Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-10x16c4-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_10x16c4__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, 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
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 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
99 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
100 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
101 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
102 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
103 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
104 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
105 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
106 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
107 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
108 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
109 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
110 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
111 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
112 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
113 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
114 __m512i vacc8x0123456789ABCDEF = vacc0x0123456789ABCDEF;
115 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
116 __m512i vacc9x0123456789ABCDEF = vacc0x0123456789ABCDEF;
117 __m512i vacc1x9x0123456789ABCDEF = _mm512_setzero_epi32();
118 w = (const int32_t*) w + 16;
119
120 size_t p = ks;
121 do {
122 const int8_t* restrict a0 = a[0];
123 if XNN_UNPREDICTABLE(a0 != zero) {
124 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
125 } else {
126 a0 = zero_data;
127 }
128 const int8_t* restrict a1 = a[1];
129 if XNN_UNPREDICTABLE(a1 != zero) {
130 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
131 } else {
132 a1 = zero_data;
133 }
134 const int8_t* restrict a2 = a[2];
135 if XNN_UNPREDICTABLE(a2 != zero) {
136 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
137 } else {
138 a2 = zero_data;
139 }
140 const int8_t* restrict a3 = a[3];
141 if XNN_UNPREDICTABLE(a3 != zero) {
142 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
143 } else {
144 a3 = zero_data;
145 }
146 const int8_t* restrict a4 = a[4];
147 if XNN_UNPREDICTABLE(a4 != zero) {
148 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
149 } else {
150 a4 = zero_data;
151 }
152 const int8_t* restrict a5 = a[5];
153 if XNN_UNPREDICTABLE(a5 != zero) {
154 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
155 } else {
156 a5 = zero_data;
157 }
158 const int8_t* restrict a6 = a[6];
159 if XNN_UNPREDICTABLE(a6 != zero) {
160 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
161 } else {
162 a6 = zero_data;
163 }
164 const int8_t* restrict a7 = a[7];
165 if XNN_UNPREDICTABLE(a7 != zero) {
166 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
167 } else {
168 a7 = zero_data;
169 }
170 const int8_t* restrict a8 = a[8];
171 if XNN_UNPREDICTABLE(a8 != zero) {
172 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
173 } else {
174 a8 = zero_data;
175 }
176 const int8_t* restrict a9 = a[9];
177 if XNN_UNPREDICTABLE(a9 != zero) {
178 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
179 } else {
180 a9 = zero_data;
181 }
182 a += 10;
183
184 size_t k = kc;
185 while (k >= 8 * sizeof(int8_t)) {
186 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
187 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
188 a0 += 8;
189 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
190 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
191 a1 += 8;
192 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
193 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
194 a2 += 8;
195 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
196 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
197 a3 += 8;
198 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
199 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
200 a4 += 8;
201 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
202 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
203 a5 += 8;
204 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
205 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
206 a6 += 8;
207 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
208 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
209 a7 += 8;
210 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
211 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
212 a8 += 8;
213 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
214 const __m512i va9x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9 + 4)), vsign_mask);
215 a9 += 8;
216
217 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
218 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
219 xnn_prefetch_to_l1((const int8_t*) w + 896);
220
221 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
222 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
223 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
224 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
225 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
226 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
227 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
228 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
229 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
230 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEFx0123);
231 xnn_prefetch_to_l1((const int8_t*) w + 960);
232 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
233 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
234 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
235 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
236 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
237 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
238 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
239 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
240 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
241 vacc1x9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x9x0123456789ABCDEF, va9x4567, vb0123456789ABCDEFx4567);
242
243 w = (const int8_t*) w + 128;
244 k -= 8 * sizeof(int8_t);
245 }
246
247 if (k != 0) {
248 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
249 a0 += 4;
250 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
251 a1 += 4;
252 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
253 a2 += 4;
254 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
255 a3 += 4;
256 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
257 a4 += 4;
258 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
259 a5 += 4;
260 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
261 a6 += 4;
262 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
263 a7 += 4;
264 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
265 a8 += 4;
266 const __m512i va9x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a9)), vsign_mask);
267 a9 += 4;
268
269 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
270 xnn_prefetch_to_l1((const int8_t*) w + 960);
271
272 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
273 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
274 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
275 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
276 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
277 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
278 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
279 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
280 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
281 vacc9x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc9x0123456789ABCDEF, va9x0123, vb0123456789ABCDEF);
282
283 w = (const int8_t*) w + 64;
284 k -= 4 * sizeof(int8_t);
285 }
286
287 p -= 10 * sizeof(void*);
288 } while (p != 0);
289
290 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
291 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
292 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
293 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
294 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
295 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
296 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
297 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
298 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
299 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, vacc1x9x0123456789ABCDEF);
300
301 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
302 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
303 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
304 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
305 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
306 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
307 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
308 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
309 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
310 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
311
312 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
313 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
314 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
315 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
316 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
317 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
318 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
319 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vinput_inv_scale);
320 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vinput_inv_scale);
321 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vinput_inv_scale);
322
323 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
324 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
325 w = (const float*) w + 32;
326
327 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
328 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
329 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
330 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
331 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
332 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
333 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
334 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
335 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
336 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
337
338 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
339 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
340 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
341 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
342 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
343 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
344 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
345 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
346 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
347 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
348
349 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
350 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
351 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
352 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
353 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
354 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
355 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
356 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
357 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
358 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
359
360 if XNN_LIKELY(nc >= 16) {
361 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
362 c9 = (float*) ((uintptr_t) c9 + cn_stride);
363 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
364 c8 = (float*) ((uintptr_t) c8 + cn_stride);
365 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
366 c7 = (float*) ((uintptr_t) c7 + cn_stride);
367 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
368 c6 = (float*) ((uintptr_t) c6 + cn_stride);
369 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
370 c5 = (float*) ((uintptr_t) c5 + cn_stride);
371 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
372 c4 = (float*) ((uintptr_t) c4 + cn_stride);
373 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
374 c3 = (float*) ((uintptr_t) c3 + cn_stride);
375 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
376 c2 = (float*) ((uintptr_t) c2 + cn_stride);
377 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
378 c1 = (float*) ((uintptr_t) c1 + cn_stride);
379 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
380 c0 = (float*) ((uintptr_t) c0 + cn_stride);
381 a = (const int8_t**restrict) ((uintptr_t) a - ks);
382 nc -= 16;
383 } else {
384 // Prepare mask for valid 32-bit elements (depends on nc).
385 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
386 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
387 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
388 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
389 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
390 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
391 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
392 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
393 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
394 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
395 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
396 nc = 0;
397 }
398 } while (nc != 0);
399 }