| 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 <stddef.h> |
| 12 |
#include <stdint.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/dwconv.h" |
| 16 |
#include "xnnpack/math.h" |
| 17 |
#include "xnnpack/microparams.h" |
| 18 |
#include "xnnpack/unaligned.h" |
| 19 |
|
| 20 |
|
| 21 |
void xnn_qs8_dwconv_minmax_fp32_ukernel_6f6m7l4c1s1r__scalar_imagic( |
| 22 |
size_t channels, |
| 23 |
size_t output_width, |
| 24 |
const int8_t** input, |
| 25 |
const void* weights, |
| 26 |
int8_t* output, |
| 27 |
intptr_t input_stride, |
| 28 |
size_t output_increment, |
| 29 |
size_t input_offset, |
| 30 |
const int8_t* zero, |
| 31 |
size_t kernel_size, |
| 32 |
int32_t* buffer, |
| 33 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 34 |
{ |
| 35 |
assert(channels != 0); |
| 36 |
assert(output_width != 0); |
| 37 |
assert(kernel_size > 6); |
| 38 |
|
| 39 |
const float vscale = params->fp32_scalar.scale; |
| 40 |
const float vmagic_bias = 12582912.0f; |
| 41 |
const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point; |
| 42 |
const int32_t output_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point; |
| 43 |
const int32_t vmagic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point); |
| 44 |
const int32_t vmagic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_point); |
| 45 |
const int32_t vmagic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point; |
| 46 |
do { |
| 47 |
const void* w = weights; |
| 48 |
|
| 49 |
// First pass to process 6 inputs. |
| 50 |
{ |
| 51 |
int32_t* b = buffer; |
| 52 |
const int8_t* i0 = input[0]; |
| 53 |
assert(i0 != NULL); |
| 54 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 55 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 56 |
} |
| 57 |
const int8_t* i1 = input[1]; |
| 58 |
assert(i1 != NULL); |
| 59 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 60 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 61 |
} |
| 62 |
const int8_t* i2 = input[2]; |
| 63 |
assert(i2 != NULL); |
| 64 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 65 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 66 |
} |
| 67 |
const int8_t* i3 = input[3]; |
| 68 |
assert(i3 != NULL); |
| 69 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 70 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 71 |
} |
| 72 |
const int8_t* i4 = input[4]; |
| 73 |
assert(i4 != NULL); |
| 74 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 75 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 76 |
} |
| 77 |
const int8_t* i5 = input[5]; |
| 78 |
assert(i5 != NULL); |
| 79 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 80 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 81 |
} |
| 82 |
input += 6; |
| 83 |
|
| 84 |
size_t c = channels; |
| 85 |
for (; c >= 4; c -= 4) { |
| 86 |
int32_t vacc0 = ((const int32_t*) w)[0]; |
| 87 |
int32_t vacc1 = ((const int32_t*) w)[1]; |
| 88 |
int32_t vacc2 = ((const int32_t*) w)[2]; |
| 89 |
int32_t vacc3 = ((const int32_t*) w)[3]; |
| 90 |
|
| 91 |
const int32_t vi0x0 = (int32_t) i0[0]; |
| 92 |
const int32_t vi0x1 = (int32_t) i0[1]; |
| 93 |
