Compute Library
 20.08
CLFFTConvolutionLayer.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2020 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"
33 
34 namespace arm_compute
35 {
36 namespace
37 {
38 int pad_decomposable(int N)
39 {
40  const auto supported_radix = CLFFTRadixStageKernel::supported_radix();
41 
42  int pad = 0;
43  bool is_decomposed = false;
44  while(!is_decomposed)
45  {
46  const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N++, supported_radix);
47  is_decomposed = !decomposed_vector.empty();
48  if(!is_decomposed)
49  {
50  ++pad;
51  }
52  }
53  return pad;
54 }
55 } // namespace
56 CLFFTConvolutionLayer::CLFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
57  : _memory_group(memory_manager),
58  _flip_weights_func(),
59  _permute_input_func(),
60  _permute_output_func(),
61  _permute_weights_func(),
62  _permute_bias_func(),
63  _pad_input_func(),
64  _pad_weights_func(),
65  _transform_input_func(memory_manager),
66  _transform_weights_func(),
67  _itransform_output_func(memory_manager),
68  _prod_func(),
69  _reduce_func(),
70  _extract_output_func(),
71  _bias_add_func(),
72  _activation_layer_func(),
73  _permuted_input(),
74  _permuted_weights(),
75  _permuted_bias(),
76  _permuted_output(),
77  _padded_input(),
78  _padded_weights(),
79  _flip_axis(),
80  _flipped_weights(),
81  _transformed_input(),
82  _transformed_weights(),
83  _input_weights_product(),
84  _output_product(),
85  _output_reduced(),
86  _itransformed_output(),
87  _reshaped_output(),
88  _bias_output(),
89  _original_weights(nullptr),
90  _original_bias(nullptr),
91  _is_activationlayer_enabled(false),
92  _needs_permute(false),
93  _has_bias(false),
94  _is_prepared(false)
95 {
96 }
97 
99  const ActivationLayerInfo &act_info)
100 {
101  configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, act_info);
102 }
103 
104 void CLFFTConvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
105  const ActivationLayerInfo &act_info)
106 {
107  _original_weights = weights;
108  _original_bias = biases;
109 
110  // Flat if bias addition is required
111  _has_bias = biases != nullptr;
112 
113  // Get indices for the width and height
114  const size_t idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
115  const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
116 
117  // Input shape, kernel size and output tile
118  const Size2D input_dims = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]);
119  const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]);
120  const Size2D pad_valid = Size2D(pad_decomposable(input_dims.x() + kernel_size.x() - 1),
121  pad_decomposable(input_dims.y() + kernel_size.y() - 1));
122  // Tensors to use
123  ICLTensor *input_to_use = input;
124  const ICLTensor *weights_to_use = weights;
125  ICLTensor *output_to_use = _has_bias ? &_bias_output : output;
126 
127  // Permute bias
128  if(biases != nullptr)
129  {
130  _permute_bias_func.configure(compile_context, biases, &_permuted_bias, PermutationVector(1U, 2U, 0U));
131  _permuted_bias.info()->set_data_layout(DataLayout::NCHW);
132  }
133 
134  // Permute input if needed
135  _needs_permute = input->info()->data_layout() == DataLayout::NHWC;
136  if(_needs_permute)
137  {
138  _memory_group.manage(&_permuted_input);
139  // Configure the function to transform the input tensor from NHWC -> NCHW
140  _permute_input_func.configure(compile_context, input, &_permuted_input, PermutationVector(1U, 2U, 0U));
141  _permuted_input.info()->set_data_layout(DataLayout::NCHW);
142 
143  // Configure the function to transform the weights tensor from HWI -> IHW
144  _permute_weights_func.configure(compile_context, weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
145  _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
146 
147  input_to_use = &_permuted_input;
148  weights_to_use = &_permuted_weights;
149  }
150 
151  // Flip weights
152  _flipped_weights.allocator()->init(weights_to_use->info()->clone()->set_is_resizable(true).reset_padding());
153  _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
154  _flip_weights_func.configure(compile_context, weights_to_use, &_flipped_weights, &_flip_axis);
155 
156  // Pad weights
157  const PaddingList padding_w = { { 0, input_dims.x() + pad_valid.x() - 1 }, { 0, input_dims.y() + pad_valid.y() - 1 } };
