1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <wasm_simd128.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_f32_dwconv_ukernel_up8x25__wasmsimd(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_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_ukernel_up8x25__wasmsimd(
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_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const float* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const float*) ((uintptr_t) i0 + input_offset);
37 }
38 const float* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const float*) ((uintptr_t) i1 + input_offset);
42 }
43 const float* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const float*) ((uintptr_t) i2 + input_offset);
47 }
48 const float* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const float*) ((uintptr_t) i3 + input_offset);
52 }
53 const float* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const float*) ((uintptr_t) i4 + input_offset);
57 }
58 const float* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const float*) ((uintptr_t) i5 + input_offset);
62 }
63 const float* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const float*) ((uintptr_t) i6 + input_offset);
67 }
68 const float* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const float*) ((uintptr_t) i7 + input_offset);
72 }
73 const float* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const float*) ((uintptr_t) i8 + input_offset);
77 }
78 const float* i9 = input[9];
79 assert(i9 != NULL);
80 if XNN_UNPREDICTABLE(i9 != zero) {
81 i9 = (const float*) ((uintptr_t) i9 + input_offset);
82 }
83 const float* i10 = input[10];
84 assert(i10 != NULL);
85 if XNN_UNPREDICTABLE(i10 != zero) {
86 i10 = (const float*) ((uintptr_t) i10 + input_offset);
87 }
88 const float* i11 = input[11];
89 assert(i11 != NULL);
90 if XNN_UNPREDICTABLE(i11 != zero) {
91 i11 = (const float*) ((uintptr_t) i11 + input_offset);
92 }
93 const float* i12 = input[12];
94 assert(i12 != NULL);
95 if XNN_UNPREDICTABLE(i12 != zero) {
96 i12 = (const float*) ((uintptr_t) i12 + input_offset);
97 }
98 const float* i13 = input[13];
99 assert(i13 != NULL);
100 if XNN_UNPREDICTABLE(i13 != zero) {
101 i13 = (const float*) ((uintptr_t) i13 + input_offset);
102 }
103 const float* i14 = input[14];
104 assert(i14 != NULL);
105 if XNN_UNPREDICTABLE(i14 != zero) {
106 i14 = (const float*) ((uintptr_t) i14 + input_offset);
107 }
108 const float* i15 = input[15];
109 assert(i15 != NULL);
110 if XNN_UNPREDICTABLE(i15 != zero) {
111 i15 = (const float*) ((uintptr_t) i15 + input_offset);
112 }
113 const float* i16 = input[16];
114 assert(i16 != NULL);
115 if XNN_UNPREDICTABLE(i16 != zero) {
116 i16 = (const float*) ((uintptr_t) i16 + input_offset);
117 }
118 const float* i17 = input[17];
119 assert(i17 != NULL);
120 if XNN_UNPREDICTABLE(i17 != zero) {
121 i17 = (const float*) ((uintptr_t) i17 + input_offset);
122 }
123 const float* i18 = input[18];
124 assert(i18 != NULL);
125 if XNN_UNPREDICTABLE(i18 != zero) {
126 i18 = (const float*) ((uintptr_t) i18 + input_offset);
127 }
128 const float* i19 = input[19];
129 assert(i19 != NULL);
130 if XNN_UNPREDICTABLE(i19 != zero) {
131 i19 = (const float*) ((uintptr_t) i19 + input_offset);
132 }
133 const float* i20 = input[20];
134 assert(i20 != NULL);
135 if XNN_UNPREDICTABLE(i20 != zero) {
136 i20 = (const float*) ((uintptr_t) i20 + input_offset);
137 }
138 const float* i21 = input[21];
139 assert(i21 != NULL);
140 if XNN_UNPREDICTABLE(i21 != zero) {
141 i21 = (const float*) ((uintptr_t) i21 + input_offset);
142 }
143 const float* i22 = input[22];
144 assert(i22 != NULL);
145 if XNN_UNPREDICTABLE(i22 != zero) {
146 i22 = (const float*) ((uintptr_t) i22 + input_offset);
147 }
148 const float* i23 = input[23];
149 assert(i23 != NULL);
150 if XNN_UNPREDICTABLE(i23 != zero) {
151 i23 = (const float*) ((uintptr_t) i23 + input_offset);
152 }
