Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-8x8c8-minmax-avx256skx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avx2.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/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_gemm_minmax_ukernel_8x8c8__avx256skx(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 8);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr < 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
65 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
71 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
72 if XNN_UNPREDICTABLE(mr < 6) {
73 a5 = a4;
74 c5 = c4;
75 }
76 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
77 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 6) {
79 a6 = a5;
80 c6 = c5;
81 }
82 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
83 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
84 if XNN_UNPREDICTABLE(mr != 8) {
85 a7 = a6;
86 c7 = c6;
87 }
88
89 const __m256 vmin = _mm256_set1_ps(params->scalar.min);
90 const __m256 vmax = _mm256_set1_ps(params->scalar.max);
91 XNN_FORCE_REALIZATION(vmin);
92 XNN_FORCE_REALIZATION(vmax);
93
94 do {
95 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
96 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
97 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
98 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
99 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
100 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
101 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
102 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
103 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
104 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
105 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
106 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
107 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
108 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
109 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
110 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
111 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
112 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
113 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
114 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
115 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
116 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
117 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
118 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
119 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
120 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
121 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
122 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
123 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
124 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
125 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
126 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
127 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point);
128 __m256i vacc4x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point4);
129 __m256i vacc4x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point4);
130 __m256i vacc4x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point4);
131 __m256i vacc4x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point4);
132 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point);
133 __m256i vacc5x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point5);
134 __m256i vacc5x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point5);
135 __m256i vacc5x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point5);
136 __m256i vacc5x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point5);
137 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point);
138 __m256i vacc6x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point6);
139 __m256i vacc6x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point6);
140 __m256i vacc6x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point6);
141 __m256i vacc6x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point6);
142 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point);
143 __m256i vacc7x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point7);
144 __m256i vacc7x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point7);
145 __m256i vacc7x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point7);
146 __m256i vacc7x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point7);
147 w = (const int32_t*) w + 8;
148
149 size_t k = kc;
150
151 while (k >= 8 * sizeof(int8_t)) {
152 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
153 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
154 a0 += 8;
155 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
156 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
157 a1 += 8;
158 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
159 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
160 a2 += 8;
161 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
162 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
163 a3 += 8;
164 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
165 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
166 a4 += 8;
167 const __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
168 const __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
169 a5 += 8;
170 const __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
171 const __m256i vxa6 = _mm256_cvtepi8_epi16(va6);
172 a6 += 8;
173 const __m128i va7 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a7));
174 const __m256i vxa7 = _mm256_cvtepi8_epi16(va7);
175 a7 += 8;
176
177 const __m256i vxb01 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) w));
178
179 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
180 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
181 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
182 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
183 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
184 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
185 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
186 vacc7x01 = _mm256_add_epi32(vacc7x01, _mm256_madd_epi16(vxa7, vxb01));
187 const __m256i vxb23 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 16)));
188
189 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
190 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
191 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
192 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
193 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
194 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
195 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
196 vacc7x23 = _mm256_add_epi32(vacc7x23, _mm256_madd_epi16(vxa7, vxb23));
197 const __m256i vxb45 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 32)));
198
199 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
200 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
201 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
202 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
203 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
204 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
205 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
206 vacc7x45 = _mm256_add_epi32(vacc7x45, _mm256_madd_epi16(vxa7, vxb45));
207 const __m256i vxb67 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 48)));
208
209 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
