Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-10x16c8-minmax-avx512vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c8-avx512vnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_f32_qc8w_igemm_minmax_ukernel_10x16c8__avx512vnni_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 float* 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_f32_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 <= 10);
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 float* c0 = c;
49 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 c1 = c0;
52 }
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 c2 = c1;
56 }
57 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 c3 = c2;
60 }
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 c4 = c3;
64 }
65 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr < 6) {
67 c5 = c4;
68 }
69 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
70 if XNN_UNPREDICTABLE(mr <= 6) {
71 c6 = c5;
72 }
73 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
74 if XNN_UNPREDICTABLE(mr < 8) {
75 c7 = c6;
76 }
77 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 8) {
79 c8 = c7;
80 }
81 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
82 if XNN_UNPREDICTABLE(mr != 10) {
83 c9 = c8;
84 }
85
86 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
87 XNN_FORCE_REALIZATION(vsign_mask);
88 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
89 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
90 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
91 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
92 // XNN_FORCE_REALIZATION(vinput_zero_point);
93 // XNN_FORCE_REALIZATION(vinput_inv_scale);
94 // XNN_FORCE_REALIZATION(voutput_min);
95 // XNN_FORCE_REALIZATION(voutput_max);
96 do {
97 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
98 const __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
99 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
100 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
101 __m512i vacc1x01234567 = vacc0x01234567;
102 __m512i vacc1x89ABCDEF = vacc0x89ABCDEF;
103 __m512i vacc2x01234567 = vacc0x01234567;
104 __m512i vacc2x89ABCDEF = vacc0x89ABCDEF;
105 __m512i vacc3x01234567 = vacc0x01234567;
106 __m512i vacc3x89ABCDEF = vacc0x89ABCDEF;
107 __m512i vacc4x01234567 = vacc0x01234567;
108 __m512i vacc4x89ABCDEF = vacc0x89ABCDEF;
109 __m512i vacc5x01234567 = vacc0x01234567;
110 __m512i vacc5x89ABCDEF = vacc0x89ABCDEF;
111 __m512i vacc6x01234567 = vacc0x01234567;
112 __m512i vacc6x89ABCDEF = vacc0x89ABCDEF;
113 __m512i vacc7x01234567 = vacc0x01234567;
114 __m512i vacc7x89ABCDEF = vacc0x89ABCDEF;
115 __m512i vacc8x01234567 = vacc0x01234567;
116 __m512i vacc8x89ABCDEF = vacc0x89ABCDEF;
117 __m512i vacc9x01234567 = vacc0x01234567;
118 __m512i vacc9x89ABCDEF = vacc0x89ABCDEF;
119 w = (const int32_t*) w + 16;
120
121 size_t p = ks;
122 do {
123 const int8_t* restrict a0 = a[0];
124 if XNN_UNPREDICTABLE(a0 != zero) {
125 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
126 } else {
127 a0 = zero_data;
128 }
129 const int8_t* restrict a1 = a[1];
130 if XNN_UNPREDICTABLE(a1 != zero) {
131 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
132 } else {
133 a1 = zero_data;
134 }
135 const int8_t* restrict a2 = a[2];
136 if XNN_UNPREDICTABLE(a2 != zero) {
137 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
138 } else {
139 a2 = zero_data;
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 } else {
145 a3 = zero_data;
146 }
147 const int8_t* restrict a4 = a[4];
148 if XNN_UNPREDICTABLE(a4 != zero) {
149 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
150 } else {
151 a4 = zero_data;
152 }
153 const int8_t* restrict a5 = a[5];
154 if XNN_UNPREDICTABLE(a5 != zero) {
155 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
156 } else {
157 a5 = zero_data;
158 }
159 const int8_t* restrict a6 = a[6];
160 if XNN_UNPREDICTABLE(a6 != zero) {
161 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
162 } else {
163 a6 = zero_data;
164 }
165 const int8_t* restrict a7 = a[7];
166 if XNN_UNPREDICTABLE(a7 != zero) {
167 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
168 } else {
169 a7 = zero_data;
170 }
171 const int8_t* restrict a8 = a[8];
172 if XNN_UNPREDICTABLE(a8 != zero) {
173 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
174 } else {
175 a8 = zero_data;
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 } else {
181 a9 = zero_data;
182 }
183 a += 10;
184
185 size_t k = kc;
186 while (k >= 16 * sizeof(int8_t)) {
187 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
188 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
189 a0 += 16;
190 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
191 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
192 a1 += 16;
193 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
194 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
195 a2 += 16;
196 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
197 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
198 a3 += 16;
199 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
200 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
201 a4 += 16;
202 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
203 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
204 a5 += 16;
205 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
206 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
207 a6 += 16;
208 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
209 const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
210 a7 += 16;
211 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
212 const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
213 a8 += 16;
214 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
215 const __m512i va9x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
216 a9 += 16;
217
218 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
219 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
220 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
221 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
222 xnn_prefetch_to_l1((const int8_t*) w + 768);
223 xnn_prefetch_to_l1((const int8_t*) w + 832);
224
225 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
226 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
227 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
228 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
229 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
230 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
231 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
232 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
233 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
234 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
235 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
236 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
237 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
238 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
239 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
240 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
241 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
242 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
243 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
244 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
245 xnn_prefetch_to_l1((const int8_t*) w + 896);
246 xnn_prefetch_to_l1((const int8_t*) w + 960);
247 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
248 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
249 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
250 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
251 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
252 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
253 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
254 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
255 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
256 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
257 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
258 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
259 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
260 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
261 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF);
262 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF);
263 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF);
264 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF);
265 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x89ABCDEF, vb01234567x89ABCDEF);
266 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x89ABCDEF, vb89ABCDEFx89ABCDEF);
267
268 w = (const int8_t*) w + 256;
269 k -= 16 * sizeof(int8_t);
270 }
271
272 if (k != 0) {
273 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
274 a0 += 8;
275 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
276 a1 += 8;
277 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
278 a2 += 8;
279 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
