Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-9x8c8-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_9x8c8__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 <= 9);
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
82 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
83 XNN_FORCE_REALIZATION(vsign_mask);
84 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
85 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
86 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
87 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
88 // XNN_FORCE_REALIZATION(vinput_zero_point);
89 // XNN_FORCE_REALIZATION(vinput_inv_scale);
90 // XNN_FORCE_REALIZATION(voutput_min);
91 // XNN_FORCE_REALIZATION(voutput_max);
92 do {
93 const __m256i vksum01234567 = _mm256_load_si256(w);
94 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
95 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
96 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
97 __m256i vacc1x0123 = vacc0x0123;
98 __m256i vacc1x4567 = vacc0x4567;
99 __m256i vacc2x0123 = vacc0x0123;
100 __m256i vacc2x4567 = vacc0x4567;
101 __m256i vacc3x0123 = vacc0x0123;
102 __m256i vacc3x4567 = vacc0x4567;
103 __m256i vacc4x0123 = vacc0x0123;
104 __m256i vacc4x4567 = vacc0x4567;
105 __m256i vacc5x0123 = vacc0x0123;
106 __m256i vacc5x4567 = vacc0x4567;
107 __m256i vacc6x0123 = vacc0x0123;
108 __m256i vacc6x4567 = vacc0x4567;
109 __m256i vacc7x0123 = vacc0x0123;
110 __m256i vacc7x4567 = vacc0x4567;
111 __m256i vacc8x0123 = vacc0x0123;
112 __m256i vacc8x4567 = vacc0x4567;
113 w = (const int32_t*) w + 8;
114
115 size_t p = ks;
116 do {
117 const int8_t* restrict a0 = a[0];
118 if XNN_UNPREDICTABLE(a0 != zero) {
119 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
120 } else {
121 a0 = zero_data;
122 }
123 const int8_t* restrict a1 = a[1];
124 if XNN_UNPREDICTABLE(a1 != zero) {
125 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
126 } else {
127 a1 = zero_data;
128 }
129 const int8_t* restrict a2 = a[2];
130 if XNN_UNPREDICTABLE(a2 != zero) {
131 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
132 } else {
133 a2 = zero_data;
134 }
135 const int8_t* restrict a3 = a[3];
136 if XNN_UNPREDICTABLE(a3 != zero) {
137 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
138 } else {
139 a3 = zero_data;
140 }
141 const int8_t* restrict a4 = a[4];
142 if XNN_UNPREDICTABLE(a4 != zero) {
143 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
144 } else {
145 a4 = zero_data;
146 }
147 const int8_t* restrict a5 = a[5];
148 if XNN_UNPREDICTABLE(a5 != zero) {
149 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
150 } else {
151 a5 = zero_data;
152 }
153 const int8_t* restrict a6 = a[6];
154 if XNN_UNPREDICTABLE(a6 != zero) {
155 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
156 } else {
157 a6 = zero_data;
158 }
159 const int8_t* restrict a7 = a[7];
160 if XNN_UNPREDICTABLE(a7 != zero) {
161 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
162 } else {
163 a7 = zero_data;
164 }
165 const int8_t* restrict a8 = a[8];
166 if XNN_UNPREDICTABLE(a8 != zero) {
167 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
168 } else {
169 a8 = zero_data;
170 }
171 a += 9;
172
173 size_t k = kc;
174 while (k >= 16 * sizeof(int8_t)) {
175 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
176 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
177 a0 += 16;
178 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
179 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
180 a1 += 16;
181 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
182 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
183 a2 += 16;
184 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
185 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
186 a3 += 16;
187 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
188 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
189 a4 += 16;
190 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
191 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
192 a5 += 16;
193 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
194 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
195 a6 += 16;
196 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
197 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
198 a7 += 16;
199 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
200 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
201 a8 += 16;
202
203 const __m256i vb01234567x0123 = _mm256_load_si256(w);
204 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
205 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
206 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
207 xnn_prefetch_to_l1((const int8_t*) w + 896);
208
209 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
210 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
211 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
212 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
213 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
214 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
215 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
216 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
217 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
218 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
219 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
220 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
221 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
222 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
223 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
224 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
225 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
226 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
227 xnn_prefetch_to_l1((const int8_t*) w + 960);
228 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
229 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
230 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
231 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
232 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
233 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
234 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
235 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
236 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
237 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
238 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
239 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
240 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
241 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
242 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
243 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
244 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
245 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
246
247 w = (const int8_t*) w + 128;
248 k -= 16 * sizeof(int8_t);
249 }
250
251 if (k != 0) {
252 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
253 a0 += 8;
254 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
255 a1 += 8;
256 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
257 a2 += 8;
258 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
259 a3 += 8;
260 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
261 a4 += 8;
262 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
263 a5 += 8;
264 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
265 a6 += 8;
266 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
267 a7 += 8;
268 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
269 a8 += 8;
270
271 const __m256i vb01234567x0123 = _mm256_load_si256(w);
272 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
273
274 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
275 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
276 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
277 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
278 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
279 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
280 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
281 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
282 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
283 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
284 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
285 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
286 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
287 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
288 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
289 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
290 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
291 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
292 xnn_prefetch_to_l1((const int8_t*) w + 960);
293
294 w = (const int8_t*) w + 64;
295 k -= 8 * sizeof(int8_t);
296 }
297
298 p -= 9 * sizeof(void*);
299 } while (p != 0);
300
301
302 // Add adjacent pairs
303 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
304 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
305 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
306 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
307 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
308 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
309 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
310 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
311 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
312 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
313 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
314 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
315 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
316 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
317 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
318 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
319 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
320 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
321
322 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
323 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
324 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
325 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
326 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
327 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
328 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
329 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
330 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
331
332 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
333 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
334 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
335 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
336 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
337 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
338 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
339 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_inv_scale);
340 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vinput_inv_scale);
341
342 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
343 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
344 w = (const float*) w + 16;
345 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
346 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
347 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
348 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
349 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
350 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
351 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
352 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
353 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
354
355 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
356 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
357 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
358 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
359 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
360 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
361 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
362 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
363 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
364
365 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
366 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
367 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
368 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
369 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
370 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
371 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
372 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
373 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
374
375 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
376 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
377 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
378 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
379 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
380 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
381 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
382 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
383 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
384 if XNN_LIKELY(nc >= 8) {
385 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
386 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
387 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
388 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
389 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
390 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
391 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
392 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
393 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
394 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
395 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
396 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
397 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
398 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
399 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
400 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
401 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
402 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
403
404 a = (const int8_t**restrict) ((uintptr_t) a - ks);
405 nc -= 8;
406 } else {
407 // Prepare mask for valid 16-bit elements (depends on nc).
408 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
409 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
410 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
411 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
412 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
413 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
414 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
415 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
416 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
417 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
418 nc = 0;
419 }
420 } while (nc != 0);
421 }