Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-12x16c4-minmax-avx512vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c4-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_12x16c4__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 <= 12);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 4 * 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 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
86 if XNN_UNPREDICTABLE(mr <= 10) {
87 c10 = c9;
88 }
89 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
90 if XNN_UNPREDICTABLE(mr != 12) {
91 c11 = c10;
92 }
93
94 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
95 XNN_FORCE_REALIZATION(vsign_mask);
96 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
97 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
98 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
99 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
100 XNN_FORCE_REALIZATION(vinput_zero_point);
101 // XNN_FORCE_REALIZATION(vinput_inv_scale);
102 // XNN_FORCE_REALIZATION(voutput_min);
103 // XNN_FORCE_REALIZATION(voutput_max);
104 do {
105 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
106 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
107 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
108 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
109 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
110 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
111 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
112 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
113 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
114 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
115 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
116 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
117 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
118 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
119 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
120 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
121 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
122 __m512i vacc8x0123456789ABCDEF = vacc0x0123456789ABCDEF;
123 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
124 __m512i vacc9x0123456789ABCDEF = vacc0x0123456789ABCDEF;
125 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
126 __m512i vacc10x0123456789ABCDEF = vacc0x0123456789ABCDEF;
127 __m512i vacc1x10x0123456789ABCDEF = _mm512_setzero_epi32();
128 __m512i vacc11x0123456789ABCDEF = vacc0x0123456789ABCDEF;
129 __m512i vacc1x11x0123456789ABCDEF = _mm512_setzero_epi32();
130 w = (const int32_t*) w + 16;
131
132 size_t p = ks;
133 do {
134 const int8_t* restrict a0 = a[0];
135 if XNN_UNPREDICTABLE(a0 != zero) {
136 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
137 } else {
138 a0 = zero_data;
139 }
140 const int8_t* restrict a1 = a[1];
141 if XNN_UNPREDICTABLE(a1 != zero) {
142 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
143 } else {
144 a1 = zero_data;
145 }
146 const int8_t* restrict a2 = a[2];
147 if XNN_UNPREDICTABLE(a2 != zero) {
148 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
149 } else {
150 a2 = zero_data;
151 }
152 const int8_t* restrict a3 = a[3];
153 if XNN_UNPREDICTABLE(a3 != zero) {
154 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
155 } else {
156 a3 = zero_data;
157 }
158 const int8_t* restrict a4 = a[4];
159 if XNN_UNPREDICTABLE(a4 != zero) {
160 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
161 } else {
162 a4 = zero_data;
163 }
164 const int8_t* restrict a5 = a[5];
165 if XNN_UNPREDICTABLE(a5 != zero) {
166 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
167 } else {
168 a5 = zero_data;
169 }
170 const int8_t* restrict a6 = a[6];
171 if XNN_UNPREDICTABLE(a6 != zero) {
172 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
173 } else {
174 a6 = zero_data;
175 }
176 const int8_t* restrict a7 = a[7];
177 if XNN_UNPREDICTABLE(a7 != zero) {
178 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
179 } else {
180 a7 = zero_data;
181 }
182 const int8_t* restrict a8 = a[8];
183 if XNN_UNPREDICTABLE(a8 != zero) {
184 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
185 } else {
186 a8 = zero_data;
187 }
188 const int8_t* restrict a9 = a[9];
189 if XNN_UNPREDICTABLE(a9 != zero) {
190 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
191 } else {
192 a9 = zero_data;
193 }
194 const int8_t* restrict a10 = a[10];
195 if XNN_UNPREDICTABLE(a10 != zero) {
196 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
197 } else {
198 a10 = zero_data;
199 }
200 const int8_t* restrict a11 = a[11];
201 if XNN_UNPREDICTABLE(a11 != zero) {
202 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
203 } else {
204 a11 = zero_data;
205 }
206 a += 12;
207
208 size_t k = kc;
209 while (k >= 8 * sizeof(int8_t)) {
210 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
211 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
212 a0 += 8;
213 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
214 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
215 a1 += 8;
216 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
217 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
218 a2 += 8;
219 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
220 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
221 a3 += 8;
222 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
223 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
224 a4 += 8;
225 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
226 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
227 a5 += 8;
228 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
229 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
230 a6 += 8;
231 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
232 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
233 a7 += 8;
234 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
235 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
236 a8 += 8;
237 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
238 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
239 a9 += 8;
240 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
241 const __m512i va10x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10 + 4)), vsign_mask);
242 a10 += 8;
243 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
244 const __m512i va11x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11 + 4)), vsign_mask);
245 a11 += 8;
246
247 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
248 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
249 xnn_prefetch_to_l1((const int8_t*) w + 896);
250
251 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
252 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
253 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
254 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
255 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
256 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
257 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
258 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
259 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
260 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
261 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEFx0123);
262 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEFx0123);
263 xnn_prefetch_to_l1((const int8_t*) w + 960);
264 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
265 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
266 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
267 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
268 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
269 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
270 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
271 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
272 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
273 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
274 vacc1x10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x10x0123456789ABCDEF, va10x4567, vb0123456789ABCDEFx4567);
275 vacc1x11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x11x0123456789ABCDEF, va11x4567, vb0123456789ABCDEFx4567);
276
277 w = (const int8_t*) w + 128;
278 k -= 8 * sizeof(int8_t);
279 }
280
281 if (k != 0) {
282 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
283 a0 += 4;
284 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
285 a1 += 4;
286 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
287 a2 += 4;
