Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-12x8c8-minmax-fp32-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 12);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr < 6) {
65 c5 = c4;
66 }
67 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 6) {
69 c6 = c5;
70 }
71 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
72 if XNN_UNPREDICTABLE(mr < 8) {
73 c7 = c6;
74 }
75 int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
76 if XNN_UNPREDICTABLE(mr <= 8) {
77 c8 = c7;
78 }
79 int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
80 if XNN_UNPREDICTABLE(mr < 10) {
81 c9 = c8;
82 }
83 int8_t* c10 = (int8_t*) ((uintptr_t) c9 + cm_stride);
84 if XNN_UNPREDICTABLE(mr <= 10) {
85 c10 = c9;
86 }
87 int8_t* c11 = (int8_t*) ((uintptr_t) c10 + cm_stride);
88 if XNN_UNPREDICTABLE(mr != 12) {
89 c11 = c10;
90 }
91
92 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
93 XNN_FORCE_REALIZATION(vsign_mask);
94 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
95 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
96 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
97 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
98 // XNN_FORCE_REALIZATION(voutput_zero_point);
99 // XNN_FORCE_REALIZATION(voutput_min);
100 do {
101 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
102 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
103 __m256i vacc1x0123 = vacc0x0123;
104 __m256i vacc1x4567 = vacc0x4567;
105 __m256i vacc2x0123 = vacc0x0123;
106 __m256i vacc2x4567 = vacc0x4567;
107 __m256i vacc3x0123 = vacc0x0123;
108 __m256i vacc3x4567 = vacc0x4567;
109 __m256i vacc4x0123 = vacc0x0123;
110 __m256i vacc4x4567 = vacc0x4567;
111 __m256i vacc5x0123 = vacc0x0123;
112 __m256i vacc5x4567 = vacc0x4567;
113 __m256i vacc6x0123 = vacc0x0123;
114 __m256i vacc6x4567 = vacc0x4567;
115 __m256i vacc7x0123 = vacc0x0123;
116 __m256i vacc7x4567 = vacc0x4567;
117 __m256i vacc8x0123 = vacc0x0123;
118 __m256i vacc8x4567 = vacc0x4567;
119 __m256i vacc9x0123 = vacc0x0123;
120 __m256i vacc9x4567 = vacc0x4567;
121 __m256i vacc10x0123 = vacc0x0123;
122 __m256i vacc10x4567 = vacc0x4567;
123 __m256i vacc11x0123 = vacc0x0123;
124 __m256i vacc11x4567 = vacc0x4567;
125 w = (const int32_t*) w + 8;
126
127 size_t p = ks;
128 do {
129 const int8_t* restrict a0 = a[0];
130 if XNN_UNPREDICTABLE(a0 != zero) {
131 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
132 }
133 const int8_t* restrict a1 = a[1];
134 if XNN_UNPREDICTABLE(a1 != zero) {
135 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
136 }
137 const int8_t* restrict a2 = a[2];
138 if XNN_UNPREDICTABLE(a2 != zero) {
139 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
140 }
141 const int8_t* restrict a3 = a[3];
142 if XNN_UNPREDICTABLE(a3 != zero) {
143 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
144 }
145 const int8_t* restrict a4 = a[4];
146 if XNN_UNPREDICTABLE(a4 != zero) {
147 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
148 }
149 const int8_t* restrict a5 = a[5];
150 if XNN_UNPREDICTABLE(a5 != zero) {
151 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
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 }
157 const int8_t* restrict a7 = a[7];
158 if XNN_UNPREDICTABLE(a7 != zero) {
159 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
160 }
161 const int8_t* restrict a8 = a[8];
162 if XNN_UNPREDICTABLE(a8 != zero) {
163 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
164 }
165 const int8_t* restrict a9 = a[9];
166 if XNN_UNPREDICTABLE(a9 != zero) {
167 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
168 }
169 const int8_t* restrict a10 = a[10];
170 if XNN_UNPREDICTABLE(a10 != zero) {
171 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
172 }
173 const int8_t* restrict a11 = a[11];
174 if XNN_UNPREDICTABLE(a11 != zero) {
175 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
176 }
177 a += 12;
178
179 size_t k = kc;
180 while (k >= 16 * sizeof(int8_t)) {
181 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
182 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
183 a0 += 16;
184 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
185 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
186 a1 += 16;
187 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
188 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
189 a2 += 16;
190 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
191 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
192 a3 += 16;
193 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
194 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
195 a4 += 16;
196 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
197 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
198 a5 += 16;
199 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
200 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
201 a6 += 16;
202 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
203 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
204 a7 += 16;
205 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
206 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
207 a8 += 16;
208 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
209 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
210 a9 += 16;
211 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
212 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
213 a10 += 16;
214 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
215 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
216 a11 += 16;
217
218 const __m256i vb01234567x0123 = _mm256_load_si256(w);
219 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
220 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
221 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
222 xnn_prefetch_to_l1((const int8_t*) w + 896);
223
224 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
225 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
226 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
227 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
228 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
229 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
230 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
231 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
232 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
233 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
234 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
235 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
236 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
237 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
238 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
239 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
240 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
