| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/MRx16c8-avx512vnni.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2023 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
#include <assert.h> |
| 11 |
|
| 12 |
#include <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/gemm.h" |
| 16 |
#include "xnnpack/intrinsics-polyfill.h" |
| 17 |
#include "xnnpack/math.h" |
| 18 |
#include "xnnpack/unaligned.h" |
| 19 |
#include "xnnpack/prefetch.h" |
| 20 |
|
| 21 |
|
| 22 |
void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_14x16c8__avx512vnni_prfm( |
| 23 |
size_t mr, |
| 24 |
size_t nc, |
| 25 |
size_t kc, |
| 26 |
const int8_t* restrict a, |
| 27 |
size_t a_stride, |
| 28 |
const void* restrict w, |
| 29 |
float* restrict c, |
| 30 |
size_t cm_stride, |
| 31 |
size_t cn_stride, |
| 32 |
const struct xnn_f32_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 33 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 34 |
{ |
| 35 |
assert(mr != 0); |
| 36 |
assert(mr <= 14); |
| 37 |
assert(nc != 0); |
| 38 |
assert(kc != 0); |
| 39 |
assert(kc % sizeof(int8_t) == 0); |
| 40 |
assert(a != NULL); |
| 41 |
assert(w != NULL); |
| 42 |
assert(c != NULL); |
| 43 |
|
| 44 |
kc = round_up_po2(kc, 8 * sizeof(int8_t)); |
| 45 |
const int8_t* a0 = a; |
| 46 |
float* c0 = c; |
| 47 |
const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 48 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 49 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 50 |
a1 = a0; |
| 51 |
c1 = c0; |
| 52 |
} |
| 53 |
const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 54 |
float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 55 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 56 |
a2 = a1; |
| 57 |
c2 = c1; |
| 58 |
} |
| 59 |
const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride); |
| 60 |
float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 61 |
if XNN_UNPREDICTABLE(mr < 4) { |
| 62 |
a3 = a2; |
| 63 |
c3 = c2; |
| 64 |
} |
| 65 |
const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride); |
| 66 |
float* c4 = (float*) ((uintptr_t) c3 + cm_stride); |
| 67 |
if XNN_UNPREDICTABLE(mr <= 4) { |
| 68 |
a4 = a3; |
| 69 |
c4 = c3; |
| 70 |
} |
| 71 |
const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride); |
| 72 |
float* c5 = (float*) ((uintptr_t) c4 + cm_stride); |
| 73 |
if XNN_UNPREDICTABLE(mr < 6) { |
| 74 |
a5 = a4; |
| 75 |
c5 = c4; |
| 76 |
} |
| 77 |
const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride); |
| 78 |
float* c6 = (float*) ((uintptr_t) c5 + cm_stride); |
| 79 |
if XNN_UNPREDICTABLE(mr <= 6) { |
| 80 |
a6 = a5; |
| 81 |
c6 = c5; |
| 82 |
} |
| 83 |
const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride); |
| 84 |
float* c7 = (float*) ((uintptr_t) c6 + cm_stride); |
| 85 |
if XNN_UNPREDICTABLE(mr < 8) { |
| 86 |
a7 = a6; |
| 87 |
c7 = c6; |
| 88 |
} |
| 89 |
const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride); |
| 90 |
float* c8 = (float*) ((uintptr_t) c7 + cm_stride); |
| 91 |
if XNN_UNPREDICTABLE(mr <= 8) { |
| 92 |
a8 = a7; |
| 93 |
c8 = c7; |
| 94 |
} |
| 95 |
const int8_t* a9 = (const int8_t*) ((uintptr_t) a8 + a_stride); |
| 96 |
float* c9 = (float*) ((uintptr_t) c8 + cm_stride); |
| 97 |
if XNN_UNPREDICTABLE(mr < 10) { |
| 98 |
a9 = a8; |
| 99 |
c9 = c8; |
| 100 |
} |
| 101 |
const int8_t* a10 = (const int8_t*) ((uintptr_t) a9 + a_stride); |
| 102 |
float* c10 = (float*) ((uintptr_t) c9 + cm_stride); |
| 103 |
if XNN_UNPREDICTABLE(mr <= 10) { |
| 104 |
a10 = a9; |
| 105 |
c10 = c9; |
| 106 |
} |
| 107 |
const int8_t* a11 = (const int8_t*) ((uintptr_t) a10 + a_stride); |
| 108 |
float* c11 = (float*) ((uintptr_t) c10 + cm_stride); |
| 109 |
if XNN_UNPREDICTABLE(mr < 12) { |
| 110 |
a11 = a10; |
| 111 |
c11 = c10; |
| 112 |
} |
| 113 |
const int8_t* a12 = (const int8_t*) ((uintptr_t) a11 + a_stride); |
| 114 |
float* c12 = (float*) ((uintptr_t) c11 + cm_stride); |
| 115 |
if XNN_UNPREDICTABLE(mr <= 12) { |
| 116 |
a12 = a11; |
| 117 |
c12 = c11; |
| 118 |
} |
| 119 |
const int8_t* a13 = (const int8_t*) ((uintptr_t) a12 + a_stride); |
| 120 |
float* c13 = (float*) ((uintptr_t) c12 + cm_stride); |
| 121 |
if XNN_UNPREDICTABLE(mr != 14) { |
| 122 |
a13 = a12; |
| 123 |
c13 = c12; |
| 124 |
} |
| 125 |
size_t bl = params->scalar.blocksize; |
| 126 |
assert(bl != 0); |
| 127 |
assert(bl <= kc); |
| 128 |
assert(kc % bl == 0); |
| 129 |
assert(bl % 32 == 0); |
| 130 |
|
| 131 |
const __m512i vsign_mask = _mm512_set1_epi8(0x80); |
| 132 |
XNN_FORCE_REALIZATION(vsign_mask); |
| 133 |
