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 #include "arm_compute/core/GLES_COMPUTE/kernels/GCDepthwiseConvolutionLayer3x3Kernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
28 #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
29 #include "arm_compute/core/GLES_COMPUTE/IGCKernel.h"
30 #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
31 #include "arm_compute/core/Helpers.h"
32 #include "arm_compute/core/TensorInfo.h"
33 #include "arm_compute/core/Types.h"
34 #include "arm_compute/core/Utils.h"
35 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
36 #include "src/core/AccessWindowStatic.h"
37 #include "src/core/helpers/AutoConfiguration.h"
38 #include "src/core/helpers/WindowHelpers.h"
39 #include "support/StringSupport.h"
40
41 using namespace arm_compute;
42 using namespace arm_compute::misc::shape_calculator;
43
GCDepthwiseConvolutionLayer3x3Kernel()44 GCDepthwiseConvolutionLayer3x3Kernel::GCDepthwiseConvolutionLayer3x3Kernel()
45 : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_x(0), _conv_stride_y(0), _conv_pad_left(0), _conv_pad_top(0), _lws(gles::NDRange(1U, 1U, 1U))
46 {
47 }
48
border_size() const49 BorderSize GCDepthwiseConvolutionLayer3x3Kernel::border_size() const
50 {
51 return _border_size;
52 }
53
configure(const IGCTensor * input,const IGCTensor * weights,const IGCTensor * biases,IGCTensor * output,const PadStrideInfo & conv_info,unsigned int depth_multiplier)54 void GCDepthwiseConvolutionLayer3x3Kernel::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info,
55 unsigned int depth_multiplier)
56 {
57 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16);
58 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
59 ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
60
61 if(biases != nullptr)
62 {
63 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
64 ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(2));
65 ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
66 }
67
68 // Get convolved dimensions
69 const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier);
70
71 // Output auto inizialitation if not yet initialized
72 auto_init_if_empty(*output->info(),
73 output_shape,
74 1,
75 input->info()->data_type());
76
77 ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
78 ARM_COMPUTE_ERROR_ON(output->info()->dimension(2) != weights->info()->dimension(2));
79
80 _input = input;
81 _output = output;
82 _weights = weights;
83 _biases = biases;
84 _conv_stride_x = conv_info.stride().first;
85 _conv_stride_y = conv_info.stride().second;
86 _conv_pad_left = conv_info.pad_left();
87 _conv_pad_top = conv_info.pad_top();
88 _border_size = BorderSize(_conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), _conv_pad_left);
89
90 // Set build options
91 ARM_COMPUTE_ERROR_ON(_conv_stride_x < 1 || _conv_stride_x > 3);
92 std::set<std::string> options;
93
94 options.emplace("#define DEPTH_MULTIPLIER " + support::cpp11::to_string(depth_multiplier));
95 options.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(_lws[0]));
96 options.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(_lws[1]));
97 options.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(_lws[2]));
98 options.emplace("#define STRIDE_X " + support::cpp11::to_string(_conv_stride_x));
99 options.emplace("#define STRIDE_Y " + support::cpp11::to_string(_conv_stride_y));
100
101 std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
102 options.emplace(("#define " + dt_name));
103
104 unsigned int num_elems_read_per_iteration_x = 8;
105 unsigned int num_elems_read_per_iteration_y = 1;
106 unsigned int num_elems_written_per_iteration_x = 4;
107 unsigned int num_elems_written_per_iteration_y = 1;
108 unsigned int num_elems_written_per_iteration_z = 1;
109
110 if((_conv_stride_x == 1) && (_conv_stride_y == 1))
111 {
112 switch(input->info()->data_type())
113 {
114 #define PROCESS_4X_3Y_1Z
115
116 case DataType::F16:
117 #if defined(PROCESS_4X_3Y_1Z)
118 options.emplace("#define PROCESS_4X_3Y_1Z");
119 num_elems_read_per_iteration_y = 5;
120 num_elems_written_per_iteration_y = 3;
121 #endif /* PROCESS_4X_3Y_1Z */
122 #undef PROCESS_4X_3Y_1Z
123 break;
124
125 default:
126 ARM_COMPUTE_ERROR("Current data type is not supported");
127 break;
128 }
129 }
130 else
131 {
132 switch(input->info()->data_type())
133 {
134 case DataType::F16:
135 options.emplace("#define PROCESS_4X_1Y_1Z");
136 break;
137
138 default:
139 ARM_COMPUTE_ERROR("Current data type is not supported");
140 break;
141 }
142 }
143
144 if(_biases != nullptr)
145 {
146 options.emplace("#define BIAS");
