Frank Barchard | df06d80 | 2019-11-20 15:53:46 -0800 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-gemm/neon-shuffle.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 <arm_neon.h> |
| 13 | |
| 14 | #include <xnnpack/gemm.h> |
| 15 | |
| 16 | |
| 17 | void xnn_f32_gemm_ukernel_6x8s4__neonfma( |
| 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, |
| 27 | const union xnn_f32_output_params params[restrict static 1]) |
| 28 | { |
| 29 | assert(mr != 0); |
| 30 | assert(mr <= 6); |
| 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 | const float* a4 = (const float*) ((uintptr_t) a3 + a_stride); |
| 59 | float* c4 = (float*) ((uintptr_t) c3 + cm_stride); |
| 60 | if XNN_UNPREDICTABLE(mr <= 4) { |
| 61 | a4 = a3; |
| 62 | c4 = c3; |
| 63 | } |
| 64 | const float* a5 = (const float*) ((uintptr_t) a4 + a_stride); |
| 65 | float* c5 = (float*) ((uintptr_t) c4 + cm_stride); |
| 66 | if XNN_UNPREDICTABLE(mr != 6) { |
| 67 | a5 = a4; |
| 68 | c5 = c4; |
| 69 | } |
| 70 | |
| 71 | do { |
| 72 | float32x4_t vacc0x0123 = vld1q_f32(w); w += 4; |
| 73 | float32x4_t vacc0x4567 = vld1q_f32(w); w += 4; |
| 74 | float32x4_t vacc1x0123 = vacc0x0123; |
| 75 | float32x4_t vacc1x4567 = vacc0x4567; |
| 76 | float32x4_t vacc2x0123 = vacc0x0123; |
| 77 | float32x4_t vacc2x4567 = vacc0x4567; |
| 78 | float32x4_t vacc3x0123 = vacc0x0123; |
| 79 | float32x4_t vacc3x4567 = vacc0x4567; |
| 80 | float32x4_t vacc4x0123 = vacc0x0123; |
| 81 | float32x4_t vacc4x4567 = vacc0x4567; |
| 82 | float32x4_t vacc5x0123 = vacc0x0123; |
| 83 | float32x4_t vacc5x4567 = vacc0x4567; |
| 84 | |
| 85 | size_t k = kc; |
| 86 | while (k >= 4 * sizeof(float)) { |
| 87 | float32x4_t va0 = vld1q_f32(a0); a0 += 4; |
| 88 | float32x4_t va1 = vld1q_f32(a1); a1 += 4; |
| 89 | float32x4_t va2 = vld1q_f32(a2); a2 += 4; |
| 90 | float32x4_t va3 = vld1q_f32(a3); a3 += 4; |
| 91 | float32x4_t va4 = vld1q_f32(a4); a4 += 4; |
| 92 | float32x4_t va5 = vld1q_f32(a5); a5 += 4; |
| 93 | |
| 94 | |
| 95 | const float32x4_t vb0123c0 = vld1q_f32(w + 0); |
| 96 | const float32x4_t vb4567c0 = vld1q_f32(w + 4); |
| 97 | |
| 98 | vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c0); |
| 99 | vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c0); |
| 100 | vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c0); |
| 101 | vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c0); |
| 102 | vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c0); |
| 103 | vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c0); |
| 104 | vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c0); |
| 105 | vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c0); |
| 106 | vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c0); |
| 107 | vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c0); |
| 108 | vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c0); |
| 109 | vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c0); |
| 110 | |
| 111 | va0 = vextq_f32(va0, va0, 1); |
| 112 | va1 = vextq_f32(va1, va1, 1); |
| 113 | va2 = vextq_f32(va2, va2, 1); |
| 114 | va3 = vextq_f32(va3, va3, 1); |
| 115 | va4 = vextq_f32(va4, va4, 1); |
| 116 | va5 = vextq_f32(va5, va5, 1); |
| 117 | |
| 118 | const float32x4_t vb0123c1 = vld1q_f32(w + 8); |
| 119 | const float32x4_t vb4567c1 = vld1q_f32(w + 12); |
| 120 | |
| 121 | vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c1); |
| 122 | vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c1); |
| 123 | vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c1); |
| 124 | vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c1); |
| 125 | vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c1); |
| 126 | vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c1); |
| 127 | vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c1); |
| 128 | vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c1); |
| 129 | vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c1); |
| 130 | vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c1); |
| 131 | vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c1); |
| 132 | vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c1); |
| 133 | |
| 134 | va0 = vextq_f32(va0, va0, 1); |
| 135 | va1 = vextq_f32(va1, va1, 1); |
| 136 | va2 = vextq_f32(va2, va2, 1); |
