| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-dwconv/unipass-neon-mul8.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 <arm_neon.h> |
| 13 |
|
| 14 |
#include "xnnpack/dwconv.h" |
| 15 |
|
| 16 |
|
| 17 |
void xnn_qs8_dwconv_minmax_rndnu_ukernel_25p16c__neon_mul8_ld64( |
| 18 |
size_t channels, |
| 19 |
size_t output_width, |
| 20 |
const int8_t** input, |
| 21 |
const void* weights, |
| 22 |
int8_t* output, |
| 23 |
intptr_t input_stride, |
| 24 |
size_t output_increment, |
| 25 |
size_t input_offset, |
| 26 |
const int8_t* zero, |
| 27 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 28 |
{ |
| 29 |
assert(channels != 0); |
| 30 |
assert(output_width != 0); |
| 31 |
|
| 32 |
const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift); |
| 33 |
const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier); |
| 34 |
const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift); |
| 35 |
const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point); |
| 36 |
const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min); |
| 37 |
const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max); |
| 38 |
do { |
| 39 |
const int8_t* i0 = input[0]; |
| 40 |
assert(i0 != NULL); |
| 41 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 42 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 43 |
} |
| 44 |
const int8_t* i1 = input[1]; |
| 45 |
assert(i1 != NULL); |
| 46 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 47 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 48 |
} |
| 49 |
const int8_t* i2 = input[2]; |
| 50 |
assert(i2 != NULL); |
| 51 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 52 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 53 |
} |
| 54 |
const int8_t* i3 = input[3]; |
| 55 |
assert(i3 != NULL); |
| 56 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 57 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 58 |
} |
| 59 |
const int8_t* i4 = input[4]; |
| 60 |
assert(i4 != NULL); |
| 61 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 62 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 63 |
} |
| 64 |
const int8_t* i5 = input[5]; |
| 65 |
assert(i5 != NULL); |
| 66 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 67 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 68 |
} |
| 69 |
const int8_t* i6 = input[6]; |
| 70 |
assert(i6 != NULL); |
| 71 |
if XNN_UNPREDICTABLE(i6 != zero) { |
| 72 |
i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| 73 |
} |
| 74 |
const int8_t* i7 = input[7]; |
| 75 |
assert(i7 != NULL); |
| 76 |
if XNN_UNPREDICTABLE(i7 != zero) { |
| 77 |
i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); |
| 78 |
} |
| 79 |
const int8_t* i8 = input[8]; |
| 80 |
assert(i8 != NULL); |
| 81 |
if XNN_UNPREDICTABLE(i8 != zero) { |
| 82 |
i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); |
| 83 |
} |
| 84 |
const int8_t* i9 = input[9]; |
| 85 |
assert(i9 != NULL); |
| 86 |
if XNN_UNPREDICTABLE(i9 != zero) { |
| 87 |
i9 = (const int8_t*) ((uintptr_t) i9 + input_offset); |
| 88 |
} |
| 89 |
const int8_t* i10 = input[10]; |
| 90 |
assert(i10 != NULL); |
| 91 |
if XNN_UNPREDICTABLE(i10 != zero) { |
| 92 |
i10 = (const int8_t*) ((uintptr_t) i10 + input_offset); |
| 93 |
} |
| 94 |
const int8_t* i11 = input[11]; |
| 95 |
assert(i11 != NULL); |
| 96 |
if XNN_UNPREDICTABLE(i11 != zero) { |
| 97 |
i11 = (const int8_t*) ((uintptr_t) i11 + input_offset); |
| 98 |
} |
| 99 |
const int8_t* i12 = input[12]; |
| 100 |
assert(i12 != NULL); |
| 101 |
if XNN_UNPREDICTABLE(i12 != zero) { |
| 102 |
i12 = (const int8_t*) ((uintptr_t) i12 + input_offset); |
| 103 |
} |
| 104 |
const int8_t* i13 = input[13]; |
| 105 |
assert(i13 != NULL); |
| 106 |
if XNN_UNPREDICTABLE(i13 != zero) { |
| 107 |
i13 = (const int8_t*) ((uintptr_t) i13 + input_offset); |
| 108 |
} |
| 109 |
const int8_t* i14 = input[14]; |
| 110 |
assert(i14 != NULL); |
| 111 |
if XNN_UNPREDICTABLE(i14 != zero) { |
| 112 |
i14 = (const int8_t*) ((uintptr_t) i14 + input_offset); |
| 113 |
} |
| 114 |
const int8_t* i15 = input[15]; |
| 115 |
assert(i15 != NULL); |
| 116 |
if XNN_UNPREDICTABLE(i15 != zero) { |
| 117 |
i15 = (const int8_t*) ((uintptr_t) i15 + input_offset); |
| 118 |
} |
| 119 |
const int8_t* i16 = input[16]; |
| 120 |
assert(i16 != NULL); |
| 121 |
if XNN_UNPREDICTABLE(i16 != zero) { |
| 122 |
i16 = (const int8_t*) ((uintptr_t) i16 + input_offset); |
| 123 |
} |
| 124 |
const int8_t* i17 = input[17]; |
| 125 |
assert(i17 != NULL); |
| 126 |
if XNN_UNPREDICTABLE(i17 != zero) { |
| 127 |
i17 = (const int8_t*) ((uintptr_t) i17 + input_offset); |
| 128 |
} |
| 129 |
const int8_t* i18 = input[18]; |
| 130 |
assert(i18 != NULL); |
| 131 |
if XNN_UNPREDICTABLE(i18 != zero) { |
| 132 |
i18 = (const int8_t*) ((uintptr_t) i18 + input_offset); |
| 133 |
} |
| 134 |
const int8_t* i19 = input[19]; |
| 135 |
assert(i19 != NULL); |
| 136 |
if XNN_UNPREDICTABLE(i19 != zero) { |
| 137 |
i19 = (const int8_t*) ((uintptr_t) i19 + input_offset); |
| 138 |
} |
| 139 |
const int8_t* i20 = input[20]; |
| 140 |
assert(i20 != NULL); |
| 141 |
if XNN_UNPREDICTABLE(i20 != zero) { |
| 142 |
i20 = (const int8_t*) ((uintptr_t) i20 + input_offset); |