280 a3 += 8;
281 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
282 a4 += 8;
283 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
284 a5 += 8;
285 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
286 a6 += 8;
287 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
288 a7 += 8;
289 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
290 a8 += 8;
291 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
292 a9 += 8;
293
294 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
295 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
296
297 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
298 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
299 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
300 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
301 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
302 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
303 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
304 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
305 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
306 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
307 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
308 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
309 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
310 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
311 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
312 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
313 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
314 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
315 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
316 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
317 xnn_prefetch_to_l1((const int8_t*) w + 896);
318 xnn_prefetch_to_l1((const int8_t*) w + 960);
319
320 w = (const int8_t*) w + 128;
321 k -= 8 * sizeof(int8_t);
322 }
323
324 p -= 10 * sizeof(void*);
325 } while (p != 0);
326
327
328 // Add adjacent pairs
329 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
330 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
331 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
332 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
333 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
334 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
335 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
336 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
337 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
338 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
339 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
340 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
341 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
342 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
343 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
344 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
345 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
346 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
347 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
348 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
349 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
350 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
351 const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32));
352 const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32));
353 __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF);
354 const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32));
355 const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32));
356 __m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF);
357 const __m512i vsum9x01234567 = _mm512_add_epi32(vacc9x01234567, _mm512_srli_epi64(vacc9x01234567, 32));
358 const __m512i vsum9x89ABCDEF = _mm512_add_epi32(vacc9x89ABCDEF, _mm512_srli_epi64(vacc9x89ABCDEF, 32));
359 __m512i vacc9x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum9x01234567, vidx, vsum9x89ABCDEF);
360
361 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
362 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
363 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
364 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
365 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
366 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
367 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
368 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
369 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
370 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
371
372 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
373 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
374 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
375 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
376 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
377 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
378 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
379 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vinput_inv_scale);
380 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vinput_inv_scale);
381 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vinput_inv_scale);
382
383 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
384 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
385 w = (const float*) w + 32;
386
387 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
388 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
389 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
390 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
391 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
392 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
393 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
394 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
395 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
396 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
397
398 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
399 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
400 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
401 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
402 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
403 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
404 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
405 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
406 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
407 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
408
409 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
410 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
411 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
412 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
413 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
414 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
415 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
416 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
417 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
418 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
419
420 if XNN_LIKELY(nc >= 16) {
421 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
422 c9 = (float*) ((uintptr_t) c9 + cn_stride);
423 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
424 c8 = (float*) ((uintptr_t) c8 + cn_stride);
425 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
426 c7 = (float*) ((uintptr_t) c7 + cn_stride);
427 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
428 c6 = (float*) ((uintptr_t) c6 + cn_stride);
429 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
430 c5 = (float*) ((uintptr_t) c5 + cn_stride);
431 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
432 c4 = (float*) ((uintptr_t) c4 + cn_stride);
433 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
434 c3 = (float*) ((uintptr_t) c3 + cn_stride);
435 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
436 c2 = (float*) ((uintptr_t) c2 + cn_stride);
437 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
438 c1 = (float*) ((uintptr_t) c1 + cn_stride);
439 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
440 c0 = (float*) ((uintptr_t) c0 + cn_stride);
441 a = (const int8_t**restrict) ((uintptr_t) a - ks);
442 nc -= 16;
443 } else {
444 // Prepare mask for valid 32-bit elements (depends on nc).
445 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
446 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
447 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
448 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
449 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
450 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
451 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
452 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
453 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
454 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
455 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
456 nc = 0;
457 }
458 } while (nc != 0);
459 }