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/CLWinogradOutputTransformKernel.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/IAccessWindow.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Types.h"
33 #include "arm_compute/core/Utils.h"
34 #include "arm_compute/core/Validate.h"
35 #include "arm_compute/core/Window.h"
36 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
37 #include "src/core/AccessWindowStatic.h"
38 #include "src/core/CL/CLValidate.h"
39 #include "src/core/helpers/AutoConfiguration.h"
40 #include "src/core/helpers/WindowHelpers.h"
41
42 #include "support/StringSupport.h"
43
44 #include <cmath>
45
46 namespace arm_compute
47 {
48 using namespace arm_compute::misc::shape_calculator;
49
50 namespace
51 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,const WinogradInfo & winograd_info,const ActivationLayerInfo & act_info)52 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info)
53 {
54 ARM_COMPUTE_UNUSED(act_info);
55 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
56 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
57
58 ARM_COMPUTE_RETURN_ERROR_ON(output->data_layout() != winograd_info.output_data_layout);
59
60 const PadStrideInfo conv_info = winograd_info.convolution_info;
61 const Size2D output_tile_size = winograd_info.output_tile_size;
62 const Size2D kernel_size = winograd_info.kernel_size;
63 const Size2D input_dimensions = winograd_info.input_dimensions;
64 const unsigned int num_channels = (winograd_info.kernel_size.width + winograd_info.output_tile_size.width - 1) * (winograd_info.kernel_size.height + winograd_info.output_tile_size.height - 1);
65
66 ARM_COMPUTE_RETURN_ERROR_ON_MSG(!cl_winograd_convolution_layer_supported(output_tile_size, kernel_size, winograd_info.output_data_layout), "Winograd output transform not supported");
67 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->dimension(2) != num_channels, "Wrong number of channels");
68
69 // Compute number of elements to process in the X and Y direction
70 // Compute the number of output tiles along the x and y direction of size "output_tile_size"
71 const Size2D num_tiles = compute_winograd_convolution_tiles(input_dimensions,
72 kernel_size,
73 output_tile_size,
74 conv_info);
75
76 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != static_cast<unsigned int>((num_tiles.area())));
77
78 if(bias != nullptr)
79 {
80 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
81 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
82 }
83
84 // Checks performed when output is configured
85 if(output->total_size() != 0)
86 {
87 const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input, winograd_info));
88
89 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
90 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
91 }
92
93 return Status{};
94 }
95
validate_and_configure_window(ITensorInfo * input,ITensorInfo * bias,ITensorInfo * output,const Size2D & output_tile_size)96 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output, const Size2D &output_tile_size)
97 {
98 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
99 ARM_COMPUTE_UNUSED(bias);
100
101 constexpr unsigned int num_elems_processed_per_iteration = 1;
102
103 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
104 bool window_changed = false;
105
106 if(output->data_layout() == DataLayout::NCHW)
107 {
108 const int output_static_window_end_x = ceil_to_multiple(output->dimension(0), output_tile_size.width);
109 const int output_static_window_end_y = ceil_to_multiple(output->dimension(1), output_tile_size.height);
110
111 AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration);
112 AccessWindowStatic output_access(output, 0, 0, output_static_window_end_x, output_static_window_end_y);
113 window_changed = update_window_and_padding(win, input_access, output_access);
114 output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
115 }
116
117 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
118 return std::make_pair(err, win);
119 }
120 } // namespace
121
CLWinogradOutputTransformKernel()122 CLWinogradOutputTransformKernel::CLWinogradOutputTransformKernel()
123 : _input(nullptr), _bias(nullptr), _output(nullptr), _is_nhwc(false)
124 {
125 }
126
configure(const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,const WinogradInfo & winograd_info,const ActivationLayerInfo & act_info)127 void CLWinogradOutputTransformKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info)
128 {
129 configure(CLKernelLibrary::get().get_compile_context(), input, bias, output, winograd_info, act_info);
130 }
131
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,const WinogradInfo & winograd_info,const ActivationLayerInfo & act_info)132 void CLWinogradOutputTransformKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info,
133 const ActivationLayerInfo &act_info)
134 {
135 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
136
137 // Output tensor auto initialization if not yet initialized
138 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input->info(), winograd_info)));
139
140 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info(), winograd_info, act_info));
141
142 auto padding_info = get_padding_info({ input, bias, output });
143
144 _input = input;
145 _bias = bias;
146 _output = output;
147 _is_nhwc = winograd_info.output_data_layout == DataLayout::NHWC;
148
149 // Compute num_tiles_x
150 const Size2D input_dimensions = winograd_info.input_dimensions;
