1 /*
2 * Copyright (c) 2021 Arm Limited.
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 "src/gpu/cl/operators/ClConcatenate.h"
25
26 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
27 #include "arm_compute/runtime/CL/CLScheduler.h"
28
29 #include "src/gpu/cl/kernels/ClBatchConcatenateKernel.h"
30 #include "src/gpu/cl/kernels/ClDepthConcatenateKernel.h"
31 #include "src/gpu/cl/kernels/ClHeightConcatenateKernel.h"
32 #include "src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h"
33 #include "src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h"
34 #include "src/gpu/cl/kernels/ClWidthConcatenateKernel.h"
35
36 #include "arm_compute/core/Error.h"
37 #include "arm_compute/core/TensorInfo.h"
38 #include "arm_compute/core/Types.h"
39
40 #include "src/common/utils/Log.h"
41 #include "src/core/helpers/AutoConfiguration.h"
42
43 namespace arm_compute
44 {
45 namespace opencl
46 {
configure(const CLCompileContext & compile_context,const std::vector<ITensorInfo * > & src_vector,ITensorInfo * dst,size_t axis)47 void ClConcatenate::configure(const CLCompileContext &compile_context, const std::vector<ITensorInfo *> &src_vector, ITensorInfo *dst, size_t axis)
48 {
49 ARM_COMPUTE_ERROR_ON(dst == nullptr);
50 ARM_COMPUTE_LOG_PARAMS(src_vector, dst, axis);
51 _axis = axis;
52 _num_inputs = src_vector.size();
53
54 TensorShape dst_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(src_vector, _axis);
55 std::vector<const ITensorInfo *> const_src_vector(src_vector.size());
56 std::transform(src_vector.begin(), src_vector.end(), const_src_vector.begin(), [](ITensorInfo * t)
57 {
58 ARM_COMPUTE_ERROR_ON_NULLPTR(t);
59 return t;
60 });
61
62 // dst auto inizialitation if not yet initialized
63 auto_init_if_empty(*dst, dst_shape, 1, src_vector[0]->data_type());
64 ARM_COMPUTE_ERROR_THROW_ON(ClConcatenate::validate(const_src_vector, dst, axis));
65
66 unsigned int offset = 0;
67 switch(_axis)
68 {
69 case Window::DimX:
70 {
71 switch(_num_inputs)
72 {
73 case 2:
74 {
75 // Configure WidthConcatenate2Tensors kernel
76 auto kernel = std::make_unique<kernels::ClWidthConcatenate2TensorsKernel>();
77 kernel->configure(compile_context, src_vector.at(0), src_vector.at(1), dst);
78 _concat_kernels.emplace_back(std::move(kernel));
79 break;
80 }
81 case 4:
82 {
83 // Configure WidthConcatenate4Tensors kernel
84 auto kernel = std::make_unique<kernels::ClWidthConcatenate4TensorsKernel>();
85 kernel->configure(compile_context, src_vector.at(0), src_vector.at(1), src_vector.at(2), src_vector.at(3), dst);
86 _concat_kernels.emplace_back(std::move(kernel));
87 break;
88 }
89 default:
90 {
91 // Configure generic case WidthConcatenate kernels
92 for(unsigned int i = 0; i < _num_inputs; ++i)
93 {
94 auto kernel = std::make_unique<kernels::ClWidthConcatenateKernel>();
95 kernel->configure(compile_context, src_vector.at(i), offset, dst);
96 offset += src_vector.at(i)->dimension(_axis);
97 _concat_kernels.emplace_back(std::move(kernel));
98 }
99 break;
100 }
101 }
102 break;
103 }
104 case Window::DimY:
105 {
106 for(unsigned int i = 0; i < _num_inputs; ++i)
107 {
108 auto kernel = std::make_unique<kernels::ClHeightConcatenateKernel>();
109 kernel->configure(compile_context, src_vector.at(i), offset, dst);
110 offset += src_vector.at(i)->dimension(_axis);
111 _concat_kernels.emplace_back(std::move(kernel));
112 }
113 break;
114 }
115 case Window::DimZ:
116 {
117 for(unsigned int i = 0; i < _num_inputs; ++i)
118 {
119 auto kernel = std::make_unique<kernels::ClDepthConcatenateKernel>();
120 kernel->configure(compile_context, src_vector.at(i), offset, dst);
121 offset += src_vector.at(i)->dimension(_axis);
122 _concat_kernels.emplace_back(std::move(kernel));
123 }
124 break;
125 }
126 case 3:
127 {
128 for(unsigned int i = 0; i < _num_inputs; ++i)
129 {
130 auto kernel = std::make_unique<kernels::ClBatchConcatenateKernel>();
