1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-neon.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 <arm_neon.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_f32_dwconv_minmax_ukernel_up16x25__neon(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up16x25__neon(
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 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
33 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const float*) ((uintptr_t) i0 + input_offset);
39 }
40 const float* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 }
45 const float* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const float*) ((uintptr_t) i2 + input_offset);
49 }
50 const float* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const float*) ((uintptr_t) i3 + input_offset);
54 }
55 const float* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const float*) ((uintptr_t) i4 + input_offset);
59 }
60 const float* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const float*) ((uintptr_t) i5 + input_offset);
64 }
65 const float* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const float*) ((uintptr_t) i6 + input_offset);
69 }
70 const float* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const float*) ((uintptr_t) i7 + input_offset);
74 }
75 const float* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const float*) ((uintptr_t) i8 + input_offset);
79 }
80 const float* i9 = input[9];
81 assert(i9 != NULL);
82 if XNN_UNPREDICTABLE(i9 != zero) {
83 i9 = (const float*) ((uintptr_t) i9 + input_offset);
84 }
85 const float* i10 = input[10];
86 assert(i10 != NULL);
87 if XNN_UNPREDICTABLE(i10 != zero) {
88 i10 = (const float*) ((uintptr_t) i10 + input_offset);
89 }
90 const float* i11 = input[11];
91 assert(i11 != NULL);
92 if XNN_UNPREDICTABLE(i11 != zero) {
93 i11 = (const float*) ((uintptr_t) i11 + input_offset);
94 }
95 const float* i12 = input[12];
96 assert(i12 != NULL);
97 if XNN_UNPREDICTABLE(i12 != zero) {
98 i12 = (const float*) ((uintptr_t) i12 + input_offset);
99 }
100 const float* i13 = input[13];
101 assert(i13 != NULL);
102 if XNN_UNPREDICTABLE(i13 != zero) {
103 i13 = (const float*) ((uintptr_t) i13 + input_offset);
104 }
105 const float* i14 = input[14];
106 assert(i14 != NULL);
107 if XNN_UNPREDICTABLE(i14 != zero) {
108 i14 = (const float*) ((uintptr_t) i14 + input_offset);
109 }
110 const float* i15 = input[15];
111 assert(i15 != NULL);
112 if XNN_UNPREDICTABLE(i15 != zero) {
113 i15 = (const float*) ((uintptr_t) i15 + input_offset);
114 }
115 const float* i16 = input[16];
116 assert(i16 != NULL);
117 if XNN_UNPREDICTABLE(i16 != zero) {
118 i16 = (const float*) ((uintptr_t) i16 + input_offset);
119 }
120 const float* i17 = input[17];
121 assert(i17 != NULL);
122 if XNN_UNPREDICTABLE(i17 != zero) {
123 i17 = (const float*) ((uintptr_t) i17 + input_offset);
124 }
125 const float* i18 = input[18];
126 assert(i18 != NULL);
127 if XNN_UNPREDICTABLE(i18 != zero) {
128 i18 = (const float*) ((uintptr_t) i18 + input_offset);
129 }
130 const float* i19 = input[19];
131 assert(i19 != NULL);
132 if XNN_UNPREDICTABLE(i19 != zero) {
133 i19 = (const float*) ((uintptr_t) i19 + input_offset);
134 }
135 const float* i20 = input[20];
136 assert(i20 != NULL);
137 if XNN_UNPREDICTABLE(i20 != zero) {
138 i20 = (const float*) ((uintptr_t) i20 + input_offset);
139 }
140 const float* i21 = input[21];
141 assert(i21 != NULL);
142 if XNN_UNPREDICTABLE(i21 != zero) {
143 i21 = (const float*) ((uintptr_t) i21 + input_offset);
144 }
145 const float* i22 = input[22];
146 assert(i22 != NULL);
147 if XNN_UNPREDICTABLE(i22 != zero) {
148 i22 = (const float*) ((uintptr_t) i22 + input_offset);
149 }
150 const float* i23 = input[23];
151 assert(i23 != NULL);
152 if XNN_UNPREDICTABLE(i23 != zero) {
153 i23 = (const float*) ((uintptr_t) i23 + input_offset);
154 }
155 const float* i24 = input[24];
156 assert(i24 != NULL);
157 if XNN_UNPREDICTABLE(i24 != zero) {
158 i24 = (const float*) ((uintptr_t) i24 + input_offset);
159 }
160
161 input = (const float**) ((uintptr_t) input + input_stride);
162
163 size_t c = channels;
164 const float* w = weights;
165 for (; c >= 16; c -= 16) {
