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
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
|
/*
* Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved
* Copyright (c) 2016, 2017 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"
#if defined(DATA_TYPE) && defined(DEPTH_OUT) && defined(BATCH_IN) && defined(HEIGHT_IN) && \
defined(WIDTH_IN) && defined(ZERO_VALUE)
/** Perform space to batch with input of 4D and NCHW format
*
* @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
* @attention Output tensor depth should be given as a preprocessor argument using -DDEPTH_OUT=size.
* e.g. -DDEPTH_OUT=16
* @attention Input tensor batch should be given as a preprocessor argument using -DBATCH_IN=size.
* e.g. -DBATCH_IN=16
* @attention Input tensor height should be given as a preprocessor argument using -DHEIGHT_IN=size.
* e.g. -DHEIGHT_IN=16
* @attention Input tensor width should be given as a preprocessor argument using -DHEIGHT_IN=size.
* e.g. -DWIDTH_IN=16
* @attention The value to be set by pad value using -DZERO_VALUE=value. e.g. -DZERO_VALUE=0
*
* @param[in] input_ptr Pointer to the source tensor. Supported
* data types: U8/S8/U16/S16/F16/U32/S32/F32
* @param[in] input_stride_x Stride of the 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 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 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_stride_w Stride of the destination tensor in W
* dimension (in bytes)
* @param[in] input_step_w input_stride_w * number of elements along
* W processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the
* 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_stride_w Stride of the destination tensor in W
* dimension (in bytes)
* @param[in] output_step_w output_stride_w * number of elements
* along W 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] block_size_ptr Pointer to the source tensor. Supported
* data types: S32
* @param[in] block_size_stride_x Stride of the source tensor in X
* dimension (in bytes)
* @param[in] block_size_step_x block_size_stride_x * number of elements
* along X processed per workitem(in bytes)
* @param[in] block_size_offset_first_element_in_bytes The offset of the first element in the
* destination tensor
* @param[in] padding_size_ptr Pointer to the source tensor. Supported
* data types: S32
* @param[in] padding_size_stride_x Stride of the source tensor in X
* dimension (in bytes)
* @param[in] padding_size_step_x padding_size_stride_x * number of
* elements along X processed per workitem
* (in bytes)
* @param[in] padding_size_stride_y Stride of the source tensor in Y
* dimension (in bytes)
* @param[in] padding_size_step_y padding_size_stride_y * number of
* elements along Y processed per workitem
* (in bytes)
* @param[in] padding_size_offset_first_element_in_bytes The offset of the first element in the
* destination tensor
*/
__kernel void space_to_batch_4d_nchw(TENSOR4D_DECLARATION(input), TENSOR4D_DECLARATION(output),
VECTOR_DECLARATION(block_size),
IMAGE_DECLARATION(padding_size))
{
Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT(input, 0);
Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(output, DEPTH_OUT);
int block_size_x = *((__global int *)(block_size_ptr));
int block_size_y = *((__global int *)(block_size_ptr + block_size_stride_x));
int shift_x = (get_global_id(2) / DEPTH_OUT / BATCH_IN) % block_size_x;
int shift_y = (get_global_id(2) / DEPTH_OUT / BATCH_IN) / block_size_x;
int in_index[4] = {
0,
};
in_index[0] = get_global_id(0) * block_size_x + shift_x - *((__global int *)(padding_size_ptr));
in_index[1] = get_global_id(1) * block_size_y + shift_y -
*((__global int *)(padding_size_ptr + padding_size_stride_y));
in_index[2] = get_global_id(2) % DEPTH_OUT;
in_index[3] = (get_global_id(2) / DEPTH_OUT) % BATCH_IN;
if (in_index[0] < 0 || in_index[0] >= WIDTH_IN || in_index[1] < 0 || in_index[1] >= HEIGHT_IN)
{
*((__global DATA_TYPE *)out.ptr) = (DATA_TYPE)ZERO_VALUE;
}
else
{
*((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(
&in, in_index[0], in_index[1], in_index[2], in_index[3]));
}
}
#endif // defined(DATA_TYPE) && defined(DEPTH_OUT) && defined(BATCH_IN) && defined(HEIGHT_IN) &&
// defined(WIDTH_IN) && defined(ZERO_VALUE)
#if defined(DATA_TYPE) && defined(HEIGHT_OUT) && defined(BATCH_IN) && defined(HEIGHT_IN) && \
defined(WIDTH_IN) && defined(ZERO_VALUE) && defined(VEC_SIZE)
/** Perform space to batch with input of 4D and NHWC format
*
* @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
* @attention Output tensor depth should be given as a preprocessor argument using
* -DHEIGHT_OUT=size. e.g. -DHEIGHT_OUT=16
* @attention Input tensor batch should be given as a preprocessor argument using -DBATCH_IN=size.
