Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-14x8c8-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_14x8c8__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 <= 14);
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 int8_t* c12 = (int8_t*) ((uintptr_t) c11 + cm_stride);
92 if XNN_UNPREDICTABLE(mr <= 12) {
93 c12 = c11;
94 }
95 int8_t* c13 = (int8_t*) ((uintptr_t) c12 + cm_stride);
96 if XNN_UNPREDICTABLE(mr != 14) {
97 c13 = c12;
98 }
99
100 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
101 XNN_FORCE_REALIZATION(vsign_mask);
102 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);
103 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
104 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
105 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
106 // XNN_FORCE_REALIZATION(voutput_zero_point);
107 // XNN_FORCE_REALIZATION(voutput_min);
108 do {
109 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
110 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
111 __m256i vacc1x0123 = vacc0x0123;
112 __m256i vacc1x4567 = vacc0x4567;
113 __m256i vacc2x0123 = vacc0x0123;
114 __m256i vacc2x4567 = vacc0x4567;
115 __m256i vacc3x0123 = vacc0x0123;
116 __m256i vacc3x4567 = vacc0x4567;
117 __m256i vacc4x0123 = vacc0x0123;
118 __m256i vacc4x4567 = vacc0x4567;
119 __m256i vacc5x0123 = vacc0x0123;
120 __m256i vacc5x4567 = vacc0x4567;
121 __m256i vacc6x0123 = vacc0x0123;
122 __m256i vacc6x4567 = vacc0x4567;
123 __m256i vacc7x0123 = vacc0x0123;
124 __m256i vacc7x4567 = vacc0x4567;
125 __m256i vacc8x0123 = vacc0x0123;
126 __m256i vacc8x4567 = vacc0x4567;
127 __m256i vacc9x0123 = vacc0x0123;
128 __m256i vacc9x4567 = vacc0x4567;
129 __m256i vacc10x0123 = vacc0x0123;
130 __m256i vacc10x4567 = vacc0x4567;
131 __m256i vacc11x0123 = vacc0x0123;
132 __m256i vacc11x4567 = vacc0x4567;
133 __m256i vacc12x0123 = vacc0x0123;
134 __m256i vacc12x4567 = vacc0x4567;
135 __m256i vacc13x0123 = vacc0x0123;
136 __m256i vacc13x4567 = vacc0x4567;
137 w = (const int32_t*) w + 8;
138
139 size_t p = ks;
140 do {
141 const int8_t* restrict a0 = a[0];
142 if XNN_UNPREDICTABLE(a0 != zero) {
143 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
144 }
145 const int8_t* restrict a1 = a[1];
146 if XNN_UNPREDICTABLE(a1 != zero) {
147 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
148 }
149 const int8_t* restrict a2 = a[2];
150 if XNN_UNPREDICTABLE(a2 != zero) {
151 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
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 }
157 const int8_t* restrict a4 = a[4];
158 if XNN_UNPREDICTABLE(a4 != zero) {
159 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
160 }
161 const int8_t* restrict a5 = a[5];
162 if XNN_UNPREDICTABLE(a5 != zero) {
163 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
164 }
165 const int8_t* restrict a6 = a[6];
166 if XNN_UNPREDICTABLE(a6 != zero) {
167 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
168 }
169 const int8_t* restrict a7 = a[7];
170 if XNN_UNPREDICTABLE(a7 != zero) {
171 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
172 }
173 const int8_t* restrict a8 = a[8];
174 if XNN_UNPREDICTABLE(a8 != zero) {
175 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
176 }
177 const int8_t* restrict a9 = a[9];
178 if XNN_UNPREDICTABLE(a9 != zero) {
179 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
180 }
181 const int8_t* restrict a10 = a[10];
182 if XNN_UNPREDICTABLE(a10 != zero) {
183 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
184 }
185 const int8_t* restrict a11 = a[11];
186 if XNN_UNPREDICTABLE(a11 != zero) {
187 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
188 }
189 const int8_t* restrict a12 = a[12];
190 if XNN_UNPREDICTABLE(a12 != zero) {
191 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
192 }
193 const int8_t* restrict a13 = a[13];
194 if XNN_UNPREDICTABLE(a13 != zero) {
195 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
196 }
197 a += 14;
198
199 size_t k = kc;
200 while (k >= 16 * sizeof(int8_t)) {
201 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
202 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
203 a0 += 16;
204 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
205 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
206 a1 += 16;
207 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
208 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
209 a2 += 16;
210 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
211 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
212 a3 += 16;
213 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
214 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
215 a4 += 16;
216 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
217 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
218 a5 += 16;
219 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
220 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
221 a6 += 16;
222 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
223 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
224 a7 += 16;
225 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
226 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
227 a8 += 16;
228 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
229 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
230 a9 += 16;
231 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
232 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
233 a10 += 16;
234 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