241 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
242 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
243 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
244 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
245 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
246 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
247 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
248 xnn_prefetch_to_l1((const int8_t*) w + 960);
249 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
250 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
251 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
252 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
253 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
254 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
255 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
256 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
257 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
258 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
259 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
260 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
261 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
262 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
263 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
264 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
265 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
266 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
267 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
268 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
269 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
270 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
271 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
272 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
273
274 w = (const int8_t*) w + 128;
275 k -= 16 * sizeof(int8_t);
276 }
277
278 if (k != 0) {
279 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
280 a0 += 8;
281 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
282 a1 += 8;
283 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
284 a2 += 8;
285 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
286 a3 += 8;
287 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
288 a4 += 8;
289 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
290 a5 += 8;
291 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
292 a6 += 8;
293 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
294 a7 += 8;
295 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
296 a8 += 8;
297 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
298 a9 += 8;
299 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
300 a10 += 8;
301 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
302 a11 += 8;
303
304 const __m256i vb01234567x0123 = _mm256_load_si256(w);
305 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
306
307 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
308 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
309 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
310 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
311 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
312 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
313 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
314 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
315 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
316 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
317 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
318 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
319 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
320 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
321 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
322 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
323 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
324 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
325 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
326 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
327 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
328 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
329 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
330 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
331 xnn_prefetch_to_l1((const int8_t*) w + 960);
332
333 w = (const int8_t*) w + 64;
334 k -= 8 * sizeof(int8_t);
335 }
336
337 p -= 12 * sizeof(void*);
338 } while (p != 0);
339
340
341 // Add adjacent pairs
342 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
343 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
344 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
345 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
346 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
347 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
348 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
349 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
350 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
351 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
352 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
353 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
354 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
355 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
356 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
357 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
358 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
359 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
360 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
361 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
362 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
363 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
364 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
365 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
366
367 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
368 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
369 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
370 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
371 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
372 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
373 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
374 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
375 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
376 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
377 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
378 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
379
380 const __m256 vscale01234567 = _mm256_load_ps(w);
381 w = (const float*) w + 8;
382 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
383 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
