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
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
|
/*
* Copyright (c) 2018-2020 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(BATCH_SIZE) && defined(DATA_TYPE) && defined(WIDTH_IN) && defined(HEIGHT_IN)
/** Calculate the space to batch conversion.
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
* @note The block shape tensor rank must be passed at compile time using -DBLOCK_SHAPE_DIM. e.g. -DBLOCK_SHAPE_DIM=2
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: All
* @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 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 first source image
* @param[in] paddings_ptr Pointer to the second source image. Supported data types: S32
* @param[in] paddings_stride_x Stride of the paddinds tensor in X dimension (in bytes)
* @param[in] paddings_step_x paddings_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] paddings_stride_y Stride of the paddinds tensor in Y dimension (in bytes)
* @param[in] paddings_step_y paddings_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] paddingse_offset_first_element_in_bytes The offset of the first element in the second source image
* @param[in] block_shape_ptr Pointer to the block shape tensor. Supported data types: S32
* @param[in] block_shape_stride_x Stride of the block shape tensor in X dimension (in bytes)
* @param[in] block_shape_step_x block_shape_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] block_shape_stride_y Stride of the block shape tensor in Y dimension (in bytes)
* @param[in] block_shape_step_y block_shape_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] block_shape_offset_first_element_in_bytes The offset of the first element in the block shapetensor
* @param[in] batch_id The output tensor batch id
* @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 image
*/
__kernel void space_to_batch_nchw(
TENSOR4D_DECLARATION(input),
IMAGE_DECLARATION(paddings),
VECTOR_DECLARATION(block_shape),
const int batch_id,
TENSOR3D_DECLARATION(output))
{
Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
Image pad = CONVERT_TO_IMAGE_STRUCT_NO_STEP(paddings);
Vector block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape);
Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
const int pad_left_x = *((__global int *)offset(&pad, 0, 0));
const int pad_right_x = *((__global int *)offset(&pad, 1, 0));
const int pad_left_y = *((__global int *)offset(&pad, 0, 1));
const int pad_right_y = *((__global int *)offset(&pad, 1, 1));
int block_x = *((__global int *)vector_offset(&block, 0));
int block_y = *((__global int *)vector_offset(&block, 1));
const int out_x = get_global_id(0);
const int out_y = get_global_id(1);
const int z = get_global_id(2);
const int pos_x = out_x * block_x + ((batch_id / BATCH_IN) % block_x);
const int pos_y = out_y * block_y + ((batch_id / BATCH_IN) / block_x);
if(((pos_y >= pad_left_y) && (pos_y < pad_left_y + HEIGHT_IN) && (pos_x >= pad_left_x) && (pos_x < pad_left_x + WIDTH_IN)))
{
const int w = batch_id % BATCH_IN;
const int in_x = pos_x - pad_left_x;
const int in_y = pos_y - pad_left_y;
*((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, in_x, in_y, z, w));
}
}
/** Calculate the space to batch conversion. (NHWC)
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
* @note The block shape tensor rank must be passed at compile time using -DBLOCK_SHAPE_DIM. e.g. -DBLOCK_SHAPE_DIM=2
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: All
* @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 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 first source image
* @param[in] paddings_ptr Pointer to the second source image. Supported data types: S32
* @param[in] paddings_stride_x Stride of the paddinds tensor in X dimension (in bytes)
* @param[in] paddings_step_x paddings_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] paddings_stride_y Stride of the paddinds tensor in Y dimension (in bytes)
* @param[in] paddings_step_y paddings_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] paddingse_offset_first_element_in_bytes The offset of the first element in the second source image
* @param[in] block_shape_ptr Pointer to the block shape tensor. Supported data types: S32
* @param[in] block_shape_stride_x Stride of the block shape tensor in X dimension (in bytes)
* @param[in] block_shape_step_x block_shape_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] block_shape_stride_y Stride of the block shape tensor in Y dimension (in bytes)
* @param[in] block_shape_step_y block_shape_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] block_shape_offset_first_element_in_bytes The offset of the first element in the block shapetensor
* @param[in] batch_id The output tensor batch id
* @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 image
*/
__kernel void space_to_batch_nhwc(
TENSOR4D_DECLARATION(input),
IMAGE_DECLARATION(paddings),
VECTOR_DECLARATION(block_shape),
const int batch_id,
TENSOR3D_DECLARATION(output))
{
Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
Image pad = CONVERT_TO_IMAGE_STRUCT_NO_STEP(paddings);
Vector block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape);
Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
const int pad_left_x = *((__global int *)offset(&pad, 0, 0));
const int pad_right_x = *((__global int *)offset(&pad, 1, 0));
const int pad_left_y = *((__global int *)offset(&pad, 0, 1));
const int pad_right_y = *((__global int *)offset(&pad, 1, 1));
int block_x = *((__global int *)vector_offset(&block, 0));
int block_y = *((__global int *)vector_offset(&block, 1));
const int out_x = get_global_id(1);
const int out_y = get_global_id(2);
const int z = get_global_id(0);
const int pos_x = out_x * block_x + ((batch_id / BATCH_IN) % block_x);
const int pos_y = out_y * block_y + ((batch_id / BATCH_IN) / block_x);
if(((pos_y >= pad_left_y) && (pos_y < pad_left_y + HEIGHT_IN) && (pos_x >= pad_left_x) && (pos_x < pad_left_x + WIDTH_IN)))
{
const int w = batch_id % BATCH_IN;
const int in_x = pos_x - pad_left_x;
const int in_y = pos_y - pad_left_y;
*((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, w));
}
}
#endif // defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(WIDTH_IN) && defined(HEIGHT_IN)
#if defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) && defined(PAD_LEFT_X) && defined(PAD_RIGHT_X) && defined(PAD_LEFT_Y) && defined(PAD_RIGHT_Y) && defined(WIDTH_IN) && defined(HEIGHT_IN)
/** Calculate the space to batch conversion.