const __m512 vinput_zero_point0 = _mm512_set1_ps((float) quantization_params[0].zero_point + 128); |
| 134 |
const __m512 vinput_zero_point1 = _mm512_set1_ps((float) quantization_params[1].zero_point + 128); |
| 135 |
const __m512 vinput_zero_point2 = _mm512_set1_ps((float) quantization_params[2].zero_point + 128); |
| 136 |
const __m512 vinput_zero_point3 = _mm512_set1_ps((float) quantization_params[3].zero_point + 128); |
| 137 |
const __m512 vinput_zero_point4 = _mm512_set1_ps((float) quantization_params[4].zero_point + 128); |
| 138 |
const __m512 vinput_zero_point5 = _mm512_set1_ps((float) quantization_params[5].zero_point + 128); |
| 139 |
const __m512 vinput_zero_point6 = _mm512_set1_ps((float) quantization_params[6].zero_point + 128); |
| 140 |
const __m512 vinput_zero_point7 = _mm512_set1_ps((float) quantization_params[7].zero_point + 128); |
| 141 |
const __m512 vinput_zero_point8 = _mm512_set1_ps((float) quantization_params[8].zero_point + 128); |
| 142 |
const __m512 vinput_zero_point9 = _mm512_set1_ps((float) quantization_params[9].zero_point + 128); |
| 143 |
const __m512 vinput_zero_point10 = _mm512_set1_ps((float) quantization_params[10].zero_point + 128); |
| 144 |
const __m512 vinput_zero_point11 = _mm512_set1_ps((float) quantization_params[11].zero_point + 128); |
| 145 |
const __m512 vinput_zero_point12 = _mm512_set1_ps((float) quantization_params[12].zero_point + 128); |
| 146 |
const __m512 vinput_zero_point13 = _mm512_set1_ps((float) quantization_params[13].zero_point + 128); |
| 147 |
const __m512 voutput_min = _mm512_set1_ps(params->scalar.min); |
| 148 |
const __m512 voutput_max = _mm512_set1_ps(params->scalar.max); |
| 149 |
// XNN_FORCE_REALIZATION(voutput_min); |
| 150 |
// XNN_FORCE_REALIZATION(voutput_max); |
| 151 |
const __m512i vmask = _mm512_set1_epi8(0xF0); |
| 152 |
XNN_FORCE_REALIZATION(vmask); |
| 153 |
do { |
| 154 |
const __m512 vksum0123456789ABCDEF = _mm512_loadu_ps(w); |
| 155 |
__m512 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point0); |
| 156 |
__m512 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point1); |
| 157 |
__m512 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point2); |
| 158 |
__m512 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point3); |
| 159 |
__m512 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point4); |
| 160 |
__m512 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point5); |
| 161 |
__m512 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point6); |
| 162 |
__m512 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point7); |
| 163 |
__m512 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point8); |
| 164 |
__m512 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point9); |
| 165 |
__m512 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point10); |
| 166 |
__m512 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point11); |
| 167 |
__m512 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point12); |
| 168 |
__m512 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point13); |
| 169 |
w = (const int32_t*) w + 16; |
| 170 |
|
| 171 |
for (size_t kb=0; kb < kc; kb+=bl) { |
| 172 |
__m512i vacc0x01234567 = _mm512_setzero_epi32(); |
| 173 |
__m512i vacc0x89ABCDEF = _mm512_setzero_epi32(); |
| 174 |
__m512i vacc1x01234567 = _mm512_setzero_epi32(); |
| 175 |
__m512i vacc1x89ABCDEF = _mm512_setzero_epi32(); |
| 176 |
__m512i vacc2x01234567 = _mm512_setzero_epi32(); |
| 177 |
__m512i vacc2x89ABCDEF = _mm512_setzero_epi32(); |
| 178 |
__m512i vacc3x01234567 = _mm512_setzero_epi32(); |
| 179 |
__m512i vacc3x89ABCDEF = _mm512_setzero_epi32(); |
| 180 |
__m512i vacc4x01234567 = _mm512_setzero_epi32(); |
| 181 |
__m512i vacc4x89ABCDEF = _mm512_setzero_epi32(); |
| 182 |
__m512i vacc5x01234567 = _mm512_setzero_epi32(); |
| 183 |
__m512i vacc5x89ABCDEF = _mm512_setzero_epi32(); |
| 184 |
__m512i vacc6x01234567 = _mm512_setzero_epi32(); |
| 185 |
__m512i vacc6x89ABCDEF = _mm512_setzero_epi32(); |
| 186 |
__m512i vacc7x01234567 = _mm512_setzero_epi32(); |
| 187 |
__m512i vacc7x89ABCDEF = _mm512_setzero_epi32(); |
| 188 |
__m512i vacc8x01234567 = _mm512_setzero_epi32(); |
| 189 |
__m512i vacc8x89ABCDEF = _mm512_setzero_epi32(); |
| 190 |
__m512i vacc9x01234567 = _mm512_setzero_epi32(); |
| 191 |
__m512i vacc9x89ABCDEF = _mm512_setzero_epi32(); |
| 192 |
__m512i vacc10x01234567 = _mm512_setzero_epi32(); |
| 193 |
__m512i vacc10x89ABCDEF = _mm512_setzero_epi32(); |
| 194 |
__m512i vacc11x01234567 = _mm512_setzero_epi32(); |
| 195 |
__m512i vacc11x89ABCDEF = _mm512_setzero_epi32(); |
| 196 |
__m512i vacc12x01234567 = _mm512_setzero_epi32(); |
| 197 |