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
167 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4;
168 float32x4_t vacc89ABp0 = vld1q_f32(w); w += 4;
169 float32x4_t vaccCDEFp0 = vld1q_f32(w); w += 4;
170
171
172 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
173 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
174 const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
175 const float32x4_t vi0xCDEF = vld1q_f32(i0); i0 += 4;
176 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
177 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4;
178 const float32x4_t vk0x89AB = vld1q_f32(w); w += 4;
179 const float32x4_t vk0xCDEF = vld1q_f32(w); w += 4;
180 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
181 vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567);
182 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi0x89AB, vk0x89AB);
183 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi0xCDEF, vk0xCDEF);
184
185 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
186 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
187 const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
188 const float32x4_t vi1xCDEF = vld1q_f32(i1); i1 += 4;
189 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
190 const float32x4_t vk1x4567 = vld1q_f32(w); w += 4;
191 const float32x4_t vk1x89AB = vld1q_f32(w); w += 4;
192 const float32x4_t vk1xCDEF = vld1q_f32(w); w += 4;
193 vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123);
194 vacc4567p0 = vmlaq_f32(vacc4567p0, vi1x4567, vk1x4567);
195 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi1x89AB, vk1x89AB);
196 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi1xCDEF, vk1xCDEF);
197
198 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
199 const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
200 const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
201 const float32x4_t vi2xCDEF = vld1q_f32(i2); i2 += 4;
202 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
203 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4;
204 const float32x4_t vk2x89AB = vld1q_f32(w); w += 4;
205 const float32x4_t vk2xCDEF = vld1q_f32(w); w += 4;
206 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
207 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567);
208 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi2x89AB, vk2x89AB);
209 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi2xCDEF, vk2xCDEF);
210
211 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
212 const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
213 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
214 const float32x4_t vi3xCDEF = vld1q_f32(i3); i3 += 4;
215 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4;
216 const float32x4_t vk3x4567 = vld1q_f32(w); w += 4;
217 const float32x4_t vk3x89AB = vld1q_f32(w); w += 4;
218 const float32x4_t vk3xCDEF = vld1q_f32(w); w += 4;
219 vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123);
220 vacc4567p0 = vmlaq_f32(vacc4567p0, vi3x4567, vk3x4567);
221 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi3x89AB, vk3x89AB);
222 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi3xCDEF, vk3xCDEF);
223
224 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4;
225 const float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;
226 const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
227 const float32x4_t vi4xCDEF = vld1q_f32(i4); i4 += 4;
228 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4;
229 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4;
230 const float32x4_t vk4x89AB = vld1q_f32(w); w += 4;
231 const float32x4_t vk4xCDEF = vld1q_f32(w); w += 4;
232 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
233 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567);
234 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi4x89AB, vk4x89AB);
235 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi4xCDEF, vk4xCDEF);
236
237 const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4;
238 const float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4;
239 const float32x4_t vi5x89AB = vld1q_f32(i5); i5 += 4;
240 const float32x4_t vi5xCDEF = vld1q_f32(i5); i5 += 4;
