Compute Library
 20.02.1
NEFFTConvolutionLayer.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019 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 namespace arm_compute
33 {
34 namespace
35 {
36 int pad_decomposable(int N)
37 {
38  const auto supported_radix = NEFFTRadixStageKernel::supported_radix();
39 
40  int pad = 0;
41  bool is_decomposed = false;
42  while(!is_decomposed)
43  {
44  const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N++, supported_radix);
45  is_decomposed = !decomposed_vector.empty();
46  if(!is_decomposed)
47  {
48  ++pad;
49  }
50  }
51  return pad;
52 }
53 } // namespace
54 
55 NEFFTConvolutionLayer::NEFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
56  : _memory_group(memory_manager),
57  _flip_weights_func(),
58  _permute_input_func(),
59  _permute_output_func(),
60  _permute_weights_func(),
61  _permute_bias_func(),
62  _pad_input_func(),
63  _pad_weights_func(),
64  _transform_input_func(memory_manager),
65  _transform_weights_func(),
66  _itransform_output_func(memory_manager),
67  _prod_func(),
68  _reduce_func(),
69  _extract_output_func(),
70  _bias_add_func(),
71  _activation_layer_func(),
72  _permuted_input(),
73  _permuted_weights(),
74  _permuted_bias(),
75  _permuted_output(),
76  _padded_input(),
77  _padded_weights(),
78  _flip_axis(),
79  _flipped_weights(),
80  _transformed_input(),
81  _transformed_weights(),
82  _input_weights_product(),
83  _output_product(),
84  _output_reduced(),
85  _itransformed_output(),
86  _reshaped_output(),
87  _bias_output(),
88  _original_weights(nullptr),
89  _original_bias(nullptr),
90  _is_activationlayer_enabled(false),
91  _needs_permute(false),
92  _has_bias(false),
93  _is_prepared(false)
94 {
95 }
96 
99 {
100  _original_weights = weights;
101  _original_bias = biases;
102 
103  // Flat if bias addition is required
104  _has_bias = biases != nullptr;
105 
106  // Get indices for the width and height
107  const size_t idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
108  const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
109 
110  // Input shape, kernel size and output tile
111  const Size2D input_dims = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]);
112  const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]);
113  const Size2D pad_valid = Size2D(pad_decomposable(input_dims.x() + kernel_size.x() - 1),
114  pad_decomposable(input_dims.y() + kernel_size.y() - 1));
115  // Tensors to use
116  ITensor *input_to_use = input;
117  const ITensor *weights_to_use = weights;
118  ITensor *output_to_use = _has_bias ? &_bias_output : output;
119 
120  // Permute bias
121  if(biases != nullptr)
122  {
123  _permute_bias_func.configure(biases, &_permuted_bias, PermutationVector(1U, 2U, 0U));
124  _permuted_bias.info()->set_data_layout(DataLayout::NCHW);
125  }
126 
127  // Permute input if needed
128  _needs_permute = input->info()->data_layout() == DataLayout::NHWC;
129  if(_needs_permute)
130  {
131  _memory_group.manage(&_permuted_input);
132  // Configure the function to transform the input tensor from NHWC -> NCHW
133  _permute_input_func.configure(input, &_permuted_input, PermutationVector(1U, 2U, 0U));
134  _permuted_input.info()->set_data_layout(DataLayout::NCHW);
135 
136  // Configure the function to transform the weights tensor from HWI -> IHW
137  _permute_weights_func.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
138  _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
139 
140  input_to_use = &_permuted_input;
141  weights_to_use = &_permuted_weights;
142  }
143 
144  // Flip weights
145  _flipped_weights.allocator()->init(weights_to_use->info()->clone()->set_is_resizable(true).reset_padding());
146  _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
147  _flip_weights_func.configure(weights_to_use, &_flipped_weights, &_flip_axis);
148 
149  // Pad weights
