| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/x32-packw/neon.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2023 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 |
|
| 11 |
#include <assert.h> |
| 12 |
#include <stddef.h> |
| 13 |
#include <stdint.h> |
| 14 |
|
| 15 |
#include <arm_neon.h> |
| 16 |
|
| 17 |
#include "xnnpack/packw.h" |
| 18 |
#include "xnnpack/prefetch.h" |
| 19 |
|
| 20 |
|
| 21 |
void xnn_x32_packw_gemm_goi_ukernel_x12__neon_ld4lane_u4_prfm( |
| 22 |
size_t g, |
| 23 |
size_t nc, |
| 24 |
size_t kc, |
| 25 |
size_t nr, |
| 26 |
size_t kr, |
| 27 |
size_t sr, |
| 28 |
const uint32_t* weights, |
| 29 |
const uint32_t* bias, |
| 30 |
const void* scale, |
| 31 |
uint32_t* packed_weights, |
| 32 |
size_t extra_bytes, |
| 33 |
const void* params) |
| 34 |
{ |
| 35 |
assert(g != 0); |
| 36 |
assert(nc != 0); |
| 37 |
assert(kc != 0); |
| 38 |
assert(nr == 12); |
| 39 |
assert(kr == 1); |
| 40 |
assert(sr == 1); |
| 41 |
assert(weights != NULL); |
| 42 |
assert(packed_weights != NULL); |
| 43 |
uint32x4x4_t vtmp0123x0123; |
| 44 |
vtmp0123x0123.val[0] = vdupq_n_u32(0); |
| 45 |
vtmp0123x0123.val[1] = vdupq_n_u32(0); |
| 46 |
vtmp0123x0123.val[2] = vdupq_n_u32(0); |
| 47 |
vtmp0123x0123.val[3] = vdupq_n_u32(0); |
| 48 |
uint32x4x4_t vtmp0123x4567; |
| 49 |
vtmp0123x4567.val[0] = vdupq_n_u32(0); |
| 50 |
vtmp0123x4567.val[1] = vdupq_n_u32(0); |
| 51 |
vtmp0123x4567.val[2] = vdupq_n_u32(0); |
| 52 |
vtmp0123x4567.val[3] = vdupq_n_u32(0); |
| 53 |
uint32x4x4_t vtmp0123x89AB; |
| 54 |
vtmp0123x89AB.val[0] = vdupq_n_u32(0); |
| 55 |
vtmp0123x89AB.val[1] = vdupq_n_u32(0); |
| 56 |
vtmp0123x89AB.val[2] = vdupq_n_u32(0); |
| 57 |
vtmp0123x89AB.val[3] = vdupq_n_u32(0); |
| 58 |
|
| 59 |
do { |
| 60 |
// NC main loop multiple of 12 |
| 61 |
const uint32_t* w0 = weights; |
| 62 |
size_t n = nc; |
| 63 |
|
| 64 |
for (; n >= 12; n -= 12) { |
| 65 |
if XNN_LIKELY(bias != NULL) { |
| 66 |
uint32x4_t vb0 = vld1q_u32(bias); bias += 4; |
| 67 |
uint32x4_t vb4 = vld1q_u32(bias); bias += 4; |
| 68 |
uint32x4_t vb8 = vld1q_u32(bias); bias += 4; |
| 69 |
vst1q_u32(packed_weights, vb0); packed_weights += 4; |
| 70 |
vst1q_u32(packed_weights, vb4); packed_weights += 4; |
| 71 |
vst1q_u32(packed_weights, vb8); packed_weights += 4; |
| 72 |
} else { |
| 73 |
const uint32x4_t vzero = vmovq_n_u32(0); |
| 74 |
vst1q_u32(packed_weights, vzero); packed_weights += 4; |
| 75 |
vst1q_u32(packed_weights, vzero); packed_weights += 4; |
| 76 |
vst1q_u32(packed_weights, vzero); packed_weights += 4; |
| 77 |
} |
| 78 |
|
| 79 |
const uint32_t* w1 = w0 + kc; |
| 80 |
const uint32_t* w2 = w1 + kc; |
| 81 |
const uint32_t* w3 = w2 + kc; |
| 82 |
const uint32_t* w4 = w3 + kc; |
| 83 |
const uint32_t* w5 = w4 + kc; |
