blob: 7559e7ae72798af6c9e3afde8d131808a3838bfb [file] [log] [blame]
giuros013175fcf2018-11-21 09:59:17 +00001/*
2 * Copyright (c) 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/core/CL/kernels/CLTileKernel.h"
25
26#include "arm_compute/core/CL/CLHelpers.h"
27#include "arm_compute/core/CL/CLKernelLibrary.h"
28#include "arm_compute/core/CL/CLValidate.h"
29#include "arm_compute/core/CL/ICLTensor.h"
30#include "arm_compute/core/Helpers.h"
31#include "arm_compute/core/IAccessWindow.h"
32#include "arm_compute/core/TensorInfo.h"
33#include "arm_compute/core/Utils.h"
34#include "arm_compute/core/Validate.h"
35#include "arm_compute/core/Window.h"
36#include "arm_compute/core/utils/misc/ShapeCalculator.h"
37
38namespace arm_compute
39{
40namespace
41{
42Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Multiples &multiples)
43{
44 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
45 ARM_COMPUTE_RETURN_ERROR_ON(multiples.size() > 4);
46 ARM_COMPUTE_RETURN_ERROR_ON(multiples.empty());
47 ARM_COMPUTE_RETURN_ERROR_ON(std::any_of(multiples.begin(), multiples.end(), [](uint32_t e)
48 {
49 return e == 0;
50 }));
51
52 // Validate output if initialized
53 if(output->total_size() != 0)
54 {
55 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_tiled_shape(input->tensor_shape(), multiples), output->tensor_shape());
56 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
57 }
58
59 return Status{};
60}
61} // namespace
62
63CLTileKernel::CLTileKernel()
64 : _input(nullptr), _output(nullptr)
65{
66}
67
68void CLTileKernel::configure(const ICLTensor *input, ICLTensor *output, const Multiples &multiples)
69{
70 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
71
72 // Auto initialize output
73 TensorShape tiled_shape = misc::shape_calculator::compute_tiled_shape(input->info()->tensor_shape(), multiples);
74 auto_init_if_empty(*output->info(), tiled_shape, 1, input->info()->data_type());
75
76 // Validate
77 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), multiples));
78
79 _input = input;
80 _output = output;
81
82 const DataType data_type = input->info()->data_type();
83 const int vec_size_x = 16 / input->info()->element_size();
84 const int input_width_x = input->info()->tensor_shape().x();
85 const unsigned int offset = ceil_to_multiple(input_width_x, vec_size_x) - input_width_x;
86 const bool multi_access_x = (input_width_x / vec_size_x > 0);
87
88 // Create kernel
89 CLBuildOptions build_opts;
90 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
91 build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input_width_x));
92 build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
93 build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
94 build_opts.add_option("-DSRC_BATCHES=" + support::cpp11::to_string(input->info()->dimension(3)));
95 build_opts.add_option("-DDST_DEPTH=" + support::cpp11::to_string(output->info()->dimension(2)));
96 build_opts.add_option_if(multi_access_x, "-DOFFSET=" + support::cpp11::to_string(offset));
97 build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
98 _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("tile", build_opts.options()));
99
100 // Configure window without padding
101 Window win = calculate_max_window(*output->info());
102
103 if(multi_access_x)
104 {
105 // If multi-access is enabled, no thread should cross the tile boundaries. This means we need
106 // as many threads as those to cover a single tile times multiples[0]. Note that if threads
107 // do not cross the boundaries of the tiles, they won't cross the boundaries of the last tile, and
108 // we don't need to pad the output
109 const unsigned int size_win_x = ceil_to_multiple(input->info()->dimension(0), vec_size_x) * multiples[0];
110 win.set(Window::DimX,
111 Window::Dimension(win.x().start(), size_win_x, vec_size_x));
112 }
113
114 ICLKernel::configure_internal(win);
115
116 // Set config_id for enabling LWS tuning
117 _config_id = "tile";
118 _config_id += "_";
119 _config_id += lower_string(string_from_data_type(input->info()->data_type()));
120 for(unsigned int i = 0; i < multiples.size(); ++i)
121 {
122 _config_id += "_";
123 _config_id += support::cpp11::to_string(input->info()->dimension(i));
124 _config_id += "_";
125 _config_id += support::cpp11::to_string(multiples[i]);
126 }
127}
128
129Status CLTileKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Multiples &multiples)
130{
131 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, multiples));
132 return Status{};
133}
134
135void CLTileKernel::run(const Window &window, cl::CommandQueue &queue)
136{
137 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
138 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
139
140 Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
141 Window slice = collapsed.first_slice_window_4D();
142
143 do
144 {
145 unsigned int idx = 0;
146 add_4D_tensor_argument(idx, _input, slice);
147 add_4D_tensor_argument(idx, _output, slice);
148 enqueue(queue, *this, slice);
149 }
150 while(collapsed.slide_window_slice_4D(slice));
151}
152} // namespace arm_compute