1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/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 <math.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f32_prelu_ukernel__wasm_2x4(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride,const union xnn_f32_output_params params[restrict static1])18 void xnn_f32_prelu_ukernel__wasm_2x4(
19 size_t rows,
20 size_t channels,
21 const float*restrict input,
22 size_t input_stride,
23 const float*restrict weights,
24 float*restrict output,
25 size_t output_stride,
26 const union xnn_f32_output_params params[restrict static 1])
27 {
28 assert(rows != 0);
29 assert(channels != 0);
30 assert(channels % sizeof(float) == 0);
31
32 const float* i0 = input;
33 float* o0 = output;
34 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
35 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
36 if XNN_UNPREDICTABLE(rows < 2) {
37 i1 = i0;
38 o1 = o0;
39 }
40
41 const size_t input_increment = input_stride * 2 - channels;
42 const size_t output_increment = output_stride * 2 - channels;
43
44 const float vmin = params->scalar.min;
45 const float vmax = params->scalar.max;
46 do {
47 const float* w = weights;
48 size_t c = channels;
49 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
50 const float vw0 = w[0];
51 const float vw1 = w[1];
52 const float vw2 = w[2];
53 const float vw3 = w[3];
54
55 const float vi0x0 = i0[0];
56 const float vi0x1 = i0[1];
57 const float vi0x2 = i0[2];
58 const float vi0x3 = i0[3];
59 i0 += 4;
60 const float vi1x0 = i1[0];
61 const float vi1x1 = i1[1];
62 const float vi1x2 = i1[2];
63 const float vi1x3 = i1[3];
64 i1 += 4;
65
66 float vacc0x0 = signbit(vi0x0) ? vi0x0 * vw0 : vi0x0;
67 float vacc0x1 = signbit(vi0x1) ? vi0x1 * vw1 : vi0x1;
68 float vacc0x2 = signbit(vi0x2) ? vi0x2 * vw2 : vi0x2;
69 float vacc0x3 = signbit(vi0x3) ? vi0x3 * vw3 : vi0x3;
70 float vacc1x0 = signbit(vi1x0) ? vi1x0 * vw0 : vi1x0;
71 float vacc1x1 = signbit(vi1x1) ? vi1x1 * vw1 : vi1x1;
72 float vacc1x2 = signbit(vi1x2) ? vi1x2 * vw2 : vi1x2;
73 float vacc1x3 = signbit(vi1x3) ? vi1x3 * vw3 : vi1x3;
74
75 vacc0x0 = __builtin_wasm_max_f32(vacc0x0, vmin);
76 vacc0x1 = __builtin_wasm_max_f32(vacc0x1, vmin);
77 vacc0x2 = __builtin_wasm_max_f32(vacc0x2, vmin);
78 vacc0x3 = __builtin_wasm_max_f32(vacc0x3, vmin);
79 vacc1x0 = __builtin_wasm_max_f32(vacc1x0, vmin);
80 vacc1x1 = __builtin_wasm_max_f32(vacc1x1, vmin);
81 vacc1x2 = __builtin_wasm_max_f32(vacc1x2, vmin);
82 vacc1x3 = __builtin_wasm_max_f32(vacc1x3, vmin);
83
84 vacc0x0 = __builtin_wasm_min_f32(vacc0x0, vmax);
85 vacc0x1 = __builtin_wasm_min_f32(vacc0x1, vmax);
86 vacc0x2 = __builtin_wasm_min_f32(vacc0x2, vmax);
87 vacc0x3 = __builtin_wasm_min_f32(vacc0x3, vmax);
88 vacc1x0 = __builtin_wasm_min_f32(vacc1x0, vmax);
89 vacc1x1 = __builtin_wasm_min_f32(vacc1x1, vmax);
90 vacc1x2 = __builtin_wasm_min_f32(vacc1x2, vmax);
91 vacc1x3 = __builtin_wasm_min_f32(vacc1x3, vmax);
92
93 o0[0] = vacc0x0;
94 o0[1] = vacc0x1;
95 o0[2] = vacc0x2;
96 o0[3] = vacc0x3;
97 o0 += 4;
98 o1[0] = vacc1x0;
99 o1[1] = vacc1x1;
100 o1[2] = vacc1x2;
101 o1[3] = vacc1x3;
102 o1 += 4;
103
104 w += 4;
105 }
106 for (; c != 0; c -= sizeof(float)) {
107 const float vw = *w++;
108
109 const float vi0 = *i0++;
110 const float vi1 = *i1++;
111
112 float vacc0 = signbit(vi0) ? vi0 * vw : vi0;
113 float vacc1 = signbit(vi1) ? vi1 * vw : vi1;
114
115 vacc0 = __builtin_wasm_max_f32(vacc0, vmin);
116 vacc1 = __builtin_wasm_max_f32(vacc1, vmin);
117
118 vacc0 = __builtin_wasm_min_f32(vacc0, vmax);
119 vacc1 = __builtin_wasm_min_f32(vacc1, vmax);
120
121 *o0++ = vacc0;
122 *o1++ = vacc1;
123 }
124 i0 = (const float*) ((uintptr_t) i0 + input_increment);
125 o0 = (float*) ((uintptr_t) o0 + output_increment);
126 i1 = (const float*) ((uintptr_t) i1 + input_increment);
127 o1 = (float*) ((uintptr_t) o1 + output_increment);
128 if XNN_UNPREDICTABLE(rows < 4) {
129 i1 = i0;
130 o1 = o0;
131 }
132 rows = doz(rows, 2);
133 } while (rows != 0);
134 }
135