241 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4;
242 const float32x4_t vk5x4567 = vld1q_f32(w); w += 4;
243 const float32x4_t vk5x89AB = vld1q_f32(w); w += 4;
244 const float32x4_t vk5xCDEF = vld1q_f32(w); w += 4;
245 vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123);
246 vacc4567p0 = vmlaq_f32(vacc4567p0, vi5x4567, vk5x4567);
247 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi5x89AB, vk5x89AB);
248 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi5xCDEF, vk5xCDEF);
249
250 const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4;
251 const float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4;
252 const float32x4_t vi6x89AB = vld1q_f32(i6); i6 += 4;
253 const float32x4_t vi6xCDEF = vld1q_f32(i6); i6 += 4;
254 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4;
255 const float32x4_t vk6x4567 = vld1q_f32(w); w += 4;
256 const float32x4_t vk6x89AB = vld1q_f32(w); w += 4;
257 const float32x4_t vk6xCDEF = vld1q_f32(w); w += 4;
258 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
259 vacc4567p0 = vmlaq_f32(vacc4567p0, vi6x4567, vk6x4567);
260 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi6x89AB, vk6x89AB);
261 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi6xCDEF, vk6xCDEF);
262
263 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4;
264 const float32x4_t vi7x4567 = vld1q_f32(i7); i7 += 4;
265 const float32x4_t vi7x89AB = vld1q_f32(i7); i7 += 4;
266 const float32x4_t vi7xCDEF = vld1q_f32(i7); i7 += 4;
267 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4;
268 const float32x4_t vk7x4567 = vld1q_f32(w); w += 4;
269 const float32x4_t vk7x89AB = vld1q_f32(w); w += 4;
270 const float32x4_t vk7xCDEF = vld1q_f32(w); w += 4;
271 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123);
272 vacc4567p0 = vmlaq_f32(vacc4567p0, vi7x4567, vk7x4567);
273 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi7x89AB, vk7x89AB);
274 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi7xCDEF, vk7xCDEF);
275
276 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4;
277 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4;
278 const float32x4_t vi8x89AB = vld1q_f32(i8); i8 += 4;
279 const float32x4_t vi8xCDEF = vld1q_f32(i8); i8 += 4;
280 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4;
281 const float32x4_t vk8x4567 = vld1q_f32(w); w += 4;
282 const float32x4_t vk8x89AB = vld1q_f32(w); w += 4;
283 const float32x4_t vk8xCDEF = vld1q_f32(w); w += 4;
284 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
285 vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567);
286 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi8x89AB, vk8x89AB);
287 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi8xCDEF, vk8xCDEF);
288
289 const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4;
290 const float32x4_t vi9x4567 = vld1q_f32(i9); i9 += 4;
291 const float32x4_t vi9x89AB = vld1q_f32(i9); i9 += 4;
292 const float32x4_t vi9xCDEF = vld1q_f32(i9); i9 += 4;
293 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4;
294 const float32x4_t vk9x4567 = vld1q_f32(w); w += 4;
295 const float32x4_t vk9x89AB = vld1q_f32(w); w += 4;
296 const float32x4_t vk9xCDEF = vld1q_f32(w); w += 4;
297 vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123);
298 vacc4567p0 = vmlaq_f32(vacc4567p0, vi9x4567, vk9x4567);
299 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi9x89AB, vk9x89AB);
300 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi9xCDEF, vk9xCDEF);
301
302 const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4;
303 const float32x4_t vi10x4567 = vld1q_f32(i10); i10 += 4;
304 const float32x4_t vi10x89AB = vld1q_f32(i10); i10 += 4;
305 const float32x4_t vi10xCDEF = vld1q_f32(i10); i10 += 4;
306 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4;
307 const float32x4_t vk10x4567 = vld1q_f32(w); w += 4;
308 const float32x4_t vk10x89AB = vld1q_f32(w); w += 4;
309 const float32x4_t vk10xCDEF = vld1q_f32(w); w += 4;
310 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123);
311 vacc4567p0 = vmlaq_f32(vacc4567p0, vi10x4567, vk10x4567);
312 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi10x89AB, vk10x89AB);
