| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/MRx8c4-avxvnni.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 |
#include <stdint.h> |
| 12 |
|
| 13 |
#include <immintrin.h> |
| 14 |
|
| 15 |
#include "xnnpack/common.h" |
| 16 |
#include "xnnpack/gemm.h" |
| 17 |
#include "xnnpack/intrinsics-polyfill.h" |
| 18 |
#include "xnnpack/math.h" |
| 19 |
#include "xnnpack/unaligned.h" |
| 20 |
#include "xnnpack/prefetch.h" |
| 21 |
|
| 22 |
|
| 23 |
void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_5x8c4__avxvnni_u4_acc4( |
| 24 |
size_t mr, |
| 25 |
size_t nc, |
| 26 |
size_t kc, |
| 27 |
const int8_t* restrict a, |
| 28 |
size_t a_stride, |
| 29 |
const void* restrict w, |
| 30 |
float* restrict c, |
| 31 |
size_t cm_stride, |
| 32 |
size_t cn_stride, |
| 33 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 34 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 35 |
{ |
| 36 |
assert(mr != 0); |
| 37 |
assert(mr <= 5); |
| 38 |
assert(nc != 0); |
| 39 |
assert(kc != 0); |
| 40 |
assert(kc % sizeof(int8_t) == 0); |
| 41 |
assert(a != NULL); |
| 42 |
assert(w != NULL); |
| 43 |
assert(c != NULL); |
| 44 |
|
| 45 |
kc = round_up_po2(kc, 4 * sizeof(int8_t)); |
| 46 |
const int8_t* a0 = a; |
| 47 |
float* c0 = c; |
| 48 |
const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 49 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 50 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 51 |
a1 = a0; |
| 52 |
c1 = c0; |
| 53 |
} |
| 54 |
const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 55 |
float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 56 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 57 |
a2 = a1; |
| 58 |
c2 = c1; |
| 59 |
} |
| 60 |
const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride); |
| 61 |
float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 62 |
if XNN_UNPREDICTABLE(mr < 4) { |
| 63 |
a3 = a2; |
| 64 |
c3 = c2; |
| 65 |
} |
| 66 |
const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride); |
| 67 |
float* c4 = (float*) ((uintptr_t) c3 + cm_stride); |
| 68 |
if XNN_UNPREDICTABLE(mr <= 4) { |
| 69 |
a4 = a3; |
| 70 |
c4 = c3; |
| 71 |
} |
| 72 |
|
| 73 |
const __m256i vsign_mask =_mm256_set1_epi8(0x80); |
| 74 |
XNN_FORCE_REALIZATION(vsign_mask); |
| 75 |
const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128); |
| 76 |
const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128); |
| 77 |
const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128); |
| 78 |
const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128); |
| 79 |
const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128); |
| 80 |
const __m256 voutput_min = _mm256_set1_ps(params->scalar.min); |
| 81 |
const __m256 voutput_max = _mm256_set1_ps(params->scalar.max); |
| 82 |
// XNN_FORCE_REALIZATION(voutput_min); |
| 83 |
// XNN_FORCE_REALIZATION(voutput_max); |
| 84 |
do { |
| 85 |
const __m256i vksum01234567 = _mm256_load_si256(w); |
| 86 |
__m256i vacc0x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0); |
| 87 |
__m256i vacc0x1x01234567 = _mm256_setzero_si256(); |
| 88 |
__m256i vacc0x2x01234567 = _mm256_setzero_si256(); |
| 89 |
__m256i vacc0x3x01234567 = _mm256_setzero_si256(); |
| 90 |
__m256i vacc1x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1); |
| 91 |
__m256i vacc1x1x01234567 = _mm256_setzero_si256(); |
| 92 |
__m256i vacc1x2x01234567 = _mm256_setzero_si256(); |
| 93 |
__m256i vacc1x3x01234567 = _mm256_setzero_si256(); |
| 94 |
