Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-12x8c8-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_12x8c8__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 <= 12);
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 uint16_t* c10 = (uint16_t*) ((uintptr_t) c9 + cm_stride);
85 if XNN_UNPREDICTABLE(mr <= 10) {
86 c10 = c9;
87 }
88 uint16_t* c11 = (uint16_t*) ((uintptr_t) c10 + cm_stride);
89 if XNN_UNPREDICTABLE(mr != 12) {
90 c11 = c10;
91 }
92
93 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
94 XNN_FORCE_REALIZATION(vsign_mask);
95 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
96 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
97 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
98 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
99 // XNN_FORCE_REALIZATION(vinput_zero_point);
100 // XNN_FORCE_REALIZATION(vinput_inv_scale);
101 // XNN_FORCE_REALIZATION(voutput_min);
102 // XNN_FORCE_REALIZATION(voutput_max);
103 do {
104 const __m256i vksum01234567 = _mm256_load_si256(w);
105 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
106 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
107 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
108 __m256i vacc1x0123 = vacc0x0123;
109 __m256i vacc1x4567 = vacc0x4567;
110 __m256i vacc2x0123 = vacc0x0123;
111 __m256i vacc2x4567 = vacc0x4567;
112 __m256i vacc3x0123 = vacc0x0123;
113 __m256i vacc3x4567 = vacc0x4567;
114 __m256i vacc4x0123 = vacc0x0123;
115 __m256i vacc4x4567 = vacc0x4567;
116 __m256i vacc5x0123 = vacc0x0123;
117 __m256i vacc5x4567 = vacc0x4567;
118 __m256i vacc6x0123 = vacc0x0123;
119 __m256i vacc6x4567 = vacc0x4567;
120 __m256i vacc7x0123 = vacc0x0123;
121 __m256i vacc7x4567 = vacc0x4567;
122 __m256i vacc8x0123 = vacc0x0123;
123 __m256i vacc8x4567 = vacc0x4567;
124 __m256i vacc9x0123 = vacc0x0123;
125 __m256i vacc9x4567 = vacc0x4567;
126 __m256i vacc10x0123 = vacc0x0123;
127 __m256i vacc10x4567 = vacc0x4567;
128 __m256i vacc11x0123 = vacc0x0123;
129 __m256i vacc11x4567 = vacc0x4567;
130 w = (const int32_t*) w + 8;
131
132 size_t p = ks;
133 do {
134 const int8_t* restrict a0 = a[0];
135 if XNN_UNPREDICTABLE(a0 != zero) {
136 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
137 } else {
138 a0 = zero_data;
139 }
140 const int8_t* restrict a1 = a[1];
141 if XNN_UNPREDICTABLE(a1 != zero) {
142 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
143 } else {
144 a1 = zero_data;
145 }
146 const int8_t* restrict a2 = a[2];
147 if XNN_UNPREDICTABLE(a2 != zero) {
148 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
149 } else {
150 a2 = zero_data;
151 }
152 const int8_t* restrict a3 = a[3];
153 if XNN_UNPREDICTABLE(a3 != zero) {
154 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
155 } else {
156 a3 = zero_data;
157 }
158 const int8_t* restrict a4 = a[4];
159 if XNN_UNPREDICTABLE(a4 != zero) {
160 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
161 } else {
162 a4 = zero_data;
163 }
164 const int8_t* restrict a5 = a[5];
165 if XNN_UNPREDICTABLE(a5 != zero) {
166 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
167 } else {
168 a5 = zero_data;
169 }
170 const int8_t* restrict a6 = a[6];
171 if XNN_UNPREDICTABLE(a6 != zero) {
172 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
173 } else {
174 a6 = zero_data;
175 }
176 const int8_t* restrict a7 = a[7];
177 if XNN_UNPREDICTABLE(a7 != zero) {
178 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
179 } else {
180 a7 = zero_data;
181 }
182 const int8_t* restrict a8 = a[8];
183 if XNN_UNPREDICTABLE(a8 != zero) {
184 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
185 } else {
186 a8 = zero_data;
187 }
188 const int8_t* restrict a9 = a[9];
189 if XNN_UNPREDICTABLE(a9 != zero) {
190 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
191 } else {
192 a9 = zero_data;
193 }
194 const int8_t* restrict a10 = a[10];
195 if XNN_UNPREDICTABLE(a10 != zero) {
196 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
197 } else {
198 a10 = zero_data;
199 }
200 const int8_t* restrict a11 = a[11];
201 if XNN_UNPREDICTABLE(a11 != zero) {
202 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
203 } else {
204 a11 = zero_data;
205 }
206 a += 12;
207
208 size_t k = kc;
209 while (k >= 16 * sizeof(int8_t)) {
210 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
211 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
212 a0 += 16;
213 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
214 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
215 a1 += 16;
216 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
217 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
218 a2 += 16;
219 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
220 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
221 a3 += 16;
222 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
223 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
224 a4 += 16;
225 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
226 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
227 a5 += 16;
228 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
229 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
230 a6 += 16;
231 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
232 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
233 a7 += 16;
234 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
235 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
236 a8 += 16;
237 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
238 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
239 a9 += 16;
240 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
241 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
242 a10 += 16;
243 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
244 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
245 a11 += 16;
246
247 const __m256i vb01234567x0123 = _mm256_load_si256(w);
248 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
249 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
250 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
251
252 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
253 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
254 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
255 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
256 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
257 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
258 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
259 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
260 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
261 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
262 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
263 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
264 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
265 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
266 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
267 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
268 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
269 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
270 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
271 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
272 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
273 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
274 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
275 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
276 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
277 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
278 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
279 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
280 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
281 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
282 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
283 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
284 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
285 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
286 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
287 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
288 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
289 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
290 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
291 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
292 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
293 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
294 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
295 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
296 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
297 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
298 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
299 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
300
301 w = (const int8_t*) w + 128;
302 k -= 16 * sizeof(int8_t);
303 }
304
305 if (k != 0) {
306 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
307 a0 += 8;
308 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
309 a1 += 8;
310 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
311 a2 += 8;
312 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
313 a3 += 8;
314 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
315 a4 += 8;
316 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
317 a5 += 8;
318 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
319 a6 += 8;
320 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
321 a7 += 8;
322 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
323 a8 += 8;
324 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
325 a9 += 8;
326 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
327 a10 += 8;
328 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
329 a11 += 8;
330
331 const __m256i vb01234567x0123 = _mm256_load_si256(w);
332 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
333
