| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-dwconv/multipass-scalar.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 |
#include <assert.h> |
| 11 |
#include <math.h> |
| 12 |
#include <stddef.h> |
| 13 |
#include <stdint.h> |
| 14 |
|
| 15 |
#include "xnnpack/common.h" |
| 16 |
#include "xnnpack/dwconv.h" |
| 17 |
#include "xnnpack/math.h" |
| 18 |
#include "xnnpack/microparams.h" |
| 19 |
#include "xnnpack/unaligned.h" |
| 20 |
|
| 21 |
|
| 22 |
void xnn_qs8_dwconv_minmax_fp32_ukernel_5f5m5l2c1s1r__scalar_lrintf( |
| 23 |
size_t channels, |
| 24 |
size_t output_width, |
| 25 |
const int8_t** input, |
| 26 |
const void* weights, |
| 27 |
int8_t* output, |
| 28 |
intptr_t input_stride, |
| 29 |
size_t output_increment, |
| 30 |
size_t input_offset, |
| 31 |
const int8_t* zero, |
| 32 |
size_t kernel_size, |
| 33 |
int32_t* buffer, |
| 34 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 35 |
{ |
| 36 |
assert(channels != 0); |
| 37 |
assert(output_width != 0); |
| 38 |
assert(kernel_size > 5); |
| 39 |
|
| 40 |
const float vscale = params->fp32_scalar.scale; |
| 41 |
const float voutput_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point; |
| 42 |
const float voutput_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point; |
| 43 |
const int32_t voutput_zero_point = params->fp32_scalar.output_zero_point; |
| 44 |
do { |
| 45 |
const void* w = weights; |
| 46 |
|
| 47 |
// First pass to process 5 inputs. |
| 48 |
{ |
| 49 |
int32_t* b = buffer; |
| 50 |
const int8_t* i0 = input[0]; |
| 51 |
assert(i0 != NULL); |
| 52 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 53 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 54 |
} |
| 55 |
const int8_t* i1 = input[1]; |
| 56 |
assert(i1 != NULL); |
| 57 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 58 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 59 |
} |
| 60 |
const int8_t* i2 = input[2]; |
| 61 |
assert(i2 != NULL); |
| 62 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 63 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 64 |
} |
| 65 |
const int8_t* i3 = input[3]; |
| 66 |
assert(i3 != NULL); |
| 67 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 68 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 69 |
} |
| 70 |
const int8_t* i4 = input[4]; |
| 71 |
assert(i4 != NULL); |
| 72 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 73 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 74 |
} |
| 75 |
input += 5; |
| 76 |
|
| 77 |
size_t c = channels; |
| 78 |
for (; c >= 2; c -= 2) { |
| 79 |
int32_t vacc0 = unaligned_indexed_load_s32(w, 0); |
| 80 |
int32_t vacc1 = unaligned_indexed_load_s32(w, 1); |
| 81 |
|
| 82 |
const int32_t vi0x0 = (int32_t) i0[0]; |
| 83 |
const int32_t vi0x1 = (int32_t) i0[1]; |
| 84 |
i0 += 2; |
| 85 |
|
| 86 |
const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0]; |
| 87 |
const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1]; |
| 88 |
|
| 89 |
vacc0 += vi0x0 * vk0x0; |