__m256i vacc2x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2); |
| 95 |
__m256i vacc2x1x01234567 = _mm256_setzero_si256(); |
| 96 |
__m256i vacc2x2x01234567 = _mm256_setzero_si256(); |
| 97 |
__m256i vacc2x3x01234567 = _mm256_setzero_si256(); |
| 98 |
__m256i vacc3x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3); |
| 99 |
__m256i vacc3x1x01234567 = _mm256_setzero_si256(); |
| 100 |
__m256i vacc3x2x01234567 = _mm256_setzero_si256(); |
| 101 |
__m256i vacc3x3x01234567 = _mm256_setzero_si256(); |
| 102 |
__m256i vacc4x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4); |
| 103 |
__m256i vacc4x1x01234567 = _mm256_setzero_si256(); |
| 104 |
__m256i vacc4x2x01234567 = _mm256_setzero_si256(); |
| 105 |
__m256i vacc4x3x01234567 = _mm256_setzero_si256(); |
| 106 |
w = (const int32_t*) w + 8; |
| 107 |
|
| 108 |
size_t k = kc; |
| 109 |
while (k >= 16 * sizeof(int8_t)) { |
| 110 |
__m256i va0x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 0)); |
| 111 |
__m256i va0x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 4)); |
| 112 |
__m256i va0x2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 8)); |
| 113 |
__m256i va0x3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 12)); |
| 114 |
a0 += 16; |
| 115 |
__m256i va1x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 0)); |
| 116 |
__m256i va1x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 4)); |
| 117 |
__m256i va1x2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 8)); |
| 118 |
__m256i va1x3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 12)); |
| 119 |
a1 += 16; |
| 120 |
__m256i va2x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 0)); |
| 121 |
__m256i va2x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 4)); |
| 122 |
__m256i va2x2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 8)); |
| 123 |
__m256i va2x3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 12)); |
| 124 |
a2 += 16; |
| 125 |
__m256i va3x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 0)); |
| 126 |
__m256i va3x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 4)); |
| 127 |
__m256i va3x2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 8)); |
| 128 |
__m256i va3x3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 12)); |
| 129 |
a3 += 16; |
| 130 |
__m256i va4x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 0)); |
| 131 |
__m256i va4x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 4)); |
| 132 |
__m256i va4x2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 8)); |
| 133 |
__m256i va4x3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 12)); |
| 134 |
a4 += 16; |
| 135 |
|
| 136 |
va0x0x0123 = _mm256_xor_si256(va0x0x0123, vsign_mask); |
| 137 |
va0x1x0123 = _mm256_xor_si256(va0x1x0123, vsign_mask); |
| 138 |
va0x2x0123 = _mm256_xor_si256(va0x2x0123, vsign_mask); |
| 139 |
va0x3x0123 = _mm256_xor_si256(va0x3x0123, vsign_mask); |
| 140 |
va1x0x0123 = _mm256_xor_si256(va1x0x0123, vsign_mask); |
| 141 |
va1x1x0123 = _mm256_xor_si256(va1x1x0123, vsign_mask); |
| 142 |
va1x2x0123 = _mm256_xor_si256(va1x2x0123, vsign_mask); |
| 143 |
va1x3x0123 = _mm256_xor_si256(va1x3x0123, vsign_mask); |
| 144 |
va2x0x0123 = _mm256_xor_si256(va2x0x0123, vsign_mask); |
| 145 |
va2x1x0123 = _mm256_xor_si256(va2x1x0123, vsign_mask); |
| 146 |
va2x2x0123 = _mm256_xor_si256(va2x2x0123, vsign_mask); |
| 147 |
va2x3x0123 = _mm256_xor_si256(va2x3x0123, vsign_mask); |
| 148 |
va3x0x0123 = _mm256_xor_si256(va3x0x0123, vsign_mask); |
| 149 |