334 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
335 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
336 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
337 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
338 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
339 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
340 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
341 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
342 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
343 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
344 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
345 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
346 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
347 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
348 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
349 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
350 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
351 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
352 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
353 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
354 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
355 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
356 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
357 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
358
359 w = (const int8_t*) w + 64;
360 k -= 8 * sizeof(int8_t);
361 }
362
363 p -= 12 * sizeof(void*);
364 } while (p != 0);
365
366
367 // Add adjacent pairs
368 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
369 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
370 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
371 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
372 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
373 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
374 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
375 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
376 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
377 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
378 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
379 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
380 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
381 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
382 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
383 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
384 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
385 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
386 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
387 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
388 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
389 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
390 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
391 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
392
393 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
394 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
395 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
396 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
397 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
398 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
399 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
400 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
401 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
402 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
403 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
404 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
405
406 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
407 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
408 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
409 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
410 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
411 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
412 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
413 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_inv_scale);
414 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vinput_inv_scale);
415 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vinput_inv_scale);
416 vout10x01234567 = _mm256_mul_ps(vout10x01234567, vinput_inv_scale);
417 vout11x01234567 = _mm256_mul_ps(vout11x01234567, vinput_inv_scale);
418
419 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
420 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
421 w = (const float*) w + 16;
422 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
423 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
424 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
425 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
426 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
427 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
428 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
429 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
430 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
431 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
432 vout10x01234567 = _mm256_fmadd_ps(vout10x01234567, vfilter_output_scale01234567, vbias01234567);
433 vout11x01234567 = _mm256_fmadd_ps(vout11x01234567, vfilter_output_scale01234567, vbias01234567);
434
435 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
436 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
437 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
438 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
439 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
440 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
441 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
442 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
443 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
444 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
445 vout10x01234567 = _mm256_max_ps(vout10x01234567, voutput_min);
446 vout11x01234567 = _mm256_max_ps(vout11x01234567, voutput_min);
447
448 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
449 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
450 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
451 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
452 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
453 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
454 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
455 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
456 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
457 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
458 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max);
459 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max);
460
461 __m128i vfp16out11x01234567 = _mm256_cvtps_ph(vout11x01234567, _MM_FROUND_TO_NEAREST_INT);
462 __m128i vfp16out10x01234567 = _mm256_cvtps_ph(vout10x01234567, _MM_FROUND_TO_NEAREST_INT);
463 __m128i vfp16out9x01234567 = _mm256_cvtps_ph(vout9x01234567, _MM_FROUND_TO_NEAREST_INT);
464 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
465 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
466 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
467 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
468 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
469 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
470 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
471 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
472 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
473 if XNN_LIKELY(nc >= 8) {
474 _mm_storeu_si128((__m128i*) c11, vfp16out11x01234567);
475 c11 = (uint16_t*) ((uintptr_t) c11 + cn_stride);
476 _mm_storeu_si128((__m128i*) c10, vfp16out10x01234567);
477 c10 = (uint16_t*) ((uintptr_t) c10 + cn_stride);
478 _mm_storeu_si128((__m128i*) c9, vfp16out9x01234567);
479 c9 = (uint16_t*) ((uintptr_t) c9 + cn_stride);
480 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
481 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
482 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
483 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
484 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
485 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
486 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
487 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
488 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
489 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
490 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
491 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
492 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
493 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
494 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
495 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
496 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
497 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
498
499 a = (const int8_t**restrict) ((uintptr_t) a - ks);
500 nc -= 8;
501 } else {
502 // Prepare mask for valid 16-bit elements (depends on nc).
503 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
504 _mm_mask_storeu_epi16(c11, vmask, vfp16out11x01234567);
505 _mm_mask_storeu_epi16(c10, vmask, vfp16out10x01234567);
506 _mm_mask_storeu_epi16(c9, vmask, vfp16out9x01234567);
507 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
508 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
509 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
510 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
511 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
512 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
513 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
514 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
515 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
516 nc = 0;
517 }
518 } while (nc != 0);
519 }