Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 110b920510a4b4c0edbf9859070a506c438f67b9 / . / src / core / CL / cl_kernels / concatenate.cl
blob: 0e8805f9b626e0c1c7f41cbb7521f496c334e2b3 [file] [log] [blame]
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]1/*
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]2 * Copyright (c) 2017-2018 ARM Limited.
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24#include "helpers.h"
25
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]26#if defined(DATA_TYPE) && defined(VEC_SIZE)
27
Michele Di Giorgio27400b92018-11-01 13:44:05 +0000[diff] [blame]28#if defined(DEPTH) && defined(ELEMENT_SIZE)
29
30#if defined(INPUT1_WIDTH)
31
32#if ELEMENT_SIZE == 1
33#define COND_DATA_TYPE char
34#elif ELEMENT_SIZE == 2
35#define COND_DATA_TYPE short
36#elif ELEMENT_SIZE == 4
37#define COND_DATA_TYPE int
38#else // ELEMENT_SIZE
39#error "Element size not supported"
40#endif // ELEMENT_SIZE
41
42#if VEC_SIZE == 2
43#define SEQ ((int2)(0, 1))
44#elif VEC_SIZE == 4
45#define SEQ ((int4)(0, 1, 2, 3))
46#elif VEC_SIZE == 8
47#define SEQ ((int8)(0, 1, 2, 3, 4, 5, 6, 7))
48#elif VEC_SIZE == 16
49#define SEQ ((int16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15))
50#else // VEC_SIZE
51#error "Vector size not supported"
52#endif // VEC_SIZE
53/** This kernel concatenates two input tensors into the output tensor along the first dimension
54 *
55 * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
56 * @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
57 * @note The offset for the first spatial dimension has to be passed at compile time using -DWIDTH_OFFSET. i.e. -DWIDTH_OFFSET=128
58 * @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
59 * @note First input tensor width should be given as a preprocessor argument using -DINPUT1_WIDTH=width. e.g. -DINPUT1_WIDTH=8
60 *
61 * @param[in] src1_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/F32
62 * @param[in] src1_stride_x Stride of the source tensor in X dimension (in bytes)
63 * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
64 * @param[in] src1_stride_y Stride of the source tensor in Y dimension (in bytes)
65 * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
66 * @param[in] src1_stride_z Stride of the source tensor in Z dimension (in bytes)
67 * @param[in] src1_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
68 * @param[in] src1_stride_w Stride of the first source tensor in Z dimension (in bytes)
69 * @param[in] src1_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
70 * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source tensor
71 * @param[in] src2_ptr Pointer to the source tensor. Supported data types: same as @p src1_ptr
72 * @param[in] src2_stride_x Stride of the source tensor in X dimension (in bytes)
73 * @param[in] src2_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
74 * @param[in] src2_stride_y Stride of the source tensor in Y dimension (in bytes)
75 * @param[in] src2_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
76 * @param[in] src2_stride_z Stride of the source tensor in Z dimension (in bytes)
77 * @param[in] src2_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
78 * @param[in] src2_stride_w Stride of the first source tensor in Z dimension (in bytes)
79 * @param[in] src2_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
80 * @param[in] src2_offset_first_element_in_bytes The offset of the first element in the source tensor
81 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src1_ptr
82 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
83 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
84 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
85 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
86 * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes)
87 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
88 * @param[in] dst_stride_w Stride of the destination tensor in Z dimension (in bytes)
89 * @param[in] dst_step_w output_stride_z * number of elements along Z processed per workitem(in bytes)
90 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
91 */
92__kernel void concatenate_width_x2(
93 TENSOR4D_DECLARATION(src1),
94 TENSOR4D_DECLARATION(src2),
95 TENSOR4D_DECLARATION(dst))
96{
97 Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
98
99 // Calculate input indices
100 const int x = get_global_id(0) * (int)VEC_SIZE;
101 const int y = get_global_id(1);
102 const int z = get_global_id(2) % (int)DEPTH;
103 const int w = get_global_id(2) / (int)DEPTH;
104 const int x1 = min(x, (int)INPUT1_WIDTH);
105 const int x2 = max(x - (int)INPUT1_WIDTH, -(int)VEC_SIZE);
106
107 // Calculate inputs and output addresses
108 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;
109 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;
110
111 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) src1_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in1_ptr);
112 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) src2_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in2_ptr);
113
114 const VEC_DATA_TYPE(int, VEC_SIZE) x_coords = SEQ + (VEC_DATA_TYPE(int, VEC_SIZE))(x);
115 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));
116 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) values = select(src2_values, src1_values, cond);
117
118 VSTORE(VEC_SIZE)
119 (values, 0, (__global DATA_TYPE *)dst.ptr);
120}
121
122#if defined(INPUT2_WIDTH) && defined(INPUT3_WIDTH)
123/** This kernel concatenates four input tensors into the output tensor along the first dimension
124 *
125 * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
126 * @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
127 * @note The offset for the first spatial dimension has to be passed at compile time using -DWIDTH_OFFSET. i.e. -DWIDTH_OFFSET=128
128 * @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
129 * @note First input tensor width should be given as a preprocessor argument using -DINPUT1_WIDTH=width. e.g. -DINPUT1_WIDTH=8
130 * @note Second input tensor width should be given as a preprocessor argument using -DINPUT2_WIDTH=width. e.g. -DINPUT2_WIDTH=8
131 * @note Third input tensor width should be given as a preprocessor argument using -DINPUT3_WIDTH=width. e.g. -DINPUT3_WIDTH=8