313 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi10xCDEF, vk10xCDEF);
314
315 const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4;
316 const float32x4_t vi11x4567 = vld1q_f32(i11); i11 += 4;
317 const float32x4_t vi11x89AB = vld1q_f32(i11); i11 += 4;
318 const float32x4_t vi11xCDEF = vld1q_f32(i11); i11 += 4;
319 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4;
320 const float32x4_t vk11x4567 = vld1q_f32(w); w += 4;
321 const float32x4_t vk11x89AB = vld1q_f32(w); w += 4;
322 const float32x4_t vk11xCDEF = vld1q_f32(w); w += 4;
323 vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123);
324 vacc4567p0 = vmlaq_f32(vacc4567p0, vi11x4567, vk11x4567);
325 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi11x89AB, vk11x89AB);
326 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi11xCDEF, vk11xCDEF);
327
328 const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4;
329 const float32x4_t vi12x4567 = vld1q_f32(i12); i12 += 4;
330 const float32x4_t vi12x89AB = vld1q_f32(i12); i12 += 4;
331 const float32x4_t vi12xCDEF = vld1q_f32(i12); i12 += 4;
332 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4;
333 const float32x4_t vk12x4567 = vld1q_f32(w); w += 4;
334 const float32x4_t vk12x89AB = vld1q_f32(w); w += 4;
335 const float32x4_t vk12xCDEF = vld1q_f32(w); w += 4;
336 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123);
337 vacc4567p0 = vmlaq_f32(vacc4567p0, vi12x4567, vk12x4567);
338 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi12x89AB, vk12x89AB);
339 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi12xCDEF, vk12xCDEF);
340
341 const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4;
342 const float32x4_t vi13x4567 = vld1q_f32(i13); i13 += 4;
343 const float32x4_t vi13x89AB = vld1q_f32(i13); i13 += 4;
344 const float32x4_t vi13xCDEF = vld1q_f32(i13); i13 += 4;
345 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4;
346 const float32x4_t vk13x4567 = vld1q_f32(w); w += 4;
347 const float32x4_t vk13x89AB = vld1q_f32(w); w += 4;
348 const float32x4_t vk13xCDEF = vld1q_f32(w); w += 4;
349 vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123);
350 vacc4567p0 = vmlaq_f32(vacc4567p0, vi13x4567, vk13x4567);
351 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi13x89AB, vk13x89AB);
352 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi13xCDEF, vk13xCDEF);
353
354 const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4;
355 const float32x4_t vi14x4567 = vld1q_f32(i14); i14 += 4;
356 const float32x4_t vi14x89AB = vld1q_f32(i14); i14 += 4;
357 const float32x4_t vi14xCDEF = vld1q_f32(i14); i14 += 4;
358 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4;
359 const float32x4_t vk14x4567 = vld1q_f32(w); w += 4;
360 const float32x4_t vk14x89AB = vld1q_f32(w); w += 4;
361 const float32x4_t vk14xCDEF = vld1q_f32(w); w += 4;
362 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123);
363 vacc4567p0 = vmlaq_f32(vacc4567p0, vi14x4567, vk14x4567);
364 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi14x89AB, vk14x89AB);
365 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi14xCDEF, vk14xCDEF);
366
367 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4;
368 const float32x4_t vi15x4567 = vld1q_f32(i15); i15 += 4;
369 const float32x4_t vi15x89AB = vld1q_f32(i15); i15 += 4;
370 const float32x4_t vi15xCDEF = vld1q_f32(i15); i15 += 4;
371 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4;
372 const float32x4_t vk15x4567 = vld1q_f32(w); w += 4;
373 const float32x4_t vk15x89AB = vld1q_f32(w); w += 4;
374 const float32x4_t vk15xCDEF = vld1q_f32(w); w += 4;
375 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123);
376 vacc4567p0 = vmlaq_f32(vacc4567p0, vi15x4567, vk15x4567);
377 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi15x89AB, vk15x89AB);
378 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi15xCDEF, vk15xCDEF);
379
380 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4;
381 const float32x4_t vi16x4567 = vld1q_f32(i16); i16 += 4;
382 const float32x4_t vi16x89AB = vld1q_f32(i16); i16 += 4;
383 const float32x4_t vi16xCDEF = vld1q_f32(i16); i16 += 4;
384 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4;
