Compute Library
 21.05
CLFFTConvolutionLayer.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2021 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"
40 
41 namespace arm_compute
42 {
43 namespace
44 {
45 int pad_decomposable(int N)
46 {
47  const auto supported_radix = CLFFTRadixStageKernel::supported_radix();
48 
49  int pad = 0;
50  bool is_decomposed = false;
51  while(!is_decomposed)
52  {
53  const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N++, supported_radix);
54  is_decomposed = !decomposed_vector.empty();
55  if(!is_decomposed)
56  {
57  ++pad;
58  }
59  }
60  return pad;
61 }
62 } // namespace
63 CLFFTConvolutionLayer::CLFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
64  : _memory_group(memory_manager),
65  _flip_weights_func(),
66  _permute_input_func(),
67  _permute_output_func(),
68  _permute_weights_func(),
69  _permute_bias_func(),
70  _pad_input_func(),
71  _pad_weights_func(),
72  _transform_input_func(memory_manager),
73  _transform_weights_func(),
74  _itransform_output_func(memory_manager),
75  _prod_func(),
76  _reduce_func(),
77  _extract_output_func(),
78  _bias_add_func(),
79  _activation_layer_func(),
80  _permuted_input(),
81  _permuted_weights(),
82  _permuted_bias(),
83  _permuted_output(),
84  _padded_input(),
85  _padded_weights(),
86  _flip_axis(),
87  _flipped_weights(),
88  _transformed_input(),
89  _transformed_weights(),
90  _input_weights_product(),
91  _output_product(),
92  _output_reduced(),
93  _itransformed_output(),
94  _reshaped_output(),
95  _bias_output(),
96  _original_weights(nullptr),
97  _original_bias(nullptr),
98  _is_activationlayer_enabled(false),
99  _needs_permute(false),
100  _has_bias(false),
101  _is_prepared(false)
102 {
103 }
104 
106  const ActivationLayerInfo &act_info, bool enable_fast_math)
107 {
108  configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, act_info, enable_fast_math);
109 }
110 
111 void CLFFTConvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
112  const ActivationLayerInfo &act_info, bool enable_fast_math)
113 {
114  ARM_COMPUTE_UNUSED(enable_fast_math);
115  ARM_COMPUTE_ERROR_THROW_ON(CLFFTConvolutionLayer::validate(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info, act_info, enable_fast_math));
116 
117  _original_weights = weights;
118  _original_bias = biases;
119 
120  // Flat if bias addition is required
121  _has_bias = biases != nullptr;
122 
123  // Get indices for the width and height
124  const size_t idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
125  const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
126 
127  // Input shape, kernel size and output tile
128  const Size2D input_dims = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]);
129  const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]);
130  const Size2D pad_valid = Size2D(pad_decomposable(input_dims.x() + kernel_size.x() - 1),
131  pad_decomposable(input_dims.y() + kernel_size.y() - 1));
132  // Tensors to use
133  ICLTensor *input_to_use = input;
134  const ICLTensor *weights_to_use = weights;
135  ICLTensor *output_to_use = _has_bias ? &_bias_output : output;
136 
137  // Permute bias
138  if(biases != nullptr)
139  {
140  _permute_bias_func.configure(compile_context, biases, &_permuted_bias, PermutationVector(1U, 2U, 0U));
141  _permuted_bias.info()->set_data_layout(DataLayout::NCHW);
142  }
143 
144  // Permute input if needed
145  _needs_permute = input->info()->data_layout() == DataLayout::NHWC;
146  if(_needs_permute)
147  {
148  _memory_group.manage(&_permuted_input);
149  // Configure the function to transform the input tensor from NHWC -> NCHW
150  _permute_input_func.configure(compile_context, input, &_permuted_input, PermutationVector(1U, 2U, 0U));
151  _permuted_input.info()->set_data_layout(DataLayout::NCHW);
152 
153  // Configure the function to transform the weights tensor from HWI -> IHW
154  _permute_weights_func.configure(compile_context, weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
155  _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
156 
157  input_to_use = &_permuted_input;
158  weights_to_use = &_permuted_weights;
159  }
160 
161  // Flip weights
162  _flipped_weights.allocator()->init(weights_to_use->info()->clone()->set_is_resizable(true).reset_padding());
163  _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
164  _flip_weights_func.configure(compile_context, weights_to_use, &_flipped_weights, &_flip_axis);
165 
166  // Pad weights
167  const PaddingList padding_w = { { 0, input_dims.x() + pad_valid.x() - 1 }, { 0, input_dims.y() + pad_valid.y() - 1 } };
168  _pad_weights_func.configure(compile_context, &_flipped_weights, &_padded_weights, padding_w);
169 
170  // Transform weights
171  _transform_weights_func = std::make_unique<CLFFT2D>();
172  _transform_weights_func->configure(compile_context, &_padded_weights, &_transformed_weights, FFT2DInfo());
173 
174  // Pad input
175  const PaddingList padding_in = { { 0, kernel_size.x() + pad_valid.x() - 1 }, { 0, kernel_size.y() + pad_valid.y() - 1 } };
176  _memory_group.manage(&_padded_input);
177  _pad_input_func.configure(compile_context, input_to_use, &_padded_input, padding_in);
178  if(_needs_permute)
