• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv/up-psimd.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 <psimd.h>
13 
14 #include <xnnpack/dwconv.h>
15 
16 
xnn_f32_dwconv_ukernel_up8x25__psimd(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_dwconv_ukernel_up8x25__psimd(
18     size_t channels,
19     size_t output_width,
20     const float** input,
21     const float* weights,
22     float* output,
23     size_t input_stride,
24     size_t output_increment,
25     const union xnn_f32_output_params params[restrict static 1])
26 {
27   assert(channels != 0);
28   assert(output_width != 0);
29 
30   const psimd_f32 vmax = psimd_load_splat_f32(&params->scalar.max);
31   const psimd_f32 vmin = psimd_load_splat_f32(&params->scalar.min);
32   do {
33     const float* i0 = input[0];
34     assert(i0 != NULL);
35     const float* i1 = input[1];
36     assert(i1 != NULL);
37     const float* i2 = input[2];
38     assert(i2 != NULL);
39     const float* i3 = input[3];
40     assert(i3 != NULL);
41     const float* i4 = input[4];
42     assert(i4 != NULL);
43     const float* i5 = input[5];
44     assert(i5 != NULL);
45     const float* i6 = input[6];
46     assert(i6 != NULL);
47     const float* i7 = input[7];
48     assert(i7 != NULL);
49     const float* i8 = input[8];
50     assert(i8 != NULL);
51     const float* i9 = input[9];
52     assert(i9 != NULL);
53     const float* i10 = input[10];
54     assert(i10 != NULL);
55     const float* i11 = input[11];
56     assert(i11 != NULL);
57     const float* i12 = input[12];
58     assert(i12 != NULL);
59     const float* i13 = input[13];
60     assert(i13 != NULL);
61     const float* i14 = input[14];
62     assert(i14 != NULL);
63     const float* i15 = input[15];
64     assert(i15 != NULL);
65     const float* i16 = input[16];
66     assert(i16 != NULL);
67     const float* i17 = input[17];
68     assert(i17 != NULL);
69     const float* i18 = input[18];
70     assert(i18 != NULL);
71     const float* i19 = input[19];
72     assert(i19 != NULL);
73     const float* i20 = input[20];
74     assert(i20 != NULL);
75     const float* i21 = input[21];
76     assert(i21 != NULL);
77     const float* i22 = input[22];
78     assert(i22 != NULL);
79     const float* i23 = input[23];
80     assert(i23 != NULL);
81     const float* i24 = input[24];
82     assert(i24 != NULL);
83     input = (const float**) ((uintptr_t) input + input_stride);
84 
85     size_t c = channels;
86     const float* w = weights;
87     for (; c >= 8; c -= 8) {
88       psimd_f32 vacc0123p0 = psimd_load_f32(w);
89       psimd_f32 vacc4567p0 = psimd_load_f32(w + 4);
90 
91 
92       const psimd_f32 vi0x0123 = psimd_load_f32(i0);
93       const psimd_f32 vi0x4567 = psimd_load_f32(i0 + 4);
94       i0 += 8;
95 
96       const psimd_f32 vk0x0123 = psimd_load_f32(w + 8);
97       const psimd_f32 vk0x4567 = psimd_load_f32(w + 12);
98       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123);
99       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi0x4567, vk0x4567);
100 
101       const psimd_f32 vi1x0123 = psimd_load_f32(i1);
102       const psimd_f32 vi1x4567 = psimd_load_f32(i1 + 4);
103       i1 += 8;
104 
105       const psimd_f32 vk1x0123 = psimd_load_f32(w + 16);
106       const psimd_f32 vk1x4567 = psimd_load_f32(w + 20);
107       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi1x0123, vk1x0123);
108       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi1x4567, vk1x4567);
109 
110       const psimd_f32 vi2x0123 = psimd_load_f32(i2);
111       const psimd_f32 vi2x4567 = psimd_load_f32(i2 + 4);
112       i2 += 8;
113 
114       const psimd_f32 vk2x0123 = psimd_load_f32(w + 24);
115       const psimd_f32 vk2x4567 = psimd_load_f32(w + 28);
116       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi2x0123, vk2x0123);
117       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi2x4567, vk2x4567);
118 
119       const psimd_f32 vi3x0123 = psimd_load_f32(i3);
120       const psimd_f32 vi3x4567 = psimd_load_f32(i3 + 4);
121       i3 += 8;
122 
123       const psimd_f32 vk3x0123 = psimd_load_f32(w + 32);
124       const psimd_f32 vk3x4567 = psimd_load_f32(w + 36);
125       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi3x0123, vk3x0123);
126       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi3x4567, vk3x4567);
