Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-14x8c8-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_14x8c8__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 <= 14);
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 uint16_t* c12 = (uint16_t*) ((uintptr_t) c11 + cm_stride);
93 if XNN_UNPREDICTABLE(mr <= 12) {
94 c12 = c11;
95 }
96 uint16_t* c13 = (uint16_t*) ((uintptr_t) c12 + cm_stride);
97 if XNN_UNPREDICTABLE(mr != 14) {
98 c13 = c12;
99 }
100
101 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
102 XNN_FORCE_REALIZATION(vsign_mask);
103 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
104 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
105 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
106 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
107 // XNN_FORCE_REALIZATION(vinput_zero_point);
108 // XNN_FORCE_REALIZATION(vinput_inv_scale);
109 // XNN_FORCE_REALIZATION(voutput_min);
110 // XNN_FORCE_REALIZATION(voutput_max);
111 do {
112 const __m256i vksum01234567 = _mm256_load_si256(w);
113 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
114 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
115 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
116 __m256i vacc1x0123 = vacc0x0123;
117 __m256i vacc1x4567 = vacc0x4567;
118 __m256i vacc2x0123 = vacc0x0123;
119 __m256i vacc2x4567 = vacc0x4567;
120 __m256i vacc3x0123 = vacc0x0123;
121 __m256i vacc3x4567 = vacc0x4567;
122 __m256i vacc4x0123 = vacc0x0123;
123 __m256i vacc4x4567 = vacc0x4567;
124 __m256i vacc5x0123 = vacc0x0123;
125 __m256i vacc5x4567 = vacc0x4567;
126 __m256i vacc6x0123 = vacc0x0123;
127 __m256i vacc6x4567 = vacc0x4567;
128 __m256i vacc7x0123 = vacc0x0123;
129 __m256i vacc7x4567 = vacc0x4567;
130 __m256i vacc8x0123 = vacc0x0123;
131 __m256i vacc8x4567 = vacc0x4567;
132 __m256i vacc9x0123 = vacc0x0123;
133 __m256i vacc9x4567 = vacc0x4567;
134 __m256i vacc10x0123 = vacc0x0123;
135 __m256i vacc10x4567 = vacc0x4567;
136 __m256i vacc11x0123 = vacc0x0123;
137 __m256i vacc11x4567 = vacc0x4567;
138 __m256i vacc12x0123 = vacc0x0123;
139 __m256i vacc12x4567 = vacc0x4567;
140 __m256i vacc13x0123 = vacc0x0123;
141 __m256i vacc13x4567 = vacc0x4567;
142 w = (const int32_t*) w + 8;
143
144 size_t p = ks;
145 do {
146 const int8_t* restrict a0 = a[0];
147 if XNN_UNPREDICTABLE(a0 != zero) {
148 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
149 } else {
150 a0 = zero_data;
151 }
152 const int8_t* restrict a1 = a[1];
153 if XNN_UNPREDICTABLE(a1 != zero) {
154 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
155 } else {
156 a1 = zero_data;
157 }
158 const int8_t* restrict a2 = a[2];
159 if XNN_UNPREDICTABLE(a2 != zero) {
160 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
161 } else {
162 a2 = zero_data;
163 }
164 const int8_t* restrict a3 = a[3];
165 if XNN_UNPREDICTABLE(a3 != zero) {
166 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
167 } else {
168 a3 = zero_data;
169 }
170 const int8_t* restrict a4 = a[4];
171 if XNN_UNPREDICTABLE(a4 != zero) {
172 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
173 } else {
174 a4 = zero_data;
175 }
176 const int8_t* restrict a5 = a[5];
177 if XNN_UNPREDICTABLE(a5 != zero) {
178 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
179 } else {
180 a5 = zero_data;
181 }
182 const int8_t* restrict a6 = a[6];
183 if XNN_UNPREDICTABLE(a6 != zero) {
184 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
185 } else {
186 a6 = zero_data;
187 }
188 const int8_t* restrict a7 = a[7];
189 if XNN_UNPREDICTABLE(a7 != zero) {
190 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
191 } else {
192 a7 = zero_data;
193 }
194 const int8_t* restrict a8 = a[8];
195 if XNN_UNPREDICTABLE(a8 != zero) {
196 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
197 } else {
198 a8 = zero_data;
199 }
200 const int8_t* restrict a9 = a[9];
201 if XNN_UNPREDICTABLE(a9 != zero) {
202 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
203 } else {
204 a9 = zero_data;
205 }
206 const int8_t* restrict a10 = a[10];
207 if XNN_UNPREDICTABLE(a10 != zero) {
208 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
209 } else {
210 a10 = zero_data;
211 }
212 const int8_t* restrict a11 = a[11];
213 if XNN_UNPREDICTABLE(a11 != zero) {
214 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
215 } else {
216 a11 = zero_data;
217 }
218 const int8_t* restrict a12 = a[12];
219 if XNN_UNPREDICTABLE(a12 != zero) {
220 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
221 } else {
222 a12 = zero_data;
223 }
224 const int8_t* restrict a13 = a[13];
225 if XNN_UNPREDICTABLE(a13 != zero) {
226 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
227 } else {
228 a13 = zero_data;
229 }
230 a += 14;
231
232 size_t k = kc;
233 while (k >= 16 * sizeof(int8_t)) {
234 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
235 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
236 a0 += 16;
237 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
238 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
239 a1 += 16;
240 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
241 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
242 a2 += 16;
243 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
244 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
245 a3 += 16;
246 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
247 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
248 a4 += 16;
249 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
250 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
251 a5 += 16;
252 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
253 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
254 a6 += 16;
255 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
256 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
257 a7 += 16;
258 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
259 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
260 a8 += 16;
261 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
262 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
263 a9 += 16;
264 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
265 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
266 a10 += 16;
267 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
268 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
269 a11 += 16;
270 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
271 const __m256i va12x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12 + 8)), vsign_mask);
272 a12 += 16;
273 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
274 const __m256i va13x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13 + 8)), vsign_mask);
275 a13 += 16;
276
277 const __m256i vb01234567x0123 = _mm256_load_si256(w);
278 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
279 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
280 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
281
282 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
283 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
284 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
285 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
286 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
287 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
288 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
289 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
290 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
291 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
292 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
293 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
294 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
295 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
296 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
