Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-9x8c8-minmax-fp32-avx256vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/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_qs8_qc8w_igemm_minmax_fp32_ukernel_9x8c8__avx256vnni(
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 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 9);
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 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63 if XNN_UNPREDICTABLE(mr < 6) {
64 c5 = c4;
65 }
66 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 6) {
68 c6 = c5;
69 }
70 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
71 if XNN_UNPREDICTABLE(mr < 8) {
72 c7 = c6;
73 }
74 int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
75 if XNN_UNPREDICTABLE(mr <= 8) {
76 c8 = c7;
77 }
78
79 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
80 XNN_FORCE_REALIZATION(vsign_mask);
81 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
82 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
83 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
84 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
85 // XNN_FORCE_REALIZATION(voutput_zero_point);
86 // XNN_FORCE_REALIZATION(voutput_min);
87 do {
88 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
89 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
90 __m256i vacc1x0123 = vacc0x0123;
91 __m256i vacc1x4567 = vacc0x4567;
92 __m256i vacc2x0123 = vacc0x0123;
93 __m256i vacc2x4567 = vacc0x4567;
94 __m256i vacc3x0123 = vacc0x0123;
95 __m256i vacc3x4567 = vacc0x4567;
96 __m256i vacc4x0123 = vacc0x0123;
97 __m256i vacc4x4567 = vacc0x4567;
98 __m256i vacc5x0123 = vacc0x0123;
99 __m256i vacc5x4567 = vacc0x4567;
100 __m256i vacc6x0123 = vacc0x0123;
101 __m256i vacc6x4567 = vacc0x4567;
102 __m256i vacc7x0123 = vacc0x0123;
103 __m256i vacc7x4567 = vacc0x4567;
104 __m256i vacc8x0123 = vacc0x0123;
105 __m256i vacc8x4567 = vacc0x4567;
106 w = (const int32_t*) w + 8;
107
108 size_t p = ks;
109 do {
110 const int8_t* restrict a0 = a[0];
111 if XNN_UNPREDICTABLE(a0 != zero) {
112 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
113 }
114 const int8_t* restrict a1 = a[1];
115 if XNN_UNPREDICTABLE(a1 != zero) {
116 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
117 }
118 const int8_t* restrict a2 = a[2];
119 if XNN_UNPREDICTABLE(a2 != zero) {
120 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
121 }
122 const int8_t* restrict a3 = a[3];
123 if XNN_UNPREDICTABLE(a3 != zero) {
124 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
125 }
126 const int8_t* restrict a4 = a[4];
127 if XNN_UNPREDICTABLE(a4 != zero) {
128 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
129 }
130 const int8_t* restrict a5 = a[5];
131 if XNN_UNPREDICTABLE(a5 != zero) {
132 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
133 }
134 const int8_t* restrict a6 = a[6];
135 if XNN_UNPREDICTABLE(a6 != zero) {
136 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
137 }
138 const int8_t* restrict a7 = a[7];
139 if XNN_UNPREDICTABLE(a7 != zero) {
140 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
141 }
142 const int8_t* restrict a8 = a[8];
143 if XNN_UNPREDICTABLE(a8 != zero) {
144 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
145 }
146 a += 9;
147
148 size_t k = kc;
149 while (k >= 16 * sizeof(int8_t)) {
150 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
151 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
152 a0 += 16;
153 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
154 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
155 a1 += 16;
156 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
157 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
158 a2 += 16;
159 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
160 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
161 a3 += 16;
162 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
163 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
164 a4 += 16;
165 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
166 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
167 a5 += 16;
168 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
169 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
170 a6 += 16;
171 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
172 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
173 a7 += 16;
174 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
175 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
176 a8 += 16;
177
178 const __m256i vb01234567x0123 = _mm256_load_si256(w);
179 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
180 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
181 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
182
183 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
184 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
185 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
186 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
187 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
188 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
189 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
190 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
191 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
192 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
193 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
194 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
195 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
196 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
197 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
198 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
199 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
200 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
201 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
202 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
203 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
204 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
205 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
206 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
207 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
208 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
209 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
210 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
211 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
212 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
213 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
214 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
215 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
216 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
217 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
218 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
219
220 w = (const int8_t*) w + 128;
221 k -= 16 * sizeof(int8_t);
222 }
223
224 if (k != 0) {
225 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
226 a0 += 8;
227 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
228 a1 += 8;
229 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
230 a2 += 8;
231 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
232 a3 += 8;
233 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
234 a4 += 8;
235 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
236 a5 += 8;
237 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
238 a6 += 8;
239 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
240 a7 += 8;
241 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
242 a8 += 8;
243
244 const __m256i vb01234567x0123 = _mm256_load_si256(w);
245 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
246
247 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
248 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
249 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