127 
128       const psimd_f32 vi4x0123 = psimd_load_f32(i4);
129       const psimd_f32 vi4x4567 = psimd_load_f32(i4 + 4);
130       i4 += 8;
131 
132       const psimd_f32 vk4x0123 = psimd_load_f32(w + 40);
133       const psimd_f32 vk4x4567 = psimd_load_f32(w + 44);
134       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi4x0123, vk4x0123);
135       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi4x4567, vk4x4567);
136 
137       const psimd_f32 vi5x0123 = psimd_load_f32(i5);
138       const psimd_f32 vi5x4567 = psimd_load_f32(i5 + 4);
139       i5 += 8;
140 
141       const psimd_f32 vk5x0123 = psimd_load_f32(w + 48);
142       const psimd_f32 vk5x4567 = psimd_load_f32(w + 52);
143       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi5x0123, vk5x0123);
144       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi5x4567, vk5x4567);
145 
146       const psimd_f32 vi6x0123 = psimd_load_f32(i6);
147       const psimd_f32 vi6x4567 = psimd_load_f32(i6 + 4);
148       i6 += 8;
149 
150       const psimd_f32 vk6x0123 = psimd_load_f32(w + 56);
151       const psimd_f32 vk6x4567 = psimd_load_f32(w + 60);
152       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi6x0123, vk6x0123);
153       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi6x4567, vk6x4567);
154 
155       const psimd_f32 vi7x0123 = psimd_load_f32(i7);
156       const psimd_f32 vi7x4567 = psimd_load_f32(i7 + 4);
157       i7 += 8;
158 
159       const psimd_f32 vk7x0123 = psimd_load_f32(w + 64);
160       const psimd_f32 vk7x4567 = psimd_load_f32(w + 68);
161       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi7x0123, vk7x0123);
162       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi7x4567, vk7x4567);
163 
164       const psimd_f32 vi8x0123 = psimd_load_f32(i8);
165       const psimd_f32 vi8x4567 = psimd_load_f32(i8 + 4);
166       i8 += 8;
167 
168       const psimd_f32 vk8x0123 = psimd_load_f32(w + 72);
169       const psimd_f32 vk8x4567 = psimd_load_f32(w + 76);
170       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi8x0123, vk8x0123);
171       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi8x4567, vk8x4567);
172 
173       const psimd_f32 vi9x0123 = psimd_load_f32(i9);
174       const psimd_f32 vi9x4567 = psimd_load_f32(i9 + 4);
175       i9 += 8;
176 
177       const psimd_f32 vk9x0123 = psimd_load_f32(w + 80);
178       const psimd_f32 vk9x4567 = psimd_load_f32(w + 84);
179       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi9x0123, vk9x0123);
180       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi9x4567, vk9x4567);
181 
182       const psimd_f32 vi10x0123 = psimd_load_f32(i10);
183       const psimd_f32 vi10x4567 = psimd_load_f32(i10 + 4);
184       i10 += 8;
185 
186       const psimd_f32 vk10x0123 = psimd_load_f32(w + 88);
187       const psimd_f32 vk10x4567 = psimd_load_f32(w + 92);
188       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi10x0123, vk10x0123);
189       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi10x4567, vk10x4567);
190 
191       const psimd_f32 vi11x0123 = psimd_load_f32(i11);
192       const psimd_f32 vi11x4567 = psimd_load_f32(i11 + 4);
193       i11 += 8;
194 
195       const psimd_f32 vk11x0123 = psimd_load_f32(w + 96);
196       const psimd_f32 vk11x4567 = psimd_load_f32(w + 100);
197       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi11x0123, vk11x0123);
198       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi11x4567, vk11x4567);
199 
200       const psimd_f32 vi12x0123 = psimd_load_f32(i12);
201       const psimd_f32 vi12x4567 = psimd_load_f32(i12 + 4);
202       i12 += 8;
203 
204       const psimd_f32 vk12x0123 = psimd_load_f32(w + 104);
205       const psimd_f32 vk12x4567 = psimd_load_f32(w + 108);
206       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi12x0123, vk12x0123);
207       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi12x4567, vk12x4567);
208 
209       const psimd_f32 vi13x0123 = psimd_load_f32(i13);
210       const psimd_f32 vi13x4567 = psimd_load_f32(i13 + 4);
211       i13 += 8;
212 
213       const psimd_f32 vk13x0123 = psimd_load_f32(w + 112);
214       const psimd_f32 vk13x4567 = psimd_load_f32(w + 116);
215       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi13x0123, vk13x0123);
