| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-igemm/neon-mlal-lane.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 |
|
| 11 |
#include <assert.h> |
| 12 |
|
| 13 |
#include <arm_neon.h> |
| 14 |
|
| 15 |
#include "xnnpack/common.h" |
| 16 |
#include "xnnpack/igemm.h" |
| 17 |
#include "xnnpack/prefetch.h" |
| 18 |
|
| 19 |
|
| 20 |
void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x16__neon_mlal_lane_prfm( |
| 21 |
size_t mr, |
| 22 |
size_t nc, |
| 23 |
size_t kc, |
| 24 |
size_t ks, |
| 25 |
const int8_t** restrict a, |
| 26 |
const void* restrict w, |
| 27 |
int8_t* restrict c, |
| 28 |
size_t cm_stride, |
| 29 |
size_t cn_stride, |
| 30 |
size_t a_offset, |
| 31 |
const int8_t* zero, |
| 32 |
const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 33 |
{ |
| 34 |
assert(mr != 0); |
| 35 |
assert(mr <= 1); |
| 36 |
assert(nc != 0); |
| 37 |
assert(kc != 0); |
| 38 |
assert(ks != 0); |
| 39 |
assert(ks % (1 * sizeof(void*)) == 0); |
| 40 |
assert(a_offset % sizeof(int8_t) == 0); |
| 41 |
assert(a != NULL); |
| 42 |
assert(w != NULL); |
| 43 |
assert(c != NULL); |
| 44 |
|
| 45 |
int8_t* c0 = c; |
| 46 |
|
| 47 |
do { |
| 48 |
int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); |
| 49 |
int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); |
| 50 |
int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); |
| 51 |
int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); |
| 52 |
|
| 53 |
size_t p = ks; |
| 54 |
do { |
| 55 |
const int8_t* restrict a0 = a[0]; |
| 56 |
if XNN_UNPREDICTABLE(a0 != zero) { |
| 57 |
a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 58 |
} |
| 59 |
a += 1; |
| 60 |
|
| 61 |
size_t k = kc; |
| 62 |
while (k >= 8 * sizeof(int8_t)) { |
| 63 |
const int8x8_t va0 = vld1_s8(a0); a0 += 8; |
| 64 |
const int16x8_t vxa0 = vmovl_s8(va0); |
| 65 |
|
| 66 |
const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 67 |
const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0); |
| 68 |
|
| 69 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); |
| 70 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); |
| 71 |
const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 72 |
const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0); |
| 73 |
|
| 74 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); |
| 75 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); |
| 76 |
const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 77 |
const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1); |
| 78 |
|
| 79 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); |
| 80 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); |
| 81 |
const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 82 |
const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1); |
| 83 |
|
| 84 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1); |
| 85 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1); |
| 86 |
const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 87 |
const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2); |
| 88 |
|
| 89 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); |
| 90 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); |
| 91 |
const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 92 |
const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2); |
| 93 |
|
| 94 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2); |
| 95 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2); |
| 96 |
const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 97 |
const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3); |
| 98 |
|
| 99 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3); |
| 100 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); |
| 101 |
const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 102 |
const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3); |
| 103 |
|
| 104 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3); |
| 105 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3); |
| 106 |
|
| 107 |
xnn_prefetch_to_l1((const int8_t*) w + 448); |
| 108 |
xnn_prefetch_to_l1((const int8_t*) w + 512); |
| 109 |
|
| 110 |
const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 111 |
const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4); |
| 112 |
|
| 113 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); |
| 114 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); |
| 115 |
const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 116 |
const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4); |
| 117 |
|
| 118 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0); |
| 119 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0); |
| 120 |
const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 121 |
const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5); |
| 122 |
|
| 123 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); |
| 124 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); |
| 125 |
const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 126 |
const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5); |
| 127 |
|
| 128 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1); |
| 129 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1); |
| 130 |
const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 131 |
const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6); |
| 132 |
|
| 133 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); |
| 134 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); |
| 135 |
const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 136 |
const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6); |
| 137 |
|
| 138 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2); |
| 139 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2); |
| 140 |
const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 141 |
const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7); |
| 142 |
|
| 143 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3); |
| 144 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3); |
| 145 |
const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 146 |
const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7); |
| 147 |
|
| 148 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3); |
| 149 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3); |
| 150 |
|
| 151 |
k -= 8 * sizeof(int8_t); |
| 152 |
} |
| 153 |
if XNN_UNLIKELY(k != 0) { |
| 154 |
const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k); |
| 155 |
const int16x8_t vxa0 = vmovl_s8(va0); |
| 156 |
|
| 157 |
const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 158 |
const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0); |
| 159 |
const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 160 |
const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0); |
| 161 |
|
| 162 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); |
| 163 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); |
| 164 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); |
| 165 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); |
| 166 |
|
| 167 |
if (k >= 2 * sizeof(int8_t)) { |
| 168 |
