1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/5x5p2-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_5x5p2__scalar_2x1_acc3(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_5x5p2__scalar_2x1_acc3(
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 == 2);
30
31 const float vmin = params->scalar.min;
32 const float vmax = params->scalar.max;
33
34 const float vbias = weights[0];
35 const float vk00 = weights[1];
36 const float vk01 = weights[2];
37 const float vk02 = weights[3];
38 const float vk03 = weights[4];
39 const float vk04 = weights[5];
40 const float vk10 = weights[6];
41 const float vk11 = weights[7];
42 const float vk12 = weights[8];
43 const float vk13 = weights[9];
44 const float vk14 = weights[10];
45 const float vk20 = weights[11];
46 const float vk21 = weights[12];
47 const float vk22 = weights[13];
48 const float vk23 = weights[14];
49 const float vk24 = weights[15];
50 const float vk30 = weights[16];
51 const float vk31 = weights[17];
52 const float vk32 = weights[18];
53 const float vk33 = weights[19];
54 const float vk34 = weights[20];
55 const float vk40 = weights[21];
56 const float vk41 = weights[22];
57 const float vk42 = weights[23];
58 const float vk43 = weights[24];
59 const float vk44 = weights[25];
60
61 const float* i0 = zero;
62 const float* i1 = zero;
63 const float* i2 = input;
64 const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
65 const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
66 const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
67
68 float* o0 = output;
69 float* o1 = (float*) ((uintptr_t) o0 + input_width);
70
71 size_t output_height = input_height;
72 do {
73 if XNN_UNPREDICTABLE(output_height < 2) {
74 i3 = zero;
75 o1 = o0;
76 }
77 if XNN_UNPREDICTABLE(output_height < 3) {
78 i4 = zero;
79 }
80 if XNN_UNPREDICTABLE(output_height < 4) {
81 i5 = zero;
82 }
83
84 float vi0x0 = 0.0f;
85 float vi1x0 = 0.0f;
86 float vi2x0 = 0.0f;
87 float vi3x0 = 0.0f;
88 float vi4x0 = 0.0f;
89 float vi5x0 = 0.0f;
90
91 float vi0x1 = 0.0f;
92 float vi1x1 = 0.0f;
93 float vi2x1 = 0.0f;
94 float vi3x1 = 0.0f;
95 float vi4x1 = 0.0f;
96 float vi5x1 = 0.0f;
97
98 float vi0x2 = *i0++;
99 float vi1x2 = *i1++;
100 float vi2x2 = *i2++;
101 float vi3x2 = *i3++;
102 float vi4x2 = *i4++;
103 float vi5x2 = *i5++;
104
105 size_t w = input_width;
106 if (w > 1 * sizeof(float)) {
107 float vi0x3 = *i0++;
108 float vi1x3 = *i1++;
109 float vi2x3 = *i2++;
110 float vi3x3 = *i3++;
111 float vi4x3 = *i4++;
112 float vi5x3 = *i5++;
113
114 for (; w > 2 * sizeof(float); w -= 1 * sizeof(float)) {
115 const float vi0x4 = *i0++;
116 const float vi1x4 = *i1++;
117 const float vi2x4 = *i2++;
118 const float vi3x4 = *i3++;
119 const float vi4x4 = *i4++;
120 const float vi5x4 = *i5++;
121