153 const float* i24 = input[24];
154 assert(i24 != NULL);
155 if XNN_UNPREDICTABLE(i24 != zero) {
156 i24 = (const float*) ((uintptr_t) i24 + input_offset);
157 }
158 input = (const float**) ((uintptr_t) input + input_stride);
159
160 size_t c = channels;
161 const float* w = weights;
162 for (; c >= 8; c -= 8) {
163 v128_t vacc0123p0 = wasm_v128_load(w);
164 v128_t vacc4567p0 = wasm_v128_load(w + 4);
165
166
167 const v128_t vi0x0123 = wasm_v128_load(i0);
168 const v128_t vi0x4567 = wasm_v128_load(i0 + 4);
169 i0 += 8;
170
171 const v128_t vk0x0123 = wasm_v128_load(w + 8);
172 const v128_t vk0x4567 = wasm_v128_load(w + 12);
173 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
174 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi0x4567, vk0x4567));
175
176 const v128_t vi1x0123 = wasm_v128_load(i1);
177 const v128_t vi1x4567 = wasm_v128_load(i1 + 4);
178 i1 += 8;
179
180 const v128_t vk1x0123 = wasm_v128_load(w + 16);
181 const v128_t vk1x4567 = wasm_v128_load(w + 20);
182 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
183 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi1x4567, vk1x4567));
184
185 const v128_t vi2x0123 = wasm_v128_load(i2);
186 const v128_t vi2x4567 = wasm_v128_load(i2 + 4);
187 i2 += 8;
188
189 const v128_t vk2x0123 = wasm_v128_load(w + 24);
190 const v128_t vk2x4567 = wasm_v128_load(w + 28);
191 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
192 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567));
193
194 const v128_t vi3x0123 = wasm_v128_load(i3);
195 const v128_t vi3x4567 = wasm_v128_load(i3 + 4);
196 i3 += 8;
197
198 const v128_t vk3x0123 = wasm_v128_load(w + 32);
199 const v128_t vk3x4567 = wasm_v128_load(w + 36);
200 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
201 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi3x4567, vk3x4567));
202
203 const v128_t vi4x0123 = wasm_v128_load(i4);
204 const v128_t vi4x4567 = wasm_v128_load(i4 + 4);
205 i4 += 8;
206
207 const v128_t vk4x0123 = wasm_v128_load(w + 40);
208 const v128_t vk4x4567 = wasm_v128_load(w + 44);
209 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));
210 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567));
211
212 const v128_t vi5x0123 = wasm_v128_load(i5);
213 const v128_t vi5x4567 = wasm_v128_load(i5 + 4);
214 i5 += 8;
215
216 const v128_t vk5x0123 = wasm_v128_load(w + 48);
217 const v128_t vk5x4567 = wasm_v128_load(w + 52);
218 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));
219 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi5x4567, vk5x4567));
220
221 const v128_t vi6x0123 = wasm_v128_load(i6);
222 const v128_t vi6x4567 = wasm_v128_load(i6 + 4);
223 i6 += 8;
224
225 const v128_t vk6x0123 = wasm_v128_load(w + 56);
226 const v128_t vk6x4567 = wasm_v128_load(w + 60);
227 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));
228 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi6x4567, vk6x4567));
229
230 const v128_t vi7x0123 = wasm_v128_load(i7);
231 const v128_t vi7x4567 = wasm_v128_load(i7 + 4);
232 i7 += 8;
233
234 const v128_t vk7x0123 = wasm_v128_load(w + 64);
235 const v128_t vk7x4567 = wasm_v128_load(w + 68);
236 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));
237 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi7x4567, vk7x4567));
238
239 const v128_t vi8x0123 = wasm_v128_load(i8);
240 const v128_t vi8x4567 = wasm_v128_load(i8 + 4);
241 i8 += 8;
242
243 const v128_t vk8x0123 = wasm_v128_load(w + 72);
244 const v128_t vk8x4567 = wasm_v128_load(w + 76);
245 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));
246 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567));
247
248 const v128_t vi9x0123 = wasm_v128_load(i9);
249 const v128_t vi9x4567 = wasm_v128_load(i9 + 4);
250 i9 += 8;
251
252 const v128_t vk9x0123 = wasm_v128_load(w + 80);
253 const v128_t vk9x4567 = wasm_v128_load(w + 84);
254 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123));
