1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/5x5s2p2-scalar.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 <xnnpack/dwconv.h>
13 #include <xnnpack/math.h>
14
15
xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2(
17 size_t input_height,
18 size_t input_width,
19 const float* input,
20 const float* weights,
21 const float* zero,
22 float* output,
23 uint32_t padding_top,
24 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(input_height != 0);
27 assert(input_width != 0);
28 assert(input_width % sizeof(float) == 0);
29 assert(padding_top >= 1);
30 assert(padding_top <= 2);
31
32 const float vmax = params->scalar.max;
33 const float vmin = params->scalar.min;
34
35 const float vbias = weights[0];
36 const float vk00 = weights[1];
37 const float vk01 = weights[2];
38 const float vk02 = weights[3];
39 const float vk03 = weights[4];
40 const float vk04 = weights[5];
41 const float vk10 = weights[6];
42 const float vk11 = weights[7];
43 const float vk12 = weights[8];
44 const float vk13 = weights[9];
45 const float vk14 = weights[10];
46 const float vk20 = weights[11];
47 const float vk21 = weights[12];
48 const float vk22 = weights[13];
49 const float vk23 = weights[14];
50 const float vk24 = weights[15];
51 const float vk30 = weights[16];
52 const float vk31 = weights[17];
53 const float vk32 = weights[18];
54 const float vk33 = weights[19];
55 const float vk34 = weights[20];
56 const float vk40 = weights[21];
57 const float vk41 = weights[22];
58 const float vk42 = weights[23];
59 const float vk43 = weights[24];
60 const float vk44 = weights[25];
61
62 const uint32_t padding_top_less_1 = padding_top - 1;
63
64 const float* i0 = zero;
65 const float* i1 = (const float*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
66 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
67 if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
68 i1 = zero;
69 }
70 const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
71 const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
72 const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
73 const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
74 const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
75 const float* i8 = (const float*) ((uintptr_t) i7 + input_width);
76
77 const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
78
79 float* o0 = output;
80 float* o1 = (float*) ((uintptr_t) o0 + output_width);
81 float* o2 = (float*) ((uintptr_t) o1 + output_width);
82
83 size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
84 size_t output_height = (padded_input_height - 5 /* kernel size */ + 2 /* subsampling */) / 2;
85 do {
86 if XNN_UNPREDICTABLE(padded_input_height < 6) {
87 i3 = zero;
88 }
89 if XNN_UNPREDICTABLE(padded_input_height < 7) {
90 i4 = zero;
91 o1 = o0;
92 }
93 if XNN_UNPREDICTABLE(padded_input_height < 8) {
94 i5 = zero;
95 }
96 if XNN_UNPREDICTABLE(padded_input_height < 9) {
97 i6 = zero;
98 o2 = o1;
99 }
100 if XNN_UNPREDICTABLE(padded_input_height < 10) {
101 i7 = zero;
102 }
103 if XNN_UNPREDICTABLE(padded_input_height < 11) {
