Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-7x16c8-minmax-avx512skx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c8-avx512skx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/igemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_7x16c8__avx512skx_prfm(
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 <= 7);
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, 8 * 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
73 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
74 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point);
75 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
76 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
77 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
78 // XNN_FORCE_REALIZATION(vinput_zero_point);
79 // XNN_FORCE_REALIZATION(vinput_inv_scale);
80 // XNN_FORCE_REALIZATION(voutput_min);
81 // XNN_FORCE_REALIZATION(voutput_max);
82 do {
83 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
84 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
85 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
86 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
87 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point);
88 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point);
89 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point);
90 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point);
91 __m512i vacc1x0123 = vacc0x0123;
92 __m512i vacc1x4567 = vacc0x4567;
93 __m512i vacc1x89AB = vacc0x89AB;
94 __m512i vacc1xCDEF = vacc0xCDEF;
95 __m512i vacc2x0123 = vacc0x0123;
96 __m512i vacc2x4567 = vacc0x4567;
97 __m512i vacc2x89AB = vacc0x89AB;
98 __m512i vacc2xCDEF = vacc0xCDEF;
99 __m512i vacc3x0123 = vacc0x0123;
100 __m512i vacc3x4567 = vacc0x4567;
101 __m512i vacc3x89AB = vacc0x89AB;
102 __m512i vacc3xCDEF = vacc0xCDEF;
103 __m512i vacc4x0123 = vacc0x0123;
104 __m512i vacc4x4567 = vacc0x4567;
105 __m512i vacc4x89AB = vacc0x89AB;
106 __m512i vacc4xCDEF = vacc0xCDEF;
107 __m512i vacc5x0123 = vacc0x0123;
108 __m512i vacc5x4567 = vacc0x4567;
109 __m512i vacc5x89AB = vacc0x89AB;
110 __m512i vacc5xCDEF = vacc0xCDEF;
111 __m512i vacc6x0123 = vacc0x0123;
112 __m512i vacc6x4567 = vacc0x4567;
113 __m512i vacc6x89AB = vacc0x89AB;
114 __m512i vacc6xCDEF = vacc0xCDEF;
115 w = (const int32_t*) w + 16;
116
117 size_t p = ks;
118 do {
119 const int8_t* restrict a0 = a[0];
120 if XNN_UNPREDICTABLE(a0 != zero) {
121 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
122 } else {
123 a0 = zero_data;
124 }
125 const int8_t* restrict a1 = a[1];
126 if XNN_UNPREDICTABLE(a1 != zero) {
127 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
128 } else {
129 a1 = zero_data;
130 }
131 const int8_t* restrict a2 = a[2];
132 if XNN_UNPREDICTABLE(a2 != zero) {
133 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
134 } else {
135 a2 = zero_data;
136 }
137 const int8_t* restrict a3 = a[3];
138 if XNN_UNPREDICTABLE(a3 != zero) {
139 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
140 } else {
141 a3 = zero_data;
142 }
143 const int8_t* restrict a4 = a[4];
144 if XNN_UNPREDICTABLE(a4 != zero) {
145 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
146 } else {
147 a4 = zero_data;
148 }
149 const int8_t* restrict a5 = a[5];
150 if XNN_UNPREDICTABLE(a5 != zero) {
151 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
152 } else {
153 a5 = zero_data;
154 }
155 const int8_t* restrict a6 = a[6];
156 if XNN_UNPREDICTABLE(a6 != zero) {
157 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
158 } else {
159 a6 = zero_data;
160 }
161 a += 7;
162
163 size_t k = kc;
164
165 while (k >= 8 * sizeof(int8_t)) {
166 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
167 a0 += 8;
168 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
169 a1 += 8;
170 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
171 a2 += 8;
172 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
173 a3 += 8;
174 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
175 a4 += 8;
176 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
177 a5 += 8;
178 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
179 a6 += 8;
180
181 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
182 xnn_prefetch_to_l1((const int8_t*) w + 896);
183
184 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
185 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
186 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
187 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
188 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
189 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
190 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
191 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
192
193 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
194 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
195 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
196 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
197 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
198 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
199 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
200 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
201 xnn_prefetch_to_l1((const int8_t*) w + 960);
202
203 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
204 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
205 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
206 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
207 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
208 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
209 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
210 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
211
212 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
213 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
214 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
215 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
216 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
217 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
218 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
219
220 w = (const int8_t*) w + 128;
221 k -= 8 * sizeof(int8_t);
222 }
223 p -= 7 * sizeof(void*);
224 } while (p != 0);
225
226 // Add 4 adjacent sums
227 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
228 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
229 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
230 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
231 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
232 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
233 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
234 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
235 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
236 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
237 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
238 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
239 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
240 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
241 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
242 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
243 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
244 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
245 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
246 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
247 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
248 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
249 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
250 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
251 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
252 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
253 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
254 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
255 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
256
257 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
258 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
259 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
260 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
261 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
262 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
263 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
264
265 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
266 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
267 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
268 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
269 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
270 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
271 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
272
273 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
274 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
275 w = (const float*) w + 32;
276
277 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
278 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
279 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
280 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
281 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
282 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
283 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
284
285 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
286 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
287 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
288 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
289 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
290 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
291 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
292
293 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
294 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
295 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
296 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
297 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
298 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
299 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
300
301 if XNN_LIKELY(nc >= 16) {
302 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
303 c6 = (float*) ((uintptr_t) c6 + cn_stride);
304 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
305 c5 = (float*) ((uintptr_t) c5 + cn_stride);
306 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
307 c4 = (float*) ((uintptr_t) c4 + cn_stride);
308 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
309 c3 = (float*) ((uintptr_t) c3 + cn_stride);
310 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
311 c2 = (float*) ((uintptr_t) c2 + cn_stride);
312 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
313 c1 = (float*) ((uintptr_t) c1 + cn_stride);
314 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
315 c0 = (float*) ((uintptr_t) c0 + cn_stride);
316 a = (const int8_t**restrict) ((uintptr_t) a - ks);
317 nc -= 16;
318 } else {
319 // Prepare mask for valid 32-bit elements (depends on nc).
320 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
321 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
322 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
323 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
324 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
325 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
326 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
327 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
328 nc = 0;
329 }
330 } while (nc != 0);
331 }