Compute Library
 20.08
GCDepthwiseConvolutionLayer3x3Kernel.cpp
Go to the documentation of this file.
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  */
25 
27 #include "arm_compute/core/Error.h"
34 #include "arm_compute/core/Types.h"
35 #include "arm_compute/core/Utils.h"
37 #include "support/StringSupport.h"
38 
39 using namespace arm_compute;
41 
43  : _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))
44 {
45 }
46 
48 {
49  return _border_size;
50 }
51 
53  unsigned int depth_multiplier)
54 {
57  ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
58 
59  if(biases != nullptr)
60  {
62  ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(2));
63  ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
64  }
65 
66  // Get convolved dimensions
68 
69  // Output auto inizialitation if not yet initialized
70  auto_init_if_empty(*output->info(),
72  1,
73  input->info()->data_type());
74 
76  ARM_COMPUTE_ERROR_ON(output->info()->dimension(2) != weights->info()->dimension(2));
77 
78  _input = input;
79  _output = output;
80  _weights = weights;
81  _biases = biases;
82  _conv_stride_x = conv_info.stride().first;
83  _conv_stride_y = conv_info.stride().second;
84  _conv_pad_left = conv_info.pad_left();
85  _conv_pad_top = conv_info.pad_top();
86  _border_size = BorderSize(_conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), _conv_pad_left);
87 
88  // Set build options
89  ARM_COMPUTE_ERROR_ON(_conv_stride_x < 1 || _conv_stride_x > 3);
90  std::set<std::string> options;
91 
92  options.emplace("#define DEPTH_MULTIPLIER " + support::cpp11::to_string(depth_multiplier));
93  options.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(_lws[0]));
94  options.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(_lws[1]));
95  options.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(_lws[2]));
96  options.emplace("#define STRIDE_X " + support::cpp11::to_string(_conv_stride_x));
97  options.emplace("#define STRIDE_Y " + support::cpp11::to_string(_conv_stride_y));
98 
99  std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
100  options.emplace(("#define " + dt_name));
101 
102  unsigned int num_elems_read_per_iteration_x = 8;
103  unsigned int num_elems_read_per_iteration_y = 1;
104  unsigned int num_elems_written_per_iteration_x = 4;
105  unsigned int num_elems_written_per_iteration_y = 1;
106  unsigned int num_elems_written_per_iteration_z = 1;
107 
108  if((_conv_stride_x == 1) && (_conv_stride_y == 1))
109  {
110  switch(input->info()->data_type())
111  {
112 #define PROCESS_4X_3Y_1Z
113 
114  case DataType::F16:
115 #if defined(PROCESS_4X_3Y_1Z)
116  options.emplace("#define PROCESS_4X_3Y_1Z");
117  num_elems_read_per_iteration_y = 5;
118  num_elems_written_per_iteration_y = 3;
119 #endif /* PROCESS_4X_3Y_1Z */
120 #undef PROCESS_4X_3Y_1Z
121  break;
122 
123  default:
124  ARM_COMPUTE_ERROR("Current data type is not supported");
125  break;
126  }
127  }
128  else
129  {
130  switch(input->info()->data_type())
131  {
132  case DataType::F16:
133  options.emplace("#define PROCESS_4X_1Y_1Z");
134  break;
135 
136  default:
137  ARM_COMPUTE_ERROR("Current data type is not supported");
138  break;
139  }
140  }
141 
142  if(_biases != nullptr)
143  {
144  options.emplace("#define BIAS");
145  }
146 
147  // Create kernel
148  std::string kernel_name = "depthwise_convolution_3x3";
149  _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel(kernel_name, options));
150 
151  // Calculate output right and bottom border
152  const int output_width = output->info()->dimension(0);
153  const int output_height = output->info()->dimension(1);
154  const int output_padding_right = ceil_to_multiple(output_width, num_elems_written_per_iteration_x * _lws[0]) - output_width;
155  const int output_padding_bottom = ceil_to_multiple(output_height, num_elems_written_per_iteration_y * _lws[1]) - output_height;
