Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-10x8c8-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_10x8c8__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 <= 10);
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 int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
80 if XNN_UNPREDICTABLE(mr != 10) {
81 c9 = c8;
82 }
83
84 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
85 XNN_FORCE_REALIZATION(vsign_mask);
86 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);
87 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
88 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
89 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
90 // XNN_FORCE_REALIZATION(voutput_zero_point);
91 // XNN_FORCE_REALIZATION(voutput_min);
92 do {
93 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
94 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
95 __m256i vacc1x0123 = vacc0x0123;
96 __m256i vacc1x4567 = vacc0x4567;
97 __m256i vacc2x0123 = vacc0x0123;
98 __m256i vacc2x4567 = vacc0x4567;
99 __m256i vacc3x0123 = vacc0x0123;
100 __m256i vacc3x4567 = vacc0x4567;
101 __m256i vacc4x0123 = vacc0x0123;
102 __m256i vacc4x4567 = vacc0x4567;
103 __m256i vacc5x0123 = vacc0x0123;
104 __m256i vacc5x4567 = vacc0x4567;
105 __m256i vacc6x0123 = vacc0x0123;
106 __m256i vacc6x4567 = vacc0x4567;
107 __m256i vacc7x0123 = vacc0x0123;
108 __m256i vacc7x4567 = vacc0x4567;
109 __m256i vacc8x0123 = vacc0x0123;
110 __m256i vacc8x4567 = vacc0x4567;
111 __m256i vacc9x0123 = vacc0x0123;
112 __m256i vacc9x4567 = vacc0x4567;
113 w = (const int32_t*) w + 8;
114
115 size_t p = ks;
116 do {
117 const int8_t* restrict a0 = a[0];
118 if XNN_UNPREDICTABLE(a0 != zero) {
119 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
120 }
121 const int8_t* restrict a1 = a[1];
122 if XNN_UNPREDICTABLE(a1 != zero) {
123 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
124 }
125 const int8_t* restrict a2 = a[2];
126 if XNN_UNPREDICTABLE(a2 != zero) {
127 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
128 }
129 const int8_t* restrict a3 = a[3];
130 if XNN_UNPREDICTABLE(a3 != zero) {
131 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
132 }
133 const int8_t* restrict a4 = a[4];
134 if XNN_UNPREDICTABLE(a4 != zero) {
135 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
136 }
137 const int8_t* restrict a5 = a[5];
138 if XNN_UNPREDICTABLE(a5 != zero) {
139 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
140 }
141 const int8_t* restrict a6 = a[6];
142 if XNN_UNPREDICTABLE(a6 != zero) {
143 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
144 }
145 const int8_t* restrict a7 = a[7];
146 if XNN_UNPREDICTABLE(a7 != zero) {
147 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
148 }
149 const int8_t* restrict a8 = a[8];
150 if XNN_UNPREDICTABLE(a8 != zero) {
151 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
152 }
153 const int8_t* restrict a9 = a[9];
154 if XNN_UNPREDICTABLE(a9 != zero) {
155 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
156 }
157 a += 10;
158
159 size_t k = kc;
160 while (k >= 16 * sizeof(int8_t)) {
161 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
162 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
163 a0 += 16;
164 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
165 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
166 a1 += 16;
167 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
168 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
169 a2 += 16;
170 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
171 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
172 a3 += 16;
173 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
174 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
175 a4 += 16;
176 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
177 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
178 a5 += 16;
179 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
180 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
181 a6 += 16;
182 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
183 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
184 a7 += 16;
185 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
186 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
187 a8 += 16;
188 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
189 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
190 a9 += 16;
191
192 const __m256i vb01234567x0123 = _mm256_load_si256(w);
193 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
194 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
195 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
196 xnn_prefetch_to_l1((const int8_t*) w + 896);
197
198 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
199 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
200 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
201 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
202 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
203 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
204 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
205 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
206 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
207 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
208 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
209 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
210 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
211 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
212 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
213 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
214 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
215 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
216 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
217 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
218 xnn_prefetch_to_l1((const int8_t*) w + 960);
219 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
220 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
221 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
222 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
223 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
224 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
225 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
226 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
227 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
228 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
229 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
230 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
231 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
232 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
233 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
234 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
235 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
236 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
237 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
238 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
239
240 w = (const int8_t*) w + 128;
241 k -= 16 * sizeof(int8_t);
242 }
243
244 if (k != 0) {
245 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
246 a0 += 8;
247 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
248 a1 += 8;
249 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
250 a2 += 8;
251 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
252 a3 += 8;
253 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
254 a4 += 8;
255 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
256 a5 += 8;
257 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
258 a6 += 8;
259 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
260 a7 += 8;
261 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
262 a8 += 8;
263 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
264 a9 += 8;
265
266 const __m256i vb01234567x0123 = _mm256_load_si256(w);
267 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
268
269 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
270 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
271 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
272 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
273 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
274 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
275 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
276 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
