Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-10x8c8-minmax-avx256vnni.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
20
21 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_10x8c8__avx256vnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 10);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 uint16_t* c0 = (uint16_t*) c;
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
65 if XNN_UNPREDICTABLE(mr < 6) {
66 c5 = c4;
67 }
68 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
69 if XNN_UNPREDICTABLE(mr <= 6) {
70 c6 = c5;
71 }
72 uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 8) {
74 c7 = c6;
75 }
76 uint16_t* c8 = (uint16_t*) ((uintptr_t) c7 + cm_stride);
77 if XNN_UNPREDICTABLE(mr <= 8) {
78 c8 = c7;
79 }
80 uint16_t* c9 = (uint16_t*) ((uintptr_t) c8 + cm_stride);
81 if XNN_UNPREDICTABLE(mr != 10) {
82 c9 = c8;
83 }
84
85 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
86 XNN_FORCE_REALIZATION(vsign_mask);
87 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
88 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
89 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
90 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
91 // XNN_FORCE_REALIZATION(vinput_zero_point);
92 // XNN_FORCE_REALIZATION(vinput_inv_scale);
93 // XNN_FORCE_REALIZATION(voutput_min);
94 // XNN_FORCE_REALIZATION(voutput_max);
95 do {
96 const __m256i vksum01234567 = _mm256_load_si256(w);
97 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
98 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
99 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
100 __m256i vacc1x0123 = vacc0x0123;
101 __m256i vacc1x4567 = vacc0x4567;
102 __m256i vacc2x0123 = vacc0x0123;
103 __m256i vacc2x4567 = vacc0x4567;
104 __m256i vacc3x0123 = vacc0x0123;
105 __m256i vacc3x4567 = vacc0x4567;
106 __m256i vacc4x0123 = vacc0x0123;
107 __m256i vacc4x4567 = vacc0x4567;
108 __m256i vacc5x0123 = vacc0x0123;
109 __m256i vacc5x4567 = vacc0x4567;
110 __m256i vacc6x0123 = vacc0x0123;
111 __m256i vacc6x4567 = vacc0x4567;
112 __m256i vacc7x0123 = vacc0x0123;
113 __m256i vacc7x4567 = vacc0x4567;
114 __m256i vacc8x0123 = vacc0x0123;
115 __m256i vacc8x4567 = vacc0x4567;
116 __m256i vacc9x0123 = vacc0x0123;
117 __m256i vacc9x4567 = vacc0x4567;
118 w = (const int32_t*) w + 8;
119
120 size_t p = ks;
121 do {
122 const int8_t* restrict a0 = a[0];
123 if XNN_UNPREDICTABLE(a0 != zero) {
124 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
125 } else {
126 a0 = zero_data;
127 }
128 const int8_t* restrict a1 = a[1];
129 if XNN_UNPREDICTABLE(a1 != zero) {
130 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
131 } else {
132 a1 = zero_data;
133 }
134 const int8_t* restrict a2 = a[2];
135 if XNN_UNPREDICTABLE(a2 != zero) {
136 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
137 } else {
138 a2 = zero_data;
139 }
140 const int8_t* restrict a3 = a[3];
141 if XNN_UNPREDICTABLE(a3 != zero) {
142 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
143 } else {
144 a3 = zero_data;
145 }
146 const int8_t* restrict a4 = a[4];
147 if XNN_UNPREDICTABLE(a4 != zero) {
148 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
149 } else {
150 a4 = zero_data;
151 }
152 const int8_t* restrict a5 = a[5];
153 if XNN_UNPREDICTABLE(a5 != zero) {
154 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
155 } else {
156 a5 = zero_data;
157 }
158 const int8_t* restrict a6 = a[6];
159 if XNN_UNPREDICTABLE(a6 != zero) {
160 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
161 } else {
162 a6 = zero_data;
163 }
164 const int8_t* restrict a7 = a[7];
165 if XNN_UNPREDICTABLE(a7 != zero) {
166 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
167 } else {
168 a7 = zero_data;
169 }
170 const int8_t* restrict a8 = a[8];
171 if XNN_UNPREDICTABLE(a8 != zero) {
172 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
173 } else {
174 a8 = zero_data;
175 }
176 const int8_t* restrict a9 = a[9];
177 if XNN_UNPREDICTABLE(a9 != zero) {
178 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
179 } else {
180 a9 = zero_data;
181 }
182 a += 10;
183
184 size_t k = kc;
185 while (k >= 16 * sizeof(int8_t)) {
186 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
187 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
188 a0 += 16;
189 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
190 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
191 a1 += 16;
192 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
193 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
194 a2 += 16;
195 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
196 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
197 a3 += 16;
198 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
199 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
200 a4 += 16;
201 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
202 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
203 a5 += 16;
204 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
205 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
206 a6 += 16;
207 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
208 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
209 a7 += 16;
210 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
211 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
212 a8 += 16;
213 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
214 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
215 a9 += 16;
216
217 const __m256i vb01234567x0123 = _mm256_load_si256(w);
218 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
219 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
220 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
221
222 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
223 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
224 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
225 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
226 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
227 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
228 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
229 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
230 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
231 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
232 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
233 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
234 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
235 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
236 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
237 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
238 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
239 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
240 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
241 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
242 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
243 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
244 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
245 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
246 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
247 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
248 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
249 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
250 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
251 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
252 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
253 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
254 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
255 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
256 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
257 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
258 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
259 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
260 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
261 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
262
263 w = (const int8_t*) w + 128;
264 k -= 16 * sizeof(int8_t);
265 }
266
267 if (k != 0) {
268 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
269 a0 += 8;
270 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
271 a1 += 8;
272 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
273 a2 += 8;
274 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
275 a3 += 8;
276 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
277 a4 += 8;
278 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
279 a5 += 8;
280 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