| 84 |
const uint32_t* w6 = w5 + kc; |
| 85 |
const uint32_t* w7 = w6 + kc; |
| 86 |
const uint32_t* w8 = w7 + kc; |
| 87 |
const uint32_t* w9 = w8 + kc; |
| 88 |
const uint32_t* w10 = w9 + kc; |
| 89 |
const uint32_t* w11 = w10 + kc; |
| 90 |
xnn_prefetch_to_l1((const int8_t*) w0); |
| 91 |
xnn_prefetch_to_l1((const int8_t*) w0 + 64); |
| 92 |
xnn_prefetch_to_l1((const int8_t*) w1); |
| 93 |
xnn_prefetch_to_l1((const int8_t*) w1 + 64); |
| 94 |
xnn_prefetch_to_l1((const int8_t*) w2); |
| 95 |
xnn_prefetch_to_l1((const int8_t*) w2 + 64); |
| 96 |
xnn_prefetch_to_l1((const int8_t*) w3); |
| 97 |
xnn_prefetch_to_l1((const int8_t*) w3 + 64); |
| 98 |
xnn_prefetch_to_l1((const int8_t*) w4); |
| 99 |
xnn_prefetch_to_l1((const int8_t*) w4 + 64); |
| 100 |
xnn_prefetch_to_l1((const int8_t*) w5); |
| 101 |
xnn_prefetch_to_l1((const int8_t*) w5 + 64); |
| 102 |
xnn_prefetch_to_l1((const int8_t*) w6); |
| 103 |
xnn_prefetch_to_l1((const int8_t*) w6 + 64); |
| 104 |
xnn_prefetch_to_l1((const int8_t*) w7); |
| 105 |
xnn_prefetch_to_l1((const int8_t*) w7 + 64); |
| 106 |
xnn_prefetch_to_l1((const int8_t*) w8); |
| 107 |
xnn_prefetch_to_l1((const int8_t*) w8 + 64); |
| 108 |
xnn_prefetch_to_l1((const int8_t*) w9); |
| 109 |
xnn_prefetch_to_l1((const int8_t*) w9 + 64); |
| 110 |
xnn_prefetch_to_l1((const int8_t*) w10); |
| 111 |
xnn_prefetch_to_l1((const int8_t*) w10 + 64); |
| 112 |
xnn_prefetch_to_l1((const int8_t*) w11); |
| 113 |
xnn_prefetch_to_l1((const int8_t*) w11 + 64); |
| 114 |
|
| 115 |
// KC main loop multiple of 4 |
| 116 |
size_t k = kc; |
| 117 |
for (; k >= 4; k -= 4) { |
| 118 |
vtmp0123x0123 = vld4q_lane_u32(w0, vtmp0123x0123, 0); w0 += 4; |
| 119 |
vtmp0123x0123 = vld4q_lane_u32(w1, vtmp0123x0123, 1); w1 += 4; |
| 120 |
vtmp0123x0123 = vld4q_lane_u32(w2, vtmp0123x0123, 2); w2 += 4; |
| 121 |
vtmp0123x0123 = vld4q_lane_u32(w3, vtmp0123x0123, 3); w3 += 4; |
| 122 |
vtmp0123x4567 = vld4q_lane_u32(w4, vtmp0123x4567, 0); w4 += 4; |
| 123 |
vtmp0123x4567 = vld4q_lane_u32(w5, vtmp0123x4567, 1); w5 += 4; |
| 124 |
vtmp0123x4567 = vld4q_lane_u32(w6, vtmp0123x4567, 2); w6 += 4; |
| 125 |
vtmp0123x4567 = vld4q_lane_u32(w7, vtmp0123x4567, 3); w7 += 4; |
| 126 |
vtmp0123x89AB = vld4q_lane_u32(w8, vtmp0123x89AB, 0); w8 += 4; |
| 127 |
vtmp0123x89AB = vld4q_lane_u32(w9, vtmp0123x89AB, 1); w9 += 4; |
| 128 |
vtmp0123x89AB = vld4q_lane_u32(w10, vtmp0123x89AB, 2); w10 += 4; |
| 129 |
vtmp0123x89AB = vld4q_lane_u32(w11, vtmp0123x89AB, 3); w11 += 4; |
| 130 |
xnn_prefetch_to_l1((const int8_t*) w0 + 128); |
| 131 |
xnn_prefetch_to_l1((const int8_t*) w1 + 128); |
| 132 |
xnn_prefetch_to_l1((const int8_t*) w2 + 128); |
| 133 |
xnn_prefetch_to_l1((const int8_t*) w3 + 128); |
| 134 |
xnn_prefetch_to_l1((const int8_t*) w4 + 128); |