277 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
278 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
279 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
280 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
281 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
282 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
283 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
284 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
285 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
286 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
287 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
288 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
289 xnn_prefetch_to_l1((const int8_t*) w + 960);
290
291 w = (const int8_t*) w + 64;
292 k -= 8 * sizeof(int8_t);
293 }
294
295 p -= 10 * sizeof(void*);
296 } while (p != 0);
297
298
299 // Add adjacent pairs
300 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
301 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
302 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
303 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
304 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
305 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
306 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
307 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
308 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
309 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
310 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
311 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
312 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
313 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
314 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
315 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
316 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
317 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
318 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
319 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
320
321 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
322 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
323 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
324 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
325 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
326 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
327 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
328 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
329 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
330 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
331
332 const __m256 vscale01234567 = _mm256_load_ps(w);
333 w = (const float*) w + 8;
334 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
335 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
336 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
337 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
338 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
339 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
340 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
341 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vscale01234567);
342 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vscale01234567);
343 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vscale01234567);
344
345 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
346 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
347 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
348 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
349 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
350 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
351 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
352 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max_less_zero_point);
353 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max_less_zero_point);
354 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max_less_zero_point);
355
356 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
357 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
358 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
359 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
360 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
361 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
362 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
363 vacc7x01234567 = _mm256_cvtps_epi32(vout7x01234567);
364 vacc8x01234567 = _mm256_cvtps_epi32(vout8x01234567);
365 vacc9x01234567 = _mm256_cvtps_epi32(vout9x01234567);
366
367 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
368 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
369 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
370 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
371 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
372 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
373 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
374 vacc7x01234567 = _mm256_add_epi32(vacc7x01234567, voutput_zero_point);
375 vacc8x01234567 = _mm256_add_epi32(vacc8x01234567, voutput_zero_point);
376 vacc9x01234567 = _mm256_add_epi32(vacc9x01234567, voutput_zero_point);
377
378 __m128i voutb0x01234567 = _mm256_cvtsepi32_epi8(vacc0x01234567);
379 __m128i voutb1x01234567 = _mm256_cvtsepi32_epi8(vacc1x01234567);
380 __m128i voutb2x01234567 = _mm256_cvtsepi32_epi8(vacc2x01234567);
381 __m128i voutb3x01234567 = _mm256_cvtsepi32_epi8(vacc3x01234567);
382 __m128i voutb4x01234567 = _mm256_cvtsepi32_epi8(vacc4x01234567);
383 __m128i voutb5x01234567 = _mm256_cvtsepi32_epi8(vacc5x01234567);
384 __m128i voutb6x01234567 = _mm256_cvtsepi32_epi8(vacc6x01234567);
385 __m128i voutb7x01234567 = _mm256_cvtsepi32_epi8(vacc7x01234567);
386 __m128i voutb8x01234567 = _mm256_cvtsepi32_epi8(vacc8x01234567);
387 __m128i voutb9x01234567 = _mm256_cvtsepi32_epi8(vacc9x01234567);
388
389 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
390 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
391 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
392 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
393 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
394 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
395 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
396 voutb7x01234567 = _mm_max_epi8(voutb7x01234567, voutput_min);
397 voutb8x01234567 = _mm_max_epi8(voutb8x01234567, voutput_min);
398 voutb9x01234567 = _mm_max_epi8(voutb9x01234567, voutput_min);
399
400 if (nc >= 8) {
401 _mm_storel_epi64((__m128i*) c9, voutb9x01234567);
402 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
403 _mm_storel_epi64((__m128i*) c8, voutb8x01234567);
404 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
405 _mm_storel_epi64((__m128i*) c7, voutb7x01234567);
406 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
407 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
408 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
409 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
410 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
411 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
412 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
413 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
414 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
415 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
416 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
417 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
418 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
419 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
420 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
421 a = (const int8_t**restrict) ((uintptr_t) a - ks);
422 nc -= 8;
423 } else {
424 // Prepare mask for valid 8-bit elements (depends on nc).
425 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
426 _mm_mask_storeu_epi8(c9, vmask, voutb9x01234567);
427 _mm_mask_storeu_epi8(c8, vmask, voutb8x01234567);
428 _mm_mask_storeu_epi8(c7, vmask, voutb7x01234567);
429 _mm_mask_storeu_epi8(c6, vmask, voutb6x01234567);
430 _mm_mask_storeu_epi8(c5, vmask, voutb5x01234567);
431 _mm_mask_storeu_epi8(c4, vmask, voutb4x01234567);
432 _mm_mask_storeu_epi8(c3, vmask, voutb3x01234567);
433 _mm_mask_storeu_epi8(c2, vmask, voutb2x01234567);
434 _mm_mask_storeu_epi8(c1, vmask, voutb1x01234567);
435 _mm_mask_storeu_epi8(c0, vmask, voutb0x01234567);
436 nc = 0;
437 }
438 } while (nc != 0);
439 }