104 i8 = zero;
105 }
106
107 float vi0x0 = 0.0f;
108 float vi1x0 = 0.0f;
109 float vi2x0 = 0.0f;
110 float vi3x0 = 0.0f;
111 float vi4x0 = 0.0f;
112 float vi5x0 = 0.0f;
113 float vi6x0 = 0.0f;
114 float vi7x0 = 0.0f;
115 float vi8x0 = 0.0f;
116
117 float vi0x1 = 0.0f;
118 float vi1x1 = 0.0f;
119 float vi2x1 = 0.0f;
120 float vi3x1 = 0.0f;
121 float vi4x1 = 0.0f;
122 float vi5x1 = 0.0f;
123 float vi6x1 = 0.0f;
124 float vi7x1 = 0.0f;
125 float vi8x1 = 0.0f;
126
127 float vi0x2 = *i0++;
128 float vi1x2 = *i1++;
129 float vi2x2 = *i2++;
130 float vi3x2 = *i3++;
131 float vi4x2 = *i4++;
132 float vi5x2 = *i5++;
133 float vi6x2 = *i6++;
134 float vi7x2 = *i7++;
135 float vi8x2 = *i8++;
136
137 size_t w = input_width;
138 for (; w > 2 * sizeof(float); w -= 2 * sizeof(float)) {
139 const float vi0x3 = i0[0];
140 const float vi1x3 = i1[0];
141 const float vi2x3 = i2[0];
142 const float vi3x3 = i3[0];
143 const float vi4x3 = i4[0];
144 const float vi5x3 = i5[0];
145 const float vi6x3 = i6[0];
146 const float vi7x3 = i7[0];
147 const float vi8x3 = i8[0];
148
149 const float vi0x4 = i0[1];
150 i0 += 2;
151 const float vi1x4 = i1[1];
152 i1 += 2;
153 const float vi2x4 = i2[1];
154 i2 += 2;
155 const float vi3x4 = i3[1];
156 i3 += 2;
157 const float vi4x4 = i4[1];
158 i4 += 2;
159 const float vi5x4 = i5[1];
160 i5 += 2;
161 const float vi6x4 = i6[1];
162 i6 += 2;
163 const float vi7x4 = i7[1];
164 i7 += 2;
165 const float vi8x4 = i8[1];
166 i8 += 2;
167
168 float vo0p0 = vbias + vi0x0 * vk00;
169 float vo1p0 = vbias + vi2x0 * vk00;
170 float vo2p0 = vbias + vi4x0 * vk00;
171 float vo0p1 = vi1x0 * vk10;
172 float vo1p1 = vi3x0 * vk10;
173 float vo2p1 = vi5x0 * vk10;
174 vo0p0 += vi2x0 * vk20;
175 vo1p0 += vi4x0 * vk20;
176 vo2p0 += vi6x0 * vk20;
177 vo0p1 += vi3x0 * vk30;
178 vo1p1 += vi5x0 * vk30;
179 vo2p1 += vi7x0 * vk30;
180 vo0p0 += vi4x0 * vk40;
181 vo1p0 += vi6x0 * vk40;
182 vo2p0 += vi8x0 * vk40;
183
184 vi0x0 = vi0x2;
185 vi1x0 = vi1x2;
186 vi2x0 = vi2x2;
187 vi3x0 = vi3x2;
188 vi4x0 = vi4x2;
189 vi5x0 = vi5x2;
190 vi6x0 = vi6x2;
191 vi7x0 = vi7x2;
192 vi8x0 = vi8x2;
193
194 vo0p1 += vi0x1 * vk01;
195 vo1p1 += vi2x1 * vk01;
196 vo2p1 += vi4x1 * vk01;
197 vo0p0 += vi1x1 * vk11;
198 vo1p0 += vi3x1 * vk11;
199 vo2p0 += vi5x1 * vk11;
200 vo0p1 += vi2x1 * vk21;
201 vo1p1 += vi4x1 * vk21;
202 vo2p1 += vi6x1 * vk21;
203 vo0p0 += vi3x1 * vk31;
204 vo1p0 += vi5x1 * vk31;
205 vo2p0 += vi7x1 * vk31;
206 vo0p1 += vi4x1 * vk41;
207 vo1p1 += vi6x1 * vk41;
208 vo2p1 += vi8x1 * vk41;
209
210 vi0x1 = vi0x3;
211 vi1x1 = vi1x3;
212 vi2x1 = vi2x3;
213 vi3x1 = vi3x3;
214 vi4x1 = vi4x3;
215 vi5x1 = vi5x3;
216 vi6x1 = vi6x3;
217 vi7x1 = vi7x3;
218 vi8x1 = vi8x3;
219
220 vo0p0 += vi0x2 * vk02;
221 vo1p0 += vi2x2 * vk02;
222 vo2p0 += vi4x2 * vk02;
223 vo0p1 += vi1x2 * vk12;
224 vo1p1 += vi3x2 * vk12;
225 vo2p1 += vi5x2 * vk12;
226 vo0p0 += vi2x2 * vk22;
227 vo1p0 += vi4x2 * vk22;
228 vo2p0 += vi6x2 * vk22;
229 vo0p1 += vi3x2 * vk32;
230 vo1p1 += vi5x2 * vk32;
231 vo2p1 += vi7x2 * vk32;
232 vo0p0 += vi4x2 * vk42;
233 vo1p0 += vi6x2 * vk42;
