | a0 = (const float*) ((uintptr_t) a0 + a_offset); |
| 85 | } |
| 86 | const float* restrict a1 = a[1]; |
| 87 | assert(a1 != NULL); |
| 88 | if XNN_UNPREDICTABLE(a1 != zero) { |
| 89 | a1 = (const float*) ((uintptr_t) a1 + a_offset); |
| 90 | } |
| 91 | const float* restrict a2 = a[2]; |
| 92 | assert(a2 != NULL); |
| 93 | if XNN_UNPREDICTABLE(a2 != zero) { |
| 94 | a2 = (const float*) ((uintptr_t) a2 + a_offset); |
| 95 | } |
| 96 | const float* restrict a3 = a[3]; |
| 97 | assert(a3 != NULL); |
| 98 | if XNN_UNPREDICTABLE(a3 != zero) { |
| 99 | a3 = (const float*) ((uintptr_t) a3 + a_offset); |
| 100 | } |
| 101 | const float* restrict a4 = a[4]; |
| 102 | assert(a4 != NULL); |
| 103 | if XNN_UNPREDICTABLE(a4 != zero) { |
| 104 | a4 = (const float*) ((uintptr_t) a4 + a_offset); |
| 105 | } |
| 106 | const float* restrict a5 = a[5]; |
| 107 | assert(a5 != NULL); |
| 108 | if XNN_UNPREDICTABLE(a5 != zero) { |
| 109 | a5 = (const float*) ((uintptr_t) a5 + a_offset); |
| 110 | } |
| 111 | const float* restrict a6 = a[6]; |
| 112 | assert(a6 != NULL); |
| 113 | if XNN_UNPREDICTABLE(a6 != zero) { |
| 114 | a6 = (const float*) ((uintptr_t) a6 + a_offset); |
| 115 | } |
| 116 | a += 7; |
| 117 | |
| 118 | size_t k = kc; |
| 119 | do { |
| 120 | const __m256 vb01234567 = _mm256_load_ps(w); |
| 121 | w += 8; |
| 122 | |
| 123 | const __m256 va0 = _mm256_broadcast_ss(a0); |
| 124 | a0 += 1; |
| 125 | const __m256 va1 = _mm256_broadcast_ss(a1); |
| 126 | a1 += 1; |
| 127 | const __m256 va2 = _mm256_broadcast_ss(a2); |
| 128 | a2 += 1; |
| 129 | const __m256 va3 = _mm256_broadcast_ss(a3); |
| 130 | a3 += 1; |
| 131 | const __m256 va4 = _mm256_broadcast_ss(a4); |
| 132 | a4 += 1; |
| 133 | const __m256 va5 = _mm256_broadcast_ss(a5); |
| 134 | a5 += 1; |
| 135 | const __m256 va6 = _mm256_broadcast_ss(a6); |
| 136 | a6 += 1; |
| 137 | |
| 138 | vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567); |
| 139 | vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567); |
| 140 | vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567); |
| 141 | vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567, vacc3x01234567); |
| 142 | vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567, vacc4x01234567); |
| 143 | vacc5x01234567 = _mm256_fmadd_ps(va5, vb01234567, vacc5x01234567); |
| 144 | vacc6x01234567 = _mm256_fmadd_ps(va6, vb01234567, vacc6x01234567); |
| 145 | k -= sizeof(float); |
| 146 | } while (k != 0); |
| 147 | p -= 7 * sizeof(void*); |
| 148 | } while (p != 0); |
| 149 | |
| 150 | const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max); |
| 151 | vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax); |
| 152 | vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax); |
| 153 | vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax); |
| 154 | vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax); |
| 155 | vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax); |
| 156 | vacc5x01234567 = _mm256_min_ps(vacc5x01234567, vmax); |
| 157 | vacc6x01234567 = _mm256_min_ps(vacc6x01234567, vmax); |
| 158 | |
| 159 | const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min); |
| 160 | vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin); |
| 161 | vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin); |
| 162 | vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin); |
| 163 | vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin); |
| 164 | vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin); |
| 165 | vacc5x01234567 = _mm256_max_ps(vacc5x01234567, vmin); |
| 166 | vacc6x01234567 = _mm256_max_ps(vacc6x01234567, vmin); |
| 167 | |
| 168 | if XNN_LIKELY(nc >= 8) { |