179  {
180  _permuted_input.allocator()->allocate();
181  }
182 
183  // Transform input
184  _memory_group.manage(&_transformed_input);
185  _transform_input_func.configure(compile_context, &_padded_input, &_transformed_input, FFT2DInfo());
186  _padded_input.allocator()->allocate();
187 
188  // Perform product
189  _memory_group.manage(&_output_product);
190  _prod_func.configure(compile_context, &_transformed_input, &_transformed_weights, &_output_product);
191  _transformed_input.allocator()->allocate();
192 
193  // Perform reduction
194  _memory_group.manage(&_output_reduced);
195  _reduce_func.configure(compile_context, &_output_product, &_output_reduced, 2, ReductionOperation::SUM);
196  _output_product.allocator()->allocate();
197 
198  // Transform output
199  _memory_group.manage(&_itransformed_output);
200  FFT2DInfo itranform_info;
201  itranform_info.direction = FFTDirection::Inverse;
202  _itransformed_output.allocator()->init(_output_reduced.info()->clone()->set_is_resizable(true).set_num_channels(1).reset_padding());
203  _itransform_output_func.configure(compile_context, &_output_reduced, &_itransformed_output, itranform_info);
204  _output_reduced.allocator()->allocate();
205 
206  // Reshape output
207  TensorShape reshaped_shape = _itransformed_output.info()->tensor_shape();
208  reshaped_shape.remove_dimension(2);
209  _reshaped_output.allocator()->init(_itransformed_output.info()->clone()->set_tensor_shape(reshaped_shape));
210 
211  // Extract correct region
212  const int start_left = kernel_size.x() - conv_info.pad_left() - 1;
213  const int start_top = kernel_size.y() - conv_info.pad_top() - 1;
214  const int end_right = _reshaped_output.info()->tensor_shape().x() - (kernel_size.x() - conv_info.pad_right() - 1) - pad_valid.x();
215  const int end_botton = _reshaped_output.info()->tensor_shape().y() - (kernel_size.y() - conv_info.pad_bottom() - 1) - pad_valid.y();
216  if(_has_bias)
217  {
218  _memory_group.manage(&_bias_output);
219  }
220  else if(_needs_permute)
221  {
222  output_to_use = &_permuted_output;
223  _memory_group.manage(&_permuted_output);
224  }
225  _extract_output_func.configure(compile_context, &_reshaped_output, output_to_use, Coordinates(start_left, start_top), Coordinates(end_right, end_botton));
226  _itransformed_output.allocator()->allocate();
227 
228  // Add bias
229  if(biases != nullptr)
230  {
231  output_to_use = output;
232  if(_needs_permute)
233  {
234  output_to_use = &_permuted_output;
235  _memory_group.manage(&_permuted_output);
236  }
237  auto_init_if_empty(*output_to_use->info(), *_bias_output.info());
238  _bias_add_func.configure(compile_context, &_bias_output, &_permuted_bias, output_to_use, ConvertPolicy::WRAP);
239  _bias_output.allocator()->allocate();
240  }
241 
242  // Permute output
243  if(_needs_permute)
244  {
245  // Configure the function to transform the convoluted output to ACL's native ordering format NCHW
246  _permuted_output.info()->set_data_layout(DataLayout::NCHW);
247  _permute_output_func.configure(compile_context, &_permuted_output, output, PermutationVector(2U, 0U, 1U));
248 
249  // Allocate tensors
250  _permuted_output.allocator()->allocate();
251  }
252 
253  // Configure Activation Layer
254  _is_activationlayer_enabled = act_info.enabled();
255  if(_is_activationlayer_enabled)
256  {
257  _activation_layer_func.configure(compile_context, output, nullptr, act_info);
258  }
259 
260  // Setup flip axis data
261  _flip_axis.allocator()->allocate();
262  _flip_axis.map(true);
263  auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
264  axis_data[0] = 0;
265  axis_data[1] = 1;
266  _flip_axis.unmap();
267 }
268 
270  const ActivationLayerInfo &act_info, bool enable_fast_math)
271 {
273  ARM_COMPUTE_RETURN_ERROR_ON((input->data_type() == DataType::F16) && !enable_fast_math);
275 
276  // Get indices for the width and height
279 
280  // Input shape, kernel size and output tile
281  const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]);
282 
283  // Strides
284  const auto strides = conv_info.stride();
285  ARM_COMPUTE_RETURN_ERROR_ON(strides.first != strides.second && strides.first != 1);
286  ARM_COMPUTE_RETURN_ERROR_ON(kernel_size.x() != kernel_size.y());
287  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_left() != (kernel_size.x() / 2) || conv_info.pad_right() != (kernel_size.x() / 2));
288  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_top() != (kernel_size.y() / 2) || conv_info.pad_bottom() != (kernel_size.y() / 2));
289 
290  // Validate biases
291  if(biases != nullptr)
292  {
294  ARM_COMPUTE_RETURN_ERROR_ON(weights->tensor_shape()[3] != biases->tensor_shape().x());
295  }
296 
297  // Checks performed when output is configured
298  if((output != nullptr) && (output->total_size() != 0))
299  {
301  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]));
302 
303  // Validate Activation Layer
304  if(act_info.enabled())
305  {
306  ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output, nullptr, act_info));
307  }
308  }
309 
310  return Status{};
311 }
312 
314 {
315  prepare();
316 
317  MemoryGroupResourceScope scope_mg(_memory_group);
318 
319  // Transform input
