| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/MRx8c8-avxvnni.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 <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/gemm.h" |
| 15 |
#include "xnnpack/intrinsics-polyfill.h" |
| 16 |
#include "xnnpack/math.h" |
| 17 |
#include "xnnpack/unaligned.h" |
| 18 |
|
| 19 |
|
| 20 |
void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x8c8__avxvnni( |
| 21 |
size_t mr, |
| 22 |
size_t nc, |
| 23 |
size_t kc, |
| 24 |
const int8_t* restrict a, |
| 25 |
size_t a_stride, |
| 26 |
const void* restrict w, |
| 27 |
float* restrict c, |
| 28 |
size_t cm_stride, |
| 29 |
size_t cn_stride, |
| 30 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 31 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 32 |
{ |
| 33 |
assert(mr != 0); |
| 34 |
assert(mr <= 4); |
| 35 |
assert(nc != 0); |
| 36 |
assert(kc != 0); |
| 37 |
assert(kc % sizeof(int8_t) == 0); |
| 38 |
assert(a != NULL); |
| 39 |
assert(w != NULL); |
| 40 |
assert(c != NULL); |
| 41 |
|
| 42 |
kc = round_up_po2(kc, 8 * sizeof(int8_t)); |
| 43 |
const int8_t* a0 = a; |
| 44 |
float* c0 = c; |
| 45 |
const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 46 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 47 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 48 |
a1 = a0; |
| 49 |
c1 = c0; |
| 50 |
} |
| 51 |
const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 52 |
float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 53 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 54 |
a2 = a1; |
| 55 |
c2 = c1; |
| 56 |
} |
| 57 |
const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride); |
| 58 |
float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 59 |
if XNN_UNPREDICTABLE(mr != 4) { |
| 60 |
a3 = a2; |
| 61 |
c3 = c2; |
| 62 |
} |
| 63 |
|
| 64 |
const __m256i vsign_mask = _mm256_set1_epi8(0x80); |
| 65 |
XNN_FORCE_REALIZATION(vsign_mask); |
| 66 |
const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128); |
| 67 |
const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128); |
| 68 |
const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128); |
| 69 |
const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128); |
| 70 |
const __m256 voutput_min = _mm256_set1_ps(params->scalar.min); |
| 71 |
const __m256 voutput_max = _mm256_set1_ps(params->scalar.max); |
| 72 |
// XNN_FORCE_REALIZATION(voutput_min); |
| 73 |
// XNN_FORCE_REALIZATION(voutput_max); |
| 74 |
do { |
| 75 |
const __m256i vksum01234567 = _mm256_load_si256(w); |
| 76 |
__m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0); |
| 77 |
__m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0)); |
| 78 |
__m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1)); |
| 79 |
__m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1); |
| 80 |
__m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0)); |
| 81 |
__m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1)); |
| 82 |
__m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2); |
| 83 |
__m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0)); |
| 84 |
__m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1)); |
| 85 |
__m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3); |
| 86 |
__m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0)); |
| 87 |
__m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1)); |
| 88 |
w = (const int32_t*) w + 8; |
| 89 |
|
| 90 |
size_t k = kc; |
| 91 |
while (k >= 16 * sizeof(int8_t)) { |
| 92 |
const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask); |
| 93 |
const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask); |
| 94 |
a0 += 16; |
| 95 |
const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask); |
| 96 |
const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask); |
| 97 |
a1 += 16; |
| 98 |
const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask); |
| 99 |
const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask); |
| 100 |
a2 += 16; |
| 101 |
const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask); |
| 102 |
const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask); |
| 103 |
a3 += 16; |
| 104 |
|
| 105 |
const __m256i vb01234567x0123 = _mm256_load_si256(w); |
| 106 |
const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)); |
| 107 |
const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)); |
| 108 |
const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)); |
| 109 |
|
| 110 |
vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123); |
| 111 |
vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123); |
| 112 |
vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123); |
| 113 |
vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123); |
| 114 |
vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123); |
| 115 |
vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123); |
| 116 |
vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123); |
| 117 |
vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123); |
| 118 |
vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567); |
| 119 |
vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567); |
| 120 |
vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567); |
| 121 |
vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567); |
| 122 |
vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567); |
| 123 |
vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567); |
| 124 |
vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567); |
| 125 |
vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567); |
| 126 |
|
| 127 |
w = (const int8_t*) w + 128; |
| 128 |
k -= 16 * sizeof(int8_t); |
| 129 |
} |
| 130 |
|
| 131 |
if (k != 0) { |
| 132 |
const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask); |
| 133 |
a0 += 8; |
| 134 |