255 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi9x4567, vk9x4567));
256
257 const v128_t vi10x0123 = wasm_v128_load(i10);
258 const v128_t vi10x4567 = wasm_v128_load(i10 + 4);
259 i10 += 8;
260
261 const v128_t vk10x0123 = wasm_v128_load(w + 88);
262 const v128_t vk10x4567 = wasm_v128_load(w + 92);
263 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123));
264 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi10x4567, vk10x4567));
265
266 const v128_t vi11x0123 = wasm_v128_load(i11);
267 const v128_t vi11x4567 = wasm_v128_load(i11 + 4);
268 i11 += 8;
269
270 const v128_t vk11x0123 = wasm_v128_load(w + 96);
271 const v128_t vk11x4567 = wasm_v128_load(w + 100);
272 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123));
273 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi11x4567, vk11x4567));
274
275 const v128_t vi12x0123 = wasm_v128_load(i12);
276 const v128_t vi12x4567 = wasm_v128_load(i12 + 4);
277 i12 += 8;
278
279 const v128_t vk12x0123 = wasm_v128_load(w + 104);
280 const v128_t vk12x4567 = wasm_v128_load(w + 108);
281 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123));
282 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi12x4567, vk12x4567));
283
284 const v128_t vi13x0123 = wasm_v128_load(i13);
285 const v128_t vi13x4567 = wasm_v128_load(i13 + 4);
286 i13 += 8;
287
288 const v128_t vk13x0123 = wasm_v128_load(w + 112);
289 const v128_t vk13x4567 = wasm_v128_load(w + 116);
290 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123));
291 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi13x4567, vk13x4567));
292
293 const v128_t vi14x0123 = wasm_v128_load(i14);
294 const v128_t vi14x4567 = wasm_v128_load(i14 + 4);
295 i14 += 8;
296
297 const v128_t vk14x0123 = wasm_v128_load(w + 120);
298 const v128_t vk14x4567 = wasm_v128_load(w + 124);
299 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123));
300 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi14x4567, vk14x4567));
301
302 const v128_t vi15x0123 = wasm_v128_load(i15);
303 const v128_t vi15x4567 = wasm_v128_load(i15 + 4);
304 i15 += 8;
305
306 const v128_t vk15x0123 = wasm_v128_load(w + 128);
307 const v128_t vk15x4567 = wasm_v128_load(w + 132);
308 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123));
309 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi15x4567, vk15x4567));
310
311 const v128_t vi16x0123 = wasm_v128_load(i16);
312 const v128_t vi16x4567 = wasm_v128_load(i16 + 4);
313 i16 += 8;
314
315 const v128_t vk16x0123 = wasm_v128_load(w + 136);
316 const v128_t vk16x4567 = wasm_v128_load(w + 140);
317 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123));
318 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi16x4567, vk16x4567));
319
320 const v128_t vi17x0123 = wasm_v128_load(i17);
321 const v128_t vi17x4567 = wasm_v128_load(i17 + 4);
322 i17 += 8;
323
324 const v128_t vk17x0123 = wasm_v128_load(w + 144);
325 const v128_t vk17x4567 = wasm_v128_load(w + 148);
326 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123));
327 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi17x4567, vk17x4567));
328
329 const v128_t vi18x0123 = wasm_v128_load(i18);
330 const v128_t vi18x4567 = wasm_v128_load(i18 + 4);
331 i18 += 8;
332
333 const v128_t vk18x0123 = wasm_v128_load(w + 152);
334 const v128_t vk18x4567 = wasm_v128_load(w + 156);
335 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123));
336 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi18x4567, vk18x4567));
337
338 const v128_t vi19x0123 = wasm_v128_load(i19);
339 const v128_t vi19x4567 = wasm_v128_load(i19 + 4);
340 i19 += 8;
341
342 const v128_t vk19x0123 = wasm_v128_load(w + 160);
343 const v128_t vk19x4567 = wasm_v128_load(w + 164);
344 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123));
345 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi19x4567, vk19x4567));
346
347 const v128_t vi20x0123 = wasm_v128_load(i20);