| 135 |
xnn_prefetch_to_l1((const int8_t*) w5 + 128); |
| 136 |
xnn_prefetch_to_l1((const int8_t*) w6 + 128); |
| 137 |
xnn_prefetch_to_l1((const int8_t*) w7 + 128); |
| 138 |
xnn_prefetch_to_l1((const int8_t*) w8 + 128); |
| 139 |
xnn_prefetch_to_l1((const int8_t*) w9 + 128); |
| 140 |
xnn_prefetch_to_l1((const int8_t*) w10 + 128); |
| 141 |
xnn_prefetch_to_l1((const int8_t*) w11 + 128); |
| 142 |
vst1q_u32(packed_weights, vtmp0123x0123.val[0]); packed_weights += 4; |
| 143 |
vst1q_u32(packed_weights, vtmp0123x4567.val[0]); packed_weights += 4; |
| 144 |
vst1q_u32(packed_weights, vtmp0123x89AB.val[0]); packed_weights += 4; |
| 145 |
vst1q_u32(packed_weights, vtmp0123x0123.val[1]); packed_weights += 4; |
| 146 |
vst1q_u32(packed_weights, vtmp0123x4567.val[1]); packed_weights += 4; |
| 147 |
vst1q_u32(packed_weights, vtmp0123x89AB.val[1]); packed_weights += 4; |
| 148 |
vst1q_u32(packed_weights, vtmp0123x0123.val[2]); packed_weights += 4; |
| 149 |
vst1q_u32(packed_weights, vtmp0123x4567.val[2]); packed_weights += 4; |
| 150 |
vst1q_u32(packed_weights, vtmp0123x89AB.val[2]); packed_weights += 4; |
| 151 |
vst1q_u32(packed_weights, vtmp0123x0123.val[3]); packed_weights += 4; |
| 152 |
vst1q_u32(packed_weights, vtmp0123x4567.val[3]); packed_weights += 4; |
| 153 |
vst1q_u32(packed_weights, vtmp0123x89AB.val[3]); packed_weights += 4; |
| 154 |
} |
| 155 |
|
| 156 |
// KC remainder of 1..3 |
| 157 |
// Same as main loop but ld1, ld2 or ld3 |
| 158 |
if XNN_UNLIKELY(k != 0) { |
| 159 |
assert(k >= 1); |
| 160 |
assert(k <= 3); |
| 161 |
switch (k) { |
| 162 |
// KC remainder of 1 |
| 163 |
case 1: |
| 164 |
{ |
| 165 |
uint32x4_t vtmp0x0123 = vdupq_n_u32(0); |
| 166 |
uint32x4_t vtmp0x4567 = vdupq_n_u32(0); |
| 167 |
uint32x4_t vtmp0x89AB = vdupq_n_u32(0); |
| 168 |
vtmp0x0123 = vld1q_lane_u32(w0, vtmp0x0123, 0); w0 += 1; |
| 169 |
vtmp0x0123 = vld1q_lane_u32(w1, vtmp0x0123, 1); w1 += 1; |
| 170 |
vtmp0x0123 = vld1q_lane_u32(w2, vtmp0x0123, 2); w2 += 1; |
| 171 |
vtmp0x0123 = vld1q_lane_u32(w3, vtmp0x0123, 3); w3 += 1; |
| 172 |
vtmp0x4567 = vld1q_lane_u32(w4, vtmp0x4567, 0); w4 += 1; |
| 173 |
vtmp0x4567 = vld1q_lane_u32(w5, vtmp0x4567, 1); w5 += 1; |
| 174 |
vtmp0x4567 = vld1q_lane_u32(w6, vtmp0x4567, 2); w6 += 1; |
| 175 |
vtmp0x4567 = vld1q_lane_u32(w7, vtmp0x4567, 3); w7 += 1; |
| 176 |
vtmp0x89AB = vld1q_lane_u32(w8, vtmp0x89AB, 0); w8 += 1; |
| 177 |
vtmp0x89AB = vld1q_lane_u32(w9, vtmp0x89AB, 1); w9 += 1; |
| 178 |
vtmp0x89AB = vld1q_lane_u32(w10, vtmp0x89AB, 2); w10 += 1; |
| 179 |
vtmp0x89AB = vld1q_lane_u32(w11, vtmp0x89AB, 3); w11 += 1; |
| 180 |
vst1q_u32(packed_weights, vtmp0x0123); packed_weights += 4; |
| 181 |
vst1q_u32(packed_weights, vtmp0x4567); packed_weights += 4; |