| 90 |
vacc1 += vi0x1 * vk0x1; |
| 91 |
|
| 92 |
const int32_t vi1x0 = (int32_t) i1[0]; |
| 93 |
const int32_t vi1x1 = (int32_t) i1[1]; |
| 94 |
i1 += 2; |
| 95 |
|
| 96 |
const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2]; |
| 97 |
const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3]; |
| 98 |
|
| 99 |
vacc0 += vi1x0 * vk1x0; |
| 100 |
vacc1 += vi1x1 * vk1x1; |
| 101 |
|
| 102 |
const int32_t vi2x0 = (int32_t) i2[0]; |
| 103 |
const int32_t vi2x1 = (int32_t) i2[1]; |
| 104 |
i2 += 2; |
| 105 |
|
| 106 |
const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4]; |
| 107 |
const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5]; |
| 108 |
|
| 109 |
vacc0 += vi2x0 * vk2x0; |
| 110 |
vacc1 += vi2x1 * vk2x1; |
| 111 |
|
| 112 |
const int32_t vi3x0 = (int32_t) i3[0]; |
| 113 |
const int32_t vi3x1 = (int32_t) i3[1]; |
| 114 |
i3 += 2; |
| 115 |
|
| 116 |
const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6]; |
| 117 |
const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7]; |
| 118 |
|
| 119 |
vacc0 += vi3x0 * vk3x0; |
| 120 |
vacc1 += vi3x1 * vk3x1; |
| 121 |
|
| 122 |
const int32_t vi4x0 = (int32_t) i4[0]; |
| 123 |
const int32_t vi4x1 = (int32_t) i4[1]; |
| 124 |
i4 += 2; |
| 125 |
|
| 126 |
const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8]; |
| 127 |
const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9]; |
| 128 |
|
| 129 |
vacc0 += vi4x0 * vk4x0; |
| 130 |
vacc1 += vi4x1 * vk4x1; |
| 131 |
|
| 132 |
w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 10 * sizeof(int8_t)); |
| 133 |
b[0] = vacc0; |
| 134 |
b[1] = vacc1; |
| 135 |
b += 2; |
| 136 |
} |
| 137 |
if XNN_UNLIKELY(c != 0) { |
| 138 |
int32_t vacc = unaligned_load_s32(w); |
| 139 |
|
| 140 |
const int32_t vi0 = (int32_t) *i0; |
| 141 |
const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[0]; |
| 142 |
vacc += vi0 * vk0; |
| 143 |
const int32_t vi1 = (int32_t) *i1; |
| 144 |
const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[1]; |
| 145 |
vacc += vi1 * vk1; |
| 146 |
const int32_t vi2 = (int32_t) *i2; |
| 147 |
const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[2]; |
| 148 |
vacc += vi2 * vk2; |
| 149 |
const int32_t vi3 = (int32_t) *i3; |
| 150 |
const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[3]; |
| 151 |
vacc += vi3 * vk3; |
| 152 |
const int32_t vi4 = (int32_t) *i4; |
| 153 |
const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[4]; |
| 154 |
vacc += vi4 * vk4; |
| 155 |
|
| 156 |
w = (const void*) ((uintptr_t) w + 1 * sizeof(int32_t) + 5 * sizeof(int8_t)); |
| 157 |
*b++ = vacc; |
| 158 |
} |
| 159 |
} |
| 160 |
|
| 161 |
// Middle pass to process 5 inputs in each iteration. |
| 162 |
for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) { |
| 163 |
int32_t* b = buffer; |
| 164 |
const int8_t* i0 = input[0]; |
| 165 |
assert(i0 != NULL); |
| 166 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 167 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 168 |
} |
| 169 |
