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_minmax_ukernel_up4x25__wasmsimd_x86_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_up4x25__wasmsimd_x86_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 v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
33 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
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 input = (const float**) ((uintptr_t) input + input_stride);
161
162 size_t c = channels;
163 const float* w = weights;
164 for (; c >= 4; c -= 4) {
165 v128_t vacc0123p0 = wasm_v128_load(w);
166
167
168 const v128_t vi0x0123 = wasm_v128_load(i0);
169 i0 += 4;
170
171 const v128_t vk0x0123 = wasm_v128_load(w + 4);
172 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
173
174 const v128_t vi1x0123 = wasm_v128_load(i1);
175 i1 += 4;
176
177 const v128_t vk1x0123 = wasm_v128_load(w + 8);
178 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123);
179
180 const v128_t vi2x0123 = wasm_v128_load(i2);
181 i2 += 4;
182
183 const v128_t vk2x0123 = wasm_v128_load(w + 12);
184 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
185
186 const v128_t vi3x0123 = wasm_v128_load(i3);
187 i3 += 4;
188
189 const v128_t vk3x0123 = wasm_v128_load(w + 16);
190 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123));
191
192 const v128_t vi4x0123 = wasm_v128_load(i4);
193 i4 += 4;
194
195 const v128_t vk4x0123 = wasm_v128_load(w + 20);
196 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));
197
198 const v128_t vi5x0123 = wasm_v128_load(i5);
199 i5 += 4;
200
201 const v128_t vk5x0123 = wasm_v128_load(w + 24);
202 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123));
203
204 const v128_t vi6x0123 = wasm_v128_load(i6);
205 i6 += 4;
206
207 const v128_t vk6x0123 = wasm_v128_load(w + 28);
208 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));
209
210 const v128_t vi7x0123 = wasm_v128_load(i7);
211 i7 += 4;
212
213 const v128_t vk7x0123 = wasm_v128_load(w + 32);
214 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123));
215
216 const v128_t vi8x0123 = wasm_v128_load(i8);
217 i8 += 4;
218
219 const v128_t vk8x0123 = wasm_v128_load(w + 36);
220 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));
221
222 const v128_t vi9x0123 = wasm_v128_load(i9);
223 i9 += 4;
224
225 const v128_t vk9x0123 = wasm_v128_load(w + 40);
226 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123));
227
228 const v128_t vi10x0123 = wasm_v128_load(i10);
229 i10 += 4;
230
231 const v128_t vk10x0123 = wasm_v128_load(w + 44);
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123));
233
234 const v128_t vi11x0123 = wasm_v128_load(i11);
235 i11 += 4;
236
237 const v128_t vk11x0123 = wasm_v128_load(w + 48);
238 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123));
239
240 const v128_t vi12x0123 = wasm_v128_load(i12);
241 i12 += 4;
242
243 const v128_t vk12x0123 = wasm_v128_load(w + 52);
244 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123));
245
246 const v128_t vi13x0123 = wasm_v128_load(i13);
247 i13 += 4;
248
249 const v128_t vk13x0123 = wasm_v128_load(w + 56);
250 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123));
251
252 const v128_t vi14x0123 = wasm_v128_load(i14);
253 i14 += 4;
254
255 const v128_t vk14x0123 = wasm_v128_load(w + 60);
256 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123));
257
258 const v128_t vi15x0123 = wasm_v128_load(i15);
259 i15 += 4;
260
261 const v128_t vk15x0123 = wasm_v128_load(w + 64);
262 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123));
263
264 const v128_t vi16x0123 = wasm_v128_load(i16);
265 i16 += 4;
266
267 const v128_t vk16x0123 = wasm_v128_load(w + 68);
268 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123));
269
270 const v128_t vi17x0123 = wasm_v128_load(i17);
271 i17 += 4;
272
273 const v128_t vk17x0123 = wasm_v128_load(w + 72);
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123));
275
276 const v128_t vi18x0123 = wasm_v128_load(i18);
277 i18 += 4;
278
279 const v128_t vk18x0123 = wasm_v128_load(w + 76);
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123));
281
282 const v128_t vi19x0123 = wasm_v128_load(i19);
283 i19 += 4;
284
285 const v128_t vk19x0123 = wasm_v128_load(w + 80);
286 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123));
287
288 const v128_t vi20x0123 = wasm_v128_load(i20);
289 i20 += 4;
290
291 const v128_t vk20x0123 = wasm_v128_load(w + 84);
292 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123));
293
294 const v128_t vi21x0123 = wasm_v128_load(i21);
295 i21 += 4;
296
297 const v128_t vk21x0123 = wasm_v128_load(w + 88);
298 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123));
299
300 const v128_t vi22x0123 = wasm_v128_load(i22);
301 i22 += 4;
302
303 const v128_t vk22x0123 = wasm_v128_load(w + 92);
304 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123));
305
306 const v128_t vi23x0123 = wasm_v128_load(i23);
307 i23 += 4;
308
309 const v128_t vk23x0123 = wasm_v128_load(w + 96);
310 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123));
311
312 const v128_t vi24x0123 = wasm_v128_load(i24);
313 i24 += 4;
314
315 const v128_t vk24x0123 = wasm_v128_load(w + 100);
316 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123));
317
318 w += 104;
319