158  _pad_weights_func.configure(compile_context, &_flipped_weights, &_padded_weights, padding_w);
159 
160  // Transform weights
161  _transform_weights_func = support::cpp14::make_unique<CLFFT2D>();
162  _transform_weights_func->configure(compile_context, &_padded_weights, &_transformed_weights, FFT2DInfo());
163 
164  // Pad input
165  const PaddingList padding_in = { { 0, kernel_size.x() + pad_valid.x() - 1 }, { 0, kernel_size.y() + pad_valid.y() - 1 } };
166  _memory_group.manage(&_padded_input);
167  _pad_input_func.configure(compile_context, input_to_use, &_padded_input, padding_in);
168  if(_needs_permute)
169  {
170  _permuted_input.allocator()->allocate();
171  }
172 
173  // Transform input
174  _memory_group.manage(&_transformed_input);
175  _transform_input_func.configure(compile_context, &_padded_input, &_transformed_input, FFT2DInfo());
176  _padded_input.allocator()->allocate();
177 
178  // Perform product
179  _memory_group.manage(&_output_product);
180  _prod_func.configure(compile_context, &_transformed_input, &_transformed_weights, &_output_product);
181  _transformed_input.allocator()->allocate();
182 
183  // Perform reduction
184  _memory_group.manage(&_output_reduced);
185  _reduce_func.configure(compile_context, &_output_product, &_output_reduced, 2, ReductionOperation::SUM);
186  _output_product.allocator()->allocate();
187 
188  // Transform output
189  _memory_group.manage(&_itransformed_output);
190  FFT2DInfo itranform_info;
191  itranform_info.direction = FFTDirection::Inverse;
192  _itransformed_output.allocator()->init(_output_reduced.info()->clone()->set_is_resizable(true).set_num_channels(1).reset_padding());
193  _itransform_output_func.configure(compile_context, &_output_reduced, &_itransformed_output, itranform_info);
194  _output_reduced.allocator()->allocate();
195 
196  // Reshape output
197  TensorShape reshaped_shape = _itransformed_output.info()->tensor_shape();
198  reshaped_shape.remove_dimension(2);
199  _reshaped_output.allocator()->init(_itransformed_output.info()->clone()->set_tensor_shape(reshaped_shape));
200 
201  // Extract correct region
202  const int start_left = kernel_size.x() - conv_info.pad_left() - 1;
203  const int start_top = kernel_size.y() - conv_info.pad_top() - 1;
204  const int end_right = _reshaped_output.info()->tensor_shape().x() - (kernel_size.x() - conv_info.pad_right() - 1) - pad_valid.x();
205  const int end_botton = _reshaped_output.info()->tensor_shape().y() - (kernel_size.y() - conv_info.pad_bottom() - 1) - pad_valid.y();
206  if(_has_bias)
207  {
208  _memory_group.manage(&_bias_output);
209  }
210  else if(_needs_permute)
211  {
212  output_to_use = &_permuted_output;
213  _memory_group.manage(&_permuted_output);
214  }
215  _extract_output_func.configure(compile_context, &_reshaped_output, output_to_use, Coordinates(start_left, start_top), Coordinates(end_right, end_botton));
216  _itransformed_output.allocator()->allocate();
217 
218  // Add bias
219  if(biases != nullptr)
220  {
221  output_to_use = output;
222  if(_needs_permute)
223  {
224  output_to_use = &_permuted_output;
225  _memory_group.manage(&_permuted_output);
226  }
227  auto_init_if_empty(*output_to_use->info(), *_bias_output.info());
228  _bias_add_func.configure(compile_context, &_bias_output, &_permuted_bias, output_to_use, ConvertPolicy::WRAP);
229  _bias_output.allocator()->allocate();
230  }
231 
232  // Permute output
233  if(_needs_permute)
234  {
235  // Configure the function to transform the convoluted output to ACL's native ordering format NCHW
236  _permuted_output.info()->set_data_layout(DataLayout::NCHW);
237  _permute_output_func.configure(compile_context, &_permuted_output, output, PermutationVector(2U, 0U, 1U));
238 
239  // Allocate tensors
240  _permuted_output.allocator()->allocate();
241  }
242 
243  // Configure Activation Layer
244  _is_activationlayer_enabled = act_info.enabled();
245  if(_is_activationlayer_enabled)
246  {
247  _activation_layer_func.configure(compile_context, output, nullptr, act_info);
248  }
249 
250  // Setup flip axis data
251  _flip_axis.allocator()->allocate();
252  _flip_axis.map(true);
253  auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
254  axis_data[0] = 0;
255  axis_data[1] = 1;