216       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi13x4567, vk13x4567);
217 
218       const psimd_f32 vi14x0123 = psimd_load_f32(i14);
219       const psimd_f32 vi14x4567 = psimd_load_f32(i14 + 4);
220       i14 += 8;
221 
222       const psimd_f32 vk14x0123 = psimd_load_f32(w + 120);
223       const psimd_f32 vk14x4567 = psimd_load_f32(w + 124);
224       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi14x0123, vk14x0123);
225       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi14x4567, vk14x4567);
226 
227       const psimd_f32 vi15x0123 = psimd_load_f32(i15);
228       const psimd_f32 vi15x4567 = psimd_load_f32(i15 + 4);
229       i15 += 8;
230 
231       const psimd_f32 vk15x0123 = psimd_load_f32(w + 128);
232       const psimd_f32 vk15x4567 = psimd_load_f32(w + 132);
233       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi15x0123, vk15x0123);
234       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi15x4567, vk15x4567);
235 
236       const psimd_f32 vi16x0123 = psimd_load_f32(i16);
237       const psimd_f32 vi16x4567 = psimd_load_f32(i16 + 4);
238       i16 += 8;
239 
240       const psimd_f32 vk16x0123 = psimd_load_f32(w + 136);
241       const psimd_f32 vk16x4567 = psimd_load_f32(w + 140);
242       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi16x0123, vk16x0123);
243       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi16x4567, vk16x4567);
244 
245       const psimd_f32 vi17x0123 = psimd_load_f32(i17);
246       const psimd_f32 vi17x4567 = psimd_load_f32(i17 + 4);
247       i17 += 8;
248 
249       const psimd_f32 vk17x0123 = psimd_load_f32(w + 144);
250       const psimd_f32 vk17x4567 = psimd_load_f32(w + 148);
251       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi17x0123, vk17x0123);
252       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi17x4567, vk17x4567);
253 
254       const psimd_f32 vi18x0123 = psimd_load_f32(i18);
255       const psimd_f32 vi18x4567 = psimd_load_f32(i18 + 4);
256       i18 += 8;
257 
258       const psimd_f32 vk18x0123 = psimd_load_f32(w + 152);
259       const psimd_f32 vk18x4567 = psimd_load_f32(w + 156);
260       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi18x0123, vk18x0123);
261       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi18x4567, vk18x4567);
262 
263       const psimd_f32 vi19x0123 = psimd_load_f32(i19);
264       const psimd_f32 vi19x4567 = psimd_load_f32(i19 + 4);
265       i19 += 8;
266 
267       const psimd_f32 vk19x0123 = psimd_load_f32(w + 160);
268       const psimd_f32 vk19x4567 = psimd_load_f32(w + 164);
269       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi19x0123, vk19x0123);
270       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi19x4567, vk19x4567);
271 
272       const psimd_f32 vi20x0123 = psimd_load_f32(i20);
273       const psimd_f32 vi20x4567 = psimd_load_f32(i20 + 4);
274       i20 += 8;
275 
276       const psimd_f32 vk20x0123 = psimd_load_f32(w + 168);
277       const psimd_f32 vk20x4567 = psimd_load_f32(w + 172);
278       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi20x0123, vk20x0123);
279       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi20x4567, vk20x4567);
280 
281       const psimd_f32 vi21x0123 = psimd_load_f32(i21);
282       const psimd_f32 vi21x4567 = psimd_load_f32(i21 + 4);
283       i21 += 8;
284 
285       const psimd_f32 vk21x0123 = psimd_load_f32(w + 176);
286       const psimd_f32 vk21x4567 = psimd_load_f32(w + 180);
287       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi21x0123, vk21x0123);
288       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi21x4567, vk21x4567);
289 
290       const psimd_f32 vi22x0123 = psimd_load_f32(i22);
291       const psimd_f32 vi22x4567 = psimd_load_f32(i22 + 4);
292       i22 += 8;
293 
294       const psimd_f32 vk22x0123 = psimd_load_f32(w + 184);
295       const psimd_f32 vk22x4567 = psimd_load_f32(w + 188);
296       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi22x0123, vk22x0123);
297       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi22x4567, vk22x4567);
298 
299       const psimd_f32 vi23x0123 = psimd_load_f32(i23);
300       const psimd_f32 vi23x4567 = psimd_load_f32(i23 + 4);
301       i23 += 8;
302 
303       const psimd_f32 vk23x0123 = psimd_load_f32(w + 192);
