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
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
|
/*
* Copyright (c) 2017-2020 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(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
#define VEC_QUANT VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
#define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
inline VEC_QUANT requantize(VEC_QUANT input, float in_offset, float out_offset, float in_scale, float out_scale)
{
const VEC_FLOAT in_f32 = (CONVERT(input, VEC_FLOAT) - (VEC_FLOAT)((float)in_offset)) * (VEC_FLOAT)((float)in_scale);
const VEC_FLOAT out_f32 = in_f32 / ((VEC_FLOAT)(float)out_scale) + ((VEC_FLOAT)((float)out_offset));
const VEC_QUANT res_q8 = CONVERT_SAT(CONVERT_DOWN(out_f32, VEC_INT), VEC_QUANT);
return res_q8;
}
#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
#if defined(DATA_TYPE) && defined(VEC_SIZE)
#if defined(DEPTH) && defined(ELEMENT_SIZE)
#if defined(INPUT1_WIDTH)
#if ELEMENT_SIZE == 1
#define COND_DATA_TYPE char
#elif ELEMENT_SIZE == 2
#define COND_DATA_TYPE short
#elif ELEMENT_SIZE == 4
#define COND_DATA_TYPE int
#else // ELEMENT_SIZE
#error "Element size not supported"
#endif // ELEMENT_SIZE
#if VEC_SIZE == 2
#define SEQ ((int2)(0, 1))
#elif VEC_SIZE == 4
#define SEQ ((int4)(0, 1, 2, 3))
#elif VEC_SIZE == 8
#define SEQ ((int8)(0, 1, 2, 3, 4, 5, 6, 7))
#elif VEC_SIZE == 16
#define SEQ ((int16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15))
#else // VEC_SIZE
#error "Vector size not supported"
#endif // VEC_SIZE
/** This kernel concatenates two input tensors into the output tensor along the first dimension
*
* @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
* @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
* @note The offset for the first spatial dimension has to be passed at compile time using -DWIDTH_OFFSET. i.e. -DWIDTH_OFFSET=128
* @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
* @note First input tensor width should be given as a preprocessor argument using -DINPUT1_WIDTH=width. e.g. -DINPUT1_WIDTH=8
*
* @param[in] src1_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/F32
* @param[in] src1_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src1_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src1_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src1_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] src2_ptr Pointer to the source tensor. Supported data types: same as @p src1_ptr
* @param[in] src2_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src2_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src2_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src2_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src2_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src2_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src2_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src2_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src2_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 src1_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 source 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_stride_w Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_w output_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] src1_pad_right Right paddings of the first input tensor in unit of elements
* @param[in] src1_pad_left Left paddings of the second input tensor in unit of elements
*/
__kernel void concatenate_width_x2(
TENSOR4D_DECLARATION(src1),
TENSOR4D_DECLARATION(src2),
TENSOR4D_DECLARATION(dst),
uint src1_pad_right,
uint src2_pad_left)
{
Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
// Calculate input indices
const int x = get_global_id(0) * (int)VEC_SIZE;
const int y = get_global_id(1);
const int z = get_global_id(2) % (int)DEPTH;
const int w = get_global_id(2) / (int)DEPTH;
const int x1 = min(x, (int)INPUT1_WIDTH + (int)src1_pad_right - (int)VEC_SIZE);
const int x2 = max(x - (int)INPUT1_WIDTH, -(int)src2_pad_left);
// Calculate inputs and output addresses
const __global uchar *in1_ptr = src1_ptr + (int)src1_offset_first_element_in_bytes + x1 * (int)src1_stride_x + y * (int)src1_stride_y + z * (int)src1_stride_z + w * (int)src1_stride_w;
const __global uchar *in2_ptr = src2_ptr + (int)src2_offset_first_element_in_bytes + x2 * (int)src2_stride_x + y * (int)src2_stride_y + z * (int)src2_stride_z + w * (int)src2_stride_w;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
src1_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in1_ptr);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
src2_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in2_ptr);
#if defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT)
src1_values = requantize(src1_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
src2_values = requantize(src2_values, OFFSET_IN2, OFFSET_OUT, SCALE_IN2, SCALE_OUT);
#endif /* defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) */
const VEC_DATA_TYPE(int, VEC_SIZE) x_coords = SEQ + (VEC_DATA_TYPE(int, VEC_SIZE))(x);
const VEC_DATA_TYPE(COND_DATA_TYPE, VEC_SIZE) cond = CONVERT(x_coords < (VEC_DATA_TYPE(int, VEC_SIZE))(INPUT1_WIDTH), VEC_DATA_TYPE(COND_DATA_TYPE, VEC_SIZE));
const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) values = select(src2_values, src1_values, cond);
VSTORE(VEC_SIZE)
(values, 0, (__global DATA_TYPE *)dst.ptr);
}
#if defined(INPUT2_WIDTH) && defined(INPUT3_WIDTH)
/** This kernel concatenates four input tensors into the output tensor along the first dimension
*
* @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
* @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