235 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
236 a11 += 16;
237 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
238 const __m256i va12x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12 + 8)), vsign_mask);
239 a12 += 16;
240 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
241 const __m256i va13x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13 + 8)), vsign_mask);
242 a13 += 16;
243
244 const __m256i vb01234567x0123 = _mm256_load_si256(w);
245 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
246 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
247 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
248 xnn_prefetch_to_l1((const int8_t*) w + 896);
249
250 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
251 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
252 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
253 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
254 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
255 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
256 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
257 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
258 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
259 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
260 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
261 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
262 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
263 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
264 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
265 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
266 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
267 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
268 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
269 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
270 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
271 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
272 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
273 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
274 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
275 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
276 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
277 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, 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 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x89ABCDEF, vb01234567x4567);
304 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x89ABCDEF, vb89ABCDEFx4567);
305 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x89ABCDEF, vb01234567x4567);
306 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x89ABCDEF, vb89ABCDEFx4567);
307
308 w = (const int8_t*) w + 128;
309 k -= 16 * sizeof(int8_t);
310 }
311
312 if (k != 0) {
313 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
314 a0 += 8;
315 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
316 a1 += 8;
317 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
318 a2 += 8;
319 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
320 a3 += 8;
321 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
322 a4 += 8;
323 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
324 a5 += 8;
325 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
326 a6 += 8;
327 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
328 a7 += 8;
329 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
330 a8 += 8;
331 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
332 a9 += 8;
333 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
334 a10 += 8;
335 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
336 a11 += 8;
337 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
338 a12 += 8;
339 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
340 a13 += 8;
341
342 const __m256i vb01234567x0123 = _mm256_load_si256(w);
343 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
344
345 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
346 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
347 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
348 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
349 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
350 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
351 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
352 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
353 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
354 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
355 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
356 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
357 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
358 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
359 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
360 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
361 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
362 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
363 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
364 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
365 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
366 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
367 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
368 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
369 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
370 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
371 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
