| 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_9p2c__scalar_fmagic( |
| 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 voutput_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point; |
| 37 |
const float voutput_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point; |
| 38 |
const float vmagic_bias = 12582912.0f; |
| 39 |
const int32_t vmagic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point; |
| 40 |
do { |
| 41 |
const int8_t* i0 = input[0]; |
| 42 |
assert(i0 != NULL); |
| 43 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 44 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 45 |
} |
| 46 |
const int8_t* i1 = input[1]; |
| 47 |
assert(i1 != NULL); |
| 48 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 49 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 50 |
} |
| 51 |
const int8_t* i2 = input[2]; |
| 52 |
assert(i2 != NULL); |
| 53 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 54 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 55 |
} |
| 56 |
const int8_t* i3 = input[3]; |
| 57 |
assert(i3 != NULL); |
| 58 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 59 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 60 |
} |
| 61 |
const int8_t* i4 = input[4]; |
| 62 |
assert(i4 != NULL); |
| 63 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 64 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 65 |
} |
| 66 |
const int8_t* i5 = input[5]; |
| 67 |
assert(i5 != NULL); |
| 68 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 69 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 70 |
} |
| 71 |
const int8_t* i6 = input[6]; |
| 72 |
assert(i6 != NULL); |
| 73 |
if XNN_UNPREDICTABLE(i6 != zero) { |
| 74 |
i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| 75 |
} |
| 76 |
const int8_t* i7 = input[7]; |
| 77 |
assert(i7 != NULL); |
| 78 |
if XNN_UNPREDICTABLE(i7 != zero) { |
| 79 |
i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); |
| 80 |
} |
| 81 |
const int8_t* i8 = input[8]; |
| 82 |
assert(i8 != NULL); |
| 83 |
if XNN_UNPREDICTABLE(i8 != zero) { |
| 84 |
i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); |
| 85 |
} |
| 86 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 87 |
|
| 88 |
size_t c = channels; |
| 89 |
const void* w = weights; |
| 90 |
for (; c >= 2; c -= 2) { |
| 91 |
int32_t vacc0 = unaligned_indexed_load_s32(w, 0); |
| 92 |
int32_t vacc1 = unaligned_indexed_load_s32(w, 1); |
| 93 |
|
| 94 |
|
| 95 |
const int32_t vi0x0 = (int32_t) i0[0]; |
| 96 |
const int32_t vi0x1 = (int32_t) i0[1]; |
| 97 |
i0 += 2; |
| 98 |
|
| 99 |
const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0]; |
| 100 |
const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1]; |
| 101 |
|
| 102 |
vacc0 += vi0x0 * vk0x0; |
| 103 |
vacc1 += vi0x1 * vk0x1; |
| 104 |
|
| 105 |
const int32_t vi1x0 = (int32_t) i1[0]; |
| 106 |
const int32_t vi1x1 = (int32_t) i1[1]; |
| 107 |
i1 += 2; |
| 108 |
|
| 109 |
const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2]; |
| 110 |
const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3]; |
| 111 |
|
| 112 |
vacc0 += vi1x0 * vk1x0; |
| 113 |
vacc1 += vi1x1 * vk1x1; |
| 114 |
|
| 115 |
const int32_t vi2x0 = (int32_t) i2[0]; |
| 116 |
const int32_t vi2x1 = (int32_t) i2[1]; |
| 117 |
i2 += 2; |
| 118 |
|
| 119 |
const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4]; |
| 120 |
const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5]; |
| 121 |
|
| 122 |
vacc0 += vi2x0 * vk2x0; |
| 123 |
vacc1 += vi2x1 * vk2x1; |
| 124 |
|
| 125 |
const int32_t vi3x0 = (int32_t) i3[0]; |
| 126 |
const int32_t vi3x1 = (int32_t) i3[1]; |
| 127 |
i3 += 2; |
| 128 |
|
| 129 |
const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6]; |
| 130 |
const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7]; |
| 131 |
|
| 132 |
vacc0 += vi3x0 * vk3x0; |
| 133 |
vacc1 += vi3x1 * vk3x1; |
| 134 |
|
| 135 |
const int32_t vi4x0 = (int32_t) i4[0]; |
| 136 |
const int32_t vi4x1 = (int32_t) i4[1]; |
| 137 |
i4 += 2; |
| 138 |
|
| 139 |
const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8]; |
| 140 |
const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9]; |
| 141 |
|
| 142 |
vacc0 += vi4x0 * vk4x0; |
| 143 |
vacc1 += vi4x1 * vk4x1; |
| 144 |
|
| 145 |