va3x1x0123 = _mm256_xor_si256(va3x1x0123, vsign_mask); |
| 150 |
va3x2x0123 = _mm256_xor_si256(va3x2x0123, vsign_mask); |
| 151 |
va3x3x0123 = _mm256_xor_si256(va3x3x0123, vsign_mask); |
| 152 |
va4x0x0123 = _mm256_xor_si256(va4x0x0123, vsign_mask); |
| 153 |
va4x1x0123 = _mm256_xor_si256(va4x1x0123, vsign_mask); |
| 154 |
va4x2x0123 = _mm256_xor_si256(va4x2x0123, vsign_mask); |
| 155 |
va4x3x0123 = _mm256_xor_si256(va4x3x0123, vsign_mask); |
| 156 |
|
| 157 |
const __m256i vb0x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0)); |
| 158 |
const __m256i vb1x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)); |
| 159 |
const __m256i vb2x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)); |
| 160 |
const __m256i vb3x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)); |
| 161 |
w = (const int8_t*) w + 128; |
| 162 |
|
| 163 |
vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0x0123, vb0x01234567); |
| 164 |
vacc0x1x01234567 = _mm256_dpbusd_avx_epi32(vacc0x1x01234567, va0x1x0123, vb1x01234567); |
| 165 |
vacc0x2x01234567 = _mm256_dpbusd_avx_epi32(vacc0x2x01234567, va0x2x0123, vb2x01234567); |
| 166 |
vacc0x3x01234567 = _mm256_dpbusd_avx_epi32(vacc0x3x01234567, va0x3x0123, vb3x01234567); |
| 167 |
vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0x0123, vb0x01234567); |
| 168 |
vacc1x1x01234567 = _mm256_dpbusd_avx_epi32(vacc1x1x01234567, va1x1x0123, vb1x01234567); |
| 169 |
vacc1x2x01234567 = _mm256_dpbusd_avx_epi32(vacc1x2x01234567, va1x2x0123, vb2x01234567); |
| 170 |
vacc1x3x01234567 = _mm256_dpbusd_avx_epi32(vacc1x3x01234567, va1x3x0123, vb3x01234567); |
| 171 |
vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0x0123, vb0x01234567); |
| 172 |
vacc2x1x01234567 = _mm256_dpbusd_avx_epi32(vacc2x1x01234567, va2x1x0123, vb1x01234567); |
| 173 |
vacc2x2x01234567 = _mm256_dpbusd_avx_epi32(vacc2x2x01234567, va2x2x0123, vb2x01234567); |
| 174 |
vacc2x3x01234567 = _mm256_dpbusd_avx_epi32(vacc2x3x01234567, va2x3x0123, vb3x01234567); |
| 175 |
vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0x0123, vb0x01234567); |
| 176 |
vacc3x1x01234567 = _mm256_dpbusd_avx_epi32(vacc3x1x01234567, va3x1x0123, vb1x01234567); |
| 177 |
vacc3x2x01234567 = _mm256_dpbusd_avx_epi32(vacc3x2x01234567, va3x2x0123, vb2x01234567); |
| 178 |
vacc3x3x01234567 = _mm256_dpbusd_avx_epi32(vacc3x3x01234567, va3x3x0123, vb3x01234567); |
| 179 |
vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0x0123, vb0x01234567); |
| 180 |
vacc4x1x01234567 = _mm256_dpbusd_avx_epi32(vacc4x1x01234567, va4x1x0123, vb1x01234567); |
| 181 |
vacc4x2x01234567 = _mm256_dpbusd_avx_epi32(vacc4x2x01234567, va4x2x0123, vb2x01234567); |
| 182 |
vacc4x3x01234567 = _mm256_dpbusd_avx_epi32(vacc4x3x01234567, va4x3x0123, vb3x01234567); |
| 183 |
|
| 184 |
k -= 16 * sizeof(int8_t); |
| 185 |
} |
| 186 |
vacc0x0x01234567 = _mm256_add_epi32(vacc0x0x01234567, vacc0x1x01234567); |
| 187 |
vacc1x0x01234567 = _mm256_add_epi32(vacc1x0x01234567, vacc1x1x01234567); |
| 188 |
vacc2x0x01234567 = _mm256_add_epi32(vacc2x0x01234567, vacc2x1x01234567); |
| 189 |
vacc3x0x01234567 = _mm256_add_epi32(vacc3x0x01234567, vacc3x1x01234567); |
| 190 |
vacc4x0x01234567 = _mm256_add_epi32(vacc4x0x01234567, vacc4x1x01234567); |
| 191 |
vacc0x2x01234567 = _mm256_add_epi32(vacc0x2x01234567, vacc0x3x01234567); |
| 192 |
vacc1x2x01234567 = _mm256_add_epi32(vacc1x2x01234567, vacc1x3x01234567); |
| 193 |
vacc2x2x01234567 = _mm256_add_epi32(vacc2x2x01234567, vacc2x3x01234567); |
| 194 |
