| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-dwconv/unipass-sse-mul32.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2020 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 |
|
| 12 |
#include <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/dwconv.h" |
| 15 |
#include "xnnpack/intrinsics-polyfill.h" |
| 16 |
#include "xnnpack/unaligned.h" |
| 17 |
|
| 18 |
|
| 19 |
void xnn_qs8_dwconv_minmax_fp32_ukernel_9p8c__avx_mul32( |
| 20 |
size_t channels, |
| 21 |
size_t output_width, |
| 22 |
const int8_t** input, |
| 23 |
const void* weights, |
| 24 |
int8_t* output, |
| 25 |
intptr_t input_stride, |
| 26 |
size_t output_increment, |
| 27 |
size_t input_offset, |
| 28 |
const int8_t* zero, |
| 29 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 30 |
{ |
| 31 |
assert(channels != 0); |
| 32 |
assert(output_width != 0); |
| 33 |
|
| 34 |
|
| 35 |
const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale); |
| 36 |
XNN_FORCE_REALIZATION(vscale); |
| 37 |
const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point); |
| 38 |
const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point); |
| 39 |
const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min); |
| 40 |
XNN_FORCE_REALIZATION(voutput_max_less_zero_point); |
| 41 |
XNN_FORCE_REALIZATION(voutput_zero_point); |
| 42 |
XNN_FORCE_REALIZATION(voutput_min); |
| 43 |
|
| 44 |
do { |
| 45 |
const int8_t* i0 = input[0]; |
| 46 |
assert(i0 != NULL); |
| 47 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 48 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 49 |
} |
| 50 |
const int8_t* i1 = input[1]; |
| 51 |
assert(i1 != NULL); |
| 52 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 53 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 54 |
} |
| 55 |
const int8_t* i2 = input[2]; |
| 56 |
assert(i2 != NULL); |
| 57 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 58 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 59 |
} |
| 60 |
const int8_t* i3 = input[3]; |
| 61 |
assert(i3 != NULL); |
| 62 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 63 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 64 |
} |
| 65 |
const int8_t* i4 = input[4]; |
| 66 |
assert(i4 != NULL); |
| 67 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 68 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 69 |
} |
| 70 |
const int8_t* i5 = input[5]; |
| 71 |
assert(i5 != NULL); |
| 72 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 73 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 74 |
} |
| 75 |
const int8_t* i6 = input[6]; |
| 76 |
assert(i6 != NULL); |
| 77 |
if XNN_UNPREDICTABLE(i6 != zero) { |
| 78 |
i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| 79 |
} |
| 80 |
const int8_t* i7 = input[7]; |
| 81 |
assert(i7 != NULL); |
| 82 |
if XNN_UNPREDICTABLE(i7 != zero) { |
| 83 |
i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); |
| 84 |
} |
| 85 |
const int8_t* i8 = input[8]; |
| 86 |
assert(i8 != NULL); |
| 87 |
if XNN_UNPREDICTABLE(i8 != zero) { |
| 88 |
i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); |
| 89 |
} |
| 90 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 91 |
|
| 92 |
size_t c = channels; |
| 93 |
const void* w = weights; |
| 94 |
for (; c >= 8; c -= 8) { |
| 95 |
__m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); |
| 96 |
__m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); |
| 97 |
|
| 98 |
|
| 99 |
const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0))); |
| 100 |
const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))))); |
| 101 |
const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4))); |
| 102 |
const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(int8_t))))); |
| 103 |
i0 += 8; |
| 104 |
|
| 105 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); |
| 106 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567)); |
| 107 |
|
| 108 |
const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1))); |
| 109 |
const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))))); |
| 110 |
const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4))); |
| 111 |
const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(int8_t))))); |
| 112 |
i1 += 8; |
| 113 |
|
| 114 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); |
| 115 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567)); |
| 116 |
|
| 117 |
const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2))); |
| 118 |
const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))))); |
| 119 |
const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4))); |
| 120 |
const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(int8_t))))); |
| 121 |
i2 += 8; |
| 122 |
|
| 123 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); |
| 124 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567)); |
| 125 |
|
| 126 |
const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3))); |
| 127 |
const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))))); |
| 128 |
const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4))); |
| 129 |
const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(int8_t))))); |
| 130 |
i3 += 8; |
| 131 |
|
| 132 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123)); |
| 133 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567)); |
| 134 |
|
| 135 |
const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4))); |
| 136 |
const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))))); |
| 137 |
const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4))); |
| 138 |
const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(int8_t))))); |
| 139 |
i4 += 8; |
| 140 |
|
| 141 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123)); |
| 142 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); |
| 143 |
|
| 144 |
const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5))); |
| 145 |
const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))))); |
| 146 |
const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4))); |
| 147 |
const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(int8_t))))); |
| 148 |
i5 += 8; |
| 149 |
|