const int32_t vi0x2 = (int32_t) i0[2]; |
| 94 |
const int32_t vi0x3 = (int32_t) i0[3]; |
| 95 |
i0 += 4; |
| 96 |
|
| 97 |
const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0]; |
| 98 |
const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1]; |
| 99 |
const int32_t vk0x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2]; |
| 100 |
const int32_t vk0x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3]; |
| 101 |
|
| 102 |
vacc0 += vi0x0 * vk0x0; |
| 103 |
vacc1 += vi0x1 * vk0x1; |
| 104 |
vacc2 += vi0x2 * vk0x2; |
| 105 |
vacc3 += vi0x3 * vk0x3; |
| 106 |
|
| 107 |
const int32_t vi1x0 = (int32_t) i1[0]; |
| 108 |
const int32_t vi1x1 = (int32_t) i1[1]; |
| 109 |
const int32_t vi1x2 = (int32_t) i1[2]; |
| 110 |
const int32_t vi1x3 = (int32_t) i1[3]; |
| 111 |
i1 += 4; |
| 112 |
|
| 113 |
const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4]; |
| 114 |
const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5]; |
| 115 |
const int32_t vk1x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6]; |
| 116 |
const int32_t vk1x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7]; |
| 117 |
|
| 118 |
vacc0 += vi1x0 * vk1x0; |
| 119 |
vacc1 += vi1x1 * vk1x1; |
| 120 |
vacc2 += vi1x2 * vk1x2; |
| 121 |
vacc3 += vi1x3 * vk1x3; |
| 122 |
|
| 123 |
const int32_t vi2x0 = (int32_t) i2[0]; |
| 124 |
const int32_t vi2x1 = (int32_t) i2[1]; |
| 125 |
const int32_t vi2x2 = (int32_t) i2[2]; |
| 126 |
const int32_t vi2x3 = (int32_t) i2[3]; |
| 127 |
i2 += 4; |
| 128 |
|
| 129 |
const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8]; |
| 130 |
const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[9]; |
| 131 |
const int32_t vk2x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[10]; |
| 132 |
const int32_t vk2x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[11]; |
| 133 |
|
| 134 |
vacc0 += vi2x0 * vk2x0; |
| 135 |
vacc1 += vi2x1 * vk2x1; |
| 136 |
vacc2 += vi2x2 * vk2x2; |
| 137 |
vacc3 += vi2x3 * vk2x3; |
| 138 |
|
| 139 |
const int32_t vi3x0 = (int32_t) i3[0]; |
| 140 |
const int32_t vi3x1 = (int32_t) i3[1]; |
| 141 |
const int32_t vi3x2 = (int32_t) i3[2]; |
| 142 |
const int32_t vi3x3 = (int32_t) i3[3]; |
| 143 |
i3 += 4; |
| 144 |
|
| 145 |
const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[12]; |
| 146 |
const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[13]; |
| 147 |
const int32_t vk3x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[14]; |
| 148 |
const int32_t vk3x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[15]; |
| 149 |
|
| 150 |
vacc0 += vi3x0 * vk3x0; |
| 151 |
vacc1 += vi3x1 * vk3x1; |
| 152 |
vacc2 += vi3x2 * vk3x2; |
| 153 |
vacc3 += vi3x3 * vk3x3; |
| 154 |
|
| 155 |
const int32_t vi4x0 = (int32_t) i4[0]; |
| 156 |
const int32_t vi4x1 = (int32_t) i4[1]; |
| 157 |
const int32_t vi4x2 = (int32_t) i4[2]; |
| 158 |
const int32_t vi4x3 = (int32_t) i4[3]; |
| 159 |
i4 += 4; |
| 160 |
|
| 161 |
const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[16]; |
| 162 |
const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[17]; |
| 163 |
const int32_t vk4x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[18]; |