* e.g. -DBATCH_IN=16
* @attention Input tensor height should be given as a preprocessor argument using -DHEIGHT_IN=size.
* e.g. -DHEIGHT_IN=16
* @attention Input tensor width should be given as a preprocessor argument using -DHEIGHT_IN=size.
* e.g. -DWIDTH_IN=16
* @attention The value to be set by pad value using -DZERO_VALUE=value. e.g. -DZERO_VALUE=0
* @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g.
* -DVEC_SIZE=16
*
* @param[in] input_ptr Pointer to the source tensor. Supported
* data types: U8/S8/U16/S16/F16/U32/S32/F32
* @param[in] input_stride_x Stride of the 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 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 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_stride_w Stride of the destination tensor in W
* dimension (in bytes)
* @param[in] input_step_w input_stride_w * number of elements along
* W processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the
* 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_stride_w Stride of the destination tensor in W
* dimension (in bytes)
* @param[in] output_step_w output_stride_w * number of elements
* along W 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] block_size_ptr Pointer to the source tensor. Supported
* data types: S32
* @param[in] block_size_stride_x Stride of the source tensor in X
* dimension (in bytes)
* @param[in] block_size_step_x block_size_stride_x * number of elements
* along X processed per workitem(in bytes)
* @param[in] block_size_offset_first_element_in_bytes The offset of the first element in the
* destination tensor
* @param[in] padding_size_ptr Pointer to the source tensor. Supported
* data types: S32
* @param[in] padding_size_stride_x Stride of the source tensor in X
* dimension (in bytes)
* @param[in] padding_size_step_x padding_size_stride_x * number of
* elements along X processed per workitem
* (in bytes)
* @param[in] padding_size_stride_y Stride of the source tensor in Y
* dimension (in bytes)
* @param[in] padding_size_step_y padding_size_stride_y * number of
* elements along Y processed per workitem
* (in bytes)
* @param[in] padding_size_offset_first_element_in_bytes The offset of the first element in the
* destination tensor
*/
__kernel void space_to_batch_4d_nhwc(TENSOR4D_DECLARATION(input), TENSOR4D_DECLARATION(output),
VECTOR_DECLARATION(block_size),
IMAGE_DECLARATION(padding_size))
{
Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT(input, 0);
Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(output, HEIGHT_OUT);
int block_size_x = *((__global int *)(block_size_ptr));
int block_size_y = *((__global int *)(block_size_ptr + block_size_stride_x));
int shift_x = (get_global_id(2) / HEIGHT_OUT / BATCH_IN) % block_size_x;
int shift_y = (get_global_id(2) / HEIGHT_OUT / BATCH_IN) / block_size_x;
int in_index[4] = {
0,
};
in_index[0] = get_global_id(0) * VEC_SIZE;
in_index[1] = get_global_id(1) * block_size_x + shift_x - *((__global int *)(padding_size_ptr));
in_index[2] = get_global_id(2) % HEIGHT_OUT * block_size_y + shift_y -
*((__global int *)(padding_size_ptr + padding_size_stride_y));
in_index[3] = (get_global_id(2) / HEIGHT_OUT) % BATCH_IN;
if (in_index[1] < 0 || in_index[1] >= WIDTH_IN || in_index[2] < 0 || in_index[2] >= HEIGHT_IN)
{
VSTORE(VEC_SIZE)
((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))ZERO_VALUE, 0, (__global DATA_TYPE *)out.ptr);
}
else
{
VSTORE(VEC_SIZE)
(CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor4D_offset(&in, in_index[0], in_index[1],
in_index[2], in_index[3])),
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)),
0, (__global DATA_TYPE *)out.ptr);
}
}
#endif // defined(DATA_TYPE) && defined(HEIGHT_OUT) && defined(BATCH_IN) && defined(HEIGHT_IN) &&
// defined(WIDTH_IN) && defined(ZERO_VALUE) && defined(VEC_SIZE)
|