256  _flip_axis.unmap();
257 }
258 
260  const ActivationLayerInfo &act_info)
261 {
264 
265  // Get indices for the width and height
266  const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
267  const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
268 
269  // Input shape, kernel size and output tile
270  const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]);
271 
272  // Strides
273  const auto strides = conv_info.stride();
274  ARM_COMPUTE_RETURN_ERROR_ON(strides.first != strides.second && strides.first != 1);
275  ARM_COMPUTE_RETURN_ERROR_ON(kernel_size.x() != kernel_size.y());
276  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_left() != (kernel_size.x() / 2) || conv_info.pad_right() != (kernel_size.x() / 2));
277  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_top() != (kernel_size.y() / 2) || conv_info.pad_bottom() != (kernel_size.y() / 2));
278 
279  // Validate biases
280  if(biases != nullptr)
281  {
282  const size_t idx_channels = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
284  ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channels] != biases->tensor_shape().x());
285  }
286 
287  // Checks performed when output is configured
288  if((output != nullptr) && (output->total_size() != 0))
289  {
291  ARM_COMPUTE_RETURN_ERROR_ON((input->tensor_shape()[idx_height] != output->tensor_shape()[idx_height]) || (input->tensor_shape()[idx_width] != output->tensor_shape()[idx_width]));
292 
293  // Validate Activation Layer
294  if(act_info.enabled())
295  {
296  ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output, nullptr, act_info));
297  }
298  }
299 
300  return Status{};
301 }
302 
304 {
305  prepare();
306 
307  MemoryGroupResourceScope scope_mg(_memory_group);
308 
309  // Transform input
310  if(_needs_permute)
311  {
312  _permute_input_func.run();
313  }
314  _pad_input_func.run();
315  _transform_input_func.run();
316 
317  // Perform operations to frequency domain
318  _prod_func.run();
319  _reduce_func.run();
320 
321  // Transform output
322  _itransform_output_func.run();
323  _reshaped_output.allocator()->import_memory(_itransformed_output.cl_buffer());
324  _extract_output_func.run();
325  // Add bias
326  if(_has_bias)
327  {
328  _bias_add_func.run();
329  }
330  if(_needs_permute)
331  {
332  _permute_output_func.run();
333  }
334 
335  // Run activation layer
336  if(_is_activationlayer_enabled)
337  {
338  _activation_layer_func.run();
339  }
340 }
341 
343 {
344  if(!_is_prepared)
345  {
346  // Permute bias to NCHW
347  if(_original_bias != nullptr)
348  {
349  _permuted_bias.allocator()->allocate();
350  _permute_bias_func.run();
351  _original_bias->mark_as_unused();
352  }
353 
354  const ICLTensor *cur_weights = _original_weights;
355  // Permute weights
356  if(_needs_permute)
357  {
358  ARM_COMPUTE_ERROR_ON(!cur_weights->is_used());
359 
360  _permuted_weights.allocator()->allocate();
361  _permute_weights_func.run();
362  cur_weights->mark_as_unused();
363  cur_weights = &_permuted_weights;
364  }
365 
366  // Flip weights
367  _flipped_weights.allocator()->allocate();
368  _flip_weights_func.run();
369  cur_weights->mark_as_unused();
370 
371  // Pad weights
372  _padded_weights.allocator()->allocate();
373  _pad_weights_func.run();
374  _flipped_weights.mark_as_unused();
375  CLScheduler::get().queue().finish();
376  _flipped_weights.allocator()->free();
377 
378  // Transform weights to frequency domain
379  _transformed_weights.allocator()->allocate();
380  _transform_weights_func->run();
381  _padded_weights.mark_as_unused();
382  CLScheduler::get().queue().finish();
383  // Delete object and release internal memory
384  _transform_weights_func.reset();
385  _padded_weights.allocator()->free();
386 
387  _is_prepared = true;
388  }
389 }
390 } // namespace arm_compute
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
Static function to check if given info will lead to a valid configuration of CLActivationLayer.
Shape of a tensor.
Definition: TensorShape.h:39
FFTDirection direction
Direction of the FFT.
void remove_dimension(size_t n)
Accessor to remove the dimension n from the tensor shape.
Definition: TensorShape.h:110
std::unique_ptr< ITensorInfo > clone() const override
Provide a clone of the current object of class T.
