Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-9x8c8-minmax-fp32-avx256vnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_9x8c8__avx256vnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 int8_t* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 9);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr < 6) {
65 c5 = c4;
66 }
67 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 6) {
69 c6 = c5;
70 }
71 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
72 if XNN_UNPREDICTABLE(mr < 8) {
73 c7 = c6;
74 }
75 int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
76 if XNN_UNPREDICTABLE(mr <= 8) {
77 c8 = c7;
78 }
79
80 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
81 XNN_FORCE_REALIZATION(vsign_mask);
82 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);
83 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
84 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
85 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
86 // XNN_FORCE_REALIZATION(voutput_zero_point);
87 // XNN_FORCE_REALIZATION(voutput_min);
88 do {
89 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
90 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
91 __m256i vacc1x0123 = vacc0x0123;
92 __m256i vacc1x4567 = vacc0x4567;
93 __m256i vacc2x0123 = vacc0x0123;
94 __m256i vacc2x4567 = vacc0x4567;
95 __m256i vacc3x0123 = vacc0x0123;
96 __m256i vacc3x4567 = vacc0x4567;
97 __m256i vacc4x0123 = vacc0x0123;
98 __m256i vacc4x4567 = vacc0x4567;
99 __m256i vacc5x0123 = vacc0x0123;
100 __m256i vacc5x4567 = vacc0x4567;
101 __m256i vacc6x0123 = vacc0x0123;
102 __m256i vacc6x4567 = vacc0x4567;
103 __m256i vacc7x0123 = vacc0x0123;
104 __m256i vacc7x4567 = vacc0x4567;
105 __m256i vacc8x0123 = vacc0x0123;
106 __m256i vacc8x4567 = vacc0x4567;
107 w = (const int32_t*) w + 8;
108
109 size_t p = ks;
110 do {
111 const int8_t* restrict a0 = a[0];
112 if XNN_UNPREDICTABLE(a0 != zero) {
113 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
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 }
119 const int8_t* restrict a2 = a[2];
120 if XNN_UNPREDICTABLE(a2 != zero) {
121 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
122 }
123 const int8_t* restrict a3 = a[3];
124 if XNN_UNPREDICTABLE(a3 != zero) {
125 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
126 }
127 const int8_t* restrict a4 = a[4];
128 if XNN_UNPREDICTABLE(a4 != zero) {
129 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
130 }
131 const int8_t* restrict a5 = a[5];
132 if XNN_UNPREDICTABLE(a5 != zero) {
133 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
134 }
135 const int8_t* restrict a6 = a[6];
136 if XNN_UNPREDICTABLE(a6 != zero) {
137 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
138 }
139 const int8_t* restrict a7 = a[7];
140 if XNN_UNPREDICTABLE(a7 != zero) {
141 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
142 }
143 const int8_t* restrict a8 = a[8];
144 if XNN_UNPREDICTABLE(a8 != zero) {
145 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
146 }
147 a += 9;
148
149 size_t k = kc;
150 while (k >= 16 * sizeof(int8_t)) {
151 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
152 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
153 a0 += 16;
154 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
155 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
156 a1 += 16;
157 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
158 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
159 a2 += 16;
160 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
161 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
162 a3 += 16;
163 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
164 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
165 a4 += 16;
166 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
167 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
168 a5 += 16;
169 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
170 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
171 a6 += 16;
172 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
173 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
174 a7 += 16;
175 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
176 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
177 a8 += 16;
178
179 const __m256i vb01234567x0123 = _mm256_load_si256(w);
180 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
181 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
182 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
183 xnn_prefetch_to_l1((const int8_t*) w + 896);
184
185 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
186 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
187 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
188 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
189 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
190 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
191 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
192 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
193 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
194 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
195 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
196 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
197 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
198 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
199 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
200 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
201 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
202 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
203 xnn_prefetch_to_l1((const int8_t*) w + 960);
204 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
205 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
206 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
207 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
208 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
209 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
210 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
211 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
212 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
213 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
214 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
215 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
216 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
217 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
218 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
219 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
220 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
221 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
222
223 w = (const int8_t*) w + 128;
224 k -= 16 * sizeof(int8_t);
225 }
226
227 if (k != 0) {
228 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
229 a0 += 8;
230 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
231 a1 += 8;
232 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
233 a2 += 8;
234 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
235 a3 += 8;
236 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
237 a4 += 8;
238 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
239 a5 += 8;
240 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
241 a6 += 8;
242 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
243 a7 += 8;
244 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
245 a8 += 8;
246
247 const __m256i vb01234567x0123 = _mm256_load_si256(w);
248 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
249
250 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
251 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
252 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
253 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
254 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
