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_up8x25__neon_acc2(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_up8x25__neon_acc2(
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 >= 8; c -= 8) {
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
167 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4;
168
169
170 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
171 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
172 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
173 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4;
174 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
175 vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567);
176
177 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
178 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
179 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
180 const float32x4_t vk1x4567 = vld1q_f32(w); w += 4;
181 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
182 float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567);
183
184 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
185 const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
186 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
187 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4;
188 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
189 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567);
190
191 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
192 const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
193 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4;
194 const float32x4_t vk3x4567 = vld1q_f32(w); w += 4;
195 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
196 vacc4567p1 = vmlaq_f32(vacc4567p1, vi3x4567, vk3x4567);
197
198 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4;
199 const float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;
200 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4;
201 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4;
202 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
203 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567);
204
205 const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4;
206 const float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4;
207 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4;
208 const float32x4_t vk5x4567 = vld1q_f32(w); w += 4;
209 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123);
210 vacc4567p1 = vmlaq_f32(vacc4567p1, vi5x4567, vk5x4567);
211
212 const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4;
213 const float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4;
214 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4;
215 const float32x4_t vk6x4567 = vld1q_f32(w); w += 4;
216 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
217 vacc4567p0 = vmlaq_f32(vacc4567p0, vi6x4567, vk6x4567);
218
219 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4;
220 const float32x4_t vi7x4567 = vld1q_f32(i7); i7 += 4;
221 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4;
222 const float32x4_t vk7x4567 = vld1q_f32(w); w += 4;
223 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123);
224 vacc4567p1 = vmlaq_f32(vacc4567p1, vi7x4567, vk7x4567);
225
226 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4;
227 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4;
228 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4;
229 const float32x4_t vk8x4567 = vld1q_f32(w); w += 4;
230 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
231 vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567);
232
233 const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4;
234 const float32x4_t vi9x4567 = vld1q_f32(i9); i9 += 4;
235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4;
236 const float32x4_t vk9x4567 = vld1q_f32(w); w += 4;
237 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123);
238 vacc4567p1 = vmlaq_f32(vacc4567p1, vi9x4567, vk9x4567);
239
240 const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4;
241 const float32x4_t vi10x4567 = vld1q_f32(i10); i10 += 4;
242 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4;
243 const float32x4_t vk10x4567 = vld1q_f32(w); w += 4;
244 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123);
245 vacc4567p0 = vmlaq_f32(vacc4567p0, vi10x4567, vk10x4567);
246
247 const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4;
248 const float32x4_t vi11x4567 = vld1q_f32(i11); i11 += 4;
249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4;
250 const float32x4_t vk11x4567 = vld1q_f32(w); w += 4;
251 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123);
252 vacc4567p1 = vmlaq_f32(vacc4567p1, vi11x4567, vk11x4567);
253
254 const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4;
255 const float32x4_t vi12x4567 = vld1q_f32(i12); i12 += 4;
256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4;
