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_2x1_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_3x3p1__scalar_2x1_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
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
50 float* o0 = output;
51 float* o1 = (float*) ((uintptr_t) o0 + input_width);
52
53 size_t output_height = input_height;
54 do {
55 if XNN_UNPREDICTABLE(output_height < 2) {
56 i2 = zero;
57 o1 = o0;
58 }
59 if XNN_UNPREDICTABLE(output_height < 3) {
60 i3 = zero;
61 }
62
63 float vi0x0 = 0.0f;
64 float vi1x0 = 0.0f;
65 float vi2x0 = 0.0f;
66 float vi3x0 = 0.0f;
67
68 float vi0x1 = *i0++;
69 float vi1x1 = *i1++;
70 float vi2x1 = *i2++;
71 float vi3x1 = *i3++;
72
73 size_t w = input_width;
74 for (; w > 1 * sizeof(float); w -= 1 * sizeof(float)) {
75 const float vi0x2 = *i0++;
76 const float vi1x2 = *i1++;
77 const float vi2x2 = *i2++;
78 const float vi3x2 = *i3++;
79
80 float vo0p0 = vbias + vi0x0 * vk00;
81 float vo1p0 = vbias + vi1x0 * vk00;
82 float vo0p1 = vi1x0 * vk10;
83 float vo1p1 = vi2x0 * vk10;
84 vo0p0 += vi2x0 * vk20;
85 vo1p0 += vi3x0 * vk20;
86
87 vi0x0 = vi0x1;
88 vi1x0 = vi1x1;
89 vi2x0 = vi2x1;
90 vi3x0 = vi3x1;
91
92 vo0p1 += vi0x1 * vk01;
93 vo1p1 += vi1x1 * vk01;
94 vo0p0 += vi1x1 * vk11;
95 vo1p0 += vi2x1 * vk11;
96 vo0p1 += vi2x1 * vk21;
97 vo1p1 += vi3x1 * vk21;
98
99 vi0x1 = vi0x2;
100 vi1x1 = vi1x2;
101 vi2x1 = vi2x2;
102 vi3x1 = vi3x2;
103
104 vo0p0 += vi0x2 * vk02;
105 vo1p0 += vi1x2 * vk02;
106 vo0p1 += vi1x2 * vk12;
107 vo1p1 += vi2x2 * vk12;
108 vo0p0 += vi2x2 * vk22;
109 vo1p0 += vi3x2 * vk22;
110
111 vo0p0 += vo0p1;
112 vo1p0 += vo1p1;
113
114 float vo0 = math_max_f32(vo0p0, vmin);
115 float vo1 = math_max_f32(vo1p0, vmin);
116
117 vo0 = math_min_f32(vo0, vmax);
118 vo1 = math_min_f32(vo1, vmax);
119
120 *o1++ = vo1;
121 *o0++ = vo0;
122 }
123 // Always process the last pixel separately to account for right edge.
124 assert(w == 1 * sizeof(float));
125 {
126 float vo0p0 = vbias + vi0x0 * vk00;
127 float vo1p0 = vbias + vi1x0 * vk00;
128 float vo0p1 = vi1x0 * vk10;
129 float vo1p1 = vi2x0 * vk10;
130 vo0p0 += vi2x0 * vk20;
131 vo1p0 += vi3x0 * vk20;
132
133 vo0p1 += vi0x1 * vk01;
134 vo1p1 += vi1x1 * vk01;
135 vo0p0 += vi1x1 * vk11;
136 vo1p0 += vi2x1 * vk11;
137 vo0p1 += vi2x1 * vk21;
138 vo1p1 += vi3x1 * vk21;
139
140 vo0p0 += vo0p1;
141 vo1p0 += vo1p1;
142
143 float vo0 = math_max_f32(vo0p0, vmin);
144 float vo1 = math_max_f32(vo1p0, vmin);
145
146 vo0 = math_min_f32(vo0, vmax);
147 vo1 = math_min_f32(vo1, vmax);
148
149 *o1++ = vo1;
150 *o0++ = vo0;
151 }
152
153 i0 = (const float*) ((uintptr_t) i2 - input_width);
154 i1 = (const float*) ((uintptr_t) i3 - input_width);
155 i2 = (const float*) ((uintptr_t) i1 + input_width);
156 i3 = (const float*) ((uintptr_t) i2 + input_width);
157
158 o0 = o1;
159 o1 = (float*) ((uintptr_t) o0 + input_width);
160
161 output_height = doz(output_height, 2);
162 } while (output_height != 0);
163 }
164