| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-igemm/MRx16c4-wasmdot.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2024 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 <wasm_simd128.h> |
| 13 |
|
| 14 |
#include "xnnpack/gemm.h" |
| 15 |
#include "xnnpack/math.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x16c4__wasmsdot_u2( |
| 19 |
size_t mr, |
| 20 |
size_t nc, |
| 21 |
size_t kc, |
| 22 |
size_t ks, |
| 23 |
const int8_t** restrict a, |
| 24 |
const void* restrict w, |
| 25 |
int8_t* restrict c, |
| 26 |
size_t cm_stride, |
| 27 |
size_t cn_stride, |
| 28 |
size_t a_offset, |
| 29 |
const int8_t* zero, |
| 30 |
const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 31 |
{ |
| 32 |
assert(mr != 0); |
| 33 |
assert(mr <= 4); |
| 34 |
assert(nc != 0); |
| 35 |
assert(kc != 0); |
| 36 |
assert(ks != 0); |
| 37 |
assert(ks % (4 * sizeof(void*)) == 0); |
| 38 |
assert(a_offset % sizeof(int8_t) == 0); |
| 39 |
assert(a != NULL); |
| 40 |
assert(w != NULL); |
| 41 |
assert(c != NULL); |
| 42 |
|
| 43 |
kc = round_up_po2(kc, 4 * sizeof(int8_t)); |
| 44 |
int8_t* c0 = c; |
| 45 |
int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 46 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 47 |
c1 = c0; |
| 48 |
} |
| 49 |
int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); |
| 50 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 51 |
c2 = c1; |
| 52 |
} |
| 53 |
int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride); |
| 54 |
if XNN_UNPREDICTABLE(mr != 4) { |
| 55 |
c3 = c2; |
| 56 |
} |
| 57 |
|
| 58 |
const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f); |
| 59 |
const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point; |
| 60 |
const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point)); |
| 61 |
const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point); |
| 62 |
const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max); |
| 63 |
//XNN_FORCE_REALIZATION(vmagic_bias); |
| 64 |
//XNN_FORCE_REALIZATION(vmagic_min); |
| 65 |
//XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point); |
| 66 |
//XNN_FORCE_REALIZATION(voutput_max); |
| 67 |
|
| 68 |
do { |
| 69 |
v128_t vacc0x0x0123 = wasm_v128_load(w); |
| 70 |
v128_t vacc0x0x4567 = wasm_v128_load((const int32_t*) w + 4); |
| 71 |
v128_t vacc0x0x89AB = wasm_v128_load((const int32_t*) w + 8); |
| 72 |
v128_t vacc0x0xCDEF = wasm_v128_load((const int32_t*) w + 12); |
| 73 |
v128_t vacc1x0x0123= vacc0x0x0123; |
| 74 |
v128_t vacc1x0x4567= vacc0x0x4567; |
| 75 |
v128_t vacc1x0x89AB= vacc0x0x89AB; |
| 76 |
v128_t vacc1x0xCDEF= vacc0x0xCDEF; |
| 77 |
v128_t vacc2x0x0123= vacc0x0x0123; |
| 78 |
v128_t vacc2x0x4567= vacc0x0x4567; |
| 79 |
v128_t vacc2x0x89AB= vacc0x0x89AB; |
| 80 |
v128_t vacc2x0xCDEF= vacc0x0xCDEF; |
| 81 |
v128_t vacc3x0x0123= vacc0x0x0123; |
| 82 |
v128_t vacc3x0x4567= vacc0x0x4567; |
| 83 |
v128_t vacc3x0x89AB= vacc0x0x89AB; |
| 84 |
v128_t vacc3x0xCDEF= vacc0x0xCDEF; |
| 85 |
w = (const int32_t*) w + 16; |
| 86 |
|
| 87 |
size_t p = ks; |
| 88 |
do { |
| 89 |
const int8_t* restrict a0 = a[0]; |
| 90 |
if XNN_UNPREDICTABLE(a0 != zero) { |
| 91 |
a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 92 |
} |
| 93 |
const int8_t* restrict a1 = a[1]; |
| 94 |
if XNN_UNPREDICTABLE(a1 != zero) { |
| 95 |
a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); |
| 96 |
} |
| 97 |
const int8_t* restrict a2 = a[2]; |
| 98 |
if XNN_UNPREDICTABLE(a2 != zero) { |
| 99 |
a2 = (const int8_t*) ((uintptr_t) a2 + a_offset); |