| 137 | va3 = vextq_f32(va3, va3, 1); |
| 138 | va4 = vextq_f32(va4, va4, 1); |
| 139 | va5 = vextq_f32(va5, va5, 1); |
| 140 | |
| 141 | const float32x4_t vb0123c2 = vld1q_f32(w + 16); |
| 142 | const float32x4_t vb4567c2 = vld1q_f32(w + 20); |
| 143 | |
| 144 | vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c2); |
| 145 | vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c2); |
| 146 | vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c2); |
| 147 | vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c2); |
| 148 | vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c2); |
| 149 | vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c2); |
| 150 | vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c2); |
| 151 | vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c2); |
| 152 | vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c2); |
| 153 | vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c2); |
| 154 | vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c2); |
| 155 | vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c2); |
| 156 | |
| 157 | va0 = vextq_f32(va0, va0, 1); |
| 158 | va1 = vextq_f32(va1, va1, 1); |
| 159 | va2 = vextq_f32(va2, va2, 1); |
| 160 | va3 = vextq_f32(va3, va3, 1); |
| 161 | va4 = vextq_f32(va4, va4, 1); |
| 162 | va5 = vextq_f32(va5, va5, 1); |
| 163 | |
| 164 | const float32x4_t vb0123c3 = vld1q_f32(w + 24); |
| 165 | const float32x4_t vb4567c3 = vld1q_f32(w + 28); |
| 166 | |
| 167 | vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c3); |
| 168 | vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c3); |
| 169 | vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c3); |
| 170 | vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c3); |
| 171 | vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c3); |
| 172 | vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c3); |
| 173 | vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c3); |
| 174 | vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c3); |
| 175 | vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c3); |
| 176 | vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c3); |
| 177 | vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c3); |
| 178 | vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c3); |
| 179 | |
| 180 | |
| 181 | w += 32; |
| 182 | k -= 4 * sizeof(float); |
| 183 | } |
| 184 | if XNN_UNLIKELY(k != 0) { |
| 185 | do { |
| 186 | const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1; |
| 187 | const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1; |
| 188 | const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1; |
| 189 | const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1; |
| 190 | const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1; |
| 191 | const float32x4_t va5 = vld1q_dup_f32(a5); a5 += 1; |
| 192 | |
| 193 | const float32x4_t vb0123 = vld1q_f32(w); w += 4; |
| 194 | const float32x4_t vb4567 = vld1q_f32(w); w += 4; |
| 195 | |
| 196 | vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123); |
| 197 | vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123); |
| 198 | vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123); |
| 199 | vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123); |
| 200 | vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123); |
| 201 | vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123); |
| 202 | vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567); |
| 203 | vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567); |
| 204 | vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567); |
| 205 | vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567); |
| 206 | vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567); |
| 207 | vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567); |
| 208 | |
| 209 | k -= sizeof(float); |
| 210 | } while (k != 0); |
| 211 | } |
| 212 | const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); |
| 213 | vacc0x0123 = vminq_f32(vacc0x0123, vmax); |
| 214 | vacc1x0123 = vminq_f32(vacc1x0123, vmax); |
| 215 | vacc2x0123 = vminq_f32(vacc2x0123, vmax); |
| 216 | vacc3x0123 = vminq_f32(vacc3x0123, vmax); |
| 217 | vacc4x0123 = vminq_f32(vacc4x0123, vmax); |
| 218 | vacc5x0123 = vminq_f32(vacc5x0123, vmax); |
