blob: 6061caf6a9d05c88a746f20cbe4319ac6b3c9b67 [file] [log] [blame]
Marat Dukhan479f87e2019-11-27 15:17:06 -08001// Auto-generated file. Do not edit!
2// Template: src/f32-dwconv/up-avx512.c.in
3// Generator: tools/xngen
4//
5// Copyright 2019 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/dwconv.h>
Marat Dukhancfb31342019-12-05 10:42:57 -080015#include <xnnpack/intrinsics-polyfill.h>
Marat Dukhan479f87e2019-11-27 15:17:06 -080016
17
18void xnn_f32_dwconv_ukernel_up32x25__avx512f(
19 size_t channels,
20 size_t output_width,
21 const float** input,
22 const float* weights,
23 float* output,
24 size_t input_stride,
25 size_t output_increment,
Marat Dukhaneb09a6b2020-04-08 17:34:32 -070026 const union xnn_f32_minmax_params params[restrict static 1])
Marat Dukhan479f87e2019-11-27 15:17:06 -080027{
28 assert(channels != 0);
29 assert(output_width != 0);
30
31 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
32 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
33 do {
34 const float* i0 = input[0];
Marat Dukhan68660992020-02-03 13:31:12 -080035 assert(i0 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080036 const float* i1 = input[1];
Marat Dukhan68660992020-02-03 13:31:12 -080037 assert(i1 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080038 const float* i2 = input[2];
Marat Dukhan68660992020-02-03 13:31:12 -080039 assert(i2 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080040 const float* i3 = input[3];
Marat Dukhan68660992020-02-03 13:31:12 -080041 assert(i3 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080042 const float* i4 = input[4];
Marat Dukhan68660992020-02-03 13:31:12 -080043 assert(i4 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080044 const float* i5 = input[5];
Marat Dukhan68660992020-02-03 13:31:12 -080045 assert(i5 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080046 const float* i6 = input[6];
Marat Dukhan68660992020-02-03 13:31:12 -080047 assert(i6 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080048 const float* i7 = input[7];
Marat Dukhan68660992020-02-03 13:31:12 -080049 assert(i7 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080050 const float* i8 = input[8];
Marat Dukhan68660992020-02-03 13:31:12 -080051 assert(i8 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080052 const float* i9 = input[9];
Marat Dukhan68660992020-02-03 13:31:12 -080053 assert(i9 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080054 const float* i10 = input[10];
Marat Dukhan68660992020-02-03 13:31:12 -080055 assert(i10 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080056 const float* i11 = input[11];
Marat Dukhan68660992020-02-03 13:31:12 -080057 assert(i11 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080058 const float* i12 = input[12];
Marat Dukhan68660992020-02-03 13:31:12 -080059 assert(i12 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080060 const float* i13 = input[13];
Marat Dukhan68660992020-02-03 13:31:12 -080061 assert(i13 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080062 const float* i14 = input[14];
Marat Dukhan68660992020-02-03 13:31:12 -080063 assert(i14 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080064 const float* i15 = input[15];
Marat Dukhan68660992020-02-03 13:31:12 -080065 assert(i15 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080066 const float* i16 = input[16];
Marat Dukhan68660992020-02-03 13:31:12 -080067 assert(i16 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080068 const float* i17 = input[17];
Marat Dukhan68660992020-02-03 13:31:12 -080069 assert(i17 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080070 const float* i18 = input[18];
Marat Dukhan68660992020-02-03 13:31:12 -080071 assert(i18 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080072 const float* i19 = input[19];
Marat Dukhan68660992020-02-03 13:31:12 -080073 assert(i19 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080074 const float* i20 = input[20];
Marat Dukhan68660992020-02-03 13:31:12 -080075 assert(i20 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080076 const float* i21 = input[21];
Marat Dukhan68660992020-02-03 13:31:12 -080077 assert(i21 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080078 const float* i22 = input[22];
Marat Dukhan68660992020-02-03 13:31:12 -080079 assert(i22 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080080 const float* i23 = input[23];
Marat Dukhan68660992020-02-03 13:31:12 -080081 assert(i23 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080082 const float* i24 = input[24];
Marat Dukhan68660992020-02-03 13:31:12 -080083 assert(i24 != NULL);
Marat Dukhan479f87e2019-11-27 15:17:06 -080084 input = (const float**) ((uintptr_t) input + input_stride);
85
86 size_t c = channels;
87 const float* w = weights;
88 for (; c >= 32; c -= 32) {
89 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
90 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16);
91
92
93 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
94 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16);
95 i0 += 32;
96
97 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
98 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48);
99 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
100 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
101
102 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
103 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16);
104 i1 += 32;
105
106 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
107 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80);
108 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
109 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
110
111 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
