Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-4x8c8-minmax-avx256skx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/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_qd8_f16_qc4w_gemm_minmax_ukernel_4x8c8__avx256skx(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 uint16_t* c0 = (uint16_t*) c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr != 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64
65 const __m128i vmask = _mm_set1_epi8(0xF0);
66 XNN_FORCE_REALIZATION(vmask);
67 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
68 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
69 XNN_FORCE_REALIZATION(vmin);
70 XNN_FORCE_REALIZATION(vmax);
71
72 do {
73 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
74 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
75 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
76 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
77 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
78 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
79 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
80 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
81 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
82 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
83 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
84 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
85 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
86 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
87 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
88 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
89 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
90 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
91 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
92 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
93 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
94 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
95 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
96 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
97 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
98 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
99 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
100 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
101 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
102 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
103 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
104 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
105 w = (const int32_t*) w + 8;
106
107 size_t k = kc;
108 while (k >= 16 * sizeof(int8_t)) {
109 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
110 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
111 a0 += 8;
112 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
113 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
114 a1 += 8;
115 __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
116 __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
117 a2 += 8;
118 __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
119 __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
120 a3 += 8;
121
122 __m128i vb01 = _mm_load_si128((const __m128i*) w);
123 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
124 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
125 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
126
127 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
128 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
129 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
130 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
131 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
132 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
133 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
134 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
135
136 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
137 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
138 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
139 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
140 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
141 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
142 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
143 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
144
145 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
146 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
147 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
148 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
149 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
150 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
151 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
152 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
153
154 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
155 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
156 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
157 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
158
159 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
160 vxa0 = _mm256_cvtepi8_epi16(va0);
161 a0 += 8;
162 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
163 vxa1 = _mm256_cvtepi8_epi16(va1);
164 a1 += 8;
165 va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
166 vxa2 = _mm256_cvtepi8_epi16(va2);
167 a2 += 8;
168 va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
169 vxa3 = _mm256_cvtepi8_epi16(va3);
170 a3 += 8;
171
172 vbm01 = _mm_and_si128(vb01, vmask);
173 vxb01 = _mm256_cvtepi8_epi16(vbm01);
174
175 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
176 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
177 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
178 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
179 vbm23 = _mm_and_si128(vb23, vmask);
180 vxb23 = _mm256_cvtepi8_epi16(vbm23);
181
182 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
183 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
184 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
185 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
186 vbm45 = _mm_and_si128(vb45, vmask);
187 vxb45 = _mm256_cvtepi8_epi16(vbm45);
188
189 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
190 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
191 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
192 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
193 vbm67 = _mm_and_si128(vb67, vmask);
194 vxb67 = _mm256_cvtepi8_epi16(vbm67);
195
196 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
197 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
198 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
199 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
200
201 w = (const int8_t*) w + 64;
202 k -= 16 * sizeof(int8_t);
203 }
204
205 while (k >= 8 * sizeof(int8_t)) {
206 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
207 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
208 a0 += 8;
209 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
210 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
211 a1 += 8;
212 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
213 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
214 a2 += 8;
215 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
216 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
217 a3 += 8;
218
219 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
220 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
221 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
222 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
223
224 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
225 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
226 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
227 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
228 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
229 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
230 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
231 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
232
233 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
234 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
235 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
236 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
237 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
238 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
239 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
240 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
241
242 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
243 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
244 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
245 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
246 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
247 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
248 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
249 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
250
251 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
252 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
253 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
254 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
255
256 w = (const int8_t*) w + 64;
257 k -= 8 * sizeof(int8_t);
258 }
259
260 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
261 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
262 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
263 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
264 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
265 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
266 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
267 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
268
269 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
270 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
271 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
272 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
273
274 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
275 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
276 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
277 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
278 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
279
280 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
281 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
282 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
283 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
284 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
285 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
286 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
287 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
288 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
289 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
290 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
291 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
292
293 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
294 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
295 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
296 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
297
298 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
299 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
300 w = (const float*) w + 16;
301 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
302 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
303 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
304 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
305
306 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
307 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
308 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
309 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
310
311 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
312 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
313 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
314 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
315 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
316 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
317 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
318 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
319 if XNN_LIKELY(nc >= 8) {
320 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
321 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
322 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
323 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
324 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
325 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
326 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
327 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
328
329 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
330 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
331 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
332 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
333
334 nc -= 8;
335 } else {
336 // Prepare mask for valid 16-bit elements (depends on nc).
337 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
338 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
339 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
340 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
341 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
342 nc = 0;
343 }
344 } while (nc != 0);
345 }