* @note The offset for the first spatial dimension has to be passed at compile time using -DWIDTH_OFFSET. i.e. -DWIDTH_OFFSET=128
* @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
* @note First input tensor width should be given as a preprocessor argument using -DINPUT1_WIDTH=width. e.g. -DINPUT1_WIDTH=8
* @note Second input tensor width should be given as a preprocessor argument using -DINPUT2_WIDTH=width. e.g. -DINPUT2_WIDTH=8
* @note Third input tensor width should be given as a preprocessor argument using -DINPUT3_WIDTH=width. e.g. -DINPUT3_WIDTH=8
*
* @param[in] src1_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/F32
* @param[in] src1_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src1_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src1_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src1_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] src2_ptr Pointer to the source tensor. Supported data types: same as @p src1_ptr
* @param[in] src2_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src2_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src2_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src2_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src2_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src2_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src2_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src2_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src2_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] src3_ptr Pointer to the source tensor. Supported data types: same as @p src1_ptr
* @param[in] src3_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src3_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src3_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src3_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src3_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src3_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src3_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src3_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src3_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] src4_ptr Pointer to the source tensor. Supported data types: same as @p src1_ptr
* @param[in] src4_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] src4_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src4_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] src4_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src4_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src4_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src4_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src4_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src4_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 src1_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 source 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_stride_w Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_w output_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] src1_pad_right Right paddings of the first input tensor in unit of elements
* @param[in] src2_pad_left Left paddings of the second input tensor in unit of elements
* @param[in] src2_pad_right Right paddings of the second input tensor in unit of elements
* @param[in] src3_pad_left Left paddings of the third input tensor in unit of elements
* @param[in] src3_pad_right Right paddings of the third input tensor in unit of elements
* @param[in] src4_pad_left Left paddings of the fourth input tensor in unit of elements
*/
__kernel void concatenate_width_x4(
TENSOR4D_DECLARATION(src1),
TENSOR4D_DECLARATION(src2),
TENSOR4D_DECLARATION(src3),
TENSOR4D_DECLARATION(src4),
TENSOR4D_DECLARATION(dst),
uint src1_pad_right,
uint src2_pad_left,
uint src2_pad_right,
uint src3_pad_left,
uint src3_pad_right,
uint src4_pad_left)
{
Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
// Calculate input indices
const int x = get_global_id(0) * (int)VEC_SIZE;
const int y = get_global_id(1);
const int z = get_global_id(2) % (int)DEPTH;
const int w = get_global_id(2) / (int)DEPTH;
const int x1 = min(x, (int)INPUT1_WIDTH + (int)src1_pad_right - (int)VEC_SIZE);
const int x2 = min(max(x - (int)INPUT1_WIDTH, -(int)src2_pad_left), (int)INPUT2_WIDTH + (int)src2_pad_right - (int)VEC_SIZE);
const int x3 = min(max(x - (int)INPUT1_WIDTH - (int)INPUT2_WIDTH, -(int)src3_pad_left), (int)INPUT3_WIDTH + (int)src3_pad_right - (int)VEC_SIZE);
const int x4 = max(x - (int)INPUT1_WIDTH - (int)INPUT2_WIDTH - (int)INPUT3_WIDTH, -(int)src4_pad_left);
// Calculate inputs and output addresses
const __global uchar *in1_ptr = src1_ptr + (int)src1_offset_first_element_in_bytes + x1 * (int)src1_stride_x + y * (int)src1_stride_y + z * (int)src1_stride_z + w * (int)src1_stride_w;
const __global uchar *in2_ptr = src2_ptr + (int)src2_offset_first_element_in_bytes + x2 * (int)src2_stride_x + y * (int)src2_stride_y + z * (int)src2_stride_z + w * (int)src2_stride_w;
const __global uchar *in3_ptr = src3_ptr + (int)src3_offset_first_element_in_bytes + x3 * (int)src3_stride_x + y * (int)src3_stride_y + z * (int)src3_stride_z + w * (int)src3_stride_w;
const __global uchar *in4_ptr = src4_ptr + (int)src4_offset_first_element_in_bytes + x4 * (int)src4_stride_x + y * (int)src4_stride_y + z * (int)src4_stride_z + w * (int)src4_stride_w;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
src1_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in1_ptr);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
src2_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in2_ptr);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
src3_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in3_ptr);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