const int8_t* i1 = input[1]; |
| 170 |
assert(i1 != NULL); |
| 171 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 172 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 173 |
} |
| 174 |
const int8_t* i2 = input[2]; |
| 175 |
assert(i2 != NULL); |
| 176 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 177 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 178 |
} |
| 179 |
const int8_t* i3 = input[3]; |
| 180 |
assert(i3 != NULL); |
| 181 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 182 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 183 |
} |
| 184 |
const int8_t* i4 = input[4]; |
| 185 |
assert(i4 != NULL); |
| 186 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 187 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 188 |
} |
| 189 |
input += 5; |
| 190 |
|
| 191 |
size_t c = channels; |
| 192 |
for (; c >= 2; c -= 2) { |
| 193 |
int32_t vacc0 = b[0]; |
| 194 |
int32_t vacc1 = b[1]; |
| 195 |
|
| 196 |
const int32_t vi0x0 = (int32_t) i0[0]; |
| 197 |
const int32_t vi0x1 = (int32_t) i0[1]; |
| 198 |
i0 += 2; |
| 199 |
|
| 200 |
const int32_t vk0x0 = (int32_t) ((const int8_t*) w)[0]; |
| 201 |
const int32_t vk0x1 = (int32_t) ((const int8_t*) w)[1]; |
| 202 |
|
| 203 |
vacc0 += vi0x0 * vk0x0; |
| 204 |
vacc1 += vi0x1 * vk0x1; |
| 205 |
|
| 206 |
const int32_t vi1x0 = (int32_t) i1[0]; |
| 207 |
const int32_t vi1x1 = (int32_t) i1[1]; |
| 208 |
i1 += 2; |
| 209 |
|
| 210 |
const int32_t vk1x0 = (int32_t) ((const int8_t*) w)[2]; |
| 211 |
const int32_t vk1x1 = (int32_t) ((const int8_t*) w)[3]; |
| 212 |
|
| 213 |
vacc0 += vi1x0 * vk1x0; |
| 214 |
vacc1 += vi1x1 * vk1x1; |
| 215 |
|
| 216 |
const int32_t vi2x0 = (int32_t) i2[0]; |
| 217 |
const int32_t vi2x1 = (int32_t) i2[1]; |
| 218 |
i2 += 2; |
| 219 |
|
| 220 |
const int32_t vk2x0 = (int32_t) ((const int8_t*) w)[4]; |
| 221 |
const int32_t vk2x1 = (int32_t) ((const int8_t*) w)[5]; |
| 222 |
|
| 223 |
vacc0 += vi2x0 * vk2x0; |
| 224 |
vacc1 += vi2x1 * vk2x1; |
| 225 |
|
| 226 |
const int32_t vi3x0 = (int32_t) i3[0]; |
| 227 |
const int32_t vi3x1 = (int32_t) i3[1]; |
| 228 |
i3 += 2; |
| 229 |
|
| 230 |
const int32_t vk3x0 = (int32_t) ((const int8_t*) w)[6]; |
| 231 |
const int32_t vk3x1 = (int32_t) ((const int8_t*) w)[7]; |
| 232 |
|
| 233 |
vacc0 += vi3x0 * vk3x0; |
| 234 |
vacc1 += vi3x1 * vk3x1; |
| 235 |
|
| 236 |
const int32_t vi4x0 = (int32_t) i4[0]; |
| 237 |
const int32_t vi4x1 = (int32_t) i4[1]; |
| 238 |
i4 += 2; |
| 239 |
|
| 240 |
const int32_t vk4x0 = (int32_t) ((const int8_t*) w)[8]; |
| 241 |
const int32_t vk4x1 = (int32_t) ((const int8_t*) w)[9]; |
| 242 |
|
| 243 |
vacc0 += vi4x0 * vk4x0; |
| 244 |
vacc1 += vi4x1 * vk4x1; |
| 245 |
|
| 246 |
w = (const void*) ((uintptr_t) w + 10 * sizeof(int8_t)); |
| 247 |
b[0] = vacc0; |
| 248 |
b[1] = vacc1; |
| 249 |
b += 2; |
| 250 |
} |
| 251 |
if XNN_UNLIKELY(c != 0) { |
| 252 |
int32_t vacc = b[0]; |
| 253 |
|
| 254 |
const int32_t vi0 = (int32_t) *i0; |
| 255 |
const int32_t vk0 = (int32_t) ((const int8_t*) w)[0]; |