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
* @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2
* @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2
* @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2
* @note The starting pad value of x must be passed at compile time using -DPAD_LEFT_X. e.g. -DPAD_LEFT_X=2
* @note The ending pad value of x must be passed at compile time using -DPAD_RIGHT_X. e.g. -DPAD_RIGHT_X=2
* @note The starting pad value of y must be passed at compile time using -DPAD_LEFT_Y. e.g. -DPAD_LEFT_Y=2
* @note The ending pad value of y must be passed at compile time using -DPAD_RIGHT_Y. e.g. -DPAD_RIGHT_X=2
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: All
* @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 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 first source image
* @param[in] batch_id The output tensor batch id
* @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 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 space_to_batch_static_nchw(
TENSOR4D_DECLARATION(input),
const int batch_id,
TENSOR3D_DECLARATION(output))
{
Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
int block_x = BLOCK_SHAPE_X;
int block_y = BLOCK_SHAPE_Y;
const int out_x = get_global_id(0);
const int out_y = get_global_id(1);
const int z = get_global_id(2);
const int pos_x = out_x * block_x + ((batch_id / BATCH_IN) % block_x);
const int pos_y = out_y * block_y + ((batch_id / BATCH_IN) / block_x);
if(pos_y >= PAD_LEFT_Y && pos_y < PAD_LEFT_Y + HEIGHT_IN && pos_x >= PAD_LEFT_X && pos_x < PAD_LEFT_X + WIDTH_IN)
{
const int w = batch_id % BATCH_IN;
const int in_x = pos_x - PAD_LEFT_X;
const int in_y = pos_y - PAD_LEFT_Y;
*((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, in_x, in_y, z, w));
}
}
/** Calculate the space to batch conversion. (NHWC)
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
* @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2
* @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2
* @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2
* @note The starting pad value of x must be passed at compile time using -DPAD_LEFT_X. e.g. -DPAD_LEFT_X=2
* @note The ending pad value of x must be passed at compile time using -DPAD_RIGHT_X. e.g. -DPAD_RIGHT_X=2
* @note The starting pad value of y must be passed at compile time using -DPAD_LEFT_Y. e.g. -DPAD_LEFT_Y=2
* @note The ending pad value of y must be passed at compile time using -DPAD_RIGHT_Y. e.g. -DPAD_RIGHT_X=2
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: All
* @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 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 first source image
* @param[in] batch_id The output tensor batch id
* @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 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 space_to_batch_static_nhwc(
TENSOR4D_DECLARATION(input),
const int batch_id,
TENSOR3D_DECLARATION(output))
{
Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
int block_x = BLOCK_SHAPE_X;
int block_y = BLOCK_SHAPE_Y;
const int out_x = get_global_id(1);
const int out_y = get_global_id(2);
const int z = get_global_id(0);
const int pos_x = out_x * block_x + ((batch_id / BATCH_IN) % block_x);
const int pos_y = out_y * block_y + ((batch_id / BATCH_IN) / block_x);
if(pos_y >= PAD_LEFT_Y && pos_y < PAD_LEFT_Y + HEIGHT_IN && pos_x >= PAD_LEFT_X && pos_x < PAD_LEFT_X + WIDTH_IN)
{
const int w = batch_id % BATCH_IN;
const int in_x = pos_x - PAD_LEFT_X;
const int in_y = pos_y - PAD_LEFT_Y;
*((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, w));
}
}
#endif // defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) && defined(PAD_LEFT_X) && defined(PAD_RIGHT_X) && defined(PAD_LEFT_Y) && defined(PAD_RIGHT_Y) && defined(WIDTH_IN) && defined(HEIGHT_IN)
|