372 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
373 xnn_prefetch_to_l1((const int8_t*) w + 960);
374
375 w = (const int8_t*) w + 64;
376 k -= 8 * sizeof(int8_t);
377 }
378
379 p -= 14 * sizeof(void*);
380 } while (p != 0);
381
382
383 // Add adjacent pairs
384 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
385 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
386 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
387 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
388 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
389 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
390 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
391 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
392 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
393 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
394 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
395 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
396 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
397 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
398 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
399 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
400 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
401 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
402 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
403 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
404 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
405 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
406 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
407 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
408 const __m256i vsum12x02134657 = _mm256_hadd_epi32(vacc12x0123, vacc12x4567);
409 __m256i vacc12x01234567 = _mm256_permute4x64_epi64(vsum12x02134657, _MM_SHUFFLE(3, 1, 2, 0));
410 const __m256i vsum13x02134657 = _mm256_hadd_epi32(vacc13x0123, vacc13x4567);
411 __m256i vacc13x01234567 = _mm256_permute4x64_epi64(vsum13x02134657, _MM_SHUFFLE(3, 1, 2, 0));
412
413 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
414 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
415 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
416 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
417 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
418 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
419 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
420 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
421 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
422 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
423 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
424 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
425 __m256 vout12x01234567 = _mm256_cvtepi32_ps(vacc12x01234567);
426 __m256 vout13x01234567 = _mm256_cvtepi32_ps(vacc13x01234567);
427
428 const __m256 vscale01234567 = _mm256_load_ps(w);
429 w = (const float*) w + 8;
430 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
431 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
432 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
433 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
434 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
435 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
436 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
437 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vscale01234567);
438 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vscale01234567);
439 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vscale01234567);
440 vout10x01234567 = _mm256_mul_ps(vout10x01234567, vscale01234567);
441 vout11x01234567 = _mm256_mul_ps(vout11x01234567, vscale01234567);
442 vout12x01234567 = _mm256_mul_ps(vout12x01234567, vscale01234567);
443 vout13x01234567 = _mm256_mul_ps(vout13x01234567, vscale01234567);
444
445 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
446 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
447 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
448 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
449 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
450 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
451 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
452 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max_less_zero_point);
453 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max_less_zero_point);
454 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max_less_zero_point);
455 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max_less_zero_point);
456 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max_less_zero_point);
457 vout12x01234567 = _mm256_min_ps(vout12x01234567, voutput_max_less_zero_point);
458 vout13x01234567 = _mm256_min_ps(vout13x01234567, voutput_max_less_zero_point);
459
460 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
461 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
462 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
463 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
464 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
465 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
466 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
467 vacc7x01234567 = _mm256_cvtps_epi32(vout7x01234567);
468 vacc8x01234567 = _mm256_cvtps_epi32(vout8x01234567);
469 vacc9x01234567 = _mm256_cvtps_epi32(vout9x01234567);
470 vacc10x01234567 = _mm256_cvtps_epi32(vout10x01234567);
471 vacc11x01234567 = _mm256_cvtps_epi32(vout11x01234567);