150  const PaddingList padding_w = { { 0, input_dims.x() + pad_valid.x() - 1 }, { 0, input_dims.y() + pad_valid.y() - 1 } };
151  _pad_weights_func.configure(&_flipped_weights, &_padded_weights, padding_w);
152 
153  // Transform weights
154  _transform_weights_func = support::cpp14::make_unique<NEFFT2D>();
155  _transform_weights_func->configure(&_padded_weights, &_transformed_weights, FFT2DInfo());
156 
157  // Pad input
158  const PaddingList padding_in = { { 0, kernel_size.x() + pad_valid.x() - 1 }, { 0, kernel_size.y() + pad_valid.y() - 1 } };
159  _memory_group.manage(&_padded_input);
160  _pad_input_func.configure(input_to_use, &_padded_input, padding_in);
161  if(_needs_permute)
162  {
163  _permuted_input.allocator()->allocate();
164  }
165 
166  // Transform input
167  _memory_group.manage(&_transformed_input);
168  _transform_input_func.configure(&_padded_input, &_transformed_input, FFT2DInfo());
169  _padded_input.allocator()->allocate();
170 
171  // Perform product
172  _memory_group.manage(&_output_product);
173  _prod_func.configure(&_transformed_input, &_transformed_weights, &_output_product);
174  _transformed_input.allocator()->allocate();
175 
176  // Perform reduction
177  _memory_group.manage(&_output_reduced);
178  _reduce_func.configure(&_output_product, &_output_reduced, 2, ReductionOperation::SUM);
179  _output_product.allocator()->allocate();
180 
181  // Transform output
182  _memory_group.manage(&_itransformed_output);
183  FFT2DInfo itranform_info;
184  itranform_info.direction = FFTDirection::Inverse;
185  _itransformed_output.allocator()->init(_output_reduced.info()->clone()->set_is_resizable(true).set_num_channels(1).reset_padding());
186  _itransform_output_func.configure(&_output_reduced, &_itransformed_output, itranform_info);
187  _output_reduced.allocator()->allocate();
188 
189  // Reshape output
190  TensorShape reshaped_shape = _itransformed_output.info()->tensor_shape();
191  reshaped_shape.remove_dimension(2);
192  _reshaped_output.allocator()->init(_itransformed_output.info()->clone()->set_tensor_shape(reshaped_shape));
193 
194  // Extract correct region
195  const int start_left = kernel_size.x() - conv_info.pad_left() - 1;
196  const int start_top = kernel_size.y() - conv_info.pad_top() - 1;
197  const int end_right = _reshaped_output.info()->tensor_shape().x() - (kernel_size.x() - conv_info.pad_right() - 1) - pad_valid.x();
198  const int end_botton = _reshaped_output.info()->tensor_shape().y() - (kernel_size.y() - conv_info.pad_bottom() - 1) - pad_valid.y();
199  if(_has_bias)
200  {
201  _memory_group.manage(&_bias_output);
202  }
203  else if(_needs_permute)
204  {
205  output_to_use = &_permuted_output;
206  _memory_group.manage(&_permuted_output);
207  }
208  _extract_output_func.configure(&_reshaped_output, output_to_use, Coordinates(start_left, start_top), Coordinates(end_right, end_botton));
209  _reshaped_output.allocator()->allocate();
210  _itransformed_output.allocator()->allocate();
211 
212  // Add bias
213  if(biases != nullptr)
214  {
215  output_to_use = output;
216  if(_needs_permute)
217  {
218  output_to_use = &_permuted_output;
219  _memory_group.manage(&_permuted_output);
220  }
221  auto_init_if_empty(*output_to_use->info(), *_bias_output.info());
222  _bias_add_func.configure(&_bias_output, &_permuted_bias, output_to_use, ConvertPolicy::WRAP);
223  _bias_output.allocator()->allocate();
224  }
225 
226  // Permute output
227  if(_needs_permute)
228  {
229  // Configure the function to transform the convoluted output to ACL's native ordering format NCHW
230  _permuted_output.info()->set_data_layout(DataLayout::NCHW);
231  _permute_output_func.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U));
232 
233  // Allocate tensors
234  _permuted_output.allocator()->allocate();
235  }
236 
237  // Configure Activation Layer
238  _is_activationlayer_enabled = act_info.enabled();
239  if(_is_activationlayer_enabled)
