xref: /aosp_15_r20/external/ComputeLibrary/src/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2018-2022 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/gpu/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 #include "support/Cast.h"
42 #include "support/StringSupport.h"
43 
44 #include <cmath>
45 
46 using namespace arm_compute::misc::shape_calculator;
47 
48 namespace arm_compute
49 {
50 namespace opencl
51 {
52 namespace kernels
53 {
54 namespace
55 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,const WinogradInfo & winograd_info,const ActivationLayerInfo & act_info)56 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info)
57 {
58     ARM_COMPUTE_UNUSED(act_info);
59     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
60     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
61 
62     ARM_COMPUTE_RETURN_ERROR_ON(output->data_layout() != winograd_info.output_data_layout);
63 
64     const PadStrideInfo conv_info        = winograd_info.convolution_info;
65     const Size2D        output_tile_size = winograd_info.output_tile_size;
66     const Size2D        kernel_size      = winograd_info.kernel_size;
67     const Size2D        input_dimensions = winograd_info.input_dimensions;
68     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);
69 
70     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");
71     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->dimension(2) != num_channels, "Wrong number of channels");
72 
73     // Compute number of elements to process in the X and Y direction
74     // Compute the number of output tiles along the x and y direction of size "output_tile_size"
75     const Size2D num_tiles = compute_winograd_convolution_tiles(input_dimensions,
76                                                                 kernel_size,
77                                                                 output_tile_size,
78                                                                 conv_info);
79 
80     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != static_cast<unsigned int>((num_tiles.area())));
81 
82     if(bias != nullptr)
83     {
84         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
85         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
86     }
87 
88     // Checks performed when output is configured
89     if(output->total_size() != 0)
90     {
91         const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input, winograd_info));
92 
93         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
94         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
95     }
96 
97     return Status{};
98 }
99 
validate_and_configure_window(ITensorInfo * input,ITensorInfo * bias,ITensorInfo * output,const Size2D & output_tile_size)100 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output, const Size2D &output_tile_size)
101 {
102     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
103     ARM_COMPUTE_UNUSED(bias);
104 
105     constexpr unsigned int num_elems_processed_per_iteration = 1;
106 
107     Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
108     bool   window_changed = false;
109 
110     if(output->data_layout() == DataLayout::NCHW)
111     {
112         const int output_static_window_end_x = ceil_to_multiple(output->dimension(0), output_tile_size.width);
113         const int output_static_window_end_y = ceil_to_multiple(output->dimension(1), output_tile_size.height);
114 
115         AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration);
116         AccessWindowStatic    output_access(output, 0, 0, output_static_window_end_x, output_static_window_end_y);
117         window_changed = update_window_and_padding(win, input_access, output_access);
118     }
119 
120     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
121     return std::make_pair(err, win);
122 }
123 } // namespace
124 
ClWinogradOutputTransformKernel()125 ClWinogradOutputTransformKernel::ClWinogradOutputTransformKernel()
126 {
127     _type = CLKernelType::WINOGRAD;
128 }
129 
configure(const ClCompileContext & compile_context,ITensorInfo * src,ITensorInfo * bias,ITensorInfo * dst,const WinogradInfo & winograd_info,const ActivationLayerInfo & act_info)130 void ClWinogradOutputTransformKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const WinogradInfo &winograd_info,
131                                                 const ActivationLayerInfo &act_info)
132 {
133     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
134 
135     // Output tensor auto initialization if not yet initialized
136     auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*src, winograd_info)));
137 
138     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, winograd_info, act_info));
139 
140     // Configure kernel window
141     auto win_config = validate_and_configure_window(src, bias, dst, winograd_info.output_tile_size);
142     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
143     IClKernel::configure_internal(win_config.second);
144 
145     auto padding_info = get_padding_info({ src, bias, dst });
146 
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 = dst->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     _num_tiles_x = num_tiles.width;
180 
181     // Conditions of -cl-fast-relaxed-math causing accuracy issues can be traced from COMPMID-5324
182     const GPUTarget gpu_target    = get_target();
183     const auto      act_function  = act_info.activation();
184     const auto      src_data_type = src->data_type();
185 
186     if((gpu_target != GPUTarget::G71 && (gpu_target & GPUTarget::GPU_ARCH_MASK) == GPUTarget::BIFROST)
187        && (act_function == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU || act_function == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
188        && (src_data_type == DataType::F32 || src_data_type == DataType::F16))
189     {
190         // -cl-fast-relaxed-math also sets -cl-finite-math-only and -cl-unsafe-math-optimizations
