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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
|
/*
* 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"
#endif // FIXED_POINT_POSITION
/** This kernel reshapes the tensor's low three dimensions to single column
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
*
* @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 Y 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 Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] dst_ptr Pointer to the destination tensor. 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] bias_ptr Pointer to the bias tensor. Same as @p src_ptr
* @param[in] bias_stride_x Stride of the bias tensor in X dimension (in bytes)
* @param[in] bias_step_x bias_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] bias_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] width The width of the input tensor
* @param[in] height The height of the input tensor
* @param[in] depth The depth of the input tensor
* @param[in] total_filters Total number of filters. 4th dimension of the weights matrix
*/
__kernel void reshape_to_columns(
TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst),
#ifdef HAS_BIAS
VECTOR_DECLARATION(bias),
#endif /* HAS_BIAS */
uint width, uint height, uint depth, uint total_filters)
{
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
bool is_last_thread = (get_global_id(0) == (get_global_size(0) - 1) && get_global_id(1) == (get_global_size(1) - 1) && get_global_id(2) == (get_global_size(2) - 1));
__global uchar *tmp_src_ptr = src.ptr;
__global uchar *tmp_dst_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(0) * dst_stride_y + get_global_id(1) * width * dst_stride_y + get_global_id(
2) * width * height * dst_stride_y;
#ifdef HAS_BIAS
__global uchar *tmp_bias_ptr = bias_ptr + bias_offset_first_element_in_bytes;
#endif /* HAS_BIAS */
if(is_last_thread)
{
for(uint i = 0; i < total_filters; ++i)
{
*((__global DATA_TYPE *)tmp_dst_ptr) = *((__global DATA_TYPE *)tmp_src_ptr);
#ifdef HAS_BIAS
*((__global DATA_TYPE *)(tmp_dst_ptr + dst_stride_y)) = *((__global DATA_TYPE *)(tmp_bias_ptr));
tmp_bias_ptr += bias_stride_x;
#endif /* HAS_BIAS */
tmp_src_ptr += depth * src_stride_z;
tmp_dst_ptr += dst_stride_x;
}
}
else
{
for(uint i = 0; i < total_filters; ++i)
{
*((__global DATA_TYPE *)tmp_dst_ptr) = *((__global DATA_TYPE *)tmp_src_ptr);
tmp_src_ptr += depth * src_stride_z;
tmp_dst_ptr += dst_stride_x;
}
}
}
#if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_X) && defined(PAD_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT)
/** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/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 Y 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[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] filter_depth The depth of the used filter
* @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
*/
__kernel void im2col_generic(
TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst),
uint filter_depth,
uint src_stride_w,
uint dst_stride_w)
{
const int xc = get_global_id(0); // x coordinate in the convolved tensor
const int yc = get_global_id(1); // y coordinate in the convolved tensor
const int ch = get_global_id(2) % filter_depth; // input feature map
const int batch = get_global_id(2) / filter_depth; // the batch
// Calculate input indeces
const int xi = xc * STRIDE_X - PAD_X;
const int yi = yc * STRIDE_Y - PAD_Y;
// Calculate output indeces
const int xo = ch * KERNEL_WIDTH * KERNEL_HEIGHT;
const int yo = xc + yc * CONVOLVED_WIDTH; // Index of the convolution
__global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch * src_stride_z + batch * src_stride_w;
__global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + yo * dst_stride_y + batch * dst_stride_w)) + xo;
// Linearize convolution elements
for(int y = yi, y_e = yi + KERNEL_HEIGHT; y < y_e; ++y)
{
for(int x = xi, x_e = xi + KERNEL_WIDTH; x < x_e; ++x, ++output_ptr)
{
#if PAD_X == 0 && PAD_Y == 0
*output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
#else // PAD_X == 0 && PAD_Y == 0
if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
{
*output_ptr = 0;
}
else
{
*output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
}
#endif // PAD_X == 0 && PAD_Y == 0
}
}
#ifdef HAS_BIAS
if(ch == (KERNEL_DEPTH - 1))
{
#ifdef FIXED_POINT_POSITION
*output_ptr = (DATA_TYPE)(1 << FIXED_POINT_POSITION);
#else // FIXED_POINT_POSITION
*output_ptr = 1.0f;
#endif // FIXED_POINT_POSITION
}
#endif // HAS_BIAS
}
/** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM when the kernel size is 3x3 and pad_x = pad_y = 0
