Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-8x16c4-minmax-avx512vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_ukernel_8x16c4__avx512vnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 8);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 4 * sizeof(int8_t));
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
65 if XNN_UNPREDICTABLE(mr < 6) {
66 c5 = c4;
67 }
68 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
69 if XNN_UNPREDICTABLE(mr <= 6) {
70 c6 = c5;
71 }
72 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
73 if XNN_UNPREDICTABLE(mr != 8) {
74 c7 = c6;
75 }
76
77 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
78 XNN_FORCE_REALIZATION(vsign_mask);
79 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
80 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
81 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
82 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
83 XNN_FORCE_REALIZATION(vinput_zero_point);
84 // XNN_FORCE_REALIZATION(vinput_inv_scale);
85 // XNN_FORCE_REALIZATION(voutput_min);
86 // XNN_FORCE_REALIZATION(voutput_max);
87 do {
88 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
89 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
90 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
91 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
92 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
93 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
94 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
95 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
96 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
97 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
98 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
99 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
100 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
101 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
102 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
103 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
104 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
105 w = (const int32_t*) w + 16;
106
107 size_t p = ks;
108 do {
109 const int8_t* restrict a0 = a[0];
110 if XNN_UNPREDICTABLE(a0 != zero) {
111 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
112 } else {
113 a0 = zero_data;
114 }
115 const int8_t* restrict a1 = a[1];
116 if XNN_UNPREDICTABLE(a1 != zero) {
117 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
118 } else {
119 a1 = zero_data;
120 }
121 const int8_t* restrict a2 = a[2];
122 if XNN_UNPREDICTABLE(a2 != zero) {
123 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
124 } else {
125 a2 = zero_data;
126 }
127 const int8_t* restrict a3 = a[3];
128 if XNN_UNPREDICTABLE(a3 != zero) {
129 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
130 } else {
131 a3 = zero_data;
132 }
133 const int8_t* restrict a4 = a[4];
134 if XNN_UNPREDICTABLE(a4 != zero) {
135 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
136 } else {
137 a4 = zero_data;
138 }
139 const int8_t* restrict a5 = a[5];
140 if XNN_UNPREDICTABLE(a5 != zero) {
141 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
142 } else {
143 a5 = zero_data;
144 }
145 const int8_t* restrict a6 = a[6];
146 if XNN_UNPREDICTABLE(a6 != zero) {
147 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
148 } else {
149 a6 = zero_data;
150 }
151 const int8_t* restrict a7 = a[7];
152 if XNN_UNPREDICTABLE(a7 != zero) {
153 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
154 } else {
155 a7 = zero_data;
156 }
157 a += 8;
158
159 size_t k = kc;
160 while (k >= 8 * sizeof(int8_t)) {
161 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
162 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
163 a0 += 8;
164 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
165 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
166 a1 += 8;
167 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
168 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
169 a2 += 8;
170 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
171 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
172 a3 += 8;
173 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
174 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
175 a4 += 8;
176 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
177 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
178 a5 += 8;
179 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
180 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
181 a6 += 8;
182 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
183 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
184 a7 += 8;
185
186 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
187 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
188
189 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
190 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
191 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
192 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
193 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
194 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
195 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
196 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
197 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
198 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
199 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
200 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
201 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
202 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
203 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
204 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
205
206 w = (const int8_t*) w + 128;
207 k -= 8 * sizeof(int8_t);
208 }
209
210 if (k != 0) {
211 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
212 a0 += 4;
213 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
214 a1 += 4;
215 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
216 a2 += 4;
217 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
218 a3 += 4;
219 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
220 a4 += 4;
221 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
222 a5 += 4;
223 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
224 a6 += 4;
225 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
226 a7 += 4;
227
228 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
229
230 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
231 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
232 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
233 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
234 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
235 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
236 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
237 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
238
239 w = (const int8_t*) w + 64;
240 k -= 4 * sizeof(int8_t);
241 }
242
243 p -= 8 * sizeof(void*);
244 } while (p != 0);
245
246 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
247 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
248 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
249 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
250 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
251 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
252 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
253 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
254
255 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
256 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
257 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
258 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
259 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
260 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
261 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
262 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
263
264 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
265 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
266 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
267 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
268 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
269 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
270 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
271 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vinput_inv_scale);
272
273 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
274 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
275 w = (const float*) w + 32;
276
277 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
278 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
279 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
280 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
281 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
282 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
283 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
284 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
285
286 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
287 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
288 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
289 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
290 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
291 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
292 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
293 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
294
295 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
296 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
297 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
298 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
299 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
300 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
301 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
302 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
303
304 if XNN_LIKELY(nc >= 16) {
305 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
306 c7 = (float*) ((uintptr_t) c7 + cn_stride);
307 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
308 c6 = (float*) ((uintptr_t) c6 + cn_stride);
309 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
310 c5 = (float*) ((uintptr_t) c5 + cn_stride);
311 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
312 c4 = (float*) ((uintptr_t) c4 + cn_stride);
313 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
314 c3 = (float*) ((uintptr_t) c3 + cn_stride);
315 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
316 c2 = (float*) ((uintptr_t) c2 + cn_stride);
317 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
318 c1 = (float*) ((uintptr_t) c1 + cn_stride);
319 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
320 c0 = (float*) ((uintptr_t) c0 + cn_stride);
321 a = (const int8_t**restrict) ((uintptr_t) a - ks);
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(c7, vmask, vscaled7x0123456789ABCDEF);
327 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
328 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
329 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
330 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
331 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
332 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
333 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
334 nc = 0;
335 }
336 } while (nc != 0);
337 }