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 "arm_compute/core/GLES_COMPUTE/kernels/GCFillBorderKernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
28 #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
29 #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
30 #include "arm_compute/core/GLES_COMPUTE/OpenGLES.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Utils.h"
33 #include "arm_compute/core/Validate.h"
34 #include "arm_compute/core/Window.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37 #include "support/StringSupport.h"
38
39 #include <cstdint>
40 #include <set>
41 #include <string>
42
43 using namespace arm_compute;
44
GCFillBorderKernel()45 GCFillBorderKernel::GCFillBorderKernel()
46 : IGCKernel(), _tensor(nullptr)
47 {
48 }
49
is_parallelisable() const50 bool GCFillBorderKernel::is_parallelisable() const
51 {
52 return false;
53 }
54
55 template <class T>
set_constant_border(unsigned int idx,const PixelValue & constant_border_value)56 void GCFillBorderKernel::set_constant_border(unsigned int idx, const PixelValue &constant_border_value)
57 {
58 T value;
59 constant_border_value.get(value);
60 _kernel.set_argument(idx, static_cast<T>(value));
61 }
62
configure(const IGCTensor * tensor,BorderSize border_size,BorderMode border_mode,const PixelValue & constant_border_value)63 void GCFillBorderKernel::configure(const IGCTensor *tensor, BorderSize border_size, BorderMode border_mode, const PixelValue &constant_border_value)
64 {
65 ARM_COMPUTE_ERROR_ON(tensor == nullptr);
66 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(tensor, 1, DataType::F32, DataType::F16);
67 ARM_COMPUTE_ERROR_ON(tensor->info()->num_channels() != 1);
68
69 border_size.limit(tensor->info()->padding());
70
71 // If there is no border: early exit
72 if(border_size.empty() || border_mode == BorderMode::UNDEFINED)
73 {
74 return;
75 }
76
77 // Select appropriate kernel
78 std::string kernel_name = "fill_image_borders_" + lower_string(string_from_border_mode(border_mode));
79
80 // Define build options
81 std::set<std::string> build_opts;
82 build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
83 build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
84 build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
85 build_opts.emplace("#define BORDER_SIZE_TOP " + support::cpp11::to_string(border_size.top));
86 build_opts.emplace("#define BORDER_SIZE_BOTTOM " + support::cpp11::to_string(border_size.bottom));
87 build_opts.emplace("#define BORDER_SIZE_LEFT " + support::cpp11::to_string(border_size.left));
88 build_opts.emplace("#define BORDER_SIZE_RIGHT " + support::cpp11::to_string(border_size.right));
89
90 if(border_mode == BorderMode::REPLICATE)
91 {
92 build_opts.emplace("#define FILL_IMAGE_BORDERS_REPLICATE\n");
93 }
94 else
95 {
96 build_opts.emplace("#define FILL_IMAGE_BORDERS_CONSTANT\n");
97 }
98
99 switch(tensor->info()->data_type())
100 {
101 case DataType::F16:
102 build_opts.emplace("#define DATA_TYPE_FP16");
103 break;
104
105 case DataType::F32:
106 build_opts.emplace("#define DATA_TYPE_FP32");
107 break;
108
109 default:
110 ARM_COMPUTE_ERROR("Current data type is not supported");
111 break;
112 }
113
114 // Create kernel
115 _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel(kernel_name, build_opts));
116 _tensor = tensor;
117
118 // Create static kernel arguments
119 const unsigned int valid_width = tensor->info()->valid_region().shape[0];
120 const unsigned int valid_height = tensor->info()->valid_region().shape[1];
121 const unsigned int total_valid_width = border_size.left + valid_width + border_size.right;
122
123 // Set static kernel arguments
124 unsigned int idx = num_arguments_per_3D_tensor(); //Skip the tensor parameters
125 _kernel.set_argument(idx++, valid_width);
126 _kernel.set_argument(idx++, valid_height);
127 _kernel.set_argument(idx++, tensor->info()->valid_region().anchor[0]);
128 _kernel.set_argument(idx++, tensor->info()->valid_region().anchor[1]);
129
130 if(BorderMode::CONSTANT == border_mode)
131 {
132 set_constant_border<float>(idx++, constant_border_value);
133 }
134
135 // Configure kernel window
136 Window win;
137 win.set(Window::DimX, Window::Dimension(0, total_valid_width + valid_height));
138 win.set(Window::DimY, Window::Dimension(0, 1, 1));
139 win.use_tensor_dimensions(tensor->info()->tensor_shape(), Window::DimZ);
140
141 IGCKernel::configure(win);
142 }
143
run(const Window & window)144 void GCFillBorderKernel::run(const Window &window)
145 {
146 // Border mode undefined or border width == 0
147 if(_kernel.get_program() == 0)
148 {
149 return;
150 }
151
152 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
153 ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IGCKernel::window(), window);
154
155 _kernel.use();
156 Window slice = window.first_slice_window_3D();
157
158 do
159 {
160 unsigned int idx = 0;
161 add_3D_tensor_argument(idx, _tensor, 1, slice);
162
163 _kernel.update_shader_params();
164
165 enqueue(*this, slice);
166 }
167 while(window.slide_window_slice_3D(slice));
168 }
169