112 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16);
113 i2 += 32;
114
115 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
116 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112);
117 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
118 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
119
120 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
121 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16);
122 i3 += 32;
123
124 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
125 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144);
126 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
127 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
128
129 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4);
130 const __m512 vi4xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i4 + 16);
131 i4 += 32;
132
133 const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 160);
134 const __m512 vk4xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 176);
135 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
136 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
137
138 const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5);
139 const __m512 vi5xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i5 + 16);
140 i5 += 32;
141
142 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192);
143 const __m512 vk5xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 208);
144 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
145 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
146
147 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6);
148 const __m512 vi6xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i6 + 16);
149 i6 += 32;
150
151 const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 224);
152 const __m512 vk6xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 240);
153 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
154 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
155
156 const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7);
157 const __m512 vi7xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i7 + 16);
158 i7 += 32;
159
160 const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 256);
161 const __m512 vk7xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 272);
162 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
163 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
164
165 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8);
166 const __m512 vi8xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i8 + 16);
167 i8 += 32;
168
169 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288);
170 const __m512 vk8xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 304);
171 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
172 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
173
174 const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9);
175 const __m512 vi9xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i9 + 16);
176 i9 += 32;
177
178 const __m512 vk9x0123456789ABCDEF = _mm512_load_ps(w + 320);
179 const __m512 vk9xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 336);
180 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
181 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi9xGHIJKLMNOPQRSTUV, vk9xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
182
183 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10);
184 const __m512 vi10xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i10 + 16);
185 i10 += 32;
186
187 const __m512 vk10x0123456789ABCDEF = _mm512_load_ps(w + 352);
188 const __m512 vk10xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 368);
189 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
190 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi10xGHIJKLMNOPQRSTUV, vk10xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
191
192 const __m512 vi11x0123456789ABCDEF = _mm512_loadu_ps(i11);
193 const __m512 vi11xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i11 + 16);
194 i11 += 32;
195
196 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384);
197 const __m512 vk11xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 400);
198 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
199 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi11xGHIJKLMNOPQRSTUV, vk11xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
200
201 const __m512 vi12x0123456789ABCDEF = _mm512_loadu_ps(i12);
202 const __m512 vi12xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i12 + 16);
203 i12 += 32;
204
205 const __m512 vk12x0123456789ABCDEF = _mm512_load_ps(w + 416);
206 const __m512 vk12xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 432);
207 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
208 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi12xGHIJKLMNOPQRSTUV, vk12xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
209
210 const __m512 vi13x0123456789ABCDEF = _mm512_loadu_ps(i13);
211 const __m512 vi13xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i13 + 16);
212 i13 += 32;
213
214 const __m512 vk13x0123456789ABCDEF = _mm512_load_ps(w + 448);
215 const __m512 vk13xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 464);
216 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
217 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi13xGHIJKLMNOPQRSTUV, vk13xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
218
219 const __m512 vi14x0123456789ABCDEF = _mm512_loadu_ps(i14);
220 const __m512 vi14xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i14 + 16);
221 i14 += 32;
222
223 const __m512 vk14x0123456789ABCDEF = _mm512_load_ps(w + 480);