| 164 |
const int32_t vk4x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[19]; |
| 165 |
|
| 166 |
vacc0 += vi4x0 * vk4x0; |
| 167 |
vacc1 += vi4x1 * vk4x1; |
| 168 |
vacc2 += vi4x2 * vk4x2; |
| 169 |
vacc3 += vi4x3 * vk4x3; |
| 170 |
|
| 171 |
const int32_t vi5x0 = (int32_t) i5[0]; |
| 172 |
const int32_t vi5x1 = (int32_t) i5[1]; |
| 173 |
const int32_t vi5x2 = (int32_t) i5[2]; |
| 174 |
const int32_t vi5x3 = (int32_t) i5[3]; |
| 175 |
i5 += 4; |
| 176 |
|
| 177 |
const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[20]; |
| 178 |
const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[21]; |
| 179 |
const int32_t vk5x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[22]; |
| 180 |
const int32_t vk5x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[23]; |
| 181 |
|
| 182 |
vacc0 += vi5x0 * vk5x0; |
| 183 |
vacc1 += vi5x1 * vk5x1; |
| 184 |
vacc2 += vi5x2 * vk5x2; |
| 185 |
vacc3 += vi5x3 * vk5x3; |
| 186 |
|
| 187 |
w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 24 * sizeof(int8_t)); |
| 188 |
b[0] = vacc0; |
| 189 |
b[1] = vacc1; |
| 190 |
b[2] = vacc2; |
| 191 |
b[3] = vacc3; |
| 192 |
b += 4; |
| 193 |
} |
| 194 |
if XNN_UNLIKELY(c != 0) { |
| 195 |
do { |
| 196 |
int32_t vacc = unaligned_load_s32(w); |
| 197 |
const int32_t vi0 = (int32_t) *i0++; |
| 198 |
const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[0]; |
| 199 |
vacc += vi0 * vk0; |
| 200 |
const int32_t vi1 = (int32_t) *i1++; |
| 201 |
const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[1]; |
| 202 |
vacc += vi1 * vk1; |
| 203 |
const int32_t vi2 = (int32_t) *i2++; |
| 204 |
const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[2]; |
| 205 |
vacc += vi2 * vk2; |
| 206 |
const int32_t vi3 = (int32_t) *i3++; |
| 207 |
const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[3]; |
| 208 |
vacc += vi3 * vk3; |
| 209 |
const int32_t vi4 = (int32_t) *i4++; |
| 210 |
const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[4]; |
| 211 |
vacc += vi4 * vk4; |
| 212 |
const int32_t vi5 = (int32_t) *i5++; |
| 213 |
const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[5]; |
| 214 |
vacc += vi5 * vk5; |
| 215 |
|
| 216 |
w = (const void*) ((uintptr_t) w + 1 * sizeof(int32_t) + 6 * sizeof(int8_t)); |
| 217 |
*b++ = vacc; |
| 218 |
} while (--c != 0); |
| 219 |
} |
| 220 |
} |
| 221 |
|
| 222 |
// Middle pass to process 6 inputs in each iteration. |
| 223 |
for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) { |
| 224 |
int32_t* b = buffer; |
| 225 |
const int8_t* i0 = input[0]; |
| 226 |
assert(i0 != NULL); |
| 227 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 228 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 229 |
} |
| 230 |
const int8_t* i1 = input[1]; |
| 231 |
assert(i1 != NULL); |
| 232 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 233 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 234 |
} |
| 235 |
const int8_t* i2 = input[2]; |
| 236 |
assert(i2 != NULL); |
| 237 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 238 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 239 |
