1 /*
2 * Copyright (c) 2017-2021 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 "arm_compute/runtime/CL/functions/CLReductionOperation.h"
25
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/PixelValue.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Validate.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "arm_compute/runtime/CL/CLScheduler.h"
33 #include "src/core/CL/kernels/CLReductionOperationKernel.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/runtime/Utils.h"
36
37 #include "src/common/utils/Log.h"
38
39 namespace arm_compute
40 {
CLReductionOperation(std::shared_ptr<IMemoryManager> memory_manager)41 CLReductionOperation::CLReductionOperation(std::shared_ptr<IMemoryManager> memory_manager)
42 : _memory_group(std::move(memory_manager)), _unreshaped_output(), _reduction_kernel(), _reshape(), _reduction_axis(), _is_reshape_required(false)
43 {
44 }
45
46 CLReductionOperation::~CLReductionOperation() = default;
47
validate(const ITensorInfo * input,const ITensorInfo * output,unsigned int axis,ReductionOperation op,bool keep_dims)48 Status CLReductionOperation::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op, bool keep_dims)
49 {
50 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
51 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
52 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
53
54 const bool is_reshape_required = !keep_dims;
55
56 if(is_reshape_required && output->total_size() != 0)
57 {
58 const TensorInfo expected_output_shape = output->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis, keep_dims));
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output_shape, output);
60 }
61
62 auto *output_internal = output;
63
64 TensorInfo output_before_reshape;
65 const auto input_shape = input->tensor_shape();
66 const auto input_num_channles = input->num_channels();
67 const auto input_qinfo = input->quantization_info();
68 const auto output_data_type = output->data_type();
69
70 auto initialize_tensorinfo = [](TensorInfo & ti, TensorShape shape, DataType data_type, int num_channels, QuantizationInfo qinfo)
71 {
72 ti.set_data_type(data_type).set_tensor_shape(shape).set_num_channels(num_channels).set_quantization_info(qinfo);
73 };
74
75 if(is_reshape_required)
76 {
77 auto shape_before_reshape = input_shape;
78 shape_before_reshape.set(axis, 1);
79 initialize_tensorinfo(output_before_reshape, shape_before_reshape, output_data_type, input_num_channles, input_qinfo);
80 output_internal = &output_before_reshape;
81 }
82
83 ARM_COMPUTE_RETURN_ON_ERROR(CLReductionOperationKernel::validate(input, output_internal, axis, op));
84
85 if(is_reshape_required)
86 {
87 ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(output_internal, output));
88 }
89
90 return Status{};
91 }
92
configure_intermediate_result_vector(ICLTensor * input,ICLTensor * output)93 ICLTensor *CLReductionOperation::configure_intermediate_result_vector(ICLTensor *input, ICLTensor *output)
94 {
95 if(!_is_reshape_required)
96 {
97 return output;
98 }
99
100 auto shape = input->info()->tensor_shape();
101 shape.set(_reduction_axis, 1);
102 _unreshaped_output.allocator()->init(input->info()->clone()->set_tensor_shape(shape));
103 return &_unreshaped_output;
104 }
105
configure(ICLTensor * input,ICLTensor * output,unsigned int axis,ReductionOperation op,bool keep_dims)106 void CLReductionOperation::configure(ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, bool keep_dims)
107 {
108 configure(CLKernelLibrary::get().get_compile_context(), input, output, axis, op, keep_dims);
109 }
110
configure(const CLCompileContext & compile_context,ICLTensor * input,ICLTensor * output,unsigned int axis,ReductionOperation op,bool keep_dims)111 void CLReductionOperation::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, bool keep_dims)
112 {
113 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
114 ARM_COMPUTE_LOG_PARAMS(input, output, axis, op, keep_dims);
115 _reduction_axis = axis;
116 _is_reshape_required = !keep_dims;
117
118 auto *output_internal = configure_intermediate_result_vector(input, output);
119
120 if(_is_reshape_required)
121 {
122 const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, false);
123 const auto output_data_type = input->info()->data_type();
124 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true));
125
126 _memory_group.manage(&_unreshaped_output);
127 }
128
129 _reduction_kernel = std::make_unique<CLReductionOperationKernel>();
130 _reduction_kernel->configure(compile_context, input, output_internal, axis, op);
131
132 if(_is_reshape_required)
133 {
134 _reshape.configure(compile_context, &_unreshaped_output, output);
135 _unreshaped_output.allocator()->allocate();
136 }
137 }
138
run()139 void CLReductionOperation::run()
140 {
141 MemoryGroupResourceScope scope_mg(_memory_group);
142
143 CLScheduler::get().enqueue(*_reduction_kernel, false);
144
145 if(_is_reshape_required)
146 {
147 _reshape.run();
148 }
149 }
150 } // namespace arm_compute
151