ConvShared.h

#include <ATen/ATen.h>

#include <ATen/cudnn/cudnn-wrapper.h>
#include <ATen/cudnn/Descriptors.h>
#include <ATen/cudnn/Types.h>
#include <ATen/native/ConvUtils.h>

namespace at { namespace native {

// ---------------------------------------------------------------------
//
// Helper classes
//
// ---------------------------------------------------------------------

// This POD struct is used to let us easily compute hashes of the
// parameters
struct ConvolutionParams
{
  cudnnDataType_t dataType;
  int input_size[2 + max_dim];
  uint8_t input_dim;
  at::MemoryFormat memory_format;
  int weight_size[2 + max_dim];
  int padding[max_dim];
  int stride[max_dim];
  int dilation[max_dim];
  int64_t groups;
  bool deterministic;
  bool allow_tf32;
  // NB: transposed purposely omitted: transposed just swaps
  // forward and backward, so you can reuse the benchmark entry,
};

std::ostream& operator<<(std::ostream & out, const ConvolutionParams& params);

// NB: This can't be a constructor, because then ConvolutionParams
// would not be a POD anymore.
// TODO: Use TensorGeometry here instead of the entire Tensor, which we
// don't actually need.  (OTOH: We can always pass in
// grad_input/grad_output, so this is not very pressing)
void setConvolutionParams(
    ConvolutionParams* params,
    const at::Tensor& input, const at::Tensor& weight,
    IntArrayRef padding, IntArrayRef stride, IntArrayRef dilation,
    int64_t groups, bool deterministic, bool allow_tf32);

std::string repro_from_args(const ConvolutionParams& args);


// ---------------------------------------------------------------------
//
// Raw functions
//
// ---------------------------------------------------------------------

void raw_cudnn_convolution_forward_out(
    const Tensor& output, const Tensor& input, const Tensor& weight,
    IntArrayRef padding, IntArrayRef stride, IntArrayRef dilation, int64_t groups,
    bool benchmark, bool deterministic, bool allow_tf32);

void raw_cudnn_convolution_backward_input_out(
    const at::Tensor& grad_input,
    const at::Tensor& grad_output,
    const at::Tensor& weight,
    IntArrayRef padding, IntArrayRef stride, IntArrayRef dilation, int64_t groups,
    bool benchmark, bool deterministic, bool allow_tf32);

void raw_cudnn_convolution_backward_weight_out(
    const Tensor& grad_weight, const Tensor& grad_output, const Tensor& input,
    IntArrayRef padding, IntArrayRef stride, IntArrayRef dilation, int64_t groups,
    bool benchmark, bool deterministic, bool allow_tf32);

}}