257 const float32x4_t vk12x4567 = vld1q_f32(w); w += 4;
258 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123);
259 vacc4567p0 = vmlaq_f32(vacc4567p0, vi12x4567, vk12x4567);
260
261 const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4;
262 const float32x4_t vi13x4567 = vld1q_f32(i13); i13 += 4;
263 const float32x4_t vk13x0123 = vld1q_f32(w); w += 4;
264 const float32x4_t vk13x4567 = vld1q_f32(w); w += 4;
265 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123);
266 vacc4567p1 = vmlaq_f32(vacc4567p1, vi13x4567, vk13x4567);
267
268 const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4;
269 const float32x4_t vi14x4567 = vld1q_f32(i14); i14 += 4;
270 const float32x4_t vk14x0123 = vld1q_f32(w); w += 4;
271 const float32x4_t vk14x4567 = vld1q_f32(w); w += 4;
272 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123);
273 vacc4567p0 = vmlaq_f32(vacc4567p0, vi14x4567, vk14x4567);
274
275 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4;
276 const float32x4_t vi15x4567 = vld1q_f32(i15); i15 += 4;
277 const float32x4_t vk15x0123 = vld1q_f32(w); w += 4;
278 const float32x4_t vk15x4567 = vld1q_f32(w); w += 4;
279 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123);
280 vacc4567p1 = vmlaq_f32(vacc4567p1, vi15x4567, vk15x4567);
281
282 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4;
283 const float32x4_t vi16x4567 = vld1q_f32(i16); i16 += 4;
284 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4;
285 const float32x4_t vk16x4567 = vld1q_f32(w); w += 4;
286 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123);
287 vacc4567p0 = vmlaq_f32(vacc4567p0, vi16x4567, vk16x4567);
288
289 const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4;
290 const float32x4_t vi17x4567 = vld1q_f32(i17); i17 += 4;
291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4;
292 const float32x4_t vk17x4567 = vld1q_f32(w); w += 4;
293 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123);
294 vacc4567p1 = vmlaq_f32(vacc4567p1, vi17x4567, vk17x4567);
295
296 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4;
297 const float32x4_t vi18x4567 = vld1q_f32(i18); i18 += 4;
298 const float32x4_t vk18x0123 = vld1q_f32(w); w += 4;
299 const float32x4_t vk18x4567 = vld1q_f32(w); w += 4;
300 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123);
301 vacc4567p0 = vmlaq_f32(vacc4567p0, vi18x4567, vk18x4567);
302
303 const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4;
304 const float32x4_t vi19x4567 = vld1q_f32(i19); i19 += 4;
305 const float32x4_t vk19x0123 = vld1q_f32(w); w += 4;
306 const float32x4_t vk19x4567 = vld1q_f32(w); w += 4;
307 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123);
308 vacc4567p1 = vmlaq_f32(vacc4567p1, vi19x4567, vk19x4567);
309
310 const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4;
311 const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4;
312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4;
313 const float32x4_t vk20x4567 = vld1q_f32(w); w += 4;
314 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123);
315 vacc4567p0 = vmlaq_f32(vacc4567p0, vi20x4567, vk20x4567);
316
317 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4;
318 const float32x4_t vi21x4567 = vld1q_f32(i21); i21 += 4;
319 const float32x4_t vk21x0123 = vld1q_f32(w); w += 4;
320 const float32x4_t vk21x4567 = vld1q_f32(w); w += 4;
321 vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123);
322 vacc4567p1 = vmlaq_f32(vacc4567p1, vi21x4567, vk21x4567);
323
324 const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4;
325 const float32x4_t vi22x4567 = vld1q_f32(i22); i22 += 4;
326 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4;
327 const float32x4_t vk22x4567 = vld1q_f32(w); w += 4;
328 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123);
329 vacc4567p0 = vmlaq_f32(vacc4567p0, vi22x4567, vk22x4567);
330
331 const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4;
332 const float32x4_t vi23x4567 = vld1q_f32(i23); i23 += 4;
333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4;
334 const float32x4_t vk23x4567 = vld1q_f32(w); w += 4;
335 vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123);
336 vacc4567p1 = vmlaq_f32(vacc4567p1, vi23x4567, vk23x4567);
337
338 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4;
339 const float32x4_t vi24x4567 = vld1q_f32(i24); i24 += 4;
340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4;
341 const float32x4_t vk24x4567 = vld1q_f32(w); w += 4;
342 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123);
343 vacc4567p0 = vmlaq_f32(vacc4567p0, vi24x4567, vk24x4567);
344
345 // Add up all accumulators to vacc01234567p0
346 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
347 vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1);
348
349 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
350 float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin);
351 vacc0123 = vminq_f32(vacc0123, vmax);
352 vacc4567 = vminq_f32(vacc4567, vmax);
