Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-gemm/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/gemm.h> |
| 15 | |
| 16 | |
| 17 | void xnn_f32_gemm_ukernel_4x16__avx_broadcast( |
| 18 | size_t mr, |
| 19 | size_t nc, |
| 20 | size_t kc, |
| 21 | const float*restrict a, |
| 22 | size_t a_stride, |
| 23 | const float*restrict w, |
| 24 | float*restrict c, |
| 25 | size_t cm_stride, |
| 26 | size_t cn_stride, |
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 | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 28 | { |
| 29 | assert(mr != 0); |
| 30 | assert(mr <= 4); |
| 31 | assert(nc != 0); |
| 32 | assert(kc != 0); |
| 33 | assert(kc % sizeof(float) == 0); |
| 34 | assert(a != NULL); |
| 35 | assert(w != NULL); |
| 36 | assert(c != NULL); |
| 37 | |
| 38 | const float* a0 = a; |
| 39 | float* c0 = c; |
| 40 | const float* a1 = (const float*) ((uintptr_t) a0 + a_stride); |
| 41 | float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 42 | if XNN_UNPREDICTABLE(mr < 2) { |
| 43 | a1 = a0; |
| 44 | c1 = c0; |
| 45 | } |
| 46 | const float* a2 = (const float*) ((uintptr_t) a1 + a_stride); |
| 47 | float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 48 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 49 | a2 = a1; |
| 50 | c2 = c1; |
| 51 | } |
| 52 | const float* a3 = (const float*) ((uintptr_t) a2 + a_stride); |
| 53 | float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 54 | if XNN_UNPREDICTABLE(mr != 4) { |
| 55 | a3 = a2; |
| 56 | c3 = c2; |
| 57 | } |
| 58 | |
| 59 | do { |
| 60 | __m256 vacc0x01234567 = _mm256_load_ps(w + 0); |
| 61 | __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8); |
| 62 | __m256 vacc1x01234567 = vacc0x01234567; |
| 63 | __m256 vacc1x89ABCDEF = vacc0x89ABCDEF; |
| 64 | __m256 vacc2x01234567 = vacc0x01234567; |
| 65 | __m256 vacc2x89ABCDEF = vacc0x89ABCDEF; |
| 66 | __m256 vacc3x01234567 = vacc0x01234567; |
| 67 | __m256 vacc3x89ABCDEF = vacc0x89ABCDEF; |
| 68 | w += 16; |
| 69 | |
| 70 | size_t k = kc; |
| 71 | do { |
| 72 | const __m256 va0 = _mm256_broadcast_ss(a0); |
| 73 | a0 += 1; |
| 74 | const __m256 va1 = _mm256_broadcast_ss(a1); |
| 75 | a1 += 1; |
| 76 | const __m256 va2 = _mm256_broadcast_ss(a2); |
| 77 | a2 += 1; |
| 78 | const __m256 va3 = _mm256_broadcast_ss(a3); |
| 79 | a3 += 1; |
| 80 | |
| 81 | const __m256 vb01234567 = _mm256_load_ps(w); |
| 82 | const __m256 vb89ABCDEF = _mm256_load_ps(w + 8); |
| 83 | w += 16; |
| 84 | |
| 85 | vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567)); |
| 86 | vacc1x01234567 = _mm256_add_ps(vacc1x01234567, _mm256_mul_ps(va1, vb01234567)); |
| 87 | vacc2x01234567 = _mm256_add_ps(vacc2x01234567, _mm256_mul_ps(va2, vb01234567)); |
| 88 | vacc3x01234567 = _mm256_add_ps(vacc3x01234567, _mm256_mul_ps(va3, vb01234567)); |
| 89 | vacc0x89ABCDEF = _mm256_add_ps(vacc0x89ABCDEF, _mm256_mul_ps(va0, vb89ABCDEF)); |
| 90 | vacc1x89ABCDEF = _mm256_add_ps(vacc1x89ABCDEF, _mm256_mul_ps(va1, vb89ABCDEF)); |
| 91 | vacc2x89ABCDEF = _mm256_add_ps(vacc2x89ABCDEF, _mm256_mul_ps(va2, vb89ABCDEF)); |
| 92 | vacc3x89ABCDEF = _mm256_add_ps(vacc3x89ABCDEF, _mm256_mul_ps(va3, vb89ABCDEF)); |
| 93 | |
| 94 | k -= sizeof(float); |
| 95 | } while (k != 0); |
| 96 | |
| 97 | const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max); |
| 98 | vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax); |
| 99 | vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax); |
| 100 | vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax); |
| 101 | vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax); |
| 102 | vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax); |
| 103 | vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax); |
| 104 | vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax); |
| 105 | vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax); |
| 106 | |
