Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-10x8c8-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_10x8c8__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 <= 10);
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 uint16_t* c8 = (uint16_t*) ((uintptr_t) c7 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 8) {
79 c8 = c7;
80 }
81 uint16_t* c9 = (uint16_t*) ((uintptr_t) c8 + cm_stride);
82 if XNN_UNPREDICTABLE(mr != 10) {
83 c9 = c8;
84 }
85
86 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
87 XNN_FORCE_REALIZATION(vsign_mask);
88 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
89 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
90 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
91 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
92 // XNN_FORCE_REALIZATION(vinput_zero_point);
93 // XNN_FORCE_REALIZATION(vinput_inv_scale);
94 // XNN_FORCE_REALIZATION(voutput_min);
95 // XNN_FORCE_REALIZATION(voutput_max);
96 do {
97 const __m256i vksum01234567 = _mm256_load_si256(w);
98 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
99 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
100 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
101 __m256i vacc1x0123 = vacc0x0123;
102 __m256i vacc1x4567 = vacc0x4567;
103 __m256i vacc2x0123 = vacc0x0123;
104 __m256i vacc2x4567 = vacc0x4567;
105 __m256i vacc3x0123 = vacc0x0123;
106 __m256i vacc3x4567 = vacc0x4567;
107 __m256i vacc4x0123 = vacc0x0123;
108 __m256i vacc4x4567 = vacc0x4567;
109 __m256i vacc5x0123 = vacc0x0123;
110 __m256i vacc5x4567 = vacc0x4567;
111 __m256i vacc6x0123 = vacc0x0123;
112 __m256i vacc6x4567 = vacc0x4567;
113 __m256i vacc7x0123 = vacc0x0123;
114 __m256i vacc7x4567 = vacc0x4567;
115 __m256i vacc8x0123 = vacc0x0123;
116 __m256i vacc8x4567 = vacc0x4567;
117 __m256i vacc9x0123 = vacc0x0123;
118 __m256i vacc9x4567 = vacc0x4567;
119 w = (const int32_t*) w + 8;
120
121 size_t p = ks;
122 do {
123 const int8_t* restrict a0 = a[0];
124 if XNN_UNPREDICTABLE(a0 != zero) {
125 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
126 } else {
127 a0 = zero_data;
128 }
129 const int8_t* restrict a1 = a[1];
130 if XNN_UNPREDICTABLE(a1 != zero) {
131 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
132 } else {
133 a1 = zero_data;
134 }
135 const int8_t* restrict a2 = a[2];
136 if XNN_UNPREDICTABLE(a2 != zero) {
137 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
138 } else {
139 a2 = zero_data;
140 }
141 const int8_t* restrict a3 = a[3];
142 if XNN_UNPREDICTABLE(a3 != zero) {
143 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
144 } else {
145 a3 = zero_data;
146 }
147 const int8_t* restrict a4 = a[4];
148 if XNN_UNPREDICTABLE(a4 != zero) {
149 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
150 } else {
151 a4 = zero_data;
152 }
153 const int8_t* restrict a5 = a[5];
154 if XNN_UNPREDICTABLE(a5 != zero) {
155 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
156 } else {
157 a5 = zero_data;
158 }
159 const int8_t* restrict a6 = a[6];
160 if XNN_UNPREDICTABLE(a6 != zero) {
161 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
162 } else {
163 a6 = zero_data;
164 }
165 const int8_t* restrict a7 = a[7];
166 if XNN_UNPREDICTABLE(a7 != zero) {
167 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
168 } else {
169 a7 = zero_data;
170 }
171 const int8_t* restrict a8 = a[8];
172 if XNN_UNPREDICTABLE(a8 != zero) {
173 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
174 } else {
175 a8 = zero_data;
176 }
177 const int8_t* restrict a9 = a[9];
178 if XNN_UNPREDICTABLE(a9 != zero) {
179 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
180 } else {
181 a9 = zero_data;
182 }
183 a += 10;
184
185 size_t k = kc;
186 while (k >= 16 * sizeof(int8_t)) {
187 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
188 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
189 a0 += 16;
190 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
191 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
192 a1 += 16;
193 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
194 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
195 a2 += 16;
196 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
197 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
198 a3 += 16;
199 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
200 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
201 a4 += 16;
202 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
203 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
204 a5 += 16;
205 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
206 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
207 a6 += 16;
208 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
209 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
210 a7 += 16;
211 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
212 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
213 a8 += 16;
214 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
215 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
216 a9 += 16;
217
218 const __m256i vb01234567x0123 = _mm256_load_si256(w);
219 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
220 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
221 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
222 xnn_prefetch_to_l1((const int8_t*) w + 896);
223
224 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
225 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
226 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
227 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
228 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
229 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
230 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
231 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
232 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
233 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
234 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
235 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
236 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
237 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
238 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
239 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
240 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
241 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
242 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
243 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
244 xnn_prefetch_to_l1((const int8_t*) w + 960);
245 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
246 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
247 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
248 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
249 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
250 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
251 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
252 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
253 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
254 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
255 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
256 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
257 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
258 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
259 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
260 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
261 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
262 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
263 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
264 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
265
266 w = (const int8_t*) w + 128;
267 k -= 16 * sizeof(int8_t);
268 }
269
270 if (k != 0) {
271 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
272 a0 += 8;
273 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
274 a1 += 8;
275 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
276 a2 += 8;
277 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
278 a3 += 8;
279 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
280 a4 += 8;
281 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
282 a5 += 8;
283 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
284 a6 += 8;
285 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
286 a7 += 8;
287 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
288 a8 += 8;
