1 /*
2 * Copyright (c) 2018-2021 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/CLNormalizePlanarYUVLayerKernel.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/TensorInfo.h"
31 #include "arm_compute/core/Utils.h"
32 #include "src/core/AccessWindowStatic.h"
33 #include "src/core/CL/CLValidate.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36
37 #include "support/StringSupport.h"
38
39 namespace arm_compute
40 {
41 namespace
42 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const ITensorInfo * mean,const ITensorInfo * std)43 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *std)
44 {
45 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
46 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
47 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
48
49 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, std);
50 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, std);
51 ARM_COMPUTE_RETURN_ERROR_ON_MSG(mean->num_dimensions() > 1, "mean and std must be vectors");
52
53 const unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
54 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(channel_idx) != mean->dimension(0));
55
56 // Checks performed when output is configured
57 if(output->total_size() != 0)
58 {
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
60 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
61 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
62 }
63
64 return Status{};
65 }
66
validate_and_configure_window_nchw(ITensorInfo * input,ITensorInfo * output)67 std::pair<Status, Window> validate_and_configure_window_nchw(ITensorInfo *input, ITensorInfo *output)
68 {
69 const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
70
71 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
72
73 AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
74 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
75
76 bool window_changed = update_window_and_padding(win, input_access, output_access);
77
78 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
79 return std::make_pair(err, win);
80 }
81 } // namespace
82
CLNormalizePlanarYUVLayerKernel()83 CLNormalizePlanarYUVLayerKernel::CLNormalizePlanarYUVLayerKernel()
84 : _input(nullptr), _output(nullptr), _mean(nullptr), _std(nullptr)
85 {
86 _type = CLKernelType::ELEMENTWISE;
87 }
88
configure(const ICLTensor * input,ICLTensor * output,const ICLTensor * mean,const ICLTensor * std)89 void CLNormalizePlanarYUVLayerKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *mean, const ICLTensor *std)
90 {
91 configure(CLKernelLibrary::get().get_compile_context(), input, output, mean, std);
92 }
93
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const ICLTensor * mean,const ICLTensor * std)94 void CLNormalizePlanarYUVLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const ICLTensor *mean, const ICLTensor *std)
95 {
96 // Perform validation step
97 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, mean, std);
98 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), mean->info(), std->info()));
99
100 // Output tensor auto initialization if not yet initialized
101 auto_init_if_empty(*output->info(), *input->info()->clone());
102
103 auto padding_info = get_padding_info({ input, output });
104
105 _input = input;
106 _output = output;
107 _mean = mean;
108 _std = std;
109
110 const DataLayout data_layout = input->info()->data_layout();
111
112 // Get number of elements to process per iterations
113 const unsigned int num_elems_processed_per_iteration = (data_layout == DataLayout::NHWC) ? adjust_vec_size(16 / input->info()->element_size(),
114 input->info()->dimension(0)) :
115 (16 / input->info()->element_size());
116 const unsigned int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
117 const DataType dt = input->info()->data_type();
118
119 // Set build options
120 CLBuildOptions build_opts;
121 build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)));
122 build_opts.add_option(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
123 build_opts.add_option(("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_processed_per_iteration)));
124 build_opts.add_option(("-DNUM_CHANNELS=" + support::cpp11::to_string(input->info()->dimension(channel_idx))));
125
126 std::string kernel_name = "normalize_planar_yuv_layer_";
127 if(is_data_type_quantized(dt))
128 {
129 const UniformQuantizationInfo qinfo = input->info()->quantization_info().uniform();
130 build_opts.add_option(("-DOFFSET=" + support::cpp11::to_string(qinfo.offset)));
131 build_opts.add_option(("-DSCALE=" + support::cpp11::to_string(qinfo.scale)));
132 kernel_name += "q8_";
133 }
134
135 // Create kernel
136 kernel_name += lower_string(string_from_data_layout(data_layout));
137 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
138
139 // Configure kernel window
140 if(data_layout == DataLayout::NHWC)
141 {
142 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
143 ICLKernel::configure_internal(win);
144 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
145 }
146 else
147 {
148 auto win_config = validate_and_configure_window_nchw(input->info(), output->info());
149 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
150 ICLKernel::configure_internal(win_config.second);
151 }
152
153 // Set config_id for enabling LWS tuning
154 _config_id = "normalize_planar_yuv_layer_";
155 _config_id += lower_string(string_from_data_layout(input->info()->data_layout()));
156 _config_id += "_";
157 _config_id += lower_string(string_from_data_type(dt));
158 _config_id += "_";
159 _config_id += support::cpp11::to_string(input->info()->dimension(0));
160 _config_id += "_";
161 _config_id += support::cpp11::to_string(input->info()->dimension(1));
162 _config_id += "_";
163 _config_id += support::cpp11::to_string(input->info()->dimension(2));
164 }
165
validate(const ITensorInfo * input,const ITensorInfo * output,const ITensorInfo * mean,const ITensorInfo * std)166 Status CLNormalizePlanarYUVLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *std)
167 {
168 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mean, std));
169 if(input->data_layout() == DataLayout::NCHW)
170 {
171 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_nchw(input->clone().get(), output->clone().get()).first);
172 }
173 return Status{};
174 }
175
run(const Window & window,cl::CommandQueue & queue)176 void CLNormalizePlanarYUVLayerKernel::run(const Window &window, cl::CommandQueue &queue)
177 {
178 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
179 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
180
181 Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
182 Window slice = collapsed.first_slice_window_3D();
183
184 Window slice_in = collapsed.first_slice_window_1D();
185 slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
186
187 unsigned int idx = 2 * num_arguments_per_3D_tensor();
188 add_1D_tensor_argument(idx, _mean, slice_in);
189 add_1D_tensor_argument(idx, _std, slice_in);
190
191 do
192 {
193 idx = 0;
194 add_3D_tensor_argument(idx, _input, slice);
195 add_3D_tensor_argument(idx, _output, slice);
196 enqueue(queue, *this, slice, lws_hint());
197 }
198 while(collapsed.slide_window_slice_3D(slice));
199 }
200 } // namespace arm_compute
201