} |
| 240 |
const int8_t* i3 = input[3]; |
| 241 |
assert(i3 != NULL); |
| 242 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 243 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 244 |
} |
| 245 |
const int8_t* i4 = input[4]; |
| 246 |
assert(i4 != NULL); |
| 247 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 248 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 249 |
} |
| 250 |
const int8_t* i5 = input[5]; |
| 251 |
assert(i5 != NULL); |
| 252 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 253 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 254 |
} |
| 255 |
input += 6; |
| 256 |
|
| 257 |
size_t c = channels; |
| 258 |
for (; c >= 4; c -= 4) { |
| 259 |
int32_t vacc0 = b[0]; |
| 260 |
int32_t vacc1 = b[1]; |
| 261 |
int32_t vacc2 = b[2]; |
| 262 |
int32_t vacc3 = b[3]; |
| 263 |
|
| 264 |
const int32_t vi0x0 = (int32_t) i0[0]; |
| 265 |
const int32_t vi0x1 = (int32_t) i0[1]; |
| 266 |
const int32_t vi0x2 = (int32_t) i0[2]; |
| 267 |
const int32_t vi0x3 = (int32_t) i0[3]; |
| 268 |
i0 += 4; |
| 269 |
|
| 270 |
const int32_t vk0x0 = (int32_t) ((const int8_t*) w)[0]; |
| 271 |
const int32_t vk0x1 = (int32_t) ((const int8_t*) w)[1]; |
| 272 |
const int32_t vk0x2 = (int32_t) ((const int8_t*) w)[2]; |
| 273 |
const int32_t vk0x3 = (int32_t) ((const int8_t*) w)[3]; |
| 274 |
|
| 275 |
vacc0 += vi0x0 * vk0x0; |
| 276 |
vacc1 += vi0x1 * vk0x1; |
| 277 |
vacc2 += vi0x2 * vk0x2; |
| 278 |
vacc3 += vi0x3 * vk0x3; |
| 279 |
|
| 280 |
const int32_t vi1x0 = (int32_t) i1[0]; |
| 281 |
const int32_t vi1x1 = (int32_t) i1[1]; |
| 282 |
const int32_t vi1x2 = (int32_t) i1[2]; |
| 283 |
const int32_t vi1x3 = (int32_t) i1[3]; |
| 284 |
i1 += 4; |
| 285 |
|
| 286 |
const int32_t vk1x0 = (int32_t) ((const int8_t*) w)[4]; |
| 287 |
const int32_t vk1x1 = (int32_t) ((const int8_t*) w)[5]; |
| 288 |
const int32_t vk1x2 = (int32_t) ((const int8_t*) w)[6]; |
| 289 |
const int32_t vk1x3 = (int32_t) ((const int8_t*) w)[7]; |
| 290 |
|
| 291 |
vacc0 += vi1x0 * vk1x0; |
| 292 |
vacc1 += vi1x1 * vk1x1; |
| 293 |
vacc2 += vi1x2 * vk1x2; |
| 294 |
vacc3 += vi1x3 * vk1x3; |
| 295 |
|
| 296 |
const int32_t vi2x0 = (int32_t) i2[0]; |
| 297 |
const int32_t vi2x1 = (int32_t) i2[1]; |
| 298 |
const int32_t vi2x2 = (int32_t) i2[2]; |
| 299 |
const int32_t vi2x3 = (int32_t) i2[3]; |
| 300 |
i2 += 4; |
| 301 |
|
| 302 |
const int32_t vk2x0 = (int32_t) ((const int8_t*) w)[8]; |
| 303 |
const int32_t vk2x1 = (int32_t) ((const int8_t*) w)[9]; |
| 304 |
const int32_t vk2x2 = (int32_t) ((const int8_t*) w)[10]; |
| 305 |
const int32_t vk2x3 = (int32_t) ((const int8_t*) w)[11]; |
| 306 |
|
| 307 |
vacc0 += vi2x0 * vk2x0; |
| 308 |
vacc1 += vi2x1 * vk2x1; |
| 309 |
vacc2 += vi2x2 * vk2x2; |
| 310 |
vacc3 += vi2x3 * vk2x3; |
| 311 |
|
| 312 |
const int32_t vi3x0 = (int32_t) i3[0]; |
| 313 |
const int32_t vi3x1 = (int32_t) i3[1]; |
| 314 |
const int32_t vi3x2 = (int32_t) i3[2]; |
| 315 |
const int32_t vi3x3 = (int32_t) i3[3]; |
| 316 |
i3 += 4; |
| 317 |
|
| 318 |
const int32_t vk3x0 = (int32_t) ((const int8_t*) w)[12]; |
| 319 |
const int32_t vk3x1 = (int32_t) ((const int8_t*) w)[13]; |