122 float vo0p0 = vbias + vi0x0 * vk00;
123 float vo1p0 = vbias + vi1x0 * vk00;
124 float vo0p1 = vi1x0 * vk10;
125 float vo1p1 = vi2x0 * vk10;
126 float vo0p2 = vi2x0 * vk20;
127 float vo1p2 = vi3x0 * vk20;
128 vo0p0 += vi3x0 * vk30;
129 vo1p0 += vi4x0 * vk30;
130 vo0p1 += vi4x0 * vk40;
131 vo1p1 += vi5x0 * vk40;
132
133 vi0x0 = vi0x1;
134 vi1x0 = vi1x1;
135 vi2x0 = vi2x1;
136 vi3x0 = vi3x1;
137 vi4x0 = vi4x1;
138 vi5x0 = vi5x1;
139
140 vo0p2 += vi0x1 * vk01;
141 vo1p2 += vi1x1 * vk01;
142 vo0p0 += vi1x1 * vk11;
143 vo1p0 += vi2x1 * vk11;
144 vo0p1 += vi2x1 * vk21;
145 vo1p1 += vi3x1 * vk21;
146 vo0p2 += vi3x1 * vk31;
147 vo1p2 += vi4x1 * vk31;
148 vo0p0 += vi4x1 * vk41;
149 vo1p0 += vi5x1 * vk41;
150
151 vi0x1 = vi0x2;
152 vi1x1 = vi1x2;
153 vi2x1 = vi2x2;
154 vi3x1 = vi3x2;
155 vi4x1 = vi4x2;
156 vi5x1 = vi5x2;
157
158 vo0p1 += vi0x2 * vk02;
159 vo1p1 += vi1x2 * vk02;
160 vo0p2 += vi1x2 * vk12;
161 vo1p2 += vi2x2 * vk12;
162 vo0p0 += vi2x2 * vk22;
163 vo1p0 += vi3x2 * vk22;
164 vo0p1 += vi3x2 * vk32;
165 vo1p1 += vi4x2 * vk32;
166 vo0p2 += vi4x2 * vk42;
167 vo1p2 += vi5x2 * vk42;
168
169 vi0x2 = vi0x3;
170 vi1x2 = vi1x3;
171 vi2x2 = vi2x3;
172 vi3x2 = vi3x3;
173 vi4x2 = vi4x3;
174 vi5x2 = vi5x3;
175
176 vo0p0 += vi0x3 * vk03;
177 vo1p0 += vi1x3 * vk03;
178 vo0p1 += vi1x3 * vk13;
179 vo1p1 += vi2x3 * vk13;
180 vo0p2 += vi2x3 * vk23;
181 vo1p2 += vi3x3 * vk23;
182 vo0p0 += vi3x3 * vk33;
183 vo1p0 += vi4x3 * vk33;
184 vo0p1 += vi4x3 * vk43;
185 vo1p1 += vi5x3 * vk43;
186
187 vi0x3 = vi0x4;
188 vi1x3 = vi1x4;
189 vi2x3 = vi2x4;
190 vi3x3 = vi3x4;
191 vi4x3 = vi4x4;
192 vi5x3 = vi5x4;
193
194 vo0p2 += vi0x4 * vk04;
195 vo1p2 += vi1x4 * vk04;
196 vo0p0 += vi1x4 * vk14;
197 vo1p0 += vi2x4 * vk14;
198 vo0p1 += vi2x4 * vk24;
199 vo1p1 += vi3x4 * vk24;
200 vo0p2 += vi3x4 * vk34;
201 vo1p2 += vi4x4 * vk34;
202 vo0p0 += vi4x4 * vk44;
203 vo1p0 += vi5x4 * vk44;
204
205 vo0p0 += vo0p1;
206 vo1p0 += vo1p1;
207 vo0p0 += vo0p2;
208 vo1p0 += vo1p2;
209
210 float vo0 = math_max_f32(vo0p0, vmin);
211 float vo1 = math_max_f32(vo1p0, vmin);
212
213 vo0 = math_min_f32(vo0, vmax);
214 vo1 = math_min_f32(vo1, vmax);
215
216 *o1++ = vo1;
217 *o0++ = vo0;
218 }
219 assert(w == 2 * sizeof(float));
220 {
221 float vo0p0 = vbias + vi0x0 * vk00;
222 float vo1p0 = vbias + vi1x0 * vk00;
223 float vo0p1 = vi1x0 * vk10;
224 float vo1p1 = vi2x0 * vk10;
225 float vo0p2 = vi2x0 * vk20;
226 float vo1p2 = vi3x0 * vk20;
227 vo0p0 += vi3x0 * vk30;
228 vo1p0 += vi4x0 * vk30;
229 vo0p1 += vi4x0 * vk40;
230 vo1p1 += vi5x0 * vk40;
231
232 vi0x0 = vi0x1;
233 vi1x0 = vi1x1;
234 vi2x0 = vi2x1;
235 vi3x0 = vi3x1;
236 vi4x0 = vi4x1;
237 vi5x0 = vi5x1;
238
239 vo0p2 += vi0x1 * vk01;
240 vo1p2 += vi1x1 * vk01;
241 vo0p0 += vi1x1 * vk11;
242 vo1p0 += vi2x1 * vk11;
243 vo0p1 += vi2x1 * vk21;
244 vo1p1 += vi3x1 * vk21;
