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
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
|
/*
* Copyright (c) 2017 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "helpers.h"
#if defined(FIXED_POINT_POSITION)
#include "fixed_point.h"
#define ADD_OP(a, b) ADD_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE)
#define SUB_OP(a, b) SUB_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE)
#define MUL_OP(a, b) MUL_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define INVSQRT_OP(a) INVSQRT_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define SQCVT_SAT(a) SQCVT_SAT_OP_EXPAND((a), DATA_TYPE, FIXED_POINT_POSITION)
#else /* FIXED_POINT_POSITION */
#define ADD_OP(a, b) ((a) + (b))
#define SUB_OP(a, b) ((a) - (b))
#define MUL_OP(a, b) ((a) * (b))
#define INVSQRT_OP(a) rsqrt((a))
#define SQCVT_SAT(a) (a)
#endif /* FIXED_POINT_POSITION */
/** Apply batch normalization.
*
* @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/QS16/F32
* @param[in] input_stride_x Stride of the first source tensor 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 first source tensor 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 first 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 first source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor 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 tensor 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 destination 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 tensor
* @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr
* @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
* @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
* @param[in] var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr
* @param[in] var_stride_x Stride of the var tensor in X dimension (in bytes)
* @param[in] var_step_x var_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] var_offset_first_element_in_bytes The offset of the first element in the var source tensor
* @param[in] beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr
* @param[in] beta_stride_x Stride of the beta source tensor in X dimension (in bytes)
* @param[in] beta_step_x beta_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor
* @param[in] gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr
* @param[in] gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes)
* @param[in] gamma_step_x gamma_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor
* @param[in] epsilon Epsilon parameter in the batch normalization equation
*/
__kernel void batchnormalization_layer(TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output),
VECTOR_DECLARATION(mean),
VECTOR_DECLARATION(var),
VECTOR_DECLARATION(beta),
VECTOR_DECLARATION(gamma),
float epsilon)
{
Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
Vector var = CONVERT_TO_VECTOR_STRUCT(var);
Vector beta = CONVERT_TO_VECTOR_STRUCT(beta);
Vector gamma = CONVERT_TO_VECTOR_STRUCT(gamma);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
_in = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
denominator = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
numerator = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
x_bar = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
gamma_vec = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
beta_vec = 0;
const int current_slice = get_global_id(2);
_in = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
denominator = *((__global DATA_TYPE *)(var.ptr + current_slice * var.stride_x));
denominator = INVSQRT_OP(ADD_OP(denominator, SQCVT_SAT(epsilon)));
// Calculate x bar and store results
numerator = *((__global DATA_TYPE *)(mean.ptr + current_slice * mean.stride_x));
numerator = SUB_OP(_in, numerator);
x_bar = MUL_OP(numerator, denominator);
gamma_vec = *((__global DATA_TYPE *)(gamma.ptr + current_slice * beta.stride_x));
beta_vec = *((__global DATA_TYPE *)(beta.ptr + current_slice * beta.stride_x));
VSTORE(VEC_SIZE)
(ADD_OP(MUL_OP(gamma_vec, x_bar), beta_vec), 0, (__global DATA_TYPE *)out.ptr);
}
|