vacc3x2x01234567 = _mm256_add_epi32(vacc3x2x01234567, vacc3x3x01234567); |
| 195 |
vacc4x2x01234567 = _mm256_add_epi32(vacc4x2x01234567, vacc4x3x01234567); |
| 196 |
vacc0x0x01234567 = _mm256_add_epi32(vacc0x0x01234567, vacc0x2x01234567); |
| 197 |
vacc1x0x01234567 = _mm256_add_epi32(vacc1x0x01234567, vacc1x2x01234567); |
| 198 |
vacc2x0x01234567 = _mm256_add_epi32(vacc2x0x01234567, vacc2x2x01234567); |
| 199 |
vacc3x0x01234567 = _mm256_add_epi32(vacc3x0x01234567, vacc3x2x01234567); |
| 200 |
vacc4x0x01234567 = _mm256_add_epi32(vacc4x0x01234567, vacc4x2x01234567); |
| 201 |
while (k != 0) { |
| 202 |
__m256i va0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0)); |
| 203 |
a0 += 4; |
| 204 |
__m256i va1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1)); |
| 205 |
a1 += 4; |
| 206 |
__m256i va2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2)); |
| 207 |
a2 += 4; |
| 208 |
__m256i va3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3)); |
| 209 |
a3 += 4; |
| 210 |
__m256i va4x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4)); |
| 211 |
a4 += 4; |
| 212 |
|
| 213 |
va0x0123 = _mm256_xor_si256(va0x0123, vsign_mask); |
| 214 |
va1x0123 = _mm256_xor_si256(va1x0123, vsign_mask); |
| 215 |
va2x0123 = _mm256_xor_si256(va2x0123, vsign_mask); |
| 216 |
va3x0123 = _mm256_xor_si256(va3x0123, vsign_mask); |
| 217 |
va4x0123 = _mm256_xor_si256(va4x0123, vsign_mask); |
| 218 |
|
| 219 |
const __m256i vb01234567 = _mm256_load_si256(w); |
| 220 |
|
| 221 |
vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0123, vb01234567); |
| 222 |
vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0123, vb01234567); |
| 223 |
vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0123, vb01234567); |
| 224 |
vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0123, vb01234567); |
| 225 |
vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0123, vb01234567); |
| 226 |
|
| 227 |
w = (const int8_t*) w + 32; |
| 228 |
k -= 4 * sizeof(int8_t); |
| 229 |
} |
| 230 |
|
| 231 |
__m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x0x01234567); |
| 232 |
__m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x0x01234567); |
| 233 |
__m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x0x01234567); |
| 234 |
__m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x0x01234567); |
| 235 |
__m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x0x01234567); |
| 236 |
|
| 237 |
vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale)); |
| 238 |
vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale)); |
| 239 |
vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale)); |
| 240 |
vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale)); |
| 241 |
vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale)); |
| 242 |
|
| 243 |
const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w); |
| 244 |
const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8); |
| 245 |
w = (const float*) w + 16; |
| 246 |
|
| 247 |
vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567); |
| 248 |
vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567); |
| 249 |
vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567); |
| 250 |
vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567); |
| 251 |
vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567); |
| 252 |
|
| 253 |
vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min); |
| 254 |
vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min); |
| 255 |
vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min); |