147 }
148
149 // Create kernel
150 std::string kernel_name = "depthwise_convolution_3x3";
151 _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel(kernel_name, options));
152
153 // Calculate output right and bottom border
154 const int output_width = output->info()->dimension(0);
155 const int output_height = output->info()->dimension(1);
156 const int output_padding_right = ceil_to_multiple(output_width, num_elems_written_per_iteration_x * _lws[0]) - output_width;
157 const int output_padding_bottom = ceil_to_multiple(output_height, num_elems_written_per_iteration_y * _lws[1]) - output_height;
158
159 // Calculate input right and bottom border
160 const int input_width = input->info()->dimension(0);
161 const int input_height = input->info()->dimension(1);
162
163 const int input_total_width = std::max(int(input->info()->padding().left), int(_conv_pad_left)) + input_width + std::max(int(input->info()->padding().right), int(_conv_pad_left));
164 const int input_total_height = std::max(int(input->info()->padding().top), int(_conv_pad_top)) + input_height + std::max(int(input->info()->padding().bottom), int(_conv_pad_top));
165
166 const int input_padding_right = ceil_to_multiple(input_total_width, num_elems_read_per_iteration_x * _lws[0]) - input_width - _conv_pad_left;
167 const int input_padding_bottom = ceil_to_multiple(input_total_height, num_elems_read_per_iteration_y * _lws[1]) - input_height - _conv_pad_top;
168
169 BorderSize border = BorderSize(0, output_padding_right, output_padding_bottom, 0);
170
171 Window win = calculate_max_enlarged_window(*output->info(), Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y, num_elems_written_per_iteration_z), border);
172
173 AccessWindowStatic input_access(input->info(), -_conv_pad_left, -_conv_pad_top, input_width + input_padding_right, input_height + input_padding_bottom);
174 AccessWindowStatic weights_access = AccessWindowStatic(nullptr, 0, 0, 0, 0);
175 AccessWindowStatic bias_access = AccessWindowStatic(nullptr, 0, 0, 0, 1);
176
177 switch(weights->info()->data_type())
178 {
179 case DataType::F16:
180 weights_access = AccessWindowStatic(weights->info(), 0, 0, 4, 3);
181 if(_biases != nullptr)
182 {
183 bias_access = AccessWindowStatic(_biases->info(), 0, 0, _biases->info()->dimension(0) + 1, 1);
184 }
185 break;
186
187 default:
188 ARM_COMPUTE_ERROR("Current data type is not supported");
189 break;
190 }
191
192 AccessWindowStatic output_access(output->info(), 0, 0, output_width + output_padding_right, output_height + output_padding_bottom);
193
194 if(_biases != nullptr)
195 {
196 update_window_and_padding(win, input_access, weights_access, bias_access, output_access);
197 }
198 else
199 {
200 update_window_and_padding(win, input_access, weights_access, output_access);
201 }
202
203 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
204
205 IGCKernel::configure(win);
206 }
207
run(const Window & window)208 void GCDepthwiseConvolutionLayer3x3Kernel::run(const Window &window)
209 {
210 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
211 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
212
213 _kernel.use();
214
215 _output->set_needs_shifting(true);
216
217 // Create input window and adjust
218 Window win_in = window;
219 win_in.adjust(Window::DimX, -_conv_pad_left, true);
220 win_in.adjust(Window::DimY, -_conv_pad_top, true);
221 win_in.set_dimension_step(Window::DimX, window.x().step() * _conv_stride_x);
222 win_in.set_dimension_step(Window::DimY, window.y().step() * _conv_stride_y);
223
224 Window slice_in = win_in.first_slice_window_3D();
225 Window slice_out = window.first_slice_window_3D();
226 Window slice_weights = window.first_slice_window_3D();
227 slice_weights.set_dimension_step(Window::DimX, 0);
228 slice_weights.set_dimension_step(Window::DimY, 0);
229
230 // Set biases
231 if(_biases != nullptr)
232 {
233 unsigned int idx = 3 * num_arguments_per_3D_tensor();
234 Window slice_biases;
235 slice_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
236 add_1D_tensor_argument(idx, _biases, 4, slice_biases);
237 }
238
239 slice_out.shift(Window::DimX, -(_output->info()->padding()).left);
240
241 do
242 {
243 unsigned int idx = 0;
244 add_3D_tensor_argument(idx, _input, 1, slice_in);
245 add_3D_tensor_argument(idx, _output, 2, slice_out);
246 add_3D_tensor_argument(idx, _weights, 3, slice_weights);
247
248 _kernel.update_shader_params();
249 enqueue(*this, slice_out, _lws);
250 }
251 while(window.slide_window_slice_3D(slice_out) && win_in.slide_window_slice_3D(slice_in));
252 }
253