Debian Codegraph @

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

raw
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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_8x16c4__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 <= 8);
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, 4 * 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
90 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
91 XNN_FORCE_REALIZATION(vsign_mask);
92 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
93 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
94 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
95 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
96 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
97 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
98 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
99 const __m512i vinput_zero_point7 = _mm512_set1_epi32((int) quantization_params[7].zero_point + 128);
100 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
101 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
102 // XNN_FORCE_REALIZATION(voutput_min);
103 // XNN_FORCE_REALIZATION(voutput_max);
104 do {
105 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
106 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
107 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
108 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
109 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
110 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
111 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
112 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
113 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
114 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
115 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
116 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
117 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
118 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
119 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
120 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point7);
121 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
122 w = (const int32_t*) w + 16;
123
124 size_t k = kc;
125 while (k >= 8 * sizeof(int8_t)) {
126 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
127 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
128 a0 += 8;
129 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
130 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
131 a1 += 8;
132 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
133 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
134 a2 += 8;
135 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
136 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
137 a3 += 8;
138 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
139 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
140 a4 += 8;
141 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
142 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
143 a5 += 8;
144 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
145 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
146 a6 += 8;
147 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
148 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
149 a7 += 8;
150
151 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
152 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
153 xnn_prefetch_to_l1((const int8_t*) w + 896);
154
155 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
156 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
157 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
158 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
159 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
160 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
161 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
162 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
163 xnn_prefetch_to_l1((const int8_t*) w + 960);
164 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
165 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
166 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
167 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
168 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
169 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
170 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
171 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
172
173 w = (const int8_t*) w + 128;
174 k -= 8 * sizeof(int8_t);
175 }
176 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
177 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
178 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
179 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
180 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
181 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
182 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
183 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
184
185 if (k != 0) {
186 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
187 a0 += 4;
188 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
189 a1 += 4;
190 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
191 a2 += 4;
192 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
193 a3 += 4;
194 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
195 a4 += 4;
196 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
197 a5 += 4;
198 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
199 a6 += 4;
200 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
201 a7 += 4;
202
203 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
204 xnn_prefetch_to_l1((const int8_t*) w + 960);
205
206 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
207 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
208 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
209 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
210 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
211 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
212 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
213 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
214
215 w = (const int8_t*) w + 64;
216 k -= 4 * sizeof(int8_t);
217 }
218
219 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
220 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
221 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
222 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
223 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
224 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
225 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
226 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
227
228 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
229 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
230 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
231 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
232 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
233 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
234 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
235 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale));
236
237 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
238 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
239 w = (const float*) w + 32;
240
241 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
242 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
243 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
244 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
245 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
246 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
247 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
248 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
249
250 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
251 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
252 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
253 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
254 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
255 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
256 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
257 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
258
259 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
260 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
261 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
262 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
263 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
264 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
265 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
266 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
267
268 if XNN_LIKELY(nc >= 16) {
269 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
270 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
271 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
272 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
273 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
274 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
275 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
276 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
277
278 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
279 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
280 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
281 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
282 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
283 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
284 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
285 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
286
287 c0 = (float*) ((uintptr_t) c0 + cn_stride);
288 c1 = (float*) ((uintptr_t) c1 + cn_stride);
289 c2 = (float*) ((uintptr_t) c2 + cn_stride);
290 c3 = (float*) ((uintptr_t) c3 + cn_stride);
291 c4 = (float*) ((uintptr_t) c4 + cn_stride);
292 c5 = (float*) ((uintptr_t) c5 + cn_stride);
293 c6 = (float*) ((uintptr_t) c6 + cn_stride);
294 c7 = (float*) ((uintptr_t) c7 + cn_stride);
295
296 nc -= 16;
297 } else {
298 // Prepare mask for valid 32-bit elements (depends on nc).
299 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
300 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
301 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
302 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
303 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
304 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
305 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
306 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
307 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
308 nc = 0;
309 }
310 } while (nc != 0);
311 }