| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-dwconv/unipass-scalar.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2021 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 |
void xnn_qs8_dwconv_minmax_fp32_ukernel_25p1c__scalar_imagic( |
| 21 |
size_t channels, |
| 22 |
size_t output_width, |
| 23 |
const int8_t** input, |
| 24 |
const void* weights, |
| 25 |
int8_t* output, |
| 26 |
intptr_t input_stride, |
| 27 |
size_t output_increment, |
| 28 |
size_t input_offset, |
| 29 |
const int8_t* zero, |
| 30 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 31 |
{ |
| 32 |
assert(channels != 0); |
| 33 |
assert(output_width != 0); |
| 34 |
|
| 35 |
const float vscale = params->fp32_scalar.scale; |
| 36 |
const float vmagic_bias = 12582912.0f; |
| 37 |
const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point; |
| 38 |
const int32_t output_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point; |
| 39 |
const int32_t vmagic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point); |
| 40 |
const int32_t vmagic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_point); |
| 41 |
const int32_t vmagic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point; |
| 42 |
do { |
| 43 |
const int8_t* i0 = input[0]; |
| 44 |
assert(i0 != NULL); |
| 45 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 46 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 47 |
} |
| 48 |
const int8_t* i1 = input[1]; |
| 49 |
assert(i1 != NULL); |
| 50 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 51 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 52 |
} |
| 53 |
const int8_t* i2 = input[2]; |
| 54 |
assert(i2 != NULL); |
| 55 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 56 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 57 |
} |
| 58 |
const int8_t* i3 = input[3]; |
| 59 |
assert(i3 != NULL); |
| 60 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 61 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 62 |
} |
| 63 |
const int8_t* i4 = input[4]; |
| 64 |
assert(i4 != NULL); |
| 65 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 66 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 67 |
} |
| 68 |
const int8_t* i5 = input[5]; |
| 69 |
assert(i5 != NULL); |
| 70 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 71 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 72 |
} |
| 73 |
const int8_t* i6 = input[6]; |
| 74 |
assert(i6 != NULL); |
| 75 |
if XNN_UNPREDICTABLE(i6 != zero) { |
| 76 |
i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| 77 |
} |
| 78 |
const int8_t* i7 = input[7]; |
| 79 |
assert(i7 != NULL); |
| 80 |
if XNN_UNPREDICTABLE(i7 != zero) { |
| 81 |
i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); |
| 82 |
} |
| 83 |
const int8_t* i8 = input[8]; |
| 84 |
assert(i8 != NULL); |
| 85 |
if XNN_UNPREDICTABLE(i8 != zero) { |
| 86 |
i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); |
| 87 |
} |
| 88 |
const int8_t* i9 = input[9]; |
| 89 |
assert(i9 != NULL); |
| 90 |
if XNN_UNPREDICTABLE(i9 != zero) { |
| 91 |
i9 = (const int8_t*) ((uintptr_t) i9 + input_offset); |
| 92 |
} |
| 93 |
const int8_t* i10 = input[10]; |
| 94 |
assert(i10 != NULL); |
| 95 |
if XNN_UNPREDICTABLE(i10 != zero) { |
| 96 |
i10 = (const int8_t*) ((uintptr_t) i10 + input_offset); |
| 97 |
} |
| 98 |
const int8_t* i11 = input[11]; |
| 99 |
assert(i11 != NULL); |
| 100 |
if XNN_UNPREDICTABLE(i11 != zero) { |
| 101 |
i11 = (const int8_t*) ((uintptr_t) i11 + input_offset); |
| 102 |
} |
| 103 |
const int8_t* i12 = input[12]; |
| 104 |
assert(i12 != NULL); |
| 105 |
if XNN_UNPREDICTABLE(i12 != zero) { |
| 106 |
