CudaTransposePoolingDeliverKernel.hpp

/*
 * CudaTransposePoolingDeliverKernel.hpp
 *
 *  Created on: Aug 16, 2016
 *      Author: pschultz
 */

#ifndef CUDATRANSPOSEPOOLINGDELIVERKERNEL_HPP_
#define CUDATRANSPOSEPOOLINGDELIVERKERNEL_HPP_

#include <cudakernels/CudaPoolingDeliverKernel.hpp>

namespace PVCuda {

class CudaTransposePoolingDeliverKernel : public CudaKernel {
  public:
   CudaTransposePoolingDeliverKernel(CudaDevice *inDevice);
   virtual ~CudaTransposePoolingDeliverKernel();
   void setArgs(
         PVLayerLoc const *preLoc,
         PVLayerLoc const *postLoc,
         PVLayerLoc const *origConnPreLoc,
         PVLayerLoc const *origConnPostLoc,
         int nxpPost,
         int nypPost,
         cudnnPoolingMode_t poolingMode,
         int multiplier,
         CudaBuffer *dataStoreBuffer,
         CudaBuffer *gSynBuffer,
         CudaBuffer *origConnDataStoreBuffer,
         CudaBuffer *origConnGSynBuffer,
         int channel);

  protected:
   virtual int do_run() override;
   int calcBorderExcess(
         int preRestricted,
         int postRestricted,
         int border,
         int patchSizePostPerspective);
   int calcManyScale(int preRestricted, int postRestricted);
   int calcStride(int preRestricted, int postRestricted);

   // Data members
  protected:
   PVLayerLoc const *mPreLoc                    = nullptr;
   PVLayerLoc const *mPostLoc                   = nullptr;
   int mBorderExcessX                           = 0;
   int mBorderExcessY                           = 0;
   cudnnPoolingMode_t mPoolingMode              = CUDNN_POOLING_MAX;
   float mMultiplier                            = 1.0f;
   cudnnPoolingDescriptor_t mPoolingDescriptor  = nullptr;
   cudnnTensorDescriptor_t mDataStoreDescriptor = nullptr;
   float *mDataStore                            = nullptr;
   CudaBuffer *mCudnnDataStore                  = nullptr;

   cudnnTensorDescriptor_t mGSynDescriptor = nullptr;
   float *mGSyn                            = nullptr;
   CudaBuffer *mCudnnGSyn                  = nullptr;

   PVLayerLoc const *mOrigConnPreLoc  = nullptr;
   PVLayerLoc const *mOrigConnPostLoc = nullptr;
   int mOrigConnBorderExcessX         = 0;
   int mOrigConnBorderExcessY         = 0;

   cudnnTensorDescriptor_t mOrigConnDataStoreDescriptor = nullptr;
   float *mOrigConnDataStore                            = nullptr;
   CudaBuffer *mCudnnOrigConnDataStore                  = nullptr;

   cudnnTensorDescriptor_t mOrigConnGSynDescriptor = nullptr;
   float *mOrigConnGSyn                            = nullptr;
   CudaBuffer *mCudnnOrigConnGSyn                  = nullptr;
}; // class CudaTransposePoolingDeliverKernel

} /* namespace PVCuda */

#endif /* CUDATRANSPOSEPOOLINGDELIVERKERNEL_HPP_ */