240  {
241  _activation_layer_func.configure(output, nullptr, act_info);
242  }
243 
244  // Setup flip axis data
245  _flip_axis.allocator()->allocate();
246 
247  auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
248  axis_data[0] = 0;
249  axis_data[1] = 1;
250 }
251 
254 {
257 
258  // Get indices for the width and height
259  const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
260  const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
261 
262  // Input shape, kernel size and output tile
263  const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]);
264 
265  // Strides
266  const auto strides = conv_info.stride();
267  ARM_COMPUTE_RETURN_ERROR_ON(strides.first != strides.second && strides.first != 1);
268  ARM_COMPUTE_RETURN_ERROR_ON(kernel_size.x() != kernel_size.y());
269  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_left() != (kernel_size.x() / 2) || conv_info.pad_right() != (kernel_size.x() / 2));
270  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_top() != (kernel_size.y() / 2) || conv_info.pad_bottom() != (kernel_size.y() / 2));
271 
272  // Validate biases
273  if(biases != nullptr)
274  {
275  const size_t idx_channels = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
277  ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channels] != biases->tensor_shape().x());
278  }
279 
280  // Checks performed when output is configured
281  if((output != nullptr) && (output->total_size() != 0))
282  {
284  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]));
285 
286  // Validate Activation Layer
287  if(act_info.enabled())
288  {
290  }
291  }
292 
293  return Status{};
294 }
295 
297 {
298  prepare();
299 
300  MemoryGroupResourceScope scope_mg(_memory_group);
301 
302  // Transform input
303  if(_needs_permute)
304  {
305  _permute_input_func.run();
306  }
307  _pad_input_func.run();
308  _transform_input_func.run();
309 
310  // Perform operations to frequency domain
311  _prod_func.run();
312 
313  _reduce_func.run();
314 
315  // Transform output
316  _itransform_output_func.run();
317  _reshaped_output.allocator()->import_memory(_itransformed_output.buffer());
318  _extract_output_func.run();
319 
320  // Add bias
321  if(_has_bias)
322  {
323  _bias_add_func.run();
324  }
325  if(_needs_permute)
326  {
327  _permute_output_func.run();
328  }
329 
330  // Run activation layer
331  if(_is_activationlayer_enabled)
332  {
333  _activation_layer_func.run();
334  }
335 }
336 
338 {
339  if(!_is_prepared)
340  {
341  // Permute bias to NCHW
342  if(_original_bias != nullptr)
343  {
344  _permuted_bias.allocator()->allocate();
345  _permute_bias_func.run();
346  _original_bias->mark_as_unused();
347  }
348 
349  const ITensor *cur_weights = _original_weights;
350 
351  // Permute weights
352  if(_needs_permute)
353  {
354  ARM_COMPUTE_ERROR_ON(!cur_weights->is_used());
355 
356  _permuted_weights.allocator()->allocate();
357  _permute_weights_func.run();
358  cur_weights->mark_as_unused();
359  cur_weights = &_permuted_weights;
360  }
361 
362  // Flip weights
363  _flipped_weights.allocator()->allocate();
364  _flip_weights_func.run();
365  cur_weights->mark_as_unused();
366 
367  // Pad weights
368  _padded_weights.allocator()->allocate();
369  _pad_weights_func.run();
370  _flipped_weights.mark_as_unused();
371  _flipped_weights.allocator()->free();
372 
373  // Transform weights to frequency domain
374  _transformed_weights.allocator()->allocate();
375  _transform_weights_func->run();
376  _transform_weights_func.reset();
377 
378  _padded_weights.mark_as_unused();
379  _padded_weights.allocator()->free();
380 
381  _is_prepared = true;
382  }
383 }
384 } // namespace arm_compute
void configure(ITensor *input1, ITensor *input2, ITensor *output, ConvertPolicy policy)
Initialise the kernel's inputs, output and conversion policy.
void prepare() override
Prepare the function for executing.
