Compute Library
 21.05
CLDirectDeconvolutionLayer.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 
28 #include "arm_compute/core/Utils.h"
37 
38 #include <memory>
39 #include <tuple>
40 
41 namespace arm_compute
42 {
44 
45 CLDirectDeconvolutionLayer::CLDirectDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
46  : _memory_group(std::move(memory_manager)),
47  _scale_f(),
48  _conv_f(),
49  _flip_weights(),
50  _scaled_output(),
51  _original_weights(nullptr),
52  _weights_flipped(),
53  _flip_axis(),
54  _is_prepared(false)
55 {
56 }
57 
60 {
64  const DataLayout data_layout = input->data_layout();
65 
69 
70  ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) != weights->dimension(idx_h));
71  ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) < 1);
72 
73  auto out_dims = deconvolution_output_dimensions(input->dimension(idx_w), input->dimension(idx_h), weights->dimension(idx_w), weights->dimension(idx_h), info);
74 
76 
78 
79  if(input->data_type() != weights->data_type())
80  {
82  }
83 
84  if(bias != nullptr)
85  {
87  {
89  }
90  else
91  {
93  }
95  }
96 
97  ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_w) != output_shape[idx_w], "Output's width is invalid.");
98  ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_h) != output_shape[idx_h], "Output's height is invalid.");
99  ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_c) != output_shape[idx_c], "Output's depth is invalid.");
100 
101  unsigned int deconv_pad_x = 0;
102  unsigned int deconv_pad_y = 0;
103  const unsigned int stride_x = info.stride().first;
104  const unsigned int stride_y = info.stride().second;
105  const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
106  TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(scale_out_shape).set_data_layout(data_layout));
107  const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
108 
110  ARM_COMPUTE_RETURN_ON_ERROR(CLConvolutionLayer::validate(&scale_out_info, weights, bias, output, conv_info, weights_info));
111 
112  return Status{};
113 }
114 
116  const WeightsInfo &weights_info)
117 {
118  configure(CLKernelLibrary::get().get_compile_context(), input, weights, bias, output, info, weights_info);
119 }
120 
121 void CLDirectDeconvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
122  const WeightsInfo &weights_info)
123 {
124  ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
125 
126  const unsigned int pad_left = info.pad_left();
127  const unsigned int pad_right = info.pad_right();
128  const unsigned int pad_top = info.pad_top();
129  const unsigned int pad_bottom = info.pad_bottom();
130  const unsigned int stride_x = info.stride().first;
131  const unsigned int stride_y = info.stride().second;
132 
133  const DataLayout data_layout = input->info()->data_layout();
134 
137 
138  _original_weights = weights;
139  _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
140  _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
141  _flip_weights.configure(compile_context, weights, &_weights_flipped, &_flip_axis);
142 
143  auto out_dims = deconvolution_output_dimensions(input->info()->dimension(idx_w), input->info()->dimension(idx_h), weights->info()->dimension(idx_w), weights->info()->dimension(idx_h), info);
144 
145  const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
146 
147  // Output auto initialization if not yet initialized
148  auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_layout(data_layout));
149 
150  // Perform validation step
151  ARM_COMPUTE_ERROR_THROW_ON(CLDirectDeconvolutionLayer::validate(input->info(), weights->info(), bias == nullptr ? nullptr : bias->info(), output->info(), info));
152 
153  _is_prepared = weights_info.retain_internal_weights();
154 
155  _memory_group.manage(&_scaled_output);
156 
157  // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
158  unsigned int deconv_pad_x = 0;
159  unsigned int deconv_pad_y = 0;
160  const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
161 
162  unsigned int deconv_pad_left = pad_right > pad_left ? pad_right - pad_left : 0;
163  unsigned int deconv_pad_right = pad_left > pad_right ? pad_left - pad_right : 0;
164  deconv_pad_x -= deconv_pad_left + deconv_pad_right;
165  ARM_COMPUTE_ERROR_ON((deconv_pad_x % 2) != 0);
166  deconv_pad_left += deconv_pad_x / 2;
167  deconv_pad_right += deconv_pad_x / 2;
168 
169  unsigned int deconv_pad_top = pad_bottom > pad_top ? pad_bottom - pad_top : 0;
170  unsigned int deconv_pad_bottom = pad_top > pad_bottom ? pad_top - pad_bottom : 0;
171  deconv_pad_y -= deconv_pad_top + deconv_pad_bottom;
172  ARM_COMPUTE_ERROR_ON((deconv_pad_y % 2) != 0);
173  deconv_pad_top += deconv_pad_y / 2;
174  deconv_pad_bottom += deconv_pad_y / 2;
175 
176  TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
177  scale_out_info.set_data_layout(data_layout);
178  _scaled_output.allocator()->init(scale_out_info);
179 
180  // configure scale function
181  const PadStrideInfo upsample_info(stride_x, stride_y, deconv_pad_left, deconv_pad_right, deconv_pad_top, deconv_pad_bottom, DimensionRoundingType::FLOOR);
182  _scale_f.configure(compile_context, input, &_scaled_output, upsample_info);
183 
184  // Setup the function to convolve the upscaled output
185  const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
186  _conv_f.configure(compile_context, &_scaled_output, &_weights_flipped, bias, output, conv_info, weights_info);
187  _scaled_output.allocator()->allocate();
188 
189  // Setup flip axis data
190  _flip_axis.allocator()->allocate();
191  _flip_axis.map(true);
192  auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
193  if(weights->info()->data_layout() == DataLayout::NHWC)
194  {
195  axis_data[0] = 1;
196  axis_data[1] = 2;
197  }
198  else
199  {
200  axis_data[0] = 0;
201  axis_data[1] = 1;
202  }
203  _flip_axis.unmap();
204 }
205 
207 {
208  prepare();
209 
210  MemoryGroupResourceScope scope_mg(_memory_group);
211 
212  _scale_f.run();
213  _conv_f.run();
214 }
215 
217 {
218  if(!_is_prepared)
219  {
220  ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
221 
222  // Run weights flipping and mark original weights tensor as unused
223  _weights_flipped.allocator()->allocate();
224  _flip_weights.run();
225  _original_weights->mark_as_unused();
226 
227  // Prepare convolution
228  _conv_f.prepare();
229 
230  // Free flipped weights
231  if(!_weights_flipped.is_used())
232  {
233  _weights_flipped.allocator()->free();
234  }
235  _is_prepared = true;
236  }
237 }
238 } // namespace arm_compute
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &info, const WeightsInfo &weights_info=WeightsInfo())
Static function to check if given info will lead to a valid configuration of CLDirectDeconvolutionLay...