281 a6 += 8;
282 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
283 a7 += 8;
284 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
285 a8 += 8;
286 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
287 a9 += 8;
288
289 const __m256i vb01234567x0123 = _mm256_load_si256(w);
290 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
291
292 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
293 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
294 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
295 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
296 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
297 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
298 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
299 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
300 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
301 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
302 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
303 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
304 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
305 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
306 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
307 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
308 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
309 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
310 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
311 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
312
313 w = (const int8_t*) w + 64;
314 k -= 8 * sizeof(int8_t);
315 }
316
317 p -= 10 * sizeof(void*);
318 } while (p != 0);
319
320
321 // Add adjacent pairs
322 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
323 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
324 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
325 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
326 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
327 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
328 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
329 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
330 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
331 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
332 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
333 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
334 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
335 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
336 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
337 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
338 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
339 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
340 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
341 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
342
343 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
344 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
345 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
346 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
347 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
348 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
349 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
350 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
351 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
352 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
353
354 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
355 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
356 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
357 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
358 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
359 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
360 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
361 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_inv_scale);
362 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vinput_inv_scale);
363 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vinput_inv_scale);
364
365 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
366 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
367 w = (const float*) w + 16;
368 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
369 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
370 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
371 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
372 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
373 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
374 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
375 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
376 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
377 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
378
379 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
380 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
381 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
382 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
383 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
384 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
385 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
386 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
387 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
388 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
389
390 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
391 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
392 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
393 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
394 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
395 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
396 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
397 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
398 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
399 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
400
401 __m128i vfp16out9x01234567 = _mm256_cvtps_ph(vout9x01234567, _MM_FROUND_TO_NEAREST_INT);
402 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
403 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
404 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
405 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
406 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
407 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
408 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
409 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
410 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
411 if XNN_LIKELY(nc >= 8) {
412 _mm_storeu_si128((__m128i*) c9, vfp16out9x01234567);
413 c9 = (uint16_t*) ((uintptr_t) c9 + cn_stride);
414 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
415 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
416 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
417 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
418 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
419 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
420 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
421 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
422 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
423 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
424 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
425 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
426 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
427 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
428 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
429 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
430 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
431 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
432
433 a = (const int8_t**restrict) ((uintptr_t) a - ks);
434 nc -= 8;
435 } else {
436 // Prepare mask for valid 16-bit elements (depends on nc).
437 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
438 _mm_mask_storeu_epi16(c9, vmask, vfp16out9x01234567);
439 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
440 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
441 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
442 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
443 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
444 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
445 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
446 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
447 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
448 nc = 0;
449 }
450 } while (nc != 0);
451 }