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
void run() override final
Run the kernels contained in the function.
void run() override
Run the kernels contained in the function.
TensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: CLTensor.cpp:41
void init(const TensorAllocator &allocator, const Coordinates &coords, TensorInfo &sub_info)
Shares the same backing memory with another tensor allocator, while the tensor info might be differen...
std::vector< PaddingInfo > PaddingList
List of padding information.
Definition: Types.h:455
void run() override final
Run the kernels contained in the function.
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
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:162
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
[NEActivationLayer snippet]
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
1 channel, 1 F32 per channel
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 NEFFTConvolutionLayer.
Strides PermutationVector
Permutation vector.
Definition: Types.h:48
#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
Store the tensor's metadata.
Definition: ITensorInfo.h:40
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:1615
Interface for NEON tensor.
Definition: ITensor.h:36
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:202
void configure(const ITensor *input, ITensor *output, const ITensor *axis)
Initialize the function.
Definition: NEReverse.cpp:31
void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Set the input and output tensors.
TensorAllocator * allocator()
Return a pointer to the tensor's allocator.
Definition: Tensor.cpp:48
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:167
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
void configure(const ITensor *input, ITensor *output, const Coordinates &starts, const Coordinates &ends)
Configure kernel.
Definition: NESlice.cpp:36
void configure(ITensor *input1, ITensor *input2, ITensor *output)
Initialise the kernel's inputs, output.
NEFFTConvolutionLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
1 channel, 1 U32 per channel
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
virtual ITensorInfo & set_data_layout(const DataLayout &data_layout)=0
Set the data layout of the tensor.
Coordinates of an item.
Definition: Coordinates.h:37
void allocate() override
Allocate size specified by TensorInfo of CPU memory.
static std::set< unsigned int > supported_radix()
Returns the radix that are support by the FFT kernel.
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:686
void free() override
Free allocated CPU memory.
Num samples, channels, height, width.
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
void run() override
Run the kernels contained in the function.
virtual size_t total_size() const =0
Returns the total size of the tensor in bytes.
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
Num samples, height, width, channels.
void configure(ITensor *input, ITensor *output, ActivationLayerInfo activation_info)
[NEActivationLayer snippet]
void run() override
Run the kernels contained in the function.
Definition: NEPadLayer.cpp:247
void run() override
Run the kernels contained in the function.
Definition: NEFFT2D.cpp:86
uint8_t * buffer() const override
Interface to be implemented by the child class to return a pointer to CPU memory.
Definition: Tensor.cpp:43
void configure(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value=PixelValue(), const PaddingMode mode=PaddingMode::CONSTANT)
Initialize the function.
Definition: NEPadLayer.cpp:164
Status import_memory(void *memory)
Import an existing memory as a tensor's backing memory.
Store the tensor's metadata.
Definition: TensorInfo.h:45
T y() const
Alias to access the size of the second dimension.
Definition: Dimensions.h:86
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:327
void configure(const ITensor *input, ITensor *output, const PermutationVector &perm)
Configure the permute NEON kernel.
Definition: NEPermute.cpp:31
const TensorShape & tensor_shape() const override
Size for each dimension of the tensor.
Definition: TensorInfo.h:261
void configure(const ITensor *input, ITensor *output, const FFT2DInfo &config)
Initialise the function's source and destinations.
Definition: NEFFT2D.cpp:37
void configure(ITensor *input, ITensor *output, unsigned int axis, ReductionOperation op, bool keep_dims=true)
Set the input and output tensors.