385 const float32x4_t vk16x4567 = vld1q_f32(w); w += 4;
386 const float32x4_t vk16x89AB = vld1q_f32(w); w += 4;
387 const float32x4_t vk16xCDEF = vld1q_f32(w); w += 4;
388 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123);
389 vacc4567p0 = vmlaq_f32(vacc4567p0, vi16x4567, vk16x4567);
390 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi16x89AB, vk16x89AB);
391 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi16xCDEF, vk16xCDEF);
392
393 const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4;
394 const float32x4_t vi17x4567 = vld1q_f32(i17); i17 += 4;
395 const float32x4_t vi17x89AB = vld1q_f32(i17); i17 += 4;
396 const float32x4_t vi17xCDEF = vld1q_f32(i17); i17 += 4;
397 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4;
398 const float32x4_t vk17x4567 = vld1q_f32(w); w += 4;
399 const float32x4_t vk17x89AB = vld1q_f32(w); w += 4;
400 const float32x4_t vk17xCDEF = vld1q_f32(w); w += 4;
401 vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123);
402 vacc4567p0 = vmlaq_f32(vacc4567p0, vi17x4567, vk17x4567);
403 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi17x89AB, vk17x89AB);
404 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi17xCDEF, vk17xCDEF);
405
406 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4;
407 const float32x4_t vi18x4567 = vld1q_f32(i18); i18 += 4;
408 const float32x4_t vi18x89AB = vld1q_f32(i18); i18 += 4;
409 const float32x4_t vi18xCDEF = vld1q_f32(i18); i18 += 4;
410 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4;
411 const float32x4_t vk18x4567 = vld1q_f32(w); w += 4;
412 const float32x4_t vk18x89AB = vld1q_f32(w); w += 4;
413 const float32x4_t vk18xCDEF = vld1q_f32(w); w += 4;
414 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123);
415 vacc4567p0 = vmlaq_f32(vacc4567p0, vi18x4567, vk18x4567);
416 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi18x89AB, vk18x89AB);
417 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi18xCDEF, vk18xCDEF);
418
419 const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4;
420 const float32x4_t vi19x4567 = vld1q_f32(i19); i19 += 4;
421 const float32x4_t vi19x89AB = vld1q_f32(i19); i19 += 4;
422 const float32x4_t vi19xCDEF = vld1q_f32(i19); i19 += 4;
423 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4;
424 const float32x4_t vk19x4567 = vld1q_f32(w); w += 4;
425 const float32x4_t vk19x89AB = vld1q_f32(w); w += 4;
426 const float32x4_t vk19xCDEF = vld1q_f32(w); w += 4;
427 vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123);
428 vacc4567p0 = vmlaq_f32(vacc4567p0, vi19x4567, vk19x4567);
429 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi19x89AB, vk19x89AB);
430 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi19xCDEF, vk19xCDEF);
431
432 const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4;
433 const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4;
434 const float32x4_t vi20x89AB = vld1q_f32(i20); i20 += 4;
435 const float32x4_t vi20xCDEF = vld1q_f32(i20); i20 += 4;
436 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4;
437 const float32x4_t vk20x4567 = vld1q_f32(w); w += 4;
438 const float32x4_t vk20x89AB = vld1q_f32(w); w += 4;
439 const float32x4_t vk20xCDEF = vld1q_f32(w); w += 4;
440 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123);
441 vacc4567p0 = vmlaq_f32(vacc4567p0, vi20x4567, vk20x4567);
442 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi20x89AB, vk20x89AB);
443 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi20xCDEF, vk20xCDEF);
444
445 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4;
446 const float32x4_t vi21x4567 = vld1q_f32(i21); i21 += 4;
447 const float32x4_t vi21x89AB = vld1q_f32(i21); i21 += 4;
448 const float32x4_t vi21xCDEF = vld1q_f32(i21); i21 += 4;
449 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4;
450 const float32x4_t vk21x4567 = vld1q_f32(w); w += 4;
451 const float32x4_t vk21x89AB = vld1q_f32(w); w += 4;
452 const float32x4_t vk21xCDEF = vld1q_f32(w); w += 4;
453 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123);
454 vacc4567p0 = vmlaq_f32(vacc4567p0, vi21x4567, vk21x4567);
455 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi21x89AB, vk21x89AB);