320  if(_needs_permute)
321  {
322  _permute_input_func.run();
323  }
324  _pad_input_func.run();
325  _transform_input_func.run();
326 
327  // Perform operations to frequency domain
328  _prod_func.run();
329  _reduce_func.run();
330 
331  // Transform output
332  _itransform_output_func.run();
333  _reshaped_output.allocator()->import_memory(_itransformed_output.cl_buffer());
334  _extract_output_func.run();
335  // Add bias
336  if(_has_bias)
337  {
338  _bias_add_func.run();
339  }
340  if(_needs_permute)
341  {
342  _permute_output_func.run();
343  }
344 
345  // Run activation layer
346  if(_is_activationlayer_enabled)
347  {
348  _activation_layer_func.run();
349  }
350 }
351 
353 {
354  if(!_is_prepared)
355  {
356  // Permute bias to NCHW
357  if(_original_bias != nullptr)
358  {
359  _permuted_bias.allocator()->allocate();
360  _permute_bias_func.run();
361  _original_bias->mark_as_unused();
362  }
363 
364  const ICLTensor *cur_weights = _original_weights;
365  // Permute weights
366  if(_needs_permute)
367  {
368  ARM_COMPUTE_ERROR_ON(!cur_weights->is_used());
369 
370  _permuted_weights.allocator()->allocate();
371  _permute_weights_func.run();
372  cur_weights->mark_as_unused();
373  cur_weights = &_permuted_weights;
374  }
375 
376  // Flip weights
377  _flipped_weights.allocator()->allocate();
378  _flip_weights_func.run();
379  cur_weights->mark_as_unused();
380 
381  // Pad weights
382  _padded_weights.allocator()->allocate();
383  _pad_weights_func.run();
384  _flipped_weights.mark_as_unused();
385  CLScheduler::get().queue().finish();
386  _flipped_weights.allocator()->free();
387 
388  // Transform weights to frequency domain
389  _transformed_weights.allocator()->allocate();
390  _transform_weights_func->run();
391  _padded_weights.mark_as_unused();
392  CLScheduler::get().queue().finish();
393  // Delete object and release internal memory
394  _transform_weights_func.reset();
395  _padded_weights.allocator()->free();
396 
397  _is_prepared = true;
398  }
399 }
400 } // 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:111
std::unique_ptr< ITensorInfo > clone() const override
Provide a clone of the current object of class T.
Definition: TensorInfo.cpp:281
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:42
bool enabled() const
Check if initialised.
Definition: Types.h:1528
std::vector< PaddingInfo > PaddingList
List of padding information.
Definition: Types.h:434
static CLScheduler & get()
Access the scheduler singleton.
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
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
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:97
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
Store the tensor's metadata.
Definition: ITensorInfo.h:40
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 CLFFTConvolutionLayer.
void run() override
Run the kernels contained in the function.
CLTensorAllocator * allocator()
Return a pointer to the tensor's allocator.
Definition: CLTensor.cpp:61
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
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:1478
unsigned int N
void init(const TensorInfo &input, size_t alignment=0)
Initialize a tensor based on the passed TensorInfo.
Copyright (c) 2017-2021 Arm Limited.
Status import_memory(cl::Buffer buffer)
Import an existing memory as a tensor's backing memory.
void run() override
Run the kernels contained in the function.
Definition: CLPermute.cpp:70
1 channel, 1 F16 per channel
void map(bool blocking=true)
Enqueue a map operation of the allocated buffer.
Definition: CLTensor.cpp:66
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:87
Interface to enqueue OpenCL kernels and get/set the OpenCL CommandQueue and ICLTuner.
ITensorInfo & set_data_layout(const DataLayout &data_layout) override
Set the data layout of the tensor.
Definition: TensorInfo.cpp:351
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
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:38
Coordinates of an item.
Definition: Coordinates.h:37
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...
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:650
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:39
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:84
Basic pool of threads to execute CPP/Neon code on several cores in parallel.
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:97
void run() override
Run the kernels contained in the function.
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
Num samples, height, width, channels.
CLFFTConvolutionLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
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:43
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:49
T y() const
Alias to access the size of the second dimension.
Definition: Dimensions.h:92
void run() override
Run the kernels contained in the function.
Definition: CLPadLayer.cpp:79
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:193
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.
const TensorShape & tensor_shape() const override
Size for each dimension of the tensor.
Definition: TensorInfo.h:234
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:31