Shape of a tensor.
Definition: TensorShape.h:39
void run() override
Run the kernels contained in the function.
void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info=WeightsInfo(), const Size2D &dilation=Size2D(1U, 1U), const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false, unsigned int num_groups=1)
Set the input and output tensors.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(...)
Definition: Validate.h:490
void configure(ICLTensor *input, ICLTensor *output, const PadStrideInfo &info)
Initialize the function's source, destination, interpolation type and border_mode.
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
std::pair< unsigned int, unsigned int > deconvolution_output_dimensions(unsigned int in_width, unsigned int in_height, unsigned int kernel_width, unsigned int kernel_height, const PadStrideInfo &pad_stride_info)
Returns expected width and height of the deconvolution's output tensor.
Definition: Utils.cpp:375
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
virtual DataType data_type() const =0
Data type used for each element of the tensor.
bool is_used() const
Flags if the tensor is used or not.
Definition: ITensor.cpp:163
void prepare() override
Prepare the function for executing.
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
const DataLayout data_layout
Definition: Im2Col.cpp:151
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
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info=WeightsInfo(), const Size2D &dilation=Size2D(1U, 1U), const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false, unsigned int num_groups=1)
Static function to check if given info will lead to a valid configuration of CLConvolutionLayer.
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
void init(const TensorInfo &input, size_t alignment=0)
Initialize a tensor based on the passed TensorInfo.
Copyright (c) 2017-2021 Arm Limited.
1 channel, 1 F16 per channel
CLDirectDeconvolutionLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Constructor.
void map(bool blocking=true)
Enqueue a map operation of the allocated buffer.
Definition: CLTensor.cpp:66
Convolution Layer Weights Information class.
Definition: Types.h:1693
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:159
TensorShape compute_deconvolution_output_shape(const std::pair< unsigned int, unsigned int > &out_dims, const ITensorInfo &input, const ITensorInfo &weights)
Calculate the output shape of the deconvolution layer.
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
1 channel, 1 S32 per channel
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
void run() override
Run the kernels contained in the function.
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
void run() override final
Run the kernels contained in the function.
1 channel, 1 U32 per channel
quantized, asymmetric fixed-point 8-bit number unsigned
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.
TensorShape compute_deconvolution_upsampled_shape(const ITensorInfo &input, const ITensorInfo &weights, unsigned int sx, unsigned int sy, std::pair< unsigned int, unsigned int > &out_dims, uint32_t &padx, uint32_t &pady)
Calculate the upsampled output shape used for deconvolution.
Padding and stride information class.
Definition: Types.h:650
void run() override
Run the kernels contained in the function.
void configure(ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info, const WeightsInfo &weights_info=WeightsInfo())
Set the input, weights, biases and output tensors.
CLCompileContext class.
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:989
quantized, symmetric per channel fixed-point 8-bit number
void allocate() override
Allocate size specified by TensorInfo of OpenCL memory.
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
void prepare() override
Prepare the function for executing.
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &info)
Static function to check if given info will lead to a valid configuration of CLDeconvolutionLayerUpsa...
Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
Num samples, height, width, channels.
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
void free() override
Free allocated OpenCL memory.
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
Store the tensor's metadata.
Definition: TensorInfo.h:43
quantized, asymmetric fixed-point 8-bit number signed
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 unmap()
Enqueue an unmap operation of the allocated and mapped buffer.
Definition: CLTensor.cpp:71
DataLayout
[DataLayout enum definition]
Definition: Types.h:114
void configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *axis)
Initialize the function.
Definition: CLReverse.cpp:31
virtual DataLayout data_layout() const =0
Get the data layout of the tensor.