| 182 |
vst1q_u32(packed_weights, vtmp0x89AB); packed_weights += 4; |
| 183 |
break; |
| 184 |
} |
| 185 |
// KC remainder of 2 |
| 186 |
case 2: |
| 187 |
{ |
| 188 |
uint32x4x2_t vtmp01x0123; |
| 189 |
vtmp01x0123.val[0] = vdupq_n_u32(0); |
| 190 |
vtmp01x0123.val[1] = vdupq_n_u32(0); |
| 191 |
uint32x4x2_t vtmp01x4567; |
| 192 |
vtmp01x4567.val[0] = vdupq_n_u32(0); |
| 193 |
vtmp01x4567.val[1] = vdupq_n_u32(0); |
| 194 |
uint32x4x2_t vtmp01x89AB; |
| 195 |
vtmp01x89AB.val[0] = vdupq_n_u32(0); |
| 196 |
vtmp01x89AB.val[1] = vdupq_n_u32(0); |
| 197 |
vtmp01x0123 = vld2q_lane_u32(w0, vtmp01x0123, 0); w0 += 2; |
| 198 |
vtmp01x0123 = vld2q_lane_u32(w1, vtmp01x0123, 1); w1 += 2; |
| 199 |
vtmp01x0123 = vld2q_lane_u32(w2, vtmp01x0123, 2); w2 += 2; |
| 200 |
vtmp01x0123 = vld2q_lane_u32(w3, vtmp01x0123, 3); w3 += 2; |
| 201 |
vtmp01x4567 = vld2q_lane_u32(w4, vtmp01x4567, 0); w4 += 2; |
| 202 |
vtmp01x4567 = vld2q_lane_u32(w5, vtmp01x4567, 1); w5 += 2; |
| 203 |
vtmp01x4567 = vld2q_lane_u32(w6, vtmp01x4567, 2); w6 += 2; |
| 204 |
vtmp01x4567 = vld2q_lane_u32(w7, vtmp01x4567, 3); w7 += 2; |
| 205 |
vtmp01x89AB = vld2q_lane_u32(w8, vtmp01x89AB, 0); w8 += 2; |
| 206 |
vtmp01x89AB = vld2q_lane_u32(w9, vtmp01x89AB, 1); w9 += 2; |
| 207 |
vtmp01x89AB = vld2q_lane_u32(w10, vtmp01x89AB, 2); w10 += 2; |
| 208 |
vtmp01x89AB = vld2q_lane_u32(w11, vtmp01x89AB, 3); w11 += 2; |
| 209 |
vst1q_u32(packed_weights, vtmp01x0123.val[0]); packed_weights += 4; |
| 210 |
vst1q_u32(packed_weights, vtmp01x4567.val[0]); packed_weights += 4; |
| 211 |
vst1q_u32(packed_weights, vtmp01x89AB.val[0]); packed_weights += 4; |
| 212 |
vst1q_u32(packed_weights, vtmp01x0123.val[1]); packed_weights += 4; |
| 213 |
vst1q_u32(packed_weights, vtmp01x4567.val[1]); packed_weights += 4; |
| 214 |
vst1q_u32(packed_weights, vtmp01x89AB.val[1]); packed_weights += 4; |
| 215 |
break; |
| 216 |
} |
| 217 |
// KC remainder of 3 |
| 218 |
case 3: |
| 219 |
{ |
| 220 |
uint32x4x3_t vtmp012x0123; |
| 221 |
vtmp012x0123.val[0] = vdupq_n_u32(0); |
| 222 |
vtmp012x0123.val[1] = vdupq_n_u32(0); |
| 223 |
vtmp012x0123.val[2] = vdupq_n_u32(0); |
| 224 |
uint32x4x3_t vtmp012x4567; |
| 225 |
vtmp012x4567.val[0] = vdupq_n_u32(0); |
| 226 |
vtmp012x4567.val[1] = vdupq_n_u32(0); |
| 227 |
vtmp012x4567.val[2] = vdupq_n_u32(0); |
| 228 |
uint32x4x3_t vtmp012x89AB; |
| 229 |
vtmp012x89AB.val[0] = vdupq_n_u32(0); |
| 230 |
vtmp012x89AB.val[1] = vdupq_n_u32(0); |
| 231 |
vtmp012x89AB.val[2] = vdupq_n_u32(0); |
| 232 |
vtmp012x0123 = vld3q_lane_u32(w0, vtmp012x0123, 0); w0 += 3; |
| 233 |
vtmp012x0123 = vld3q_lane_u32(w1, vtmp012x0123, 1); w1 += 3; |
| 234 |
vtmp012x0123 = vld3q_lane_u32(w2, vtmp012x0123, 2); w2 += 3; |
| 235 |
vtmp012x0123 = vld3q_lane_u32(w3, vtmp012x0123, 3); w3 += 3; |
