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_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_dwconv_minmax_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_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 const float* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const float*) ((uintptr_t) i3 + input_offset);
53 }
54 input = (const float**) ((uintptr_t) input + input_stride);
55
56 size_t c = channels;
57 const float* w = weights;
58 for (; c >= 2; c -= 2) {
59 float vacc0p0 = w[0];
60 float vacc1p0 = w[1];
61
62
63 const float vi0x0 = i0[0];
64 const float vi0x1 = i0[1];
65 i0 += 2;
66
67 const float vk0x0 = w[2];
68 vacc0p0 = math_muladd_f32(vi0x0, vk0x0, vacc0p0);
69 const float vk0x1 = w[3];
70 vacc1p0 = math_muladd_f32(vi0x1, vk0x1, vacc1p0);
71
72 const float vi1x0 = i1[0];
73 const float vi1x1 = i1[1];
74 i1 += 2;
75
76 const float vk1x0 = w[4];
77 vacc0p0 = math_muladd_f32(vi1x0, vk1x0, vacc0p0);
78 const float vk1x1 = w[5];
79 vacc1p0 = math_muladd_f32(vi1x1, vk1x1, vacc1p0);
80
81 const float vi2x0 = i2[0];
82 const float vi2x1 = i2[1];
83 i2 += 2;
84
85 const float vk2x0 = w[6];
86 vacc0p0 = math_muladd_f32(vi2x0, vk2x0, vacc0p0);
87 const float vk2x1 = w[7];
88 vacc1p0 = math_muladd_f32(vi2x1, vk2x1, vacc1p0);
89
90 const float vi3x0 = i3[0];
91 const float vi3x1 = i3[1];
92 i3 += 2;
93
94 const float vk3x0 = w[8];
95 vacc0p0 = math_muladd_f32(vi3x0, vk3x0, vacc0p0);
96 const float vk3x1 = w[9];
97 vacc1p0 = math_muladd_f32(vi3x1, vk3x1, vacc1p0);
98
99 w += 10;
100
101
102 float vacc0 = math_max_f32(vacc0p0, vmin);
103 float vacc1 = math_max_f32(vacc1p0, vmin);
104
105 vacc0 = math_min_f32(vacc0, vmax);
106 vacc1 = math_min_f32(vacc1, vmax);
107
108 output[0] = vacc0;
109 output[1] = vacc1;
110 output += 2;
111 }
112 for (; c >= 1; c -= 1) {
113 float vacc0p0 = *w++;
114
115 const float vi0 = *i0++;
116 const float vk0 = w[1];
117 vacc0p0 = math_muladd_f32(vi0, vk0, vacc0p0);
118 const float vi1 = *i1++;
119 const float vk1 = w[3];
120 vacc0p0 = math_muladd_f32(vi1, vk1, vacc0p0);
121 const float vi2 = *i2++;
122 const float vk2 = w[5];
123 vacc0p0 = math_muladd_f32(vi2, vk2, vacc0p0);
124 const float vi3 = *i3++;
125 const float vk3 = w[7];
126 vacc0p0 = math_muladd_f32(vi3, vk3, vacc0p0);
127
128
129 float vacc0 = math_max_f32(vacc0p0, vmin);
130 vacc0 = math_min_f32(vacc0, vmax);
131 *output++ = vacc0;
132 }
133
134 output = (float*) ((uintptr_t) output + output_increment);
135 } while (--output_width != 0);
136 }
137