Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-8x16c8-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_8x16c8__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 <= 8);
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 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
73 if XNN_UNPREDICTABLE(mr != 8) {
74 c7 = c6;
75 }
76
77 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
78 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point);
79 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
80 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
81 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
82 // XNN_FORCE_REALIZATION(vinput_zero_point);
83 // XNN_FORCE_REALIZATION(vinput_inv_scale);
84 // XNN_FORCE_REALIZATION(voutput_min);
85 // XNN_FORCE_REALIZATION(voutput_max);
86 do {
87 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
88 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
89 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
90 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
91 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point);
92 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point);
93 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point);
94 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point);
95 __m512i vacc1x0123 = vacc0x0123;
96 __m512i vacc1x4567 = vacc0x4567;
97 __m512i vacc1x89AB = vacc0x89AB;
98 __m512i vacc1xCDEF = vacc0xCDEF;
99 __m512i vacc2x0123 = vacc0x0123;
100 __m512i vacc2x4567 = vacc0x4567;
101 __m512i vacc2x89AB = vacc0x89AB;
102 __m512i vacc2xCDEF = vacc0xCDEF;
103 __m512i vacc3x0123 = vacc0x0123;
104 __m512i vacc3x4567 = vacc0x4567;
105 __m512i vacc3x89AB = vacc0x89AB;
106 __m512i vacc3xCDEF = vacc0xCDEF;
107 __m512i vacc4x0123 = vacc0x0123;
108 __m512i vacc4x4567 = vacc0x4567;
109 __m512i vacc4x89AB = vacc0x89AB;
110 __m512i vacc4xCDEF = vacc0xCDEF;
111 __m512i vacc5x0123 = vacc0x0123;
112 __m512i vacc5x4567 = vacc0x4567;
113 __m512i vacc5x89AB = vacc0x89AB;
114 __m512i vacc5xCDEF = vacc0xCDEF;
115 __m512i vacc6x0123 = vacc0x0123;
116 __m512i vacc6x4567 = vacc0x4567;
117 __m512i vacc6x89AB = vacc0x89AB;
118 __m512i vacc6xCDEF = vacc0xCDEF;
119 __m512i vacc7x0123 = vacc0x0123;
120 __m512i vacc7x4567 = vacc0x4567;
121 __m512i vacc7x89AB = vacc0x89AB;
122 __m512i vacc7xCDEF = vacc0xCDEF;
123 w = (const int32_t*) w + 16;
124
125 size_t p = ks;
126 do {
127 const int8_t* restrict a0 = a[0];
128 if XNN_UNPREDICTABLE(a0 != zero) {
129 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
130 } else {
131 a0 = zero_data;
132 }
133 const int8_t* restrict a1 = a[1];
134 if XNN_UNPREDICTABLE(a1 != zero) {
135 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
136 } else {
137 a1 = zero_data;
138 }
139 const int8_t* restrict a2 = a[2];
140 if XNN_UNPREDICTABLE(a2 != zero) {
141 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
142 } else {
143 a2 = zero_data;
144 }
145 const int8_t* restrict a3 = a[3];
146 if XNN_UNPREDICTABLE(a3 != zero) {
147 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
148 } else {
149 a3 = zero_data;
150 }
151 const int8_t* restrict a4 = a[4];
152 if XNN_UNPREDICTABLE(a4 != zero) {
153 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
154 } else {
155 a4 = zero_data;
156 }
157 const int8_t* restrict a5 = a[5];
158 if XNN_UNPREDICTABLE(a5 != zero) {
159 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
160 } else {
161 a5 = zero_data;
162 }
163 const int8_t* restrict a6 = a[6];
164 if XNN_UNPREDICTABLE(a6 != zero) {
165 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
166 } else {
167 a6 = zero_data;
168 }
169 const int8_t* restrict a7 = a[7];
170 if XNN_UNPREDICTABLE(a7 != zero) {
171 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
172 } else {
173 a7 = zero_data;
174 }
175 a += 8;
176
177 size_t k = kc;
178
179 while (k >= 8 * sizeof(int8_t)) {
180 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
181 a0 += 8;
182 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
183 a1 += 8;
184 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
185 a2 += 8;
186 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
187 a3 += 8;
188 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
189 a4 += 8;
190 const __m512i va5 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a5)));
191 a5 += 8;
192 const __m512i va6 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a6)));
193 a6 += 8;
194 const __m512i va7 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a7)));
195 a7 += 8;
196
197 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
198 xnn_prefetch_to_l1((const int8_t*) w + 896);
199
200 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
201 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
202 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
203 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
204 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
205 vacc5x0123 = _mm512_add_epi32(vacc5x0123, _mm512_madd_epi16(va5, vb0123));
206 vacc6x0123 = _mm512_add_epi32(vacc6x0123, _mm512_madd_epi16(va6, vb0123));
207 vacc7x0123 = _mm512_add_epi32(vacc7x0123, _mm512_madd_epi16(va7, vb0123));
208 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
209
210 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
211 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
212 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
213 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
214 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
215 vacc5x4567 = _mm512_add_epi32(vacc5x4567, _mm512_madd_epi16(va5, vb4567));
216 vacc6x4567 = _mm512_add_epi32(vacc6x4567, _mm512_madd_epi16(va6, vb4567));
217 vacc7x4567 = _mm512_add_epi32(vacc7x4567, _mm512_madd_epi16(va7, vb4567));
218 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
219 xnn_prefetch_to_l1((const int8_t*) w + 960);
220
221 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
222 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
223 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
224 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
225 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
226 vacc5x89AB = _mm512_add_epi32(vacc5x89AB, _mm512_madd_epi16(va5, vb89AB));
227 vacc6x89AB = _mm512_add_epi32(vacc6x89AB, _mm512_madd_epi16(va6, vb89AB));
228 vacc7x89AB = _mm512_add_epi32(vacc7x89AB, _mm512_madd_epi16(va7, vb89AB));
229 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
230
231 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
232 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
233 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
234 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
235 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
236 vacc5xCDEF = _mm512_add_epi32(vacc5xCDEF, _mm512_madd_epi16(va5, vbCDEF));
237 vacc6xCDEF = _mm512_add_epi32(vacc6xCDEF, _mm512_madd_epi16(va6, vbCDEF));
238 vacc7xCDEF = _mm512_add_epi32(vacc7xCDEF, _mm512_madd_epi16(va7, vbCDEF));
239
240 w = (const int8_t*) w + 128;
241 k -= 8 * sizeof(int8_t);
242 }