| 236 |
vtmp012x4567 = vld3q_lane_u32(w4, vtmp012x4567, 0); w4 += 3; |
| 237 |
vtmp012x4567 = vld3q_lane_u32(w5, vtmp012x4567, 1); w5 += 3; |
| 238 |
vtmp012x4567 = vld3q_lane_u32(w6, vtmp012x4567, 2); w6 += 3; |
| 239 |
vtmp012x4567 = vld3q_lane_u32(w7, vtmp012x4567, 3); w7 += 3; |
| 240 |
vtmp012x89AB = vld3q_lane_u32(w8, vtmp012x89AB, 0); w8 += 3; |
| 241 |
vtmp012x89AB = vld3q_lane_u32(w9, vtmp012x89AB, 1); w9 += 3; |
| 242 |
vtmp012x89AB = vld3q_lane_u32(w10, vtmp012x89AB, 2); w10 += 3; |
| 243 |
vtmp012x89AB = vld3q_lane_u32(w11, vtmp012x89AB, 3); w11 += 3; |
| 244 |
vst1q_u32(packed_weights, vtmp012x0123.val[0]); packed_weights += 4; |
| 245 |
vst1q_u32(packed_weights, vtmp012x4567.val[0]); packed_weights += 4; |
| 246 |
vst1q_u32(packed_weights, vtmp012x89AB.val[0]); packed_weights += 4; |
| 247 |
vst1q_u32(packed_weights, vtmp012x0123.val[1]); packed_weights += 4; |
| 248 |
vst1q_u32(packed_weights, vtmp012x4567.val[1]); packed_weights += 4; |
| 249 |
vst1q_u32(packed_weights, vtmp012x89AB.val[1]); packed_weights += 4; |
| 250 |
vst1q_u32(packed_weights, vtmp012x0123.val[2]); packed_weights += 4; |
| 251 |
vst1q_u32(packed_weights, vtmp012x4567.val[2]); packed_weights += 4; |
| 252 |
vst1q_u32(packed_weights, vtmp012x89AB.val[2]); packed_weights += 4; |
| 253 |
break; |
| 254 |
} |
| 255 |
default: |
| 256 |
XNN_UNREACHABLE; |
| 257 |
} |
| 258 |
} |
| 259 |
packed_weights = (uint32_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 260 |
w0 = w11; |
| 261 |
} |
| 262 |
|
| 263 |
// NC remainder (1..11) |
| 264 |
if XNN_UNLIKELY(n != 0) { |
| 265 |
assert(n >= 1); |
| 266 |
assert(n <= 11); |
| 267 |
if XNN_LIKELY(bias != NULL) { |
| 268 |
size_t nb = n; |
| 269 |
do { |
| 270 |
*packed_weights++ = *bias++; |
| 271 |
} while (--nb != 0); |
| 272 |
packed_weights += (12 - n); |
| 273 |
} else { |
| 274 |
const uint32x4_t vzero = vmovq_n_u32(0); |
| 275 |
vst1q_u32(packed_weights, vzero); packed_weights += 4; |
| 276 |
vst1q_u32(packed_weights, vzero); packed_weights += 4; |
| 277 |
vst1q_u32(packed_weights, vzero); packed_weights += 4; |
| 278 |
} |
| 279 |
|
| 280 |
// NR remainder has less than 12 rows so last row is not loaded |
| 281 |
const uint32_t* w1 = w0 + kc; |
| 282 |
if XNN_UNPREDICTABLE(n < 2) { |
| 283 |
w1 = w0; |
| 284 |
} |
| 285 |
const uint32_t* w2 = w1 + kc; |
| 286 |
if XNN_UNPREDICTABLE(n <= 2) { |
| 287 |
w2 = w1; |
| 288 |
} |
| 289 |
const uint32_t* w3 = w2 + kc; |
| 290 |
if XNN_UNPREDICTABLE(n < 4) { |
| 291 |
w3 = w2; |
| 292 |
} |
| 293 |
const uint32_t* w4 = w3 + kc; |
| 294 |
if XNN_UNPREDICTABLE(n <= 4) { |
| 295 |
w4 = w3; |
| 296 |
} |
| 297 |
const uint32_t* w5 = w4 + kc; |
| 298 |
if XNN_UNPREDICTABLE(n < 6) { |
| 299 |
w5 = w4; |
| 300 |
