Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-8x8c8-minmax-avx256skx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_8x8c8__avx256skx(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 8);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (8 * sizeof(void*)) == 0);
42 assert(a_offset % 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 vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point);
79 const __m256 vinput_scale = _mm256_broadcast_ss(&quantization_params->inv_scale);
80 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
81 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
82 // XNN_FORCE_REALIZATION(vinput_zero_point);
83 // XNN_FORCE_REALIZATION(vinput_scale);
84 // XNN_FORCE_REALIZATION(vmin);
85 // XNN_FORCE_REALIZATION(vmax);
86
87 do {
88 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
89 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
90 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
91 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
92 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
93 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
94 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
95 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
96 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
97 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
98 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
99 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
100 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point);
101 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point);
102 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point);
103 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point);
104 __m256i vacc1x01 = vacc0x01;
105 __m256i vacc1x23 = vacc0x23;
106 __m256i vacc1x45 = vacc0x45;
107 __m256i vacc1x67 = vacc0x67;
108 __m256i vacc2x01 = vacc0x01;
109 __m256i vacc2x23 = vacc0x23;
110 __m256i vacc2x45 = vacc0x45;
111 __m256i vacc2x67 = vacc0x67;
112 __m256i vacc3x01 = vacc0x01;
113 __m256i vacc3x23 = vacc0x23;
114 __m256i vacc3x45 = vacc0x45;
115 __m256i vacc3x67 = vacc0x67;
116 __m256i vacc4x01 = vacc0x01;
117 __m256i vacc4x23 = vacc0x23;
118 __m256i vacc4x45 = vacc0x45;
119 __m256i vacc4x67 = vacc0x67;
120 __m256i vacc5x01 = vacc0x01;
121 __m256i vacc5x23 = vacc0x23;
122 __m256i vacc5x45 = vacc0x45;
123 __m256i vacc5x67 = vacc0x67;
124 __m256i vacc6x01 = vacc0x01;
125 __m256i vacc6x23 = vacc0x23;
126 __m256i vacc6x45 = vacc0x45;
127 __m256i vacc6x67 = vacc0x67;
128 __m256i vacc7x01 = vacc0x01;
129 __m256i vacc7x23 = vacc0x23;
130 __m256i vacc7x45 = vacc0x45;
131 __m256i vacc7x67 = vacc0x67;
132 w = (const int32_t*) w + 8;
133
134 size_t p = ks;
135 do {
136 const int8_t* restrict a0 = a[0];
137 if XNN_UNPREDICTABLE(a0 != zero) {
138 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
139 } else {
140 a0 = zero_data;
141 }
142 const int8_t* restrict a1 = a[1];
143 if XNN_UNPREDICTABLE(a1 != zero) {
144 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
145 } else {
146 a1 = zero_data;
147 }
148 const int8_t* restrict a2 = a[2];
149 if XNN_UNPREDICTABLE(a2 != zero) {
150 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
151 } else {
152 a2 = zero_data;
153 }
154 const int8_t* restrict a3 = a[3];
155 if XNN_UNPREDICTABLE(a3 != zero) {
156 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
157 } else {
158 a3 = zero_data;
159 }
160 const int8_t* restrict a4 = a[4];
161 if XNN_UNPREDICTABLE(a4 != zero) {
162 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
163 } else {
164 a4 = zero_data;
165 }
166 const int8_t* restrict a5 = a[5];
167 if XNN_UNPREDICTABLE(a5 != zero) {
168 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
169 } else {
170 a5 = zero_data;
171 }
172 const int8_t* restrict a6 = a[6];
173 if XNN_UNPREDICTABLE(a6 != zero) {
174 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
175 } else {
176 a6 = zero_data;
177 }
178 const int8_t* restrict a7 = a[7];
179 if XNN_UNPREDICTABLE(a7 != zero) {
180 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
181 } else {
182 a7 = zero_data;
183 }
184 a += 8;
185
186 size_t k = 0;
187 while (k < kc) {
188 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
189 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
190 a0 += 8;
191 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
192 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
193 a1 += 8;
194 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
195 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
196 a2 += 8;
197 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
198 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
199 a3 += 8;
200 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
201 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
202 a4 += 8;
203 const __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5));
204 const __m256i vxa5 = _mm256_cvtepi8_epi16(va5);
205 a5 += 8;
206 const __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6));
207 const __m256i vxa6 = _mm256_cvtepi8_epi16(va6);
208 a6 += 8;
209 const __m128i va7 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a7));
210 const __m256i vxa7 = _mm256_cvtepi8_epi16(va7);
211 a7 += 8;
212
213 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
214 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
215
216 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
217 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
218 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
219 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
220 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
221 vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01));
222 vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01));
223 vacc7x01 = _mm256_add_epi32(vacc7x01, _mm256_madd_epi16(vxa7, vxb01));
224 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
225 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
226
227 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
228 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
229 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
230 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
231 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
232 vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23));
233 vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23));
234 vacc7x23 = _mm256_add_epi32(vacc7x23, _mm256_madd_epi16(vxa7, vxb23));
235 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
236 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
237
238 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
239 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
240 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
241 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
242 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
243 vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45));
244 vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45));
245 vacc7x45 = _mm256_add_epi32(vacc7x45, _mm256_madd_epi16(vxa7, vxb45));
246 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
247 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
248
249 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
250 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