348 const v128_t vi20x4567 = wasm_v128_load(i20 + 4);
349 i20 += 8;
350
351 const v128_t vk20x0123 = wasm_v128_load(w + 168);
352 const v128_t vk20x4567 = wasm_v128_load(w + 172);
353 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123));
354 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi20x4567, vk20x4567));
355
356 const v128_t vi21x0123 = wasm_v128_load(i21);
357 const v128_t vi21x4567 = wasm_v128_load(i21 + 4);
358 i21 += 8;
359
360 const v128_t vk21x0123 = wasm_v128_load(w + 176);
361 const v128_t vk21x4567 = wasm_v128_load(w + 180);
362 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123));
363 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi21x4567, vk21x4567));
364
365 const v128_t vi22x0123 = wasm_v128_load(i22);
366 const v128_t vi22x4567 = wasm_v128_load(i22 + 4);
367 i22 += 8;
368
369 const v128_t vk22x0123 = wasm_v128_load(w + 184);
370 const v128_t vk22x4567 = wasm_v128_load(w + 188);
371 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123));
372 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi22x4567, vk22x4567));
373
374 const v128_t vi23x0123 = wasm_v128_load(i23);
375 const v128_t vi23x4567 = wasm_v128_load(i23 + 4);
376 i23 += 8;
377
378 const v128_t vk23x0123 = wasm_v128_load(w + 192);
379 const v128_t vk23x4567 = wasm_v128_load(w + 196);
380 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123));
381 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi23x4567, vk23x4567));
382
383 const v128_t vi24x0123 = wasm_v128_load(i24);
384 const v128_t vi24x4567 = wasm_v128_load(i24 + 4);
385 i24 += 8;
386
387 const v128_t vk24x0123 = wasm_v128_load(w + 200);
388 const v128_t vk24x4567 = wasm_v128_load(w + 204);
389 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123));
390 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi24x4567, vk24x4567));
391
392 w += 208;
393
394
395 const v128_t vacc0123 = vacc0123p0;
396 const v128_t vacc4567 = vacc4567p0;
397
398 wasm_v128_store(output, vacc0123);
399 wasm_v128_store(output + 4, vacc4567);
400 output += 8;
401 }
402 for (; c >= 4; c -= 4) {
403 v128_t vacc0123p0 = wasm_v128_load(w);
404
405 const v128_t vi0x0123 = wasm_v128_load(i0);
406 i0 += 4;
407
408 const v128_t vk0x0123 = wasm_v128_load(w + 8);
409 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
410
411 const v128_t vi1x0123 = wasm_v128_load(i1);
412 i1 += 4;
413
414 const v128_t vk1x0123 = wasm_v128_load(w + 16);
415 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
416
417 const v128_t vi2x0123 = wasm_v128_load(i2);
418 i2 += 4;
419
420 const v128_t vk2x0123 = wasm_v128_load(w + 24);
421 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
422
423 const v128_t vi3x0123 = wasm_v128_load(i3);
424 i3 += 4;
425
426 const v128_t vk3x0123 = wasm_v128_load(w + 32);
427 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
428
429 const v128_t vi4x0123 = wasm_v128_load(i4);
430 i4 += 4;
431
432 const v128_t vk4x0123 = wasm_v128_load(w + 40);
433 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));
434
435 const v128_t vi5x0123 = wasm_v128_load(i5);
436 i5 += 4;
437
438 const v128_t vk5x0123 = wasm_v128_load(w + 48);
439 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));
440
441 const v128_t vi6x0123 = wasm_v128_load(i6);
442 i6 += 4;
443
444 const v128_t vk6x0123 = wasm_v128_load(w + 56);
445 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));
446
447 const v128_t vi7x0123 = wasm_v128_load(i7);
448 i7 += 4;
449
450 const v128_t vk7x0123 = wasm_v128_load(w + 64);
451 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));
452
453 const v128_t vi8x0123 = wasm_v128_load(i8);
454 i8 += 4;
455
456 const v128_t vk8x0123 = wasm_v128_load(w + 72);
457 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));
458
459 const v128_t vi9x0123 = wasm_v128_load(i9);
460 i9 += 4;
461
462 const v128_t vk9x0123 = wasm_v128_load(w + 80);
463 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123));