__m512i vacc12x89ABCDEF = _mm512_setzero_epi32(); |
| 198 |
__m512i vacc13x01234567 = _mm512_setzero_epi32(); |
| 199 |
__m512i vacc13x89ABCDEF = _mm512_setzero_epi32(); |
| 200 |
size_t k = bl; |
| 201 |
while (k >= 16 * sizeof(int8_t)) { |
| 202 |
const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask); |
| 203 |
const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask); |
| 204 |
a0 += 16; |
| 205 |
const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask); |
| 206 |
const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask); |
| 207 |
a1 += 16; |
| 208 |
const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask); |
| 209 |
const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask); |
| 210 |
a2 += 16; |
| 211 |
const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask); |
| 212 |
const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask); |
| 213 |
a3 += 16; |
| 214 |
const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask); |
| 215 |
const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask); |
| 216 |
a4 += 16; |
| 217 |
const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask); |
| 218 |
const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask); |
| 219 |
a5 += 16; |
| 220 |
const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask); |
| 221 |
const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask); |
| 222 |
a6 += 16; |
| 223 |
const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask); |
| 224 |
const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask); |
| 225 |
a7 += 16; |
| 226 |
const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask); |
| 227 |
const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask); |
| 228 |
a8 += 16; |
| 229 |
const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask); |
| 230 |
const __m512i va9x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask); |
| 231 |
a9 += 16; |
| 232 |
const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask); |
| 233 |
const __m512i va10x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask); |
| 234 |
a10 += 16; |
| 235 |
const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask); |
| 236 |
const __m512i va11x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask); |
| 237 |
a11 += 16; |
| 238 |
const __m512i va12x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a12)), vsign_mask); |
| 239 |
const __m512i va12x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a12 + 8)), vsign_mask); |
| 240 |
a12 += 16; |
| 241 |
const __m512i va13x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a13)), vsign_mask); |
| 242 |
const __m512i va13x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a13 + 8)), vsign_mask); |
| 243 |
a13 += 16; |
| 244 |
|
| 245 |
const __m512i vbb01234567x01234567 = _mm512_loadu_si512(w); |
| 246 |
const __m512i vbb89ABCDEFx01234567 = _mm512_loadu_si512((const int8_t*) w + 64); |
| 247 |
const __m512i vbs01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4); |
| 248 |
const __m512i vbs89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4); |
| 249 |
const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbb01234567x01234567, vmask); |
| 250 |
const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbb89ABCDEFx01234567, vmask); |
| 251 |
const __m512i vb01234567x01234567 = _mm512_and_si512(vbs01234567x01234567, vmask); |
| 252 |
const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbs89ABCDEFx01234567, vmask); |
| 253 |
|
| 254 |
vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567); |
| 255 |
vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567); |
| 256 |
vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567); |
| 257 |
vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567); |
| 258 |
vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567); |
| 259 |
vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567); |
| 260 |
vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567); |
| 261 |
vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567); |
| 262 |
vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567); |
| 263 |
vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567); |
| 264 |
vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567); |
| 265 |
vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567); |
| 266 |
vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567); |
| 267 |
vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567); |
| 268 |
vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567); |
| 269 |
vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567); |