| 143 |
} |
| 144 |
const int8_t* i21 = input[21]; |
| 145 |
assert(i21 != NULL); |
| 146 |
if XNN_UNPREDICTABLE(i21 != zero) { |
| 147 |
i21 = (const int8_t*) ((uintptr_t) i21 + input_offset); |
| 148 |
} |
| 149 |
const int8_t* i22 = input[22]; |
| 150 |
assert(i22 != NULL); |
| 151 |
if XNN_UNPREDICTABLE(i22 != zero) { |
| 152 |
i22 = (const int8_t*) ((uintptr_t) i22 + input_offset); |
| 153 |
} |
| 154 |
const int8_t* i23 = input[23]; |
| 155 |
assert(i23 != NULL); |
| 156 |
if XNN_UNPREDICTABLE(i23 != zero) { |
| 157 |
i23 = (const int8_t*) ((uintptr_t) i23 + input_offset); |
| 158 |
} |
| 159 |
const int8_t* i24 = input[24]; |
| 160 |
assert(i24 != NULL); |
| 161 |
if XNN_UNPREDICTABLE(i24 != zero) { |
| 162 |
i24 = (const int8_t*) ((uintptr_t) i24 + input_offset); |
| 163 |
} |
| 164 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 165 |
|
| 166 |
size_t c = channels; |
| 167 |
const void* w = weights; |
| 168 |
for (; c >= 16; c -= 16) { |
| 169 |
int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4; |
| 170 |
int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4; |
| 171 |
int32x4_t vacc89AB = vld1q_s32(w); w = (const int32_t*) w + 4; |
| 172 |
int32x4_t vaccCDEF = vld1q_s32(w); w = (const int32_t*) w + 4; |
| 173 |
|
| 174 |
const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; |
| 175 |
const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 176 |
const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8; |
| 177 |
const int8x8_t vk0x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 178 |
|
| 179 |
int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567); |
| 180 |
int16x8_t vprod89ABCDEF = vmull_s8(vi0x89ABCDEF, vk0x89ABCDEF); |
| 181 |
|
| 182 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 183 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 184 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 185 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 186 |
const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; |
| 187 |
const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 188 |
const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8; |
| 189 |
const int8x8_t vk1x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 190 |
|
| 191 |
vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567); |
| 192 |
vprod89ABCDEF = vmull_s8(vi1x89ABCDEF, vk1x89ABCDEF); |
| 193 |
|
| 194 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 195 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 196 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 197 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 198 |
const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; |
| 199 |
const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 200 |
const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; |
| 201 |
const int8x8_t vk2x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 202 |
|
| 203 |
vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567); |
| 204 |
vprod89ABCDEF = vmull_s8(vi2x89ABCDEF, vk2x89ABCDEF); |
| 205 |
|
| 206 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 207 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 208 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 209 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 210 |
const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8; |
| 211 |
const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 212 |
const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8; |
| 213 |
const int8x8_t vk3x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 214 |
|
| 215 |
vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567); |
| 216 |
vprod89ABCDEF = vmull_s8(vi3x89ABCDEF, vk3x89ABCDEF); |
| 217 |
|
| 218 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 219 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 220 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 221 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 222 |
const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8; |
| 223 |
const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 224 |
const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; |
| 225 |
const int8x8_t vk4x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 226 |
|
| 227 |
vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567); |
| 228 |
vprod89ABCDEF = vmull_s8(vi4x89ABCDEF, vk4x89ABCDEF); |
| 229 |
|
| 230 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 231 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 232 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 233 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 234 |
const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8; |
| 235 |
const int8x8_t vk5x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 236 |
const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8; |
| 237 |
const int8x8_t vk5x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 238 |
|
| 239 |
vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567); |
| 240 |
vprod89ABCDEF = vmull_s8(vi5x89ABCDEF, vk5x89ABCDEF); |
| 241 |
|
