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/GCNormalizationLayerKernel.h"
25
26 #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
27 #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
28 #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
29 #include "arm_compute/core/Helpers.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Utils.h"
32 #include "arm_compute/core/Validate.h"
33 #include "arm_compute/core/Window.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36 #include "support/StringSupport.h"
37
38 #include <string>
39
40 using namespace arm_compute;
41
GCNormalizationLayerKernel()42 GCNormalizationLayerKernel::GCNormalizationLayerKernel()
43 : _input(nullptr), _squared_input(nullptr), _output(nullptr), _border_size(0)
44 {
45 }
46
border_size() const47 BorderSize GCNormalizationLayerKernel::border_size() const
48 {
49 return _border_size;
50 }
51
configure(const IGCTensor * input,const IGCTensor * squared_input,IGCTensor * output,NormalizationLayerInfo norm_info)52 void GCNormalizationLayerKernel::configure(const IGCTensor *input, const IGCTensor *squared_input, IGCTensor *output, NormalizationLayerInfo norm_info)
53 {
54 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
55 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
56 ARM_COMPUTE_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd");
57 ARM_COMPUTE_ERROR_ON_MSG(norm_info.type() == NormType::IN_MAP_2D, "2D In-Map Normalization not implemented");
58
59 // Set build options
60 std::set<std::string> build_opts;
61
62 _input = input;
63 _squared_input = squared_input;
64 _output = output;
65
66 const bool is_in_map = norm_info.is_in_map();
67 const unsigned int border_width = is_in_map ? std::min(norm_info.norm_size() / 2, 3U) : 0;
68 _border_size = BorderSize(0, border_width);
69
70 // Set kernel static arguments
71 std::string func_name = ((norm_info.type() == NormType::IN_MAP_1D) ? "IN_MAP_1D" : "CROSS_MAP");
72 build_opts.emplace(("#define " + func_name));
73 build_opts.emplace(("#define COEFF " + float_to_string_with_full_precision(norm_info.scale_coeff())));
74 build_opts.emplace(("#define BETA " + float_to_string_with_full_precision(norm_info.beta())));
75 build_opts.emplace(("#define KAPPA " + float_to_string_with_full_precision(norm_info.kappa())));
76 build_opts.emplace(("#define RADIUS " + support::cpp11::to_string(norm_info.norm_size() / 2)));
77 build_opts.emplace(("#define LOCAL_SIZE_X " + support::cpp11::to_string(1)));
78 build_opts.emplace(("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)));
79 build_opts.emplace(("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)));
80
81 // Create kernel
82 _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("normalization_layer", build_opts));
83
84 // Configure kernel window
85 const unsigned int num_elems_processed_per_iteration = 1;
86 const unsigned int num_elems_read_per_iteration = num_elems_processed_per_iteration + 2 * (norm_info.norm_size() / 2);
87
88 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
89
90 AccessWindowHorizontal input_access(input->info(), -_border_size.left, num_elems_read_per_iteration);
91 AccessWindowHorizontal squared_input_access(squared_input->info(), -_border_size.left, num_elems_read_per_iteration);
92 AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
93
94 update_window_and_padding(win, input_access, squared_input_access, output_access);
95
96 output_access.set_valid_region(win, input->info()->valid_region());
97
98 IGCKernel::configure(win);
99 }
100
run(const Window & window)101 void GCNormalizationLayerKernel::run(const Window &window)
102 {
103 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
104 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
105
106 _kernel.use();
107
108 Window slice = window.first_slice_window_3D();
109
110 do
111 {
112 unsigned int idx = 0;
113 unsigned int binding = 1;
114 add_3D_tensor_argument(idx, _input, binding++, slice);
115 add_3D_tensor_argument(idx, _squared_input, binding++, slice);
116 add_3D_tensor_argument(idx, _output, binding++, slice);
117
118 _kernel.update_shader_params();
119
120 enqueue(*this, slice);
121 }
122 while(window.slide_window_slice_3D(slice));
123 }
124