Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-9x16c4-minmax-avx512vnni.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
20
21 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_9x16c4__avx512vnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 9);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 4 * sizeof(int8_t));
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
65 if XNN_UNPREDICTABLE(mr < 6) {
66 c5 = c4;
67 }
68 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
69 if XNN_UNPREDICTABLE(mr <= 6) {
70 c6 = c5;
71 }
72 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 8) {
74 c7 = c6;
75 }
76 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
77 if XNN_UNPREDICTABLE(mr <= 8) {
78 c8 = c7;
79 }
80
81 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
82 XNN_FORCE_REALIZATION(vsign_mask);
83 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
84 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
85 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
86 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
87 XNN_FORCE_REALIZATION(vinput_zero_point);
88 // XNN_FORCE_REALIZATION(vinput_inv_scale);
89 // XNN_FORCE_REALIZATION(voutput_min);
90 // XNN_FORCE_REALIZATION(voutput_max);
91 do {
92 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
93 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
94 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
95 __m512i vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
96 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
97 __m512i vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
98 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
99 __m512i vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
100 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
101 __m512i vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
102 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
103 __m512i vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
104 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
105 __m512i vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
106 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
107 __m512i vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
108 __m512i vacc1x7x0123456789ABCDEF = _mm512_setzero_epi32();
109 __m512i vacc8x0123456789ABCDEF = vacc0x0123456789ABCDEF;
110 __m512i vacc1x8x0123456789ABCDEF = _mm512_setzero_epi32();
111 w = (const int32_t*) w + 16;
112
113 size_t p = ks;
114 do {
115 const int8_t* restrict a0 = a[0];
116 if XNN_UNPREDICTABLE(a0 != zero) {
117 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
118 } else {
119 a0 = zero_data;
120 }
121 const int8_t* restrict a1 = a[1];
122 if XNN_UNPREDICTABLE(a1 != zero) {
123 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
124 } else {
125 a1 = zero_data;
126 }
127 const int8_t* restrict a2 = a[2];
128 if XNN_UNPREDICTABLE(a2 != zero) {
129 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
130 } else {
131 a2 = zero_data;
132 }
133 const int8_t* restrict a3 = a[3];
134 if XNN_UNPREDICTABLE(a3 != zero) {
135 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
136 } else {
137 a3 = zero_data;
138 }
139 const int8_t* restrict a4 = a[4];
140 if XNN_UNPREDICTABLE(a4 != zero) {
141 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
142 } else {
143 a4 = zero_data;
144 }
145 const int8_t* restrict a5 = a[5];
146 if XNN_UNPREDICTABLE(a5 != zero) {
147 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
148 } else {
149 a5 = zero_data;
150 }
151 const int8_t* restrict a6 = a[6];
152 if XNN_UNPREDICTABLE(a6 != zero) {
153 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
154 } else {
155 a6 = zero_data;
156 }
157 const int8_t* restrict a7 = a[7];
158 if XNN_UNPREDICTABLE(a7 != zero) {
159 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
160 } else {
161 a7 = zero_data;
162 }
163 const int8_t* restrict a8 = a[8];
164 if XNN_UNPREDICTABLE(a8 != zero) {
165 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
166 } else {
167 a8 = zero_data;
168 }
169 a += 9;
170
171 size_t k = kc;
172 while (k >= 8 * sizeof(int8_t)) {
173 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
174 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
175 a0 += 8;
176 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
177 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
178 a1 += 8;
179 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
180 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
181 a2 += 8;
182 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
183 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
184 a3 += 8;
185 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
186 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
187 a4 += 8;
188 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
189 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
190 a5 += 8;
191 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
192 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
193 a6 += 8;
194 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
195 const __m512i va7x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7 + 4)), vsign_mask);
196 a7 += 8;
197 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
198 const __m512i va8x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8 + 4)), vsign_mask);
199 a8 += 8;
200
201 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
202 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
203
204 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
205 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
206 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
207 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
208 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
209 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
210 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
211 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEFx0123);
212 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEFx0123);
213 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
214 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
215 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
216 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
217 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
218 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
219 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
220 vacc1x7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x7x0123456789ABCDEF, va7x4567, vb0123456789ABCDEFx4567);
221 vacc1x8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x8x0123456789ABCDEF, va8x4567, vb0123456789ABCDEFx4567);
222
223 w = (const int8_t*) w + 128;
224 k -= 8 * sizeof(int8_t);
225 }
226
227 if (k != 0) {
228 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
229 a0 += 4;
230 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
231 a1 += 4;
232 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
233 a2 += 4;
234 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
235 a3 += 4;
236 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
237 a4 += 4;
238 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
239 a5 += 4;
240 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
241 a6 += 4;
242 const __m512i va7x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a7)), vsign_mask);
243 a7 += 4;
244 const __m512i va8x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a8)), vsign_mask);
245 a8 += 4;
246
247 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
248
249 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
250 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
251 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
252 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
253 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
254 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
255 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
256 vacc7x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc7x0123456789ABCDEF, va7x0123, vb0123456789ABCDEF);
257 vacc8x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc8x0123456789ABCDEF, va8x0123, vb0123456789ABCDEF);
258
259 w = (const int8_t*) w + 64;
260 k -= 4 * sizeof(int8_t);
261 }
262
263 p -= 9 * sizeof(void*);
264 } while (p != 0);
265
266 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
267 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
268 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
269 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
270 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
271 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
272 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
273 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, vacc1x7x0123456789ABCDEF);
274 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, vacc1x8x0123456789ABCDEF);
275
276 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
277 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
278 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
279 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
280 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
281 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
282 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
283 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
284 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
285
286 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
287 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
288 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
289 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
290 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
291 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
292 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
293 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vinput_inv_scale);
294 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vinput_inv_scale);
295
296 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
297 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
298 w = (const float*) w + 32;
299
300 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
301 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
302 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
303 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
304 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
305 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
306 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
307 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
308 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
309
310 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
311 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
312 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
313 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
314 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
315 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
316 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
317 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
318 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
319
320 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
321 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
322 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
323 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
324 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
325 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
326 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
327 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
328 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
329
330 if XNN_LIKELY(nc >= 16) {
331 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
332 c8 = (float*) ((uintptr_t) c8 + cn_stride);
333 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
334 c7 = (float*) ((uintptr_t) c7 + cn_stride);
335 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
336 c6 = (float*) ((uintptr_t) c6 + cn_stride);
337 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
338 c5 = (float*) ((uintptr_t) c5 + cn_stride);
339 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
340 c4 = (float*) ((uintptr_t) c4 + cn_stride);
341 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
342 c3 = (float*) ((uintptr_t) c3 + cn_stride);
343 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
344 c2 = (float*) ((uintptr_t) c2 + cn_stride);
345 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
346 c1 = (float*) ((uintptr_t) c1 + cn_stride);
347 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
348 c0 = (float*) ((uintptr_t) c0 + cn_stride);
349 a = (const int8_t**restrict) ((uintptr_t) a - ks);
350 nc -= 16;
351 } else {
352 // Prepare mask for valid 32-bit elements (depends on nc).
353 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
354 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
355 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
356 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
357 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
358 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
359 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
360 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
361 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
362 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
363 nc = 0;
364 }
365 } while (nc != 0);
366 }