| 100 |
} |
| 101 |
const int8_t* restrict a3 = a[3]; |
| 102 |
if XNN_UNPREDICTABLE(a3 != zero) { |
| 103 |
a3 = (const int8_t*) ((uintptr_t) a3 + a_offset); |
| 104 |
} |
| 105 |
a += 4; |
| 106 |
|
| 107 |
size_t k = kc; |
| 108 |
while (k >= 8 * sizeof(int8_t)) { |
| 109 |
v128_t va0x0x0123 = wasm_v128_load32_splat(a0); |
| 110 |
v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4); |
| 111 |
a0 += 8; |
| 112 |
v128_t va1x0x0123 = wasm_v128_load32_splat(a1); |
| 113 |
v128_t va1x1x0123 = wasm_v128_load32_splat(a1 + 4); |
| 114 |
a1 += 8; |
| 115 |
v128_t va2x0x0123 = wasm_v128_load32_splat(a2); |
| 116 |
v128_t va2x1x0123 = wasm_v128_load32_splat(a2 + 4); |
| 117 |
a2 += 8; |
| 118 |
v128_t va3x0x0123 = wasm_v128_load32_splat(a3); |
| 119 |
v128_t va3x1x0123 = wasm_v128_load32_splat(a3 + 4); |
| 120 |
a3 += 8; |
| 121 |
|
| 122 |
|
| 123 |
const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w); |
| 124 |
const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16); |
| 125 |
const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32); |
| 126 |
const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48); |
| 127 |
const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64); |
| 128 |
const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80); |
| 129 |
const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96); |
| 130 |
const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112); |
| 131 |
|
| 132 |
vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123); |
| 133 |
vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567); |
| 134 |
vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB); |
| 135 |
vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF); |
| 136 |
vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va1x0x0123, vacc1x0x0123); |
| 137 |
vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va1x0x0123, vacc1x0x4567); |
| 138 |
vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va1x0x0123, vacc1x0x89AB); |
| 139 |
vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va1x0x0123, vacc1x0xCDEF); |
| 140 |
vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va2x0x0123, vacc2x0x0123); |
| 141 |
vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va2x0x0123, vacc2x0x4567); |
| 142 |
vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va2x0x0123, vacc2x0x89AB); |
| 143 |
vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va2x0x0123, vacc2x0xCDEF); |
| 144 |
vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va3x0x0123, vacc3x0x0123); |
| 145 |
vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va3x0x0123, vacc3x0x4567); |
| 146 |
vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va3x0x0123, vacc3x0x89AB); |
| 147 |
vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va3x0x0123, vacc3x0xCDEF); |
| 148 |
vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x0x0123); |
| 149 |
vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x0x4567); |
| 150 |
vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x0x89AB); |
| 151 |
vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x0xCDEF); |
| 152 |
vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va1x1x0123, vacc1x0x0123); |
| 153 |
vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va1x1x0123, vacc1x0x4567); |
| 154 |
vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va1x1x0123, vacc1x0x89AB); |
| 155 |
vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va1x1x0123, vacc1x0xCDEF); |
| 156 |
vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va2x1x0123, vacc2x0x0123); |
| 157 |
vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va2x1x0123, vacc2x0x4567); |
| 158 |
vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va2x1x0123, vacc2x0x89AB); |
| 159 |
vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va2x1x0123, vacc2x0xCDEF); |
| 160 |
vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va3x1x0123, vacc3x0x0123); |
| 161 |
vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va3x1x0123, vacc3x0x4567); |
| 162 |
vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va3x1x0123, vacc3x0x89AB); |
| 163 |
vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va3x1x0123, vacc3x0xCDEF); |
| 164 |
|
| 165 |
w = (const int8_t*) w + 128; |
| 166 |
k -= 8 * sizeof(int8_t); |
| 167 |
} |
| 168 |
|
| 169 |
while (k != 0) { |
| 170 |
v128_t va0x0123 = wasm_v128_load32_splat(a0); |
| 171 |
a0 += 4; |
| 172 |
v128_t va1x0123 = wasm_v128_load32_splat(a1); |
| 173 |
a1 += 4; |
| 174 |
v128_t va2x0123 = wasm_v128_load32_splat(a2); |
| 175 |
a2 += 4; |
| 176 |
v128_t va3x0123 = wasm_v128_load32_splat(a3); |
| 177 |
a3 += 4; |
| 178 |
|
| 179 |
|
| 180 |
const v128_t vb0123 = wasm_v128_load((const int8_t*) w); |
| 181 |
const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16); |
| 182 |
const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32); |
| 183 |
const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48); |
| 184 |
|
| 185 |
vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123); |
| 186 |
vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567); |
| 187 |
vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB); |
| 188 |
vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF); |
| 189 |
vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0x0123); |
| 190 |
vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x0x4567); |
| 191 |
vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x0x89AB); |
| 192 |
vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1x0xCDEF); |
| 193 |
vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0x0123); |
| 194 |
vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x0x4567); |
| 195 |
vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x0x89AB); |
| 196 |
vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2x0xCDEF); |
| 197 |
vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va3x0123, vacc3x0x0123); |
| 198 |
vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va3x0123, vacc3x0x4567); |
| 199 |
vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va3x0123, vacc3x0x89AB); |
| 200 |
vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va3x0123, vacc3x0xCDEF); |
| 201 |
|
| 202 |
w = (const int8_t*) w + 64; |
| 203 |
k -= 4 * sizeof(int8_t); |
| 204 |
} |
| 205 |
|
| 206 |
p -= 4 * sizeof(void*); |
| 207 |
} while (p != 0); |
| 208 |
|
| 209 |
vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123); |
| 210 |
vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567); |
| 211 |
vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB); |
| 212 |
vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF); |
| 213 |
vacc1x0x0123 = wasm_f32x4_convert_i32x4(vacc1x0x0123); |
| 214 |
vacc1x0x4567 = wasm_f32x4_convert_i32x4(vacc1x0x4567); |
| 215 |
vacc1x0x89AB = wasm_f32x4_convert_i32x4(vacc1x0x89AB); |
| 216 |
vacc1x0xCDEF = wasm_f32x4_convert_i32x4(vacc1x0xCDEF); |
| 217 |
vacc2x0x0123 = wasm_f32x4_convert_i32x4(vacc2x0x0123); |
| 218 |
vacc2x0x4567 = wasm_f32x4_convert_i32x4(vacc2x0x4567); |