156 
157  // Calculate input right and bottom border
158  const int input_width = input->info()->dimension(0);
159  const int input_height = input->info()->dimension(1);
160 
161  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));
162  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));
163 
164  const int input_padding_right = ceil_to_multiple(input_total_width, num_elems_read_per_iteration_x * _lws[0]) - input_width - _conv_pad_left;
165  const int input_padding_bottom = ceil_to_multiple(input_total_height, num_elems_read_per_iteration_y * _lws[1]) - input_height - _conv_pad_top;
166 
167  BorderSize border = BorderSize(0, output_padding_right, output_padding_bottom, 0);
168 
169  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);
170 
171  AccessWindowStatic input_access(input->info(), -_conv_pad_left, -_conv_pad_top, input_width + input_padding_right, input_height + input_padding_bottom);
172  AccessWindowStatic weights_access = AccessWindowStatic(nullptr, 0, 0, 0, 0);
173  AccessWindowStatic bias_access = AccessWindowStatic(nullptr, 0, 0, 0, 1);
174 
175  switch(weights->info()->data_type())
176  {
177  case DataType::F16:
178  weights_access = AccessWindowStatic(weights->info(), 0, 0, 4, 3);
179  if(_biases != nullptr)
180  {
181  bias_access = AccessWindowStatic(_biases->info(), 0, 0, _biases->info()->dimension(0) + 1, 1);
182  }
183  break;
184 
185  default:
186  ARM_COMPUTE_ERROR("Current data type is not supported");
187  break;
188  }
189 
190  AccessWindowStatic output_access(output->info(), 0, 0, output_width + output_padding_right, output_height + output_padding_bottom);
191 
192  if(_biases != nullptr)
193  {
194  update_window_and_padding(win, input_access, weights_access, bias_access, output_access);
195  }
196  else
197  {
198  update_window_and_padding(win, input_access, weights_access, output_access);
199  }
200 
201  output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
202 
203  IGCKernel::configure(win);
204 }
205 
207 {
210 
211  _kernel.use();
212 
213  _output->set_needs_shifting(true);
214 
215  // Create input window and adjust
216  Window win_in = window;
217  win_in.adjust(Window::DimX, -_conv_pad_left, true);
218  win_in.adjust(Window::DimY, -_conv_pad_top, true);
219  win_in.set_dimension_step(Window::DimX, window.x().step() * _conv_stride_x);
220  win_in.set_dimension_step(Window::DimY, window.y().step() * _conv_stride_y);
221 
222  Window slice_in = win_in.first_slice_window_3D();
223  Window slice_out = window.first_slice_window_3D();
224  Window slice_weights = window.first_slice_window_3D();
225  slice_weights.set_dimension_step(Window::DimX, 0);
226  slice_weights.set_dimension_step(Window::DimY, 0);
227 
228  // Set biases
229  if(_biases != nullptr)
230  {
231  unsigned int idx = 3 * num_arguments_per_3D_tensor();
232  Window slice_biases;
233  slice_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
234  add_1D_tensor_argument(idx, _biases, 4, slice_biases);
235  }
236 
237  slice_out.shift(Window::DimX, -(_output->info()->padding()).left);
238 
239  do
240  {
241  unsigned int idx = 0;
242  add_3D_tensor_argument(idx, _input, 1, slice_in);
243  add_3D_tensor_argument(idx, _output, 2, slice_out);
244  add_3D_tensor_argument(idx, _weights, 3, slice_weights);
245 
246  _kernel.update_shader_params();
247  enqueue(*this, slice_out, _lws);
248  }
249  while(window.slide_window_slice_3D(slice_out) && win_in.slide_window_slice_3D(slice_in));
250 }
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
Shape of a tensor.
Definition: TensorShape.h:39
void add_3D_tensor_argument(unsigned int &idx, const IGCTensor *tensor, const unsigned int binding_point, const Window &window)
Add the passed 3D tensor's parameters to the object's kernel's arguments starting from the index idx.