234 vo2p0 += vi8x2 * vk42;
235
236 vi0x2 = vi0x4;
237 vi1x2 = vi1x4;
238 vi2x2 = vi2x4;
239 vi3x2 = vi3x4;
240 vi4x2 = vi4x4;
241 vi5x2 = vi5x4;
242 vi6x2 = vi6x4;
243 vi7x2 = vi7x4;
244 vi8x2 = vi8x4;
245
246 vo0p1 += vi0x3 * vk03;
247 vo1p1 += vi2x3 * vk03;
248 vo2p1 += vi4x3 * vk03;
249 vo0p0 += vi1x3 * vk13;
250 vo1p0 += vi3x3 * vk13;
251 vo2p0 += vi5x3 * vk13;
252 vo0p1 += vi2x3 * vk23;
253 vo1p1 += vi4x3 * vk23;
254 vo2p1 += vi6x3 * vk23;
255 vo0p0 += vi3x3 * vk33;
256 vo1p0 += vi5x3 * vk33;
257 vo2p0 += vi7x3 * vk33;
258 vo0p1 += vi4x3 * vk43;
259 vo1p1 += vi6x3 * vk43;
260 vo2p1 += vi8x3 * vk43;
261
262 vo0p0 += vi0x4 * vk04;
263 vo1p0 += vi2x4 * vk04;
264 vo2p0 += vi4x4 * vk04;
265 vo0p1 += vi1x4 * vk14;
266 vo1p1 += vi3x4 * vk14;
267 vo2p1 += vi5x4 * vk14;
268 vo0p0 += vi2x4 * vk24;
269 vo1p0 += vi4x4 * vk24;
270 vo2p0 += vi6x4 * vk24;
271 vo0p1 += vi3x4 * vk34;
272 vo1p1 += vi5x4 * vk34;
273 vo2p1 += vi7x4 * vk34;
274 vo0p0 += vi4x4 * vk44;
275 vo1p0 += vi6x4 * vk44;
276 vo2p0 += vi8x4 * vk44;
277
278 vo0p0 += vo0p1;
279 vo1p0 += vo1p1;
280 vo2p0 += vo2p1;
281
282 float vo0 = math_max_f32(vo0p0, vmin);
283 float vo1 = math_max_f32(vo1p0, vmin);
284 float vo2 = math_max_f32(vo2p0, vmin);
285
286 vo0 = math_min_f32(vo0, vmax);
287 vo1 = math_min_f32(vo1, vmax);
288 vo2 = math_min_f32(vo2, vmax);
289
290 *o2++ = vo2;
291 *o1++ = vo1;
292 *o0++ = vo0;
293 }
294 if XNN_LIKELY(w == 2 * sizeof(float)) {
295 const float vi0x3 = *i0++;
296 const float vi1x3 = *i1++;
297 const float vi2x3 = *i2++;
298 const float vi3x3 = *i3++;
299 const float vi4x3 = *i4++;
300 const float vi5x3 = *i5++;
301 const float vi6x3 = *i6++;
302 const float vi7x3 = *i7++;
303 const float vi8x3 = *i8++;
304
305 float vo0p0 = vbias + vi0x0 * vk00;
306 float vo1p0 = vbias + vi2x0 * vk00;
307 float vo2p0 = vbias + vi4x0 * vk00;
308 float vo0p1 = vi1x0 * vk10;
309 float vo1p1 = vi3x0 * vk10;
310 float vo2p1 = vi5x0 * vk10;
311 vo0p0 += vi2x0 * vk20;
312 vo1p0 += vi4x0 * vk20;
313 vo2p0 += vi6x0 * vk20;
314 vo0p1 += vi3x0 * vk30;
315 vo1p1 += vi5x0 * vk30;
316 vo2p1 += vi7x0 * vk30;
317 vo0p0 += vi4x0 * vk40;
318 vo1p0 += vi6x0 * vk40;
319 vo2p0 += vi8x0 * vk40;
320
321 vo0p1 += vi0x1 * vk01;
322 vo1p1 += vi2x1 * vk01;
323 vo2p1 += vi4x1 * vk01;
324 vo0p0 += vi1x1 * vk11;
325 vo1p0 += vi3x1 * vk11;
326 vo2p0 += vi5x1 * vk11;
327 vo0p1 += vi2x1 * vk21;
328 vo1p1 += vi4x1 * vk21;
329 vo2p1 += vi6x1 * vk21;
330 vo0p0 += vi3x1 * vk31;
331 vo1p0 += vi5x1 * vk31;
332 vo2p0 += vi7x1 * vk31;
333 vo0p1 += vi4x1 * vk41;
334 vo1p1 += vi6x1 * vk41;
335 vo2p1 += vi8x1 * vk41;
336
337 vo0p0 += vi0x2 * vk02;
338 vo1p0 += vi2x2 * vk02;
339 vo2p0 += vi4x2 * vk02;
340 vo0p1 += vi1x2 * vk12;
341 vo1p1 += vi3x2 * vk12;
342 vo2p1 += vi5x2 * vk12;
343 vo0p0 += vi2x2 * vk22;
344 vo1p0 += vi4x2 * vk22;
345 vo2p0 += vi6x2 * vk22;
346 vo0p1 += vi3x2 * vk32;
347 vo1p1 += vi5x2 * vk32;
348 vo2p1 += vi7x2 * vk32;
349 vo0p0 += vi4x2 * vk42;
350 vo1p0 += vi6x2 * vk42;
351 vo2p0 += vi8x2 * vk42;
352
353 vo0p1 += vi0x3 * vk03;
354 vo1p1 += vi2x3 * vk03;