255 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
256 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
257 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
258 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
259 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
260 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
261 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
262 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
263 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
264 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
265 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
266 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
267 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
268 xnn_prefetch_to_l1((const int8_t*) w + 960);
269
270 w = (const int8_t*) w + 64;
271 k -= 8 * sizeof(int8_t);
272 }
273
274 p -= 9 * sizeof(void*);
275 } while (p != 0);
276
277
278 // Add adjacent pairs
279 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
280 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
281 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
282 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
283 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
284 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
285 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
286 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
287 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
288 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
289 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
290 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
291 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
292 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
293 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
294 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
295 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
296 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
297
298 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
299 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
300 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
301 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
302 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
303 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
304 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
305 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
306 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
307
308 const __m256 vscale01234567 = _mm256_load_ps(w);
309 w = (const float*) w + 8;
310 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
311 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
312 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
313 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
314 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
315 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
316 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
317 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vscale01234567);
318 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vscale01234567);
319
320 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
321 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
322 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
323 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
324 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
325 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
326 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
327 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max_less_zero_point);
328 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max_less_zero_point);
329
330 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
331 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
332 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
333 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
334 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
335 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
336 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
337 vacc7x01234567 = _mm256_cvtps_epi32(vout7x01234567);
338 vacc8x01234567 = _mm256_cvtps_epi32(vout8x01234567);
339
340 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
341 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
342 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
343 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
344 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
345 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
346 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
347 vacc7x01234567 = _mm256_add_epi32(vacc7x01234567, voutput_zero_point);
348 vacc8x01234567 = _mm256_add_epi32(vacc8x01234567, voutput_zero_point);
349
350 __m128i voutb0x01234567 = _mm256_cvtsepi32_epi8(vacc0x01234567);
351 __m128i voutb1x01234567 = _mm256_cvtsepi32_epi8(vacc1x01234567);
352 __m128i voutb2x01234567 = _mm256_cvtsepi32_epi8(vacc2x01234567);
353 __m128i voutb3x01234567 = _mm256_cvtsepi32_epi8(vacc3x01234567);
354 __m128i voutb4x01234567 = _mm256_cvtsepi32_epi8(vacc4x01234567);
355 __m128i voutb5x01234567 = _mm256_cvtsepi32_epi8(vacc5x01234567);
356 __m128i voutb6x01234567 = _mm256_cvtsepi32_epi8(vacc6x01234567);
357 __m128i voutb7x01234567 = _mm256_cvtsepi32_epi8(vacc7x01234567);
358 __m128i voutb8x01234567 = _mm256_cvtsepi32_epi8(vacc8x01234567);
359
360 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
361 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
362 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
363 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
364 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
365 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
366 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
367 voutb7x01234567 = _mm_max_epi8(voutb7x01234567, voutput_min);
368 voutb8x01234567 = _mm_max_epi8(voutb8x01234567, voutput_min);
369
370 if (nc >= 8) {
371 _mm_storel_epi64((__m128i*) c8, voutb8x01234567);
372 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
373 _mm_storel_epi64((__m128i*) c7, voutb7x01234567);
374 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
375 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
376 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
377 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
378 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
379 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
380 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
381 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
382 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
383 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
384 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
385 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
386 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
387 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
388 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
389 a = (const int8_t**restrict) ((uintptr_t) a - ks);
390 nc -= 8;
391 } else {
392 // Prepare mask for valid 8-bit elements (depends on nc).
393 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
394 _mm_mask_storeu_epi8(c8, vmask, voutb8x01234567);
395 _mm_mask_storeu_epi8(c7, vmask, voutb7x01234567);
396 _mm_mask_storeu_epi8(c6, vmask, voutb6x01234567);
397 _mm_mask_storeu_epi8(c5, vmask, voutb5x01234567);
398 _mm_mask_storeu_epi8(c4, vmask, voutb4x01234567);
399 _mm_mask_storeu_epi8(c3, vmask, voutb3x01234567);
400 _mm_mask_storeu_epi8(c2, vmask, voutb2x01234567);
401 _mm_mask_storeu_epi8(c1, vmask, voutb1x01234567);
402 _mm_mask_storeu_epi8(c0, vmask, voutb0x01234567);
403 nc = 0;
404 }
405 } while (nc != 0);
406 }