224 const __m512 vk14xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 496);
225 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
226 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi14xGHIJKLMNOPQRSTUV, vk14xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
227
228 const __m512 vi15x0123456789ABCDEF = _mm512_loadu_ps(i15);
229 const __m512 vi15xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i15 + 16);
230 i15 += 32;
231
232 const __m512 vk15x0123456789ABCDEF = _mm512_load_ps(w + 512);
233 const __m512 vk15xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 528);
234 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
235 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi15xGHIJKLMNOPQRSTUV, vk15xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
236
237 const __m512 vi16x0123456789ABCDEF = _mm512_loadu_ps(i16);
238 const __m512 vi16xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i16 + 16);
239 i16 += 32;
240
241 const __m512 vk16x0123456789ABCDEF = _mm512_load_ps(w + 544);
242 const __m512 vk16xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 560);
243 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
244 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi16xGHIJKLMNOPQRSTUV, vk16xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
245
246 const __m512 vi17x0123456789ABCDEF = _mm512_loadu_ps(i17);
247 const __m512 vi17xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i17 + 16);
248 i17 += 32;
249
250 const __m512 vk17x0123456789ABCDEF = _mm512_load_ps(w + 576);
251 const __m512 vk17xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 592);
252 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
253 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi17xGHIJKLMNOPQRSTUV, vk17xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
254
255 const __m512 vi18x0123456789ABCDEF = _mm512_loadu_ps(i18);
256 const __m512 vi18xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i18 + 16);
257 i18 += 32;
258
259 const __m512 vk18x0123456789ABCDEF = _mm512_load_ps(w + 608);
260 const __m512 vk18xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 624);
261 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
262 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi18xGHIJKLMNOPQRSTUV, vk18xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
263
264 const __m512 vi19x0123456789ABCDEF = _mm512_loadu_ps(i19);
265 const __m512 vi19xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i19 + 16);
266 i19 += 32;
267
268 const __m512 vk19x0123456789ABCDEF = _mm512_load_ps(w + 640);
269 const __m512 vk19xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 656);
270 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
271 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi19xGHIJKLMNOPQRSTUV, vk19xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
272
273 const __m512 vi20x0123456789ABCDEF = _mm512_loadu_ps(i20);
274 const __m512 vi20xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i20 + 16);
275 i20 += 32;
276
277 const __m512 vk20x0123456789ABCDEF = _mm512_load_ps(w + 672);
278 const __m512 vk20xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 688);
279 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
280 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi20xGHIJKLMNOPQRSTUV, vk20xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
281
282 const __m512 vi21x0123456789ABCDEF = _mm512_loadu_ps(i21);
283 const __m512 vi21xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i21 + 16);
284 i21 += 32;
285
286 const __m512 vk21x0123456789ABCDEF = _mm512_load_ps(w + 704);
287 const __m512 vk21xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 720);
288 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
289 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi21xGHIJKLMNOPQRSTUV, vk21xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
290
291 const __m512 vi22x0123456789ABCDEF = _mm512_loadu_ps(i22);
292 const __m512 vi22xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i22 + 16);
293 i22 += 32;
294
295 const __m512 vk22x0123456789ABCDEF = _mm512_load_ps(w + 736);
296 const __m512 vk22xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 752);
297 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
298 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi22xGHIJKLMNOPQRSTUV, vk22xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
299
300 const __m512 vi23x0123456789ABCDEF = _mm512_loadu_ps(i23);
301 const __m512 vi23xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i23 + 16);
302 i23 += 32;
303
304 const __m512 vk23x0123456789ABCDEF = _mm512_load_ps(w + 768);
305 const __m512 vk23xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 784);
306 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
307 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi23xGHIJKLMNOPQRSTUV, vk23xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
308
309 const __m512 vi24x0123456789ABCDEF = _mm512_loadu_ps(i24);
310 const __m512 vi24xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i24 + 16);
311 i24 += 32;
312
313 const __m512 vk24x0123456789ABCDEF = _mm512_load_ps(w + 800);
314 const __m512 vk24xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 816);
315 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
316 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi24xGHIJKLMNOPQRSTUV, vk24xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
317
318 w += 832;
319
320
321 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
322 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin);
323 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
324 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax);
325
326 _mm512_storeu_ps(output, vacc0123456789ABCDEF);