| 242 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 243 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 244 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 245 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 246 |
const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8; |
| 247 |
const int8x8_t vk6x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 248 |
const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8; |
| 249 |
const int8x8_t vk6x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 250 |
|
| 251 |
vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567); |
| 252 |
vprod89ABCDEF = vmull_s8(vi6x89ABCDEF, vk6x89ABCDEF); |
| 253 |
|
| 254 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 255 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 256 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 257 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 258 |
const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8; |
| 259 |
const int8x8_t vk7x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 260 |
const int8x8_t vi7x89ABCDEF = vld1_s8(i7); i7 += 8; |
| 261 |
const int8x8_t vk7x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 262 |
|
| 263 |
vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567); |
| 264 |
vprod89ABCDEF = vmull_s8(vi7x89ABCDEF, vk7x89ABCDEF); |
| 265 |
|
| 266 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 267 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 268 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 269 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 270 |
const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8; |
| 271 |
const int8x8_t vk8x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 272 |
const int8x8_t vi8x89ABCDEF = vld1_s8(i8); i8 += 8; |
| 273 |
const int8x8_t vk8x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 274 |
|
| 275 |
vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567); |
| 276 |
vprod89ABCDEF = vmull_s8(vi8x89ABCDEF, vk8x89ABCDEF); |
| 277 |
|
| 278 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 279 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 280 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 281 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 282 |
const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8; |
| 283 |
const int8x8_t vk9x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 284 |
const int8x8_t vi9x89ABCDEF = vld1_s8(i9); i9 += 8; |
| 285 |
const int8x8_t vk9x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 286 |
|
| 287 |
vprod01234567 = vmull_s8(vi9x01234567, vk9x01234567); |
| 288 |
vprod89ABCDEF = vmull_s8(vi9x89ABCDEF, vk9x89ABCDEF); |
| 289 |
|
| 290 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 291 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 292 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 293 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 294 |
const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8; |
| 295 |
const int8x8_t vk10x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 296 |
const int8x8_t vi10x89ABCDEF = vld1_s8(i10); i10 += 8; |
| 297 |
const int8x8_t vk10x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 298 |
|
| 299 |
vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567); |
| 300 |
vprod89ABCDEF = vmull_s8(vi10x89ABCDEF, vk10x89ABCDEF); |
| 301 |
|
| 302 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 303 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 304 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 305 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 306 |
const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8; |
| 307 |
const int8x8_t vk11x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 308 |
const int8x8_t vi11x89ABCDEF = vld1_s8(i11); i11 += 8; |
| 309 |
const int8x8_t vk11x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 310 |
|
| 311 |
vprod01234567 = vmull_s8(vi11x01234567, vk11x01234567); |
| 312 |
vprod89ABCDEF = vmull_s8(vi11x89ABCDEF, vk11x89ABCDEF); |
| 313 |
|
| 314 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 315 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 316 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 317 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 318 |
const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8; |
| 319 |
const int8x8_t vk12x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 320 |
const int8x8_t vi12x89ABCDEF = vld1_s8(i12); i12 += 8; |
| 321 |
const int8x8_t vk12x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 322 |
|
| 323 |
vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567); |
| 324 |
vprod89ABCDEF = vmull_s8(vi12x89ABCDEF, vk12x89ABCDEF); |
| 325 |
|
| 326 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 327 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 328 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 329 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 330 |
const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8; |
| 331 |
const int8x8_t vk13x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 332 |
const int8x8_t vi13x89ABCDEF = vld1_s8(i13); i13 += 8; |
| 333 |
