Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-7x8c8-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_7x8c8__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 <= 7);
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
74 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
75 XNN_FORCE_REALIZATION(vsign_mask);
76 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
77 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
78 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
79 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
80 // XNN_FORCE_REALIZATION(vinput_zero_point);
81 // XNN_FORCE_REALIZATION(vinput_inv_scale);
82 // XNN_FORCE_REALIZATION(voutput_min);
83 // XNN_FORCE_REALIZATION(voutput_max);
84 do {
85 const __m256i vksum01234567 = _mm256_load_si256(w);
86 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
87 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
88 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
89 __m256i vacc1x0123 = vacc0x0123;
90 __m256i vacc1x4567 = vacc0x4567;
91 __m256i vacc2x0123 = vacc0x0123;
92 __m256i vacc2x4567 = vacc0x4567;
93 __m256i vacc3x0123 = vacc0x0123;
94 __m256i vacc3x4567 = vacc0x4567;
95 __m256i vacc4x0123 = vacc0x0123;
96 __m256i vacc4x4567 = vacc0x4567;
97 __m256i vacc5x0123 = vacc0x0123;
98 __m256i vacc5x4567 = vacc0x4567;
99 __m256i vacc6x0123 = vacc0x0123;
100 __m256i vacc6x4567 = vacc0x4567;
101 w = (const int32_t*) w + 8;
102
103 size_t p = ks;
104 do {
105 const int8_t* restrict a0 = a[0];
106 if XNN_UNPREDICTABLE(a0 != zero) {
107 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
108 } else {
109 a0 = zero_data;
110 }
111 const int8_t* restrict a1 = a[1];
112 if XNN_UNPREDICTABLE(a1 != zero) {
113 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
114 } else {
115 a1 = zero_data;
116 }
117 const int8_t* restrict a2 = a[2];
118 if XNN_UNPREDICTABLE(a2 != zero) {
119 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
120 } else {
121 a2 = zero_data;
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 } else {
127 a3 = zero_data;
128 }
129 const int8_t* restrict a4 = a[4];
130 if XNN_UNPREDICTABLE(a4 != zero) {
131 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
132 } else {
133 a4 = zero_data;
134 }
135 const int8_t* restrict a5 = a[5];
136 if XNN_UNPREDICTABLE(a5 != zero) {
137 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
138 } else {
139 a5 = zero_data;
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 } else {
145 a6 = zero_data;
146 }
147 a += 7;
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
173 const __m256i vb01234567x0123 = _mm256_load_si256(w);
174 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
175 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
176 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
177 xnn_prefetch_to_l1((const int8_t*) w + 896);
178
179 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
180 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
181 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
182 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
183 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
184 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
185 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
186 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
187 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
188 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
189 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
190 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
191 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
192 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
193 xnn_prefetch_to_l1((const int8_t*) w + 960);
194 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
195 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
196 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
197 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
198 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
199 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
200 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
201 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
202 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
203 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
204 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
205 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
206 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
207 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
208
209 w = (const int8_t*) w + 128;
210 k -= 16 * sizeof(int8_t);
211 }
212
213 if (k != 0) {
214 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
215 a0 += 8;
216 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
217 a1 += 8;
218 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
219 a2 += 8;
220 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
221 a3 += 8;
222 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
223 a4 += 8;
224 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
225 a5 += 8;
226 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
227 a6 += 8;
228
229 const __m256i vb01234567x0123 = _mm256_load_si256(w);
230 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
231
232 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
233 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
234 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
235 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
236 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
237 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
238 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
239 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
240 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
241 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
242 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
243 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
244 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
245 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
246 xnn_prefetch_to_l1((const int8_t*) w + 960);
247
248 w = (const int8_t*) w + 64;
249 k -= 8 * sizeof(int8_t);
250 }
251
252 p -= 7 * sizeof(void*);
253 } while (p != 0);
254
255
256 // Add adjacent pairs
257 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
258 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
259 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
260 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
261 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
262 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
263 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
264 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
265 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
266 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
267 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
268 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
269 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
270 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
271
272 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
273 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
274 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
275 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
276 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
277 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
278 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
279
280 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
281 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
282 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
283 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
284 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
285 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
286 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
287
288 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
289 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
290 w = (const float*) w + 16;
291 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
292 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
293 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
294 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
295 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
296 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
297 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
298
299 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
300 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
301 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
302 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
303 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
304 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
305 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
306
307 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
308 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
309 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
310 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
311 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
312 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
313 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
314
315 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
316 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
317 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
318 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
319 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
320 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
321 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
322 if XNN_LIKELY(nc >= 8) {
323 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
324 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
325 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
326 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
327 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
328 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
329 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
330 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
331 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
332 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
333 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
334 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
335 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
336 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
337
338 a = (const int8_t**restrict) ((uintptr_t) a - ks);
339 nc -= 8;
340 } else {
341 // Prepare mask for valid 16-bit elements (depends on nc).
342 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
343 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
344 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
345 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
346 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
347 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
348 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
349 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
350 nc = 0;
351 }
352 } while (nc != 0);
353 }