| 270 |
vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567); |
| 271 |
vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567); |
| 272 |
vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567); |
| 273 |
vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567); |
| 274 |
vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567); |
| 275 |
vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567); |
| 276 |
vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567); |
| 277 |
vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567); |
| 278 |
vacc12x01234567 = _mm512_dpbusd_epi32(vacc12x01234567, va12x01234567, vb01234567x01234567); |
| 279 |
vacc12x89ABCDEF = _mm512_dpbusd_epi32(vacc12x89ABCDEF, va12x01234567, vb89ABCDEFx01234567); |
| 280 |
vacc13x01234567 = _mm512_dpbusd_epi32(vacc13x01234567, va13x01234567, vb01234567x01234567); |
| 281 |
vacc13x89ABCDEF = _mm512_dpbusd_epi32(vacc13x89ABCDEF, va13x01234567, vb89ABCDEFx01234567); |
| 282 |
xnn_prefetch_to_l1((const int8_t*) w + 896); |
| 283 |
xnn_prefetch_to_l1((const int8_t*) w + 960); |
| 284 |
vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF); |
| 285 |
vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 286 |
vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF); |
| 287 |
vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 288 |
vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF); |
| 289 |
vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 290 |
vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF); |
| 291 |
vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 292 |
vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF); |
| 293 |
vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 294 |
vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF); |
| 295 |
vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 296 |
vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF); |
| 297 |
vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 298 |
vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF); |
| 299 |
vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 300 |
vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF); |
| 301 |
vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 302 |
vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x89ABCDEF, vb01234567x89ABCDEF); |
| 303 |
vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 304 |
vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x89ABCDEF, vb01234567x89ABCDEF); |
| 305 |
vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 306 |
vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x89ABCDEF, vb01234567x89ABCDEF); |
| 307 |
vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 308 |
vacc12x01234567 = _mm512_dpbusd_epi32(vacc12x01234567, va12x89ABCDEF, vb01234567x89ABCDEF); |
| 309 |
vacc12x89ABCDEF = _mm512_dpbusd_epi32(vacc12x89ABCDEF, va12x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 310 |
vacc13x01234567 = _mm512_dpbusd_epi32(vacc13x01234567, va13x89ABCDEF, vb01234567x89ABCDEF); |
| 311 |
vacc13x89ABCDEF = _mm512_dpbusd_epi32(vacc13x89ABCDEF, va13x89ABCDEF, vb89ABCDEFx89ABCDEF); |
| 312 |
|
| 313 |
w = (const int8_t*) w + 128; |
| 314 |
k -= 16 * sizeof(int8_t); |
| 315 |
} |
| 316 |
|
| 317 |
if (k != 0) { |
| 318 |
const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask); |
| 319 |
a0 += 8; |
| 320 |
const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask); |
| 321 |
a1 += 8; |
| 322 |
const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask); |
| 323 |
a2 += 8; |
| 324 |
const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask); |
| 325 |
a3 += 8; |
| 326 |
const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask); |
| 327 |
a4 += 8; |
| 328 |
const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask); |
| 329 |
a5 += 8; |
| 330 |
const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask); |
| 331 |
a6 += 8; |
| 332 |
const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask); |
| 333 |
a7 += 8; |
| 334 |
const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask); |
| 335 |
a8 += 8; |
| 336 |
const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask); |
| 337 |
a9 += 8; |
| 338 |
const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask); |
| 339 |
a10 += 8; |
| 340 |
const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask); |
| 341 |
a11 += 8; |
| 342 |
const __m512i va12x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a12)), vsign_mask); |
| 343 |
a12 += 8; |
| 344 |
const __m512i va13x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a13)), vsign_mask); |