| 219 | vacc0x4567 = vminq_f32(vacc0x4567, vmax); |
| 220 | vacc1x4567 = vminq_f32(vacc1x4567, vmax); |
| 221 | vacc2x4567 = vminq_f32(vacc2x4567, vmax); |
| 222 | vacc3x4567 = vminq_f32(vacc3x4567, vmax); |
| 223 | vacc4x4567 = vminq_f32(vacc4x4567, vmax); |
| 224 | vacc5x4567 = vminq_f32(vacc5x4567, vmax); |
| 225 | |
| 226 | const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); |
| 227 | vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); |
| 228 | vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); |
| 229 | vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); |
| 230 | vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); |
| 231 | vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); |
| 232 | vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); |
| 233 | vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); |
| 234 | vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); |
| 235 | vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); |
| 236 | vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); |
| 237 | vacc4x4567 = vmaxq_f32(vacc4x4567, vmin); |
| 238 | vacc5x4567 = vmaxq_f32(vacc5x4567, vmin); |
| 239 | |
| 240 | if XNN_LIKELY(nc >= 8) { |
| 241 | vst1q_f32(c5, vacc5x0123); |
| 242 | vst1q_f32(c5 + 4, vacc5x4567); |
| 243 | c5 = (float*) ((uintptr_t) c5 + cn_stride); |
| 244 | vst1q_f32(c4, vacc4x0123); |
| 245 | vst1q_f32(c4 + 4, vacc4x4567); |
| 246 | c4 = (float*) ((uintptr_t) c4 + cn_stride); |
| 247 | vst1q_f32(c3, vacc3x0123); |
| 248 | vst1q_f32(c3 + 4, vacc3x4567); |
| 249 | c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 250 | vst1q_f32(c2, vacc2x0123); |
| 251 | vst1q_f32(c2 + 4, vacc2x4567); |
| 252 | c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 253 | vst1q_f32(c1, vacc1x0123); |
| 254 | vst1q_f32(c1 + 4, vacc1x4567); |
| 255 | c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 256 | vst1q_f32(c0, vacc0x0123); |
| 257 | vst1q_f32(c0 + 4, vacc0x4567); |
| 258 | c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 259 | |
| 260 | a5 = (const float*) ((uintptr_t) a5 - kc); |
| 261 | a4 = (const float*) ((uintptr_t) a4 - kc); |
| 262 | a3 = (const float*) ((uintptr_t) a3 - kc); |
| 263 | a2 = (const float*) ((uintptr_t) a2 - kc); |
| 264 | a1 = (const float*) ((uintptr_t) a1 - kc); |
| 265 | a0 = (const float*) ((uintptr_t) a0 - kc); |
| 266 | |
| 267 | nc -= 8; |
| 268 | |
| 269 | } else { |
| 270 | if (nc & 4) { |
| 271 | vst1q_f32(c5, vacc5x0123); c5 += 4; |
| 272 | vst1q_f32(c4, vacc4x0123); c4 += 4; |
| 273 | vst1q_f32(c3, vacc3x0123); c3 += 4; |
| 274 | vst1q_f32(c2, vacc2x0123); c2 += 4; |
| 275 | vst1q_f32(c1, vacc1x0123); c1 += 4; |
| 276 | vst1q_f32(c0, vacc0x0123); c0 += 4; |
| 277 | |
| 278 | vacc5x0123 = vacc5x4567; |
| 279 | vacc4x0123 = vacc4x4567; |
| 280 | vacc3x0123 = vacc3x4567; |
| 281 | vacc2x0123 = vacc2x4567; |
| 282 | vacc1x0123 = vacc1x4567; |
| 283 | vacc0x0123 = vacc0x4567; |
| 284 | } |
| 285 | float32x2_t vacc5x01 = vget_low_f32(vacc5x0123); |
| 286 | float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); |
| 287 | float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); |
| 288 | float32x2_t vacc2x01 = vget_low_f32(vacc2x0123); |
| 289 | float32x2_t vacc1x01 = vget_low_f32(vacc1x0123); |
| 290 | float32x2_t vacc0x01 = vget_low_f32(vacc0x0123); |
| 291 | if (nc & 2) { |
| 292 | vst1_f32(c5, vacc5x01); c5 += 2; |
| 293 | vst1_f32(c4, vacc4x01); c4 += 2; |
| 294 | vst1_f32(c3, vacc3x01); c3 += 2; |
| 295 | vst1_f32(c2, vacc2x01); c2 += 2; |
| 296 | vst1_f32(c1, vacc1x01); c1 += 2; |
| 297 | vst1_f32(c0, vacc0x01); c0 += 2; |
| 298 | |
| 299 | vacc5x01 = vget_high_f32(vacc5x0123); |
| 300 | vacc4x01 = vget_high_f32(vacc4x0123); |
| 301 | vacc3x01 = vget_high_f32(vacc3x0123); |
| 302 | vacc2x01 = vget_high_f32(vacc2x0123); |
| 303 | vacc1x01 = vget_high_f32(vacc1x0123); |
| 304 | vacc0x01 = vget_high_f32(vacc0x0123); |
| 305 | } |
| 306 | if (nc & 1) { |
| 307 | vst1_lane_f32(c5, vacc5x01, 0); |
| 308 | vst1_lane_f32(c4, vacc4x01, 0); |
| 309 | vst1_lane_f32(c3, vacc3x01, 0); |
| 310 | vst1_lane_f32(c2, vacc2x01, 0); |
| 311 | vst1_lane_f32(c1, vacc1x01, 0); |
| 312 | vst1_lane_f32(c0, vacc0x01, 0); |
| 313 | } |
| 314 | |
| 315 | nc = 0; |
| 316 | } |
| 317 | } while (nc != 0); |
| 318 | } |