Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-8x16c8-minmax-avx512skx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c8-avx512skx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/prefetch.h"
18
19 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_8x16c8__avx512skx_prfm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 8);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 float* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
57 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 a3 = a2;
60 c3 = c2;
61 }
62 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
63 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
64 if XNN_UNPREDICTABLE(mr <= 4) {
65 a4 = a3;
66 c4 = c3;
67 }
68 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
69 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
70 if XNN_UNPREDICTABLE(mr < 6) {
71 a5 = a4;
72 c5 = c4;
73 }
74 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
75 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
76 if XNN_UNPREDICTABLE(mr <= 6) {
77 a6 = a5;
78 c6 = c5;
79 }
80 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
81 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
82 if XNN_UNPREDICTABLE(mr != 8) {
83 a7 = a6;
84 c7 = c6;
85 }
86
87 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
88 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
89 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point);
90 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point);
91 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point);
92 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point);
93 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point);
94 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point);
95 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point);
96 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
97 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
98 // XNN_FORCE_REALIZATION(voutput_min);
99 // XNN_FORCE_REALIZATION(voutput_max);
100
101 do {
102 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
103 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
104 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
105 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
106
107 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
108 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
109 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
110 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
111 __m512i vacc1x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point1);
112 __m512i vacc1x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point1);
113 __m512i vacc1x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point1);
114 __m512i vacc1xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point1);
115 __m512i vacc2x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point2);
116 __m512i vacc2x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point2);
117 __m512i vacc2x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point2);
118 __m512i vacc2xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point2);
119 __m512i vacc3x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point3);
120 __m512i vacc3x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point3);
121 __m512i vacc3x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point3);
122 __m512i vacc3xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point3);
123 __m512i vacc4x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point4);
124 __m512i vacc4x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point4);
125 __m512i vacc4x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point4);
126 __m512i vacc4xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point4);
127 __m512i vacc5x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point5);
128 __m512i vacc5x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point5);
129 __m512i vacc5x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point5);
130 __m512i vacc5xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point5);
131 __m512i vacc6x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point6);
132 __m512i vacc6x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point6);
133 __m512i vacc6x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point6);
134 __m512i vacc6xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point6);
135 __m512i vacc7x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point7);
136 __m512i vacc7x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point7);
137 __m512i vacc7x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point7);
138 __m512i vacc7xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point7);
139 w = (const int32_t*) w + 16;
140
141 size_t k = kc;
142
143 while (k >= 8 * sizeof(int8_t)) {
144 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
145 a0 += 8;
146 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
147 a1 += 8;
148 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
149 a2 += 8;
150 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
151 a3 += 8;
152 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
153 a4 += 8;
154 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
155 a5 += 8;
156 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
157 a6 += 8;
158 const __m512i va7 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a7)));
159 a7 += 8;
160
161 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
162 xnn_prefetch_to_l1((const int8_t*) w + 896);
163
164 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
165 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
166 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
167 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
168 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
169 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
170 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
171 vacc7x0123 = _mm512_add_epi32(vacc7x0123, _mm512_madd_epi16(va7, vb0123));
172 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
173
174 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
175 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
176 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
177 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
178 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
179 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
180 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
181 vacc7x4567 = _mm512_add_epi32(vacc7x4567, _mm512_madd_epi16(va7, vb4567));
182 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
183 xnn_prefetch_to_l1((const int8_t*) w + 960);
184
185 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
186 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
187 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
188 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
189 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
190 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
191 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
192 vacc7x89AB = _mm512_add_epi32(vacc7x89AB, _mm512_madd_epi16(va7, vb89AB));
193 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
194
195 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
196 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
197 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
198 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
199 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
200 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
201 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
202 vacc7xCDEF = _mm512_add_epi32(vacc7xCDEF, _mm512_madd_epi16(va7, vbCDEF));
203
204 w = (const int8_t*) w + 128;
205 k -= 8 * sizeof(int8_t);
206 }
207
208 // Add 4 adjacent sums
209 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
210 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
211 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
212 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
213 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
214 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
215 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
216 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
217 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
218 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
219 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
220 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
221 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
222 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
223 const __m512i vacc7x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x0123, vacc7x4567), _mm512_unpackhi_epi32(vacc7x0123, vacc7x4567));
224 const __m512i vacc7x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x89AB, vacc7xCDEF), _mm512_unpackhi_epi32(vacc7x89AB, vacc7xCDEF));
225
226 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
227 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
228 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
229 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
230 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
231 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
232 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
233 const __m512i vacc7x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x04152637, vacc7x8C9DAEBF), _mm512_unpackhi_epi32(vacc7x04152637, vacc7x8C9DAEBF));
234
235 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
236 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
237 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
238 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
239 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
240 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
241 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
242 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
243 __m512i vacc7x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc7x084C195D2A6E3B7F);
244
245 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
246 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
247 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
248 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
249 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
250 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
251 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
252 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
253
254 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
255 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
256 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
257 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
258 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
259 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
260 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
261 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
262
263 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
264 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
265 w = (const float*) w + 32;
266
267 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
268 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
269 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
270 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
271 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
272 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
273 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
274 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
275
276 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
277 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
278 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
279 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
280 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
281 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
282 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
283 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
284
285 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
286 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
287 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
288 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
289 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
290 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
291 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
292 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
293
294 if XNN_LIKELY(nc >= 16) {
295 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
296 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
297 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
298 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
299 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
300 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
301 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
302 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
303
304 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
305 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
306 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
307 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
308 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
309 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
310 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
311 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
312
313 c0 = (float*) ((uintptr_t) c0 + cn_stride);
314 c1 = (float*) ((uintptr_t) c1 + cn_stride);
315 c2 = (float*) ((uintptr_t) c2 + cn_stride);
316 c3 = (float*) ((uintptr_t) c3 + cn_stride);
317 c4 = (float*) ((uintptr_t) c4 + cn_stride);
318 c5 = (float*) ((uintptr_t) c5 + cn_stride);
319 c6 = (float*) ((uintptr_t) c6 + cn_stride);
320 c7 = (float*) ((uintptr_t) c7 + cn_stride);
321
322 nc -= 16;
323 } else {
324 // Prepare mask for valid 32-bit elements (depends on nc).
325 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
326 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
327 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
328 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
329 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
330 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
331 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
332 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
333 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
334 nc = 0;
335 }
336 } while (nc != 0);
337 }