const int8x8_t vk13x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 334 |
|
| 335 |
vprod01234567 = vmull_s8(vi13x01234567, vk13x01234567); |
| 336 |
vprod89ABCDEF = vmull_s8(vi13x89ABCDEF, vk13x89ABCDEF); |
| 337 |
|
| 338 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 339 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 340 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 341 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 342 |
const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8; |
| 343 |
const int8x8_t vk14x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 344 |
const int8x8_t vi14x89ABCDEF = vld1_s8(i14); i14 += 8; |
| 345 |
const int8x8_t vk14x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 346 |
|
| 347 |
vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567); |
| 348 |
vprod89ABCDEF = vmull_s8(vi14x89ABCDEF, vk14x89ABCDEF); |
| 349 |
|
| 350 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 351 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 352 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 353 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 354 |
const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8; |
| 355 |
const int8x8_t vk15x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 356 |
const int8x8_t vi15x89ABCDEF = vld1_s8(i15); i15 += 8; |
| 357 |
const int8x8_t vk15x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 358 |
|
| 359 |
vprod01234567 = vmull_s8(vi15x01234567, vk15x01234567); |
| 360 |
vprod89ABCDEF = vmull_s8(vi15x89ABCDEF, vk15x89ABCDEF); |
| 361 |
|
| 362 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 363 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 364 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 365 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 366 |
const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8; |
| 367 |
const int8x8_t vk16x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 368 |
const int8x8_t vi16x89ABCDEF = vld1_s8(i16); i16 += 8; |
| 369 |
const int8x8_t vk16x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 370 |
|
| 371 |
vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567); |
| 372 |
vprod89ABCDEF = vmull_s8(vi16x89ABCDEF, vk16x89ABCDEF); |
| 373 |
|
| 374 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 375 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 376 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 377 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 378 |
const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8; |
| 379 |
const int8x8_t vk17x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 380 |
const int8x8_t vi17x89ABCDEF = vld1_s8(i17); i17 += 8; |
| 381 |
const int8x8_t vk17x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 382 |
|
| 383 |
vprod01234567 = vmull_s8(vi17x01234567, vk17x01234567); |
| 384 |
vprod89ABCDEF = vmull_s8(vi17x89ABCDEF, vk17x89ABCDEF); |
| 385 |
|
| 386 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 387 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 388 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 389 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 390 |
const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8; |
| 391 |
const int8x8_t vk18x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 392 |
const int8x8_t vi18x89ABCDEF = vld1_s8(i18); i18 += 8; |
| 393 |
const int8x8_t vk18x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 394 |
|
| 395 |
vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567); |
| 396 |
vprod89ABCDEF = vmull_s8(vi18x89ABCDEF, vk18x89ABCDEF); |
| 397 |
|
| 398 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 399 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 400 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 401 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 402 |
const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8; |
| 403 |
const int8x8_t vk19x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 404 |
const int8x8_t vi19x89ABCDEF = vld1_s8(i19); i19 += 8; |
| 405 |
const int8x8_t vk19x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 406 |
|
| 407 |
vprod01234567 = vmull_s8(vi19x01234567, vk19x01234567); |
| 408 |
vprod89ABCDEF = vmull_s8(vi19x89ABCDEF, vk19x89ABCDEF); |
| 409 |
|
| 410 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 411 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 412 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 413 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 414 |
const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8; |
| 415 |
const int8x8_t vk20x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 416 |
const int8x8_t vi20x89ABCDEF = vld1_s8(i20); i20 += 8; |
| 417 |
const int8x8_t vk20x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 418 |
|
| 419 |
vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567); |
| 420 |
vprod89ABCDEF = vmull_s8(vi20x89ABCDEF, vk20x89ABCDEF); |
| 421 |
|
| 422 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 423 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 424 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 425 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 426 |
