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
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
|
/*
* Copyright (c) 2016-2019 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"
/** This kernel performs l2 normalization on x-axis
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along X processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] sum_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] sum_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_step_y sum_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] epsilon Epsilon value
*/
__kernel void l2_normalize_x(
IMAGE_DECLARATION(src),
IMAGE_DECLARATION(sum),
IMAGE_DECLARATION(dst),
DATA_TYPE epsilon)
{
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image sum = CONVERT_TO_IMAGE_STRUCT(sum);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
VEC_DATA_TYPE(DATA_TYPE, 16)
in = vload16(0, (__global DATA_TYPE *)src.ptr);
VEC_DATA_TYPE(DATA_TYPE, 16)
normalize_value = (VEC_DATA_TYPE(DATA_TYPE, 16))rsqrt(fmax(((__global DATA_TYPE *)sum.ptr)[0], epsilon));
vstore16(in * normalize_value, 0, (__global DATA_TYPE *)dst.ptr);
}
/** This kernel performs l2 normalization on y-axis.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along Y processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along X processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] sum_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] sum_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_step_y sum_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] epsilon Epsilon value
*/
__kernel void l2_normalize_y(
IMAGE_DECLARATION(src),
IMAGE_DECLARATION(sum),
IMAGE_DECLARATION(dst),
DATA_TYPE epsilon)
{
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image sum = CONVERT_TO_IMAGE_STRUCT(sum);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
VEC_DATA_TYPE(DATA_TYPE, 16)
in = vload16(0, (__global DATA_TYPE *)src.ptr);
VEC_DATA_TYPE(DATA_TYPE, 16)
sums = vload16(0, (__global DATA_TYPE *)sum.ptr);
VEC_DATA_TYPE(DATA_TYPE, 16)
normalize_value = (VEC_DATA_TYPE(DATA_TYPE, 16))rsqrt(fmax(sums, epsilon));
vstore16(in * normalize_value, 0, (__global DATA_TYPE *)dst.ptr);
}
/** This kernel performs l2 normalization on z-axis.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along Y processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] sum_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] sum_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_step_y sum_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] sum_step_z sum_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] sum_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] epsilon Epsilon value
*/
__kernel void l2_normalize_z(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(sum),
TENSOR3D_DECLARATION(dst),
DATA_TYPE epsilon)
{
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Tensor3D sum = CONVERT_TO_TENSOR3D_STRUCT(sum);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
VEC_DATA_TYPE(DATA_TYPE, 16)
in = vload16(0, (__global DATA_TYPE *)src.ptr);
VEC_DATA_TYPE(DATA_TYPE, 16)
sums = vload16(0, (__global DATA_TYPE *)sum.ptr);
VEC_DATA_TYPE(DATA_TYPE, 16)
normalize_value = (VEC_DATA_TYPE(DATA_TYPE, 16))rsqrt(fmax(sums, epsilon));
vstore16(in * normalize_value, 0, (__global DATA_TYPE *)dst.ptr);
}
|