i12 = (const int8_t*) ((uintptr_t) i12 + input_offset); |
| 107 |
} |
| 108 |
const int8_t* i13 = input[13]; |
| 109 |
assert(i13 != NULL); |
| 110 |
if XNN_UNPREDICTABLE(i13 != zero) { |
| 111 |
i13 = (const int8_t*) ((uintptr_t) i13 + input_offset); |
| 112 |
} |
| 113 |
const int8_t* i14 = input[14]; |
| 114 |
assert(i14 != NULL); |
| 115 |
if XNN_UNPREDICTABLE(i14 != zero) { |
| 116 |
i14 = (const int8_t*) ((uintptr_t) i14 + input_offset); |
| 117 |
} |
| 118 |
const int8_t* i15 = input[15]; |
| 119 |
assert(i15 != NULL); |
| 120 |
if XNN_UNPREDICTABLE(i15 != zero) { |
| 121 |
i15 = (const int8_t*) ((uintptr_t) i15 + input_offset); |
| 122 |
} |
| 123 |
const int8_t* i16 = input[16]; |
| 124 |
assert(i16 != NULL); |
| 125 |
if XNN_UNPREDICTABLE(i16 != zero) { |
| 126 |
i16 = (const int8_t*) ((uintptr_t) i16 + input_offset); |
| 127 |
} |
| 128 |
const int8_t* i17 = input[17]; |
| 129 |
assert(i17 != NULL); |
| 130 |
if XNN_UNPREDICTABLE(i17 != zero) { |
| 131 |
i17 = (const int8_t*) ((uintptr_t) i17 + input_offset); |
| 132 |
} |
| 133 |
const int8_t* i18 = input[18]; |
| 134 |
assert(i18 != NULL); |
| 135 |
if XNN_UNPREDICTABLE(i18 != zero) { |
| 136 |
i18 = (const int8_t*) ((uintptr_t) i18 + input_offset); |
| 137 |
} |
| 138 |
const int8_t* i19 = input[19]; |
| 139 |
assert(i19 != NULL); |
| 140 |
if XNN_UNPREDICTABLE(i19 != zero) { |
| 141 |
i19 = (const int8_t*) ((uintptr_t) i19 + input_offset); |
| 142 |
} |
| 143 |
const int8_t* i20 = input[20]; |
| 144 |
assert(i20 != NULL); |
| 145 |
if XNN_UNPREDICTABLE(i20 != zero) { |
| 146 |
i20 = (const int8_t*) ((uintptr_t) i20 + input_offset); |
| 147 |
} |
| 148 |
const int8_t* i21 = input[21]; |
| 149 |
assert(i21 != NULL); |
| 150 |
if XNN_UNPREDICTABLE(i21 != zero) { |
| 151 |
i21 = (const int8_t*) ((uintptr_t) i21 + input_offset); |
| 152 |
} |
| 153 |
const int8_t* i22 = input[22]; |
| 154 |
assert(i22 != NULL); |
| 155 |
if XNN_UNPREDICTABLE(i22 != zero) { |
| 156 |
i22 = (const int8_t*) ((uintptr_t) i22 + input_offset); |
| 157 |
} |
| 158 |
const int8_t* i23 = input[23]; |
| 159 |
assert(i23 != NULL); |
| 160 |
if XNN_UNPREDICTABLE(i23 != zero) { |
| 161 |
i23 = (const int8_t*) ((uintptr_t) i23 + input_offset); |
| 162 |
} |
| 163 |
const int8_t* i24 = input[24]; |
| 164 |
assert(i24 != NULL); |
| 165 |
if XNN_UNPREDICTABLE(i24 != zero) { |
| 166 |
i24 = (const int8_t*) ((uintptr_t) i24 + input_offset); |
| 167 |
} |
| 168 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 169 |
|
| 170 |
size_t c = channels; |
| 171 |
const void* w = weights; |
| 172 |
do { |
| 173 |
int32_t vacc = unaligned_load_s32(w); |
| 174 |
|
| 175 |
const int32_t vi0 = (int32_t) *i0++; |
| 176 |
const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0]; |
| 177 |
vacc += vi0 * vk0; |
| 178 |
const int32_t vi1 = (int32_t) *i1++; |
| 179 |
const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1]; |
| 180 |
vacc += vi1 * vk1; |
| 181 |
const int32_t vi2 = (int32_t) *i2++; |
| 182 |
const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2]; |
| 183 |
vacc += vi2 * vk2; |
| 184 |
const int32_t vi3 = (int32_t) *i3++; |
| 185 |
const int32_t vk3 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[3]; |
| 186 |
vacc += vi3 * vk3; |
| 187 |
const int32_t vi4 = (int32_t) *i4++; |
| 188 |
const int32_t vk4 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[4]; |
| 189 |
vacc += vi4 * vk4; |
| 190 |
const int32_t vi5 = (int32_t) *i5++; |
| 191 |
const int32_t vk5 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[5]; |
| 192 |
vacc += vi5 * vk5; |
| 193 |
const int32_t vi6 = (int32_t) *i6++; |
| 194 |
const int32_t vk6 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[6]; |