456 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi21xCDEF, vk21xCDEF);
457
458 const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4;
459 const float32x4_t vi22x4567 = vld1q_f32(i22); i22 += 4;
460 const float32x4_t vi22x89AB = vld1q_f32(i22); i22 += 4;
461 const float32x4_t vi22xCDEF = vld1q_f32(i22); i22 += 4;
462 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4;
463 const float32x4_t vk22x4567 = vld1q_f32(w); w += 4;
464 const float32x4_t vk22x89AB = vld1q_f32(w); w += 4;
465 const float32x4_t vk22xCDEF = vld1q_f32(w); w += 4;
466 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123);
467 vacc4567p0 = vmlaq_f32(vacc4567p0, vi22x4567, vk22x4567);
468 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi22x89AB, vk22x89AB);
469 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi22xCDEF, vk22xCDEF);
470
471 const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4;
472 const float32x4_t vi23x4567 = vld1q_f32(i23); i23 += 4;
473 const float32x4_t vi23x89AB = vld1q_f32(i23); i23 += 4;
474 const float32x4_t vi23xCDEF = vld1q_f32(i23); i23 += 4;
475 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4;
476 const float32x4_t vk23x4567 = vld1q_f32(w); w += 4;
477 const float32x4_t vk23x89AB = vld1q_f32(w); w += 4;
478 const float32x4_t vk23xCDEF = vld1q_f32(w); w += 4;
479 vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123);
480 vacc4567p0 = vmlaq_f32(vacc4567p0, vi23x4567, vk23x4567);
481 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi23x89AB, vk23x89AB);
482 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi23xCDEF, vk23xCDEF);
483
484 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4;
485 const float32x4_t vi24x4567 = vld1q_f32(i24); i24 += 4;
486 const float32x4_t vi24x89AB = vld1q_f32(i24); i24 += 4;
487 const float32x4_t vi24xCDEF = vld1q_f32(i24); i24 += 4;
488 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4;
489 const float32x4_t vk24x4567 = vld1q_f32(w); w += 4;
490 const float32x4_t vk24x89AB = vld1q_f32(w); w += 4;
491 const float32x4_t vk24xCDEF = vld1q_f32(w); w += 4;
492 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123);
493 vacc4567p0 = vmlaq_f32(vacc4567p0, vi24x4567, vk24x4567);
494 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi24x89AB, vk24x89AB);
495 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi24xCDEF, vk24xCDEF);
496
497
498 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
499 float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin);
500 float32x4_t vacc89AB = vmaxq_f32(vacc89ABp0, vmin);
501 float32x4_t vaccCDEF = vmaxq_f32(vaccCDEFp0, vmin);
502 vacc0123 = vminq_f32(vacc0123, vmax);
503 vacc4567 = vminq_f32(vacc4567, vmax);
504 vacc89AB = vminq_f32(vacc89AB, vmax);
505 vaccCDEF = vminq_f32(vaccCDEF, vmax);
506
507 vst1q_f32(output, vacc0123); output += 4;
508 vst1q_f32(output, vacc4567); output += 4;
509 vst1q_f32(output, vacc89AB); output += 4;
510 vst1q_f32(output, vaccCDEF); output += 4;
511 }
512 for (; c >= 4; c -= 4) {
513 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
514
515
516 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
517 const float32x4_t vk0x0123 = vld1q_f32(w + 12);
518 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
519
520 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
521 const float32x4_t vk1x0123 = vld1q_f32(w + 28);
522 vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123);
523
524 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
525 const float32x4_t vk2x0123 = vld1q_f32(w + 44);
526 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
527
528 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
529 const float32x4_t vk3x0123 = vld1q_f32(w + 60);
530 vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123);
531
532 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4;
533 const float32x4_t vk4x0123 = vld1q_f32(w + 76);
534 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
535
536 const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4;
537 const float32x4_t vk5x0123 = vld1q_f32(w + 92);
538 vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123);
539
540 const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4;