const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 169 |
const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1); |
| 170 |
const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 171 |
const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1); |
| 172 |
|
| 173 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); |
| 174 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); |
| 175 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1); |
| 176 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1); |
| 177 |
|
| 178 |
if (k > 2 * sizeof(int8_t)) { |
| 179 |
const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 180 |
const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2); |
| 181 |
const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 182 |
const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2); |
| 183 |
|
| 184 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); |
| 185 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); |
| 186 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2); |
| 187 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2); |
| 188 |
|
| 189 |
if (k >= 4 * sizeof(int8_t)) { |
| 190 |
const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 191 |
const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3); |
| 192 |
const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 193 |
const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3); |
| 194 |
|
| 195 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3); |
| 196 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); |
| 197 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3); |
| 198 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3); |
| 199 |
|
| 200 |
if (k > 4 * sizeof(int8_t)) { |
| 201 |
const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 202 |
const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4); |
| 203 |
const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 204 |
const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4); |
| 205 |
|
| 206 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); |
| 207 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); |
| 208 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0); |
| 209 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0); |
| 210 |
|
| 211 |
if (k >= 6 * sizeof(int8_t)) { |
| 212 |
const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 213 |
const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5); |
| 214 |
const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 215 |
const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5); |
| 216 |
|
| 217 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); |
| 218 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); |
| 219 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1); |
| 220 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1); |
| 221 |
|
| 222 |
if (k > 6 * sizeof(int8_t)) { |
| 223 |
const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 224 |
const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6); |
| 225 |
const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 226 |
const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6); |
| 227 |
|
| 228 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); |
| 229 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); |
| 230 |
vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2); |
| 231 |
vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2); |
| 232 |
} |
| 233 |
} |
| 234 |
} |
| 235 |
} |
| 236 |
} |
| 237 |
} |
| 238 |
} |
| 239 |
p -= 1 * sizeof(void*); |
| 240 |
} while (p != 0); |
| 241 |
|
| 242 |
// Post-accumulation work |
| 243 |
float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123); |
| 244 |
float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567); |
| 245 |
float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB); |
| 246 |
float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF); |
| 247 |
|
| 248 |
const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 249 |
vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123); |
| 250 |
const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 251 |
vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567); |
| 252 |
const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 253 |
vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB); |
| 254 |
const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 255 |
vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF); |
| 256 |
|
| 257 |
const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias); |
| 258 |
vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias)); |
| 259 |
vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias)); |
| 260 |
vacc0x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x89AB, vmagic_bias)); |
| 261 |
vacc0xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc0xCDEF, vmagic_bias)); |
| 262 |
|
| 263 |
const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point); |
| 264 |
vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point); |
| 265 |
vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point); |
| 266 |
vacc0x89AB = vqsubq_s32(vacc0x89AB, vmagic_bias_less_output_zero_point); |
| 267 |
vacc0xCDEF = vqsubq_s32(vacc0xCDEF, vmagic_bias_less_output_zero_point); |
| 268 |
|
| 269 |
#if XNN_ARCH_ARM64 |
| 270 |
int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567); |
| 271 |
int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF); |
| 272 |
|
| 273 |
|
| 274 |
int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); |
| 275 |
#else |
| 276 |
int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)); |
| 277 |
int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)); |
| 278 |
|
| 279 |
|
| 280 |
int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF)); |
| 281 |
#endif |
| 282 |
|
| 283 |
const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min); |
| 284 |
vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min); |
| 285 |
|
| 286 |
const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max); |
| 287 |
vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max); |
| 288 |
|
| 289 |
if (nc >= 16) { |
| 290 |
vst1q_s8(c0 + 0, vout0x0123456789ABCDEF); |
| 291 |
|
| 292 |
c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 293 |
|
| 294 |
a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 295 |
|
| 296 |
nc -= 16; |
| 297 |
} else { |
| 298 |
int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF); |
| 299 |
if (nc & 8) { |
| 300 |
vst1_s8(c0, vout0x01234567); c0 += 8; |
| 301 |
vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF); |
| 302 |
} |
| 303 |
if (nc & 4) { |
| 304 |
vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4; |
| 305 |
vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4); |
| 306 |
} |
| 307 |
if (nc & 2) { |
| 308 |
vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2; |
| 309 |
vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2); |
| 310 |
} |
| 311 |
if (nc & 1) { |
| 312 |
vst1_lane_s8(c0, vout0x01234567, 0); |
| 313 |
} |
| 314 |
|
| 315 |
nc = 0; |
| 316 |
} |
| 317 |
} while (nc != 0); |
| 318 |
} |