1 /*
2 * Copyright (c) 2017-2020 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
25 #include "arm_compute/core/GLES_COMPUTE/kernels/GCIm2ColKernel.h"
26
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
29 #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
30 #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
31 #include "arm_compute/core/GLES_COMPUTE/OpenGLES.h"
32 #include "arm_compute/core/Helpers.h"
33 #include "arm_compute/core/Size2D.h"
34 #include "arm_compute/core/TensorInfo.h"
35 #include "arm_compute/core/Types.h"
36 #include "arm_compute/core/Validate.h"
37 #include "src/core/AccessWindowStatic.h"
38 #include "src/core/helpers/AutoConfiguration.h"
39 #include "src/core/helpers/WindowHelpers.h"
40 #include "support/StringSupport.h"
41
42 #include <cmath>
43 #include <tuple>
44
45 using namespace arm_compute;
46
47 namespace
48 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output)49 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
50 {
51 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
52 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
53
54 // Checks performed when output is configured
55 if(output->total_size() != 0)
56 {
57 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
58 }
59
60 return Status{};
61 }
62 } // namespace
63
GCIm2ColKernel()64 GCIm2ColKernel::GCIm2ColKernel()
65 : _input(nullptr), _output(nullptr), _convolved_dims(), _kernel_dims(), _num_elems_processed_per_iteration(1), _run_func(nullptr)
66 {
67 }
68
configure(const IGCTensor * input,IGCTensor * output,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation)69 void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation)
70 {
71 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
72
73 // Perform validation step
74 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
75
76 _input = input;
77 _output = output;
78
79 // Create kernel
80 std::set<std::string> build_opts;
81 std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
82 build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
83 build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
84 build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
85 build_opts.insert("#define " + dt_name);
86
87 if(has_bias)
88 {
89 build_opts.emplace("#define HAS_BIAS");
90 }
91
92 int stride_x = 0;
93 int stride_y = 0;
94
95 std::tie(stride_x, stride_y) = conv_info.stride();
96 _kernel_dims = std::make_pair(kernel_dims.width, kernel_dims.height);
97
98 const bool run_img2col_reduced = (output->info()->dimension(0) == (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))) && (TensorShape::num_max_dimensions >= 4)
99 && (std::equal(input->info()->tensor_shape().cbegin() + 3,
100 input->info()->tensor_shape().cend(),
101 output->info()->tensor_shape().cbegin() + 1))
102 && ((stride_x == 1) && (stride_y == 1) && !conv_info.has_padding())
103 && (dilation == Size2D(1U, 1U));
104
105 std::string kernel_name = "im2col_generic";
106 if(!run_img2col_reduced)
107 {
108 if(input->info()->data_type() == DataType::F16 && _kernel_dims == std::pair<unsigned int, unsigned int>(1, 1))
109 {
110 build_opts.emplace("#define KERNEL_1x1");
111 }
112
113 build_opts.emplace("#define IM2COL_GENERIC");
114 _convolved_dims = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1),
115 kernel_dims.width, kernel_dims.height,
116 conv_info, dilation);
117 _num_elems_processed_per_iteration = (input->info()->data_type() == DataType::F32) ? 1 : 2;
118
119 build_opts.emplace("#define KERNEL_WIDTH " + support::cpp11::to_string(kernel_dims.width));
120 build_opts.emplace("#define KERNEL_HEIGHT " + support::cpp11::to_string(kernel_dims.height));
121 build_opts.emplace("#define KERNEL_DEPTH " + support::cpp11::to_string(input->info()->dimension(2)));
122 build_opts.emplace("#define CONVOLVED_WIDTH " + support::cpp11::to_string(_convolved_dims.first));
