File size: 12,764 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 |
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <assert.h>
#include <arm_neon.h>
#include <xnnpack/argmaxpool.h>
void xnn_f32_argmaxpool_ukernel_9p8x__neon_c4(
size_t output_pixels,
size_t pooling_elements,
size_t channels,
const float** input,
size_t input_offset,
float* accumulation_buffer,
uint32_t* index_buffer,
float* output,
uint32_t* index,
size_t input_increment,
size_t output_increment) XNN_OOB_READS
{
assert(output_pixels != 0);
assert(pooling_elements != 0);
assert(pooling_elements > 9);
assert(channels != 0);
do {
{
float* ab = accumulation_buffer;
uint32_t* ib = index_buffer;
const float* i0 = *input++;
const float* i1 = *input++;
const float* i2 = *input++;
const float* i3 = *input++;
const float* i4 = *input++;
const float* i5 = *input++;
const float* i6 = *input++;
const float* i7 = *input++;
const float* i8 = *input++;
i0 = (const float*) ((uintptr_t) i0 + input_offset);
i1 = (const float*) ((uintptr_t) i1 + input_offset);
i2 = (const float*) ((uintptr_t) i2 + input_offset);
i3 = (const float*) ((uintptr_t) i3 + input_offset);
i4 = (const float*) ((uintptr_t) i4 + input_offset);
i5 = (const float*) ((uintptr_t) i5 + input_offset);
i6 = (const float*) ((uintptr_t) i6 + input_offset);
i7 = (const float*) ((uintptr_t) i7 + input_offset);
i8 = (const float*) ((uintptr_t) i8 + input_offset);
for (size_t c = 0; c < channels; c += 4) {
const float32x4_t vi0 = vld1q_f32(i0); i0 += 4;
const float32x4_t vi1 = vld1q_f32(i1); i1 += 4;
const float32x4_t vi2 = vld1q_f32(i2); i2 += 4;
const float32x4_t vi3 = vld1q_f32(i3); i3 += 4;
const float32x4_t vi4 = vld1q_f32(i4); i4 += 4;
const float32x4_t vi5 = vld1q_f32(i5); i5 += 4;
const float32x4_t vi6 = vld1q_f32(i6); i6 += 4;
const float32x4_t vi7 = vld1q_f32(i7); i7 += 4;
const float32x4_t vi8 = vld1q_f32(i8); i8 += 4;
float32x4_t vmax = vi0;
uint32x4_t vidx = vmovq_n_u32(0);
const uint32x4_t vm1 = vcgtq_f32(vi1, vmax);
vmax = vbslq_f32(vm1, vi1, vmax);
vidx = vbslq_u32(vm1, vmovq_n_u32(1), vidx);
const uint32x4_t vm2 = vcgtq_f32(vi2, vmax);
vmax = vbslq_f32(vm2, vi2, vmax);
vidx = vbslq_u32(vm2, vmovq_n_u32(2), vidx);
const uint32x4_t vm3 = vcgtq_f32(vi3, vmax);
vmax = vbslq_f32(vm3, vi3, vmax);
vidx = vbslq_u32(vm3, vmovq_n_u32(3), vidx);
const uint32x4_t vm4 = vcgtq_f32(vi4, vmax);
vmax = vbslq_f32(vm4, vi4, vmax);
vidx = vbslq_u32(vm4, vmovq_n_u32(4), vidx);
const uint32x4_t vm5 = vcgtq_f32(vi5, vmax);
vmax = vbslq_f32(vm5, vi5, vmax);
vidx = vbslq_u32(vm5, vmovq_n_u32(5), vidx);
const uint32x4_t vm6 = vcgtq_f32(vi6, vmax);
vmax = vbslq_f32(vm6, vi6, vmax);