304       const psimd_f32 vk23x4567 = psimd_load_f32(w + 196);
305       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi23x0123, vk23x0123);
306       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi23x4567, vk23x4567);
307 
308       const psimd_f32 vi24x0123 = psimd_load_f32(i24);
309       const psimd_f32 vi24x4567 = psimd_load_f32(i24 + 4);
310       i24 += 8;
311 
312       const psimd_f32 vk24x0123 = psimd_load_f32(w + 200);
313       const psimd_f32 vk24x4567 = psimd_load_f32(w + 204);
314       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123);
315       vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi24x4567, vk24x4567);
316 
317       w += 208;
318 
319 
320       psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin);
321       psimd_f32 vacc4567 = psimd_max_f32(vacc4567p0, vmin);
322       vacc0123 = psimd_min_f32(vacc0123, vmax);
323       vacc4567 = psimd_min_f32(vacc4567, vmax);
324 
325       psimd_store_f32(output, vacc0123);
326       psimd_store_f32(output + 4, vacc4567);
327       output += 8;
328     }
329     for (; c >= 4; c -= 4) {
330       psimd_f32 vacc0123p0 = psimd_load_f32(w);
331 
332       const psimd_f32 vi0x0123 = psimd_load_f32(i0);
333       i0 += 4;
334 
335       const psimd_f32 vk0x0123 = psimd_load_f32(w + 8);
336       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123);
337 
338       const psimd_f32 vi1x0123 = psimd_load_f32(i1);
339       i1 += 4;
340 
341       const psimd_f32 vk1x0123 = psimd_load_f32(w + 16);
342       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi1x0123, vk1x0123);
343 
344       const psimd_f32 vi2x0123 = psimd_load_f32(i2);
345       i2 += 4;
346 
347       const psimd_f32 vk2x0123 = psimd_load_f32(w + 24);
348       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi2x0123, vk2x0123);
349 
350       const psimd_f32 vi3x0123 = psimd_load_f32(i3);
351       i3 += 4;
352 
353       const psimd_f32 vk3x0123 = psimd_load_f32(w + 32);
354       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi3x0123, vk3x0123);
355 
356       const psimd_f32 vi4x0123 = psimd_load_f32(i4);
357       i4 += 4;
358 
359       const psimd_f32 vk4x0123 = psimd_load_f32(w + 40);
360       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi4x0123, vk4x0123);
361 
362       const psimd_f32 vi5x0123 = psimd_load_f32(i5);
363       i5 += 4;
364 
365       const psimd_f32 vk5x0123 = psimd_load_f32(w + 48);
366       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi5x0123, vk5x0123);
367 
368       const psimd_f32 vi6x0123 = psimd_load_f32(i6);
369       i6 += 4;
370 
371       const psimd_f32 vk6x0123 = psimd_load_f32(w + 56);
372       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi6x0123, vk6x0123);
373 
374       const psimd_f32 vi7x0123 = psimd_load_f32(i7);
375       i7 += 4;
376 
377       const psimd_f32 vk7x0123 = psimd_load_f32(w + 64);
378       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi7x0123, vk7x0123);
379 
380       const psimd_f32 vi8x0123 = psimd_load_f32(i8);
381       i8 += 4;
382 
383       const psimd_f32 vk8x0123 = psimd_load_f32(w + 72);
384       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi8x0123, vk8x0123);
385 
386       const psimd_f32 vi9x0123 = psimd_load_f32(i9);
387       i9 += 4;
388 
389       const psimd_f32 vk9x0123 = psimd_load_f32(w + 80);
390       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi9x0123, vk9x0123);
391 
392       const psimd_f32 vi10x0123 = psimd_load_f32(i10);
393       i10 += 4;
394 
395       const psimd_f32 vk10x0123 = psimd_load_f32(w + 88);
396       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi10x0123, vk10x0123);
397 
398       const psimd_f32 vi11x0123 = psimd_load_f32(i11);
399       i11 += 4;
400 
401       const psimd_f32 vk11x0123 = psimd_load_f32(w + 96);
402       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi11x0123, vk11x0123);
403 
404       const psimd_f32 vi12x0123 = psimd_load_f32(i12);
405       i12 += 4;
406 
407       const psimd_f32 vk12x0123 = psimd_load_f32(w + 104);
408       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi12x0123, vk12x0123);
409 
410       const psimd_f32 vi13x0123 = psimd_load_f32(i13);
411       i13 += 4;
412 
413       const psimd_f32 vk13x0123 = psimd_load_f32(w + 112);
414       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi13x0123, vk13x0123);
