Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-12x16c8-minmax-avx512vnni-prfm.c

raw
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_qc8w_gemm_minmax_ukernel_12x16c8__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 union xnn_f32_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 <= 12);
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
114 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
115 XNN_FORCE_REALIZATION(vsign_mask);
116 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
117 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
118 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
119 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
120 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
121 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
122 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
123 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point + 128);
124 const __m512i vinput_zero_point8 = _mm512_set1_epi32((int) quantization_params[8].zero_point + 128);
125 const __m512i vinput_zero_point9 = _mm512_set1_epi32((int) quantization_params[9].zero_point + 128);
126 const __m512i vinput_zero_point10 = _mm512_set1_epi32((int) quantization_params[10].zero_point + 128);
127 const __m512i vinput_zero_point11 = _mm512_set1_epi32((int) quantization_params[11].zero_point + 128);
128 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
129 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
130 // XNN_FORCE_REALIZATION(voutput_min);
131 // XNN_FORCE_REALIZATION(voutput_max);
132 do {
133 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
134 __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
135 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
136 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
137 __m512i vsum1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
138 __m512i vacc1x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum1x0123456789ABCDEF, 0));
139 __m512i vacc1x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum1x0123456789ABCDEF, 1));
140 __m512i vsum2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
141 __m512i vacc2x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum2x0123456789ABCDEF, 0));
142 __m512i vacc2x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum2x0123456789ABCDEF, 1));
143 __m512i vsum3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
144 __m512i vacc3x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum3x0123456789ABCDEF, 0));
145 __m512i vacc3x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum3x0123456789ABCDEF, 1));
146 __m512i vsum4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
147 __m512i vacc4x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum4x0123456789ABCDEF, 0));
148 __m512i vacc4x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum4x0123456789ABCDEF, 1));
149 __m512i vsum5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
150 __m512i vacc5x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum5x0123456789ABCDEF, 0));
151 __m512i vacc5x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum5x0123456789ABCDEF, 1));
152 __m512i vsum6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
153 __m512i vacc6x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum6x0123456789ABCDEF, 0));
154 __m512i vacc6x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum6x0123456789ABCDEF, 1));
155 __m512i vsum7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point7);
156 __m512i vacc7x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum7x0123456789ABCDEF, 0));
157 __m512i vacc7x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum7x0123456789ABCDEF, 1));
158 __m512i vsum8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point8);
159 __m512i vacc8x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum8x0123456789ABCDEF, 0));
160 __m512i vacc8x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum8x0123456789ABCDEF, 1));
161 __m512i vsum9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point9);
162 __m512i vacc9x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum9x0123456789ABCDEF, 0));
163 __m512i vacc9x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum9x0123456789ABCDEF, 1));
164 __m512i vsum10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point10);
165 __m512i vacc10x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum10x0123456789ABCDEF, 0));
166 __m512i vacc10x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum10x0123456789ABCDEF, 1));
167 __m512i vsum11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point11);
168 __m512i vacc11x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum11x0123456789ABCDEF, 0));
169 __m512i vacc11x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum11x0123456789ABCDEF, 1));
170 w = (const int32_t*) w + 16;
171
172 size_t k = kc;
173 while (k >= 16 * sizeof(int8_t)) {
174 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
175 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
176 a0 += 16;
177 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
178 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
179 a1 += 16;
180 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
181 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
182 a2 += 16;
183 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
184 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
185 a3 += 16;
186 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
187 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
188 a4 += 16;
189 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
190 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
191 a5 += 16;
192 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
193 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
194 a6 += 16;
195 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
196 const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
197 a7 += 16;
198 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
199 const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
200 a8 += 16;
201 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
202 const __m512i va9x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
203 a9 += 16;
204 const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask);
205 const __m512i va10x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
206 a10 += 16;
207 const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask);
208 const __m512i va11x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
209 a11 += 16;
210
211 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
212 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
213 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
214 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
215 xnn_prefetch_to_l1((const int8_t*) w + 768);
216 xnn_prefetch_to_l1((const int8_t*) w + 832);
217
218 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
219 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
220 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
221 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
222 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
223 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
224 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
225 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
226 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
227 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
228 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
229 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
230 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
231 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
232 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
233 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
234 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
235 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
236 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
237 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
238 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567);
239 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567);
240 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567);
241 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567);
242 xnn_prefetch_to_l1((const int8_t*) w + 896);
243 xnn_prefetch_to_l1((const int8_t*) w + 960);
244 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
245 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
246 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
247 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
248 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
249 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
250 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
251 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
252 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
253 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
254 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
255 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
256 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
257 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
258 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF);
259 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF);
260 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF);
261 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF);
262 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x89ABCDEF, vb01234567x89ABCDEF);
263 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x89ABCDEF, vb89ABCDEFx89ABCDEF);
264 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x89ABCDEF, vb01234567x89ABCDEF);
265 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x89ABCDEF, vb89ABCDEFx89ABCDEF);
266 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x89ABCDEF, vb01234567x89ABCDEF);
267 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x89ABCDEF, vb89ABCDEFx89ABCDEF);
268
269 w = (const int8_t*) w + 256;
270 k -= 16 * sizeof(int8_t);
271 }
272
273 if (k != 0) {
274 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
275 a0 += 8;
276 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
277 a1 += 8;
278 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
279 a2 += 8;
280 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
281 a3 += 8;
282 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
283 a4 += 8;
284 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
285 a5 += 8;
286 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
287 a6 += 8;
288 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
289 a7 += 8;
290 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
291 a8 += 8;
292 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
293 a9 += 8;
294 const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask);
295 a10 += 8;
296 const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask);
297 a11 += 8;
298
299 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
