1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/3x3p1-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_3x3p1__scalar_6x1(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_3x3p1__scalar_6x1(
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
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 vk10 = weights[4];
39 const float vk11 = weights[5];
40 const float vk12 = weights[6];
41 const float vk20 = weights[7];
42 const float vk21 = weights[8];
43 const float vk22 = weights[9];
44
45 const float* i0 = zero;
46 const float* i1 = input;
47 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
48 const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
49 const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
50 const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
51 const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
52 const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
53
54 float* o0 = output;
55 float* o1 = (float*) ((uintptr_t) o0 + input_width);
56 float* o2 = (float*) ((uintptr_t) o1 + input_width);
57 float* o3 = (float*) ((uintptr_t) o2 + input_width);
58 float* o4 = (float*) ((uintptr_t) o3 + input_width);
59 float* o5 = (float*) ((uintptr_t) o4 + input_width);
60
61 size_t output_height = input_height;
62 do {
63 if XNN_UNPREDICTABLE(output_height < 2) {
64 i2 = zero;
65 o1 = o0;
66 }
67 if XNN_UNPREDICTABLE(output_height < 3) {
68 i3 = zero;
69 o2 = o1;
70 }
71 if XNN_UNPREDICTABLE(output_height < 4) {
72 i4 = zero;
73 o3 = o2;
74 }
75 if XNN_UNPREDICTABLE(output_height < 5) {
76 i5 = zero;
77 o4 = o3;
78 }
79 if XNN_UNPREDICTABLE(output_height < 6) {
80 i6 = zero;
81 o5 = o4;
82 }
83 if XNN_UNPREDICTABLE(output_height < 7) {
84 i7 = zero;
85 }
86
87 float vi0x0 = 0.0f;
88 float vi1x0 = 0.0f;
89 float vi2x0 = 0.0f;
90 float vi3x0 = 0.0f;
91 float vi4x0 = 0.0f;
92 float vi5x0 = 0.0f;
93 float vi6x0 = 0.0f;
94 float vi7x0 = 0.0f;
95
96 float vi0x1 = *i0++;
97 float vi1x1 = *i1++;
98 float vi2x1 = *i2++;
99 float vi3x1 = *i3++;
100 float vi4x1 = *i4++;
101 float vi5x1 = *i5++;
102 float vi6x1 = *i6++;
103 float vi7x1 = *i7++;
104
105 size_t w = input_width;
106 for (; w > 1 * sizeof(float); w -= 1 * sizeof(float)) {
107 const float vi0x2 = *i0++;
108 const float vi1x2 = *i1++;
109 const float vi2x2 = *i2++;
110 const float vi3x2 = *i3++;
111 const float vi4x2 = *i4++;
112 const float vi5x2 = *i5++;
113 const float vi6x2 = *i6++;
114 const float vi7x2 = *i7++;
115
116 float vo0p0 = vbias + vi0x0 * vk00;
117 float vo1p0 = vbias + vi1x0 * vk00;
118 float vo2p0 = vbias + vi2x0 * vk00;
119 float vo3p0 = vbias + vi3x0 * vk00;
120 float vo4p0 = vbias + vi4x0 * vk00;
121 float vo5p0 = vbias + vi5x0 * vk00;
122 vo0p0 += vi1x0 * vk10;
123 vo1p0 += vi2x0 * vk10;
124 vo2p0 += vi3x0 * vk10;
125 vo3p0 += vi4x0 * vk10;
126 vo4p0 += vi5x0 * vk10;
127 vo5p0 += vi6x0 * vk10;
128 vo0p0 += vi2x0 * vk20;
129 vo1p0 += vi3x0 * vk20;
130 vo2p0 += vi4x0 * vk20;
131 vo3p0 += vi5x0 * vk20;
132 vo4p0 += vi6x0 * vk20;
133 vo5p0 += vi7x0 * vk20;
134
135 vi0x0 = vi0x1;
136 vi1x0 = vi1x1;
137 vi2x0 = vi2x1;
138 vi3x0 = vi3x1;
139 vi4x0 = vi4x1;
140 vi5x0 = vi5x1;
141 vi6x0 = vi6x1;
142 vi7x0 = vi7x1;
143
144 vo0p0 += vi0x1 * vk01;
145 vo1p0 += vi1x1 * vk01;
146 vo2p0 += vi2x1 * vk01;
147 vo3p0 += vi3x1 * vk01;
148 vo4p0 += vi4x1 * vk01;
149 vo5p0 += vi5x1 * vk01;
150 vo0p0 += vi1x1 * vk11;
151 vo1p0 += vi2x1 * vk11;
152 vo2p0 += vi3x1 * vk11;
153 vo3p0 += vi4x1 * vk11;
154 vo4p0 += vi5x1 * vk11;
155 vo5p0 += vi6x1 * vk11;
156 vo0p0 += vi2x1 * vk21;
157 vo1p0 += vi3x1 * vk21;
158 vo2p0 += vi4x1 * vk21;
159 vo3p0 += vi5x1 * vk21;
160 vo4p0 += vi6x1 * vk21;
161 vo5p0 += vi7x1 * vk21;
162
163 vi0x1 = vi0x2;
164 vi1x1 = vi1x2;
165 vi2x1 = vi2x2;
166 vi3x1 = vi3x2;
167 vi4x1 = vi4x2;
168 vi5x1 = vi5x2;
169 vi6x1 = vi6x2;
170 vi7x1 = vi7x2;
171
172 vo0p0 += vi0x2 * vk02;
173 vo1p0 += vi1x2 * vk02;
174 vo2p0 += vi2x2 * vk02;
175 vo3p0 += vi3x2 * vk02;
176 vo4p0 += vi4x2 * vk02;
177 vo5p0 += vi5x2 * vk02;
178 vo0p0 += vi1x2 * vk12;
179 vo1p0 += vi2x2 * vk12;
180 vo2p0 += vi3x2 * vk12;
181 vo3p0 += vi4x2 * vk12;
182 vo4p0 += vi5x2 * vk12;
183 vo5p0 += vi6x2 * vk12;
184 vo0p0 += vi2x2 * vk22;
185 vo1p0 += vi3x2 * vk22;
186 vo2p0 += vi4x2 * vk22;
187 vo3p0 += vi5x2 * vk22;
188 vo4p0 += vi6x2 * vk22;
189 vo5p0 += vi7x2 * vk22;
190
191
192 float vo0 = math_max_f32(vo0p0, vmin);
193 float vo1 = math_max_f32(vo1p0, vmin);
194 float vo2 = math_max_f32(vo2p0, vmin);
195 float vo3 = math_max_f32(vo3p0, vmin);
196 float vo4 = math_max_f32(vo4p0, vmin);
197 float vo5 = math_max_f32(vo5p0, vmin);
198
199 vo0 = math_min_f32(vo0, vmax);
200 vo1 = math_min_f32(vo1, vmax);
201 vo2 = math_min_f32(vo2, vmax);
202 vo3 = math_min_f32(vo3, vmax);
203 vo4 = math_min_f32(vo4, vmax);
204 vo5 = math_min_f32(vo5, vmax);
205
206 *o5++ = vo5;
207 *o4++ = vo4;
208 *o3++ = vo3;
209 *o2++ = vo2;
210 *o1++ = vo1;
211 *o0++ = vo0;
212 }
213 // Always process the last pixel separately to account for right edge.