const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8; |
| 427 |
const int8x8_t vk21x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 428 |
const int8x8_t vi21x89ABCDEF = vld1_s8(i21); i21 += 8; |
| 429 |
const int8x8_t vk21x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 430 |
|
| 431 |
vprod01234567 = vmull_s8(vi21x01234567, vk21x01234567); |
| 432 |
vprod89ABCDEF = vmull_s8(vi21x89ABCDEF, vk21x89ABCDEF); |
| 433 |
|
| 434 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 435 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 436 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 437 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 438 |
const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8; |
| 439 |
const int8x8_t vk22x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 440 |
const int8x8_t vi22x89ABCDEF = vld1_s8(i22); i22 += 8; |
| 441 |
const int8x8_t vk22x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 442 |
|
| 443 |
vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567); |
| 444 |
vprod89ABCDEF = vmull_s8(vi22x89ABCDEF, vk22x89ABCDEF); |
| 445 |
|
| 446 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 447 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 448 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 449 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 450 |
const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8; |
| 451 |
const int8x8_t vk23x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 452 |
const int8x8_t vi23x89ABCDEF = vld1_s8(i23); i23 += 8; |
| 453 |
const int8x8_t vk23x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 454 |
|
| 455 |
vprod01234567 = vmull_s8(vi23x01234567, vk23x01234567); |
| 456 |
vprod89ABCDEF = vmull_s8(vi23x89ABCDEF, vk23x89ABCDEF); |
| 457 |
|
| 458 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 459 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 460 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 461 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 462 |
const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8; |
| 463 |
const int8x8_t vk24x01234567 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 464 |
const int8x8_t vi24x89ABCDEF = vld1_s8(i24); i24 += 8; |
| 465 |
const int8x8_t vk24x89ABCDEF = vld1_s8(w); w = (const int8_t*) w + 8; |
| 466 |
|
| 467 |
vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567); |
| 468 |
vprod89ABCDEF = vmull_s8(vi24x89ABCDEF, vk24x89ABCDEF); |
| 469 |
|
| 470 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 471 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 472 |
vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF)); |
| 473 |
vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF)); |
| 474 |
|
| 475 |
vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift); |
| 476 |
vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift); |
| 477 |
vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift); |
| 478 |
vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift); |
| 479 |
|
| 480 |
vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier); |
| 481 |
vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier); |
| 482 |
vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier); |
| 483 |
vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier); |
| 484 |
|
| 485 |
vacc0123 = vrshlq_s32(vacc0123, vright_post_shift); |
| 486 |
vacc4567 = vrshlq_s32(vacc4567, vright_post_shift); |
| 487 |
vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift); |
| 488 |
vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift); |
| 489 |
|
| 490 |
#if XNN_ARCH_ARM64 |
| 491 |
int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); |
| 492 |
int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF); |
| 493 |
|
| 494 |
vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point); |
| 495 |
vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point); |
| 496 |
|
| 497 |
int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF); |
| 498 |
#else // !XNN_ARCH_ARM64 |
| 499 |
int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); |
| 500 |
int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)); |
| 501 |
|
| 502 |
vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point); |
| 503 |
vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point); |
| 504 |
|
| 505 |
int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); |
| 506 |
#endif // !XNN_ARCH_ARM64 |
| 507 |
|
| 508 |
vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min); |
| 509 |
|
| 510 |
vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max); |
| 511 |
|
| 512 |
vst1q_s8(output, vout0123456789ABCDEF); output += 16; |
| 513 |
} |
| 514 |
if XNN_UNLIKELY(c != 0) { |
| 515 |
const int8_t* k = (const int8_t*) ((const int32_t*) w + 16); |
| 516 |
do { |
| 517 |
int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4; |
| 518 |
int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4; |
| 519 |
|
| 520 |
const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; |
| 521 |
const int8x8_t vk0x01234567 = vld1_s8(k); k += 8; |