| 195 |
vacc += vi6 * vk6; |
| 196 |
const int32_t vi7 = (int32_t) *i7++; |
| 197 |
const int32_t vk7 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[7]; |
| 198 |
vacc += vi7 * vk7; |
| 199 |
const int32_t vi8 = (int32_t) *i8++; |
| 200 |
const int32_t vk8 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[8]; |
| 201 |
vacc += vi8 * vk8; |
| 202 |
const int32_t vi9 = (int32_t) *i9++; |
| 203 |
const int32_t vk9 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[9]; |
| 204 |
vacc += vi9 * vk9; |
| 205 |
const int32_t vi10 = (int32_t) *i10++; |
| 206 |
const int32_t vk10 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[10]; |
| 207 |
vacc += vi10 * vk10; |
| 208 |
const int32_t vi11 = (int32_t) *i11++; |
| 209 |
const int32_t vk11 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[11]; |
| 210 |
vacc += vi11 * vk11; |
| 211 |
const int32_t vi12 = (int32_t) *i12++; |
| 212 |
const int32_t vk12 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[12]; |
| 213 |
vacc += vi12 * vk12; |
| 214 |
const int32_t vi13 = (int32_t) *i13++; |
| 215 |
const int32_t vk13 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[13]; |
| 216 |
vacc += vi13 * vk13; |
| 217 |
const int32_t vi14 = (int32_t) *i14++; |
| 218 |
const int32_t vk14 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[14]; |
| 219 |
vacc += vi14 * vk14; |
| 220 |
const int32_t vi15 = (int32_t) *i15++; |
| 221 |
const int32_t vk15 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[15]; |
| 222 |
vacc += vi15 * vk15; |
| 223 |
const int32_t vi16 = (int32_t) *i16++; |
| 224 |
const int32_t vk16 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[16]; |
| 225 |
vacc += vi16 * vk16; |
| 226 |
const int32_t vi17 = (int32_t) *i17++; |
| 227 |
const int32_t vk17 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[17]; |
| 228 |
vacc += vi17 * vk17; |
| 229 |
const int32_t vi18 = (int32_t) *i18++; |
| 230 |
const int32_t vk18 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[18]; |
| 231 |
vacc += vi18 * vk18; |
| 232 |
const int32_t vi19 = (int32_t) *i19++; |
| 233 |
const int32_t vk19 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[19]; |
| 234 |
vacc += vi19 * vk19; |
| 235 |
const int32_t vi20 = (int32_t) *i20++; |
| 236 |
const int32_t vk20 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[20]; |
| 237 |
vacc += vi20 * vk20; |
| 238 |
const int32_t vi21 = (int32_t) *i21++; |
| 239 |
const int32_t vk21 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[21]; |
| 240 |
vacc += vi21 * vk21; |
| 241 |
const int32_t vi22 = (int32_t) *i22++; |
| 242 |
const int32_t vk22 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[22]; |
| 243 |
vacc += vi22 * vk22; |
| 244 |
const int32_t vi23 = (int32_t) *i23++; |
| 245 |
const int32_t vk23 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[23]; |
| 246 |
vacc += vi23 * vk23; |
| 247 |
const int32_t vi24 = (int32_t) *i24++; |
| 248 |
const int32_t vk24 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[24]; |
| 249 |
vacc += vi24 * vk24; |
| 250 |
|
| 251 |
w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 25 * sizeof(int8_t)); |
| 252 |
|
| 253 |
float vfpacc = (float) vacc * vscale; |
| 254 |
|
| 255 |
vfpacc += vmagic_bias; |
| 256 |
int32_t vout = (int32_t) float_as_uint32(vfpacc); |
| 257 |
vout = math_max_s32(vout, vmagic_min); |
| 258 |
vout = math_min_s32(vout, vmagic_max); |
| 259 |
vout -= vmagic_bias_less_zero_point; |
| 260 |
|
| 261 |
*output++ = (int8_t) vout; |
| 262 |
} while (--c != 0); |
| 263 |
|
| 264 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 265 |
} while (--output_width != 0); |
| 266 |
} |