| 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_5f5m5l1c1s1r__wasm_fmagic( |
| 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 > 5); |
| 38 |
|
| 39 |
const float vscale = params->fp32_scalar.scale; |
| 40 |
const float voutput_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point; |
| 41 |
const float voutput_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point; |
| 42 |
const float vmagic_bias = 12582912.0f; |
| 43 |
const int32_t vmagic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) 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 |
do { |
| 79 |
int32_t vacc = unaligned_load_s32(w); |
| 80 |
const int32_t vi0 = (int32_t) *i0++; |
| 81 |
const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0]; |
| 82 |
vacc += vi0 * vk0; |
| 83 |
const int32_t vi1 = (int32_t) *i1++; |
| 84 |
const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1]; |
| 85 |
vacc += vi1 * vk1; |
| 86 |
const int32_t vi2 = (int32_t) *i2++; |
| 87 |
const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2]; |
| 88 |
vacc += vi2 * vk2; |
| 89 |
const int32_t vi3 = (int32_t) *i3++; |
| 90 |
const int32_t vk3 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[3]; |
| 91 |
vacc += vi3 * vk3; |
| 92 |
const int32_t vi4 = (int32_t) *i4++; |
| 93 |
const int32_t vk4 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[4]; |
| 94 |
vacc += vi4 * vk4; |
| 95 |
|
| 96 |
w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 5 * sizeof(int8_t)); |
| 97 |
*b++ = vacc; |
| 98 |
} while (--c != 0); |
| 99 |
} |
| 100 |
|
| 101 |
// Middle pass to process 5 inputs in each iteration. |
| 102 |
for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) { |
| 103 |
int32_t* b = buffer; |
| 104 |
const int8_t* i0 = input[0]; |
| 105 |
assert(i0 != NULL); |
| 106 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 107 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 108 |
} |
| 109 |
const int8_t* i1 = input[1]; |
| 110 |
assert(i1 != NULL); |
| 111 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 112 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 113 |
} |
| 114 |
const int8_t* i2 = input[2]; |
| 115 |
assert(i2 != NULL); |
| 116 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 117 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 118 |
} |
| 119 |
const int8_t* i3 = input[3]; |
| 120 |
assert(i3 != NULL); |
| 121 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 122 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 123 |
} |
| 124 |
const int8_t* i4 = input[4]; |
| 125 |
assert(i4 != NULL); |
| 126 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 127 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 128 |
} |
| 129 |
input += 5; |
| 130 |
|
| 131 |
size_t c = channels; |
| 132 |
do { |
| 133 |
int32_t vacc = *b; |
| 134 |
const int32_t vi0 = (int32_t) *i0++; |
| 135 |
const int32_t vk0 = ((const int8_t*) w)[0]; |
| 136 |
vacc += vi0 * vk0; |
| 137 |
const int32_t vi1 = (int32_t) *i1++; |
| 138 |
const int32_t vk1 = ((const int8_t*) w)[1]; |
| 139 |
vacc += vi1 * vk1; |
| 140 |
const int32_t vi2 = (int32_t) *i2++; |
| 141 |
const int32_t vk2 = ((const int8_t*) w)[2]; |
| 142 |
vacc += vi2 * vk2; |
| 143 |
const int32_t vi3 = (int32_t) *i3++; |
| 144 |
const int32_t vk3 = ((const int8_t*) w)[3]; |
| 145 |
vacc += vi3 * vk3; |
| 146 |
const int32_t vi4 = (int32_t) *i4++; |
| 147 |
const int32_t vk4 = ((const int8_t*) w)[4]; |
| 148 |
vacc += vi4 * vk4; |
| 149 |
|
| 150 |
w = (const void*) ((uintptr_t) w + 5 * sizeof(int8_t)); |
| 151 |
*b++ = vacc; |
| 152 |
} while (--c != 0); |
| 153 |
} |
| 154 |
|
| 155 |
// Last pass to process up to 5 inputs. |
| 156 |
{ |
| 157 |
const int32_t* b = buffer; |
| 158 |
const int8_t* i0 = input[0]; |
| 159 |
assert(i0 != NULL); |
| 160 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 161 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 162 |
} |
| 163 |
const int8_t* i1 = input[1]; |
| 164 |
assert(i1 != NULL); |
| 165 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 166 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 167 |
} |
| 168 |
const int8_t* i2 = input[2]; |
| 169 |
assert(i2 != NULL); |
| 170 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 171 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 172 |
} |
| 173 |
const int8_t* i3 = input[3]; |
| 174 |
assert(i3 != NULL); |
| 175 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 176 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 177 |
} |
| 178 |
const int8_t* i4 = input[4]; |
| 179 |
assert(i4 != NULL); |
| 180 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 181 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 182 |
} |
| 183 |
|
| 184 |
size_t c = channels; |
| 185 |
do { |
| 186 |
int32_t vacc = unaligned_load_s32(b++); |
| 187 |
const int32_t vi0 = (int32_t) *i0++; |
| 188 |
const int32_t vk0 = ((const int8_t*) w)[0]; |
| 189 |
vacc += vi0 * vk0; |
| 190 |
const int32_t vi1 = (int32_t) *i1++; |
| 191 |
const int32_t vk1 = ((const int8_t*) w)[1]; |
| 192 |
vacc += vi1 * vk1; |
| 193 |
const int32_t vi2 = (int32_t) *i2++; |
| 194 |
const int32_t vk2 = ((const int8_t*) w)[2]; |
| 195 |
vacc += vi2 * vk2; |
| 196 |
const int32_t vi3 = (int32_t) *i3++; |
| 197 |
const int32_t vk3 = ((const int8_t*) w)[3]; |
| 198 |
vacc += vi3 * vk3; |
| 199 |
const int32_t vi4 = (int32_t) *i4++; |
| 200 |
const int32_t vk4 = ((const int8_t*) w)[4]; |
| 201 |
vacc += vi4 * vk4; |
| 202 |
|
| 203 |
w = (const void*) ((uintptr_t) w + 5 * sizeof(int8_t)); |
| 204 |
|
| 205 |
float vfpacc = (float) vacc * vscale; |
| 206 |
|
| 207 |
vfpacc = __builtin_wasm_max_f32(vfpacc, voutput_min_less_zero_point); |
| 208 |
vfpacc = __builtin_wasm_min_f32(vfpacc, voutput_max_less_zero_point); |
| 209 |
vfpacc += vmagic_bias; |
| 210 |
int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point; |
| 211 |
|
| 212 |
*output++ = (int8_t) vout; |
| 213 |
} while (--c != 0); |
| 214 |
} |
| 215 |
|
| 216 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 217 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 218 |
} while (--output_width != 0); |
| 219 |
} |