adlucem/ffi-experimental

Code generation / FFI binding experiments using libtorch 1.0

Hasktorch 0.2 Libtorch FFI

Work on ffi bindings into the c++ libtorch library in preparation for 0.2 which targets the pytorch's post 1.0libtorch backend.

General approach is to use generated Declarations.yaml spec instead of header parsing for code generation.

Project Structure

• codegen/ - code generation, parses Declarations.yaml spec from pytorch and produces ffi/ contents
• deps/ - submodules for dependencies - libtorch, mklml, pytorch
• examples/ - high level example models (xor mlp, typed cnn)
• ffi/- low level FFI bindings to libtorch
• hasktorch/ - higher level user-facing library, calls into ffi/, used by examples/
• inline-c/ - submodule to inline-cpp fork used for C++ FFI
• spec/ - specification files used for codegen/

Getting dependencies

deps/ holds several external dependencies that are retrieved using the deps/get-deps.sh script.

This should be run prior to building

XOR MLP Example

The following steps should run the xor mlp example:

# Download libtorch-binary and other shared library dependencies
pushd deps
# For CPU
./get-deps.sh
# For CUDA-9
# ./get-deps.sh -a cu90
# For CUDA-10
# ./get-deps.sh -a cu100
popd

# Set shared library environment variables
source setenv

stack build examples

stack exec xor_mlp

Running code generation

Code generation is used to build low-level FFI functions.

Note that the code is already generated in this repo under ffi, running this is only needed if changes are being made to the code generation process.

To run:

stack build codegen
stack exec codegen-exe

To get CLI options:

stack exec codegen-exe -- --help

Additional Information

See the wiki for developer information.

Contributions

Contributions/PRs are welcome.