| 150 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123)); |
| 151 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567)); |
| 152 |
|
| 153 |
const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6))); |
| 154 |
const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))))); |
| 155 |
const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4))); |
| 156 |
const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(int8_t))))); |
| 157 |
i6 += 8; |
| 158 |
|
| 159 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123)); |
| 160 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567)); |
| 161 |
|
| 162 |
const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7))); |
| 163 |
const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))))); |
| 164 |
const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4))); |
| 165 |
const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(int8_t))))); |
| 166 |
i7 += 8; |
| 167 |
|
| 168 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123)); |
| 169 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567)); |
| 170 |
|
| 171 |
const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8))); |
| 172 |
const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))))); |
| 173 |
const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4))); |
| 174 |
const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 68 * sizeof(int8_t))))); |
| 175 |
i8 += 8; |
| 176 |
|
| 177 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123)); |
| 178 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567)); |
| 179 |
|
| 180 |
w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)); |
| 181 |
|
| 182 |
__m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); |
| 183 |
__m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); |
| 184 |
|
| 185 |
vscaled0123 = _mm_mul_ps(vscaled0123, vscale); |
| 186 |
vscaled4567 = _mm_mul_ps(vscaled4567, vscale); |
| 187 |
|
| 188 |
vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); |
| 189 |
vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); |
| 190 |
|
| 191 |
vacc0123 = _mm_cvtps_epi32(vscaled0123); |
| 192 |
vacc4567 = _mm_cvtps_epi32(vscaled4567); |
| 193 |
|
| 194 |
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); |
| 195 |
|
| 196 |
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); |
| 197 |
vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min); |
| 198 |
|
| 199 |
_mm_storel_epi64((__m128i*) output, vout0123456701234567); |
| 200 |
output += 8; |
| 201 |
} |
| 202 |
if XNN_UNLIKELY(c != 0) { |
| 203 |
const int8_t* k = (const int8_t*) ((const int32_t*) w + 8); |
| 204 |
do { |
| 205 |
__m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); |
| 206 |
|
| 207 |
const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0))); |
| 208 |
const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k))); |
| 209 |
i0 += 4; |
| 210 |
|
| 211 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123)); |
| 212 |
const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1))); |
| 213 |
const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 8)))); |
| 214 |
i1 += 4; |
| 215 |
|
| 216 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123)); |
| 217 |
const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2))); |
| 218 |
const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16)))); |
| 219 |
i2 += 4; |
| 220 |
|
| 221 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123)); |
| 222 |
const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3))); |
| 223 |
const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24)))); |
| 224 |
i3 += 4; |
| 225 |
|
| 226 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123)); |
| 227 |
const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4))); |
| 228 |
const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32)))); |
| 229 |
i4 += 4; |
| 230 |
|
| 231 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123)); |
| 232 |
const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5))); |
| 233 |
const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 40)))); |
| 234 |
i5 += 4; |
| 235 |
|
| 236 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123)); |
| 237 |
const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6))); |
| 238 |
const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48)))); |
| 239 |
i6 += 4; |
| 240 |
|
| 241 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123)); |
| 242 |
const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7))); |
| 243 |
const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 56)))); |
| 244 |
i7 += 4; |
| 245 |
|
| 246 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123)); |
| 247 |
const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8))); |
| 248 |
const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64)))); |
| 249 |
i8 += 4; |
| 250 |
|
| 251 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123)); |
| 252 |
|
| 253 |
k += 4; |
| 254 |
|
| 255 |
__m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); |
| 256 |
vscaled0123 = _mm_mul_ps(vscaled0123, vscale); |
| 257 |
vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); |
| 258 |
vacc0123 = _mm_cvtps_epi32(vscaled0123); |
| 259 |
|
| 260 |
w = (const void*) ((const int32_t*) w + 4); |
| 261 |
|
| 262 |
__m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point); |
| 263 |
|
| 264 |
vout0123 = _mm_packs_epi16(vout0123, vout0123); |
| 265 |
vout0123 = _mm_max_epi8(vout0123, voutput_min); |
| 266 |
|
| 267 |
if XNN_LIKELY(c >= 4) { |
| 268 |
_mm_storeu_si32(output, vout0123); |
| 269 |
output += 4; |
| 270 |
c -= 4; |
| 271 |
} else { |
| 272 |
if (c & 2) { |
| 273 |
unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0)); |
| 274 |
vout0123 = _mm_srli_epi32(vout0123, 16); |
| 275 |
output += 2; |
| 276 |
} |
| 277 |
if (c & 1) { |
| 278 |
*output = (int8_t) _mm_extract_epi8(vout0123, 0); |
| 279 |
output += 1; |
| 280 |
} |
| 281 |
c = 0; |
| 282 |
} |
| 283 |
} while (c != 0); |
| 284 |
} |
| 285 |
|
| 286 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 287 |
} while (--output_width != 0); |
| 288 |
} |