} |
| 301 |
const uint32_t* w6 = w5 + kc; |
| 302 |
if XNN_UNPREDICTABLE(n <= 6) { |
| 303 |
w6 = w5; |
| 304 |
} |
| 305 |
const uint32_t* w7 = w6 + kc; |
| 306 |
if XNN_UNPREDICTABLE(n < 8) { |
| 307 |
w7 = w6; |
| 308 |
} |
| 309 |
const uint32_t* w8 = w7 + kc; |
| 310 |
if XNN_UNPREDICTABLE(n <= 8) { |
| 311 |
w8 = w7; |
| 312 |
} |
| 313 |
const uint32_t* w9 = w8 + kc; |
| 314 |
if XNN_UNPREDICTABLE(n < 10) { |
| 315 |
w9 = w8; |
| 316 |
} |
| 317 |
const uint32_t* w10 = w9 + kc; |
| 318 |
if XNN_UNPREDICTABLE(n <= 10) { |
| 319 |
w10 = w9; |
| 320 |
} |
| 321 |
|
| 322 |
// KC main loop multiple of 4 |
| 323 |
size_t k = kc; |
| 324 |
for (; k >= 4; k -= 4) { |
| 325 |
vtmp0123x0123 = vld4q_lane_u32(w0, vtmp0123x0123, 0); w0 += 4; |
| 326 |
vtmp0123x0123 = vld4q_lane_u32(w1, vtmp0123x0123, 1); w1 += 4; |
| 327 |
vtmp0123x0123 = vld4q_lane_u32(w2, vtmp0123x0123, 2); w2 += 4; |
| 328 |
vtmp0123x0123 = vld4q_lane_u32(w3, vtmp0123x0123, 3); w3 += 4; |
| 329 |
vtmp0123x4567 = vld4q_lane_u32(w4, vtmp0123x4567, 0); w4 += 4; |
| 330 |
vtmp0123x4567 = vld4q_lane_u32(w5, vtmp0123x4567, 1); w5 += 4; |
| 331 |
vtmp0123x4567 = vld4q_lane_u32(w6, vtmp0123x4567, 2); w6 += 4; |
| 332 |
vtmp0123x4567 = vld4q_lane_u32(w7, vtmp0123x4567, 3); w7 += 4; |
| 333 |
vtmp0123x89AB = vld4q_lane_u32(w8, vtmp0123x89AB, 0); w8 += 4; |
| 334 |
vtmp0123x89AB = vld4q_lane_u32(w9, vtmp0123x89AB, 1); w9 += 4; |
| 335 |
vtmp0123x89AB = vld4q_lane_u32(w10, vtmp0123x89AB, 2); w10 += 4; |
| 336 |
xnn_prefetch_to_l1((const int8_t*) w0 + 128); |
| 337 |
xnn_prefetch_to_l1((const int8_t*) w1 + 128); |
| 338 |
xnn_prefetch_to_l1((const int8_t*) w2 + 128); |
| 339 |
xnn_prefetch_to_l1((const int8_t*) w3 + 128); |
| 340 |
xnn_prefetch_to_l1((const int8_t*) w4 + 128); |
| 341 |
xnn_prefetch_to_l1((const int8_t*) w5 + 128); |
| 342 |
xnn_prefetch_to_l1((const int8_t*) w6 + 128); |
| 343 |
xnn_prefetch_to_l1((const int8_t*) w7 + 128); |
| 344 |
xnn_prefetch_to_l1((const int8_t*) w8 + 128); |
| 345 |
xnn_prefetch_to_l1((const int8_t*) w9 + 128); |
| 346 |
xnn_prefetch_to_l1((const int8_t*) w10 + 128); |
| 347 |
vst1q_u32(packed_weights, vtmp0123x0123.val[0]); packed_weights += 4; |
| 348 |
vst1q_u32(packed_weights, vtmp0123x4567.val[0]); packed_weights += 4; |
| 349 |
vst1q_u32(packed_weights, vtmp0123x89AB.val[0]); packed_weights += 4; |
| 350 |
vst1q_u32(packed_weights, vtmp0123x0123.val[1]); packed_weights += 4; |
| 351 |
vst1q_u32(packed_weights, vtmp0123x4567.val[1]); packed_weights += 4; |
| 352 |
vst1q_u32(packed_weights, vtmp0123x89AB.val[1]); packed_weights += 4; |
| 353 |
vst1q_u32(packed_weights, vtmp0123x0123.val[2]); packed_weights += 4; |
| 354 |
vst1q_u32(packed_weights, vtmp0123x4567.val[2]); packed_weights += 4; |
| 355 |