Definition: IGCKernel.cpp:132
TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const ITensorInfo &weights, PadStrideInfo conv_info, unsigned int depth_multiplier, const Size2D &dilation=Size2D(1U, 1U))
Calculate the depthwise convolution output shape of a tensor.
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
void shift(size_t dimension, int shift_value)
Shift the values of a given dimension by the given shift_value.
Definition: Window.inl:133
#define ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:543
Container for 2D border size.
Definition: Types.h:272
void enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Window &window, const cl::NDRange &lws_hint=CLKernelLibrary::get().default_ndrange(), bool use_dummy_work_items=false)
Add the kernel to the command queue with the given window.
Definition: ICLKernel.cpp:39
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
constexpr int step() const
Return the step of the dimension.
Definition: Window.h:102
std::string to_string(T &&value)
Convert integer and float values to string.
virtual DataType data_type() const =0
Data type used for each element of the tensor.
1 channel, 1 F32 per channel
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
Interface for GLES Compute tensor.
Definition: IGCTensor.h:35
unsigned int num_arguments_per_3D_tensor() const
Returns the number of arguments enqueued per 3D tensor object.
Definition: IGCKernel.cpp:147
void use_tensor_dimensions(const TensorShape &shape, size_t first_dimension=Window::DimX)
Use the tensor's dimensions to fill the window dimensions.
Definition: Window.inl:276
Copyright (c) 2017-2020 Arm Limited.
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
Definition: Helpers.inl:207
1 channel, 1 F16 per channel
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: Tensor.cpp:33
Implementation of a static rectangular access pattern.
Window calculate_max_enlarged_window(const ValidRegion &valid_region, const Steps &steps=Steps(), BorderSize border_size=BorderSize())
Calculate the maximum window for a given tensor shape and border setting.
Definition: Helpers.cpp:82
cl::Kernel create_kernel(const CLCompileContext &ctx, const std::string &kernel_name, const std::set< std::string > &build_opts=std::set< std::string >())
Creates an opencl kernel using a compile context.
Definition: CLHelpers.cpp:403
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
bool update_window_and_padding(Window &win, Ts &&... patterns)
Update window and padding size for each of the access patterns.
Definition: Helpers.h:437
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
auto ceil_to_multiple(S value, T divisor) -> decltype(((value+divisor - 1)/divisor) *divisor)
Computes the smallest number larger or equal to value that is a multiple of divisor.
Definition: Utils.h:67
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
void set_needs_shifting(bool needs_shifting)
Set the flag indicating whether or not a tensor needs shifting.
Definition: IGCTensor.cpp:61
Coordinates of an item.
Definition: Coordinates.h:37
std::string kernel_name
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
#define ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(...)
Definition: Validate.h:286
Padding and stride information class.
Definition: Types.h:689
virtual PaddingSize padding() const =0
Padding of tensor.
bool slide_window_slice_3D(Window &slice) const
Slide the passed 3D window slice.
Definition: Window.h:333
#define ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:790
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
static GCKernelLibrary & get()
Get the static instance of GCKernelLibrary.
void set_dimension_step(size_t dimension, int step)
Set the step of a given dimension.
Definition: Window.inl:167
constexpr const Dimension & y() const
Alias to access the second dimension of the window.
Definition: Window.h:152
void configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier=1)
Initialize the function's source, destination, conv and border_size.
void run(const Window &window) override
Enqueue the OpenGL ES shader to process the given window.
BorderSize border_size() const override
The size of the border for that kernel.
Container for valid region of a window.
Definition: Types.h:187
void adjust(size_t dimension, int adjust_value, bool is_at_start)
Adjust the start or end of a given dimension by the given value.
Definition: Window.inl:140
Window first_slice_window_3D() const
First 3D slice of the window.
Definition: Window.h:289
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:941
void add_1D_tensor_argument(unsigned int &idx, const IGCTensor *tensor, const unsigned int binding_point, const Window &window)
Add the passed 1D tensor's parameters to the object's kernel's arguments starting from the index idx.
Definition: IGCKernel.cpp:122
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:143