1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
|
/*
* Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved
* Copyright (c) 2016-2018 ARM Limited.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "helpers.h"
#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
#define CONST_ONE 1.f
#define DIV_OP(a, b) ((a) / (b))
#define RSQRT_OP(a) DIV_OP(CONST_ONE, sqrt((a)))
// Inverse Square-root Activation
inline TYPE rsqrt_op(TYPE x)
{
return RSQRT_OP(x);
}
#define ACTIVATION_OP2(op, x) op##_op(x)
#define ACTIVATION_OP(op, x) ACTIVATION_OP2(op, x)
#if defined(ACT)
/** This performs an activation function floating point inputs.
*
* @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
* @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
* @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH
* @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
*
* @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image
*/
__kernel void activation_layer_ex(
TENSOR3D_DECLARATION(input)
#ifndef IN_PLACE
,
TENSOR3D_DECLARATION(output)
#endif /* not IN_PLACE */
)
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
#ifdef IN_PLACE
Tensor3D output = input;
#else /* IN_PLACE */
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
#endif /* IN_PLACE */
// Load data
TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
// Perform activation
data = ACTIVATION_OP(ACT, data);
// Store result
VSTORE(VEC_SIZE)
(data, 0, (__global DATA_TYPE *)output.ptr);
}
#endif /* defined(ACT) */
|