bugfix: CLDeconvolutionLayer::validate fails if bias==NULL (#439)
[platform/upstream/armcl.git] / src / runtime / CL / functions / CLDeconvolutionLayer.cpp
1 /*
2  * Copyright (c) 2017-2018 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/CLDeconvolutionLayer.h"
25
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/Utils.h"
28 #include "arm_compute/core/Validate.h"
29 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
30
31 #include <memory>
32 #include <tuple>
33
34 using namespace arm_compute;
35 using namespace arm_compute::misc::shape_calculator;
36
37 CLDeconvolutionLayer::CLDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
38     : _memory_group(std::move(memory_manager)),
39       _scale_f(),
40       _conv_f(),
41       _scaled_output()
42 {
43 }
44
45 Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &info,
46                                       unsigned int inner_border_right, unsigned int inner_border_top)
47 {
48     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
49     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
50     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != weights->dimension(1));
51     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) < 1);
52     ARM_COMPUTE_RETURN_ERROR_ON(!info.padding_is_symmetric());
53
54     const unsigned int stride_x = info.stride().first;
55     const unsigned int stride_y = info.stride().second;
56
57     ARM_COMPUTE_RETURN_ERROR_ON_MSG(inner_border_right > stride_x - 1, "inner_border_right must be smaller than stride_x");
58     ARM_COMPUTE_RETURN_ERROR_ON_MSG(inner_border_top > stride_y - 1, "inner_border_top must be smaller than stride_y");
59
60     auto out_dims = deconvolution_output_dimensions(input->dimension(0), input->dimension(1), weights->dimension(0), weights->dimension(1),
61                                                     info.pad().first, info.pad().second, inner_border_right, inner_border_top, stride_x, stride_y);
62
63     const TensorShape output_shape = deconvolution_output_shape(out_dims, input->tensor_shape(), weights->tensor_shape());
64
65     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
66     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output, weights);
67
68     if(bias != nullptr)
69     {
70         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
71         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, bias);
72     }
73
74     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
75     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
76     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
77
78     TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_deconvolution_shape(*input, stride_x, stride_y, inner_border_right, inner_border_top,
79                                                                                                       info)));
80     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
81
82     ARM_COMPUTE_RETURN_ON_ERROR(CLDeconvolutionLayerUpsample::validate(input, &scale_out_info, BorderSize(inner_border_right, inner_border_top), info));
83     ARM_COMPUTE_RETURN_ON_ERROR(CLConvolutionLayer::validate(&scale_out_info, weights, bias, output, info, WeightsInfo()));
84
85     return Status{};
86 }
87
88 void CLDeconvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
89                                      unsigned int inner_border_right, unsigned int inner_border_top)
90 {
91     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
92
93     const unsigned int stride_x = info.stride().first;
94     const unsigned int stride_y = info.stride().second;
95
96     auto out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
97                                                     info.pad().first, info.pad().second, inner_border_top, inner_border_right, stride_x, stride_y);
98
99     const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
100
101     // Output auto initialization if not yet initialized
102     auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
103
104     // Perform validation step
105     ARM_COMPUTE_ERROR_THROW_ON(CLDeconvolutionLayer::validate(input->info(), weights->info(), bias == nullptr ? nullptr : bias->info(), output->info(), info, inner_border_right, inner_border_top));
106
107     _memory_group.manage(&_scaled_output);
108
109     // configure scale function
110     // Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor
111     TensorShape        scale_out_shape(input->info()->tensor_shape());
112     const unsigned int out_x = input->info()->dimension(0) + (input->info()->dimension(0) - 1) * (stride_x - 1) + inner_border_right + 2 * info.pad().first;
113     const unsigned int out_y = input->info()->dimension(1) + (input->info()->dimension(1) - 1) * (stride_y - 1) + inner_border_top + 2 * info.pad().second;
114     scale_out_shape.set(0, out_x);
115     scale_out_shape.set(1, out_y);
116     TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
117     _scaled_output.allocator()->init(scale_out_info);
118
119     _scale_f.configure(input, &_scaled_output, BorderSize(inner_border_top, inner_border_right), info);
120
121     // setup the function to convolve the upscaled output
122     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
123     _conv_f.configure(&_scaled_output, weights, bias, output, conv_info);
124     _scaled_output.allocator()->allocate();
125 }
126
127 void CLDeconvolutionLayer::run()
128 {
129     _memory_group.acquire();
130     _scale_f.run();
131     _conv_f.run();
132     _memory_group.release();
133 }