250 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
251 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
252 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
253 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
254 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
255 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
256 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
257 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
258 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
259 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
260 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
261 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
262 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
263 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
264 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
265
266 w = (const int8_t*) w + 64;
267 k -= 8 * sizeof(int8_t);
268 }
269
270 p -= 9 * sizeof(void*);
271 } while (p != 0);
272
273
274 // Add adjacent pairs
275 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
276 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
277 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
278 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
279 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
280 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
281 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
282 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
283 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
284 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
285 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
286 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
287 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
288 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
289 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
290 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
291 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
292 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
293
294 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
295 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
296 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
297 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
298 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
299 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
300 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
301 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
302 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
303
304 const __m256 vscale01234567 = _mm256_load_ps(w);
305 w = (const float*) w + 8;
306 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
307 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
308 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
309 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
310 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
311 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
312 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
313 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vscale01234567);
314 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vscale01234567);
315
316 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
317 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
318 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
319 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
320 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
321 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
322 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
323 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max_less_zero_point);
324 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max_less_zero_point);
325
326 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
327 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
328 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
329 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
330 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
331 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
332 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
333 vacc7x01234567 = _mm256_cvtps_epi32(vout7x01234567);
334 vacc8x01234567 = _mm256_cvtps_epi32(vout8x01234567);
335
336 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
337 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
338 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
339 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
340 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
341 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
342 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
343 vacc7x01234567 = _mm256_add_epi32(vacc7x01234567, voutput_zero_point);
344 vacc8x01234567 = _mm256_add_epi32(vacc8x01234567, voutput_zero_point);
345
346 __m128i voutb0x01234567 = _mm256_cvtsepi32_epi8(vacc0x01234567);
347 __m128i voutb1x01234567 = _mm256_cvtsepi32_epi8(vacc1x01234567);
348 __m128i voutb2x01234567 = _mm256_cvtsepi32_epi8(vacc2x01234567);
349 __m128i voutb3x01234567 = _mm256_cvtsepi32_epi8(vacc3x01234567);
350 __m128i voutb4x01234567 = _mm256_cvtsepi32_epi8(vacc4x01234567);
351 __m128i voutb5x01234567 = _mm256_cvtsepi32_epi8(vacc5x01234567);
352 __m128i voutb6x01234567 = _mm256_cvtsepi32_epi8(vacc6x01234567);
353 __m128i voutb7x01234567 = _mm256_cvtsepi32_epi8(vacc7x01234567);
354 __m128i voutb8x01234567 = _mm256_cvtsepi32_epi8(vacc8x01234567);
355
356 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
357 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
358 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
359 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
360 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
361 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
362 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
363 voutb7x01234567 = _mm_max_epi8(voutb7x01234567, voutput_min);
364 voutb8x01234567 = _mm_max_epi8(voutb8x01234567, voutput_min);
365
366 if (nc >= 8) {
367 _mm_storel_epi64((__m128i*) c8, voutb8x01234567);
368 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
369 _mm_storel_epi64((__m128i*) c7, voutb7x01234567);
370 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
371 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
372 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
373 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
374 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
375 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
376 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
377 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
378 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
379 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
380 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
381 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
382 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
383 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
384 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
385 a = (const int8_t**restrict) ((uintptr_t) a - ks);
386 nc -= 8;
387 } else {
388 // Prepare mask for valid 8-bit elements (depends on nc).
389 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
390 _mm_mask_storeu_epi8(c8, vmask, voutb8x01234567);
391 _mm_mask_storeu_epi8(c7, vmask, voutb7x01234567);
392 _mm_mask_storeu_epi8(c6, vmask, voutb6x01234567);
393 _mm_mask_storeu_epi8(c5, vmask, voutb5x01234567);
394 _mm_mask_storeu_epi8(c4, vmask, voutb4x01234567);
395 _mm_mask_storeu_epi8(c3, vmask, voutb3x01234567);
396 _mm_mask_storeu_epi8(c2, vmask, voutb2x01234567);
397 _mm_mask_storeu_epi8(c1, vmask, voutb1x01234567);
398 _mm_mask_storeu_epi8(c0, vmask, voutb0x01234567);
399 nc = 0;
400 }
401 } while (nc != 0);
402 }