| 522 |
|
| 523 |
int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567); |
| 524 |
|
| 525 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 526 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 527 |
const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; |
| 528 |
const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8)); |
| 529 |
|
| 530 |
vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567); |
| 531 |
|
| 532 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 533 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 534 |
const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; |
| 535 |
const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24)); |
| 536 |
|
| 537 |
vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567); |
| 538 |
|
| 539 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 540 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 541 |
const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8; |
| 542 |
const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40)); |
| 543 |
|
| 544 |
vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567); |
| 545 |
|
| 546 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 547 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 548 |
const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8; |
| 549 |
const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56)); |
| 550 |
|
| 551 |
vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567); |
| 552 |
|
| 553 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 554 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 555 |
const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8; |
| 556 |
const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72)); |
| 557 |
|
| 558 |
vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567); |
| 559 |
|
| 560 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 561 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 562 |
const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8; |
| 563 |
const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88)); |
| 564 |
|
| 565 |
vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567); |
| 566 |
|
| 567 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 568 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 569 |
const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8; |
| 570 |
const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104)); |
| 571 |
|
| 572 |
vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567); |
| 573 |
|
| 574 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 575 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 576 |
const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8; |
| 577 |
const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120)); |
| 578 |
|
| 579 |
vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567); |
| 580 |
|
| 581 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 582 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 583 |
const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8; |
| 584 |
const int8x8_t vk9x01234567 = vld1_s8((const void*) (k + 136)); |
| 585 |
|
| 586 |
vprod01234567 = vmull_s8(vi9x01234567, vk9x01234567); |
| 587 |
|
| 588 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 589 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 590 |
const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8; |
| 591 |
const int8x8_t vk10x01234567 = vld1_s8((const void*) (k + 152)); |
| 592 |
|
| 593 |
vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567); |
| 594 |
|
| 595 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 596 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 597 |
const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8; |
| 598 |
const int8x8_t vk11x01234567 = vld1_s8((const void*) (k + 168)); |
| 599 |
|
| 600 |
vprod01234567 = vmull_s8(vi11x01234567, vk11x01234567); |
| 601 |
|
| 602 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 603 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 604 |
const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8; |
| 605 |
const int8x8_t vk12x01234567 = vld1_s8((const void*) (k + 184)); |
| 606 |
|
| 607 |
vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567); |
| 608 |
|
| 609 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 610 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 611 |
const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8; |
| 612 |
const int8x8_t vk13x01234567 = vld1_s8((const void*) (k + 200)); |
| 613 |
|
| 614 |
vprod01234567 = vmull_s8(vi13x01234567, vk13x01234567); |
| 615 |
|
| 616 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 617 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 618 |
const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8; |
| 619 |
const int8x8_t vk14x01234567 = vld1_s8((const void*) (k + 216)); |
| 620 |
|
| 621 |
vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567); |
| 622 |
|
| 623 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 624 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 625 |
const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8; |
