Compute Library
 20.08
CLWinogradConvolutionLayer.cpp
Go to the documentation of this file.
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  */
25 
27 #include "arm_compute/core/Utils.h"
31 
32 using namespace arm_compute;
33 
34 namespace
35 {
36 Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataLayout data_layout)
37 {
38  Size2D output_tile = Size2D{};
39 
40  const unsigned int kernel_max_dim = std::max(kernel_dims.width, kernel_dims.height);
41 
42  // Check if the input spatial dimensions are smaller than 4
43  const bool is_input_lt4_nchw = (input_dims.width <= 4 && input_dims.height <= 4) && (data_layout == DataLayout::NCHW);
44 
45  if(kernel_max_dim == 3U)
46  {
47  if(kernel_dims == Size2D(3U, 3U))
48  {
49  output_tile = is_input_lt4_nchw ? Size2D(2U, 2U) : Size2D(4U, 4U);
50  }
51  else if(kernel_dims == Size2D(3U, 1U))
52  {
53  output_tile = is_input_lt4_nchw ? Size2D(2U, 1U) : Size2D(4U, 1U);
54  }
55  else
56  {
57  output_tile = is_input_lt4_nchw ? Size2D(1U, 2U) : Size2D(1U, 4U);
58  }
59  }
60  else if(kernel_max_dim == 5U)
61  {
62  output_tile = Size2D(kernel_dims.width == 1 ? 1U : 4U,
63  kernel_dims.height == 1 ? 1U : 4U);
64  }
65  else if(kernel_max_dim == 7U)
66  {
67  output_tile = Size2D(kernel_dims.width == 1 ? 1U : 2U,
68  kernel_dims.height == 1 ? 1U : 2U);
69  }
70 
71  return output_tile;
72 }
73 
74 bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size)
75 {
76  // Check if we want to configure a Winograd configuration which requires fast math
77  using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>;
78 
79  std::vector<WinogradConfiguration> fast_math_winograd =
80  {
81  WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)),
82  WinogradConfiguration(std::pair<int, int>(2, 2), std::pair<int, int>(7, 7))
83  };
84 
85  auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height),
86  std::pair<int, int>(kernel_size.width, kernel_size.height));
87 
88  return std::find(fast_math_winograd.begin(), fast_math_winograd.end(), p) != fast_math_winograd.end();
89 }
90 } // namespace
91 
92 CLWinogradConvolutionLayer::CLWinogradConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
93  : _memory_group(memory_manager), _batched_mm(memory_manager), _input_transform(), _filter_transform(), _output_transform(), _input0(), _input1(), _batched_mm_output(), _original_weights(nullptr),
94  _is_prepared(false)
95 {
96 }
97 
99  bool enable_fast_math)
100 {
101  configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, act_info, enable_fast_math);
102 }
103 
104 void CLWinogradConvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
105  const PadStrideInfo &conv_info,
106  const ActivationLayerInfo &act_info, bool enable_fast_math)
107 {
108  // Get indices for the width and height
109  const size_t idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
110  const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
111 
112  // Input shape, kernel size and output tile
113  const Size2D input_dims = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]);
114  const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]);
115  const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->info()->data_layout());
116 
117  // Check if the Winograd configuration requires fast math
118  if(!enable_fast_math)
119  {
120  ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); //disable winograd for fp16 if fast math is false.
121  ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true");
122  }
123  const WinogradInfo winograd_info = WinogradInfo(output_tile,
124  kernel_size,
125  input_dims,
126  conv_info,
127  input->info()->data_layout());
128 
129  _is_prepared = false;
130  _original_weights = weights;
131 
132  // Manage intermediate tensors
133  _memory_group.manage(&_input0);
134  _memory_group.manage(&_batched_mm_output);
135 
136  // Do not manage _input1 as it contains the weights
137 
138  // Configure input transform
139  _input_transform.configure(compile_context, input, &_input0, winograd_info);
140 
141  // Configure filter transform
142  _filter_transform.configure(compile_context, weights, &_input1, winograd_info);
143 
144  // Configure batched matrix multiply
145  _batched_mm.configure(compile_context, &_input0, &_input1, nullptr, &_batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, 0, false, false,
147  (input->info()->data_type() == DataType::F16)));
148 
149  // Configure output transform
150  _output_transform.configure(compile_context, &_batched_mm_output, biases, output, winograd_info, act_info);
151 
152  // Allocate temporary tensors
153  _input0.allocator()->allocate();
154  _batched_mm_output.allocator()->allocate();
155 }
156 
158  const ActivationLayerInfo &act_info, bool enable_fast_math)
159 {
160  // Get indeces for the width and height
161  const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
162  const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
163 
164  // Input shape, kernel size and output tile
165  const Size2D input_dims = Size2D(input->tensor_shape()[idx_width], input->tensor_shape()[idx_height]);
166  const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]);
167  const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->data_layout());
168 
169  ARM_COMPUTE_RETURN_ERROR_ON_MSG(((conv_info.pad_left() > (kernel_size.x() / 2u)) || (conv_info.pad_right() > (kernel_size.x() / 2u))), "Winograd only supports padding up to half kernel size");
170  ARM_COMPUTE_RETURN_ERROR_ON_MSG(((conv_info.pad_top() > (kernel_size.y() / 2u)) || (conv_info.pad_bottom() > (kernel_size.y() / 2u))), "Winograd only supports padding up to half kernel size");
171 
172  // Check if the Winograd configuration requires fast math
173  if(!enable_fast_math)
174  {
175  ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); //disable winograd for fp16 if fast math is false.