vst1q_u32(packed_weights, vtmp0123x89AB.val[2]); packed_weights += 4; |
| 356 |
vst1q_u32(packed_weights, vtmp0123x0123.val[3]); packed_weights += 4; |
| 357 |
vst1q_u32(packed_weights, vtmp0123x4567.val[3]); packed_weights += 4; |
| 358 |
vst1q_u32(packed_weights, vtmp0123x89AB.val[3]); packed_weights += 4; |
| 359 |
} |
| 360 |
|
| 361 |
// KC remainder of 1..3 |
| 362 |
// Same as main loop but ld1, ld2 or ld3 |
| 363 |
if XNN_UNLIKELY(k != 0) { |
| 364 |
assert(k >= 1); |
| 365 |
assert(k <= 3); |
| 366 |
switch (k) { |
| 367 |
// KC remainder of 1 |
| 368 |
case 1: |
| 369 |
{ |
| 370 |
uint32x4_t vtmp0x0123 = vdupq_n_u32(0); |
| 371 |
uint32x4_t vtmp0x4567 = vdupq_n_u32(0); |
| 372 |
uint32x4_t vtmp0x89AB = vdupq_n_u32(0); |
| 373 |
vtmp0x0123 = vld1q_lane_u32(w0, vtmp0x0123, 0); |
| 374 |
vtmp0x0123 = vld1q_lane_u32(w1, vtmp0x0123, 1); |
| 375 |
vtmp0x0123 = vld1q_lane_u32(w2, vtmp0x0123, 2); |
| 376 |
vtmp0x0123 = vld1q_lane_u32(w3, vtmp0x0123, 3); |
| 377 |
vtmp0x4567 = vld1q_lane_u32(w4, vtmp0x4567, 0); |
| 378 |
vtmp0x4567 = vld1q_lane_u32(w5, vtmp0x4567, 1); |
| 379 |
vtmp0x4567 = vld1q_lane_u32(w6, vtmp0x4567, 2); |
| 380 |
vtmp0x4567 = vld1q_lane_u32(w7, vtmp0x4567, 3); |
| 381 |
vtmp0x89AB = vld1q_lane_u32(w8, vtmp0x89AB, 0); |
| 382 |
vtmp0x89AB = vld1q_lane_u32(w9, vtmp0x89AB, 1); |
| 383 |
vtmp0x89AB = vld1q_lane_u32(w10, vtmp0x89AB, 2); |
| 384 |
vst1q_u32(packed_weights, vtmp0x0123); packed_weights += 4; |
| 385 |
vst1q_u32(packed_weights, vtmp0x4567); packed_weights += 4; |
| 386 |
vst1q_u32(packed_weights, vtmp0x89AB); packed_weights += 4; |
| 387 |
break; |
| 388 |
} |
| 389 |
// KC remainder of 2 |
| 390 |
case 2: |
| 391 |
{ |
| 392 |
uint32x4x2_t vtmp01x0123; |
| 393 |
vtmp01x0123.val[0] = vdupq_n_u32(0); |
| 394 |
vtmp01x0123.val[1] = vdupq_n_u32(0); |
| 395 |
uint32x4x2_t vtmp01x4567; |
| 396 |
vtmp01x4567.val[0] = vdupq_n_u32(0); |
| 397 |
vtmp01x4567.val[1] = vdupq_n_u32(0); |
| 398 |
uint32x4x2_t vtmp01x89AB; |
| 399 |
vtmp01x89AB.val[0] = vdupq_n_u32(0); |
| 400 |
vtmp01x89AB.val[1] = vdupq_n_u32(0); |
| 401 |
vtmp01x0123 = vld2q_lane_u32(w0, vtmp01x0123, 0); |
| 402 |
vtmp01x0123 = vld2q_lane_u32(w1, vtmp01x0123, 1); |
| 403 |
vtmp01x0123 = vld2q_lane_u32(w2, vtmp01x0123, 2); |
| 404 |
vtmp01x0123 = vld2q_lane_u32(w3, vtmp01x0123, 3); |
| 405 |
vtmp01x4567 = vld2q_lane_u32(w4, vtmp01x4567, 0); |
| 406 |
vtmp01x4567 = vld2q_lane_u32(w5, vtmp01x4567, 1); |
| 407 |
vtmp01x4567 = vld2q_lane_u32(w6, vtmp01x4567, 2); |
| 408 |
vtmp01x4567 = vld2q_lane_u32(w7, vtmp01x4567, 3); |
| 409 |
vtmp01x89AB = vld2q_lane_u32(w8, vtmp01x89AB, 0); |
| 410 |
vtmp01x89AB = vld2q_lane_u32(w9, vtmp01x89AB, 1); |
| 411 |
vtmp01x89AB = vld2q_lane_u32(w10, vtmp01x89AB, 2); |
| 412 |