300 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
301
302 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
303 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
304 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
305 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
306 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
307 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
308 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
309 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
310 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
311 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
312 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
313 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
314 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
315 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
316 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
317 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
318 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
319 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
320 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
321 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
322 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567);
323 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567);
324 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567);
325 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567);
326 xnn_prefetch_to_l1((const int8_t*) w + 896);
327 xnn_prefetch_to_l1((const int8_t*) w + 960);
328
329 w = (const int8_t*) w + 128;
330 k -= 8 * sizeof(int8_t);
331 }
332
333 // Add adjacent pairs
334 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
335 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
336 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
337 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
338 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
339 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
340 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
341 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
342 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
343 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
344 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
345 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
346 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
347 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
348 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
349 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
350 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
351 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
352 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
353 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
354 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
355 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
356 const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32));
357 const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32));
358 __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF);
359 const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32));
360 const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32));
361 __m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF);
362 const __m512i vsum9x01234567 = _mm512_add_epi32(vacc9x01234567, _mm512_srli_epi64(vacc9x01234567, 32));
363 const __m512i vsum9x89ABCDEF = _mm512_add_epi32(vacc9x89ABCDEF, _mm512_srli_epi64(vacc9x89ABCDEF, 32));
364 __m512i vacc9x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum9x01234567, vidx, vsum9x89ABCDEF);
365 const __m512i vsum10x01234567 = _mm512_add_epi32(vacc10x01234567, _mm512_srli_epi64(vacc10x01234567, 32));
366 const __m512i vsum10x89ABCDEF = _mm512_add_epi32(vacc10x89ABCDEF, _mm512_srli_epi64(vacc10x89ABCDEF, 32));
367 __m512i vacc10x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum10x01234567, vidx, vsum10x89ABCDEF);
368 const __m512i vsum11x01234567 = _mm512_add_epi32(vacc11x01234567, _mm512_srli_epi64(vacc11x01234567, 32));
369 const __m512i vsum11x89ABCDEF = _mm512_add_epi32(vacc11x89ABCDEF, _mm512_srli_epi64(vacc11x89ABCDEF, 32));
370 __m512i vacc11x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum11x01234567, vidx, vsum11x89ABCDEF);
371
372 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
373 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
374 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
375 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
376 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
377 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
378 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
379 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
380 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
381 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
382 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
383 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
384
385 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
386 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
387 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
388 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
389 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
390 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
391 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
392 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
393 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale));
394 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params[9].inv_scale));
395 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params[10].inv_scale));
396 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params[11].inv_scale));
397
398 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
399 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
400 w = (const float*) w + 32;
401
402 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
403 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
404 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
405 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
406 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
407 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
408 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
409 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
410 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
411 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
412 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
413 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
414
415 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
416 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
417 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
418 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
419 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
420 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
421 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
422 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
423 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
424 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
425 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
426 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
427
428 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
429 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
430 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
431 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
432 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
433 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
434 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
435 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
436 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
437 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
438 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
439 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
440
441 if XNN_LIKELY(nc >= 16) {
442 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
443 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
444 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
445 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
446 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
447 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
448 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
449 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
450 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
451 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
452 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
453 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
454
455 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
456 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
457 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
458 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
459 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
460 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
461 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
462 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
463 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
464 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
465 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
466 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
467
468 c0 = (float*) ((uintptr_t) c0 + cn_stride);
469 c1 = (float*) ((uintptr_t) c1 + cn_stride);
470 c2 = (float*) ((uintptr_t) c2 + cn_stride);
471 c3 = (float*) ((uintptr_t) c3 + cn_stride);
472 c4 = (float*) ((uintptr_t) c4 + cn_stride);
473 c5 = (float*) ((uintptr_t) c5 + cn_stride);
474 c6 = (float*) ((uintptr_t) c6 + cn_stride);
475 c7 = (float*) ((uintptr_t) c7 + cn_stride);
476 c8 = (float*) ((uintptr_t) c8 + cn_stride);
477 c9 = (float*) ((uintptr_t) c9 + cn_stride);
478 c10 = (float*) ((uintptr_t) c10 + cn_stride);
479 c11 = (float*) ((uintptr_t) c11 + cn_stride);
480
481 nc -= 16;
482 } else {
483 // Prepare mask for valid 32-bit elements (depends on nc).
484 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
485 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
486 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
487 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
488 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
489 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
490 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
491 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
492 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
493 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
494 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
495 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
496 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
497 nc = 0;
498 }
499 } while (nc != 0);
500 }