| 320 |
const int32_t vk3x2 = (int32_t) ((const int8_t*) w)[14]; |
| 321 |
const int32_t vk3x3 = (int32_t) ((const int8_t*) w)[15]; |
| 322 |
|
| 323 |
vacc0 += vi3x0 * vk3x0; |
| 324 |
vacc1 += vi3x1 * vk3x1; |
| 325 |
vacc2 += vi3x2 * vk3x2; |
| 326 |
vacc3 += vi3x3 * vk3x3; |
| 327 |
|
| 328 |
const int32_t vi4x0 = (int32_t) i4[0]; |
| 329 |
const int32_t vi4x1 = (int32_t) i4[1]; |
| 330 |
const int32_t vi4x2 = (int32_t) i4[2]; |
| 331 |
const int32_t vi4x3 = (int32_t) i4[3]; |
| 332 |
i4 += 4; |
| 333 |
|
| 334 |
const int32_t vk4x0 = (int32_t) ((const int8_t*) w)[16]; |
| 335 |
const int32_t vk4x1 = (int32_t) ((const int8_t*) w)[17]; |
| 336 |
const int32_t vk4x2 = (int32_t) ((const int8_t*) w)[18]; |
| 337 |
const int32_t vk4x3 = (int32_t) ((const int8_t*) w)[19]; |
| 338 |
|
| 339 |
vacc0 += vi4x0 * vk4x0; |
| 340 |
vacc1 += vi4x1 * vk4x1; |
| 341 |
vacc2 += vi4x2 * vk4x2; |
| 342 |
vacc3 += vi4x3 * vk4x3; |
| 343 |
|
| 344 |
const int32_t vi5x0 = (int32_t) i5[0]; |
| 345 |
const int32_t vi5x1 = (int32_t) i5[1]; |
| 346 |
const int32_t vi5x2 = (int32_t) i5[2]; |
| 347 |
const int32_t vi5x3 = (int32_t) i5[3]; |
| 348 |
i5 += 4; |
| 349 |
|
| 350 |
const int32_t vk5x0 = (int32_t) ((const int8_t*) w)[20]; |
| 351 |
const int32_t vk5x1 = (int32_t) ((const int8_t*) w)[21]; |
| 352 |
const int32_t vk5x2 = (int32_t) ((const int8_t*) w)[22]; |
| 353 |
const int32_t vk5x3 = (int32_t) ((const int8_t*) w)[23]; |
| 354 |
|
| 355 |
vacc0 += vi5x0 * vk5x0; |
| 356 |
vacc1 += vi5x1 * vk5x1; |
| 357 |
vacc2 += vi5x2 * vk5x2; |
| 358 |
vacc3 += vi5x3 * vk5x3; |
| 359 |
|
| 360 |
w = (const void*) ((uintptr_t) w + 24 * sizeof(int8_t)); |
| 361 |
b[0] = vacc0; |
| 362 |
b[1] = vacc1; |
| 363 |
b[2] = vacc2; |
| 364 |
b[3] = vacc3; |
| 365 |
b += 4; |
| 366 |
} |
| 367 |
if XNN_UNLIKELY(c != 0) { |
| 368 |
do { |
| 369 |
int32_t vacc = *b; |
| 370 |
const int32_t vi0 = (int32_t) *i0++; |
| 371 |
const int32_t vk0 = (int32_t) ((const int8_t*) w)[0]; |
| 372 |
vacc += vi0 * vk0; |
| 373 |
const int32_t vi1 = (int32_t) *i1++; |
| 374 |
const int32_t vk1 = (int32_t) ((const int8_t*) w)[1]; |
| 375 |
vacc += vi1 * vk1; |
| 376 |
const int32_t vi2 = (int32_t) *i2++; |
| 377 |
const int32_t vk2 = (int32_t) ((const int8_t*) w)[2]; |
| 378 |
vacc += vi2 * vk2; |
| 379 |
const int32_t vi3 = (int32_t) *i3++; |
| 380 |
const int32_t vk3 = (int32_t) ((const int8_t*) w)[3]; |
| 381 |
vacc += vi3 * vk3; |
| 382 |
const int32_t vi4 = (int32_t) *i4++; |
| 383 |
const int32_t vk4 = (int32_t) ((const int8_t*) w)[4]; |
| 384 |
vacc += vi4 * vk4; |
| 385 |
const int32_t vi5 = (int32_t) *i5++; |
| 386 |
const int32_t vk5 = (int32_t) ((const int8_t*) w)[5]; |
| 387 |
vacc += vi5 * vk5; |
| 388 |
|
| 389 |
w = (const void*) ((uintptr_t) w + 6 * sizeof(int8_t)); |
| 390 |
*b++ = vacc; |
| 391 |
} while (--c != 0); |
| 392 |
} |
| 393 |
} |
| 394 |
|
| 395 |
// Last pass to process up to 7 inputs. |
| 396 |
{ |
| 397 |
const int32_t* b = buffer; |
| 398 |
const int8_t* i0 = input[0]; |
| 399 |
assert(i0 != NULL); |
