1 /*
2 * Copyright (c) 2018-2020 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 "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 const unsigned int num_elems_processed_per_iteration_x = is_nhwc ? 4 : max_cl_vector_width / input->element_size();
71 constexpr unsigned int num_elems_processed_per_iteration_y = 2;
72
73 // Configure kernel window
74 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
75 AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
76 AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
77
78 const bool window_changed = update_window_and_padding(win, input_access, output_access);
79 output_access.set_valid_region(win, input->valid_region());
80
81 Window win_collapsed = win.collapse(win, Window::DimZ);
82
83 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
84 return std::make_pair(err, win_collapsed);
85 }
86 } // namespace
87
CLChannelShuffleLayerKernel()88 CLChannelShuffleLayerKernel::CLChannelShuffleLayerKernel()
89 : _input(nullptr), _output(nullptr)
90 {
91 }
92
configure(const ICLTensor * input,ICLTensor * output,unsigned int num_groups)93 void CLChannelShuffleLayerKernel::configure(const ICLTensor *input, ICLTensor *output, unsigned int num_groups)
94 {
95 configure(CLKernelLibrary::get().get_compile_context(), input, output, num_groups);
96 }
97
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,unsigned int num_groups)98 void CLChannelShuffleLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, unsigned int num_groups)
99 {
100 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
101
102 _input = input;
103 _output = output;
104
105 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), num_groups));
106
107 const DataLayout data_layout = input->info()->data_layout();
108 const bool is_nhwc = data_layout == DataLayout::NHWC;
109 const unsigned int channels = input->info()->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL));
110 const unsigned int vec_size = is_nhwc ? 4 : max_cl_vector_width / input->info()->element_size();
111
112 // Set kernel build options
113 CLBuildOptions build_opts;
114 build_opts.add_option("-DNUM_GROUPS=" + support::cpp11::to_string(num_groups));
115 build_opts.add_option("-DK=" + support::cpp11::to_string(channels / num_groups));
116 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size));
117 build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(input->info()->dimension(2)));
118 build_opts.add_option("-DLAST_ACCESSED=" + support::cpp11::to_string(std::max(static_cast<int>(channels - vec_size), 0)));
119 build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
120
121 // Create kernel
122 std::string kernel_name = "channel_shuffle_" + lower_string(string_from_data_layout(data_layout));
123
124 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
125
126 // Configure kernel window
127 auto win_config = validate_and_configure_window(input->info(), output->info());
128 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
129 ICLKernel::configure_internal(win_config.second);
130
131 // Set config_id for enabling LWS tuning
132 _config_id = kernel_name;
133 _config_id += "_";
134 _config_id += lower_string(string_from_data_type(input->info()->data_type()));
135 _config_id += "_";
136 _config_id += support::cpp11::to_string(num_groups);
137 _config_id += "_";
138 _config_id += support::cpp11::to_string(input->info()->dimension(0));
139 _config_id += "_";
140 _config_id += support::cpp11::to_string(input->info()->dimension(1));
141 _config_id += "_";
142 _config_id += support::cpp11::to_string(input->info()->dimension(2));
143 _config_id += "_";
144 _config_id += support::cpp11::to_string(output->info()->dimension(0));
145 _config_id += "_";
146 _config_id += support::cpp11::to_string(output->info()->dimension(1));
147 _config_id += "_";
148 _config_id += support::cpp11::to_string(output->info()->dimension(2));
149 }
150
validate(const ITensorInfo * input,const ITensorInfo * output,unsigned int num_groups)151 Status CLChannelShuffleLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups)
152 {
153 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, num_groups));
154 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
155
156 return Status{};
157 }
158
run(const Window & window,cl::CommandQueue & queue)159 void CLChannelShuffleLayerKernel::run(const Window &window, cl::CommandQueue &queue)
160 {
161 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
162 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
163
164 unsigned int idx = 0;
165 add_4D_tensor_argument(idx, _input, window);
166 add_4D_tensor_argument(idx, _output, window);
167 enqueue(queue, *this, window, lws_hint());
168 }
169 } // namespace arm_compute
170