151 const Size2D kernel_size = winograd_info.kernel_size;
152 const Size2D output_tile_size = winograd_info.output_tile_size;
153 const PadStrideInfo conv_info = winograd_info.convolution_info;
154 const int idx_width = get_data_layout_dimension_index(winograd_info.output_data_layout, DataLayoutDimension::WIDTH);
155 const int idx_height = get_data_layout_dimension_index(winograd_info.output_data_layout, DataLayoutDimension::HEIGHT);
156
157 // Compute the number of output tiles along the x and y direction of size "output_tile_size"
158 const Size2D num_tiles = compute_winograd_convolution_tiles(input_dimensions,
159 kernel_size,
160 output_tile_size,
161 conv_info);
162 const size_t total_batches = output->info()->tensor_shape().total_size_upper(3);
163
164 // Set build options
165 CLBuildOptions build_opts;
166 build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation())));
167 build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
168 build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
169
170 if((output_tile_size.x() == 2) || (output_tile_size.x() == 1 && output_tile_size.y() == 2))
171 {
172 build_opts.add_option("-DVEC_SIZE=2");
173 }
174 else if((output_tile_size.x() == 4) || (output_tile_size.x() == 1 && output_tile_size.y() == 4))
175 {
176 build_opts.add_option("-DVEC_SIZE=4");
177 }
178
179 build_opts.add_option_if(_bias != nullptr, std::string("-DHAS_BIAS"));
180 build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width));
181 build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
182 build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
183 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
184 build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(idx_width)));
185 build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(_output->info()->dimension(idx_height)));
186 build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(_input->info()->dimension(2)));
187 build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL");
188 build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL");
189
190 // Create kernel
191 std::string kernel_name = "winograd_output_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string() + "_" + lower_string(string_from_data_layout(winograd_info.output_data_layout));
192 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
193
194 // Configure kernel window
195 auto win_config = validate_and_configure_window(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info(), winograd_info.output_tile_size);
196 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
197 ICLKernel::configure_internal(win_config.second);
198
199 // Set config_id for enabling LWS tuning
200 _config_id = kernel_name;
201 _config_id += "_";
202 _config_id += lower_string(string_from_data_type(input->info()->data_type()));
203 _config_id += "_";
204 _config_id += support::cpp11::to_string(input->info()->dimension(0));
205 _config_id += "_";
206 _config_id += support::cpp11::to_string(input->info()->dimension(1));
207 _config_id += "_";
208 _config_id += support::cpp11::to_string(output->info()->dimension(0));
209 _config_id += "_";
210 _config_id += support::cpp11::to_string(output->info()->dimension(1));
211 _config_id += "_";
212 _config_id += lower_string(string_from_data_layout(winograd_info.output_data_layout));
213
214 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info) && _is_nhwc);
215 }
216
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,const WinogradInfo & winograd_info,const ActivationLayerInfo & act_info)217 Status CLWinogradOutputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info)
218 {
219 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, (bias != nullptr ? bias->clone().get() : nullptr), output, winograd_info, act_info));
220 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), output->clone().get(), winograd_info.output_tile_size).first);
221
222 return Status{};
223 }
224
run(const Window & window,cl::CommandQueue & queue)225 void CLWinogradOutputTransformKernel::run(const Window &window, cl::CommandQueue &queue)
226 {
227 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
228 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
229
230 // Collapse window
231 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
232
233 // Get initial windows
234 Window slice = window_collapsed.first_slice_window_4D();
235 slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
236
237 // Setup output slice
238 Window slice_out(slice);
239 slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
240 slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
241
242 if(_bias != nullptr)
243 {
244 unsigned int idx1 = 2 * num_arguments_per_4D_tensor();
245 Window slice_biases;
246 slice_biases.use_tensor_dimensions(_bias->info()->tensor_shape());
247 add_1D_tensor_argument(idx1, _bias, slice_biases);
248 }
249
250 if(_is_nhwc)
251 {
252 unsigned int idx2 = 2 * num_arguments_per_4D_tensor() + ((_bias != nullptr) ? num_arguments_per_1D_tensor() : 0);
253 _kernel.setArg(idx2, static_cast<int>(_output->info()->total_size() - _output->info()->strides_in_bytes().y()));
254 }
255
256 do
257 {
258 unsigned int idx = 0;
259 add_4D_tensor_argument(idx, _input, slice);
260 add_4D_tensor_argument(idx, _output, slice_out);
261 enqueue(queue, *this, slice, lws_hint());
262 }
263 while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_out));
264 }
265 } // namespace arm_compute
266