| 400 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 401 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 402 |
} |
| 403 |
const int8_t* i1 = input[1]; |
| 404 |
assert(i1 != NULL); |
| 405 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 406 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 407 |
} |
| 408 |
const int8_t* i2 = input[2]; |
| 409 |
assert(i2 != NULL); |
| 410 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 411 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 412 |
} |
| 413 |
const int8_t* i3 = input[3]; |
| 414 |
assert(i3 != NULL); |
| 415 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 416 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 417 |
} |
| 418 |
const int8_t* i4 = input[4]; |
| 419 |
assert(i4 != NULL); |
| 420 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 421 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 422 |
} |
| 423 |
const int8_t* i5 = input[5]; |
| 424 |
assert(i5 != NULL); |
| 425 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 426 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 427 |
} |
| 428 |
const int8_t* i6 = input[6]; |
| 429 |
assert(i6 != NULL); |
| 430 |
if XNN_UNPREDICTABLE(i6 != zero) { |
| 431 |
i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| 432 |
} |
| 433 |
|
| 434 |
size_t c = channels; |
| 435 |
for (; c >= 4; c -= 4) { |
| 436 |
int32_t vacc0 = b[0]; |
| 437 |
int32_t vacc1 = b[1]; |
| 438 |
int32_t vacc2 = b[2]; |
| 439 |
int32_t vacc3 = b[3]; |
| 440 |
b += 4; |
| 441 |
|
| 442 |
const int32_t vi0x0 = (int32_t) i0[0]; |
| 443 |
const int32_t vi0x1 = (int32_t) i0[1]; |
| 444 |
const int32_t vi0x2 = (int32_t) i0[2]; |
| 445 |
const int32_t vi0x3 = (int32_t) i0[3]; |
| 446 |
i0 += 4; |
| 447 |
|
| 448 |
const int32_t vk0x0 = (int32_t) ((const int8_t*) w)[0]; |
| 449 |
const int32_t vk0x1 = (int32_t) ((const int8_t*) w)[1]; |
| 450 |
const int32_t vk0x2 = (int32_t) ((const int8_t*) w)[2]; |
| 451 |
const int32_t vk0x3 = (int32_t) ((const int8_t*) w)[3]; |
| 452 |
|
| 453 |
vacc0 += vi0x0 * vk0x0; |
| 454 |
vacc1 += vi0x1 * vk0x1; |
| 455 |
vacc2 += vi0x2 * vk0x2; |
| 456 |
vacc3 += vi0x3 * vk0x3; |
| 457 |
|
| 458 |
const int32_t vi1x0 = (int32_t) i1[0]; |
| 459 |
const int32_t vi1x1 = (int32_t) i1[1]; |
| 460 |
const int32_t vi1x2 = (int32_t) i1[2]; |
| 461 |
const int32_t vi1x3 = (int32_t) i1[3]; |
| 462 |
i1 += 4; |
| 463 |
|
| 464 |
const int32_t vk1x0 = (int32_t) ((const int8_t*) w)[4]; |
| 465 |
const int32_t vk1x1 = (int32_t) ((const int8_t*) w)[5]; |
| 466 |
const int32_t vk1x2 = (int32_t) ((const int8_t*) w)[6]; |
| 467 |
const int32_t vk1x3 = (int32_t) ((const int8_t*) w)[7]; |
| 468 |
|
| 469 |
vacc0 += vi1x0 * vk1x0; |
| 470 |
vacc1 += vi1x1 * vk1x1; |
| 471 |
vacc2 += vi1x2 * vk1x2; |
| 472 |
vacc3 += vi1x3 * vk1x3; |
| 473 |
|
| 474 |
const int32_t vi2x0 = (int32_t) i2[0]; |
| 475 |
const int32_t vi2x1 = (int32_t) i2[1]; |
| 476 |
const int32_t vi2x2 = (int32_t) i2[2]; |
| 477 |
const int32_t vi2x3 = (int32_t) i2[3]; |
| 478 |
i2 += 4; |
| 479 |
|
| 480 |
const int32_t vk2x0 = (int32_t) ((const int8_t*) w)[8]; |
| 481 |
const int32_t vk2x1 = (int32_t) ((const int8_t*) w)[9]; |
