* Double-dispatch copy.

In order to split ATen's CPU/CUDA code into two separate libraries
which don't require a build flag (AT_CUDA_ENABLED) to separate them,
we need to be able to split source files based on whether or not they
handle CPU functionality only, or also touch CUDA.  Copy poses a unique
challenge here, because the naive implementation involves writing
a matrix for all combinations of CPU/GPU in a single file.

This PR splits up Copy.cpp into CPUCopy.cpp and CUDACopy.cpp, respecting
the following matrix:

    to\from    CPU           CUDA
          +---------------------------
    CPU   | CPUCopy.cpp   CUDACopy.cpp
    CUDA  | CUDACopy.cpp  CUDACopy.cpp

When you run x.copy_(y) where x is CPU and y is CUDA, we do a second
virtual dispatch to copy_from(y, x) on y's type, so that we can get
from CPUCopy.cpp to CUDACopy.cpp

The new autogenerated code for CPU looks like this:

Tensor & CPUByteType::s_copy_(Tensor & dst, const Tensor & src, bool non_blocking) const {
  // code generated by copy_wrapper
  checked_cast_tensor<CPUByteTensor>(dst.pImpl, "dst", 0, false);
  switch (src.type().ID()) {
    case TypeID::CPUByte:
        THByteTensor_copyByte(static_cast<CPUByteTensor*>(dst.pImpl)->tensor, static_cast<CPUByteTensor*>(src.pImpl)->tensor);
        break;
    case TypeID::CPUChar:
        THByteTensor_copyChar(static_cast<CPUByteTensor*>(dst.pImpl)->tensor, static_cast<CPUCharTensor*>(src.pImpl)->tensor);
        break;
    ...
    default:
      return src.type().s_copy_from(src, dst, non_blocking);

Notice that the fall through goes to s_copy_from.  s_copy_from is like s_copy
but the arguments are reversed.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Lintfix and no-CUDA fix

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Fix compilation erorr.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* CR

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Double-dispatch copy.

In order to split ATen's CPU/CUDA code into two separate libraries
which don't require a build flag (AT_CUDA_ENABLED) to separate them,
we need to be able to split source files based on whether or not they
handle CPU functionality only, or also touch CUDA.  Copy poses a unique
challenge here, because the naive implementation involves writing
a matrix for all combinations of CPU/GPU in a single file.

This PR splits up Copy.cpp into CPUCopy.cpp and CUDACopy.cpp, respecting
the following matrix:

    to\from    CPU           CUDA
          +---------------------------
    CPU   | CPUCopy.cpp   CUDACopy.cpp
    CUDA  | CUDACopy.cpp  CUDACopy.cpp

When you run x.copy_(y) where x is CPU and y is CUDA, we do a second
virtual dispatch to copy_from(y, x) on y's type, so that we can get
from CPUCopy.cpp to CUDACopy.cpp

The new autogenerated code for CPU looks like this:

Tensor & CPUByteType::s_copy_(Tensor & dst, const Tensor & src, bool non_blocking) const {
  // code generated by copy_wrapper
  checked_cast_tensor<CPUByteTensor>(dst.pImpl, "dst", 0, false);
  switch (src.type().ID()) {
    case TypeID::CPUByte:
        THByteTensor_copyByte(static_cast<CPUByteTensor*>(dst.pImpl)->tensor, static_cast<CPUByteTensor*>(src.pImpl)->tensor);
        break;
    case TypeID::CPUChar:
        THByteTensor_copyChar(static_cast<CPUByteTensor*>(dst.pImpl)->tensor, static_cast<CPUCharTensor*>(src.pImpl)->tensor);
        break;
    ...
    default:
      return src.type().s_copy_from(src, dst, non_blocking);

Notice that the fall through goes to s_copy_from.  s_copy_from is like s_copy
but the arguments are reversed.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Lintfix and no-CUDA fix

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Fix compilation erorr.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* CR

Signed-off-by: Edward Z. Yang <ezyang@fb.com>