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_minmax_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_dwconv_minmax_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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(channels != 0);
29 assert(output_width != 0);
30
31 const float vmin = params->scalar.min;
32 const float vmax = params->scalar.max;
33 do {
34 const float* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const float*) ((uintptr_t) i0 + input_offset);
38 }
39 const float* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const float*) ((uintptr_t) i1 + input_offset);
43 }
44 const float* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const float*) ((uintptr_t) i2 + input_offset);
48 }
49 input = (const float**) ((uintptr_t) input + input_stride);
50
51 size_t c = channels;
52 const float* w = weights;
53 for (; c >= 2; c -= 2) {
54 float vacc0p0 = w[0];
55 float vacc1p0 = w[1];
56
57
58 const float vi0x0 = i0[0];
59 const float vi0x1 = i0[1];
60 i0 += 2;
61
62 const float vk0x0 = w[2];
63 vacc0p0 = math_muladd_f32(vi0x0, vk0x0, vacc0p0);
64 const float vk0x1 = w[3];
65 vacc1p0 = math_muladd_f32(vi0x1, vk0x1, vacc1p0);
66
67 const float vi1x0 = i1[0];
68 const float vi1x1 = i1[1];
69 i1 += 2;
70
71 const float vk1x0 = w[4];
72 vacc0p0 = math_muladd_f32(vi1x0, vk1x0, vacc0p0);
73 const float vk1x1 = w[5];
74 vacc1p0 = math_muladd_f32(vi1x1, vk1x1, vacc1p0);
75
76 const float vi2x0 = i2[0];
77 const float vi2x1 = i2[1];
78 i2 += 2;
79
80 const float vk2x0 = w[6];
81 vacc0p0 = math_muladd_f32(vi2x0, vk2x0, vacc0p0);
82 const float vk2x1 = w[7];
83 vacc1p0 = math_muladd_f32(vi2x1, vk2x1, vacc1p0);
84
85 w += 8;
86
87
88 float vacc0 = math_max_f32(vacc0p0, vmin);
89 float vacc1 = math_max_f32(vacc1p0, vmin);
90
91 vacc0 = math_min_f32(vacc0, vmax);
92 vacc1 = math_min_f32(vacc1, vmax);
93
94 output[0] = vacc0;
95 output[1] = vacc1;
96 output += 2;
97 }
98 for (; c >= 1; c -= 1) {
99 float vacc0p0 = *w++;
100
101 const float vi0 = *i0++;
102 const float vk0 = w[1];
103 vacc0p0 = math_muladd_f32(vi0, vk0, vacc0p0);
104 const float vi1 = *i1++;
105 const float vk1 = w[3];
106 vacc0p0 = math_muladd_f32(vi1, vk1, vacc0p0);
107 const float vi2 = *i2++;
108 const float vk2 = w[5];
109 vacc0p0 = math_muladd_f32(vi2, vk2, vacc0p0);
110
111
112 float vacc0 = math_max_f32(vacc0p0, vmin);
113 vacc0 = math_min_f32(vacc0, vmax);
114 *output++ = vacc0;
115 }
116
117 output = (float*) ((uintptr_t) output + output_increment);
118 } while (--output_width != 0);
119 }
120