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/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 Y 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[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] filter_depth The depth of the used filter
* @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
*/
__kernel void im2col_kernel3x3_padx0_pady0(
TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst),
uint filter_depth,
uint src_stride_w,
uint dst_stride_w)
{
const int xc = get_global_id(0); // x coordinate in the convolved tensor
const int yc = get_global_id(1); // y coordinate in the convolved tensor
const int ch = get_global_id(2) % filter_depth; // input feature map
const int batch = get_global_id(2) / filter_depth; // the batch
// Calculate input indeces
const int xi = xc * STRIDE_X;
const int yi = yc * STRIDE_Y;
// Calculate output indeces
const int xo = ch * KERNEL_WIDTH * KERNEL_HEIGHT;
const int yo = xc + yc * CONVOLVED_WIDTH; // Index of the convolution
// Get input and output address
__global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * src_stride_x + yi * src_stride_y + ch * src_stride_z + batch * src_stride_w;
__global DATA_TYPE *output_ptr = (__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + yo * dst_stride_y + batch * dst_stride_w) + xo;
VEC_DATA_TYPE(DATA_TYPE, 3)
row0 = vload3(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y));
VEC_DATA_TYPE(DATA_TYPE, 3)
row1 = vload3(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y));
VEC_DATA_TYPE(DATA_TYPE, 3)
row2 = vload3(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y));
vstore8((VEC_DATA_TYPE(DATA_TYPE, 8))(row0.s012, row1.s012, row2.s01), 0, output_ptr);
*(output_ptr + 8) = row2.s2;
#ifdef HAS_BIAS
if(ch == (KERNEL_DEPTH - 1))
{
#ifdef FIXED_POINT_POSITION
*(output_ptr + 9) = (DATA_TYPE)(1 << FIXED_POINT_POSITION);
#else // FIXED_POINT_POSITION
*(output_ptr + 9) = 1.0f;
#endif // FIXED_POINT_POSITION
}
#endif // HAS_BIAS
}
#endif //defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_X) && defined(PAD_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT)
#if defined(WIDTH_OUTPUT)
/** This kernel performs a reshaping of the output of the convolution layer.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/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 Y 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[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 Z 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] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
*/
__kernel void col2im(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
uint dst_stride_w)
{
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(dst);
// Compute output offset
int idx = get_global_id(0) * dst.stride_z + (get_global_id(1) / WIDTH_OUTPUT) * dst_stride_y + (get_global_id(1) % WIDTH_OUTPUT) * dst_stride_x + get_global_id(2) * dst_stride_w;
// Store value
*((__global DATA_TYPE *)(dst.ptr + idx)) = *((__global DATA_TYPE *)(src.ptr));
}
#endif // defined(WIDTH_OUTPUT)
/** This kernel reshapes the tensor's low three dimensions to single row for GEMM operation
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
* @note In case biases will be added in late stage, -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/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 Y 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 Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] dst_ptr Pointer to the destination tensor. 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_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] width The width of the input tensor
* @param[in] height The height of the input tensor
*/
__kernel void im2col_reduced(
TENSOR3D_DECLARATION(src),
VECTOR_DECLARATION(dst),
uint width, uint height)
{
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
const uint image_size = width * height;
__global uchar *tmp_out_ptr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) + get_global_id(1) * width + get_global_id(2) * image_size) * dst_stride_x;
*((__global DATA_TYPE *)tmp_out_ptr) = *((__global DATA_TYPE *)src.ptr);
#ifdef HAS_BIAS
// If it is the last thread in the 3 dimensional workgroup
if(get_global_id(0) == (get_global_size(0) - 1) && get_global_id(1) == (get_global_size(1) - 1) && get_global_id(2) == (get_global_size(2) - 1))
{
tmp_out_ptr += dst_stride_x;
#ifdef FIXED_POINT_POSITION
*((__global DATA_TYPE *)tmp_out_ptr) = (DATA_TYPE)(1 << FIXED_POINT_POSITION);
#else // FIXED_POINT_POSITION
*((__global DATA_TYPE *)tmp_out_ptr) = (DATA_TYPE)1;
#endif // FIXED_POINT_POSITION
}
#endif // HAS_BIAS
}
|
