Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-7x8c8-minmax-avx256vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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/unaligned.h"
18
19
20 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_7x8c8__avx256vnni(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 7);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
58 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr < 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
64 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
65 if XNN_UNPREDICTABLE(mr <= 4) {
66 a4 = a3;
67 c4 = c3;
68 }
69 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
70 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
71 if XNN_UNPREDICTABLE(mr < 6) {
72 a5 = a4;
73 c5 = c4;
74 }
75 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
76 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
77 if XNN_UNPREDICTABLE(mr <= 6) {
78 a6 = a5;
79 c6 = c5;
80 }
81
82 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
83 XNN_FORCE_REALIZATION(vsign_mask);
84 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
85 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
86 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
87 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
88 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
89 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
90 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
91 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
92 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
93 // XNN_FORCE_REALIZATION(voutput_min);
94 // XNN_FORCE_REALIZATION(voutput_max);
95 do {
96 const __m256i vksum01234567 = _mm256_load_si256(w);
97 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
98 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
99 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
100 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
101 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
102 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
103 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
104 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
105 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
106 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
107 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
108 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
109 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
110 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
111 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
112 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
113 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
114 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
115 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
116 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
117 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
118 w = (const int32_t*) w + 8;
119
120 size_t k = kc;
121 while (k >= 16 * sizeof(int8_t)) {
122 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
123 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
124 a0 += 16;
125 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
126 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
127 a1 += 16;
128 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
129 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
130 a2 += 16;
131 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
132 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
133 a3 += 16;
134 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
135 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
136 a4 += 16;
137 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
138 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
139 a5 += 16;
140 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
141 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
142 a6 += 16;
143
144 const __m256i vb01234567x0123 = _mm256_load_si256(w);
145 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
146 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
147 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
148
149 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
150 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
151 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
152 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
153 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
154 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
155 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
156 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
157 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
158 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
159 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
160 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
161 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
162 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
163 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
164 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
165 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
166 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
167 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
168 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
169 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
170 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
171 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
172 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
173 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
174 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
175 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
176 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
177
178 w = (const int8_t*) w + 128;
179 k -= 16 * sizeof(int8_t);
180 }
181
182 if (k != 0) {
183 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
184 a0 += 8;
185 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
186 a1 += 8;
187 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
188 a2 += 8;
189 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
190 a3 += 8;
191 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
192 a4 += 8;
193 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
194 a5 += 8;
195 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
196 a6 += 8;
197
198 const __m256i vb01234567x0123 = _mm256_load_si256(w);
199 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
200
201 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
202 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
203 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
204 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
205 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
206 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
207 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
208 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
209 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
210 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
211 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
212 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
213 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
214 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
215
216 w = (const int8_t*) w + 64;
217 k -= 8 * sizeof(int8_t);
218 }
219
220 // Add adjacent pairs
221 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
222 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
223 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
224 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
225 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
226 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
227 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
228 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
229 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
230 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
231 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
232 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
233 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
234 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
235
236 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
237 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
238 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
239 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
240 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
241 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
242 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
243
244 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
245 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
246 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
247 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
248 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
249 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
250 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
251
252 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
253 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
254 w = (const float*) w + 16;
255
256 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
257 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
258 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
259 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
260 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
261 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
262 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
263
264 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
265 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
266 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
267 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
268 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
269 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
270 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
271
272 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
273 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
274 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
275 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
276 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
277 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
278 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
279
280 if XNN_LIKELY(nc >= 8) {
281 _mm256_storeu_ps(c0, vout0x01234567);
282 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
283 c0 = (float*) ((uintptr_t) c0 + cn_stride);
284 _mm256_storeu_ps(c1, vout1x01234567);
285 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
286 c1 = (float*) ((uintptr_t) c1 + cn_stride);
287 _mm256_storeu_ps(c2, vout2x01234567);
288 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
289 c2 = (float*) ((uintptr_t) c2 + cn_stride);
290 _mm256_storeu_ps(c3, vout3x01234567);
291 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
292 c3 = (float*) ((uintptr_t) c3 + cn_stride);
293 _mm256_storeu_ps(c4, vout4x01234567);
294 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
295 c4 = (float*) ((uintptr_t) c4 + cn_stride);
296 _mm256_storeu_ps(c5, vout5x01234567);
297 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
298 c5 = (float*) ((uintptr_t) c5 + cn_stride);
299 _mm256_storeu_ps(c6, vout6x01234567);
300 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
301 c6 = (float*) ((uintptr_t) c6 + cn_stride);
302 nc -= 8;
303 } else {
304 // Prepare mask for valid 32-bit elements (depends on nc).
305 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
306 _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
307 _mm256_mask_storeu_ps(c1, vmask, vout1x01234567);
308 _mm256_mask_storeu_ps(c2, vmask, vout2x01234567);
309 _mm256_mask_storeu_ps(c3, vmask, vout3x01234567);
310 _mm256_mask_storeu_ps(c4, vmask, vout4x01234567);
311 _mm256_mask_storeu_ps(c5, vmask, vout5x01234567);
312 _mm256_mask_storeu_ps(c6, vmask, vout6x01234567);
313 nc = 0;
314 }
315 } while (nc != 0);
316 }