176  ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true");
177  }
178 
179  const WinogradInfo winograd_info = WinogradInfo(output_tile,
180  kernel_size,
181  input_dims,
182  conv_info,
183  input->data_layout());
184 
185  // Validate input transform
187  const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape);
189 
190  // Validate filter transform
192  const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape);
194 
195  // Validate batched matrix multiply
196  TensorShape batched_mm_output_shape = input0.tensor_shape();
197  batched_mm_output_shape[0] = input1.tensor_shape()[0];
198  const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape);
199  ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, 0, false, false,
200  GEMMLowpOutputStageInfo(), (input->data_type() == DataType::F16))));
201 
202  // Configure output transform
203  ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradOutputTransformKernel::validate(&batched_mm_output, biases, output, winograd_info, act_info));
204 
205  return Status{};
206 }
207 
209 {
210  prepare();
211 
212  MemoryGroupResourceScope scope_mg(_memory_group);
213 
214  // Run input transform
215  _input_transform.run();
216 
217  // Run batched matrix multiplication
218  _batched_mm.run();
219 
220  // Run output transform
221  CLScheduler::get().enqueue(_output_transform);
222 }
223 
225 {
226  if(!_is_prepared)
227  {
228  // Run filter transform and mark original weights as unused
229  _input1.allocator()->allocate();
230  CLScheduler::get().enqueue(_filter_transform, false);
231  _original_weights->mark_as_unused();
232 
233  // Prepare GEMM and release reshaped weights if marked unused by CLGEMM
234  _batched_mm.prepare();
235  if(!_input1.is_used())
236  {
237  _input1.allocator()->free();
238  }
239 
240  CLScheduler::get().queue().finish();
241  _is_prepared = true;
242  }
243 }
Shape of a tensor.
Definition: TensorShape.h:39
const DataLayout data_layout
Definition: Im2Col.cpp:146
void prepare() override
Prepare the function for executing.
Definition: CLGEMM.cpp:683
TensorShape compute_winograd_input_transform_shape(const ITensorInfo &input, const WinogradInfo &winograd_info)
Calculate the winograd input transform shape.
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
Static function to check if given info will lead to a valid configuration of CLWinogradFilterTransfor...
static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Static function to check if given info will lead to a valid configuration of CLWinogradOutputTransfor...
std::unique_ptr< ITensorInfo > clone() const override
Provide a clone of the current object of class T.
Definition: TensorInfo.cpp:314
void run() override
Run the kernels contained in the function.
Definition: CLGEMM.cpp:602
CLWinogradConvolutionLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
static CLScheduler & get()
Access the scheduler singleton.
Definition: CLScheduler.cpp:99
Winograd information.
Definition: Types.h:2111
void run() override
Run the kernels contained in the function.
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
bool is_used() const
Flags if the tensor is used or not.
Definition: ITensor.cpp:163
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
1 channel, 1 F32 per channel
void prepare() override
Prepare the function for executing.
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
Store the tensor's metadata.
Definition: ITensorInfo.h:40
CLTensorAllocator * allocator()
Return a pointer to the tensor's allocator.
Definition: CLTensor.cpp:61
void configure(const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info)
Set the input and output tensor.
Status class.
Definition: Error.h:52
Activation Layer Information class.
Definition: Types.h:1517
Copyright (c) 2017-2020 Arm Limited.
size_t height
Height of the image region or rectangle.
Definition: Size2D.h:90
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
void mark_as_unused() const
Marks a tensor as unused.
Definition: ITensor.cpp:168
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
void run() override final
Run the kernels contained in the function.
void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Set the input and output tensor.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
#define ARM_COMPUTE_ERROR_ON_MSG(cond, msg)
Definition: Error.h:456
GEMMLowp output stage info.
Definition: Types.h:1881
Padding and stride information class.
Definition: Types.h:689
TensorShape compute_winograd_filter_transform_shape(const ITensorInfo &input, const WinogradInfo &winograd_info)
Calculate the winograd filter transform shape.
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false)
Static function to check if given info will lead to a valid configuration of CLWinogradConvolutionLay...
cl::CommandQueue & queue()
Accessor for the associated CL command queue.
Definition: CLScheduler.cpp:41
void enqueue(ICLKernel &kernel, bool flush=true)
Schedule the execution of the passed kernel if possible.
Num samples, channels, height, width.
CLCompileContext class.
#define ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:790
void allocate() override
Allocate size specified by TensorInfo of OpenCL memory.
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
size_t width
Width of the image region or rectangle.
Definition: Size2D.h:89
void configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Initialise the kernel's inputs and output.
Definition: CLGEMM.cpp:497
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
void free() override
Free allocated OpenCL memory.
void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false)
Set the input and output tensors.
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of CLGEMM.
Definition: CLGEMM.cpp:556
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
Store the tensor's metadata.
Definition: TensorInfo.h:45
GEMM information class.
Definition: Types.h:1932
void configure(ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info)
Set the input and output tensors.
size_t get_data_layout_dimension_index(const DataLayout data_layout, const DataLayoutDimension data_layout_dimension)
Get the index of the given dimension.
Definition: Helpers.inl:332
const TensorShape & tensor_shape() const override
Size for each dimension of the tensor.
Definition: TensorInfo.h:261
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
Static function to check if given info will lead to a valid configuration of CLWinogradInputTransform...
DataLayout
[DataLayout enum definition]
Definition: Types.h:120