| 345 |
a13 += 8; |
| 346 |
|
| 347 |
const __m512i vbb01234567x01234567 = _mm512_load_si512(w); |
| 348 |
const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64); |
| 349 |
const __m512i vb01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4); |
| 350 |
const __m512i vb89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4); |
| 351 |
|
| 352 |
vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567); |
| 353 |
vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567); |
| 354 |
vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567); |
| 355 |
vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567); |
| 356 |
vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567); |
| 357 |
vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567); |
| 358 |
vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567); |
| 359 |
vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567); |
| 360 |
vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567); |
| 361 |
vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567); |
| 362 |
vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567); |
| 363 |
vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567); |
| 364 |
vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567); |
| 365 |
vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567); |
| 366 |
vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567); |
| 367 |
vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567); |
| 368 |
vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567); |
| 369 |
vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567); |
| 370 |
vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567); |
| 371 |
vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567); |
| 372 |
vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567); |
| 373 |
vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567); |
| 374 |
vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567); |
| 375 |
vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567); |
| 376 |
vacc12x01234567 = _mm512_dpbusd_epi32(vacc12x01234567, va12x01234567, vb01234567x01234567); |
| 377 |
vacc12x89ABCDEF = _mm512_dpbusd_epi32(vacc12x89ABCDEF, va12x01234567, vb89ABCDEFx01234567); |
| 378 |
vacc13x01234567 = _mm512_dpbusd_epi32(vacc13x01234567, va13x01234567, vb01234567x01234567); |
| 379 |
vacc13x89ABCDEF = _mm512_dpbusd_epi32(vacc13x89ABCDEF, va13x01234567, vb89ABCDEFx01234567); |
| 380 |
xnn_prefetch_to_l1((const int8_t*) w + 896); |
| 381 |
xnn_prefetch_to_l1((const int8_t*) w + 960); |
| 382 |
|
| 383 |
w = (const int8_t*) w + 128; |
| 384 |
k -= 8 * sizeof(int8_t); |
| 385 |
} |
| 386 |
const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_castsi512_ps(_mm512_slli_epi32( |
| 387 |
_mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) w)), 16)); |
| 388 |
w = (const uint16_t*) w + 16; |
| 389 |
|
| 390 |
// Add adjacent pairs |
| 391 |
const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0); |
| 392 |
const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32)); |
| 393 |
const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32)); |
| 394 |
__m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF); |
| 395 |
const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32)); |
| 396 |
const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32)); |
| 397 |
__m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF); |
| 398 |
const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32)); |
| 399 |
const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32)); |
| 400 |
__m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF); |
| 401 |
const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32)); |
| 402 |
const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32)); |
| 403 |
__m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF); |
| 404 |
const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32)); |
| 405 |
const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32)); |
| 406 |
__m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF); |
| 407 |
const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32)); |
| 408 |
const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32)); |
| 409 |
__m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF); |
| 410 |
const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32)); |
| 411 |
const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32)); |
| 412 |
__m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF); |
| 413 |
const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32)); |