327 _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV);
328 output += 32;
329 }
330 for (; c >= 16; c -= 16) {
331 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
332
333 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
334 i0 += 16;
335
336 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
337 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
338
339 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
340 i1 += 16;
341
342 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
343 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
344
345 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
346 i2 += 16;
347
348 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
349 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
350
351 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
352 i3 += 16;
353
354 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
355 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
356
357 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4);
358 i4 += 16;
359
360 const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 160);
361 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
362
363 const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5);
364 i5 += 16;
365
366 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192);
367 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
368
369 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6);
370 i6 += 16;
371
372 const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 224);
373 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
374
375 const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7);
376 i7 += 16;
377
378 const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 256);
379 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
380
381 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8);
382 i8 += 16;
383
384 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288);
385 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
386
387 const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9);
388 i9 += 16;
389
390 const __m512 vk9x0123456789ABCDEF = _mm512_load_ps(w + 320);
391 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
392
393 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10);
394 i10 += 16;
395
396 const __m512 vk10x0123456789ABCDEF = _mm512_load_ps(w + 352);
397 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
398
399 const __m512 vi11x0123456789ABCDEF = _mm512_loadu_ps(i11);
400 i11 += 16;
401
402 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384);
403 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
404
405 const __m512 vi12x0123456789ABCDEF = _mm512_loadu_ps(i12);
406 i12 += 16;
407
408 const __m512 vk12x0123456789ABCDEF = _mm512_load_ps(w + 416);
409 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
410
411 const __m512 vi13x0123456789ABCDEF = _mm512_loadu_ps(i13);
412 i13 += 16;
413
414 const __m512 vk13x0123456789ABCDEF = _mm512_load_ps(w + 448);
415 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
416
417 const __m512 vi14x0123456789ABCDEF = _mm512_loadu_ps(i14);
418 i14 += 16;
419
420 const __m512 vk14x0123456789ABCDEF = _mm512_load_ps(w + 480);
421 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
422
423 const __m512 vi15x0123456789ABCDEF = _mm512_loadu_ps(i15);
424 i15 += 16;
425
426 const __m512 vk15x0123456789ABCDEF = _mm512_load_ps(w + 512);
427 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
428
429 const __m512 vi16x0123456789ABCDEF = _mm512_loadu_ps(i16);
430 i16 += 16;
431
432 const __m512 vk16x0123456789ABCDEF = _mm512_load_ps(w + 544);
433 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
434
435 const __m512 vi17x0123456789ABCDEF = _mm512_loadu_ps(i17);
436 i17 += 16;
437
438 const __m512 vk17x0123456789ABCDEF = _mm512_load_ps(w + 576);
439 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
440
441 const __m512 vi18x0123456789ABCDEF = _mm512_loadu_ps(i18);
442 i18 += 16;
443
444 const __m512 vk18x0123456789ABCDEF = _mm512_load_ps(w + 608);
445 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
446
447 const __m512 vi19x0123456789ABCDEF = _mm512_loadu_ps(i19);
448 i19 += 16;
449
450 const __m512 vk19x0123456789ABCDEF = _mm512_load_ps(w + 640);
451 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
452
453 const __m512 vi20x0123456789ABCDEF = _mm512_loadu_ps(i20);
454 i20 += 16;
455
456 const __m512 vk20x0123456789ABCDEF = _mm512_load_ps(w + 672);
457 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
458
459 const __m512 vi21x0123456789ABCDEF = _mm512_loadu_ps(i21);
460 i21 += 16;
461
462 const __m512 vk21x0123456789ABCDEF = _mm512_load_ps(w + 704);
463 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
464
465 const __m512 vi22x0123456789ABCDEF = _mm512_loadu_ps(i22);
466 i22 += 16;
467
468 const __m512 vk22x0123456789ABCDEF = _mm512_load_ps(w + 736);
469 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
470
471 const __m512 vi23x0123456789ABCDEF = _mm512_loadu_ps(i23);
472 i23 += 16;
473
474 const __m512 vk23x0123456789ABCDEF = _mm512_load_ps(w + 768);
475 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
476
477 const __m512 vi24x0123456789ABCDEF = _mm512_loadu_ps(i24);
478 i24 += 16;
479
480 const __m512 vk24x0123456789ABCDEF = _mm512_load_ps(w + 800);
481 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
482
483 w += 16;
484
485
486 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
487 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
488
489 _mm512_storeu_ps(output, vacc0123456789ABCDEF);
490 output += 16;
491 }
492 if XNN_UNLIKELY(c != 0) {
493 assert(c >= 1);
494 assert(c <= 16);
495 // Prepare mask for valid 32-bit elements (depends on nc).
