xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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/CLChannelShuffleLayerKernel.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 "src/core/CL/CLValidate.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34 #include "support/StringSupport.h"
35 
36 namespace arm_compute
37 {
38 namespace
39 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,unsigned int num_groups)40 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups)
41 {
42     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
43     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
44     ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups < 2, "Channel shuffling with less than 2 groups would be inefficient");
45 
46     const unsigned int channels = input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL));
47 
48     ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups == channels, "Channel shuffling with same number of groups as number of channels would be inefficient");
49     // There cannot be more groups than channels
50     ARM_COMPUTE_RETURN_ERROR_ON(num_groups > channels);
51     ARM_COMPUTE_RETURN_ERROR_ON_MSG((channels % num_groups) != 0, "The number of channels must be a multiple of the number of groups");
52 
53     // Checks performed when output is configured
54     if(output->total_size() != 0)
55     {
56         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
57         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
58         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
59     }
60 
61     return Status{};
62 }
63 
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output)64 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
65 {
66     // Output tensor auto initialization if not yet initialized
67     auto_init_if_empty(*output, *input->clone());
68 
69     const bool is_nhwc = input->data_layout() == DataLayout::NHWC;
70     if(is_nhwc)
71     {
72         unsigned int num_elems_processed_per_iteration_x = adjust_vec_size(max_cl_vector_width / input->element_size(), input->dimension(0));
73         Window       win                                 = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x));
74         Window       win_collapsed                       = win.collapse(win, Window::DimZ);
75         return std::make_pair(Status{}, win_collapsed);
76     }
77     else
78     {
79         const unsigned int     num_elems_processed_per_iteration_x = max_cl_vector_width / input->element_size();
80         constexpr unsigned int num_elems_processed_per_iteration_y = 2;
81 
82         // Configure kernel window
83         Window                win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
84         AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
85         AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
86 
87         const bool window_changed = update_window_and_padding(win, input_access, output_access);
88 
89         Window win_collapsed = win.collapse(win, Window::DimZ);
90 
91         Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
92         return std::make_pair(err, win_collapsed);
93     }
94 }
95 } // namespace
96 
CLChannelShuffleLayerKernel()97 CLChannelShuffleLayerKernel::CLChannelShuffleLayerKernel()
98     : _input(nullptr), _output(nullptr)
99 {
100     _type = CLKernelType::ELEMENTWISE;
101 }
102 
configure(const ICLTensor * input,ICLTensor * output,unsigned int num_groups)103 void CLChannelShuffleLayerKernel::configure(const ICLTensor *input, ICLTensor *output, unsigned int num_groups)
104 {
105     configure(CLKernelLibrary::get().get_compile_context(), input, output, num_groups);
106 }
107 
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,unsigned int num_groups)108 void CLChannelShuffleLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, unsigned int num_groups)
109 {
110     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
111     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), num_groups));
112     auto padding_info = get_padding_info({ input, output });
113 
114     _input  = input;
115     _output = output;
116 
117     const DataLayout   data_layout          = input->info()->data_layout();
118     const bool         is_nhwc              = data_layout == DataLayout::NHWC;
119     const unsigned int channels             = input->info()->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL));
120     unsigned int       vec_size_x           = 0;
121     unsigned int       vec_size_x_leftovers = 0;
122     if(is_nhwc)
123     {
124         vec_size_x           = adjust_vec_size(max_cl_vector_width / input->info()->element_size(), input->info()->dimension(0));
125         vec_size_x_leftovers = input->info()->dimension(0) % vec_size_x;
126     }
127     else
128     {
129         vec_size_x = max_cl_vector_width / input->info()->element_size();
130     }
131 
132     // Set kernel build options
133     CLBuildOptions build_opts;
134     build_opts.add_option("-DNUM_GROUPS=" + support::cpp11::to_string(num_groups));
135     build_opts.add_option("-DK=" + support::cpp11::to_string(channels / num_groups));
136     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
137     build_opts.add_option_if(is_nhwc, "-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftovers));
138     build_opts.add_option_if(is_nhwc, "-DSRC_DIM_X=" + support::cpp11::to_string(input->info()->dimension(0)));
139     build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(input->info()->dimension(2)));
140     build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
141 
142     // Create kernel
143     std::string kernel_name = "channel_shuffle_" + lower_string(string_from_data_layout(data_layout));
144 
145     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
146 
147     // Configure kernel window
148     auto win_config = validate_and_configure_window(input->info(), output->info());
149     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
150     ICLKernel::configure_internal(win_config.second);
151 
152     // Set config_id for enabling LWS tuning
153     _config_id = kernel_name;
154     _config_id += "_";
155     _config_id += lower_string(string_from_data_type(input->info()->data_type()));
156     _config_id += "_";
157     _config_id += support::cpp11::to_string(num_groups);
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     _config_id += "_";
165     _config_id += support::cpp11::to_string(output->info()->dimension(0));
166     _config_id += "_";
167     _config_id += support::cpp11::to_string(output->info()->dimension(1));
168     _config_id += "_";
169     _config_id += support::cpp11::to_string(output->info()->dimension(2));
170     if(data_layout == DataLayout::NHWC)
171     {
172         ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
173     }
174 }
175 
validate(const ITensorInfo * input,const ITensorInfo * output,unsigned int num_groups)176 Status CLChannelShuffleLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups)
177 {
178     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, num_groups));
179     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
180 
181     return Status{};
182 }
183 
run(const Window & window,cl::CommandQueue & queue)184 void CLChannelShuffleLayerKernel::run(const Window &window, cl::CommandQueue &queue)
185 {
186     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
187     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
188 
189     unsigned int idx = 0;
190     add_4D_tensor_argument(idx, _input, window);
191     add_4D_tensor_argument(idx, _output, window);
192     enqueue(queue, *this, window, lws_hint());
193 }
194 } // namespace arm_compute
195