251 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
252 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
253 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
254 vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67));
255 vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67));
256 vacc7x67 = _mm256_add_epi32(vacc7x67, _mm256_madd_epi16(vxa7, vxb67));
257
258 w = (const void*) ((const int8_t*) w + 64);
259 k += 8 * sizeof(int8_t);
260 }
261 p -= 8 * sizeof(void*);
262 } while (p != 0);
263
264 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
265 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
266 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
267 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
268 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
269 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
270 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
271 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
272 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
273 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
274 const __m256i vacc5x0213 = _mm256_hadd_epi32(vacc5x01, vacc5x23);
275 const __m256i vacc5x4657 = _mm256_hadd_epi32(vacc5x45, vacc5x67);
276 const __m256i vacc6x0213 = _mm256_hadd_epi32(vacc6x01, vacc6x23);
277 const __m256i vacc6x4657 = _mm256_hadd_epi32(vacc6x45, vacc6x67);
278 const __m256i vacc7x0213 = _mm256_hadd_epi32(vacc7x01, vacc7x23);
279 const __m256i vacc7x4657 = _mm256_hadd_epi32(vacc7x45, vacc7x67);
280
281 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
282 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
283 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
284 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
285 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
286 const __m256i vacc5x02461357 = _mm256_hadd_epi32(vacc5x0213, vacc5x4657);
287 const __m256i vacc6x02461357 = _mm256_hadd_epi32(vacc6x0213, vacc6x4657);
288 const __m256i vacc7x02461357 = _mm256_hadd_epi32(vacc7x0213, vacc7x4657);
289
290 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
291 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
292 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
293 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
294 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
295 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
296 __m256i vacc5x01234567 = _mm256_permutevar8x32_epi32(vacc5x02461357, vpermute_mask);
297 __m256i vacc6x01234567 = _mm256_permutevar8x32_epi32(vacc6x02461357, vpermute_mask);
298 __m256i vacc7x01234567 = _mm256_permutevar8x32_epi32(vacc7x02461357, vpermute_mask);
299
300 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
301 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
302 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
303 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
304 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
305 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
306 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
307 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
308
309 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale);
310 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale);
311 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale);
312 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale);
313 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale);
314 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_scale);
315 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_scale);
316 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_scale);
317
318 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
319 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
320 w = (const float*) w + 16;
321 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
322 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
323 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
324 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
325 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
326 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
327 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
328 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
329
330 vout7x01234567 = _mm256_max_ps(vout7x01234567, vmin);
331 vout6x01234567 = _mm256_max_ps(vout6x01234567, vmin);
332 vout5x01234567 = _mm256_max_ps(vout5x01234567, vmin);
333 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
334 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
335 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
336 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
337 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
338
339 vout7x01234567 = _mm256_min_ps(vout7x01234567, vmax);
340 vout6x01234567 = _mm256_min_ps(vout6x01234567, vmax);
341 vout5x01234567 = _mm256_min_ps(vout5x01234567, vmax);
342 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
343 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
344 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
345 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
346 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
347 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
348 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
349 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
350 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
351 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
352 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
353 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
354 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
355 if XNN_LIKELY(nc >= 8) {
356 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
357 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
358 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
359 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
360 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
361 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
362 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
363 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
364 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
365 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
366 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
367 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
368 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
369 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
370 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
371 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
372
373 a = (const int8_t**restrict) ((uintptr_t) a - ks);
374 nc -= 8;
375 } else {
376 // Prepare mask for valid 16-bit elements (depends on nc).
377 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
378 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
379 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
380 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
381 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
382 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
383 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
384 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
385 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
386 nc = 0;
387 }
388 } while (nc != 0);
389 }