const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask); |
| 135 |
a1 += 8; |
| 136 |
const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask); |
| 137 |
a2 += 8; |
| 138 |
const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask); |
| 139 |
a3 += 8; |
| 140 |
|
| 141 |
const __m256i vb01234567x0123 = _mm256_load_si256(w); |
| 142 |
const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)); |
| 143 |
|
| 144 |
vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123); |
| 145 |
vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123); |
| 146 |
vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123); |
| 147 |
vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123); |
| 148 |
vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123); |
| 149 |
vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123); |
| 150 |
vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123); |
| 151 |
vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123); |
| 152 |
|
| 153 |
w = (const int8_t*) w + 64; |
| 154 |
k -= 8 * sizeof(int8_t); |
| 155 |
} |
| 156 |
|
| 157 |
// Add adjacent pairs |
| 158 |
const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567); |
| 159 |
__m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0)); |
| 160 |
const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567); |
| 161 |
__m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0)); |
| 162 |
const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567); |
| 163 |
__m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0)); |
| 164 |
const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567); |
| 165 |
__m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0)); |
| 166 |
|
| 167 |
__m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567); |
| 168 |
__m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567); |
| 169 |
__m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567); |
| 170 |
__m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567); |
| 171 |
|
| 172 |
vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale)); |
| 173 |
vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale)); |
| 174 |
vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale)); |
| 175 |
vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale)); |
| 176 |
|
| 177 |
const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w); |
| 178 |
const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8); |
| 179 |
w = (const float*) w + 16; |
| 180 |
|
| 181 |
vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567); |
| 182 |
vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567); |
| 183 |
vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567); |
| 184 |
vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567); |
| 185 |
|
| 186 |
vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min); |
| 187 |
vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min); |
| 188 |
vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min); |
| 189 |
vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min); |
| 190 |
|
| 191 |
vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max); |
| 192 |
vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max); |
| 193 |
vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max); |
| 194 |
vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max); |
| 195 |
|
| 196 |
if XNN_LIKELY(nc >= 8) { |
| 197 |
_mm256_storeu_ps(c0, vout0x01234567); |
| 198 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 199 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 200 |
_mm256_storeu_ps(c1, vout1x01234567); |
| 201 |
a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 202 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 203 |
_mm256_storeu_ps(c2, vout2x01234567); |
| 204 |
a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 205 |
c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 206 |
_mm256_storeu_ps(c3, vout3x01234567); |
| 207 |
a3 = (const int8_t*) ((uintptr_t) a3 - kc); |
| 208 |
c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 209 |
nc -= 8; |
| 210 |
} else { |
| 211 |
__m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567); |
| 212 |
__m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567); |
| 213 |
__m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567); |
| 214 |
__m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567); |
| 215 |
if (nc & 4) { |
| 216 |
_mm_storeu_ps(c0, vout0x0123); |
| 217 |
c0 += 4; |
| 218 |
vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1); |
| 219 |
_mm_storeu_ps(c1, vout1x0123); |
| 220 |
c1 += 4; |
| 221 |
vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1); |
| 222 |
_mm_storeu_ps(c2, vout2x0123); |
| 223 |
c2 += 4; |
| 224 |
vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1); |
| 225 |
_mm_storeu_ps(c3, vout3x0123); |
| 226 |
c3 += 4; |
| 227 |
vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1); |
| 228 |
} |
| 229 |
if (nc & 2) { |
| 230 |
_mm_storel_pi((__m64*) c0, vout0x0123); |
| 231 |
c0 += 2; |
| 232 |
vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123); |
| 233 |
_mm_storel_pi((__m64*) c1, vout1x0123); |
| 234 |
c1 += 2; |
| 235 |
vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123); |
| 236 |
_mm_storel_pi((__m64*) c2, vout2x0123); |
| 237 |
c2 += 2; |
| 238 |
vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123); |
| 239 |
_mm_storel_pi((__m64*) c3, vout3x0123); |
| 240 |
c3 += 2; |
| 241 |
vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123); |
| 242 |
} |
| 243 |
if (nc & 1) { |
| 244 |
_mm_store_ss(c0, vout0x0123); |
| 245 |
_mm_store_ss(c1, vout1x0123); |
| 246 |
_mm_store_ss(c2, vout2x0123); |
| 247 |
_mm_store_ss(c3, vout3x0123); |
| 248 |
} |
| 249 |
nc = 0; |
| 250 |
} |
| 251 |
} while (nc != 0); |
| 252 |
} |