132 *
133 * @param[in] src1_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/F32
134 * @param[in] src1_stride_x Stride of the source tensor in X dimension (in bytes)
135 * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
136 * @param[in] src1_stride_y Stride of the source tensor in Y dimension (in bytes)
137 * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
138 * @param[in] src1_stride_z Stride of the source tensor in Z dimension (in bytes)
139 * @param[in] src1_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
140 * @param[in] src1_stride_w Stride of the first source tensor in Z dimension (in bytes)
141 * @param[in] src1_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
142 * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source tensor
143 * @param[in] src2_ptr Pointer to the source tensor. Supported data types: same as @p src1_ptr
144 * @param[in] src2_stride_x Stride of the source tensor in X dimension (in bytes)
145 * @param[in] src2_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
146 * @param[in] src2_stride_y Stride of the source tensor in Y dimension (in bytes)
147 * @param[in] src2_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
148 * @param[in] src2_stride_z Stride of the source tensor in Z dimension (in bytes)
149 * @param[in] src2_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
150 * @param[in] src2_stride_w Stride of the first source tensor in Z dimension (in bytes)
151 * @param[in] src2_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
152 * @param[in] src2_offset_first_element_in_bytes The offset of the first element in the source tensor
153 * @param[in] src3_ptr Pointer to the source tensor. Supported data types: same as @p src1_ptr
154 * @param[in] src3_stride_x Stride of the source tensor in X dimension (in bytes)
155 * @param[in] src3_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
156 * @param[in] src3_stride_y Stride of the source tensor in Y dimension (in bytes)
157 * @param[in] src3_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
158 * @param[in] src3_stride_z Stride of the source tensor in Z dimension (in bytes)
159 * @param[in] src3_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
160 * @param[in] src3_stride_w Stride of the first source tensor in Z dimension (in bytes)
161 * @param[in] src3_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
162 * @param[in] src3_offset_first_element_in_bytes The offset of the first element in the source tensor
163 * @param[in] src4_ptr Pointer to the source tensor. Supported data types: same as @p src1_ptr
164 * @param[in] src4_stride_x Stride of the source tensor in X dimension (in bytes)
165 * @param[in] src4_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
166 * @param[in] src4_stride_y Stride of the source tensor in Y dimension (in bytes)
167 * @param[in] src4_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
168 * @param[in] src4_stride_z Stride of the source tensor in Z dimension (in bytes)
169 * @param[in] src4_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
170 * @param[in] src4_stride_w Stride of the first source tensor in Z dimension (in bytes)
171 * @param[in] src4_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
172 * @param[in] src4_offset_first_element_in_bytes The offset of the first element in the source tensor
173 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src1_ptr
174 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
175 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
176 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
177 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
178 * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes)
179 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
180 * @param[in] dst_stride_w Stride of the destination tensor in Z dimension (in bytes)
181 * @param[in] dst_step_w output_stride_z * number of elements along Z processed per workitem(in bytes)
182 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
183 */
184__kernel void concatenate_width_x4(
185 TENSOR4D_DECLARATION(src1),
186 TENSOR4D_DECLARATION(src2),
187 TENSOR4D_DECLARATION(src3),
188 TENSOR4D_DECLARATION(src4),
189 TENSOR4D_DECLARATION(dst))
190{
191 Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
192
193 // Calculate input indices
194 const int x = get_global_id(0) * (int)VEC_SIZE;
195 const int y = get_global_id(1);
196 const int z = get_global_id(2) % (int)DEPTH;
197 const int w = get_global_id(2) / (int)DEPTH;
198
199 const int x1 = min(x, (int)INPUT1_WIDTH);
200 const int x2 = min(max(x - (int)INPUT1_WIDTH, -(int)VEC_SIZE), (int)INPUT2_WIDTH);
201 const int x3 = min(max(x - (int)INPUT1_WIDTH - (int)INPUT2_WIDTH, -(int)VEC_SIZE), (int)INPUT3_WIDTH);
202 const int x4 = max(x - (int)INPUT1_WIDTH - (int)INPUT2_WIDTH - (int)INPUT3_WIDTH, -(int)VEC_SIZE);
203
204 // Calculate inputs and output addresses
205 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;
206 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;
207 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;
208 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;
209
210 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) src1_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in1_ptr);
211 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) src2_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in2_ptr);
212 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) src3_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in3_ptr);
213 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) src4_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in4_ptr);
214
215 const VEC_DATA_TYPE(int, VEC_SIZE) x_coords = SEQ + (VEC_DATA_TYPE(int, VEC_SIZE))(x);
216
217 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));
218 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));
219 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));
220
221 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
222 values = select(src2_values, src1_values, cond_in2);
223 values = select(src3_values, values, cond_in3);
224 values = select(src4_values, values, cond_in4);
225
226 VSTORE(VEC_SIZE)