| 107 | const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min); |
| 108 | vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin); |
| 109 | vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin); |
| 110 | vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin); |
| 111 | vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin); |
| 112 | vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin); |
| 113 | vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin); |
| 114 | vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin); |
| 115 | vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin); |
| 116 | |
| 117 | if XNN_LIKELY(nc >= 16) { |
| 118 | _mm256_storeu_ps(c3, vacc3x01234567); |
| 119 | _mm256_storeu_ps(c3 + 8, vacc3x89ABCDEF); |
| 120 | c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 121 | _mm256_storeu_ps(c2, vacc2x01234567); |
| 122 | _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF); |
| 123 | c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 124 | _mm256_storeu_ps(c1, vacc1x01234567); |
| 125 | _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF); |
| 126 | c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 127 | _mm256_storeu_ps(c0, vacc0x01234567); |
| 128 | _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF); |
| 129 | c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 130 | |
| 131 | a3 = (const float*) ((uintptr_t) a3 - kc); |
| 132 | a2 = (const float*) ((uintptr_t) a2 - kc); |
| 133 | a1 = (const float*) ((uintptr_t) a1 - kc); |
| 134 | a0 = (const float*) ((uintptr_t) a0 - kc); |
| 135 | |
| 136 | nc -= 16; |
| 137 | } else { |
| 138 | if (nc & 8) { |
| 139 | _mm256_storeu_ps(c3, vacc3x01234567); |
| 140 | _mm256_storeu_ps(c2, vacc2x01234567); |
| 141 | _mm256_storeu_ps(c1, vacc1x01234567); |
| 142 | _mm256_storeu_ps(c0, vacc0x01234567); |
| 143 | |
| 144 | vacc3x01234567 = vacc3x89ABCDEF; |
| 145 | vacc2x01234567 = vacc2x89ABCDEF; |
| 146 | vacc1x01234567 = vacc1x89ABCDEF; |
| 147 | vacc0x01234567 = vacc0x89ABCDEF; |
| 148 | |
| 149 | c3 += 8; |
| 150 | c2 += 8; |
| 151 | c1 += 8; |
| 152 | c0 += 8; |
| 153 | } |
| 154 | __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567); |
| 155 | __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567); |
| 156 | __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567); |
| 157 | __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567); |
| 158 | if (nc & 4) { |
| 159 | _mm_storeu_ps(c3, vacc3x0123); |
| 160 | _mm_storeu_ps(c2, vacc2x0123); |
| 161 | _mm_storeu_ps(c1, vacc1x0123); |
| 162 | _mm_storeu_ps(c0, vacc0x0123); |
| 163 | |
| 164 | vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1); |
| 165 | vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1); |
| 166 | vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1); |
| 167 | vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1); |
| 168 | |
| 169 | c3 += 4; |
| 170 | c2 += 4; |
| 171 | c1 += 4; |
| 172 | c0 += 4; |
| 173 | } |
| 174 | if (nc & 2) { |
| 175 | _mm_storel_pi((__m64*) c3, vacc3x0123); |
| 176 | _mm_storel_pi((__m64*) c2, vacc2x0123); |
| 177 | _mm_storel_pi((__m64*) c1, vacc1x0123); |
| 178 | _mm_storel_pi((__m64*) c0, vacc0x0123); |
| 179 | |
| 180 | vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123); |
| 181 | vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123); |
| 182 | vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123); |
| 183 | vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123); |
| 184 | |
| 185 | c3 += 2; |
| 186 | c2 += 2; |
| 187 | c1 += 2; |
| 188 | c0 += 2; |
| 189 | } |
| 190 | if (nc & 1) { |
| 191 | _mm_store_ss(c3, vacc3x0123); |
| 192 | _mm_store_ss(c2, vacc2x0123); |
| 193 | _mm_store_ss(c1, vacc1x0123); |
| 194 | _mm_store_ss(c0, vacc0x0123); |
| 195 | } |
| 196 | |
| 197 | nc = 0; |
| 198 | } |
| 199 | } while (nc != 0); |
| 200 | } |