| 169 | _mm256_storeu_ps(c6, vacc6x01234567); |
| 170 | c6 = (float*) ((uintptr_t) c6 + cn_stride); |
| 171 | _mm256_storeu_ps(c5, vacc5x01234567); |
| 172 | c5 = (float*) ((uintptr_t) c5 + cn_stride); |
| 173 | _mm256_storeu_ps(c4, vacc4x01234567); |
| 174 | c4 = (float*) ((uintptr_t) c4 + cn_stride); |
| 175 | _mm256_storeu_ps(c3, vacc3x01234567); |
| 176 | c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 177 | _mm256_storeu_ps(c2, vacc2x01234567); |
| 178 | c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 179 | _mm256_storeu_ps(c1, vacc1x01234567); |
| 180 | c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 181 | _mm256_storeu_ps(c0, vacc0x01234567); |
| 182 | c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 183 | |
| 184 | a = (const float**restrict) ((uintptr_t) a - ks); |
| 185 | nc -= 8; |
| 186 | } else { |
| 187 | __m128 vacc6x0123 = _mm256_castps256_ps128(vacc6x01234567); |
| 188 | __m128 vacc5x0123 = _mm256_castps256_ps128(vacc5x01234567); |
| 189 | __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567); |
| 190 | __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567); |
| 191 | __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567); |
| 192 | __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567); |
| 193 | __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567); |
| 194 | if (nc & 4) { |
| 195 | _mm_storeu_ps(c6, vacc6x0123); |
| 196 | _mm_storeu_ps(c5, vacc5x0123); |
| 197 | _mm_storeu_ps(c4, vacc4x0123); |
| 198 | _mm_storeu_ps(c3, vacc3x0123); |
| 199 | _mm_storeu_ps(c2, vacc2x0123); |
| 200 | _mm_storeu_ps(c1, vacc1x0123); |
| 201 | _mm_storeu_ps(c0, vacc0x0123); |
| 202 | |
| 203 | vacc6x0123 = _mm256_extractf128_ps(vacc6x01234567, 1); |
| 204 | vacc5x0123 = _mm256_extractf128_ps(vacc5x01234567, 1); |
| 205 | vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1); |
| 206 | vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1); |
| 207 | vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1); |
| 208 | vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1); |
| 209 | vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1); |
| 210 | |
| 211 | c6 += 4; |
| 212 | c5 += 4; |
| 213 | c4 += 4; |
| 214 | c3 += 4; |
| 215 | c2 += 4; |
| 216 | c1 += 4; |
| 217 | c0 += 4; |
| 218 | } |
| 219 | if (nc & 2) { |
| 220 | _mm_storel_pi((__m64*) c6, vacc6x0123); |
| 221 | _mm_storel_pi((__m64*) c5, vacc5x0123); |
| 222 | _mm_storel_pi((__m64*) c4, vacc4x0123); |
| 223 | _mm_storel_pi((__m64*) c3, vacc3x0123); |
| 224 | _mm_storel_pi((__m64*) c2, vacc2x0123); |
| 225 | _mm_storel_pi((__m64*) c1, vacc1x0123); |
| 226 | _mm_storel_pi((__m64*) c0, vacc0x0123); |
| 227 | |
| 228 | vacc6x0123 = _mm_movehl_ps(vacc6x0123, vacc6x0123); |
| 229 | vacc5x0123 = _mm_movehl_ps(vacc5x0123, vacc5x0123); |
| 230 | vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123); |
| 231 | vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123); |
| 232 | vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123); |
| 233 | vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123); |
| 234 | vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123); |
| 235 | |
| 236 | c6 += 2; |
| 237 | c5 += 2; |
| 238 | c4 += 2; |
| 239 | c3 += 2; |
| 240 | c2 += 2; |
| 241 | c1 += 2; |
| 242 | c0 += 2; |
| 243 | } |
| 244 | if (nc & 1) { |
| 245 | _mm_store_ss(c6, vacc6x0123); |
| 246 | _mm_store_ss(c5, vacc5x0123); |
| 247 | _mm_store_ss(c4, vacc4x0123); |
| 248 | _mm_store_ss(c3, vacc3x0123); |
| 249 | _mm_store_ss(c2, vacc2x0123); |
| 250 | _mm_store_ss(c1, vacc1x0123); |
| 251 | _mm_store_ss(c0, vacc0x0123); |
| 252 | } |
| 253 | |
| 254 | nc = 0; |
| 255 | } |
| 256 | } while (nc != 0); |
| 257 | } |