vidx = vbslq_u32(vm6, vmovq_n_u32(6), vidx);
const uint32x4_t vm7 = vcgtq_f32(vi7, vmax);
vmax = vbslq_f32(vm7, vi7, vmax);
vidx = vbslq_u32(vm7, vmovq_n_u32(7), vidx);
const uint32x4_t vm8 = vcgtq_f32(vi8, vmax);
vmax = vbslq_f32(vm8, vi8, vmax);
vidx = vbslq_u32(vm8, vmovq_n_u32(8), vidx);
vst1q_f32(ab, vmax); ab += 4;
vst1q_u32(ib, vidx); ib += 4;
}
}
const uint32x4_t v1 = vmovq_n_u32(1);
const uint32x4_t v8 = vmovq_n_u32(8);
uint32x4_t vidx0 = vaddq_u32(v1, v8);
size_t k = pooling_elements;
for (k -= 9; k > 8; k -= 8) {
const float* i0 = *input++;
const float* i1 = *input++;
const float* i2 = *input++;
const float* i3 = *input++;
const float* i4 = *input++;
const float* i5 = *input++;
const float* i6 = *input++;
const float* i7 = *input++;
i0 = (const float*) ((uintptr_t) i0 + input_offset);
i1 = (const float*) ((uintptr_t) i1 + input_offset);
i2 = (const float*) ((uintptr_t) i2 + input_offset);
i3 = (const float*) ((uintptr_t) i3 + input_offset);
i4 = (const float*) ((uintptr_t) i4 + input_offset);
i5 = (const float*) ((uintptr_t) i5 + input_offset);
i6 = (const float*) ((uintptr_t) i6 + input_offset);
i7 = (const float*) ((uintptr_t) i7 + input_offset);
float* ab = accumulation_buffer;
uint32_t* ib = index_buffer;
for (size_t c = 0; c < channels; c += 4) {
const float32x4_t vi0 = vld1q_f32(i0); i0 += 4;
const float32x4_t vi1 = vld1q_f32(i1); i1 += 4;
const float32x4_t vi2 = vld1q_f32(i2); i2 += 4;
const float32x4_t vi3 = vld1q_f32(i3); i3 += 4;
const float32x4_t vi4 = vld1q_f32(i4); i4 += 4;
const float32x4_t vi5 = vld1q_f32(i5); i5 += 4;
const float32x4_t vi6 = vld1q_f32(i6); i6 += 4;
const float32x4_t vi7 = vld1q_f32(i7); i7 += 4;
float32x4_t vmax = vld1q_f32(ab);
uint32x4_t vidx = vld1q_u32(ib);
const uint32x4_t vm0 = vcgtq_f32(vi0, vmax);
vmax = vbslq_f32(vm0, vi0, vmax);
vidx = vbslq_u32(vm0, vidx0, vidx);
const uint32x4_t vm1 = vcgtq_f32(vi1, vmax);
const uint32x4_t vidx1 = vaddq_u32(vidx0, v1);
vmax = vbslq_f32(vm1, vi1, vmax);
vidx = vbslq_u32(vm1, vidx1, vidx);
const uint32x4_t vm2 = vcgtq_f32(vi2, vmax);
const uint32x4_t vidx2 = vaddq_u32(vidx1, v1);
vmax = vbslq_f32(vm2, vi2, vmax);
vidx = vbslq_u32(vm2, vidx2, vidx);
const uint32x4_t vm3 = vcgtq_f32(vi3, vmax);
const uint32x4_t vidx3 = vaddq_u32(vidx2, v1);
vmax = vbslq_f32(vm3, vi3, vmax);
vidx = vbslq_u32(vm3, vidx3, vidx);
const uint32x4_t vm4 = vcgtq_f32(vi4, vmax);
const uint32x4_t vidx4 = vaddq_u32(vidx3, v1);
vmax = vbslq_f32(vm4, vi4, vmax);
vidx = vbslq_u32(vm4, vidx4, vidx);
const uint32x4_t vm5 = vcgtq_f32(vi5, vmax);
const uint32x4_t vidx5 = vaddq_u32(vidx4, v1);
vmax = vbslq_f32(vm5, vi5, vmax);
vidx = vbslq_u32(vm5, vidx5, vidx);