214 assert(w == 1 * sizeof(float));
215 {
216 float vo0p0 = vbias + vi0x0 * vk00;
217 float vo1p0 = vbias + vi1x0 * vk00;
218 float vo2p0 = vbias + vi2x0 * vk00;
219 float vo3p0 = vbias + vi3x0 * vk00;
220 float vo4p0 = vbias + vi4x0 * vk00;
221 float vo5p0 = vbias + vi5x0 * vk00;
222 vo0p0 += vi1x0 * vk10;
223 vo1p0 += vi2x0 * vk10;
224 vo2p0 += vi3x0 * vk10;
225 vo3p0 += vi4x0 * vk10;
226 vo4p0 += vi5x0 * vk10;
227 vo5p0 += vi6x0 * vk10;
228 vo0p0 += vi2x0 * vk20;
229 vo1p0 += vi3x0 * vk20;
230 vo2p0 += vi4x0 * vk20;
231 vo3p0 += vi5x0 * vk20;
232 vo4p0 += vi6x0 * vk20;
233 vo5p0 += vi7x0 * vk20;
234
235 vo0p0 += vi0x1 * vk01;
236 vo1p0 += vi1x1 * vk01;
237 vo2p0 += vi2x1 * vk01;
238 vo3p0 += vi3x1 * vk01;
239 vo4p0 += vi4x1 * vk01;
240 vo5p0 += vi5x1 * vk01;
241 vo0p0 += vi1x1 * vk11;
242 vo1p0 += vi2x1 * vk11;
243 vo2p0 += vi3x1 * vk11;
244 vo3p0 += vi4x1 * vk11;
245 vo4p0 += vi5x1 * vk11;
246 vo5p0 += vi6x1 * vk11;
247 vo0p0 += vi2x1 * vk21;
248 vo1p0 += vi3x1 * vk21;
249 vo2p0 += vi4x1 * vk21;
250 vo3p0 += vi5x1 * vk21;
251 vo4p0 += vi6x1 * vk21;
252 vo5p0 += vi7x1 * vk21;
253
254
255 float vo0 = math_max_f32(vo0p0, vmin);
256 float vo1 = math_max_f32(vo1p0, vmin);
257 float vo2 = math_max_f32(vo2p0, vmin);
258 float vo3 = math_max_f32(vo3p0, vmin);
259 float vo4 = math_max_f32(vo4p0, vmin);
260 float vo5 = math_max_f32(vo5p0, vmin);
261
262 vo0 = math_min_f32(vo0, vmax);
263 vo1 = math_min_f32(vo1, vmax);
264 vo2 = math_min_f32(vo2, vmax);
265 vo3 = math_min_f32(vo3, vmax);
266 vo4 = math_min_f32(vo4, vmax);
267 vo5 = math_min_f32(vo5, vmax);
268
269 *o5++ = vo5;
270 *o4++ = vo4;
271 *o3++ = vo3;
272 *o2++ = vo2;
273 *o1++ = vo1;
274 *o0++ = vo0;
275 }
276
277 i0 = (const float*) ((uintptr_t) i6 - input_width);
278 i1 = (const float*) ((uintptr_t) i7 - input_width);
279 i2 = (const float*) ((uintptr_t) i1 + input_width);
280 i3 = (const float*) ((uintptr_t) i2 + input_width);
281 i4 = (const float*) ((uintptr_t) i3 + input_width);
282 i5 = (const float*) ((uintptr_t) i4 + input_width);
283 i6 = (const float*) ((uintptr_t) i5 + input_width);
284 i7 = (const float*) ((uintptr_t) i6 + input_width);
285
286 o0 = o5;
287 o1 = (float*) ((uintptr_t) o0 + input_width);
288 o2 = (float*) ((uintptr_t) o1 + input_width);
289 o3 = (float*) ((uintptr_t) o2 + input_width);
290 o4 = (float*) ((uintptr_t) o3 + input_width);
291 o5 = (float*) ((uintptr_t) o4 + input_width);
292
293 output_height = doz(output_height, 6);
294 } while (output_height != 0);
295 }
296