| 256 |
vacc += vi0 * vk0; |
| 257 |
const int32_t vi1 = (int32_t) *i1; |
| 258 |
const int32_t vk1 = (int32_t) ((const int8_t*) w)[1]; |
| 259 |
vacc += vi1 * vk1; |
| 260 |
const int32_t vi2 = (int32_t) *i2; |
| 261 |
const int32_t vk2 = (int32_t) ((const int8_t*) w)[2]; |
| 262 |
vacc += vi2 * vk2; |
| 263 |
const int32_t vi3 = (int32_t) *i3; |
| 264 |
const int32_t vk3 = (int32_t) ((const int8_t*) w)[3]; |
| 265 |
vacc += vi3 * vk3; |
| 266 |
const int32_t vi4 = (int32_t) *i4; |
| 267 |
const int32_t vk4 = (int32_t) ((const int8_t*) w)[4]; |
| 268 |
vacc += vi4 * vk4; |
| 269 |
w = (const void*) ((uintptr_t) w + 5 * sizeof(int8_t)); |
| 270 |
*b++ = vacc; |
| 271 |
} |
| 272 |
} |
| 273 |
|
| 274 |
// Last pass to process up to 5 inputs. |
| 275 |
{ |
| 276 |
const int32_t* b = buffer; |
| 277 |
const int8_t* i0 = input[0]; |
| 278 |
assert(i0 != NULL); |
| 279 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 280 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 281 |
} |
| 282 |
const int8_t* i1 = input[1]; |
| 283 |
assert(i1 != NULL); |
| 284 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 285 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 286 |
} |
| 287 |
const int8_t* i2 = input[2]; |
| 288 |
assert(i2 != NULL); |
| 289 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 290 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 291 |
} |
| 292 |
const int8_t* i3 = input[3]; |
| 293 |
assert(i3 != NULL); |
| 294 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 295 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 296 |
} |
| 297 |
const int8_t* i4 = input[4]; |
| 298 |
assert(i4 != NULL); |
| 299 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 300 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 301 |
} |
| 302 |
|
| 303 |
size_t c = channels; |
| 304 |
for (; c >= 2; c -= 2) { |
| 305 |
int32_t vacc0 = b[0]; |
| 306 |
int32_t vacc1 = b[1]; |
| 307 |
b += 2; |
| 308 |
|
| 309 |
const int32_t vi0x0 = (int32_t) i0[0]; |
| 310 |
const int32_t vi0x1 = (int32_t) i0[1]; |
| 311 |
i0 += 2; |
| 312 |
|
| 313 |
const int32_t vk0x0 = (int32_t) ((const int8_t*) w)[0]; |
| 314 |
const int32_t vk0x1 = (int32_t) ((const int8_t*) w)[1]; |
| 315 |
|
| 316 |
vacc0 += vi0x0 * vk0x0; |
| 317 |
vacc1 += vi0x1 * vk0x1; |
| 318 |
|
| 319 |
const int32_t vi1x0 = (int32_t) i1[0]; |
| 320 |
const int32_t vi1x1 = (int32_t) i1[1]; |
| 321 |
i1 += 2; |
| 322 |
|
| 323 |
const int32_t vk1x0 = (int32_t) ((const int8_t*) w)[2]; |
| 324 |
const int32_t vk1x1 = (int32_t) ((const int8_t*) w)[3]; |
| 325 |
|
| 326 |
vacc0 += vi1x0 * vk1x0; |
| 327 |
vacc1 += vi1x1 * vk1x1; |
| 328 |
|
| 329 |
const int32_t vi2x0 = (int32_t) i2[0]; |
| 330 |
const int32_t vi2x1 = (int32_t) i2[1]; |
| 331 |
i2 += 2; |
| 332 |
|
| 333 |
const int32_t vk2x0 = (int32_t) ((const int8_t*) w)[4]; |
| 334 |
const int32_t vk2x1 = (int32_t) ((const int8_t*) w)[5]; |
| 335 |
|
| 336 |
vacc0 += vi2x0 * vk2x0; |
| 337 |