297 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
298 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
299 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
300 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
301 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
302 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
303 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
304 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
305 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
306 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
307 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
308 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
309 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
310 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
311 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
312 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
313 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
314 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
315 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
316 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
317 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
318 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
319 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
320 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
321 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
322 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
323 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
324 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
325 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
326 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
327 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
328 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
329 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
330 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
331 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
332 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
333 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
334 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x89ABCDEF, vb01234567x4567);
335 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x89ABCDEF, vb89ABCDEFx4567);
336 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x89ABCDEF, vb01234567x4567);
337 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x89ABCDEF, vb89ABCDEFx4567);
338
339 w = (const int8_t*) w + 128;
340 k -= 16 * sizeof(int8_t);
341 }
342
343 if (k != 0) {
344 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
345 a0 += 8;
346 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
347 a1 += 8;
348 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
349 a2 += 8;
350 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
351 a3 += 8;
352 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
353 a4 += 8;
354 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
355 a5 += 8;
356 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
357 a6 += 8;
358 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
359 a7 += 8;
360 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
361 a8 += 8;
362 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
363 a9 += 8;
364 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
365 a10 += 8;
366 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
367 a11 += 8;
368 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
369 a12 += 8;
370 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
371 a13 += 8;
372
373 const __m256i vb01234567x0123 = _mm256_load_si256(w);
374 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
375
376 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
377 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
378 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
379 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
380 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
381 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
382 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
383 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
384 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
385 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
386 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
387 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
388 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
389 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
390 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
391 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
392 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
393 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
394 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
395 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
396 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
397 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
398 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
399 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
400 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
401 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
402 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
403 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
404
405 w = (const int8_t*) w + 64;
406 k -= 8 * sizeof(int8_t);
407 }
408
409 p -= 14 * sizeof(void*);
410 } while (p != 0);
411
412
413 // Add adjacent pairs
414 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
415 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
416 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
417 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
418 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
419 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
420 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
421 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
422 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
423 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
424 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
425 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
426 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
427 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
428 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
429 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
430 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
431 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
432 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
433 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
434 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
435 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
436 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
437 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
438 const __m256i vsum12x02134657 = _mm256_hadd_epi32(vacc12x0123, vacc12x4567);
439 __m256i vacc12x01234567 = _mm256_permute4x64_epi64(vsum12x02134657, _MM_SHUFFLE(3, 1, 2, 0));
440 const __m256i vsum13x02134657 = _mm256_hadd_epi32(vacc13x0123, vacc13x4567);
441 __m256i vacc13x01234567 = _mm256_permute4x64_epi64(vsum13x02134657, _MM_SHUFFLE(3, 1, 2, 0));
442
443 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
444 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
445 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
446 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
447 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
448 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
449 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
450 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
451 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
