1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-scalar.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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_dwconv_ukernel_up2x4__scalar(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_dwconv_ukernel_up2x4__scalar(
17 size_t channels,
18 size_t output_width,
19 const float** input,
20 const float* weights,
21 float* output,
22 size_t input_stride,
23 size_t output_increment,
24 size_t input_offset,
25 const float* zero,
26 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(channels != 0);
29 assert(output_width != 0);
30
31 do {
32 const float* i0 = input[0];
33 assert(i0 != NULL);
34 if XNN_UNPREDICTABLE(i0 != zero) {
35 i0 = (const float*) ((uintptr_t) i0 + input_offset);
36 }
37 const float* i1 = input[1];
38 assert(i1 != NULL);
39 if XNN_UNPREDICTABLE(i1 != zero) {
40 i1 = (const float*) ((uintptr_t) i1 + input_offset);
41 }
42 const float* i2 = input[2];
43 assert(i2 != NULL);
44 if XNN_UNPREDICTABLE(i2 != zero) {
45 i2 = (const float*) ((uintptr_t) i2 + input_offset);
46 }
47 const float* i3 = input[3];
48 assert(i3 != NULL);
49 if XNN_UNPREDICTABLE(i3 != zero) {
50 i3 = (const float*) ((uintptr_t) i3 + input_offset);
51 }
52 input = (const float**) ((uintptr_t) input + input_stride);
53
54 size_t c = channels;
55 const float* w = weights;
56 for (; c >= 2; c -= 2) {
57 float vacc0p0 = w[0];
58 float vacc1p0 = w[1];
59
60
61 const float vi0x0 = i0[0];
62 const float vi0x1 = i0[1];
63 i0 += 2;
64
65 const float vk0x0 = w[2];
66 vacc0p0 += vi0x0 * vk0x0;
67 const float vk0x1 = w[3];
68 vacc1p0 += vi0x1 * vk0x1;
69
70 const float vi1x0 = i1[0];
71 const float vi1x1 = i1[1];
72 i1 += 2;
73
74 const float vk1x0 = w[4];
75 vacc0p0 += vi1x0 * vk1x0;
76 const float vk1x1 = w[5];
77 vacc1p0 += vi1x1 * vk1x1;
78
79 const float vi2x0 = i2[0];
80 const float vi2x1 = i2[1];
81 i2 += 2;
82
83 const float vk2x0 = w[6];
84 vacc0p0 += vi2x0 * vk2x0;
85 const float vk2x1 = w[7];
86 vacc1p0 += vi2x1 * vk2x1;
87
88 const float vi3x0 = i3[0];
89 const float vi3x1 = i3[1];
90 i3 += 2;
91
92 const float vk3x0 = w[8];
93 vacc0p0 += vi3x0 * vk3x0;
94 const float vk3x1 = w[9];
95 vacc1p0 += vi3x1 * vk3x1;
96
97 w += 10;
98
99
100 output[0] = vacc0p0;
101 output[1] = vacc1p0;
102 output += 2;
103 }
104 for (; c >= 1; c -= 1) {
105 float vacc0p0 = *w++;
106
107 const float vi0 = *i0++;
108 const float vk0 = w[1];
109 vacc0p0 += vi0 * vk0;
110 const float vi1 = *i1++;
111 const float vk1 = w[3];
112 vacc0p0 += vi1 * vk1;
113 const float vi2 = *i2++;
114 const float vk2 = w[5];
115 vacc0p0 += vi2 * vk2;
116 const float vi3 = *i3++;
117 const float vk3 = w[7];
118 vacc0p0 += vi3 * vk3;
119
120
121 *output++ = vacc0p0;
122 }
123
124 output = (float*) ((uintptr_t) output + output_increment);
125 } while (--output_width != 0);
126 }
127