arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h

/*
 * Copyright (c) 2017-2018 ARM Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#ifndef __ARM_COMPUTE_NEGEMMWINOGRADLAYERKERNEL_H__
#define __ARM_COMPUTE_NEGEMMWINOGRADLAYERKERNEL_H__

#include "arm_compute/core/NEON/INEKernel.h"
#include "arm_compute/core/NEON/kernels/winograd/convolution.hpp"
#include "arm_compute/core/NEON/kernels/winograd/tensor.hpp"

namespace arm_compute
{
class ITensor;
class NEWinogradLayerKernel;

class Winograd3x3F32 final
{
public:
    friend class NEWinogradLayerKernel;
    Winograd3x3F32(
        const int          n_batches,         /** Number of batches in the input and output tensors. */
        const int          n_input_channels,  /** Number of feature maps in a batch of the input tensor. */
        const int          n_input_rows,      /** Number of rows in a feature map of the input tensor. */
        const int          n_input_cols,      /** Number of columns in a feature map of the input tensor. */
        const int          n_output_channels, /** Number of feature maps in the output tensor. */
        const bool         same_padding,      /** Use "SAME" padding, otherwise use "VALID". */
        const float *const weights,           /** Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. */
        float *const       weights_storage,   /** Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size`. */
        const float *const input,             /** Pointer to NHWC ordered input tensor, in the spatial domain. */
        float *const       winograd_input,    /** Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. */
        float *const       output,            /** Pointer to NHWC ordered output tensor, in the spatial domain. */
        float *const       winograd_output    /** Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. */
    );

    ~Winograd3x3F32();
    void transform_weights();
    void transform_input();
    void transform_output();

private:
    class Private;
    std::unique_ptr<Private> _pimpl;
};

class NEWinogradLayerKernel : public INEKernel
{
public:
    /** Constructor */
    NEWinogradLayerKernel();

    /** Prevent instances of this class from being copied (As this class contains pointers) */
    NEWinogradLayerKernel(const NEWinogradLayerKernel &) = delete;
    /** Prevent instances of this class from being copied (As this class contains pointers) */
    NEWinogradLayerKernel &operator=(const NEWinogradLayerKernel &) = delete;
    /** Allow instances of this class to be moved */
    NEWinogradLayerKernel(NEWinogradLayerKernel &&) = default;
    /** Allow instances of this class to be moved */
    NEWinogradLayerKernel &operator=(NEWinogradLayerKernel &&) = default;

    virtual ~NEWinogradLayerKernel() = default;

    /** Initialise the kernel
     *
     * @param[in] convolver A pointer to the winograd convolver, this object must have been configured and is ready to execute 16 GEMMS .
     */
    void configure(Winograd3x3F32 *convolver);

    // Inherited methods overridden:
    void run(const Window &window, const ThreadInfo &info) override;

    /* Get the memory required to instantiate a new Winograd operator.
       */
    static size_t get_weight_storage_size(
        const int n_output_channels, /** Number of output feature maps. */
        const int n_input_channels   /** Number of input feature maps. */
    );

    static unsigned int get_input_storage_size(
        const int  n_batches,   /** Number of batches in the input tensor. */
        const int  n_channels,  /** Number of feature maps in the input tensor. */
        const int  n_rows,      /** Number of rows in each feature map. */
        const int  n_cols,      /** Number of columns in each feature map. */
        const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */
    );

    /** Determine how much memory (in units of TOut) to allocate for the
     * (Winograd domain) output.
     */
    static unsigned int get_output_storage_size(
        const int  n_batches,         /** Number of batches in the output tensor. */
        const int  n_rows,            /** Number of rows in each feature map of the input tensor. */
        const int  n_cols,            /** Number of columns in each feature map of the input tensor. */
        const int  n_output_channels, /** Number of feature maps in the output tensor. */
        const bool same_padding       /** Use "SAME" padding, otherwise use "VALID". */
    );

protected:
    Winograd3x3F32 *_convolver;
};

} // namespace arm_compute
#endif /*__ARM_COMPUTE_NEGEMMWINOGRADLAYERKERNEL_H__*/