245 vo0p2 += vi3x1 * vk31;
246 vo1p2 += vi4x1 * vk31;
247 vo0p0 += vi4x1 * vk41;
248 vo1p0 += vi5x1 * vk41;
249
250 vi0x1 = vi0x2;
251 vi1x1 = vi1x2;
252 vi2x1 = vi2x2;
253 vi3x1 = vi3x2;
254 vi4x1 = vi4x2;
255 vi5x1 = vi5x2;
256
257 vo0p1 += vi0x2 * vk02;
258 vo1p1 += vi1x2 * vk02;
259 vo0p2 += vi1x2 * vk12;
260 vo1p2 += vi2x2 * vk12;
261 vo0p0 += vi2x2 * vk22;
262 vo1p0 += vi3x2 * vk22;
263 vo0p1 += vi3x2 * vk32;
264 vo1p1 += vi4x2 * vk32;
265 vo0p2 += vi4x2 * vk42;
266 vo1p2 += vi5x2 * vk42;
267
268 vi0x2 = vi0x3;
269 vi1x2 = vi1x3;
270 vi2x2 = vi2x3;
271 vi3x2 = vi3x3;
272 vi4x2 = vi4x3;
273 vi5x2 = vi5x3;
274
275 vo0p0 += vi0x3 * vk03;
276 vo1p0 += vi1x3 * vk03;
277 vo0p1 += vi1x3 * vk13;
278 vo1p1 += vi2x3 * vk13;
279 vo0p2 += vi2x3 * vk23;
280 vo1p2 += vi3x3 * vk23;
281 vo0p0 += vi3x3 * vk33;
282 vo1p0 += vi4x3 * vk33;
283 vo0p1 += vi4x3 * vk43;
284 vo1p1 += vi5x3 * vk43;
285
286 vo0p0 += vo0p1;
287 vo1p0 += vo1p1;
288 vo0p0 += vo0p2;
289 vo1p0 += vo1p2;
290
291 float vo0 = math_max_f32(vo0p0, vmin);
292 float vo1 = math_max_f32(vo1p0, vmin);
293
294 vo0 = math_min_f32(vo0, vmax);
295 vo1 = math_min_f32(vo1, vmax);
296
297 *o1++ = vo1;
298 *o0++ = vo0;
299 }
300 w -= 1 * sizeof(float);
301 }
302 assert(w == 1 * sizeof(float));
303 {
304 float vo0p0 = vbias + vi0x0 * vk00;
305 float vo1p0 = vbias + vi1x0 * vk00;
306 float vo0p1 = vi1x0 * vk10;
307 float vo1p1 = vi2x0 * vk10;
308 float vo0p2 = vi2x0 * vk20;
309 float vo1p2 = vi3x0 * vk20;
310 vo0p0 += vi3x0 * vk30;
311 vo1p0 += vi4x0 * vk30;
312 vo0p1 += vi4x0 * vk40;
313 vo1p1 += vi5x0 * vk40;
314
315 vo0p2 += vi0x1 * vk01;
316 vo1p2 += vi1x1 * vk01;
317 vo0p0 += vi1x1 * vk11;
318 vo1p0 += vi2x1 * vk11;
319 vo0p1 += vi2x1 * vk21;
320 vo1p1 += vi3x1 * vk21;
321 vo0p2 += vi3x1 * vk31;
322 vo1p2 += vi4x1 * vk31;
323 vo0p0 += vi4x1 * vk41;
324 vo1p0 += vi5x1 * vk41;
325
326 vo0p1 += vi0x2 * vk02;
327 vo1p1 += vi1x2 * vk02;
328 vo0p2 += vi1x2 * vk12;
329 vo1p2 += vi2x2 * vk12;
330 vo0p0 += vi2x2 * vk22;
331 vo1p0 += vi3x2 * vk22;
332 vo0p1 += vi3x2 * vk32;
333 vo1p1 += vi4x2 * vk32;
334 vo0p2 += vi4x2 * vk42;
335 vo1p2 += vi5x2 * vk42;
336
337 vo0p0 += vo0p1;
338 vo1p0 += vo1p1;
339 vo0p0 += vo0p2;
340 vo1p0 += vo1p2;
341
342 float vo0 = math_max_f32(vo0p0, vmin);
343 float vo1 = math_max_f32(vo1p0, vmin);
344
345 vo0 = math_min_f32(vo0, vmax);
346 vo1 = math_min_f32(vo1, vmax);
347
348 *o1++ = vo1;
349 *o0++ = vo0;
350 }
351
352 i0 = (const float*) ((uintptr_t) i2 - input_width);
353 i1 = (const float*) ((uintptr_t) i3 - input_width);
354 i2 = i3;
355 i3 = i4;
356 i4 = i5;
357 i5 = (const float*) ((uintptr_t) i4 + input_width);
358
359 o0 = o1;
360 o1 = (float*) ((uintptr_t) o0 + input_width);
361
362 output_height = doz(output_height, 2);
363 } while (output_height != 0);
364 }
365