const int32_t vi5x0 = (int32_t) i5[0]; |
| 146 |
const int32_t vi5x1 = (int32_t) i5[1]; |
| 147 |
i5 += 2; |
| 148 |
|
| 149 |
const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10]; |
| 150 |
const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11]; |
| 151 |
|
| 152 |
vacc0 += vi5x0 * vk5x0; |
| 153 |
vacc1 += vi5x1 * vk5x1; |
| 154 |
|
| 155 |
const int32_t vi6x0 = (int32_t) i6[0]; |
| 156 |
const int32_t vi6x1 = (int32_t) i6[1]; |
| 157 |
i6 += 2; |
| 158 |
|
| 159 |
const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12]; |
| 160 |
const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[13]; |
| 161 |
|
| 162 |
vacc0 += vi6x0 * vk6x0; |
| 163 |
vacc1 += vi6x1 * vk6x1; |
| 164 |
|
| 165 |
const int32_t vi7x0 = (int32_t) i7[0]; |
| 166 |
const int32_t vi7x1 = (int32_t) i7[1]; |
| 167 |
i7 += 2; |
| 168 |
|
| 169 |
const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14]; |
| 170 |
const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[15]; |
| 171 |
|
| 172 |
vacc0 += vi7x0 * vk7x0; |
| 173 |
vacc1 += vi7x1 * vk7x1; |
| 174 |
|
| 175 |
const int32_t vi8x0 = (int32_t) i8[0]; |
| 176 |
const int32_t vi8x1 = (int32_t) i8[1]; |
| 177 |
i8 += 2; |
| 178 |
|
| 179 |
const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16]; |
| 180 |
const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[17]; |
| 181 |
|
| 182 |
vacc0 += vi8x0 * vk8x0; |
| 183 |
vacc1 += vi8x1 * vk8x1; |
| 184 |
|
| 185 |
w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 18 * sizeof(int8_t)); |
| 186 |
|
| 187 |
float vfpacc0 = (float) vacc0; |
| 188 |
float vfpacc1 = (float) vacc1; |
| 189 |
|
| 190 |
vfpacc0 *= vscale; |
| 191 |
vfpacc1 *= vscale; |
| 192 |
|
| 193 |
vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point); |
| 194 |
vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point); |
| 195 |
|
| 196 |
vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point); |
| 197 |
vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point); |
| 198 |
|
| 199 |
vfpacc0 += vmagic_bias; |
| 200 |
vfpacc1 += vmagic_bias; |
| 201 |
|
| 202 |
int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point; |
| 203 |
int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point; |
| 204 |
|
| 205 |
output[0] = (int8_t) vout0; |
| 206 |
output[1] = (int8_t) vout1; |
| 207 |
output += 2; |
| 208 |
} |
| 209 |
if XNN_UNLIKELY(c != 0) { |
| 210 |
int32_t vacc = unaligned_load_s32(w); |
| 211 |
|
| 212 |
const int32_t vi0 = (int32_t) *i0; |
| 213 |
const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0]; |
| 214 |
vacc += vi0 * vk0; |
| 215 |
const int32_t vi1 = (int32_t) *i1; |
| 216 |
const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2]; |
| 217 |
vacc += vi1 * vk1; |
| 218 |
const int32_t vi2 = (int32_t) *i2; |
| 219 |
const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4]; |
| 220 |
vacc += vi2 * vk2; |
| 221 |
const int32_t vi3 = (int32_t) *i3; |
| 222 |
const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6]; |
| 223 |
vacc += vi3 * vk3; |
| 224 |
const int32_t vi4 = (int32_t) *i4; |
| 225 |
const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8]; |
| 226 |
vacc += vi4 * vk4; |
| 227 |
const int32_t vi5 = (int32_t) *i5; |
| 228 |
const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10]; |
| 229 |
vacc += vi5 * vk5; |
| 230 |
const int32_t vi6 = (int32_t) *i6; |
| 231 |
const int32_t vk6 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12]; |
| 232 |
vacc += vi6 * vk6; |
| 233 |
const int32_t vi7 = (int32_t) *i7; |
| 234 |
const int32_t vk7 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14]; |
| 235 |
vacc += vi7 * vk7; |
| 236 |
const int32_t vi8 = (int32_t) *i8; |
| 237 |
const int32_t vk8 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16]; |
| 238 |
vacc += vi8 * vk8; |
| 239 |
|
| 240 |
float vfpacc = (float) vacc * vscale; |
| 241 |
|
| 242 |
vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); |
| 243 |
vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); |
| 244 |
vfpacc += vmagic_bias; |
| 245 |
int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point; |
| 246 |
|
| 247 |
*output++ = (int8_t) vout; |
| 248 |
} |
| 249 |
|
| 250 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 251 |
} while (--output_width != 0); |
| 252 |
} |