123 build_opts.emplace("#define CONVOLVED_HEIGHT " + support::cpp11::to_string(_convolved_dims.second));
124 build_opts.emplace("#define STRIDE_X " + support::cpp11::to_string(conv_info.stride().first));
125 build_opts.emplace("#define STRIDE_Y " + support::cpp11::to_string(conv_info.stride().second));
126 build_opts.emplace("#define PAD_LEFT " + support::cpp11::to_string(conv_info.pad_left()));
127 build_opts.emplace("#define PAD_TOP " + support::cpp11::to_string(conv_info.pad_top()));
128 build_opts.emplace("#define PAD_RIGHT " + support::cpp11::to_string(conv_info.pad_right()));
129 build_opts.emplace("#define PAD_BOTTOM " + support::cpp11::to_string(conv_info.pad_bottom()));
130 build_opts.emplace("#define SRC_WIDTH " + support::cpp11::to_string(input->info()->dimension(0)));
131 build_opts.emplace("#define SRC_HEIGHT " + support::cpp11::to_string(input->info()->dimension(1)));
132 build_opts.emplace("#define DILATION_X " + support::cpp11::to_string(dilation.x()));
133 build_opts.emplace("#define DILATION_Y " + support::cpp11::to_string(dilation.y()));
134
135 _run_func = &GCIm2ColKernel::run_generic;
136 }
137 else
138 {
139 build_opts.emplace("#define IM2COL_REDUCED");
140 kernel_name = "im2col_reduced";
141
142 if(input->info()->data_type() == DataType::F32)
143 {
144 _num_elems_processed_per_iteration = 4 / input->info()->element_size();
145 }
146 else if(input->info()->data_type() == DataType::F16)
147 {
148 int input_width = input->info()->dimension(0);
149 int input_height = input->info()->dimension(1);
150
151 build_opts.emplace("#define IMAGE_SIZE " + support::cpp11::to_string(input_width * input_height));
152 if(input_width % 8 == 0)
153 {
154 _num_elems_processed_per_iteration = 8;
155 build_opts.emplace("#define IM2COL_REDUCED_8X");
156 }
157 else if(input_width % 4 == 0)
158 {
159 _num_elems_processed_per_iteration = 4;
160 build_opts.emplace("#define IM2COL_REDUCED_4X");
161 }
162 else if(input_width % 2 == 0)
163 {
164 _num_elems_processed_per_iteration = 2;
165 build_opts.emplace("#define IM2COL_REDUCED_2X");
166 }
167 else
168 {
169 _num_elems_processed_per_iteration = 2;
170 build_opts.emplace("#define IM2COL_REDUCED_GENERIC");
171 }
172 }
173
174 _run_func = &GCIm2ColKernel::run_reduced;
175 }
176
177 // Create kernel
178 _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel(kernel_name, build_opts));
179
180 // Configure kernel window
181 Window win = calculate_max_window(*input->info(), Steps(_num_elems_processed_per_iteration));
182
183 if(input->info()->data_type() == DataType::F16)
184 {
185 // Calculate input right and bottom border
186 const int input_width = input->info()->dimension(0);
187 const int input_height = input->info()->dimension(1);
188 int input_total_width = input->info()->padding().left + input_width + input->info()->padding().right;
189 int input_padding_right = ceil_to_multiple(input_total_width, _num_elems_processed_per_iteration) - input_total_width;
190 input_total_width = input_width + input_padding_right + input->info()->padding().right;
191 AccessWindowStatic input_access(input->info(), 0, 0, input_total_width, input_height);
192
193 // Calculate output right and bottom border
194 const int output_width = output->info()->dimension(0);
195 const int output_height = output->info()->dimension(1);
196 const int output_padding_right = ceil_to_multiple(output_width, _num_elems_processed_per_iteration) - output_width;
197 AccessWindowStatic output_access(output->info(), 0, 0, output_width + output_padding_right, output_height);
198
199 update_window_and_padding(win, input_access, output_access);
200 }
201
202 output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
203
204 if(!run_img2col_reduced)
205 {
206 // set the Z dimension's step same size as the whole dimension so that one can't split across the Z dimension