464
465 const v128_t vi10x0123 = wasm_v128_load(i10);
466 i10 += 4;
467
468 const v128_t vk10x0123 = wasm_v128_load(w + 88);
469 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123));
470
471 const v128_t vi11x0123 = wasm_v128_load(i11);
472 i11 += 4;
473
474 const v128_t vk11x0123 = wasm_v128_load(w + 96);
475 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123));
476
477 const v128_t vi12x0123 = wasm_v128_load(i12);
478 i12 += 4;
479
480 const v128_t vk12x0123 = wasm_v128_load(w + 104);
481 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123));
482
483 const v128_t vi13x0123 = wasm_v128_load(i13);
484 i13 += 4;
485
486 const v128_t vk13x0123 = wasm_v128_load(w + 112);
487 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123));
488
489 const v128_t vi14x0123 = wasm_v128_load(i14);
490 i14 += 4;
491
492 const v128_t vk14x0123 = wasm_v128_load(w + 120);
493 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123));
494
495 const v128_t vi15x0123 = wasm_v128_load(i15);
496 i15 += 4;
497
498 const v128_t vk15x0123 = wasm_v128_load(w + 128);
499 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123));
500
501 const v128_t vi16x0123 = wasm_v128_load(i16);
502 i16 += 4;
503
504 const v128_t vk16x0123 = wasm_v128_load(w + 136);
505 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123));
506
507 const v128_t vi17x0123 = wasm_v128_load(i17);
508 i17 += 4;
509
510 const v128_t vk17x0123 = wasm_v128_load(w + 144);
511 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123));
512
513 const v128_t vi18x0123 = wasm_v128_load(i18);
514 i18 += 4;
515
516 const v128_t vk18x0123 = wasm_v128_load(w + 152);
517 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123));
518
519 const v128_t vi19x0123 = wasm_v128_load(i19);
520 i19 += 4;
521
522 const v128_t vk19x0123 = wasm_v128_load(w + 160);
523 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123));
524
525 const v128_t vi20x0123 = wasm_v128_load(i20);
526 i20 += 4;
527
528 const v128_t vk20x0123 = wasm_v128_load(w + 168);
529 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123));
530
531 const v128_t vi21x0123 = wasm_v128_load(i21);
532 i21 += 4;
533
534 const v128_t vk21x0123 = wasm_v128_load(w + 176);
535 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123));
536
537 const v128_t vi22x0123 = wasm_v128_load(i22);
538 i22 += 4;
539
540 const v128_t vk22x0123 = wasm_v128_load(w + 184);
541 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123));
542
543 const v128_t vi23x0123 = wasm_v128_load(i23);
544 i23 += 4;
545
546 const v128_t vk23x0123 = wasm_v128_load(w + 192);
547 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123));
548
549 const v128_t vi24x0123 = wasm_v128_load(i24);
550 i24 += 4;
551
552 const v128_t vk24x0123 = wasm_v128_load(w + 200);
553 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123));
554
555 w += 4;
556
557
558 const v128_t vacc0123 = vacc0123p0;
559
560 wasm_v128_store(output, vacc0123);
561 output += 4;
562 }
563 if XNN_UNLIKELY(c != 0) {
564 v128_t vacc0123p0 = wasm_v128_load(w);
565
566 const v128_t vi0x0123 = wasm_v128_load(i0);
567 const v128_t vk0x0123 = wasm_v128_load(w + 8);
568 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
569
570 const v128_t vi1x0123 = wasm_v128_load(i1);
571 const v128_t vk1x0123 = wasm_v128_load(w + 16);
572 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
573
574 const v128_t vi2x0123 = wasm_v128_load(i2);
575 const v128_t vk2x0123 = wasm_v128_load(w + 24);
576 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
577
578 const v128_t vi3x0123 = wasm_v128_load(i3);
579 const v128_t vk3x0123 = wasm_v128_load(w + 32);
580 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
581
582 const v128_t vi4x0123 = wasm_v128_load(i4);
583 const v128_t vk4x0123 = wasm_v128_load(w + 40);
584 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));
585