const uint32x4_t vm6 = vcgtq_f32(vi6, vmax);
const uint32x4_t vidx6 = vaddq_u32(vidx5, v1);
vmax = vbslq_f32(vm6, vi6, vmax);
vidx = vbslq_u32(vm6, vidx6, vidx);
const uint32x4_t vm7 = vcgtq_f32(vi7, vmax);
const uint32x4_t vidx7 = vaddq_u32(vidx6, v1);
vmax = vbslq_f32(vm7, vi7, vmax);
vidx = vbslq_u32(vm7, vidx7, vidx);
vst1q_f32(ab, vmax); ab += 4;
vst1q_u32(ib, vidx); ib += 4;
}
vidx0 = vaddq_u32(vidx0, v8);
}
float* o = output;
uint32_t* i = index;
{
const float* i0 = input[0];
const float* i1 = input[1];
const float* i2 = input[2];
const float* i3 = input[3];
const float* i4 = input[4];
const float* i5 = input[5];
const float* i6 = input[6];
const float* i7 = input[7];
i0 = (const float*) ((uintptr_t) i0 + input_offset);
i1 = (const float*) ((uintptr_t) i1 + input_offset);
i2 = (const float*) ((uintptr_t) i2 + input_offset);
i3 = (const float*) ((uintptr_t) i3 + input_offset);
i4 = (const float*) ((uintptr_t) i4 + input_offset);
i5 = (const float*) ((uintptr_t) i5 + input_offset);
i6 = (const float*) ((uintptr_t) i6 + input_offset);
i7 = (const float*) ((uintptr_t) i7 + input_offset);
input = (const float**) ((uintptr_t) input + input_increment);
if (k < 2) {
i1 = i0;
}
if (k <= 2) {
i2 = i0;
}
if (k < 4) {
i3 = i0;
}
if (k <= 4) {
i4 = i0;
}
if (k < 6) {
i5 = i0;
}
if (k <= 6) {
i6 = i0;
}
if (k != 8) {
i7 = i0;
}
size_t c = channels;
float* ab = accumulation_buffer;
uint32_t* ib = index_buffer;
for (; c >= 4; c -= 4) {
const float32x4_t vi0 = vld1q_f32(i0); i0 += 4;
const float32x4_t vi1 = vld1q_f32(i1); i1 += 4;
const float32x4_t vi2 = vld1q_f32(i2); i2 += 4;
const float32x4_t vi3 = vld1q_f32(i3); i3 += 4;
const float32x4_t vi4 = vld1q_f32(i4); i4 += 4;
const float32x4_t vi5 = vld1q_f32(i5); i5 += 4;
const float32x4_t vi6 = vld1q_f32(i6); i6 += 4;
const float32x4_t vi7 = vld1q_f32(i7); i7 += 4;
float32x4_t vmax = vld1q_f32(ab); ab += 4;
uint32x4_t vidx = vld1q_u32(ib); ib += 4;
const uint32x4_t vm0 = vcgtq_f32(vi0, vmax);
vmax = vbslq_f32(vm0, vi0, vmax);
vidx = vbslq_u32(vm0, vidx0, vidx);
const uint32x4_t vm1 = vcgtq_f32(vi1, vmax);
const uint32x4_t vidx1 = vaddq_u32(vidx0, v1);
vmax = vbslq_f32(vm1, vi1, vmax);
vidx = vbslq_u32(vm1, vidx1, vidx);
const uint32x4_t vm2 = vcgtq_f32(vi2, vmax);
const uint32x4_t vidx2 = vaddq_u32(vidx1, v1);
vmax = vbslq_f32(vm2, vi2, vmax);
vidx = vbslq_u32(vm2, vidx2, vidx);
const uint32x4_t vm3 = vcgtq_f32(vi3, vmax);
const uint32x4_t vidx3 = vaddq_u32(vidx2, v1);
vmax = vbslq_f32(vm3, vi3, vmax);
vidx = vbslq_u32(vm3, vidx3, vidx);
const uint32x4_t vm4 = vcgtq_f32(vi4, vmax);
const uint32x4_t vidx4 = vaddq_u32(vidx3, v1);
vmax = vbslq_f32(vm4, vi4, vmax);