| 482 |
const int32_t vk2x2 = (int32_t) ((const int8_t*) w)[10]; |
| 483 |
const int32_t vk2x3 = (int32_t) ((const int8_t*) w)[11]; |
| 484 |
|
| 485 |
vacc0 += vi2x0 * vk2x0; |
| 486 |
vacc1 += vi2x1 * vk2x1; |
| 487 |
vacc2 += vi2x2 * vk2x2; |
| 488 |
vacc3 += vi2x3 * vk2x3; |
| 489 |
|
| 490 |
const int32_t vi3x0 = (int32_t) i3[0]; |
| 491 |
const int32_t vi3x1 = (int32_t) i3[1]; |
| 492 |
const int32_t vi3x2 = (int32_t) i3[2]; |
| 493 |
const int32_t vi3x3 = (int32_t) i3[3]; |
| 494 |
i3 += 4; |
| 495 |
|
| 496 |
const int32_t vk3x0 = (int32_t) ((const int8_t*) w)[12]; |
| 497 |
const int32_t vk3x1 = (int32_t) ((const int8_t*) w)[13]; |
| 498 |
const int32_t vk3x2 = (int32_t) ((const int8_t*) w)[14]; |
| 499 |
const int32_t vk3x3 = (int32_t) ((const int8_t*) w)[15]; |
| 500 |
|
| 501 |
vacc0 += vi3x0 * vk3x0; |
| 502 |
vacc1 += vi3x1 * vk3x1; |
| 503 |
vacc2 += vi3x2 * vk3x2; |
| 504 |
vacc3 += vi3x3 * vk3x3; |
| 505 |
|
| 506 |
const int32_t vi4x0 = (int32_t) i4[0]; |
| 507 |
const int32_t vi4x1 = (int32_t) i4[1]; |
| 508 |
const int32_t vi4x2 = (int32_t) i4[2]; |
| 509 |
const int32_t vi4x3 = (int32_t) i4[3]; |
| 510 |
i4 += 4; |
| 511 |
|
| 512 |
const int32_t vk4x0 = (int32_t) ((const int8_t*) w)[16]; |
| 513 |
const int32_t vk4x1 = (int32_t) ((const int8_t*) w)[17]; |
| 514 |
const int32_t vk4x2 = (int32_t) ((const int8_t*) w)[18]; |
| 515 |
const int32_t vk4x3 = (int32_t) ((const int8_t*) w)[19]; |
| 516 |
|
| 517 |
vacc0 += vi4x0 * vk4x0; |
| 518 |
vacc1 += vi4x1 * vk4x1; |
| 519 |
vacc2 += vi4x2 * vk4x2; |
| 520 |
vacc3 += vi4x3 * vk4x3; |
| 521 |
|
| 522 |
const int32_t vi5x0 = (int32_t) i5[0]; |
| 523 |
const int32_t vi5x1 = (int32_t) i5[1]; |
| 524 |
const int32_t vi5x2 = (int32_t) i5[2]; |
| 525 |
const int32_t vi5x3 = (int32_t) i5[3]; |
| 526 |
i5 += 4; |
| 527 |
|
| 528 |
const int32_t vk5x0 = (int32_t) ((const int8_t*) w)[20]; |
| 529 |
const int32_t vk5x1 = (int32_t) ((const int8_t*) w)[21]; |
| 530 |
const int32_t vk5x2 = (int32_t) ((const int8_t*) w)[22]; |
| 531 |
const int32_t vk5x3 = (int32_t) ((const int8_t*) w)[23]; |
| 532 |
|
| 533 |
vacc0 += vi5x0 * vk5x0; |
| 534 |
vacc1 += vi5x1 * vk5x1; |
| 535 |
vacc2 += vi5x2 * vk5x2; |
| 536 |
vacc3 += vi5x3 * vk5x3; |
| 537 |
|
| 538 |
const int32_t vi6x0 = (int32_t) i6[0]; |
| 539 |
const int32_t vi6x1 = (int32_t) i6[1]; |
| 540 |
const int32_t vi6x2 = (int32_t) i6[2]; |
| 541 |
const int32_t vi6x3 = (int32_t) i6[3]; |
| 542 |
i6 += 4; |
| 543 |
|
| 544 |
const int32_t vk6x0 = (int32_t) ((const int8_t*) w)[24]; |
| 545 |
const int32_t vk6x1 = (int32_t) ((const int8_t*) w)[25]; |
| 546 |
const int32_t vk6x2 = (int32_t) ((const int8_t*) w)[26]; |
| 547 |
const int32_t vk6x3 = (int32_t) ((const int8_t*) w)[27]; |
| 548 |
|
| 549 |
vacc0 += vi6x0 * vk6x0; |
| 550 |
vacc1 += vi6x1 * vk6x1; |
| 551 |
vacc2 += vi6x2 * vk6x2; |
| 552 |
vacc3 += vi6x3 * vk6x3; |
| 553 |
|
| 554 |
w = (const void*) ((uintptr_t) w + 28 * sizeof(int8_t)); |
| 555 |
|
| 556 |
float vfpacc0 = (float) vacc0; |
| 557 |
float vfpacc1 = (float) vacc1; |
| 558 |
float vfpacc2 = (float) vacc2; |
| 559 |