415 
416       const psimd_f32 vi14x0123 = psimd_load_f32(i14);
417       i14 += 4;
418 
419       const psimd_f32 vk14x0123 = psimd_load_f32(w + 120);
420       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi14x0123, vk14x0123);
421 
422       const psimd_f32 vi15x0123 = psimd_load_f32(i15);
423       i15 += 4;
424 
425       const psimd_f32 vk15x0123 = psimd_load_f32(w + 128);
426       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi15x0123, vk15x0123);
427 
428       const psimd_f32 vi16x0123 = psimd_load_f32(i16);
429       i16 += 4;
430 
431       const psimd_f32 vk16x0123 = psimd_load_f32(w + 136);
432       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi16x0123, vk16x0123);
433 
434       const psimd_f32 vi17x0123 = psimd_load_f32(i17);
435       i17 += 4;
436 
437       const psimd_f32 vk17x0123 = psimd_load_f32(w + 144);
438       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi17x0123, vk17x0123);
439 
440       const psimd_f32 vi18x0123 = psimd_load_f32(i18);
441       i18 += 4;
442 
443       const psimd_f32 vk18x0123 = psimd_load_f32(w + 152);
444       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi18x0123, vk18x0123);
445 
446       const psimd_f32 vi19x0123 = psimd_load_f32(i19);
447       i19 += 4;
448 
449       const psimd_f32 vk19x0123 = psimd_load_f32(w + 160);
450       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi19x0123, vk19x0123);
451 
452       const psimd_f32 vi20x0123 = psimd_load_f32(i20);
453       i20 += 4;
454 
455       const psimd_f32 vk20x0123 = psimd_load_f32(w + 168);
456       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi20x0123, vk20x0123);
457 
458       const psimd_f32 vi21x0123 = psimd_load_f32(i21);
459       i21 += 4;
460 
461       const psimd_f32 vk21x0123 = psimd_load_f32(w + 176);
462       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi21x0123, vk21x0123);
463 
464       const psimd_f32 vi22x0123 = psimd_load_f32(i22);
465       i22 += 4;
466 
467       const psimd_f32 vk22x0123 = psimd_load_f32(w + 184);
468       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi22x0123, vk22x0123);
469 
470       const psimd_f32 vi23x0123 = psimd_load_f32(i23);
471       i23 += 4;
472 
473       const psimd_f32 vk23x0123 = psimd_load_f32(w + 192);
474       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi23x0123, vk23x0123);
475 
476       const psimd_f32 vi24x0123 = psimd_load_f32(i24);
477       i24 += 4;
478 
479       const psimd_f32 vk24x0123 = psimd_load_f32(w + 200);
480       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123);
481 
482       w += 4;
483 
484 
485       psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin);
486       vacc0123 = psimd_min_f32(vacc0123, vmax);
487 
488       psimd_store_f32(output, vacc0123);
489       output += 4;
490     }
491     if XNN_UNLIKELY(c != 0) {
492       psimd_f32 vacc0123p0 = psimd_load_f32(w);
493 
494       const psimd_f32 vi0x0123 = psimd_load_f32(i0);
495       const psimd_f32 vk0x0123 = psimd_load_f32(w + 8);
496       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123);
497 
498       const psimd_f32 vi1x0123 = psimd_load_f32(i1);
499       const psimd_f32 vk1x0123 = psimd_load_f32(w + 16);
500       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi1x0123, vk1x0123);
501 
502       const psimd_f32 vi2x0123 = psimd_load_f32(i2);
503       const psimd_f32 vk2x0123 = psimd_load_f32(w + 24);
504       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi2x0123, vk2x0123);
505 
506       const psimd_f32 vi3x0123 = psimd_load_f32(i3);
507       const psimd_f32 vk3x0123 = psimd_load_f32(w + 32);
508       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi3x0123, vk3x0123);
509 
510       const psimd_f32 vi4x0123 = psimd_load_f32(i4);
511       const psimd_f32 vk4x0123 = psimd_load_f32(w + 40);
512       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi4x0123, vk4x0123);
513 
514       const psimd_f32 vi5x0123 = psimd_load_f32(i5);
515       const psimd_f32 vk5x0123 = psimd_load_f32(w + 48);
516       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi5x0123, vk5x0123);
517 
518       const psimd_f32 vi6x0123 = psimd_load_f32(i6);
519       const psimd_f32 vk6x0123 = psimd_load_f32(w + 56);
520       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi6x0123, vk6x0123);