| 256 |
vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min); |
| 257 |
vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min); |
| 258 |
|
| 259 |
vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max); |
| 260 |
vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max); |
| 261 |
vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max); |
| 262 |
vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max); |
| 263 |
vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max); |
| 264 |
|
| 265 |
if XNN_LIKELY(nc >= 8) { |
| 266 |
_mm256_storeu_ps(c0, vout0x01234567); |
| 267 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 268 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 269 |
_mm256_storeu_ps(c1, vout1x01234567); |
| 270 |
a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 271 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 272 |
_mm256_storeu_ps(c2, vout2x01234567); |
| 273 |
a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 274 |
c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 275 |
_mm256_storeu_ps(c3, vout3x01234567); |
| 276 |
a3 = (const int8_t*) ((uintptr_t) a3 - kc); |
| 277 |
c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 278 |
_mm256_storeu_ps(c4, vout4x01234567); |
| 279 |
a4 = (const int8_t*) ((uintptr_t) a4 - kc); |
| 280 |
c4 = (float*) ((uintptr_t) c4 + cn_stride); |
| 281 |
nc -= 8; |
| 282 |
} else { |
| 283 |
__m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567); |
| 284 |
__m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567); |
| 285 |
__m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567); |
| 286 |
__m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567); |
| 287 |
__m128 vout4x0123 = _mm256_castps256_ps128(vout4x01234567); |
| 288 |
if (nc & 4) { |
| 289 |
_mm_storeu_ps(c0, vout0x0123); |
| 290 |
c0 += 4; |
| 291 |
_mm_storeu_ps(c1, vout1x0123); |
| 292 |
c1 += 4; |
| 293 |
_mm_storeu_ps(c2, vout2x0123); |
| 294 |
c2 += 4; |
| 295 |
_mm_storeu_ps(c3, vout3x0123); |
| 296 |
c3 += 4; |
| 297 |
_mm_storeu_ps(c4, vout4x0123); |
| 298 |
c4 += 4; |
| 299 |
vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1); |
| 300 |
vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1); |
| 301 |
vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1); |
| 302 |
vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1); |
| 303 |
vout4x0123 = _mm256_extractf128_ps(vout4x01234567, 1); |
| 304 |
} |
| 305 |
if (nc & 2) { |
| 306 |
_mm_storel_pi((__m64*) c0, vout0x0123); |
| 307 |
c0 += 2; |
| 308 |
_mm_storel_pi((__m64*) c1, vout1x0123); |
| 309 |
c1 += 2; |
| 310 |
_mm_storel_pi((__m64*) c2, vout2x0123); |
| 311 |
c2 += 2; |
| 312 |
_mm_storel_pi((__m64*) c3, vout3x0123); |
| 313 |
c3 += 2; |
| 314 |
_mm_storel_pi((__m64*) c4, vout4x0123); |
| 315 |
c4 += 2; |
| 316 |
vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123); |
| 317 |
vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123); |
| 318 |
vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123); |
| 319 |
vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123); |
| 320 |
vout4x0123 = _mm_movehl_ps(vout4x0123, vout4x0123); |
| 321 |
} |
| 322 |
if (nc & 1) { |
| 323 |
_mm_store_ss(c0, vout0x0123); |
| 324 |
_mm_store_ss(c1, vout1x0123); |
| 325 |
_mm_store_ss(c2, vout2x0123); |
| 326 |
_mm_store_ss(c3, vout3x0123); |
| 327 |
_mm_store_ss(c4, vout4x0123); |
| 328 |
} |
| 329 |
nc = 0; |
| 330 |
} |
| 331 |
} while (nc != 0); |
| 332 |
} |