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_1x1_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_3x3s2p1__scalar_1x1_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 >= 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
47 const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
48 const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
49 if XNN_UNPREDICTABLE(padding_top != 0) {
50 i0 = zero;
51 }
52 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
53
54 float* o0 = output;
55
56 size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
57 size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
58 do {
59 if XNN_UNPREDICTABLE(padded_input_height < 4) {
60 i2 = zero;
61 }
62
63 float vi0x0 = 0.0f;
64 float vi1x0 = 0.0f;
65 float vi2x0 = 0.0f;
66
67 size_t w = input_width;
68 for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
69 const float vi0x1 = i0[0];
70 const float vi1x1 = i1[0];
71 const float vi2x1 = i2[0];
72
73 float vo0p0 = vbias + vi0x0 * vk00;
74 float vo0p1 = vi1x0 * vk10;
75 float vo0p2 = vi2x0 * vk20;
76
77 const float vi0x2 = i0[1];
78 i0 += 2;
79 const float vi1x2 = i1[1];
80 i1 += 2;
81 const float vi2x2 = i2[1];
82 i2 += 2;
83
84 vo0p0 += vi0x1 * vk01;
85 vo0p1 += vi1x1 * vk11;
86 vo0p2 += vi2x1 * vk21;
87
88 vi0x0 = vi0x2;
89 vi1x0 = vi1x2;
90 vi2x0 = vi2x2;
91
92 vo0p0 += vi0x2 * vk02;
93 vo0p1 += vi1x2 * vk12;
94 vo0p2 += vi2x2 * vk22;
95
96 vo0p0 += vo0p1;
97 vo0p0 += vo0p2;
98
99 float vo0 = math_max_f32(vo0p0, vmin);
100
101 vo0 = math_min_f32(vo0, vmax);
102
103 *o0++ = vo0;
104 }
105 // Potentially process the last pixel.
106 assert(w <= 1 * sizeof(float));
107 if (w != 0) {
108 const float vi0x1 = *i0++;
109 const float vi1x1 = *i1++;
110 const float vi2x1 = *i2++;
111
112 float vo0p0 = vbias + vi0x0 * vk00;
113 float vo0p1 = vi1x0 * vk10;
114 float vo0p2 = vi2x0 * vk20;
115
116 vo0p0 += vi0x1 * vk01;
117 vo0p1 += vi1x1 * vk11;
118 vo0p2 += vi2x1 * vk21;
119
120 vo0p0 += vo0p1;
121 vo0p0 += vo0p2;
122
123 float vo0 = math_max_f32(vo0p0, vmin);
124
125 vo0 = math_min_f32(vo0, vmax);
126
127 *o0++ = vo0;
128 }
129
130 i0 = (const float*) ((uintptr_t) i1);
131 i1 = (const float*) ((uintptr_t) i2);
132 i2 = (const float*) ((uintptr_t) i1 + input_width);
133
134
135 output_height -= 1;
136 padded_input_height -= 2;
137 } while (output_height != 0);
138 }
139