Definition: TensorInfo.cpp:314
TensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: CLTensor.cpp:41
void run() override
Run the kernels contained in the function.
void configure(const ICLTensor *input, ICLTensor *output, const FFT2DInfo &config)
Initialise the function's source, destinations and border mode.
Definition: CLFFT2D.cpp:37
bool enabled() const
Check if initialised.
Definition: Types.h:1567
std::vector< PaddingInfo > PaddingList
List of padding information.
Definition: Types.h:458
static CLScheduler & get()
Access the scheduler singleton.
Definition: CLScheduler.cpp:99
std::vector< unsigned int > decompose_stages(unsigned int N, const std::set< unsigned int > &supported_factors)
Decompose a given 1D input size using the provided supported factors.
Definition: fft.cpp:34
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
void run() override
Run the kernels contained in the function.
const cl::Buffer & cl_buffer() const override
Interface to be implemented by the child class to return a reference to the OpenCL buffer containing ...
Definition: CLTensor.cpp:51
Descriptor used by the FFT2D 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
Strides PermutationVector
Permutation vector.
Definition: Types.h:49
#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
void run() override
Run the kernels contained in the function.
Definition: CLFFT2D.cpp:91
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
Store the tensor's metadata.
Definition: ITensorInfo.h:40
void run() override
Run the kernels contained in the function.
CLTensorAllocator * allocator()
Return a pointer to the tensor's allocator.
Definition: CLTensor.cpp:61
void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Set the input and output tensors.
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Activation Layer Information class.
Definition: Types.h:1517
void init(const TensorInfo &input, size_t alignment=0)
Initialize a tensor based on the passed TensorInfo.
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
Status import_memory(cl::Buffer buffer)
Import an existing memory as a tensor's backing memory.
void map(bool blocking=true)
Enqueue a map operation of the allocated buffer.
Definition: CLTensor.cpp:66
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
uint8_t * buffer() const override
Interface to be implemented by the child class to return a pointer to CPU memory.
Definition: ICLTensor.cpp:53
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
T x() const
Alias to access the size of the first dimension.
Definition: Dimensions.h:81
ITensorInfo & set_data_layout(const DataLayout &data_layout) override
Set the data layout of the tensor.
Definition: TensorInfo.cpp:378
void run() override final
Run the kernels contained in the function.
1 channel, 1 U32 per channel
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
void run() override
Run the kernels contained in the function.
void configure(ICLTensor *input, ICLTensor *output, const PaddingList &padding, PixelValue constant_value=PixelValue(), PaddingMode mode=PaddingMode::CONSTANT)
Initialize the function.
Definition: CLPadLayer.cpp:33
Coordinates of an item.
Definition: Coordinates.h:37
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
Padding and stride information class.
Definition: Types.h:689
void configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, ConvertPolicy policy, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Initialise the kernel's inputs, output and conversion policy.
cl::CommandQueue & queue()
Accessor for the associated CL command queue.
Definition: CLScheduler.cpp:41
Num samples, channels, height, width.
CLCompileContext class.
static std::set< unsigned int > supported_radix()
Returns the radix that are support by the FFT kernel.
void configure(const ICLTensor *input, ICLTensor *output, const Coordinates &starts, const Coordinates &ends)
Configure kernel.
Definition: CLSlice.cpp:85
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
virtual size_t total_size() const =0
Returns the total size of the tensor in bytes.
void run() override
Run the kernels contained in the function.
Definition: CLSlice.cpp:98
void run() override
Run the kernels contained in the function.
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
Num samples, height, width, channels.
CLFFTConvolutionLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Static function to check if given info will lead to a valid configuration of CLFFTConvolutionLayer.
void free() override
Free allocated OpenCL memory.
void configure(ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info)
Set the input and output tensor.
Store the tensor's metadata.
Definition: TensorInfo.h:45
void configure(ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, bool keep_dims=true)
Set the input and output tensors.
void configure(const ICLTensor *input, ICLTensor *output, const PermutationVector &perm)
Set the input and output tensors.
Definition: CLPermute.cpp:33
T y() const
Alias to access the size of the second dimension.
Definition: Dimensions.h:86
void run() override
Run the kernels contained in the function.
Definition: CLPadLayer.cpp:76
void prepare() override
Prepare the function for executing.
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
void unmap()
Enqueue an unmap operation of the allocated and mapped buffer.
Definition: CLTensor.cpp:71
void configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Initialise the kernel's inputs, output.
void configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *axis)
Initialize the function.
Definition: CLReverse.cpp:32