131 kernel->configure(compile_context, src_vector.at(i), offset, dst);
132 offset += src_vector.at(i)->dimension(_axis);
133 _concat_kernels.emplace_back(std::move(kernel));
134 }
135 break;
136 }
137 default:
138 ARM_COMPUTE_ERROR("Axis not supported");
139 }
140 }
141
validate(const std::vector<const ITensorInfo * > & src_vector,const ITensorInfo * dst,size_t axis)142 Status ClConcatenate::validate(const std::vector<const ITensorInfo *> &src_vector, const ITensorInfo *dst, size_t axis)
143 {
144 ARM_COMPUTE_RETURN_ERROR_ON(dst == nullptr);
145 const unsigned int num_inputs = src_vector.size();
146
147 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst);
148 ARM_COMPUTE_RETURN_ERROR_ON(num_inputs < 2);
149
150 unsigned int offset = 0;
151 switch(axis)
152 {
153 case Window::DimX:
154 {
155 switch(num_inputs)
156 {
157 case 2:
158 // Validate WidthConcatenate2Tensors kernels if there are 2 inputs
159 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src_vector[0], src_vector[1]);
160 ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWidthConcatenate2TensorsKernel::validate(src_vector[0], src_vector[1], dst));
161 break;
162 case 4:
163 // Validate WidthConcatenate4Tensors kernels if there are 4 inputs
164 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src_vector[0], src_vector[1], src_vector[2], src_vector[3]);
165 ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWidthConcatenate4TensorsKernel::validate(src_vector[0], src_vector[1], src_vector[2], src_vector[3], dst));
166 break;
167 default:
168 // Validate generic case of WidthConcatenate kernel
169 for(const auto &src : src_vector)
170 {
171 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
172 ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWidthConcatenateKernel::validate(src, offset, dst));
173 offset += src->dimension(axis);
174 }
175 break;
176 }
177 break;
178 }
179 case Window::DimY:
180 {
181 for(const auto &src : src_vector)
182 {
183 ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClHeightConcatenateKernel::validate(src, offset, dst));
184 offset += src->dimension(axis);
185 }
186 break;
187 }
188 case Window::DimZ:
189 {
190 for(const auto &src : src_vector)
191 {
192 ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClDepthConcatenateKernel::validate(src, offset, dst));
193 offset += src->dimension(axis);
194 }
195 break;
196 }
197 case 3:
198 {
199 for(const auto &src : src_vector)
200 {
201 ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClBatchConcatenateKernel::validate(src, offset, dst));
202 offset += src->dimension(axis);
203 }
204 break;
205 }
206 default:
207 ARM_COMPUTE_ERROR("Axis not supported");
208 }
209
210 if(dst->total_size() != 0)
211 {
212 TensorShape dst_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(src_vector, axis);
213 ARM_COMPUTE_RETURN_ERROR_ON(dst_shape.total_size() != dst->tensor_shape().total_size());
214 }
215
216 return Status{};
217 }
218
run(ITensorPack & tensors)219 void ClConcatenate::run(ITensorPack &tensors)
220 {
221 if(tensors.empty())
222 {
223 ARM_COMPUTE_ERROR("No inputs provided");
224 }
225
226 if(static_cast<int>(tensors.size()) - 1 != static_cast<int>(_num_inputs))
227 {
228 ARM_COMPUTE_ERROR("Configured with different number of inputs");
229 }
230
231 if(_axis == Window::DimX && (_num_inputs == 2 || _num_inputs == 4))
232 {
233 ARM_COMPUTE_ERROR_ON(_concat_kernels.empty());
234 CLScheduler::get().enqueue_op(*_concat_kernels.at(0), tensors, true);
235 }
236 else
237 {
238 int i = 0;
239 for(auto &k : _concat_kernels)
240 {
241 ITensorPack pack;
242 pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC_VEC + i));
243 pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_DST));
244 CLScheduler::get().enqueue_op(*k, pack, true);
245 ++i;
246 }
247 }
248 }
249 } // namespace opencl
250 } // namespace arm_compute
251