353
354 vst1q_f32(output, vacc0123); output += 4;
355 vst1q_f32(output, vacc4567); output += 4;
356 }
357 for (; c >= 4; c -= 4) {
358 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
359
360
361 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
362 const float32x4_t vk0x0123 = vld1q_f32(w + 4);
363 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
364
365 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
366 const float32x4_t vk1x0123 = vld1q_f32(w + 12);
367 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
368
369 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
370 const float32x4_t vk2x0123 = vld1q_f32(w + 20);
371 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
372
373 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
374 const float32x4_t vk3x0123 = vld1q_f32(w + 28);
375 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
376
377 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4;
378 const float32x4_t vk4x0123 = vld1q_f32(w + 36);
379 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
380
381 const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4;
382 const float32x4_t vk5x0123 = vld1q_f32(w + 44);
383 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123);
384
385 const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4;
386 const float32x4_t vk6x0123 = vld1q_f32(w + 52);
387 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
388
389 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4;
390 const float32x4_t vk7x0123 = vld1q_f32(w + 60);
391 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123);
392
393 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4;
394 const float32x4_t vk8x0123 = vld1q_f32(w + 68);
395 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
396
397 const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4;
398 const float32x4_t vk9x0123 = vld1q_f32(w + 76);
399 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123);
400
401 const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4;
402 const float32x4_t vk10x0123 = vld1q_f32(w + 84);
403 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123);
404
405 const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4;
406 const float32x4_t vk11x0123 = vld1q_f32(w + 92);
407 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123);
408
409 const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4;
410 const float32x4_t vk12x0123 = vld1q_f32(w + 100);
411 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123);
412
413 const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4;
414 const float32x4_t vk13x0123 = vld1q_f32(w + 108);
415 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123);
416
417 const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4;
418 const float32x4_t vk14x0123 = vld1q_f32(w + 116);
419 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123);
420
421 const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4;
422 const float32x4_t vk15x0123 = vld1q_f32(w + 124);
423 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123);
424
425 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4;
426 const float32x4_t vk16x0123 = vld1q_f32(w + 132);
427 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123);
428
429 const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4;
430 const float32x4_t vk17x0123 = vld1q_f32(w + 140);
431 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123);
432
433 const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4;
434 const float32x4_t vk18x0123 = vld1q_f32(w + 148);
435 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123);
436
437 const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4;
438 const float32x4_t vk19x0123 = vld1q_f32(w + 156);
439 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123);
440
441 const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4;
442 const float32x4_t vk20x0123 = vld1q_f32(w + 164);
443 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123);
444
445 const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4;
446 const float32x4_t vk21x0123 = vld1q_f32(w + 172);
447 vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123);
448
449 const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4;
450 const float32x4_t vk22x0123 = vld1q_f32(w + 180);
451 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123);
452
453 const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4;
454 const float32x4_t vk23x0123 = vld1q_f32(w + 188);
455 vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123);
456
457 const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4;
458 const float32x4_t vk24x0123 = vld1q_f32(w + 196);
459 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123);
460
461 // Add up all accumulators to vacc0123p0
462 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
463
464 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
