1 /*
2 * Copyright (c) 2017-2020 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/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/Helpers.h"
30 #include "arm_compute/core/Validate.h"
31 #include "src/core/CL/CLValidate.h"
32 #include "src/core/helpers/WindowHelpers.h"
33
34 namespace arm_compute
35 {
CLDeconvolutionLayerUpsampleKernel()36 CLDeconvolutionLayerUpsampleKernel::CLDeconvolutionLayerUpsampleKernel()
37 : _input(nullptr), _output(nullptr), _info(), _data_layout(DataLayout::UNKNOWN)
38 {
39 }
40
validate(const ITensorInfo * input,const ITensorInfo * output,const PadStrideInfo & info)41 Status CLDeconvolutionLayerUpsampleKernel::validate(const ITensorInfo *input, const ITensorInfo *output,
42 const PadStrideInfo &info)
43 {
44 ARM_COMPUTE_UNUSED(info);
45 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
46 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
47 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
48 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
49 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
50
51 const DataLayout data_layout = input->data_layout();
52
53 const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
54 const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
55 const size_t idx_c = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
56
57 ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(idx_w) == 0);
58 ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(idx_h) == 0);
59
60 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_c) != output->dimension(idx_c));
61 for(size_t i = 3; i < Coordinates::num_max_dimensions; ++i)
62 {
63 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i));
64 }
65
66 return Status{};
67 }
68
configure(const ICLTensor * input,ICLTensor * output,const PadStrideInfo & info)69 void CLDeconvolutionLayerUpsampleKernel::configure(const ICLTensor *input, ICLTensor *output,
70 const PadStrideInfo &info)
71 {
72 configure(CLKernelLibrary::get().get_compile_context(), input, output, info);
73 }
74
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const PadStrideInfo & info)75 void CLDeconvolutionLayerUpsampleKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output,
76 const PadStrideInfo &info)
77 {
78 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
79
80 // Perform validation step
81 ARM_COMPUTE_ERROR_THROW_ON(CLDeconvolutionLayerUpsampleKernel::validate(input->info(), output->info(), info));
82
83 _input = input;
84 _output = output;
85 _info = info;
86 _data_layout = input->info()->data_layout();
87
88 // Create kernel
89 CLBuildOptions build_opts;
90 build_opts.add_option(("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size())));
91 _kernel = create_kernel(compile_context, "deconvolution_upsample", build_opts.options());
92
93 constexpr unsigned int num_elems_processed_per_iteration = 1;
94
95 // Configure kernel window
96 Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
97 AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
98 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
99
100 ICLKernel::configure_internal(win);
101 }
102
run(const Window & window,cl::CommandQueue & queue)103 void CLDeconvolutionLayerUpsampleKernel::run(const Window &window, cl::CommandQueue &queue)
104 {
105 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
106 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
107
108 const size_t idx_w = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
109 const size_t idx_h = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
110
111 const int out_start_x = _info.pad_left();
112 const int out_end_x = _output->info()->dimension(idx_w) - _info.pad_right() + _info.stride().first - 1;
113 const int out_step_x = _info.stride().first;
114
115 const int out_start_y = _info.pad_top();
116 const int out_end_y = _output->info()->dimension(idx_h) - _info.pad_bottom() + _info.stride().second - 1;
117 const int out_step_y = _info.stride().second;
118
119 switch(_data_layout)
120 {
121 case DataLayout::NCHW:
122 {
123 Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
124
125 Window slice_out = collapsed.first_slice_window_3D();
126 slice_out.set(Window::DimX, Window::Dimension(out_start_x, out_end_x, out_step_x));
127 slice_out.set(Window::DimY, Window::Dimension(out_start_y, out_end_y, out_step_y));
128
129 Window slice_in = collapsed.first_slice_window_3D();
130
131 do
132 {
133 unsigned int idx = 0;
134 add_3D_tensor_argument(idx, _input, slice_in);
135 add_3D_tensor_argument(idx, _output, slice_out);
136 enqueue(queue, *this, slice_out, lws_hint());
137 }
138 while(collapsed.slide_window_slice_3D(slice_in) && collapsed.slide_window_slice_3D(slice_out));
139 break;
140 }
141 case DataLayout::NHWC:
142 {
143 // NOTE: not collapsing in NHWC
144 Window slice_out = window.first_slice_window_3D();
145 slice_out.set(Window::DimY, Window::Dimension(out_start_x, out_end_x, out_step_x));
146 slice_out.set(Window::DimZ, Window::Dimension(out_start_y, out_end_y, out_step_y));
147
148 Window slice_in = window.first_slice_window_3D();
149
150 do
151 {
152 unsigned int idx = 0;
153 add_3D_tensor_argument(idx, _input, slice_in);
154 add_3D_tensor_argument(idx, _output, slice_out);
155 enqueue(queue, *this, slice_out, lws_hint());
156 }
157 while(window.slide_window_slice_3D(slice_in) && window.slide_window_slice_3D(slice_out));
158 break;
159 }
160 default:
161 ARM_COMPUTE_ERROR("Unsupported data layout");
162 }
163 }
164 } // namespace arm_compute
165