Compute Library
 22.05
CLFFTConvolutionLayer.h
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  */
24 #ifndef ARM_COMPUTE_CLFFTCONVOLUTIONLAYER_H
25 #define ARM_COMPUTE_CLFFTCONVOLUTIONLAYER_H
26 
28 
29 #include "arm_compute/core/Types.h"
40 
41 namespace arm_compute
42 {
43 // Forward declarations
44 class ICLTensor;
45 
46 /** Basic function to execute FFT-based convolution on OpenCL. This function calls the following OpenCL functions/kernels:
47  *
48  * -# @ref CLPermute Permute input if NHWC(only NCHW is supported).
49  * -# @ref CLPadLayer Pad input.
50  * -# @ref CLFFT2D Forward transform to the frequency domain.
51  * -# @ref CLComplexPixelWiseMultiplication Complex element-wise product of input and the weights.
52  * -# @ref CLReductionOperation Reduction across channels.
53  * -# @ref CLFFT2D Inverse transform back to the time domain.
54  * -# @ref CLStridedSlice Extract valid output.
55  * -# @ref CLArithmeticAddition Add bias.
56  * -# @ref CLActivationLayer Perform activation.
57  * -# @ref CLPermute Permute output if NHWC(only NCHW is supported).
58  */
60 {
61 public:
62  /** Default constructor */
63  CLFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
64  /** Prevent instances of this class from being copied (As this class contains pointers) */
66  /** Default move constructor */
68  /** Prevent instances of this class from being copied (As this class contains pointers) */
70  /** Default move assignment operator */
72  /** Set the input and output tensors.
73  *
74  * Valid data layouts:
75  * - All
76  *
77  * Valid data type configurations:
78  * |src |dst |
79  * |:------|:------|
80  * |F32 |F32 |
81  * |F16 |F16 |
82  *
83  * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
84  *
85  * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
86  * while every optional dimension from 4 and above represent a batch of inputs.
87  * Data types supported: F16/F32.
88  * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
89  * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
90  * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
91  * Data types supported: Same as @p input.
92  * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
93  * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
94  * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
95  * available which may introduce a drop of accuracy as well. Default is false
96  */
97  void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
98  const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
99  /** Set the input and output tensors.
100  *
101  * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
102  *
103  * @param[in] compile_context The compile context to be used.
104  * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
105  * while every optional dimension from 4 and above represent a batch of inputs.
106  * Data types supported: F16/F32.
107  * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
108  * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
109  * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
110  * Data types supported: Same as @p input.
111  * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
112  * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
113  * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
114  * available which may introduce a drop of accuracy as well. Default is false
115  */
116  void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
117  const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
118  /** Static function to check if given info will lead to a valid configuration of @ref CLFFTConvolutionLayer
119  *
120  * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
121  *
122  * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
123  * while every optional dimension from 4 and above represent a batch of inputs.
124  * Data types supported: F16/F32.
125  * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
126  * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
127  * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
128  * Data types supported: Same as @p input.
129  * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
130  * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
131  * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
132  * available which may introduce a drop of accuracy as well. Default is false
133  *
134  * @return a status
135  */
136  static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
137  const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
138 
139  // Inherited methods overridden:
140  void run() override;
141  void prepare() override;
142 
143 private:
144  MemoryGroup _memory_group;
145  CLReverse _flip_weights_func;
146  CLPermute _permute_input_func;
147  CLPermute _permute_output_func;
148  CLPermute _permute_weights_func;
149  CLPermute _permute_bias_func;
150  CLPadLayer _pad_input_func;
151  CLPadLayer _pad_weights_func;
152  CLFFT2D _transform_input_func;
153  std::unique_ptr<CLFFT2D> _transform_weights_func;
154  CLFFT2D _itransform_output_func;
156  CLReductionOperation _reduce_func;
157  CLSlice _extract_output_func;
158  CLArithmeticAddition _bias_add_func;
159  CLActivationLayer _activation_layer_func;
160 
161  CLTensor _permuted_input;
162  CLTensor _permuted_weights;
163  CLTensor _permuted_bias;
164  CLTensor _permuted_output;
165  CLTensor _padded_input;
166  CLTensor _padded_weights;
167  CLTensor _flip_axis;
168  CLTensor _flipped_weights;
169  CLTensor _transformed_input;
170  CLTensor _transformed_weights;
171  CLTensor _input_weights_product;
172  CLTensor _output_product;
173  CLTensor _output_reduced;
174  CLTensor _itransformed_output;
175  CLTensor _reshaped_output;
176  CLTensor _bias_output;
177 
178  const ICLTensor *_original_weights;
179  const ICLTensor *_original_bias;
180  bool _is_activationlayer_enabled;
181  bool _needs_permute;
182  bool _has_bias;
183  bool _is_prepared;
184 };
185 } // namespace arm_compute
186 #endif /* ARM_COMPUTE_CLFFTCONVOLUTIONLAYER_H */
Basic function to execute FFT-based convolution on OpenCL.
Basic function to run CLReverseKernel.
Definition: CLReverse.h:37
Base class for all functions.
Definition: IFunction.h:30
void run() override
Run the kernels contained in the function.
Basic function to pad a tensor.
Definition: CLPadLayer.h:44
Basic function to run opencl::ClComplexMul.
Basic function to run opencl::kernels::ClSaturatedArithmeticKernel for addition.
Store the tensor&#39;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.
Basic function to run opencl::kernels::ClActivationKernel.
Status class.
Definition: Error.h:52
Activation Layer Information class.
Definition: Types.h:1625
Basic function to perform tensor slicing.
Definition: CLSlice.h:38
Copyright (c) 2017-2022 Arm Limited.
Basic function to execute an opencl::kernels::ClPermuteKernel.
Definition: CLPermute.h:39
Padding and stride information class.
Definition: Types.h:669
CLCompileContext class.
Basic function to execute two dimensional FFT.
Definition: CLFFT2D.h:44
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
CLFFTConvolutionLayer & operator=(const CLFFTConvolutionLayer &)=delete
Prevent instances of this class from being copied (As this class contains pointers) ...
Perform reduction operation.
CLFFTConvolutionLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
void prepare() override
Prepare the function for executing.
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.
Basic implementation of the OpenCL tensor interface.
Definition: CLTensor.h:41