541 const float32x4_t vk6x0123 = vld1q_f32(w + 108);
542 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
543
544 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4;
545 const float32x4_t vk7x0123 = vld1q_f32(w + 124);
546 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123);
547
548 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4;
549 const float32x4_t vk8x0123 = vld1q_f32(w + 140);
550 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
551
552 const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4;
553 const float32x4_t vk9x0123 = vld1q_f32(w + 156);
554 vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123);
555
556 const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4;
557 const float32x4_t vk10x0123 = vld1q_f32(w + 172);
558 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123);
559
560 const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4;
561 const float32x4_t vk11x0123 = vld1q_f32(w + 188);
562 vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123);
563
564 const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4;
565 const float32x4_t vk12x0123 = vld1q_f32(w + 204);
566 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123);
567
568 const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4;
569 const float32x4_t vk13x0123 = vld1q_f32(w + 220);
570 vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123);
571
572 const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4;
573 const float32x4_t vk14x0123 = vld1q_f32(w + 236);
574 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123);
575
576 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4;
577 const float32x4_t vk15x0123 = vld1q_f32(w + 252);
578 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123);
579
580 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4;
581 const float32x4_t vk16x0123 = vld1q_f32(w + 268);
582 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123);
583
584 const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4;
585 const float32x4_t vk17x0123 = vld1q_f32(w + 284);
586 vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123);
587
588 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4;
589 const float32x4_t vk18x0123 = vld1q_f32(w + 300);
590 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123);
591
592 const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4;
593 const float32x4_t vk19x0123 = vld1q_f32(w + 316);
594 vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123);
595
596 const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4;
597 const float32x4_t vk20x0123 = vld1q_f32(w + 332);
598 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123);
599
600 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4;
601 const float32x4_t vk21x0123 = vld1q_f32(w + 348);
602 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123);
603
604 const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4;
605 const float32x4_t vk22x0123 = vld1q_f32(w + 364);
606 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123);
607
608 const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4;
609 const float32x4_t vk23x0123 = vld1q_f32(w + 380);
610 vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123);
611
612 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4;
613 const float32x4_t vk24x0123 = vld1q_f32(w + 396);
614 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123);
615
616
617 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
618 vacc0123 = vminq_f32(vacc0123, vmax);
619
620 vst1q_f32(output, vacc0123); output += 4;
621 }
622 if XNN_UNLIKELY(c != 0) {
623 float32x4_t vacc0123p0 = vld1q_f32(w);
624
625
626 const float32x4_t vi0x0123 = vld1q_f32(i0);
627 const float32x4_t vk0x0123 = vld1q_f32(w + 16);
628 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
629
630 const float32x4_t vi1x0123 = vld1q_f32(i1);
631 const float32x4_t vk1x0123 = vld1q_f32(w + 32);
632 vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123);
633
634 const float32x4_t vi2x0123 = vld1q_f32(i2);
635 const float32x4_t vk2x0123 = vld1q_f32(w + 48);
636 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
637