472 vacc12x01234567 = _mm256_cvtps_epi32(vout12x01234567);
473 vacc13x01234567 = _mm256_cvtps_epi32(vout13x01234567);
474
475 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
476 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
477 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
478 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
479 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
480 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
481 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
482 vacc7x01234567 = _mm256_add_epi32(vacc7x01234567, voutput_zero_point);
483 vacc8x01234567 = _mm256_add_epi32(vacc8x01234567, voutput_zero_point);
484 vacc9x01234567 = _mm256_add_epi32(vacc9x01234567, voutput_zero_point);
485 vacc10x01234567 = _mm256_add_epi32(vacc10x01234567, voutput_zero_point);
486 vacc11x01234567 = _mm256_add_epi32(vacc11x01234567, voutput_zero_point);
487 vacc12x01234567 = _mm256_add_epi32(vacc12x01234567, voutput_zero_point);
488 vacc13x01234567 = _mm256_add_epi32(vacc13x01234567, voutput_zero_point);
489
490 __m128i voutb0x01234567 = _mm256_cvtsepi32_epi8(vacc0x01234567);
491 __m128i voutb1x01234567 = _mm256_cvtsepi32_epi8(vacc1x01234567);
492 __m128i voutb2x01234567 = _mm256_cvtsepi32_epi8(vacc2x01234567);
493 __m128i voutb3x01234567 = _mm256_cvtsepi32_epi8(vacc3x01234567);
494 __m128i voutb4x01234567 = _mm256_cvtsepi32_epi8(vacc4x01234567);
495 __m128i voutb5x01234567 = _mm256_cvtsepi32_epi8(vacc5x01234567);
496 __m128i voutb6x01234567 = _mm256_cvtsepi32_epi8(vacc6x01234567);
497 __m128i voutb7x01234567 = _mm256_cvtsepi32_epi8(vacc7x01234567);
498 __m128i voutb8x01234567 = _mm256_cvtsepi32_epi8(vacc8x01234567);
499 __m128i voutb9x01234567 = _mm256_cvtsepi32_epi8(vacc9x01234567);
500 __m128i voutb10x01234567 = _mm256_cvtsepi32_epi8(vacc10x01234567);
501 __m128i voutb11x01234567 = _mm256_cvtsepi32_epi8(vacc11x01234567);
502 __m128i voutb12x01234567 = _mm256_cvtsepi32_epi8(vacc12x01234567);
503 __m128i voutb13x01234567 = _mm256_cvtsepi32_epi8(vacc13x01234567);
504
505 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
506 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
507 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
508 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
509 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
510 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
511 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
512 voutb7x01234567 = _mm_max_epi8(voutb7x01234567, voutput_min);
513 voutb8x01234567 = _mm_max_epi8(voutb8x01234567, voutput_min);
514 voutb9x01234567 = _mm_max_epi8(voutb9x01234567, voutput_min);
515 voutb10x01234567 = _mm_max_epi8(voutb10x01234567, voutput_min);
516 voutb11x01234567 = _mm_max_epi8(voutb11x01234567, voutput_min);
517 voutb12x01234567 = _mm_max_epi8(voutb12x01234567, voutput_min);
518 voutb13x01234567 = _mm_max_epi8(voutb13x01234567, voutput_min);
519
520 if (nc >= 8) {
521 _mm_storel_epi64((__m128i*) c13, voutb13x01234567);
522 c13 = (int8_t*) ((uintptr_t) c13 + cn_stride);
523 _mm_storel_epi64((__m128i*) c12, voutb12x01234567);
524 c12 = (int8_t*) ((uintptr_t) c12 + cn_stride);
525 _mm_storel_epi64((__m128i*) c11, voutb11x01234567);
526 c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
527 _mm_storel_epi64((__m128i*) c10, voutb10x01234567);
528 c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
529 _mm_storel_epi64((__m128i*) c9, voutb9x01234567);
530 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
531 _mm_storel_epi64((__m128i*) c8, voutb8x01234567);
532 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
533 _mm_storel_epi64((__m128i*) c7, voutb7x01234567);
534 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
535 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
536 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
537 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
538 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
539 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
540 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
541 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
542 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
543 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
544 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
545 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
546 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
547 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
548 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
549 a = (const int8_t**restrict) ((uintptr_t) a - ks);
550 nc -= 8;
551 } else {
552 // Prepare mask for valid 8-bit elements (depends on nc).
553 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
554 _mm_mask_storeu_epi8(c13, vmask, voutb13x01234567);
555 _mm_mask_storeu_epi8(c12, vmask, voutb12x01234567);
556 _mm_mask_storeu_epi8(c11, vmask, voutb11x01234567);
557 _mm_mask_storeu_epi8(c10, vmask, voutb10x01234567);
558 _mm_mask_storeu_epi8(c9, vmask, voutb9x01234567);
559 _mm_mask_storeu_epi8(c8, vmask, voutb8x01234567);
560 _mm_mask_storeu_epi8(c7, vmask, voutb7x01234567);
561 _mm_mask_storeu_epi8(c6, vmask, voutb6x01234567);
562 _mm_mask_storeu_epi8(c5, vmask, voutb5x01234567);
563 _mm_mask_storeu_epi8(c4, vmask, voutb4x01234567);
564 _mm_mask_storeu_epi8(c3, vmask, voutb3x01234567);
565 _mm_mask_storeu_epi8(c2, vmask, voutb2x01234567);
566 _mm_mask_storeu_epi8(c1, vmask, voutb1x01234567);
567 _mm_mask_storeu_epi8(c0, vmask, voutb0x01234567);
568 nc = 0;
569 }
570 } while (nc != 0);
571 }