float vfpacc3 = (float) vacc3; |
| 560 |
|
| 561 |
vfpacc0 *= vscale; |
| 562 |
vfpacc1 *= vscale; |
| 563 |
vfpacc2 *= vscale; |
| 564 |
vfpacc3 *= vscale; |
| 565 |
|
| 566 |
vfpacc0 += vmagic_bias; |
| 567 |
vfpacc1 += vmagic_bias; |
| 568 |
vfpacc2 += vmagic_bias; |
| 569 |
vfpacc3 += vmagic_bias; |
| 570 |
|
| 571 |
int32_t vout0 = (int32_t) float_as_uint32(vfpacc0); |
| 572 |
int32_t vout1 = (int32_t) float_as_uint32(vfpacc1); |
| 573 |
int32_t vout2 = (int32_t) float_as_uint32(vfpacc2); |
| 574 |
int32_t vout3 = (int32_t) float_as_uint32(vfpacc3); |
| 575 |
|
| 576 |
vout0 = math_max_s32(vout0, vmagic_min); |
| 577 |
vout1 = math_max_s32(vout1, vmagic_min); |
| 578 |
vout2 = math_max_s32(vout2, vmagic_min); |
| 579 |
vout3 = math_max_s32(vout3, vmagic_min); |
| 580 |
|
| 581 |
vout0 = math_min_s32(vout0, vmagic_max); |
| 582 |
vout1 = math_min_s32(vout1, vmagic_max); |
| 583 |
vout2 = math_min_s32(vout2, vmagic_max); |
| 584 |
vout3 = math_min_s32(vout3, vmagic_max); |
| 585 |
|
| 586 |
vout0 -= vmagic_bias_less_zero_point; |
| 587 |
vout1 -= vmagic_bias_less_zero_point; |
| 588 |
vout2 -= vmagic_bias_less_zero_point; |
| 589 |
vout3 -= vmagic_bias_less_zero_point; |
| 590 |
|
| 591 |
output[0] = (int8_t) vout0; |
| 592 |
output[1] = (int8_t) vout1; |
| 593 |
output[2] = (int8_t) vout2; |
| 594 |
output[3] = (int8_t) vout3; |
| 595 |
output += 4; |
| 596 |
} |
| 597 |
if XNN_UNLIKELY(c != 0) { |
| 598 |
do { |
| 599 |
int32_t vacc = *b++; |
| 600 |
const int32_t vi0 = (int32_t) *i0++; |
| 601 |
const int32_t vk0 = (int32_t) ((const int8_t*) w)[0]; |
| 602 |
vacc += vi0 * vk0; |
| 603 |
const int32_t vi1 = (int32_t) *i1++; |
| 604 |
const int32_t vk1 = (int32_t) ((const int8_t*) w)[1]; |
| 605 |
vacc += vi1 * vk1; |
| 606 |
const int32_t vi2 = (int32_t) *i2++; |
| 607 |
const int32_t vk2 = (int32_t) ((const int8_t*) w)[2]; |
| 608 |
vacc += vi2 * vk2; |
| 609 |
const int32_t vi3 = (int32_t) *i3++; |
| 610 |
const int32_t vk3 = (int32_t) ((const int8_t*) w)[3]; |
| 611 |
vacc += vi3 * vk3; |
| 612 |
const int32_t vi4 = (int32_t) *i4++; |
| 613 |
const int32_t vk4 = (int32_t) ((const int8_t*) w)[4]; |
| 614 |
vacc += vi4 * vk4; |
| 615 |
const int32_t vi5 = (int32_t) *i5++; |
| 616 |
const int32_t vk5 = (int32_t) ((const int8_t*) w)[5]; |
| 617 |
vacc += vi5 * vk5; |
| 618 |
const int32_t vi6 = (int32_t) *i6++; |
| 619 |
const int32_t vk6 = (int32_t) ((const int8_t*) w)[6]; |
| 620 |
vacc += vi6 * vk6; |
| 621 |
|
| 622 |
w = (const void*) ((uintptr_t) w + 7 * sizeof(int8_t)); |
| 623 |
|
| 624 |
float vfpacc = (float) vacc * vscale; |
| 625 |
|
| 626 |
vfpacc += vmagic_bias; |
| 627 |
int32_t vout = (int32_t) float_as_uint32(vfpacc); |
| 628 |
vout = math_max_s32(vout, vmagic_min); |
| 629 |
vout = math_min_s32(vout, vmagic_max); |
| 630 |
vout -= vmagic_bias_less_zero_point; |
| 631 |
|
| 632 |
*output++ = (int8_t) vout; |
| 633 |
} while (--c != 0); |
| 634 |
} |
| 635 |
} |
| 636 |
|
| 637 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 638 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 639 |
} while (--output_width != 0); |
| 640 |
} |