Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-12x8c8-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_12x8c8__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 <= 12);
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 uint16_t* c10 = (uint16_t*) ((uintptr_t) c9 + cm_stride);
86 if XNN_UNPREDICTABLE(mr <= 10) {
87 c10 = c9;
88 }
89 uint16_t* c11 = (uint16_t*) ((uintptr_t) c10 + cm_stride);
90 if XNN_UNPREDICTABLE(mr != 12) {
91 c11 = c10;
92 }
93
94 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
95 XNN_FORCE_REALIZATION(vsign_mask);
96 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
97 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
98 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
99 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
100 // XNN_FORCE_REALIZATION(vinput_zero_point);
101 // XNN_FORCE_REALIZATION(vinput_inv_scale);
102 // XNN_FORCE_REALIZATION(voutput_min);
103 // XNN_FORCE_REALIZATION(voutput_max);
104 do {
105 const __m256i vksum01234567 = _mm256_load_si256(w);
106 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
107 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
108 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
109 __m256i vacc1x0123 = vacc0x0123;
110 __m256i vacc1x4567 = vacc0x4567;
111 __m256i vacc2x0123 = vacc0x0123;
112 __m256i vacc2x4567 = vacc0x4567;
113 __m256i vacc3x0123 = vacc0x0123;
114 __m256i vacc3x4567 = vacc0x4567;
115 __m256i vacc4x0123 = vacc0x0123;
116 __m256i vacc4x4567 = vacc0x4567;
117 __m256i vacc5x0123 = vacc0x0123;
118 __m256i vacc5x4567 = vacc0x4567;
119 __m256i vacc6x0123 = vacc0x0123;
120 __m256i vacc6x4567 = vacc0x4567;
121 __m256i vacc7x0123 = vacc0x0123;
122 __m256i vacc7x4567 = vacc0x4567;
123 __m256i vacc8x0123 = vacc0x0123;
124 __m256i vacc8x4567 = vacc0x4567;
125 __m256i vacc9x0123 = vacc0x0123;
126 __m256i vacc9x4567 = vacc0x4567;
127 __m256i vacc10x0123 = vacc0x0123;
128 __m256i vacc10x4567 = vacc0x4567;
129 __m256i vacc11x0123 = vacc0x0123;
130 __m256i vacc11x4567 = vacc0x4567;
131 w = (const int32_t*) w + 8;
132
133 size_t p = ks;
134 do {
135 const int8_t* restrict a0 = a[0];
136 if XNN_UNPREDICTABLE(a0 != zero) {
137 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
138 } else {
139 a0 = zero_data;
140 }
141 const int8_t* restrict a1 = a[1];
142 if XNN_UNPREDICTABLE(a1 != zero) {
143 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
144 } else {
145 a1 = zero_data;
146 }
147 const int8_t* restrict a2 = a[2];
148 if XNN_UNPREDICTABLE(a2 != zero) {
149 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
150 } else {
151 a2 = zero_data;
152 }
153 const int8_t* restrict a3 = a[3];
154 if XNN_UNPREDICTABLE(a3 != zero) {
155 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
156 } else {
157 a3 = zero_data;
158 }
159 const int8_t* restrict a4 = a[4];
160 if XNN_UNPREDICTABLE(a4 != zero) {
161 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
162 } else {
163 a4 = zero_data;
164 }
165 const int8_t* restrict a5 = a[5];
166 if XNN_UNPREDICTABLE(a5 != zero) {
167 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
168 } else {
169 a5 = zero_data;
170 }
171 const int8_t* restrict a6 = a[6];
172 if XNN_UNPREDICTABLE(a6 != zero) {
173 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
174 } else {
175 a6 = zero_data;
176 }
177 const int8_t* restrict a7 = a[7];
178 if XNN_UNPREDICTABLE(a7 != zero) {
179 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
180 } else {
181 a7 = zero_data;
182 }
183 const int8_t* restrict a8 = a[8];
184 if XNN_UNPREDICTABLE(a8 != zero) {
185 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
186 } else {
187 a8 = zero_data;
188 }
189 const int8_t* restrict a9 = a[9];
190 if XNN_UNPREDICTABLE(a9 != zero) {
191 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
192 } else {
193 a9 = zero_data;
194 }
195 const int8_t* restrict a10 = a[10];
196 if XNN_UNPREDICTABLE(a10 != zero) {
197 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
198 } else {
199 a10 = zero_data;
200 }
201 const int8_t* restrict a11 = a[11];
202 if XNN_UNPREDICTABLE(a11 != zero) {
203 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
204 } else {
205 a11 = zero_data;
206 }
207 a += 12;
208
209 size_t k = kc;
210 while (k >= 16 * sizeof(int8_t)) {
211 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
212 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
213 a0 += 16;
214 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
215 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
216 a1 += 16;
217 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
218 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
219 a2 += 16;
220 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
221 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
222 a3 += 16;
223 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
224 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
225 a4 += 16;
226 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
227 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
228 a5 += 16;
229 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
230 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
231 a6 += 16;
232 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
233 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
234 a7 += 16;
235 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
236 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
237 a8 += 16;
238 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
239 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
240 a9 += 16;
241 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
242 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
243 a10 += 16;
244 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
245 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
246 a11 += 16;
247
248 const __m256i vb01234567x0123 = _mm256_load_si256(w);
249 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
250 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
251 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
252 xnn_prefetch_to_l1((const int8_t*) w + 896);
253
254 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
255 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
256 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
257 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
258 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
259 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
260 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
261 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
262 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
263 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
264 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
265 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
266 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
267 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
268 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
269 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
270 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
271 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
272 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
273 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
274 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
275 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
276 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
277 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
278 xnn_prefetch_to_l1((const int8_t*) w + 960);
279 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
280 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
281 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
282 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
283 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
284 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
285 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
286 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
287 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
288 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
289 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
290 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
291 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
292 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
293 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
294 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
295 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
296 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
297 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
298 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
299 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
300 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
301 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
302 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
303
304 w = (const int8_t*) w + 128;
305 k -= 16 * sizeof(int8_t);
306 }
307
308 if (k != 0) {
309 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
310 a0 += 8;
311 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
312 a1 += 8;
313 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
314 a2 += 8;
315 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
316 a3 += 8;
317 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
318 a4 += 8;
319 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
320 a5 += 8;
321 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
322 a6 += 8;
323 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
324 a7 += 8;
325 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
326 a8 += 8;
327 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
328 a9 += 8;
329 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
330 a10 += 8;
331 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
332 a11 += 8;
333
334 const __m256i vb01234567x0123 = _mm256_load_si256(w);
335 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