vidx = vbslq_u32(vm4, vidx4, vidx);
const uint32x4_t vm5 = vcgtq_f32(vi5, vmax);
const uint32x4_t vidx5 = vaddq_u32(vidx4, v1);
vmax = vbslq_f32(vm5, vi5, vmax);
vidx = vbslq_u32(vm5, vidx5, vidx);
const uint32x4_t vm6 = vcgtq_f32(vi6, vmax);
const uint32x4_t vidx6 = vaddq_u32(vidx5, v1);
vmax = vbslq_f32(vm6, vi6, vmax);
vidx = vbslq_u32(vm6, vidx6, vidx);
const uint32x4_t vm7 = vcgtq_f32(vi7, vmax);
const uint32x4_t vidx7 = vaddq_u32(vidx6, v1);
vmax = vbslq_f32(vm7, vi7, vmax);
vidx = vbslq_u32(vm7, vidx7, vidx);
vst1q_f32(o, vmax); o += 4;
vst1q_u32(i, vidx); i += 4;
}
if (c != 0) {
const float32x4_t vi0 = vld1q_f32(i0);
const float32x4_t vi1 = vld1q_f32(i1);
const float32x4_t vi2 = vld1q_f32(i2);
const float32x4_t vi3 = vld1q_f32(i3);
const float32x4_t vi4 = vld1q_f32(i4);
const float32x4_t vi5 = vld1q_f32(i5);
const float32x4_t vi6 = vld1q_f32(i6);
const float32x4_t vi7 = vld1q_f32(i7);
float32x4_t vmax = vld1q_f32(ab);
uint32x4_t vidx = vld1q_u32(ib);
const uint32x4_t vm0 = vcgtq_f32(vi0, vmax);
vmax = vbslq_f32(vm0, vi0, vmax);
vidx = vbslq_u32(vm0, vidx0, vidx);
const uint32x4_t vm1 = vcgtq_f32(vi1, vmax);
const uint32x4_t vidx1 = vaddq_u32(vidx0, v1);
vmax = vbslq_f32(vm1, vi1, vmax);
vidx = vbslq_u32(vm1, vidx1, vidx);
const uint32x4_t vm2 = vcgtq_f32(vi2, vmax);
const uint32x4_t vidx2 = vaddq_u32(vidx1, v1);
vmax = vbslq_f32(vm2, vi2, vmax);
vidx = vbslq_u32(vm2, vidx2, vidx);
const uint32x4_t vm3 = vcgtq_f32(vi3, vmax);
const uint32x4_t vidx3 = vaddq_u32(vidx2, v1);
vmax = vbslq_f32(vm3, vi3, vmax);
vidx = vbslq_u32(vm3, vidx3, vidx);
const uint32x4_t vm4 = vcgtq_f32(vi4, vmax);
const uint32x4_t vidx4 = vaddq_u32(vidx3, v1);
vmax = vbslq_f32(vm4, vi4, vmax);
vidx = vbslq_u32(vm4, vidx4, vidx);
const uint32x4_t vm5 = vcgtq_f32(vi5, vmax);
const uint32x4_t vidx5 = vaddq_u32(vidx4, v1);
vmax = vbslq_f32(vm5, vi5, vmax);
vidx = vbslq_u32(vm5, vidx5, vidx);
const uint32x4_t vm6 = vcgtq_f32(vi6, vmax);
const uint32x4_t vidx6 = vaddq_u32(vidx5, v1);
vmax = vbslq_f32(vm6, vi6, vmax);
vidx = vbslq_u32(vm6, vidx6, vidx);
const uint32x4_t vm7 = vcgtq_f32(vi7, vmax);
const uint32x4_t vidx7 = vaddq_u32(vidx6, v1);
vmax = vbslq_f32(vm7, vi7, vmax);
vidx = vbslq_u32(vm7, vidx7, vidx);
float32x2_t vmax_lo = vget_low_f32(vmax);
uint32x2_t vidx_lo = vget_low_u32(vidx);
if (c & 2) {
vst1_f32(o, vmax_lo); o += 2;
vst1_u32(i, vidx_lo); i += 2;
vmax_lo = vget_high_f32(vmax);
vidx_lo = vget_high_u32(vidx);
}
if (c & 1) {
vst1_lane_f32(o, vmax_lo, 0); o += 1;
vst1_lane_u32(i, vidx_lo, 0); i += 1;
}
}
}
output = (float*) ((uintptr_t) o + output_increment);
index = (uint32_t*) i;
} while (--output_pixels != 0);
}
|