320 // Add up all accumulators to vacc0123p0
321 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1);
322
323 v128_t vacc0123 = wasm_f32x4_pmax(vmin, vacc0123p0);
324
325 vacc0123 = wasm_f32x4_pmin(vmax, vacc0123);
326
327 wasm_v128_store(output, vacc0123);
328 output += 4;
329 }
330 if XNN_UNLIKELY(c != 0) {
331 v128_t vacc0123p0 = wasm_v128_load(w);
332
333 const v128_t vi0x0123 = wasm_v128_load(i0);
334 const v128_t vk0x0123 = wasm_v128_load(w + 4);
335 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
336
337 const v128_t vi1x0123 = wasm_v128_load(i1);
338 const v128_t vk1x0123 = wasm_v128_load(w + 8);
339 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123);
340
341 const v128_t vi2x0123 = wasm_v128_load(i2);
342 const v128_t vk2x0123 = wasm_v128_load(w + 12);
343 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
344
345 const v128_t vi3x0123 = wasm_v128_load(i3);
346 const v128_t vk3x0123 = wasm_v128_load(w + 16);
347 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123));
348
349 const v128_t vi4x0123 = wasm_v128_load(i4);
350 const v128_t vk4x0123 = wasm_v128_load(w + 20);
351 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));
352
353 const v128_t vi5x0123 = wasm_v128_load(i5);
354 const v128_t vk5x0123 = wasm_v128_load(w + 24);
355 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123));
356
357 const v128_t vi6x0123 = wasm_v128_load(i6);
358 const v128_t vk6x0123 = wasm_v128_load(w + 28);
359 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));
360
361 const v128_t vi7x0123 = wasm_v128_load(i7);
362 const v128_t vk7x0123 = wasm_v128_load(w + 32);
363 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123));
364
365 const v128_t vi8x0123 = wasm_v128_load(i8);
366 const v128_t vk8x0123 = wasm_v128_load(w + 36);
367 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));
368
369 const v128_t vi9x0123 = wasm_v128_load(i9);
370 const v128_t vk9x0123 = wasm_v128_load(w + 40);
371 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123));
372
373 const v128_t vi10x0123 = wasm_v128_load(i10);
374 const v128_t vk10x0123 = wasm_v128_load(w + 44);
375 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123));
376
377 const v128_t vi11x0123 = wasm_v128_load(i11);
378 const v128_t vk11x0123 = wasm_v128_load(w + 48);
379 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123));
380
381 const v128_t vi12x0123 = wasm_v128_load(i12);
382 const v128_t vk12x0123 = wasm_v128_load(w + 52);
383 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123));
384
385 const v128_t vi13x0123 = wasm_v128_load(i13);
386 const v128_t vk13x0123 = wasm_v128_load(w + 56);
387 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123));
388
389 const v128_t vi14x0123 = wasm_v128_load(i14);
390 const v128_t vk14x0123 = wasm_v128_load(w + 60);
391 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi14x0123, vk14x0123));
392
393 const v128_t vi15x0123 = wasm_v128_load(i15);
394 const v128_t vk15x0123 = wasm_v128_load(w + 64);
395 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123));
396
397 const v128_t vi16x0123 = wasm_v128_load(i16);
398 const v128_t vk16x0123 = wasm_v128_load(w + 68);
399 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123));
400
401 const v128_t vi17x0123 = wasm_v128_load(i17);
402 const v128_t vk17x0123 = wasm_v128_load(w + 72);
403 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123));
404
405 const v128_t vi18x0123 = wasm_v128_load(i18);
406 const v128_t vk18x0123 = wasm_v128_load(w + 76);
407 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi18x0123, vk18x0123));
408
409 const v128_t vi19x0123 = wasm_v128_load(i19);
410 const v128_t vk19x0123 = wasm_v128_load(w + 80);
411 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123));
412
413 const v128_t vi20x0123 = wasm_v128_load(i20);
414 const v128_t vk20x0123 = wasm_v128_load(w + 84);
415 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi20x0123, vk20x0123));
416
417 const v128_t vi21x0123 = wasm_v128_load(i21);
418 const v128_t vk21x0123 = wasm_v128_load(w + 88);
419 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi21x0123, vk21x0123));
420
421 const v128_t vi22x0123 = wasm_v128_load(i22);
422 const v128_t vk22x0123 = wasm_v128_load(w + 92);
423 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123));
424
425 const v128_t vi23x0123 = wasm_v128_load(i23);
426 const v128_t vk23x0123 = wasm_v128_load(w + 96);
427 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi23x0123, vk23x0123));
428
429 const v128_t vi24x0123 = wasm_v128_load(i24);
430 const v128_t vk24x0123 = wasm_v128_load(w + 100);
431 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi24x0123, vk24x0123));
432
433 // Add up all accumulators to vacc0123p0
434 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1);
435
436 v128_t vacc0123 = wasm_f32x4_pmax(vmin, vacc0123p0);
437 vacc0123 = wasm_f32x4_pmin(vmax, vacc0123);
438
439 if (c & 2) {
440 *((double*) output) = wasm_f64x2_extract_lane(vacc0123, 0);
441 vacc0123 = wasm_v32x4_shuffle(vacc0123, vacc0123, 2, 3, 2, 3);
442 output += 2;
443 }
444 if (c & 1) {
445 *output = wasm_f32x4_extract_lane(vacc0123, 0);
446 output += 1;
447 }
448 }
449
450 output = (float*) ((uintptr_t) output + output_increment);
451 } while (--output_width != 0);
452 }
453