/*!
 * Copyright (c) 2017 by Contributors
 * \file convolution.cc
 * \brief
 * \author Bing Xu, Jun Wu
*/

#include "./convolution-inl.h"
#if MXNET_USE_MKL2017 == 1
#include <mkl_memory.h>
#include "./mkl/mkl_memory-inl.h"
#include "./mkl/mkl_convolution-inl.h"
#endif  // MXNET_USE_MKL2017
#if MXNET_USE_NNPACK == 1
#include "./nnpack/nnpack_convolution-inl.h"
#endif  // MXNET_USE_NNPACK

namespace mxnet {
namespace op {
DMLC_REGISTER_PARAMETER(ConvolutionParam);

template<>
Operator* CreateOp<cpu>(ConvolutionParam param, int dtype,
                        std::vector<TShape> *in_shape,
                        std::vector<TShape> *out_shape,
                        Context ctx) {
  Operator *op = NULL;
  // If 1D convolution, use MXNet implementation
  if (param.kernel.ndim() == 1) {
    MSHADOW_REAL_TYPE_SWITCH(dtype, DType, {
      op = new ConvolutionOp<cpu, DType>(param);
    })
    return op;
  }
#if MXNET_USE_MKL2017 == 1
  if ((param.dilate[0] == 1 && param.dilate[1] == 1)
      && param.kernel.ndim() == 2) {
    switch (dtype) {
    case mshadow::kFloat32:
      return new MKLConvolutionOp<cpu, float>(param);
    case mshadow::kFloat64:
      return new MKLConvolutionOp<cpu, double>(param);
    default:
      break;
    }
  }
  LOG(INFO) << MKLConvolutionOp<cpu, float>::getName() << " Skip MKL optimization";
#endif
#if MXNET_USE_NNPACK == 1
  const size_t batch_size = (*in_shape)[0][0];
  if ((param.dilate[0] == 1 && param.dilate[1] == 1)
      && param.kernel.ndim() == 2 && (!param.no_bias)
      && param.num_group == 1 && (batch_size == 1 ||
      ((batch_size > 1) && (param.stride[0] == 1) &&
      (param.stride[1] == 1)))) {
    switch (dtype) {
    case mshadow::kFloat32:
      return new NNPACKConvolutionOp<cpu, float>(param);
    default:
      break;
    }
  }
#endif
  MSHADOW_REAL_TYPE_SWITCH(dtype, DType, {
    op = new ConvolutionOp<cpu, DType>(param);
  })
  return op;
}

// DO_BIND_DISPATCH comes from operator_common.h
Operator *ConvolutionProp::CreateOperatorEx(Context ctx,
                                            std::vector<TShape> *in_shape,
                                            std::vector<int> *in_type) const {
  std::vector<TShape> out_shape, aux_shape;
  std::vector<int> out_type, aux_type;
  CHECK(InferType(in_type, &out_type, &aux_type));
  CHECK(InferShape(in_shape, &out_shape, &aux_shape));
  DO_BIND_DISPATCH(CreateOp, param_, (*in_type)[0], in_shape, &out_shape, ctx);
}

MXNET_REGISTER_OP_PROPERTY(Convolution, ConvolutionProp)
.describe(R"code(Compute *N*-D convolution on *(N+2)*-D input.

In the simplest 2-D convolution, given input data with shape *(batch_size,
channel, height, weight)*, the output is computed by

.. math::

   out[n,i,:,:] = bias[i] + \sum_{j=0}^{num\_filter} data[n,j,:,:] \star
   weight[i,j,:,:]

where :math:`\star` is the 2-D cross-correlation operator.

For general 2-D convolution, the shapes are

- **data**: *(batch_size, channel, height, weight)*
- **weight**: *(num_filter, channel, kernel[0], kernel[1])*
- **bias**: *(num_filter,)*
- **out**: *(batch_size, num_filter, out_height, out_weight)*.

Define::

  f(x,k,p,s,d) = floor((x+2*p-d*(k-1)-1)/s)+1

then we have::

  out_height=f(height, kernel[0], pad[0], stride[0], dilate[0])
  out_weight=f(weight, kernel[1], pad[1], stride[1], dilate[1])

If ``no_bias`` is set to be true, then the ``bias`` term is ignored.

The default data ``layout`` is *NCHW*, namely *(batch_size, channle, height,
weight)*. We can choose other layouts such as *NHWC*.

If ``num_group`` is larger than 1, denoted by *g*, then split the input ``data``
evenly into *g* parts along the channel axis, and also evenly split ``weight``
along the first dimension. Next compute the convolution on the *i*-th part of
the data with the *i*-th weight part. The output is obtained by concating all
the *g* results.

To perform 1-D convolution, simply use 2-D convolution but set the last axis
size to be 1 for both data and weight.

3-D convolution adds an additional depth dimension besides height and
weight. The shapes are

- **data**: *(batch_size, channel, depth, height, weight)*
- **weight**: *(num_filter, channel, kernel[0], kernel[1], kernel[2])*
- **bias**: *(num_filter,)*
- **out**: *(batch_size, num_filter, out_depth, out_height, out_weight)*.

Both ``weight`` and ``bias`` are learnable parameters.

There are other options to tune the performance.

- **cudnn_tune**: enable this option leads to higher startup time but may give
  faster speed. Options are

  - **off**: no tuning
  - **limited_workspace**:run test and pick the fastest algorithm that doesn't
    exceed workspace limit.
  - **fastest**: pick the fastest algorithm and ignore workspace limit.
  - **None** (default): the behavior is determined by environment variable
    ``MXNET_CUDNN_AUTOTUNE_DEFAULT``. 0 for off, 1 for limited workspace
    (default), 2 for fastest.

- **workspace**: A large number leads to more (GPU) memory usage but may improve
  the performance.

)code" ADD_FILELINE)
.add_argument("data", "ndarray-or-symbol", "Input data to the ConvolutionOp.")
.add_argument("weight", "ndarray-or-symbol", "Weight matrix.")
.add_argument("bias", "ndarray-or-symbol", "Bias parameter.")
.add_arguments(ConvolutionParam::__FIELDS__());

}  // namespace op
}  // namespace mxnet