| 219 |
vacc2x0x89AB = wasm_f32x4_convert_i32x4(vacc2x0x89AB); |
| 220 |
vacc2x0xCDEF = wasm_f32x4_convert_i32x4(vacc2x0xCDEF); |
| 221 |
vacc3x0x0123 = wasm_f32x4_convert_i32x4(vacc3x0x0123); |
| 222 |
vacc3x0x4567 = wasm_f32x4_convert_i32x4(vacc3x0x4567); |
| 223 |
vacc3x0x89AB = wasm_f32x4_convert_i32x4(vacc3x0x89AB); |
| 224 |
vacc3x0xCDEF = wasm_f32x4_convert_i32x4(vacc3x0xCDEF); |
| 225 |
|
| 226 |
const v128_t vscale0123 = wasm_v128_load(w); |
| 227 |
w = (const float*) w + 4; |
| 228 |
const v128_t vscale4567 = wasm_v128_load(w); |
| 229 |
w = (const float*) w + 4; |
| 230 |
const v128_t vscale89AB = wasm_v128_load(w); |
| 231 |
w = (const float*) w + 4; |
| 232 |
const v128_t vscaleCDEF = wasm_v128_load(w); |
| 233 |
w = (const float*) w + 4; |
| 234 |
vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vscale0123); |
| 235 |
vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vscale4567); |
| 236 |
vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vscale89AB); |
| 237 |
vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vscaleCDEF); |
| 238 |
vacc1x0x0123 = wasm_f32x4_mul(vacc1x0x0123, vscale0123); |
| 239 |
vacc1x0x4567 = wasm_f32x4_mul(vacc1x0x4567, vscale4567); |
| 240 |
vacc1x0x89AB = wasm_f32x4_mul(vacc1x0x89AB, vscale89AB); |
| 241 |
vacc1x0xCDEF = wasm_f32x4_mul(vacc1x0xCDEF, vscaleCDEF); |
| 242 |
vacc2x0x0123 = wasm_f32x4_mul(vacc2x0x0123, vscale0123); |
| 243 |
vacc2x0x4567 = wasm_f32x4_mul(vacc2x0x4567, vscale4567); |
| 244 |
vacc2x0x89AB = wasm_f32x4_mul(vacc2x0x89AB, vscale89AB); |
| 245 |
vacc2x0xCDEF = wasm_f32x4_mul(vacc2x0xCDEF, vscaleCDEF); |
| 246 |
vacc3x0x0123 = wasm_f32x4_mul(vacc3x0x0123, vscale0123); |
| 247 |
vacc3x0x4567 = wasm_f32x4_mul(vacc3x0x4567, vscale4567); |
| 248 |
vacc3x0x89AB = wasm_f32x4_mul(vacc3x0x89AB, vscale89AB); |
| 249 |
vacc3x0xCDEF = wasm_f32x4_mul(vacc3x0xCDEF, vscaleCDEF); |
| 250 |
|
| 251 |
vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vmagic_bias); |
| 252 |
vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vmagic_bias); |
| 253 |
vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vmagic_bias); |
| 254 |
vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vmagic_bias); |
| 255 |
vacc1x0x0123 = wasm_f32x4_add(vacc1x0x0123, vmagic_bias); |
| 256 |
vacc1x0x4567 = wasm_f32x4_add(vacc1x0x4567, vmagic_bias); |
| 257 |
vacc1x0x89AB = wasm_f32x4_add(vacc1x0x89AB, vmagic_bias); |
| 258 |
vacc1x0xCDEF = wasm_f32x4_add(vacc1x0xCDEF, vmagic_bias); |
| 259 |
vacc2x0x0123 = wasm_f32x4_add(vacc2x0x0123, vmagic_bias); |
| 260 |
vacc2x0x4567 = wasm_f32x4_add(vacc2x0x4567, vmagic_bias); |
| 261 |
vacc2x0x89AB = wasm_f32x4_add(vacc2x0x89AB, vmagic_bias); |
| 262 |
vacc2x0xCDEF = wasm_f32x4_add(vacc2x0xCDEF, vmagic_bias); |
| 263 |
vacc3x0x0123 = wasm_f32x4_add(vacc3x0x0123, vmagic_bias); |
| 264 |
vacc3x0x4567 = wasm_f32x4_add(vacc3x0x4567, vmagic_bias); |
| 265 |
vacc3x0x89AB = wasm_f32x4_add(vacc3x0x89AB, vmagic_bias); |
| 266 |
vacc3x0xCDEF = wasm_f32x4_add(vacc3x0xCDEF, vmagic_bias); |
| 267 |
|
| 268 |
vacc0x0x0123 = wasm_i32x4_max(vacc0x0x0123, vmagic_min); |
| 269 |
vacc0x0x4567 = wasm_i32x4_max(vacc0x0x4567, vmagic_min); |
| 270 |
vacc0x0x89AB = wasm_i32x4_max(vacc0x0x89AB, vmagic_min); |
| 271 |
vacc0x0xCDEF = wasm_i32x4_max(vacc0x0xCDEF, vmagic_min); |
| 272 |
vacc1x0x0123 = wasm_i32x4_max(vacc1x0x0123, vmagic_min); |
| 273 |
vacc1x0x4567 = wasm_i32x4_max(vacc1x0x4567, vmagic_min); |
| 274 |
vacc1x0x89AB = wasm_i32x4_max(vacc1x0x89AB, vmagic_min); |
| 275 |
