| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/MRx16c4-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 |
|
| 20 |
|
| 21 |
void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x16c4__avx512vnni( |
| 22 |
size_t mr, |
| 23 |
size_t nc, |
| 24 |
size_t kc, |
| 25 |
const int8_t* restrict a, |
| 26 |
size_t a_stride, |
| 27 |
const void* restrict w, |
| 28 |
float* restrict c, |
| 29 |
size_t cm_stride, |
| 30 |
size_t cn_stride, |
| 31 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 32 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 33 |
{ |
| 34 |
assert(mr != 0); |
| 35 |
assert(mr <= 4); |
| 36 |
assert(nc != 0); |
| 37 |
assert(kc != 0); |
| 38 |
assert(kc % 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 |
const int8_t* a0 = a; |
| 45 |
float* c0 = c; |
| 46 |
const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 47 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 48 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 49 |
a1 = a0; |
| 50 |
c1 = c0; |
| 51 |
} |
| 52 |
const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 53 |
float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 54 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 55 |
a2 = a1; |
| 56 |
c2 = c1; |
| 57 |
} |
| 58 |
const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride); |
| 59 |
float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 60 |
if XNN_UNPREDICTABLE(mr != 4) { |
| 61 |
a3 = a2; |
| 62 |
c3 = c2; |
| 63 |
} |
| 64 |
|
| 65 |
const __m512i vsign_mask = _mm512_set1_epi8(0x80); |
| 66 |
XNN_FORCE_REALIZATION(vsign_mask); |
| 67 |
const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128); |
| 68 |
const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128); |
| 69 |
const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128); |
| 70 |
const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128); |
| 71 |
const __m512 voutput_min = _mm512_set1_ps(params->scalar.min); |
| 72 |
const __m512 voutput_max = _mm512_set1_ps(params->scalar.max); |
| 73 |
// XNN_FORCE_REALIZATION(voutput_min); |
| 74 |
// XNN_FORCE_REALIZATION(voutput_max); |
| 75 |
do { |
| 76 |
const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w); |
| 77 |
__m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0); |
| 78 |
__m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32(); |
| 79 |
__m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1); |
| 80 |
__m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32(); |
| 81 |
__m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2); |
| 82 |
__m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32(); |
| 83 |
__m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3); |
| 84 |
__m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32(); |
| 85 |
w = (const int32_t*) w + 16; |
| 86 |
|
| 87 |
size_t k = kc; |
| 88 |
while (k >= 8 * sizeof(int8_t)) { |
| 89 |
const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask); |
| 90 |
const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask); |
| 91 |
a0 += 8; |
| 92 |
const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask); |
| 93 |
const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask); |
| 94 |
a1 += 8; |
| 95 |
const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask); |
| 96 |
const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask); |
| 97 |
a2 += 8; |
| 98 |
const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask); |
| 99 |
const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask); |
| 100 |
a3 += 8; |
| 101 |
|
| 102 |
const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w); |
| 103 |
const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64); |
| 104 |
|
| 105 |
vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123); |
| 106 |
vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123); |
| 107 |
vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123); |
| 108 |
vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123); |
| 109 |
vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567); |
| 110 |
vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567); |
| 111 |
vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567); |
| 112 |
vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567); |
| 113 |
|
| 114 |
w = (const int8_t*) w + 128; |
| 115 |
k -= 8 * sizeof(int8_t); |
| 116 |
} |
| 117 |
vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF); |
| 118 |
vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF); |
| 119 |
vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF); |
| 120 |
vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF); |
| 121 |
|
| 122 |
if (k != 0) { |
| 123 |
const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask); |
| 124 |
a0 += 4; |
| 125 |
const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask); |
| 126 |
a1 += 4; |
| 127 |
const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask); |
| 128 |
a2 += 4; |
| 129 |
const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask); |
| 130 |
a3 += 4; |
| 131 |
|
| 132 |
const __m512i vb0123456789ABCDEF = _mm512_load_si512(w); |
| 133 |
|
| 134 |
vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF); |
| 135 |
vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF); |
| 136 |
vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF); |
| 137 |
vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF); |
| 138 |
|
| 139 |
w = (const int8_t*) w + 64; |
| 140 |
k -= 4 * sizeof(int8_t); |
| 141 |
} |
| 142 |
|
| 143 |
__m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF); |
| 144 |
__m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF); |
| 145 |
__m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF); |
| 146 |
__m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF); |
| 147 |
|
| 148 |
vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale)); |
| 149 |
vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale)); |
| 150 |
vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale)); |
| 151 |
vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale)); |
| 152 |
|
| 153 |
const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w); |
| 154 |
const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16); |
| 155 |
w = (const float*) w + 32; |
| 156 |
|
| 157 |
vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF); |
| 158 |
vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF); |
| 159 |
vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF); |
| 160 |
vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF); |
| 161 |
|
| 162 |
vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min); |
| 163 |
vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min); |
| 164 |
vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min); |
| 165 |
vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min); |
| 166 |
|
| 167 |
vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max); |
| 168 |
vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max); |
| 169 |
vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max); |
| 170 |
vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max); |
| 171 |
|
| 172 |
if XNN_LIKELY(nc >= 16) { |
| 173 |
_mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF); |
| 174 |
_mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF); |
| 175 |
_mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF); |
| 176 |
_mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF); |
| 177 |
|
| 178 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 179 |
a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 180 |
a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 181 |
a3 = (const int8_t*) ((uintptr_t) a3 - kc); |
| 182 |
|
| 183 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 184 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 185 |
c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 186 |
c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 187 |
|
| 188 |
nc -= 16; |
| 189 |
} else { |
| 190 |
// Prepare mask for valid 32-bit elements (depends on nc). |
| 191 |
const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1); |
| 192 |
_mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF); |
| 193 |
_mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF); |
| 194 |
_mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF); |
| 195 |
_mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF); |
| 196 |
nc = 0; |
| 197 |
} |
| 198 |
} while (nc != 0); |
| 199 |
} |