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__wasm(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,const union xnn_f32_output_params params[restrict static1])16 void xnn_f32_dwconv_ukernel_up2x4__wasm(
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 const union xnn_f32_output_params params[restrict static 1])
25 {
26 assert(channels != 0);
27 assert(output_width != 0);
28
29 const float vmin = params->scalar.min;
30 const float vmax = params->scalar.max;
31 do {
32 const float* i0 = input[0];
33 assert(i0 != NULL);
34 const float* i1 = input[1];
35 assert(i1 != NULL);
36 const float* i2 = input[2];
37 assert(i2 != NULL);
38 const float* i3 = input[3];
39 assert(i3 != NULL);
40 input = (const float**) ((uintptr_t) input + input_stride);
41
42 size_t c = channels;
43 const float* w = weights;
44 for (; c >= 2; c -= 2) {
45 float vacc0p0 = w[0];
46 float vacc1p0 = w[1];
47
48
49 const float vi0x0 = i0[0];
50 const float vi0x1 = i0[1];
51 i0 += 2;
52
53 const float vk0x0 = w[2];
54 vacc0p0 += vi0x0 * vk0x0;
55 const float vk0x1 = w[3];
56 vacc1p0 += vi0x1 * vk0x1;
57
58 const float vi1x0 = i1[0];
59 const float vi1x1 = i1[1];
60 i1 += 2;
61
62 const float vk1x0 = w[4];
63 vacc0p0 += vi1x0 * vk1x0;
64 const float vk1x1 = w[5];
65 vacc1p0 += vi1x1 * vk1x1;
66
67 const float vi2x0 = i2[0];
68 const float vi2x1 = i2[1];
69 i2 += 2;
70
71 const float vk2x0 = w[6];
72 vacc0p0 += vi2x0 * vk2x0;
73 const float vk2x1 = w[7];
74 vacc1p0 += vi2x1 * vk2x1;
75
76 const float vi3x0 = i3[0];
77 const float vi3x1 = i3[1];
78 i3 += 2;
79
80 const float vk3x0 = w[8];
81 vacc0p0 += vi3x0 * vk3x0;
82 const float vk3x1 = w[9];
83 vacc1p0 += vi3x1 * vk3x1;
84
85 w += 10;
86
87
88 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin);
89 float vacc1 = __builtin_wasm_max_f32(vacc1p0, vmin);
90
91 vacc0 = __builtin_wasm_min_f32(vacc0, vmax);
92 vacc1 = __builtin_wasm_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 += vi0 * vk0;
104 const float vi1 = *i1++;
105 const float vk1 = w[3];
106 vacc0p0 += vi1 * vk1;
107 const float vi2 = *i2++;
108 const float vk2 = w[5];
109 vacc0p0 += vi2 * vk2;
110 const float vi3 = *i3++;
111 const float vk3 = w[7];
112 vacc0p0 += vi3 * vk3;
113
114
115 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin);
116 vacc0 = __builtin_wasm_min_f32(vacc0, vmax);
117 *output++ = vacc0;
118 }
119
120 output = (float*) ((uintptr_t) output + output_increment);
121 } while (--output_width != 0);
122 }
123