| 414 |
const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32)); |
| 415 |
__m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF); |
| 416 |
const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32)); |
| 417 |
const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32)); |
| 418 |
__m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF); |
| 419 |
const __m512i vsum9x01234567 = _mm512_add_epi32(vacc9x01234567, _mm512_srli_epi64(vacc9x01234567, 32)); |
| 420 |
const __m512i vsum9x89ABCDEF = _mm512_add_epi32(vacc9x89ABCDEF, _mm512_srli_epi64(vacc9x89ABCDEF, 32)); |
| 421 |
__m512i vacc9x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum9x01234567, vidx, vsum9x89ABCDEF); |
| 422 |
const __m512i vsum10x01234567 = _mm512_add_epi32(vacc10x01234567, _mm512_srli_epi64(vacc10x01234567, 32)); |
| 423 |
const __m512i vsum10x89ABCDEF = _mm512_add_epi32(vacc10x89ABCDEF, _mm512_srli_epi64(vacc10x89ABCDEF, 32)); |
| 424 |
__m512i vacc10x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum10x01234567, vidx, vsum10x89ABCDEF); |
| 425 |
const __m512i vsum11x01234567 = _mm512_add_epi32(vacc11x01234567, _mm512_srli_epi64(vacc11x01234567, 32)); |
| 426 |
const __m512i vsum11x89ABCDEF = _mm512_add_epi32(vacc11x89ABCDEF, _mm512_srli_epi64(vacc11x89ABCDEF, 32)); |
| 427 |
__m512i vacc11x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum11x01234567, vidx, vsum11x89ABCDEF); |
| 428 |
const __m512i vsum12x01234567 = _mm512_add_epi32(vacc12x01234567, _mm512_srli_epi64(vacc12x01234567, 32)); |
| 429 |
const __m512i vsum12x89ABCDEF = _mm512_add_epi32(vacc12x89ABCDEF, _mm512_srli_epi64(vacc12x89ABCDEF, 32)); |
| 430 |
__m512i vacc12x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum12x01234567, vidx, vsum12x89ABCDEF); |
| 431 |
const __m512i vsum13x01234567 = _mm512_add_epi32(vacc13x01234567, _mm512_srli_epi64(vacc13x01234567, 32)); |
| 432 |
const __m512i vsum13x89ABCDEF = _mm512_add_epi32(vacc13x89ABCDEF, _mm512_srli_epi64(vacc13x89ABCDEF, 32)); |
| 433 |
__m512i vacc13x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum13x01234567, vidx, vsum13x89ABCDEF); |
| 434 |
__m512 vf0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF); |
| 435 |
__m512 vf1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF); |
| 436 |
__m512 vf2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF); |
| 437 |
__m512 vf3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF); |
| 438 |
__m512 vf4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF); |
| 439 |
__m512 vf5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF); |
| 440 |
__m512 vf6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF); |
| 441 |
__m512 vf7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF); |
| 442 |
__m512 vf8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF); |
| 443 |
__m512 vf9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF); |
| 444 |
__m512 vf10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF); |
| 445 |
__m512 vf11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF); |
| 446 |
__m512 vf12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF); |
| 447 |
__m512 vf13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF); |
| 448 |
|
| 449 |
vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vf0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled0x0123456789ABCDEF); |
| 450 |
vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vf1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled1x0123456789ABCDEF); |
| 451 |
vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vf2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled2x0123456789ABCDEF); |
| 452 |
vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vf3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled3x0123456789ABCDEF); |
| 453 |
vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vf4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled4x0123456789ABCDEF); |
| 454 |
vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vf5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled5x0123456789ABCDEF); |
| 455 |
vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vf6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled6x0123456789ABCDEF); |
| 456 |
vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vf7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled7x0123456789ABCDEF); |
| 457 |
vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vf8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled8x0123456789ABCDEF); |
| 458 |
vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vf9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled9x0123456789ABCDEF); |
| 459 |
vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vf10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled10x0123456789ABCDEF); |
| 460 |
vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vf11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled11x0123456789ABCDEF); |
| 461 |
vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vf12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled12x0123456789ABCDEF); |
| 462 |
vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vf13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled13x0123456789ABCDEF); |
| 463 |
} |
| 464 |
|
| 465 |
vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale)); |
| 466 |
vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale)); |
| 467 |
vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale)); |
| 468 |
vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale)); |
| 469 |
vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale)); |
| 470 |
vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale)); |
| 471 |
vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale)); |
| 472 |
vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale)); |
| 473 |
vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale)); |
| 474 |
vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale)); |
| 475 |
vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale)); |
| 476 |
vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale)); |
| 477 |
vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params[12].inv_scale)); |
| 478 |
vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params[13].inv_scale)); |
| 479 |
|
| 480 |
const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w); |
| 481 |
w = (const float*) w + 16; |
| 482 |
|
| 483 |
vscaled0x0123456789ABCDEF = _mm512_add_ps(vscaled0x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 484 |
vscaled1x0123456789ABCDEF = _mm512_add_ps(vscaled1x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 485 |
vscaled2x0123456789ABCDEF = _mm512_add_ps(vscaled2x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 486 |
vscaled3x0123456789ABCDEF = _mm512_add_ps(vscaled3x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 487 |
vscaled4x0123456789ABCDEF = _mm512_add_ps(vscaled4x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 488 |
vscaled5x0123456789ABCDEF = _mm512_add_ps(vscaled5x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 489 |
vscaled6x0123456789ABCDEF = _mm512_add_ps(vscaled6x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 490 |
vscaled7x0123456789ABCDEF = _mm512_add_ps(vscaled7x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 491 |
vscaled8x0123456789ABCDEF = _mm512_add_ps(vscaled8x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 492 |
vscaled9x0123456789ABCDEF = _mm512_add_ps(vscaled9x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 493 |
vscaled10x0123456789ABCDEF = _mm512_add_ps(vscaled10x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 494 |
vscaled11x0123456789ABCDEF = _mm512_add_ps(vscaled11x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 495 |
vscaled12x0123456789ABCDEF = _mm512_add_ps(vscaled12x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 496 |
vscaled13x0123456789ABCDEF = _mm512_add_ps(vscaled13x0123456789ABCDEF, vbias0123456789ABCDEF); |
| 497 |
|
| 498 |
vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min); |
| 499 |
vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min); |
| 500 |
vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min); |
| 501 |
vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min); |
| 502 |
vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min); |
| 503 |
vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min); |
| 504 |
vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min); |
| 505 |
vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min); |
| 506 |
vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min); |
| 507 |
vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min); |
| 508 |
vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min); |
| 509 |
vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min); |
| 510 |
vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min); |
| 511 |
vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min); |
| 512 |
|
| 513 |
vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max); |
| 514 |
vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max); |
| 515 |
vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max); |