| 626 |
const int8x8_t vk15x01234567 = vld1_s8((const void*) (k + 232)); |
| 627 |
|
| 628 |
vprod01234567 = vmull_s8(vi15x01234567, vk15x01234567); |
| 629 |
|
| 630 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 631 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 632 |
const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8; |
| 633 |
const int8x8_t vk16x01234567 = vld1_s8((const void*) (k + 248)); |
| 634 |
|
| 635 |
vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567); |
| 636 |
|
| 637 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 638 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 639 |
const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8; |
| 640 |
const int8x8_t vk17x01234567 = vld1_s8((const void*) (k + 264)); |
| 641 |
|
| 642 |
vprod01234567 = vmull_s8(vi17x01234567, vk17x01234567); |
| 643 |
|
| 644 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 645 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 646 |
const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8; |
| 647 |
const int8x8_t vk18x01234567 = vld1_s8((const void*) (k + 280)); |
| 648 |
|
| 649 |
vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567); |
| 650 |
|
| 651 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 652 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 653 |
const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8; |
| 654 |
const int8x8_t vk19x01234567 = vld1_s8((const void*) (k + 296)); |
| 655 |
|
| 656 |
vprod01234567 = vmull_s8(vi19x01234567, vk19x01234567); |
| 657 |
|
| 658 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 659 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 660 |
const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8; |
| 661 |
const int8x8_t vk20x01234567 = vld1_s8((const void*) (k + 312)); |
| 662 |
|
| 663 |
vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567); |
| 664 |
|
| 665 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 666 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 667 |
const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8; |
| 668 |
const int8x8_t vk21x01234567 = vld1_s8((const void*) (k + 328)); |
| 669 |
|
| 670 |
vprod01234567 = vmull_s8(vi21x01234567, vk21x01234567); |
| 671 |
|
| 672 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 673 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 674 |
const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8; |
| 675 |
const int8x8_t vk22x01234567 = vld1_s8((const void*) (k + 344)); |
| 676 |
|
| 677 |
vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567); |
| 678 |
|
| 679 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 680 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 681 |
const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8; |
| 682 |
const int8x8_t vk23x01234567 = vld1_s8((const void*) (k + 360)); |
| 683 |
|
| 684 |
vprod01234567 = vmull_s8(vi23x01234567, vk23x01234567); |
| 685 |
|
| 686 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 687 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 688 |
const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8; |
| 689 |
const int8x8_t vk24x01234567 = vld1_s8((const void*) (k + 376)); |
| 690 |
|
| 691 |
vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567); |
| 692 |
|
| 693 |
vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567)); |
| 694 |
vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567)); |
| 695 |
|
| 696 |
vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift); |
| 697 |
vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift); |
| 698 |
|
| 699 |
vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier); |
| 700 |
vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier); |
| 701 |
|
| 702 |
vacc0123 = vrshlq_s32(vacc0123, vright_post_shift); |
| 703 |
vacc4567 = vrshlq_s32(vacc4567, vright_post_shift); |
| 704 |
|
| 705 |
#if XNN_ARCH_ARM64 |
| 706 |
int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); |
| 707 |
#else |
| 708 |
int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); |
| 709 |
#endif |
| 710 |
vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point); |
| 711 |
|
| 712 |
int8x8_t vout01234567 = vqmovn_s16(vacc01234567); |
| 713 |
vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min)); |
| 714 |
vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max)); |
| 715 |
|
| 716 |
if XNN_LIKELY(c >= 8) { |
| 717 |
vst1_s8(output, vout01234567); output += 8; |
| 718 |
c -= 8; |
| 719 |
} else { |
| 720 |
if (c & 4) { |
| 721 |
vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; |
| 722 |
vout01234567 = vext_s8(vout01234567, vout01234567, 4); |
| 723 |
} |
| 724 |
if (c & 2) { |
| 725 |
vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2; |
| 726 |
vout01234567 = vext_s8(vout01234567, vout01234567, 2); |
| 727 |
} |
| 728 |
if (c & 1) { |
| 729 |
vst1_lane_s8(output, vout01234567, 0); output += 1; |
| 730 |
} |
| 731 |
c = 0; |
| 732 |
} |
| 733 |
} while (c != 0); |
| 734 |
} |
| 735 |
|
| 736 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 737 |
} while (--output_width != 0); |
| 738 |
} |