288 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
289 a3 += 4;
290 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
291 a4 += 4;
292 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
293 a5 += 4;
294 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
295 a6 += 4;
296 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
297 a7 += 4;
298 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
299 a8 += 4;
300 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
301 a9 += 4;
302 const __m512i va10x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a10)), vsign_mask);
303 a10 += 4;
304 const __m512i va11x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a11)), vsign_mask);
305 a11 += 4;
306
307 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
308 xnn_prefetch_to_l1((const int8_t*) w + 960);
309
310 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
311 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
312 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
313 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
314 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
315 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
316 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
317 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
318 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
319 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
320 vacc10x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc10x0123456789ABCDEF, va10x0123, vb0123456789ABCDEF);
321 vacc11x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc11x0123456789ABCDEF, va11x0123, vb0123456789ABCDEF);
322
323 w = (const int8_t*) w + 64;
324 k -= 4 * sizeof(int8_t);
325 }
326
327 p -= 12 * sizeof(void*);
328 } while (p != 0);
329
330 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
331 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
332 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
333 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
334 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
335 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
336 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
337 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
338 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
339 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
340 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, vacc1x10x0123456789ABCDEF);
341 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, vacc1x11x0123456789ABCDEF);
342
343 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
344 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
345 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
346 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
347 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
348 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
349 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
350 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
351 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
352 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
353 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
354 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
355
356 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
357 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
358 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
359 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
360 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
361 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
362 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
363 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vinput_inv_scale);
364 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vinput_inv_scale);
365 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vinput_inv_scale);
366 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, vinput_inv_scale);
367 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, vinput_inv_scale);
368
369 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
370 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
371 w = (const float*) w + 32;
372
373 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
374 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
375 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
376 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
377 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
378 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
379 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
380 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
381 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
382 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
383 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
384 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
385
386 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
387 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
388 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
389 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
390 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
391 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
392 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
393 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
394 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
395 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
396 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
397 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
398
399 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
400 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
401 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
402 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
403 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
404 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
405 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
406 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
407 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
408 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
409 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
410 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
411
412 if XNN_LIKELY(nc >= 16) {
413 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
414 c11 = (float*) ((uintptr_t) c11 + cn_stride);
415 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
416 c10 = (float*) ((uintptr_t) c10 + cn_stride);
417 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
418 c9 = (float*) ((uintptr_t) c9 + cn_stride);
419 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
420 c8 = (float*) ((uintptr_t) c8 + cn_stride);
421 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
422 c7 = (float*) ((uintptr_t) c7 + cn_stride);
423 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
424 c6 = (float*) ((uintptr_t) c6 + cn_stride);
425 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
426 c5 = (float*) ((uintptr_t) c5 + cn_stride);
427 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
428 c4 = (float*) ((uintptr_t) c4 + cn_stride);
429 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
430 c3 = (float*) ((uintptr_t) c3 + cn_stride);
431 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
432 c2 = (float*) ((uintptr_t) c2 + cn_stride);
433 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
434 c1 = (float*) ((uintptr_t) c1 + cn_stride);
435 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
436 c0 = (float*) ((uintptr_t) c0 + cn_stride);
437 a = (const int8_t**restrict) ((uintptr_t) a - ks);
438 nc -= 16;
439 } else {
440 // Prepare mask for valid 32-bit elements (depends on nc).
441 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
442 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
443 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
444 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
445 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
446 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
447 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
448 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
449 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
450 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
451 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
452 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
453 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
454 nc = 0;
455 }
456 } while (nc != 0);
457 }