210 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
211 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
212 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
213 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
214 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
215 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
216 vacc7x67 = _mm256_add_epi32(vacc7x67, _mm256_madd_epi16(vxa7, vxb67));
217
218 w = (const int8_t*) w + 64;
219 k -= 8 * sizeof(int8_t);
220 }
221
222 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
223 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
224 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
225 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
226 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
227 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
228 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
229 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
230 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
231 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
232 const __m256i vacc5x0213 = _mm256_hadd_epi32(vacc5x01, vacc5x23);
233 const __m256i vacc5x4657 = _mm256_hadd_epi32(vacc5x45, vacc5x67);
234 const __m256i vacc6x0213 = _mm256_hadd_epi32(vacc6x01, vacc6x23);
235 const __m256i vacc6x4657 = _mm256_hadd_epi32(vacc6x45, vacc6x67);
236 const __m256i vacc7x0213 = _mm256_hadd_epi32(vacc7x01, vacc7x23);
237 const __m256i vacc7x4657 = _mm256_hadd_epi32(vacc7x45, vacc7x67);
238
239 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
240 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
241 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
242 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
243 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
244 const __m256i vacc5x02461357 = _mm256_hadd_epi32(vacc5x0213, vacc5x4657);
245 const __m256i vacc6x02461357 = _mm256_hadd_epi32(vacc6x0213, vacc6x4657);
246 const __m256i vacc7x02461357 = _mm256_hadd_epi32(vacc7x0213, vacc7x4657);
247
248 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
249 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
250 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
251 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
252 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
253 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
254 __m256i vacc5x01234567 = _mm256_permutevar8x32_epi32(vacc5x02461357, vpermute_mask);
255 __m256i vacc6x01234567 = _mm256_permutevar8x32_epi32(vacc6x02461357, vpermute_mask);
256 __m256i vacc7x01234567 = _mm256_permutevar8x32_epi32(vacc7x02461357, vpermute_mask);
257
258 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
259 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
260 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
261 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
262 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
263 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
264 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
265 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
266 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
267 const __m256 vinput_scale4 = _mm256_broadcast_ss(&quantization_params[4].inv_scale);
268 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
269 const __m256 vinput_scale5 = _mm256_broadcast_ss(&quantization_params[5].inv_scale);
270 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
271 const __m256 vinput_scale6 = _mm256_broadcast_ss(&quantization_params[6].inv_scale);
272 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
273 const __m256 vinput_scale7 = _mm256_broadcast_ss(&quantization_params[7].inv_scale);
274
275 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
276 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
277 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
278 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
279 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale4);
280 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_scale5);
281 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_scale6);
282 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_scale7);
283
284 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
285 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
286 w = (const float*) w + 16;
287 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
288 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
289 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
290 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
291 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
292 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
293 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
294 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
295
296 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
297 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
298 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
299 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
300 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
301 vout5x01234567 = _mm256_max_ps(vout5x01234567, vmin);
302 vout6x01234567 = _mm256_max_ps(vout6x01234567, vmin);
303 vout7x01234567 = _mm256_max_ps(vout7x01234567, vmin);
304
305 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
306 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
307 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
308 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
309 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
310 vout5x01234567 = _mm256_min_ps(vout5x01234567, vmax);
311 vout6x01234567 = _mm256_min_ps(vout6x01234567, vmax);
312 vout7x01234567 = _mm256_min_ps(vout7x01234567, vmax);
313
314 if XNN_LIKELY(nc >= 8) {
315 _mm256_storeu_ps(c0, vout0x01234567);
316 c0 = (float*) ((uintptr_t) c0 + cn_stride);
317 _mm256_storeu_ps(c1, vout1x01234567);
318 c1 = (float*) ((uintptr_t) c1 + cn_stride);
319 _mm256_storeu_ps(c2, vout2x01234567);
320 c2 = (float*) ((uintptr_t) c2 + cn_stride);
321 _mm256_storeu_ps(c3, vout3x01234567);
322 c3 = (float*) ((uintptr_t) c3 + cn_stride);
323 _mm256_storeu_ps(c4, vout4x01234567);
324 c4 = (float*) ((uintptr_t) c4 + cn_stride);
325 _mm256_storeu_ps(c5, vout5x01234567);
326 c5 = (float*) ((uintptr_t) c5 + cn_stride);
327 _mm256_storeu_ps(c6, vout6x01234567);
328 c6 = (float*) ((uintptr_t) c6 + cn_stride);
329 _mm256_storeu_ps(c7, vout7x01234567);
330 c7 = (float*) ((uintptr_t) c7 + cn_stride);
331
332 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
333 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
334 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
335 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
336 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
337 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
338 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
339 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
340
341 nc -= 8;
342 } else {
343 // Prepare mask for valid 32-bit elements (depends on nc).
344 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
345 _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
346 _mm256_mask_storeu_ps(c1, vmask, vout1x01234567);
347 _mm256_mask_storeu_ps(c2, vmask, vout2x01234567);
348 _mm256_mask_storeu_ps(c3, vmask, vout3x01234567);
349 _mm256_mask_storeu_ps(c4, vmask, vout4x01234567);
350 _mm256_mask_storeu_ps(c5, vmask, vout5x01234567);
351 _mm256_mask_storeu_ps(c6, vmask, vout6x01234567);
352 _mm256_mask_storeu_ps(c7, vmask, vout7x01234567);
353 nc = 0;
354 }
355 } while (nc != 0);
356 }