336
337 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
338 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
339 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
340 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
341 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
342 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
343 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
344 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
345 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
346 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
347 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
348 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
349 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
350 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
351 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
352 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
353 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
354 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
355 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
356 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
357 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
358 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
359 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
360 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
361 xnn_prefetch_to_l1((const int8_t*) w + 960);
362
363 w = (const int8_t*) w + 64;
364 k -= 8 * sizeof(int8_t);
365 }
366
367 p -= 12 * sizeof(void*);
368 } while (p != 0);
369
370
371 // Add adjacent pairs
372 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
373 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
374 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
375 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
376 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
377 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
378 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
379 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
380 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
381 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
382 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
383 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
384 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
385 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
386 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
387 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
388 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
389 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
390 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
391 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
392 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
393 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
394 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
395 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
396
397 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
398 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
399 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
400 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
401 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
402 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
403 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
404 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
405 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
406 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
407 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
408 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
409
410 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
411 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
412 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
413 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
414 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
415 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
416 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
417 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_inv_scale);
418 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vinput_inv_scale);
419 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vinput_inv_scale);
420 vout10x01234567 = _mm256_mul_ps(vout10x01234567, vinput_inv_scale);
421 vout11x01234567 = _mm256_mul_ps(vout11x01234567, vinput_inv_scale);
422
423 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
424 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
425 w = (const float*) w + 16;
426 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
427 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
428 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
429 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
430 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
431 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
432 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
433 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
434 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
435 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
436 vout10x01234567 = _mm256_fmadd_ps(vout10x01234567, vfilter_output_scale01234567, vbias01234567);
437 vout11x01234567 = _mm256_fmadd_ps(vout11x01234567, vfilter_output_scale01234567, vbias01234567);
438
439 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
440 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
441 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
442 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
443 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
444 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
445 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
446 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
447 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
448 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
449 vout10x01234567 = _mm256_max_ps(vout10x01234567, voutput_min);
450 vout11x01234567 = _mm256_max_ps(vout11x01234567, voutput_min);
451
452 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
453 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
454 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
455 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
456 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
457 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
458 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
459 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
460 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
461 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
462 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max);
463 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max);
464
465 __m128i vfp16out11x01234567 = _mm256_cvtps_ph(vout11x01234567, _MM_FROUND_TO_NEAREST_INT);
466 __m128i vfp16out10x01234567 = _mm256_cvtps_ph(vout10x01234567, _MM_FROUND_TO_NEAREST_INT);
467 __m128i vfp16out9x01234567 = _mm256_cvtps_ph(vout9x01234567, _MM_FROUND_TO_NEAREST_INT);
468 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
469 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
470 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
471 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
472 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
473 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
474 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
475 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
476 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
477 if XNN_LIKELY(nc >= 8) {
478 _mm_storeu_si128((__m128i*) c11, vfp16out11x01234567);
479 c11 = (uint16_t*) ((uintptr_t) c11 + cn_stride);
480 _mm_storeu_si128((__m128i*) c10, vfp16out10x01234567);
481 c10 = (uint16_t*) ((uintptr_t) c10 + cn_stride);
482 _mm_storeu_si128((__m128i*) c9, vfp16out9x01234567);
483 c9 = (uint16_t*) ((uintptr_t) c9 + cn_stride);
484 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
485 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
486 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
487 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
488 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
489 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
490 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
491 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
492 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
493 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
494 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
495 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
496 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
497 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
498 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
499 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
500 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
501 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
502
503 a = (const int8_t**restrict) ((uintptr_t) a - ks);
504 nc -= 8;
505 } else {
506 // Prepare mask for valid 16-bit elements (depends on nc).
507 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
508 _mm_mask_storeu_epi16(c11, vmask, vfp16out11x01234567);
509 _mm_mask_storeu_epi16(c10, vmask, vfp16out10x01234567);
510 _mm_mask_storeu_epi16(c9, vmask, vfp16out9x01234567);
511 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
512 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
513 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
514 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
515 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
516 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
517 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
518 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
519 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
520 nc = 0;
521 }
522 } while (nc != 0);
523 }