452 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
453 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
454 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
455 __m256 vout12x01234567 = _mm256_cvtepi32_ps(vacc12x01234567);
456 __m256 vout13x01234567 = _mm256_cvtepi32_ps(vacc13x01234567);
457
458 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
459 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
460 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
461 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
462 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
463 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
464 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
465 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_inv_scale);
466 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vinput_inv_scale);
467 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vinput_inv_scale);
468 vout10x01234567 = _mm256_mul_ps(vout10x01234567, vinput_inv_scale);
469 vout11x01234567 = _mm256_mul_ps(vout11x01234567, vinput_inv_scale);
470 vout12x01234567 = _mm256_mul_ps(vout12x01234567, vinput_inv_scale);
471 vout13x01234567 = _mm256_mul_ps(vout13x01234567, vinput_inv_scale);
472
473 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
474 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
475 w = (const float*) w + 16;
476 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
477 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
478 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
479 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
480 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
481 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
482 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
483 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
484 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
485 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
486 vout10x01234567 = _mm256_fmadd_ps(vout10x01234567, vfilter_output_scale01234567, vbias01234567);
487 vout11x01234567 = _mm256_fmadd_ps(vout11x01234567, vfilter_output_scale01234567, vbias01234567);
488 vout12x01234567 = _mm256_fmadd_ps(vout12x01234567, vfilter_output_scale01234567, vbias01234567);
489 vout13x01234567 = _mm256_fmadd_ps(vout13x01234567, vfilter_output_scale01234567, vbias01234567);
490
491 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
492 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
493 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
494 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
495 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
496 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
497 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
498 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
499 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
500 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
501 vout10x01234567 = _mm256_max_ps(vout10x01234567, voutput_min);
502 vout11x01234567 = _mm256_max_ps(vout11x01234567, voutput_min);
503 vout12x01234567 = _mm256_max_ps(vout12x01234567, voutput_min);
504 vout13x01234567 = _mm256_max_ps(vout13x01234567, voutput_min);
505
506 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
507 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
508 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
509 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
510 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
511 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
512 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
513 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
514 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
515 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
516 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max);
517 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max);
518 vout12x01234567 = _mm256_min_ps(vout12x01234567, voutput_max);
519 vout13x01234567 = _mm256_min_ps(vout13x01234567, voutput_max);
520
521 __m128i vfp16out13x01234567 = _mm256_cvtps_ph(vout13x01234567, _MM_FROUND_TO_NEAREST_INT);
522 __m128i vfp16out12x01234567 = _mm256_cvtps_ph(vout12x01234567, _MM_FROUND_TO_NEAREST_INT);
523 __m128i vfp16out11x01234567 = _mm256_cvtps_ph(vout11x01234567, _MM_FROUND_TO_NEAREST_INT);
524 __m128i vfp16out10x01234567 = _mm256_cvtps_ph(vout10x01234567, _MM_FROUND_TO_NEAREST_INT);
525 __m128i vfp16out9x01234567 = _mm256_cvtps_ph(vout9x01234567, _MM_FROUND_TO_NEAREST_INT);
526 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
527 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
528 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
529 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
530 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
531 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
532 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
533 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
534 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
535 if XNN_LIKELY(nc >= 8) {
536 _mm_storeu_si128((__m128i*) c13, vfp16out13x01234567);
537 c13 = (uint16_t*) ((uintptr_t) c13 + cn_stride);
538 _mm_storeu_si128((__m128i*) c12, vfp16out12x01234567);
539 c12 = (uint16_t*) ((uintptr_t) c12 + cn_stride);
540 _mm_storeu_si128((__m128i*) c11, vfp16out11x01234567);
541 c11 = (uint16_t*) ((uintptr_t) c11 + cn_stride);
542 _mm_storeu_si128((__m128i*) c10, vfp16out10x01234567);
543 c10 = (uint16_t*) ((uintptr_t) c10 + cn_stride);
544 _mm_storeu_si128((__m128i*) c9, vfp16out9x01234567);
545 c9 = (uint16_t*) ((uintptr_t) c9 + cn_stride);
546 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
547 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
548 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
549 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
550 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
551 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
552 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
553 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
554 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
555 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
556 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
557 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
558 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
559 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
560 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
561 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
562 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
563 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
564
565 a = (const int8_t**restrict) ((uintptr_t) a - ks);
566 nc -= 8;
567 } else {
568 // Prepare mask for valid 16-bit elements (depends on nc).
569 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
570 _mm_mask_storeu_epi16(c13, vmask, vfp16out13x01234567);
571 _mm_mask_storeu_epi16(c12, vmask, vfp16out12x01234567);
572 _mm_mask_storeu_epi16(c11, vmask, vfp16out11x01234567);
573 _mm_mask_storeu_epi16(c10, vmask, vfp16out10x01234567);
574 _mm_mask_storeu_epi16(c9, vmask, vfp16out9x01234567);
575 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
576 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
577 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
578 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
579 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
580 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
581 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
582 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
583 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
584 nc = 0;
585 }
586 } while (nc != 0);
587 }