vacc1 += vi2x1 * vk2x1; |
| 338 |
|
| 339 |
const int32_t vi3x0 = (int32_t) i3[0]; |
| 340 |
const int32_t vi3x1 = (int32_t) i3[1]; |
| 341 |
i3 += 2; |
| 342 |
|
| 343 |
const int32_t vk3x0 = (int32_t) ((const int8_t*) w)[6]; |
| 344 |
const int32_t vk3x1 = (int32_t) ((const int8_t*) w)[7]; |
| 345 |
|
| 346 |
vacc0 += vi3x0 * vk3x0; |
| 347 |
vacc1 += vi3x1 * vk3x1; |
| 348 |
|
| 349 |
const int32_t vi4x0 = (int32_t) i4[0]; |
| 350 |
const int32_t vi4x1 = (int32_t) i4[1]; |
| 351 |
i4 += 2; |
| 352 |
|
| 353 |
const int32_t vk4x0 = (int32_t) ((const int8_t*) w)[8]; |
| 354 |
const int32_t vk4x1 = (int32_t) ((const int8_t*) w)[9]; |
| 355 |
|
| 356 |
vacc0 += vi4x0 * vk4x0; |
| 357 |
vacc1 += vi4x1 * vk4x1; |
| 358 |
|
| 359 |
w = (const void*) ((uintptr_t) w + 10 * sizeof(int8_t)); |
| 360 |
|
| 361 |
float vfpacc0 = (float) vacc0; |
| 362 |
float vfpacc1 = (float) vacc1; |
| 363 |
|
| 364 |
vfpacc0 *= vscale; |
| 365 |
vfpacc1 *= vscale; |
| 366 |
|
| 367 |
vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point); |
| 368 |
vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point); |
| 369 |
|
| 370 |
vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point); |
| 371 |
vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point); |
| 372 |
|
| 373 |
const int32_t vrndacc0 = (int32_t) lrintf(vfpacc0); |
| 374 |
const int32_t vrndacc1 = (int32_t) lrintf(vfpacc1); |
| 375 |
|
| 376 |
int32_t vout0 = (int32_t) vrndacc0 + voutput_zero_point; |
| 377 |
int32_t vout1 = (int32_t) vrndacc1 + voutput_zero_point; |
| 378 |
|
| 379 |
output[0] = (int8_t) vout0; |
| 380 |
output[1] = (int8_t) vout1; |
| 381 |
output += 2; |
| 382 |
} |
| 383 |
if XNN_UNLIKELY(c != 0) { |
| 384 |
int32_t vacc = b[0]; |
| 385 |
|
| 386 |
const int32_t vi0 = (int32_t) *i0; |
| 387 |
const int32_t vk0 = (int32_t) ((const int8_t*) w)[0]; |
| 388 |
vacc += vi0 * vk0; |
| 389 |
const int32_t vi1 = (int32_t) *i1; |
| 390 |
const int32_t vk1 = (int32_t) ((const int8_t*) w)[1]; |
| 391 |
vacc += vi1 * vk1; |
| 392 |
const int32_t vi2 = (int32_t) *i2; |
| 393 |
const int32_t vk2 = (int32_t) ((const int8_t*) w)[2]; |
| 394 |
vacc += vi2 * vk2; |
| 395 |
const int32_t vi3 = (int32_t) *i3; |
| 396 |
const int32_t vk3 = (int32_t) ((const int8_t*) w)[3]; |
| 397 |
vacc += vi3 * vk3; |
| 398 |
const int32_t vi4 = (int32_t) *i4; |
| 399 |
const int32_t vk4 = (int32_t) ((const int8_t*) w)[4]; |
| 400 |
vacc += vi4 * vk4; |
| 401 |
|
| 402 |
float vfpacc = (float) vacc * vscale; |
| 403 |
|
| 404 |
vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); |
| 405 |
vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); |
| 406 |
const int32_t vrndacc = (int32_t) lrintf(vfpacc); |
| 407 |
int32_t vout = vrndacc + voutput_zero_point; |
| 408 |
|
| 409 |
*output++ = (int8_t) vout; |
| 410 |
} |
| 411 |
} |
| 412 |
|
| 413 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 414 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 415 |
} while (--output_width != 0); |
| 416 |
} |