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_up2x3__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_up2x3__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 input = (const float**) ((uintptr_t) input + input_stride);
48
49 size_t c = channels;
50 const float* w = weights;
51 for (; c >= 2; c -= 2) {
52 float vacc0p0 = w[0];
53 float vacc1p0 = w[1];
54
55
56 const float vi0x0 = i0[0];
57 const float vi0x1 = i0[1];
58 i0 += 2;
59
60 const float vk0x0 = w[2];
61 vacc0p0 = math_muladd_f32(vi0x0, vk0x0, vacc0p0);
62 const float vk0x1 = w[3];
63 vacc1p0 = math_muladd_f32(vi0x1, vk0x1, vacc1p0);
64
65 const float vi1x0 = i1[0];
66 const float vi1x1 = i1[1];
67 i1 += 2;
68
69 const float vk1x0 = w[4];
70 vacc0p0 = math_muladd_f32(vi1x0, vk1x0, vacc0p0);
71 const float vk1x1 = w[5];
72 vacc1p0 = math_muladd_f32(vi1x1, vk1x1, vacc1p0);
73
74 const float vi2x0 = i2[0];
75 const float vi2x1 = i2[1];
76 i2 += 2;
77
78 const float vk2x0 = w[6];
79 vacc0p0 = math_muladd_f32(vi2x0, vk2x0, vacc0p0);
80 const float vk2x1 = w[7];
81 vacc1p0 = math_muladd_f32(vi2x1, vk2x1, vacc1p0);
82
83 w += 8;
84
85
86 output[0] = vacc0p0;
87 output[1] = vacc1p0;
88 output += 2;
89 }
90 for (; c >= 1; c -= 1) {
91 float vacc0p0 = *w++;
92
93 const float vi0 = *i0++;
94 const float vk0 = w[1];
95 vacc0p0 = math_muladd_f32(vi0, vk0, vacc0p0);
96 const float vi1 = *i1++;
97 const float vk1 = w[3];
98 vacc0p0 = math_muladd_f32(vi1, vk1, vacc0p0);
99 const float vi2 = *i2++;
100 const float vk2 = w[5];
101 vacc0p0 = math_muladd_f32(vi2, vk2, vacc0p0);
102
103
104 *output++ = vacc0p0;
105 }
106
107 output = (float*) ((uintptr_t) output + output_increment);
108 } while (--output_width != 0);
109 }
110