Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-8x8c8-minmax-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_qd8_f16_qc8w_igemm_minmax_ukernel_8x8c8__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 xnn_float16* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 8);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 8 * sizeof(int8_t));
48 uint16_t* c0 = (uint16_t*) c;
49 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 c1 = c0;
52 }
53 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 c2 = c1;
56 }
57 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 c3 = c2;
60 }
61 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 c4 = c3;
64 }
65 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr < 6) {
67 c5 = c4;
68 }
69 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
70 if XNN_UNPREDICTABLE(mr <= 6) {
71 c6 = c5;
72 }
73 uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride);
74 if XNN_UNPREDICTABLE(mr != 8) {
75 c7 = c6;
76 }
77
78 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
79 XNN_FORCE_REALIZATION(vsign_mask);
80 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
81 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
82 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
83 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
84 // XNN_FORCE_REALIZATION(vinput_zero_point);
85 // XNN_FORCE_REALIZATION(vinput_inv_scale);
86 // XNN_FORCE_REALIZATION(voutput_min);
87 // XNN_FORCE_REALIZATION(voutput_max);
88 do {
89 const __m256i vksum01234567 = _mm256_load_si256(w);
90 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
91 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
92 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
93 __m256i vacc1x0123 = vacc0x0123;
94 __m256i vacc1x4567 = vacc0x4567;
95 __m256i vacc2x0123 = vacc0x0123;
96 __m256i vacc2x4567 = vacc0x4567;
97 __m256i vacc3x0123 = vacc0x0123;
98 __m256i vacc3x4567 = vacc0x4567;
99 __m256i vacc4x0123 = vacc0x0123;
100 __m256i vacc4x4567 = vacc0x4567;
101 __m256i vacc5x0123 = vacc0x0123;
102 __m256i vacc5x4567 = vacc0x4567;
103 __m256i vacc6x0123 = vacc0x0123;
104 __m256i vacc6x4567 = vacc0x4567;
105 __m256i vacc7x0123 = vacc0x0123;
106 __m256i vacc7x4567 = 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 } else {
115 a0 = zero_data;
116 }
117 const int8_t* restrict a1 = a[1];
118 if XNN_UNPREDICTABLE(a1 != zero) {
119 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
120 } else {
121 a1 = zero_data;
122 }
123 const int8_t* restrict a2 = a[2];
124 if XNN_UNPREDICTABLE(a2 != zero) {
125 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
126 } else {
127 a2 = zero_data;
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 } else {
133 a3 = zero_data;
134 }
135 const int8_t* restrict a4 = a[4];
136 if XNN_UNPREDICTABLE(a4 != zero) {
137 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
138 } else {
139 a4 = zero_data;
140 }
141 const int8_t* restrict a5 = a[5];
142 if XNN_UNPREDICTABLE(a5 != zero) {
143 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
144 } else {
145 a5 = zero_data;
146 }
147 const int8_t* restrict a6 = a[6];
148 if XNN_UNPREDICTABLE(a6 != zero) {
149 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
150 } else {
151 a6 = zero_data;
152 }
153 const int8_t* restrict a7 = a[7];
154 if XNN_UNPREDICTABLE(a7 != zero) {
155 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
156 } else {
157 a7 = zero_data;
158 }
159 a += 8;
160
161 size_t k = kc;
162 while (k >= 16 * sizeof(int8_t)) {
163 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
164 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
165 a0 += 16;
166 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
167 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
168 a1 += 16;
169 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
170 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
171 a2 += 16;
172 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
173 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
174 a3 += 16;
175 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
176 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
177 a4 += 16;
178 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
179 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
180 a5 += 16;
181 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
182 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
183 a6 += 16;
184 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
185 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
186 a7 += 16;
187
188 const __m256i vb01234567x0123 = _mm256_load_si256(w);
189 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
190 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
191 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
192 xnn_prefetch_to_l1((const int8_t*) w + 896);
193
194 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
195 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
196 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
197 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
198 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
199 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
200 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
201 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
202 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
203 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
204 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
205 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
206 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
207 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
208 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
209 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
210 xnn_prefetch_to_l1((const int8_t*) w + 960);
211 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
212 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
213 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
214 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
215 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
216 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
217 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
218 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
219 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
220 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
221 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
222 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
223 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
224 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
225 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
226 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
227
228 w = (const int8_t*) w + 128;
229 k -= 16 * sizeof(int8_t);
230 }
231
232 if (k != 0) {
233 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
234 a0 += 8;
235 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
236 a1 += 8;
237 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
238 a2 += 8;
239 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
240 a3 += 8;
241 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
242 a4 += 8;
243 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
244 a5 += 8;
245 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
246 a6 += 8;
247 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
248 a7 += 8;
249
250 const __m256i vb01234567x0123 = _mm256_load_si256(w);
251 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
252
253 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
254 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
255 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
256 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
257 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
258 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
259 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
260 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
261 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
262 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
263 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
264 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
265 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
266 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
267 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
268 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
269 xnn_prefetch_to_l1((const int8_t*) w + 960);
270
271 w = (const int8_t*) w + 64;
272 k -= 8 * sizeof(int8_t);
273 }
274
275 p -= 8 * sizeof(void*);
276 } while (p != 0);
277
278
279 // Add adjacent pairs
280 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
281 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
282 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
283 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
284 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
285 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
286 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
287 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
288 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
289 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
290 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
291 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
292 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
293 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
294 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
295 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
296
297 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
298 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
299 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
300 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
301 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
302 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
303 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
304 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
305
306 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
307 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
308 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
309 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
310 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
311 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
312 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
313 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_inv_scale);
314
315 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
316 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
317 w = (const float*) w + 16;
318 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
319 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
320 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
321 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
322 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
323 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
324 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
325 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
326
327 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
328 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
329 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
330 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
331 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
332 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
333 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
334 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
335
336 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
337 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
338 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
339 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
340 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
341 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
342 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
343 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
344
345 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
346 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
347 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
348 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
349 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
350 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
351 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
352 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
353 if XNN_LIKELY(nc >= 8) {
354 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
355 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
356 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
357 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
358 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
359 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
360 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
361 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
362 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
363 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
364 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
365 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
366 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
367 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
368 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
369 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
370
371 a = (const int8_t**restrict) ((uintptr_t) a - ks);
372 nc -= 8;
373 } else {
374 // Prepare mask for valid 16-bit elements (depends on nc).
375 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
376 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
377 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
378 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
379 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
380 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
381 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
382 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
383 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
384 nc = 0;
385 }
386 } while (nc != 0);
387 }