| 516 |
vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max); |
| 517 |
vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max); |
| 518 |
vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max); |
| 519 |
vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max); |
| 520 |
vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max); |
| 521 |
vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max); |
| 522 |
vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max); |
| 523 |
vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max); |
| 524 |
vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max); |
| 525 |
vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max); |
| 526 |
vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max); |
| 527 |
|
| 528 |
if XNN_LIKELY(nc >= 16) { |
| 529 |
_mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF); |
| 530 |
_mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF); |
| 531 |
_mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF); |
| 532 |
_mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF); |
| 533 |
_mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF); |
| 534 |
_mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF); |
| 535 |
_mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF); |
| 536 |
_mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF); |
| 537 |
_mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF); |
| 538 |
_mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF); |
| 539 |
_mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF); |
| 540 |
_mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF); |
| 541 |
_mm512_storeu_ps(c12, vscaled12x0123456789ABCDEF); |
| 542 |
_mm512_storeu_ps(c13, vscaled13x0123456789ABCDEF); |
| 543 |
|
| 544 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 545 |
a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 546 |
a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 547 |
a3 = (const int8_t*) ((uintptr_t) a3 - kc); |
| 548 |
a4 = (const int8_t*) ((uintptr_t) a4 - kc); |
| 549 |
a5 = (const int8_t*) ((uintptr_t) a5 - kc); |
| 550 |
a6 = (const int8_t*) ((uintptr_t) a6 - kc); |
| 551 |
a7 = (const int8_t*) ((uintptr_t) a7 - kc); |
| 552 |
a8 = (const int8_t*) ((uintptr_t) a8 - kc); |
| 553 |
a9 = (const int8_t*) ((uintptr_t) a9 - kc); |
| 554 |
a10 = (const int8_t*) ((uintptr_t) a10 - kc); |
| 555 |
a11 = (const int8_t*) ((uintptr_t) a11 - kc); |
| 556 |
a12 = (const int8_t*) ((uintptr_t) a12 - kc); |
| 557 |
a13 = (const int8_t*) ((uintptr_t) a13 - kc); |
| 558 |
|
| 559 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 560 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 561 |
c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 562 |
c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 563 |
c4 = (float*) ((uintptr_t) c4 + cn_stride); |
| 564 |
c5 = (float*) ((uintptr_t) c5 + cn_stride); |
| 565 |
c6 = (float*) ((uintptr_t) c6 + cn_stride); |
| 566 |
c7 = (float*) ((uintptr_t) c7 + cn_stride); |
| 567 |
c8 = (float*) ((uintptr_t) c8 + cn_stride); |
| 568 |
c9 = (float*) ((uintptr_t) c9 + cn_stride); |
| 569 |
c10 = (float*) ((uintptr_t) c10 + cn_stride); |
| 570 |
c11 = (float*) ((uintptr_t) c11 + cn_stride); |
| 571 |
c12 = (float*) ((uintptr_t) c12 + cn_stride); |
| 572 |
c13 = (float*) ((uintptr_t) c13 + cn_stride); |
| 573 |
|
| 574 |
nc -= 16; |
| 575 |
} else { |
| 576 |
// Prepare mask for valid 32-bit elements (depends on nc). |
| 577 |
const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1); |
| 578 |
_mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF); |
| 579 |
_mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF); |
| 580 |
_mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF); |
| 581 |
_mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF); |
| 582 |
_mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF); |
| 583 |
_mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF); |
| 584 |
_mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF); |
| 585 |
_mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF); |
| 586 |
_mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF); |
| 587 |
_mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF); |
| 588 |
_mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF); |
| 589 |
_mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF); |
| 590 |
_mm512_mask_storeu_ps(c12, vmask, vscaled12x0123456789ABCDEF); |
| 591 |
_mm512_mask_storeu_ps(c13, vmask, vscaled13x0123456789ABCDEF); |
| 592 |
nc = 0; |
| 593 |
} |
| 594 |
} while (nc != 0); |
| 595 |
} |