227 (values, 0, (__global DATA_TYPE *)dst.ptr);
228}
229#endif /* defined(INPUT2_WIDTH) && defined(INPUT3_WIDTH) */
230#endif /* defined(INPUT1_WIDTH) */
231#endif /* defined(DEPTH) && defined(ELEMENT_SIZE) */
232
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]233#if defined(WIDTH_OFFSET) && defined(DEPTH)
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]234/** This kernel concatenates the input tensor into the output tensor along the first dimension
235 *
Gian Marco Iodice1d1f32c2018-08-10 09:34:11 +0100[diff] [blame]236 * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]237 * @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
Gian Marco Iodice1d1f32c2018-08-10 09:34:11 +0100[diff] [blame]238 * @note The offset for the first spatial dimension has to be passed at compile time using -DWIDTH_OFFSET. i.e. -DWIDTH_OFFSET=128
Michele Di Giorgio27400b92018-11-01 13:44:05 +0000[diff] [blame]239 * @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
Gian Marco Iodice1d1f32c2018-08-10 09:34:11 +0100[diff] [blame]240 *
241 * @param[in] src_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/F32
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]242 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
243 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
244 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
245 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
246 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
247 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]248 * @param[in] src_stride_w Stride of the first source tensor in Z dimension (in bytes)
249 * @param[in] src_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]250 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
251 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
252 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
253 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
254 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
255 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
256 * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes)
257 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]258 * @param[in] dst_stride_w Stride of the destination tensor in Z dimension (in bytes)
259 * @param[in] dst_step_w output_stride_z * number of elements along Z processed per workitem(in bytes)
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]260 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]261 */
262__kernel void concatenate_width(
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]263 TENSOR4D_DECLARATION(src),
264 TENSOR4D_DECLARATION(dst))
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]265{
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]266 Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, DEPTH);
267 Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]268
269 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
270 source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
271
272 VSTORE(VEC_SIZE)
Gian Marco Iodice1d1f32c2018-08-10 09:34:11 +0100[diff] [blame]273 (source_values, 0, (__global DATA_TYPE *)(dst.ptr) + WIDTH_OFFSET);
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]274}
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]275#endif /* defined(WIDTH_OFFSET) && defined(DEPTH) */
Michalis Spyrou55b3d122018-05-09 09:59:23 +0100[diff] [blame]276
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]277/** This kernel concatenates the input tensor into the output tensor along the third dimension
278 *
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]279 * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
280 * @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
281 *
Vidhya Sudhan Loganathan7485d5a2018-07-04 09:34:00 +0100[diff] [blame]282 * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16, F32
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]283 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
284 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
285 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
286 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
Georgios Pinitasac4e8732017-07-05 17:02:25 +0100[diff] [blame]287 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
288 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]289 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
Georgios Pinitasac4e8732017-07-05 17:02:25 +0100[diff] [blame]290 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]291 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
292 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
293 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
294 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
Georgios Pinitasac4e8732017-07-05 17:02:25 +0100[diff] [blame]295 * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes)
296 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]297 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
Georgios Pinitasac4e8732017-07-05 17:02:25 +0100[diff] [blame]298 * @param[in] offsets The offsets to the first valid element of the output tensor in bytes
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]299 */
300__kernel void concatenate_depth(
Georgios Pinitasac4e8732017-07-05 17:02:25 +0100[diff] [blame]301 TENSOR3D_DECLARATION(src),
302 TENSOR3D_DECLARATION(dst),
303 int3 offsets)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]304{
Georgios Pinitasac4e8732017-07-05 17:02:25 +0100[diff] [blame]305 Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
306 Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]307
Georgios Pinitasac4e8732017-07-05 17:02:25 +0100[diff] [blame]308 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
309 source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, -offsets.x, -offsets.y, 0));
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]310
Georgios Pinitasac4e8732017-07-05 17:02:25 +0100[diff] [blame]311 VSTORE(VEC_SIZE)
312 (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offsets.z));
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]313}
Michele Di Giorgioe6dbde02018-10-19 15:46:19 +0100[diff] [blame]314#endif /* defined(DATA_TYPE) && defined(VEC_SIZE) */