207 win.set_dimension_step(Window::DimZ, win[Window::DimZ].end() - win[Window::DimZ].start());
208 }
209
210 IGCKernel::configure(win);
211 }
212
validate(const ITensorInfo * input,const ITensorInfo * output,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation)213 Status GCIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation)
214 {
215 ARM_COMPUTE_UNUSED(kernel_dims);
216 ARM_COMPUTE_UNUSED(conv_info);
217 ARM_COMPUTE_UNUSED(has_bias);
218 ARM_COMPUTE_UNUSED(dilation);
219 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
220 return Status{};
221 }
222
run(const Window & window)223 void GCIm2ColKernel::run(const Window &window)
224 {
225 ARM_COMPUTE_ERROR_ON(_run_func == nullptr);
226 (this->*_run_func)(window);
227 }
228
run_generic(const Window & window)229 void GCIm2ColKernel::run_generic(const Window &window)
230 {
231 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
232 ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IGCKernel::window(), window);
233
234 // Get initial windows
235 Window window_collapsed = window.collapse_if_possible(IGCKernel::window(), Window::DimZ);
236
237 // Change the Z dimension's step back to 1
238 window_collapsed.set_dimension_step(Window::DimZ, 1);
239
240 Window slice = window_collapsed.first_slice_window_3D();
241 Window slice_in = window_collapsed.first_slice_window_3D();
242 Window slice_out = window_collapsed.first_slice_window_3D();
243
244 // Setup slice
245 slice.set(Window::DimX, Window::Dimension(0, static_cast<int>(_convolved_dims.first), 1));
246 slice.set(Window::DimY, Window::Dimension(0, static_cast<int>(_convolved_dims.second), 1));
247
248 // Setup output slice
249 slice_out.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _num_elems_processed_per_iteration));
250 slice_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 1));
251 slice_out.set(Window::DimZ, Window::Dimension(0, 1, 1));
252
253 // we need top/left pad to be included in valid region
254 if(_input->info()->data_type() == DataType::F16)
255 {
256 (dynamic_cast<TensorInfo *>(_input->info()))->init(_input->info()->tensor_shape(), _input->info()->num_channels(), _input->info()->data_type(), _input->info()->strides_in_bytes(), 0,
257 _input->info()->total_size());
258 }
259
260 _kernel.use();
261
262 do
263 {
264 unsigned int idx = 0;
265 add_3D_tensor_argument(idx, _input, 1, slice_in);
266 add_2D_tensor_argument(idx, _output, 2, slice_out);
267 _kernel.set_argument(idx++, static_cast<unsigned int>(_input->info()->strides_in_bytes()[3]));
268 _kernel.set_argument(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[3]));
269 _kernel.update_shader_params();
270
271 enqueue(*this, slice);
272 }
273 while(window_collapsed.slide_window_slice_3D(slice) && window_collapsed.slide_window_slice_3D(slice_out) && window_collapsed.slide_window_slice_3D(slice_in));
274 }
275
run_reduced(const Window & window)276 void GCIm2ColKernel::run_reduced(const Window &window)
277 {
278 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
279 ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IGCKernel::window(), window);
280
281 Window out_window;
282 out_window.use_tensor_dimensions(_output->info()->tensor_shape());
283
284 Window out_slice = out_window.first_slice_window_1D();
285 Window in_slice = window.first_slice_window_3D();
286
287 _kernel.use();
288
289 // Run kernel
290 do
291 {
292 // Set arguments
293 unsigned int idx = 0;
294
295 add_3D_tensor_argument(idx, _input, 1, in_slice);
296 add_1D_tensor_argument(idx, _output, 2, out_slice);
297 _kernel.set_argument(idx++, _input->info()->dimension(0));
298 _kernel.set_argument(idx++, _input->info()->dimension(1));
299 _kernel.update_shader_params();
300
301 enqueue(*this, in_slice);
302 }
303 while(window.slide_window_slice_3D(in_slice) && out_window.slide_window_slice_1D(out_slice));
304 }
305