src4_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in4_ptr);
#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) && defined(OFFSET_IN2) && defined(SCALE_IN2) && defined(OFFSET_IN3) && defined(SCALE_IN3) && defined(OFFSET_IN4) && defined(SCALE_IN4)
src1_values = requantize(src1_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
src2_values = requantize(src2_values, OFFSET_IN2, OFFSET_OUT, SCALE_IN2, SCALE_OUT);
src3_values = requantize(src3_values, OFFSET_IN3, OFFSET_OUT, SCALE_IN3, SCALE_OUT);
src4_values = requantize(src4_values, OFFSET_IN4, OFFSET_OUT, SCALE_IN4, SCALE_OUT);
#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) && defined(OFFSET_IN2) && defined(SCALE_IN2) && defined(OFFSET_IN3) && defined(SCALE_IN3) && defined(OFFSET_IN4) && defined(SCALE_IN4) */
const VEC_DATA_TYPE(int, VEC_SIZE) x_coords = SEQ + (VEC_DATA_TYPE(int, VEC_SIZE))(x);
const VEC_DATA_TYPE(COND_DATA_TYPE, VEC_SIZE) cond_in2 = CONVERT(x_coords < (VEC_DATA_TYPE(int, VEC_SIZE))(INPUT1_WIDTH), VEC_DATA_TYPE(COND_DATA_TYPE, VEC_SIZE));
const VEC_DATA_TYPE(COND_DATA_TYPE, VEC_SIZE) cond_in3 = CONVERT(x_coords < (VEC_DATA_TYPE(int, VEC_SIZE))(INPUT1_WIDTH + INPUT2_WIDTH), VEC_DATA_TYPE(COND_DATA_TYPE, VEC_SIZE));
const VEC_DATA_TYPE(COND_DATA_TYPE, VEC_SIZE) cond_in4 = CONVERT(x_coords < (VEC_DATA_TYPE(int, VEC_SIZE))(INPUT1_WIDTH + INPUT2_WIDTH + INPUT3_WIDTH), VEC_DATA_TYPE(COND_DATA_TYPE, VEC_SIZE));
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
values = select(src2_values, src1_values, cond_in2);
values = select(src3_values, values, cond_in3);
values = select(src4_values, values, cond_in4);
VSTORE(VEC_SIZE)
(values, 0, (__global DATA_TYPE *)dst.ptr);
}
#endif /* defined(INPUT2_WIDTH) && defined(INPUT3_WIDTH) */
#endif /* defined(INPUT1_WIDTH) */
#endif /* defined(DEPTH) && defined(ELEMENT_SIZE) */
#if defined(WIDTH_OFFSET) && defined(DEPTH)
/** This kernel concatenates the input tensor into the output tensor along the first dimension
*
* @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
* @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
* @note The offset for the first spatial dimension has to be passed at compile time using -DWIDTH_OFFSET. i.e. -DWIDTH_OFFSET=128
* @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/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_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src_step_w 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 source 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_stride_w Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_w output_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
*/
__kernel void concatenate_width(
TENSOR4D_DECLARATION(src),
TENSOR4D_DECLARATION(dst))
{
Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, DEPTH);
Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
const VEC_QUANT out = requantize(source_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
VSTORE(VEC_SIZE)
(out, 0, (__global DATA_TYPE *)(dst.ptr) + WIDTH_OFFSET);
#else /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
VSTORE(VEC_SIZE)
(source_values, 0, (__global DATA_TYPE *)(dst.ptr) + WIDTH_OFFSET);
#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
}
#endif /* defined(WIDTH_OFFSET) && defined(DEPTH) */
#if defined(HEIGHT_OFFSET) && defined(DEPTH) && defined(VEC_SIZE)
/** This kernel concatenates the input tensor into the output tensor along the second dimension
*
* @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
* @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
* @note Vector sizes supported are 2,4,8 and 16.
* @note The offset for the second spatial dimension has to be passed at compile time using -DHEIGHT_OFFSET. i.e. -DHEIGHT_OFFSET=128
* @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/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_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src_step_w 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 source 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_stride_w Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_w output_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
*/
__kernel void concatenate_height(
TENSOR4D_DECLARATION(src),
TENSOR4D_DECLARATION(dst))
{
Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, DEPTH);
Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
const VEC_QUANT out = requantize(source_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
VSTORE(VEC_SIZE)
(out, 0, (__global DATA_TYPE *)(dst.ptr + HEIGHT_OFFSET * dst_stride_y));
#else /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
VSTORE(VEC_SIZE)
(source_values, 0, (__global DATA_TYPE *)(dst.ptr + HEIGHT_OFFSET * dst_stride_y));
#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
}
#endif /* defined(HEIGHT_OFFSET) && defined(DEPTH) */
/** This kernel concatenates the input tensor into the output tensor along the third dimension
*
* @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
* @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
*
* @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 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 source 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] offsets The offsets to the first valid element of the output tensor in bytes
*/
__kernel void concatenate(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
int offset)
{
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
source_values = requantize(source_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
VSTORE(VEC_SIZE)
(source_values, 0, (__global DATA_TYPE *)(dst.ptr + offset));
}
#endif /* defined(DATA_TYPE) && defined(VEC_SIZE) */
|
