1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/3x3s2p1-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_3x3s2p1__scalar_3x1(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_3x3s2p1__scalar_3x1(
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 >= 0);
30 assert(padding_top <= 1);
31
32 const float vmin = params->scalar.min;
33 const float vmax = params->scalar.max;
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 vk10 = weights[4];
40 const float vk11 = weights[5];
41 const float vk12 = weights[6];
42 const float vk20 = weights[7];
43 const float vk21 = weights[8];
44 const float vk22 = weights[9];
45
46 const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
47
48 const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
49 const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
50 if XNN_UNPREDICTABLE(padding_top != 0) {
51 i0 = zero;
52 }
53 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
54 const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
55 const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
56 const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
57 const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
58
59 float* o0 = output;
60 float* o1 = (float*) ((uintptr_t) o0 + output_width);
61 float* o2 = (float*) ((uintptr_t) o1 + output_width);
62
63 size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
64 size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
65 do {
66 if XNN_UNPREDICTABLE(padded_input_height < 4) {
67 i2 = zero;
68 }
69 if XNN_UNPREDICTABLE(padded_input_height < 5) {
70 i3 = zero;
71 o1 = o0;
72 }
73 if XNN_UNPREDICTABLE(padded_input_height < 6) {
74 i4 = zero;
75 }
76 if XNN_UNPREDICTABLE(padded_input_height < 7) {
77 i5 = zero;
78 o2 = o1;
79 }
80 if XNN_UNPREDICTABLE(padded_input_height < 8) {
81 i6 = 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 float vi6x0 = 0.0f;
91
92 size_t w = input_width;
93 for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
94 const float vi0x1 = i0[0];
95 const float vi1x1 = i1[0];
96 const float vi2x1 = i2[0];
97 const float vi3x1 = i3[0];
98 const float vi4x1 = i4[0];
99 const float vi5x1 = i5[0];
100 const float vi6x1 = i6[0];
101
102 float vo0p0 = vbias + vi0x0 * vk00;
103 float vo1p0 = vbias + vi2x0 * vk00;
104 float vo2p0 = vbias + vi4x0 * vk00;
105 vo0p0 += vi1x0 * vk10;
106 vo1p0 += vi3x0 * vk10;
107 vo2p0 += vi5x0 * vk10;
108 vo0p0 += vi2x0 * vk20;
109 vo1p0 += vi4x0 * vk20;
110 vo2p0 += vi6x0 * vk20;
111
112 const float vi0x2 = i0[1];
113 i0 += 2;
114 const float vi1x2 = i1[1];
115 i1 += 2;
116 const float vi2x2 = i2[1];
117 i2 += 2;
118 const float vi3x2 = i3[1];
119 i3 += 2;
120 const float vi4x2 = i4[1];
121 i4 += 2;
122 const float vi5x2 = i5[1];
123 i5 += 2;
124 const float vi6x2 = i6[1];
125 i6 += 2;
126
127 vo0p0 += vi0x1 * vk01;
128 vo1p0 += vi2x1 * vk01;
129 vo2p0 += vi4x1 * vk01;
130 vo0p0 += vi1x1 * vk11;
131 vo1p0 += vi3x1 * vk11;
132 vo2p0 += vi5x1 * vk11;
133 vo0p0 += vi2x1 * vk21;
134 vo1p0 += vi4x1 * vk21;
135 vo2p0 += vi6x1 * vk21;
136
137 vi0x0 = vi0x2;
138 vi1x0 = vi1x2;
139 vi2x0 = vi2x2;
140 vi3x0 = vi3x2;
141 vi4x0 = vi4x2;
142 vi5x0 = vi5x2;
143 vi6x0 = vi6x2;
144
145 vo0p0 += vi0x2 * vk02;
146 vo1p0 += vi2x2 * vk02;
147 vo2p0 += vi4x2 * vk02;
148 vo0p0 += vi1x2 * vk12;
149 vo1p0 += vi3x2 * vk12;
150 vo2p0 += vi5x2 * vk12;
151 vo0p0 += vi2x2 * vk22;
152 vo1p0 += vi4x2 * vk22;
153 vo2p0 += vi6x2 * vk22;
154
155
156 float vo0 = math_max_f32(vo0p0, vmin);
157 float vo1 = math_max_f32(vo1p0, vmin);
158 float vo2 = math_max_f32(vo2p0, vmin);
159
160 vo0 = math_min_f32(vo0, vmax);
161 vo1 = math_min_f32(vo1, vmax);
162 vo2 = math_min_f32(vo2, vmax);
163
164 *o2++ = vo2;
165 *o1++ = vo1;
166 *o0++ = vo0;
167 }
168 // Potentially process the last pixel.
169 assert(w <= 1 * sizeof(float));
170 if (w != 0) {
171 const float vi0x1 = *i0++;
172 const float vi1x1 = *i1++;
173 const float vi2x1 = *i2++;
174 const float vi3x1 = *i3++;
175 const float vi4x1 = *i4++;
176 const float vi5x1 = *i5++;
177 const float vi6x1 = *i6++;
178
179 float vo0p0 = vbias + vi0x0 * vk00;
180 float vo1p0 = vbias + vi2x0 * vk00;
181 float vo2p0 = vbias + vi4x0 * vk00;
182 vo0p0 += vi1x0 * vk10;
183 vo1p0 += vi3x0 * vk10;
184 vo2p0 += vi5x0 * vk10;
185 vo0p0 += vi2x0 * vk20;
186 vo1p0 += vi4x0 * vk20;
187 vo2p0 += vi6x0 * vk20;
188
189 vo0p0 += vi0x1 * vk01;
190 vo1p0 += vi2x1 * vk01;
191 vo2p0 += vi4x1 * vk01;
192 vo0p0 += vi1x1 * vk11;
193 vo1p0 += vi3x1 * vk11;
194 vo2p0 += vi5x1 * vk11;
195 vo0p0 += vi2x1 * vk21;
196 vo1p0 += vi4x1 * vk21;
197 vo2p0 += vi6x1 * vk21;
198
199
200 float vo0 = math_max_f32(vo0p0, vmin);
201 float vo1 = math_max_f32(vo1p0, vmin);
202 float vo2 = math_max_f32(vo2p0, vmin);
203
204 vo0 = math_min_f32(vo0, vmax);
205 vo1 = math_min_f32(vo1, vmax);
206 vo2 = math_min_f32(vo2, vmax);
207
208 *o2++ = vo2;
209 *o1++ = vo1;
210 *o0++ = vo0;
211 }
212
213 i0 = (const float*) ((uintptr_t) i5);
214 i1 = (const float*) ((uintptr_t) i6);
215 i2 = (const float*) ((uintptr_t) i1 + input_width);
216 i3 = (const float*) ((uintptr_t) i2 + input_width);
217 i4 = (const float*) ((uintptr_t) i3 + input_width);
218 i5 = (const float*) ((uintptr_t) i4 + input_width);
219 i6 = (const float*) ((uintptr_t) i5 + input_width);
220
221 o0 = o2;
222 o1 = (float*) ((uintptr_t) o0 + output_width);
223 o2 = (float*) ((uintptr_t) o1 + output_width);
224
225 output_height = doz(output_height, 3);
226 padded_input_height = doz(padded_input_height, 6);
227 } while (output_height != 0);
228 }
229