638 const float32x4_t vi3x0123 = vld1q_f32(i3);
639 const float32x4_t vk3x0123 = vld1q_f32(w + 64);
640 vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123);
641
642 const float32x4_t vi4x0123 = vld1q_f32(i4);
643 const float32x4_t vk4x0123 = vld1q_f32(w + 80);
644 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
645
646 const float32x4_t vi5x0123 = vld1q_f32(i5);
647 const float32x4_t vk5x0123 = vld1q_f32(w + 96);
648 vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123);
649
650 const float32x4_t vi6x0123 = vld1q_f32(i6);
651 const float32x4_t vk6x0123 = vld1q_f32(w + 112);
652 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
653
654 const float32x4_t vi7x0123 = vld1q_f32(i7);
655 const float32x4_t vk7x0123 = vld1q_f32(w + 128);
656 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123);
657
658 const float32x4_t vi8x0123 = vld1q_f32(i8);
659 const float32x4_t vk8x0123 = vld1q_f32(w + 144);
660 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
661
662 const float32x4_t vi9x0123 = vld1q_f32(i9);
663 const float32x4_t vk9x0123 = vld1q_f32(w + 160);
664 vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123);
665
666 const float32x4_t vi10x0123 = vld1q_f32(i10);
667 const float32x4_t vk10x0123 = vld1q_f32(w + 176);
668 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123);
669
670 const float32x4_t vi11x0123 = vld1q_f32(i11);
671 const float32x4_t vk11x0123 = vld1q_f32(w + 192);
672 vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123);
673
674 const float32x4_t vi12x0123 = vld1q_f32(i12);
675 const float32x4_t vk12x0123 = vld1q_f32(w + 208);
676 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123);
677
678 const float32x4_t vi13x0123 = vld1q_f32(i13);
679 const float32x4_t vk13x0123 = vld1q_f32(w + 224);
680 vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123);
681
682 const float32x4_t vi14x0123 = vld1q_f32(i14);
683 const float32x4_t vk14x0123 = vld1q_f32(w + 240);
684 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123);
685
686 const float32x4_t vi15x0123 = vld1q_f32(i15);
687 const float32x4_t vk15x0123 = vld1q_f32(w + 256);
688 vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123);
689
690 const float32x4_t vi16x0123 = vld1q_f32(i16);
691 const float32x4_t vk16x0123 = vld1q_f32(w + 272);
692 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123);
693
694 const float32x4_t vi17x0123 = vld1q_f32(i17);
695 const float32x4_t vk17x0123 = vld1q_f32(w + 288);
696 vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123);
697
698 const float32x4_t vi18x0123 = vld1q_f32(i18);
699 const float32x4_t vk18x0123 = vld1q_f32(w + 304);
700 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123);
701
702 const float32x4_t vi19x0123 = vld1q_f32(i19);
703 const float32x4_t vk19x0123 = vld1q_f32(w + 320);
704 vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123);
705
706 const float32x4_t vi20x0123 = vld1q_f32(i20);
707 const float32x4_t vk20x0123 = vld1q_f32(w + 336);
708 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123);
709
710 const float32x4_t vi21x0123 = vld1q_f32(i21);
711 const float32x4_t vk21x0123 = vld1q_f32(w + 352);
712 vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123);
713
714 const float32x4_t vi22x0123 = vld1q_f32(i22);
715 const float32x4_t vk22x0123 = vld1q_f32(w + 368);
716 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123);
717
718 const float32x4_t vi23x0123 = vld1q_f32(i23);
719 const float32x4_t vk23x0123 = vld1q_f32(w + 384);
720 vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123);
721
722 const float32x4_t vi24x0123 = vld1q_f32(i24);
723 const float32x4_t vk24x0123 = vld1q_f32(w + 400);
724 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123);
725
726
727 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
728 vacc0123 = vminq_f32(vacc0123, vmax);
729
730 float32x2_t vacc01 = vget_low_f32(vacc0123);
731 if (c & 2) {
732 vst1_f32(output, vacc01); output += 2;
733 vacc01 = vget_high_f32(vacc0123);
734 }
735 if (c & 1) {
736 vst1_lane_f32(output, vacc01, 0); output += 1;
737 }
738 }
739
740 output = (float*) ((uintptr_t) output + output_increment);
741 } while (--output_width != 0);
742 }
743