586 const v128_t vi5x0123 = wasm_v128_load(i5);
587 const v128_t vk5x0123 = wasm_v128_load(w + 48);
588 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));
589
590 const v128_t vi6x0123 = wasm_v128_load(i6);
591 const v128_t vk6x0123 = wasm_v128_load(w + 56);
592 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));
593
594 const v128_t vi7x0123 = wasm_v128_load(i7);
595 const v128_t vk7x0123 = wasm_v128_load(w + 64);
596 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));
597
598 const v128_t vi8x0123 = wasm_v128_load(i8);
599 const v128_t vk8x0123 = wasm_v128_load(w + 72);
600 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));
601
602 const v128_t vi9x0123 = wasm_v128_load(i9);
603 const v128_t vk9x0123 = wasm_v128_load(w + 80);
604 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi9x0123, vk9x0123));
605
606 const v128_t vi10x0123 = wasm_v128_load(i10);
607 const v128_t vk10x0123 = wasm_v128_load(w + 88);
608 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123));
609
610 const v128_t vi11x0123 = wasm_v128_load(i11);
611 const v128_t vk11x0123 = wasm_v128_load(w + 96);
612 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123));
613
614 const v128_t vi12x0123 = wasm_v128_load(i12);
615 const v128_t vk12x0123 = wasm_v128_load(w + 104);
616 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123));
617
618 const v128_t vi13x0123 = wasm_v128_load(i13);
619 const v128_t vk13x0123 = wasm_v128_load(w + 112);
620 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi13x0123, vk13x0123));
621
622 const v128_t vi14x0123 = wasm_v128_load(i14);
623 const v128_t vk14x0123 = wasm_v128_load(w + 120);
624 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123));
625
626 const v128_t vi15x0123 = wasm_v128_load(i15);
627 const v128_t vk15x0123 = wasm_v128_load(w + 128);
628 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi15x0123, vk15x0123));
629
630 const v128_t vi16x0123 = wasm_v128_load(i16);
631 const v128_t vk16x0123 = wasm_v128_load(w + 136);
632 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123));
633
634 const v128_t vi17x0123 = wasm_v128_load(i17);
635 const v128_t vk17x0123 = wasm_v128_load(w + 144);
636 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123));
637
638 const v128_t vi18x0123 = wasm_v128_load(i18);
639 const v128_t vk18x0123 = wasm_v128_load(w + 152);
640 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123));
641
642 const v128_t vi19x0123 = wasm_v128_load(i19);
643 const v128_t vk19x0123 = wasm_v128_load(w + 160);
644 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi19x0123, vk19x0123));
645
646 const v128_t vi20x0123 = wasm_v128_load(i20);
647 const v128_t vk20x0123 = wasm_v128_load(w + 168);
648 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123));
649
650 const v128_t vi21x0123 = wasm_v128_load(i21);
651 const v128_t vk21x0123 = wasm_v128_load(w + 176);
652 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi21x0123, vk21x0123));
653
654 const v128_t vi22x0123 = wasm_v128_load(i22);
655 const v128_t vk22x0123 = wasm_v128_load(w + 184);
656 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123));
657
658 const v128_t vi23x0123 = wasm_v128_load(i23);
659 const v128_t vk23x0123 = wasm_v128_load(w + 192);
660 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi23x0123, vk23x0123));
661
662 const v128_t vi24x0123 = wasm_v128_load(i24);
663 const v128_t vk24x0123 = wasm_v128_load(w + 200);
664 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123));
665
666
667 v128_t vacc0123 = vacc0123p0;
668
669 if (c & 2) {
670 *((double*) output) = wasm_f64x2_extract_lane(vacc0123, 0);
671 vacc0123 = wasm_v32x4_shuffle(vacc0123, vacc0123, 2, 3, 2, 3);
672 output += 2;
673 }
674 if (c & 1) {
675 *output = wasm_f32x4_extract_lane(vacc0123, 0);
676 output += 1;
677 }
678 }
679
680 output = (float*) ((uintptr_t) output + output_increment);
681 } while (--output_width != 0);
682 }
683