496 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));
497
Marat Dukhan90dff802020-02-10 17:23:53 -0800498 __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800499
500 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0);
Marat Dukhan90dff802020-02-10 17:23:53 -0800501 const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800502 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
503
504 const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1);
Marat Dukhan90dff802020-02-10 17:23:53 -0800505 const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800506 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
507
508 const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2);
Marat Dukhan90dff802020-02-10 17:23:53 -0800509 const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800510 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
511
512 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3);
Marat Dukhan90dff802020-02-10 17:23:53 -0800513 const __m512 vk3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 128);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800514 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
515
516 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4);
Marat Dukhan90dff802020-02-10 17:23:53 -0800517 const __m512 vk4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 160);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800518 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
519
520 const __m512 vi5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i5);
Marat Dukhan90dff802020-02-10 17:23:53 -0800521 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800522 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
523
524 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6);
Marat Dukhan90dff802020-02-10 17:23:53 -0800525 const __m512 vk6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 224);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800526 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
527
528 const __m512 vi7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i7);
Marat Dukhan90dff802020-02-10 17:23:53 -0800529 const __m512 vk7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 256);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800530 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
531
532 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8);
Marat Dukhan90dff802020-02-10 17:23:53 -0800533 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 288);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800534 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
535
536 const __m512 vi9x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i9);
Marat Dukhan90dff802020-02-10 17:23:53 -0800537 const __m512 vk9x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 320);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800538 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
539
540 const __m512 vi10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i10);
Marat Dukhan90dff802020-02-10 17:23:53 -0800541 const __m512 vk10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 352);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800542 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
543
544 const __m512 vi11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i11);
Marat Dukhan90dff802020-02-10 17:23:53 -0800545 const __m512 vk11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 384);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800546 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
547
548 const __m512 vi12x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i12);
Marat Dukhan90dff802020-02-10 17:23:53 -0800549 const __m512 vk12x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 416);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800550 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
551
552 const __m512 vi13x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i13);
Marat Dukhan90dff802020-02-10 17:23:53 -0800553 const __m512 vk13x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 448);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800554 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
555
556 const __m512 vi14x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i14);
Marat Dukhan90dff802020-02-10 17:23:53 -0800557 const __m512 vk14x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 480);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800558 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
559
560 const __m512 vi15x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i15);
Marat Dukhan90dff802020-02-10 17:23:53 -0800561 const __m512 vk15x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 512);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800562 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
563
564 const __m512 vi16x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i16);
Marat Dukhan90dff802020-02-10 17:23:53 -0800565 const __m512 vk16x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 544);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800566 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
567
568 const __m512 vi17x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i17);
Marat Dukhan90dff802020-02-10 17:23:53 -0800569 const __m512 vk17x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 576);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800570 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
571
572 const __m512 vi18x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i18);
Marat Dukhan90dff802020-02-10 17:23:53 -0800573 const __m512 vk18x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 608);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800574 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
575
576 const __m512 vi19x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i19);
Marat Dukhan90dff802020-02-10 17:23:53 -0800577 const __m512 vk19x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 640);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800578 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
579
580 const __m512 vi20x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i20);
Marat Dukhan90dff802020-02-10 17:23:53 -0800581 const __m512 vk20x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 672);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800582 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
583
584 const __m512 vi21x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i21);
Marat Dukhan90dff802020-02-10 17:23:53 -0800585 const __m512 vk21x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 704);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800586 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
587
588 const __m512 vi22x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i22);
Marat Dukhan90dff802020-02-10 17:23:53 -0800589 const __m512 vk22x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 736);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800590 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
591
592 const __m512 vi23x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i23);
Marat Dukhan90dff802020-02-10 17:23:53 -0800593 const __m512 vk23x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 768);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800594 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
595
596 const __m512 vi24x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i24);
Marat Dukhan90dff802020-02-10 17:23:53 -0800597 const __m512 vk24x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 800);
Marat Dukhan479f87e2019-11-27 15:17:06 -0800598 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
599
600
601 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
602 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
603
604 _mm512_mask_storeu_ps(output, vmask, vacc0123456789ABCDEF);
605 output += c;
606 }
607
608 output = (float*) ((uintptr_t) output + output_increment);
609 } while (--output_width != 0);
610}