289 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
290 a9 += 8;
291
292 const __m256i vb01234567x0123 = _mm256_load_si256(w);
293 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
294
295 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
296 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
297 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
298 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
299 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
300 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
301 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
302 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
303 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
304 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
305 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
306 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
307 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
308 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
309 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
310 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
311 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
312 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
313 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
314 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
315 xnn_prefetch_to_l1((const int8_t*) w + 960);
316
317 w = (const int8_t*) w + 64;
318 k -= 8 * sizeof(int8_t);
319 }
320
321 p -= 10 * sizeof(void*);
322 } while (p != 0);
323
324
325 // Add adjacent pairs
326 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
327 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
328 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
329 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
330 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
331 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
332 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
333 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
334 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
335 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
336 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
337 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
338 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
339 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
340 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
341 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
342 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
343 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
344 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
345 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
346
347 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
348 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
349 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
350 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
351 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
352 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
353 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
354 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
355 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
356 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
357
358 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
359 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
360 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
361 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
362 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
363 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
364 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
365 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_inv_scale);
366 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vinput_inv_scale);
367 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vinput_inv_scale);
368
369 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
370 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
371 w = (const float*) w + 16;
372 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
373 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
374 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
375 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
376 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
377 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
378 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
379 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
380 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
381 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
382
383 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
384 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
385 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
386 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
387 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
388 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
389 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
390 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
391 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
392 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
393
394 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
395 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
396 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
397 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
398 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
399 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
400 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
401 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
402 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
403 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
404
405 __m128i vfp16out9x01234567 = _mm256_cvtps_ph(vout9x01234567, _MM_FROUND_TO_NEAREST_INT);
406 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
407 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
408 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
409 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
410 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
411 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
412 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
413 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
414 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
415 if XNN_LIKELY(nc >= 8) {
416 _mm_storeu_si128((__m128i*) c9, vfp16out9x01234567);
417 c9 = (uint16_t*) ((uintptr_t) c9 + cn_stride);
418 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
419 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
420 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
421 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
422 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
423 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
424 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
425 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
426 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
427 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
428 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
429 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
430 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
431 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
432 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
433 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
434 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
435 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
436
437 a = (const int8_t**restrict) ((uintptr_t) a - ks);
438 nc -= 8;
439 } else {
440 // Prepare mask for valid 16-bit elements (depends on nc).
441 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
442 _mm_mask_storeu_epi16(c9, vmask, vfp16out9x01234567);
443 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
444 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
445 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
446 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
447 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
448 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
449 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
450 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
451 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
452 nc = 0;
453 }
454 } while (nc != 0);
455 }