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 <xnnpack/math.h>
13 #include <xnnpack/prelu.h>
14
15
xnn_f32_prelu_ukernel__scalar_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)16 void xnn_f32_prelu_ukernel__scalar_2x4(
17 size_t rows,
18 size_t channels,
19 const float*restrict input,
20 size_t input_stride,
21 const float*restrict weights,
22 float*restrict output,
23 size_t output_stride)
24 {
25 assert(rows != 0);
26 assert(channels != 0);
27 assert(channels % sizeof(float) == 0);
28
29 const float* i0 = input;
30 float* o0 = output;
31 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
32 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
33 if XNN_UNPREDICTABLE(rows < 2) {
34 i1 = i0;
35 o1 = o0;
36 }
37
38 const size_t input_increment = input_stride * 2 - channels;
39 const size_t output_increment = output_stride * 2 - channels;
40
41 do {
42 const float* w = weights;
43 size_t c = channels;
44 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
45 const float vw0 = w[0];
46 const float vw1 = w[1];
47 const float vw2 = w[2];
48 const float vw3 = w[3];
49
50 const float vi0x0 = i0[0];
51 const float vi0x1 = i0[1];
52 const float vi0x2 = i0[2];
53 const float vi0x3 = i0[3];
54 i0 += 4;
55 const float vi1x0 = i1[0];
56 const float vi1x1 = i1[1];
57 const float vi1x2 = i1[2];
58 const float vi1x3 = i1[3];
59 i1 += 4;
60
61 const float vacc0x0 = XNN_UNPREDICTABLE(vi0x0 < 0.0f) ? vi0x0 * vw0 : vi0x0;
62 const float vacc0x1 = XNN_UNPREDICTABLE(vi0x1 < 0.0f) ? vi0x1 * vw1 : vi0x1;
63 const float vacc0x2 = XNN_UNPREDICTABLE(vi0x2 < 0.0f) ? vi0x2 * vw2 : vi0x2;
64 const float vacc0x3 = XNN_UNPREDICTABLE(vi0x3 < 0.0f) ? vi0x3 * vw3 : vi0x3;
65 const float vacc1x0 = XNN_UNPREDICTABLE(vi1x0 < 0.0f) ? vi1x0 * vw0 : vi1x0;
66 const float vacc1x1 = XNN_UNPREDICTABLE(vi1x1 < 0.0f) ? vi1x1 * vw1 : vi1x1;
67 const float vacc1x2 = XNN_UNPREDICTABLE(vi1x2 < 0.0f) ? vi1x2 * vw2 : vi1x2;
68 const float vacc1x3 = XNN_UNPREDICTABLE(vi1x3 < 0.0f) ? vi1x3 * vw3 : vi1x3;
69
70 o0[0] = vacc0x0;
71 o0[1] = vacc0x1;
72 o0[2] = vacc0x2;
73 o0[3] = vacc0x3;
74 o0 += 4;
75 o1[0] = vacc1x0;
76 o1[1] = vacc1x1;
77 o1[2] = vacc1x2;
78 o1[3] = vacc1x3;
79 o1 += 4;
80
81 w += 4;
82 }
83 for (; c != 0; c -= sizeof(float)) {
84 const float vw = *w++;
85
86 const float vi0 = *i0++;
87 const float vi1 = *i1++;
88
89 const float vacc0 = XNN_UNPREDICTABLE(vi0 < 0.0f) ? vi0 * vw : vi0;
90 const float vacc1 = XNN_UNPREDICTABLE(vi1 < 0.0f) ? vi1 * vw : vi1;
91
92 *o0++ = vacc0;
93 *o1++ = vacc1;
94 }
95 i0 = (const float*) ((uintptr_t) i0 + input_increment);
96 o0 = (float*) ((uintptr_t) o0 + output_increment);
97 i1 = (const float*) ((uintptr_t) i1 + input_increment);
98 o1 = (float*) ((uintptr_t) o1 + output_increment);
99 if XNN_UNPREDICTABLE(rows < 4) {
100 i1 = i0;
101 o1 = o0;
102 }
103 rows = doz(rows, 2);
104 } while (rows != 0);
105 }
106