vst1q_u32(packed_weights, vtmp01x0123.val[0]); packed_weights += 4; |
| 413 |
vst1q_u32(packed_weights, vtmp01x4567.val[0]); packed_weights += 4; |
| 414 |
vst1q_u32(packed_weights, vtmp01x89AB.val[0]); packed_weights += 4; |
| 415 |
vst1q_u32(packed_weights, vtmp01x0123.val[1]); packed_weights += 4; |
| 416 |
vst1q_u32(packed_weights, vtmp01x4567.val[1]); packed_weights += 4; |
| 417 |
vst1q_u32(packed_weights, vtmp01x89AB.val[1]); packed_weights += 4; |
| 418 |
break; |
| 419 |
} |
| 420 |
// KC remainder of 3 |
| 421 |
case 3: |
| 422 |
{ |
| 423 |
uint32x4x3_t vtmp012x0123; |
| 424 |
vtmp012x0123.val[0] = vdupq_n_u32(0); |
| 425 |
vtmp012x0123.val[1] = vdupq_n_u32(0); |
| 426 |
vtmp012x0123.val[2] = vdupq_n_u32(0); |
| 427 |
uint32x4x3_t vtmp012x4567; |
| 428 |
vtmp012x4567.val[0] = vdupq_n_u32(0); |
| 429 |
vtmp012x4567.val[1] = vdupq_n_u32(0); |
| 430 |
vtmp012x4567.val[2] = vdupq_n_u32(0); |
| 431 |
uint32x4x3_t vtmp012x89AB; |
| 432 |
vtmp012x89AB.val[0] = vdupq_n_u32(0); |
| 433 |
vtmp012x89AB.val[1] = vdupq_n_u32(0); |
| 434 |
vtmp012x89AB.val[2] = vdupq_n_u32(0); |
| 435 |
vtmp012x0123 = vld3q_lane_u32(w0, vtmp012x0123, 0); |
| 436 |
vtmp012x0123 = vld3q_lane_u32(w1, vtmp012x0123, 1); |
| 437 |
vtmp012x0123 = vld3q_lane_u32(w2, vtmp012x0123, 2); |
| 438 |
vtmp012x0123 = vld3q_lane_u32(w3, vtmp012x0123, 3); |
| 439 |
vtmp012x4567 = vld3q_lane_u32(w4, vtmp012x4567, 0); |
| 440 |
vtmp012x4567 = vld3q_lane_u32(w5, vtmp012x4567, 1); |
| 441 |
vtmp012x4567 = vld3q_lane_u32(w6, vtmp012x4567, 2); |
| 442 |
vtmp012x4567 = vld3q_lane_u32(w7, vtmp012x4567, 3); |
| 443 |
vtmp012x89AB = vld3q_lane_u32(w8, vtmp012x89AB, 0); |
| 444 |
vtmp012x89AB = vld3q_lane_u32(w9, vtmp012x89AB, 1); |
| 445 |
vtmp012x89AB = vld3q_lane_u32(w10, vtmp012x89AB, 2); |
| 446 |
vst1q_u32(packed_weights, vtmp012x0123.val[0]); packed_weights += 4; |
| 447 |
vst1q_u32(packed_weights, vtmp012x4567.val[0]); packed_weights += 4; |
| 448 |
vst1q_u32(packed_weights, vtmp012x89AB.val[0]); packed_weights += 4; |
| 449 |
vst1q_u32(packed_weights, vtmp012x0123.val[1]); packed_weights += 4; |
| 450 |
vst1q_u32(packed_weights, vtmp012x4567.val[1]); packed_weights += 4; |
| 451 |
vst1q_u32(packed_weights, vtmp012x89AB.val[1]); packed_weights += 4; |
| 452 |
vst1q_u32(packed_weights, vtmp012x0123.val[2]); packed_weights += 4; |
| 453 |
vst1q_u32(packed_weights, vtmp012x4567.val[2]); packed_weights += 4; |
| 454 |
vst1q_u32(packed_weights, vtmp012x89AB.val[2]); packed_weights += 4; |
| 455 |
break; |
| 456 |
} |
| 457 |
default: |
| 458 |
XNN_UNREACHABLE; |
| 459 |
} |
| 460 |
} |
| 461 |
packed_weights = (uint32_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 462 |
} |
| 463 |
|
| 464 |
weights += nc * kc; |
| 465 |
} while (--g != 0); |
| 466 |
} |