465 vacc0123 = vminq_f32(vacc0123, vmax);
466
467 vst1q_f32(output, vacc0123); output += 4;
468 }
469 if XNN_UNLIKELY(c != 0) {
470 float32x4_t vacc0123p0 = vld1q_f32(w);
471
472
473 const float32x4_t vi0x0123 = vld1q_f32(i0);
474 const float32x4_t vk0x0123 = vld1q_f32(w + 8);
475 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
476
477 const float32x4_t vi1x0123 = vld1q_f32(i1);
478 const float32x4_t vk1x0123 = vld1q_f32(w + 16);
479 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
480
481 const float32x4_t vi2x0123 = vld1q_f32(i2);
482 const float32x4_t vk2x0123 = vld1q_f32(w + 24);
483 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
484
485 const float32x4_t vi3x0123 = vld1q_f32(i3);
486 const float32x4_t vk3x0123 = vld1q_f32(w + 32);
487 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
488
489 const float32x4_t vi4x0123 = vld1q_f32(i4);
490 const float32x4_t vk4x0123 = vld1q_f32(w + 40);
491 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
492
493 const float32x4_t vi5x0123 = vld1q_f32(i5);
494 const float32x4_t vk5x0123 = vld1q_f32(w + 48);
495 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123);
496
497 const float32x4_t vi6x0123 = vld1q_f32(i6);
498 const float32x4_t vk6x0123 = vld1q_f32(w + 56);
499 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
500
501 const float32x4_t vi7x0123 = vld1q_f32(i7);
502 const float32x4_t vk7x0123 = vld1q_f32(w + 64);
503 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123);
504
505 const float32x4_t vi8x0123 = vld1q_f32(i8);
506 const float32x4_t vk8x0123 = vld1q_f32(w + 72);
507 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
508
509 const float32x4_t vi9x0123 = vld1q_f32(i9);
510 const float32x4_t vk9x0123 = vld1q_f32(w + 80);
511 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123);
512
513 const float32x4_t vi10x0123 = vld1q_f32(i10);
514 const float32x4_t vk10x0123 = vld1q_f32(w + 88);
515 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123);
516
517 const float32x4_t vi11x0123 = vld1q_f32(i11);
518 const float32x4_t vk11x0123 = vld1q_f32(w + 96);
519 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123);
520
521 const float32x4_t vi12x0123 = vld1q_f32(i12);
522 const float32x4_t vk12x0123 = vld1q_f32(w + 104);
523 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123);
524
525 const float32x4_t vi13x0123 = vld1q_f32(i13);
526 const float32x4_t vk13x0123 = vld1q_f32(w + 112);
527 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123);
528
529 const float32x4_t vi14x0123 = vld1q_f32(i14);
530 const float32x4_t vk14x0123 = vld1q_f32(w + 120);
531 vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123);
532
533 const float32x4_t vi15x0123 = vld1q_f32(i15);
534 const float32x4_t vk15x0123 = vld1q_f32(w + 128);
535 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123);
536
537 const float32x4_t vi16x0123 = vld1q_f32(i16);
538 const float32x4_t vk16x0123 = vld1q_f32(w + 136);
539 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123);
540
541 const float32x4_t vi17x0123 = vld1q_f32(i17);
542 const float32x4_t vk17x0123 = vld1q_f32(w + 144);
543 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123);
544
545 const float32x4_t vi18x0123 = vld1q_f32(i18);
546 const float32x4_t vk18x0123 = vld1q_f32(w + 152);
547 vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123);
548
549 const float32x4_t vi19x0123 = vld1q_f32(i19);
550 const float32x4_t vk19x0123 = vld1q_f32(w + 160);
551 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123);
552
553 const float32x4_t vi20x0123 = vld1q_f32(i20);
554 const float32x4_t vk20x0123 = vld1q_f32(w + 168);
555 vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123);
556
557 const float32x4_t vi21x0123 = vld1q_f32(i21);
558 const float32x4_t vk21x0123 = vld1q_f32(w + 176);
559 vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123);
560
561 const float32x4_t vi22x0123 = vld1q_f32(i22);
562 const float32x4_t vk22x0123 = vld1q_f32(w + 184);
563 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123);
564
565 const float32x4_t vi23x0123 = vld1q_f32(i23);
566 const float32x4_t vk23x0123 = vld1q_f32(w + 192);
567 vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123);
568
569 const float32x4_t vi24x0123 = vld1q_f32(i24);
570 const float32x4_t vk24x0123 = vld1q_f32(w + 200);
571 vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123);
572
573 // Add up all accumulators to vacc0123p0
574 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
575
576 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
577 vacc0123 = vminq_f32(vacc0123, vmax);
578
579 float32x2_t vacc01 = vget_low_f32(vacc0123);
580 if (c & 2) {
581 vst1_f32(output, vacc01); output += 2;
582 vacc01 = vget_high_f32(vacc0123);
583 }
584 if (c & 1) {
585 vst1_lane_f32(output, vacc01, 0); output += 1;
586 }
587 }
588
589 output = (float*) ((uintptr_t) output + output_increment);
590 } while (--output_width != 0);
591 }
592