243 p -= 8 * sizeof(void*);
244 } while (p != 0);
245
246 // Add 4 adjacent sums
247 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
248 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
249 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
250 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
251 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
252 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
253 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
254 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
255 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
256 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
257 const __m512i vacc5x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x0123, vacc5x4567), _mm512_unpackhi_epi32(vacc5x0123, vacc5x4567));
258 const __m512i vacc5x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x89AB, vacc5xCDEF), _mm512_unpackhi_epi32(vacc5x89AB, vacc5xCDEF));
259 const __m512i vacc6x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x0123, vacc6x4567), _mm512_unpackhi_epi32(vacc6x0123, vacc6x4567));
260 const __m512i vacc6x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x89AB, vacc6xCDEF), _mm512_unpackhi_epi32(vacc6x89AB, vacc6xCDEF));
261 const __m512i vacc7x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x0123, vacc7x4567), _mm512_unpackhi_epi32(vacc7x0123, vacc7x4567));
262 const __m512i vacc7x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x89AB, vacc7xCDEF), _mm512_unpackhi_epi32(vacc7x89AB, vacc7xCDEF));
263 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
264 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
265 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
266 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
267 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
268 const __m512i vacc5x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc5x04152637, vacc5x8C9DAEBF), _mm512_unpackhi_epi32(vacc5x04152637, vacc5x8C9DAEBF));
269 const __m512i vacc6x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc6x04152637, vacc6x8C9DAEBF), _mm512_unpackhi_epi32(vacc6x04152637, vacc6x8C9DAEBF));
270 const __m512i vacc7x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc7x04152637, vacc7x8C9DAEBF), _mm512_unpackhi_epi32(vacc7x04152637, vacc7x8C9DAEBF));
271 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
272 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
273 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
274 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
275 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
276 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
277 __m512i vacc5x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc5x084C195D2A6E3B7F);
278 __m512i vacc6x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc6x084C195D2A6E3B7F);
279 __m512i vacc7x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc7x084C195D2A6E3B7F);
280
281 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
282 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
283 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
284 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
285 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
286 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
287 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
288 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
289
290 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
291 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
292 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
293 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
294 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
295 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
296 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
297 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vinput_inv_scale);
298
299 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
300 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
301 w = (const float*) w + 32;
302
303 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
304 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
305 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
306 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
307 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
308 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
309 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
310 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
311
312 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
313 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
314 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
315 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
316 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
317 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
318 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
319 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
320
321 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
322 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
323 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
324 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
325 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
326 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
327 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
328 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
329
330 if XNN_LIKELY(nc >= 16) {
331 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
332 c7 = (float*) ((uintptr_t) c7 + cn_stride);
333 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
334 c6 = (float*) ((uintptr_t) c6 + cn_stride);
335 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
336 c5 = (float*) ((uintptr_t) c5 + cn_stride);
337 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
338 c4 = (float*) ((uintptr_t) c4 + cn_stride);
339 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
340 c3 = (float*) ((uintptr_t) c3 + cn_stride);
341 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
342 c2 = (float*) ((uintptr_t) c2 + cn_stride);
343 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
344 c1 = (float*) ((uintptr_t) c1 + cn_stride);
345 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
346 c0 = (float*) ((uintptr_t) c0 + cn_stride);
347 a = (const int8_t**restrict) ((uintptr_t) a - ks);
348 nc -= 16;
349 } else {
350 // Prepare mask for valid 32-bit elements (depends on nc).
351 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
352 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
353 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
354 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
355 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
356 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
357 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
358 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
359 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
360 nc = 0;
361 }
362 } while (nc != 0);
363 }