355 vo2p1 += vi4x3 * vk03;
356 vo0p0 += vi1x3 * vk13;
357 vo1p0 += vi3x3 * vk13;
358 vo2p0 += vi5x3 * vk13;
359 vo0p1 += vi2x3 * vk23;
360 vo1p1 += vi4x3 * vk23;
361 vo2p1 += vi6x3 * vk23;
362 vo0p0 += vi3x3 * vk33;
363 vo1p0 += vi5x3 * vk33;
364 vo2p0 += vi7x3 * vk33;
365 vo0p1 += vi4x3 * vk43;
366 vo1p1 += vi6x3 * vk43;
367 vo2p1 += vi8x3 * vk43;
368
369 vo0p0 += vo0p1;
370 vo1p0 += vo1p1;
371 vo2p0 += vo2p1;
372
373 float vo0 = math_max_f32(vo0p0, vmin);
374 float vo1 = math_max_f32(vo1p0, vmin);
375 float vo2 = math_max_f32(vo2p0, vmin);
376
377 vo0 = math_min_f32(vo0, vmax);
378 vo1 = math_min_f32(vo1, vmax);
379 vo2 = math_min_f32(vo2, vmax);
380
381 *o2++ = vo2;
382 *o1++ = vo1;
383 *o0++ = vo0;
384 } else {
385 float vo0p0 = vbias + vi0x0 * vk00;
386 float vo1p0 = vbias + vi2x0 * vk00;
387 float vo2p0 = vbias + vi4x0 * vk00;
388 float vo0p1 = vi1x0 * vk10;
389 float vo1p1 = vi3x0 * vk10;
390 float vo2p1 = vi5x0 * vk10;
391 vo0p0 += vi2x0 * vk20;
392 vo1p0 += vi4x0 * vk20;
393 vo2p0 += vi6x0 * vk20;
394 vo0p1 += vi3x0 * vk30;
395 vo1p1 += vi5x0 * vk30;
396 vo2p1 += vi7x0 * vk30;
397 vo0p0 += vi4x0 * vk40;
398 vo1p0 += vi6x0 * vk40;
399 vo2p0 += vi8x0 * vk40;
400
401 vo0p1 += vi0x1 * vk01;
402 vo1p1 += vi2x1 * vk01;
403 vo2p1 += vi4x1 * vk01;
404 vo0p0 += vi1x1 * vk11;
405 vo1p0 += vi3x1 * vk11;
406 vo2p0 += vi5x1 * vk11;
407 vo0p1 += vi2x1 * vk21;
408 vo1p1 += vi4x1 * vk21;
409 vo2p1 += vi6x1 * vk21;
410 vo0p0 += vi3x1 * vk31;
411 vo1p0 += vi5x1 * vk31;
412 vo2p0 += vi7x1 * vk31;
413 vo0p1 += vi4x1 * vk41;
414 vo1p1 += vi6x1 * vk41;
415 vo2p1 += vi8x1 * vk41;
416
417 vo0p0 += vi0x2 * vk02;
418 vo1p0 += vi2x2 * vk02;
419 vo2p0 += vi4x2 * vk02;
420 vo0p1 += vi1x2 * vk12;
421 vo1p1 += vi3x2 * vk12;
422 vo2p1 += vi5x2 * vk12;
423 vo0p0 += vi2x2 * vk22;
424 vo1p0 += vi4x2 * vk22;
425 vo2p0 += vi6x2 * vk22;
426 vo0p1 += vi3x2 * vk32;
427 vo1p1 += vi5x2 * vk32;
428 vo2p1 += vi7x2 * vk32;
429 vo0p0 += vi4x2 * vk42;
430 vo1p0 += vi6x2 * vk42;
431 vo2p0 += vi8x2 * vk42;
432
433 vo0p0 += vo0p1;
434 vo1p0 += vo1p1;
435 vo2p0 += vo2p1;
436
437 float vo0 = math_max_f32(vo0p0, vmin);
438 float vo1 = math_max_f32(vo1p0, vmin);
439 float vo2 = math_max_f32(vo2p0, vmin);
440
441 vo0 = math_min_f32(vo0, vmax);
442 vo1 = math_min_f32(vo1, vmax);
443 vo2 = math_min_f32(vo2, vmax);
444
445 *o2++ = vo2;
446 *o1++ = vo1;
447 *o0++ = vo0;
448 }
449
450 i0 = (const float*) ((uintptr_t) i5);
451 i1 = (const float*) ((uintptr_t) i6);
452 i2 = (const float*) ((uintptr_t) i7);
453 i3 = (const float*) ((uintptr_t) i8);
454 i4 = (const float*) ((uintptr_t) i3 + input_width);
455 i5 = (const float*) ((uintptr_t) i4 + input_width);
456 i6 = (const float*) ((uintptr_t) i5 + input_width);
457 i7 = (const float*) ((uintptr_t) i6 + input_width);
458 i8 = (const float*) ((uintptr_t) i7 + input_width);
459
460 o0 = o2;
461 o1 = (float*) ((uintptr_t) o0 + output_width);
462 o2 = (float*) ((uintptr_t) o1 + output_width);
463
464 output_height = doz(output_height, 3);
465 padded_input_height = doz(padded_input_height, 6);
466 } while (output_height != 0);
467 }
468