vacc1x0xCDEF = wasm_i32x4_max(vacc1x0xCDEF, vmagic_min); |
| 276 |
vacc2x0x0123 = wasm_i32x4_max(vacc2x0x0123, vmagic_min); |
| 277 |
vacc2x0x4567 = wasm_i32x4_max(vacc2x0x4567, vmagic_min); |
| 278 |
vacc2x0x89AB = wasm_i32x4_max(vacc2x0x89AB, vmagic_min); |
| 279 |
vacc2x0xCDEF = wasm_i32x4_max(vacc2x0xCDEF, vmagic_min); |
| 280 |
vacc3x0x0123 = wasm_i32x4_max(vacc3x0x0123, vmagic_min); |
| 281 |
vacc3x0x4567 = wasm_i32x4_max(vacc3x0x4567, vmagic_min); |
| 282 |
vacc3x0x89AB = wasm_i32x4_max(vacc3x0x89AB, vmagic_min); |
| 283 |
vacc3x0xCDEF = wasm_i32x4_max(vacc3x0xCDEF, vmagic_min); |
| 284 |
|
| 285 |
vacc0x0x0123 = wasm_i32x4_sub(vacc0x0x0123, vmagic_bias_less_output_zero_point); |
| 286 |
vacc0x0x4567 = wasm_i32x4_sub(vacc0x0x4567, vmagic_bias_less_output_zero_point); |
| 287 |
vacc0x0x89AB = wasm_i32x4_sub(vacc0x0x89AB, vmagic_bias_less_output_zero_point); |
| 288 |
vacc0x0xCDEF = wasm_i32x4_sub(vacc0x0xCDEF, vmagic_bias_less_output_zero_point); |
| 289 |
vacc1x0x0123 = wasm_i32x4_sub(vacc1x0x0123, vmagic_bias_less_output_zero_point); |
| 290 |
vacc1x0x4567 = wasm_i32x4_sub(vacc1x0x4567, vmagic_bias_less_output_zero_point); |
| 291 |
vacc1x0x89AB = wasm_i32x4_sub(vacc1x0x89AB, vmagic_bias_less_output_zero_point); |
| 292 |
vacc1x0xCDEF = wasm_i32x4_sub(vacc1x0xCDEF, vmagic_bias_less_output_zero_point); |
| 293 |
vacc2x0x0123 = wasm_i32x4_sub(vacc2x0x0123, vmagic_bias_less_output_zero_point); |
| 294 |
vacc2x0x4567 = wasm_i32x4_sub(vacc2x0x4567, vmagic_bias_less_output_zero_point); |
| 295 |
vacc2x0x89AB = wasm_i32x4_sub(vacc2x0x89AB, vmagic_bias_less_output_zero_point); |
| 296 |
vacc2x0xCDEF = wasm_i32x4_sub(vacc2x0xCDEF, vmagic_bias_less_output_zero_point); |
| 297 |
vacc3x0x0123 = wasm_i32x4_sub(vacc3x0x0123, vmagic_bias_less_output_zero_point); |
| 298 |
vacc3x0x4567 = wasm_i32x4_sub(vacc3x0x4567, vmagic_bias_less_output_zero_point); |
| 299 |
vacc3x0x89AB = wasm_i32x4_sub(vacc3x0x89AB, vmagic_bias_less_output_zero_point); |
| 300 |
vacc3x0xCDEF = wasm_i32x4_sub(vacc3x0xCDEF, vmagic_bias_less_output_zero_point); |
| 301 |
|
| 302 |
v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0x0123, vacc0x0x4567); |
| 303 |
v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x0x89AB, vacc0x0xCDEF); |
| 304 |
v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0x0123, vacc1x0x4567); |
| 305 |
v128_t vacc1x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc1x0x89AB, vacc1x0xCDEF); |
| 306 |
v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0x0123, vacc2x0x4567); |
| 307 |
v128_t vacc2x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc2x0x89AB, vacc2x0xCDEF); |
| 308 |
v128_t vacc3x01234567 = wasm_i16x8_narrow_i32x4(vacc3x0x0123, vacc3x0x4567); |
| 309 |
v128_t vacc3x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc3x0x89AB, vacc3x0xCDEF); |
| 310 |
|
| 311 |
vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567); |
| 312 |
vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF); |
| 313 |
vacc1x01234567 = wasm_i8x16_narrow_i16x8(vacc1x01234567, vacc1x01234567); |
| 314 |
vacc1x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc1x89ABCDEF, vacc1x89ABCDEF); |
| 315 |
vacc2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567); |
| 316 |
vacc2x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc2x89ABCDEF, vacc2x89ABCDEF); |
| 317 |
vacc3x01234567 = wasm_i8x16_narrow_i16x8(vacc3x01234567, vacc3x01234567); |
| 318 |
vacc3x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc3x89ABCDEF, vacc3x89ABCDEF); |
| 319 |
|
| 320 |
vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max); |
| 321 |
vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max); |
| 322 |
vacc1x01234567 = wasm_i8x16_min(vacc1x01234567, voutput_max); |
| 323 |
vacc1x89ABCDEF = wasm_i8x16_min(vacc1x89ABCDEF, voutput_max); |
| 324 |
vacc2x01234567 = wasm_i8x16_min(vacc2x01234567, voutput_max); |
| 325 |
vacc2x89ABCDEF = wasm_i8x16_min(vacc2x89ABCDEF, voutput_max); |
| 326 |
vacc3x01234567 = wasm_i8x16_min(vacc3x01234567, voutput_max); |
| 327 |
vacc3x89ABCDEF = wasm_i8x16_min(vacc3x89ABCDEF, voutput_max); |
| 328 |
|
| 329 |
if (nc >= 16) { |
| 330 |
wasm_v128_store64_lane(c3, vacc3x01234567, 0); |
| 331 |
wasm_v128_store64_lane(c3 + 8, vacc3x89ABCDEF, 0); |
| 332 |
wasm_v128_store64_lane(c2, vacc2x01234567, 0); |
| 333 |
wasm_v128_store64_lane(c2 + 8, vacc2x89ABCDEF, 0); |
| 334 |
wasm_v128_store64_lane(c1, vacc1x01234567, 0); |
| 335 |
wasm_v128_store64_lane(c1 + 8, vacc1x89ABCDEF, 0); |
| 336 |
wasm_v128_store64_lane(c0, vacc0x01234567, 0); |
| 337 |
wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0); |
| 338 |
|
| 339 |
c3 = (int8_t*) ((uintptr_t) c3 + cn_stride); |
| 340 |
c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| 341 |
c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 342 |
c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 343 |
|
| 344 |
a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 345 |
|
| 346 |
nc -= 16; |
| 347 |
} else { |
| 348 |
if (nc & 8) { |
| 349 |
wasm_v128_store64_lane(c3, vacc3x01234567, 0); |
| 350 |
c3 += 8; |
| 351 |
wasm_v128_store64_lane(c2, vacc2x01234567, 0); |
| 352 |
c2 += 8; |
| 353 |
wasm_v128_store64_lane(c1, vacc1x01234567, 0); |
| 354 |
c1 += 8; |
| 355 |
wasm_v128_store64_lane(c0, vacc0x01234567, 0); |
| 356 |
c0 += 8; |
| 357 |
|
| 358 |
vacc0x01234567 = vacc0x89ABCDEF; |
| 359 |
vacc1x01234567 = vacc1x89ABCDEF; |
| 360 |
vacc2x01234567 = vacc2x89ABCDEF; |
| 361 |
vacc3x01234567 = vacc3x89ABCDEF; |
| 362 |
} |
| 363 |
if (nc & 4) { |
| 364 |
wasm_v128_store32_lane(c3, vacc3x01234567, 0); |
| 365 |
c3 += 4; |
| 366 |
wasm_v128_store32_lane(c2, vacc2x01234567, 0); |
| 367 |
c2 += 4; |
| 368 |
wasm_v128_store32_lane(c1, vacc1x01234567, 0); |
| 369 |
c1 += 4; |
| 370 |
wasm_v128_store32_lane(c0, vacc0x01234567, 0); |
| 371 |
c0 += 4; |
| 372 |
|
| 373 |
vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32); |
| 374 |
vacc1x01234567 = wasm_u64x2_shr(vacc1x01234567, 32); |
| 375 |
vacc2x01234567 = wasm_u64x2_shr(vacc2x01234567, 32); |
| 376 |
vacc3x01234567 = wasm_u64x2_shr(vacc3x01234567, 32); |
| 377 |
} |
| 378 |
if (nc & 2) { |
| 379 |
wasm_v128_store16_lane(c3, vacc3x01234567, 0); |
| 380 |
c3 += 2; |
| 381 |
wasm_v128_store16_lane(c2, vacc2x01234567, 0); |
| 382 |
c2 += 2; |
| 383 |
wasm_v128_store16_lane(c1, vacc1x01234567, 0); |
| 384 |
c1 += 2; |
| 385 |
wasm_v128_store16_lane(c0, vacc0x01234567, 0); |
| 386 |
c0 += 2; |
| 387 |
|
| 388 |
vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16); |
| 389 |
vacc1x01234567 = wasm_u32x4_shr(vacc1x01234567, 16); |
| 390 |
vacc2x01234567 = wasm_u32x4_shr(vacc2x01234567, 16); |
| 391 |
vacc3x01234567 = wasm_u32x4_shr(vacc3x01234567, 16); |
| 392 |
} |
| 393 |
if (nc & 1) { |
| 394 |
wasm_v128_store8_lane(c3, vacc3x01234567, 0); |
| 395 |
wasm_v128_store8_lane(c2, vacc2x01234567, 0); |
| 396 |
wasm_v128_store8_lane(c1, vacc1x01234567, 0); |
| 397 |
wasm_v128_store8_lane(c0, vacc0x01234567, 0); |
| 398 |
} |
| 399 |
|
| 400 |
nc = 0; |
| 401 |
} |
| 402 |
} while (nc != 0); |
| 403 |
} |