191         // to disable -cl-finite-math-only, we only include -cl-unsafe-math-optimizations
192         build_opts.add_option("-cl-unsafe-math-optimizations");
193     }
194     else
195     {
196         build_opts.add_option("-cl-fast-relaxed-math");
197     }
198 
199     if(_is_nhwc)
200     {
201         build_opts.add_option_if(bias != nullptr, std::string("-DHAS_BIAS"));
202         build_opts.add_option("-DN0=" + support::cpp11::to_string(win_config.second.x().step()));
203         build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
204         build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
205         build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src_data_type));
206         build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(2)));
207         build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL");
208         build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL");
209         build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(_num_tiles_x));
210     }
211     else
212     {
213         build_opts.add_option_if(bias != nullptr, std::string("-DHAS_BIAS"));
214         build_opts.add_option("-DN0=" + support::cpp11::to_string(win_config.second.x().step()));
215         build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width));
216         build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
217         build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
218         build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src_data_type));
219         build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(1)));
220         build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst->dimension(idx_width)));
221         build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(idx_height)));
222         build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(2)));
223         build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL");
224         build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL");
225     }
226 
227     // Storing tensor dimensions to be sent later as kernel arguments
228     _src_height = src->dimension(1);
229     _dst_width  = dst->dimension(idx_width);
230     _dst_height = dst->dimension(idx_height);
231 
232     // Create kernel
233     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));
234 
235     // A macro guard to compile ONLY the kernel of interest
236     build_opts.add_option("-D" + upper_string(kernel_name));
237     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
238 
239     // Set config_id for enabling LWS tuning
240     _config_id = kernel_name;
241     _config_id += "_";
242     _config_id += lower_string(string_from_data_type(src_data_type));
243     _config_id += "_";
244     _config_id += support::cpp11::to_string(src->dimension(0));
245     _config_id += "_";
246     _config_id += support::cpp11::to_string(src->dimension(1));
247     _config_id += "_";
248     _config_id += support::cpp11::to_string(dst->dimension(0));
249     _config_id += "_";
250     _config_id += support::cpp11::to_string(dst->dimension(1));
251     _config_id += "_";
252     _config_id += lower_string(string_from_data_layout(winograd_info.output_data_layout));
253 
254     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info) && _is_nhwc);
255 }
256 
validate(const ITensorInfo * src,const ITensorInfo * bias,const ITensorInfo * dst,const WinogradInfo & winograd_info,const ActivationLayerInfo & act_info)257 Status ClWinogradOutputTransformKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info)
258 {
259     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, (bias != nullptr ? bias->clone().get() : nullptr), dst, winograd_info, act_info));
260     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), dst->clone().get(), winograd_info.output_tile_size).first);
261     return Status{};
262 }
263 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)264 void ClWinogradOutputTransformKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
265 {
266     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
267     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IClKernel::window(), window);
268 
269     auto src  = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
270     auto bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
271     auto dst  = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
272 
273     // Collapse window
274     Window window_collapsed = window.collapse_if_possible(IClKernel::window(), Window::DimZ);
275 
276     // Get initial windows
277     Window slice = window_collapsed.first_slice_window_4D();
278     slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
279 
280     // Setup output slice
281     Window slice_out(slice);
282     slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
283     slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
284 
285     if(bias != nullptr)
286     {
287         unsigned int idx1 = 2 * num_arguments_per_4D_tensor();
288         Window       slice_biases;
289         slice_biases.use_tensor_dimensions(bias->info()->tensor_shape());
290         add_1D_tensor_argument(idx1, bias, slice_biases);
291     }
292 
293     if(_is_nhwc)
294     {
295         unsigned int idx2 = 2 * num_arguments_per_4D_tensor() + ((bias != nullptr) ? num_arguments_per_1D_tensor() : 0);
296         _kernel.setArg(idx2++, static_cast<int>(dst->info()->total_size() - dst->info()->strides_in_bytes().y()));
297         _kernel.setArg<cl_int>(idx2++, _src_height);
298         _kernel.setArg<cl_int>(idx2++, _dst_width);
299         _kernel.setArg<cl_int>(idx2++, _dst_height);
300     }
301 
302     do
303     {
304         unsigned int idx = 0;
305         add_4D_tensor_argument(idx, src, slice);
306         add_4D_tensor_argument(idx, dst, slice_out);
307         enqueue(queue, *this, slice, lws_hint());
308     }
309     while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_out));
310 }
311 } // namespace kernels
312 } // namespace opencl
313 } // namespace arm_compute
314