521 
522       const psimd_f32 vi7x0123 = psimd_load_f32(i7);
523       const psimd_f32 vk7x0123 = psimd_load_f32(w + 64);
524       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi7x0123, vk7x0123);
525 
526       const psimd_f32 vi8x0123 = psimd_load_f32(i8);
527       const psimd_f32 vk8x0123 = psimd_load_f32(w + 72);
528       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi8x0123, vk8x0123);
529 
530       const psimd_f32 vi9x0123 = psimd_load_f32(i9);
531       const psimd_f32 vk9x0123 = psimd_load_f32(w + 80);
532       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi9x0123, vk9x0123);
533 
534       const psimd_f32 vi10x0123 = psimd_load_f32(i10);
535       const psimd_f32 vk10x0123 = psimd_load_f32(w + 88);
536       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi10x0123, vk10x0123);
537 
538       const psimd_f32 vi11x0123 = psimd_load_f32(i11);
539       const psimd_f32 vk11x0123 = psimd_load_f32(w + 96);
540       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi11x0123, vk11x0123);
541 
542       const psimd_f32 vi12x0123 = psimd_load_f32(i12);
543       const psimd_f32 vk12x0123 = psimd_load_f32(w + 104);
544       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi12x0123, vk12x0123);
545 
546       const psimd_f32 vi13x0123 = psimd_load_f32(i13);
547       const psimd_f32 vk13x0123 = psimd_load_f32(w + 112);
548       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi13x0123, vk13x0123);
549 
550       const psimd_f32 vi14x0123 = psimd_load_f32(i14);
551       const psimd_f32 vk14x0123 = psimd_load_f32(w + 120);
552       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi14x0123, vk14x0123);
553 
554       const psimd_f32 vi15x0123 = psimd_load_f32(i15);
555       const psimd_f32 vk15x0123 = psimd_load_f32(w + 128);
556       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi15x0123, vk15x0123);
557 
558       const psimd_f32 vi16x0123 = psimd_load_f32(i16);
559       const psimd_f32 vk16x0123 = psimd_load_f32(w + 136);
560       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi16x0123, vk16x0123);
561 
562       const psimd_f32 vi17x0123 = psimd_load_f32(i17);
563       const psimd_f32 vk17x0123 = psimd_load_f32(w + 144);
564       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi17x0123, vk17x0123);
565 
566       const psimd_f32 vi18x0123 = psimd_load_f32(i18);
567       const psimd_f32 vk18x0123 = psimd_load_f32(w + 152);
568       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi18x0123, vk18x0123);
569 
570       const psimd_f32 vi19x0123 = psimd_load_f32(i19);
571       const psimd_f32 vk19x0123 = psimd_load_f32(w + 160);
572       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi19x0123, vk19x0123);
573 
574       const psimd_f32 vi20x0123 = psimd_load_f32(i20);
575       const psimd_f32 vk20x0123 = psimd_load_f32(w + 168);
576       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi20x0123, vk20x0123);
577 
578       const psimd_f32 vi21x0123 = psimd_load_f32(i21);
579       const psimd_f32 vk21x0123 = psimd_load_f32(w + 176);
580       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi21x0123, vk21x0123);
581 
582       const psimd_f32 vi22x0123 = psimd_load_f32(i22);
583       const psimd_f32 vk22x0123 = psimd_load_f32(w + 184);
584       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi22x0123, vk22x0123);
585 
586       const psimd_f32 vi23x0123 = psimd_load_f32(i23);
587       const psimd_f32 vk23x0123 = psimd_load_f32(w + 192);
588       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi23x0123, vk23x0123);
589 
590       const psimd_f32 vi24x0123 = psimd_load_f32(i24);
591       const psimd_f32 vk24x0123 = psimd_load_f32(w + 200);
592       vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123);
593 
594 
595       psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin);
596       vacc0123 = psimd_min_f32(vacc0123, vmax);
597 
598       if (c & 2) {
599         psimd_store2_f32(output, vacc0123);
600         vacc0123 = psimd_concat_hi_f32(vacc0123, vacc0123);
601         output += 2;
602       }
603       if (c & 1) {
604         psimd_store1_f32(output, vacc0123);
605         output += 1;
606       }
607     }
608 
609     output = (float*) ((uintptr_t) output + output_increment);
610   } while (--output_width != 0);
611 }
612