384 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
385 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
386 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
387 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
388 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
389 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vscale01234567);
390 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vscale01234567);
391 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vscale01234567);
392 vout10x01234567 = _mm256_mul_ps(vout10x01234567, vscale01234567);
393 vout11x01234567 = _mm256_mul_ps(vout11x01234567, vscale01234567);
394
395 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
396 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
397 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
398 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
399 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
400 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
401 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
402 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max_less_zero_point);
403 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max_less_zero_point);
404 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max_less_zero_point);
405 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max_less_zero_point);
406 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max_less_zero_point);
407
408 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
409 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
410 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
411 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
412 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
413 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
414 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
415 vacc7x01234567 = _mm256_cvtps_epi32(vout7x01234567);
416 vacc8x01234567 = _mm256_cvtps_epi32(vout8x01234567);
417 vacc9x01234567 = _mm256_cvtps_epi32(vout9x01234567);
418 vacc10x01234567 = _mm256_cvtps_epi32(vout10x01234567);
419 vacc11x01234567 = _mm256_cvtps_epi32(vout11x01234567);
420
421 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
422 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
423 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
424 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
425 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
426 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
427 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
428 vacc7x01234567 = _mm256_add_epi32(vacc7x01234567, voutput_zero_point);
429 vacc8x01234567 = _mm256_add_epi32(vacc8x01234567, voutput_zero_point);
430 vacc9x01234567 = _mm256_add_epi32(vacc9x01234567, voutput_zero_point);
431 vacc10x01234567 = _mm256_add_epi32(vacc10x01234567, voutput_zero_point);
432 vacc11x01234567 = _mm256_add_epi32(vacc11x01234567, voutput_zero_point);
433
434 __m128i voutb0x01234567 = _mm256_cvtsepi32_epi8(vacc0x01234567);
435 __m128i voutb1x01234567 = _mm256_cvtsepi32_epi8(vacc1x01234567);
436 __m128i voutb2x01234567 = _mm256_cvtsepi32_epi8(vacc2x01234567);
437 __m128i voutb3x01234567 = _mm256_cvtsepi32_epi8(vacc3x01234567);
438 __m128i voutb4x01234567 = _mm256_cvtsepi32_epi8(vacc4x01234567);
439 __m128i voutb5x01234567 = _mm256_cvtsepi32_epi8(vacc5x01234567);
440 __m128i voutb6x01234567 = _mm256_cvtsepi32_epi8(vacc6x01234567);
441 __m128i voutb7x01234567 = _mm256_cvtsepi32_epi8(vacc7x01234567);
442 __m128i voutb8x01234567 = _mm256_cvtsepi32_epi8(vacc8x01234567);
443 __m128i voutb9x01234567 = _mm256_cvtsepi32_epi8(vacc9x01234567);
444 __m128i voutb10x01234567 = _mm256_cvtsepi32_epi8(vacc10x01234567);
445 __m128i voutb11x01234567 = _mm256_cvtsepi32_epi8(vacc11x01234567);
446
447 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
448 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
449 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
450 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
451 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
452 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
453 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
454 voutb7x01234567 = _mm_max_epi8(voutb7x01234567, voutput_min);
455 voutb8x01234567 = _mm_max_epi8(voutb8x01234567, voutput_min);
456 voutb9x01234567 = _mm_max_epi8(voutb9x01234567, voutput_min);
457 voutb10x01234567 = _mm_max_epi8(voutb10x01234567, voutput_min);
458 voutb11x01234567 = _mm_max_epi8(voutb11x01234567, voutput_min);
459
460 if (nc >= 8) {
461 _mm_storel_epi64((__m128i*) c11, voutb11x01234567);
462 c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
463 _mm_storel_epi64((__m128i*) c10, voutb10x01234567);
464 c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
465 _mm_storel_epi64((__m128i*) c9, voutb9x01234567);
466 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
467 _mm_storel_epi64((__m128i*) c8, voutb8x01234567);
468 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
469 _mm_storel_epi64((__m128i*) c7, voutb7x01234567);
470 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
471 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
472 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
473 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
474 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
475 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
476 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
477 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
478 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
479 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
480 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
481 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
482 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
483 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
484 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
485 a = (const int8_t**restrict) ((uintptr_t) a - ks);
486 nc -= 8;
487 } else {
488 // Prepare mask for valid 8-bit elements (depends on nc).
489 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
490 _mm_mask_storeu_epi8(c11, vmask, voutb11x01234567);
491 _mm_mask_storeu_epi8(c10, vmask, voutb10x01234567);
492 _mm_mask_storeu_epi8(c9, vmask, voutb9x01234567);
493 _mm_mask_storeu_epi8(c8, vmask, voutb8x01234567);
494 _mm_mask_storeu_epi8(c7, vmask, voutb7x01234567);
495 _mm_mask_storeu_epi8(c6, vmask, voutb6x01234567);
496 _mm_mask_storeu_epi8(c5, vmask, voutb5x01234567);
497 _mm_mask_storeu_epi8(c4, vmask, voutb4x01234567);
498 _mm_mask_storeu_epi8(c3, vmask, voutb3x01234567);
499 _mm_mask_storeu_epi8(c2, vmask, voutb2x01234567);
500 _mm_mask_storeu_epi8(c1, vmask, voutb1x01234567);
501 _mm_mask_storeu_epi8(c0, vmask, voutb0x01234567);
502 nc = 0;
503 }
504 } while (nc != 0);
505 }