Experimental C binding for libtorch

Experimental C binding for libtorch C++ interface.

C binding is useful for rapid-prototyping with libtorch library, reduce linkage issue, embedding libtorch to other languages through ffi, etc.

This project implements minimal and partial(features required for our usecases) C binding for libtorch API. There is no plan to implement solid and full C binding for libtorch(We think pytorch will officially release C binding in the future).

Supported platform

• Linux
  • [x] x86-64
  • [ ] aarch64 should work
• Windows
  • [x] Visual Studio 2019
  • [ ] llvm-mingw
  • [ ] MinGW
• macOS(Limited support)
  • [x] Big Sur x86/arm64
• Optional: CUDA

Supported libtorch version

• libtorch 1.10.0 or later

Status

• [x] Minimal float32/float64 tensor.
• [ ] Consider CUDA(Acclerator) devices.

Requirements

• CMake 3.2 or later
• C11 + C++14 compiler
• (optional)CUDA and cuDNN

Build

Install pytorch/libtorch

Prepare libtorch libraries. You can install pytorch or download prebuilt libtorch package, or build it from source code. If your pytorch/libtorch build uses CUDA, you'll need to install CUDA SDK and cuDNN package also.

Linux

Note that pytorch packages from https://pytorch.org/get-started/locally/ (conda or pip package) contains libtorch headers and libraries, and these are compiled with gcc + pre-C++11 ABI for Linux platform.

Edit path to libtorch in scripts/bootstrap-linux.sh.

If your libtorch is compiled with pre-C++11 ABI(e.g. pytorch build), you need to specify _GLIBCXX_USE_CXX11_ABI=0 when you compile c-libtorch with clang.

Then,

$ ./scripts/bootstrap-linux.sh
$ cd build
$ make

Windows

Visual Studio 2019 is supported at the momenet.

Run vcsetup2019.bat.

Solution file will be generated in build directory.

To run unit test, you'll need to set PATH to pytorch/libtordh dlls(or copy dlls to your working directory).

macOS(Big Sur or later)

CPU build only.

Assume some dependencies(e.g. protobuf) are installed through macports or homebrew.

If you want to build c-libtorch with pytorch package, you may need to delete linking with libopenblas.dylib in TORCH_DIR/python3.9/site-packages/torch/share/cmake/Caffe2/Caffe2Targets.cmake

Then,

$ ./scripts/bootstrap-macos.sh
$ cd build
$ make

Example

#include "c_libtorch.h"

int main(int argc, char **argv) {

  int major, minor, patch;

  c_torch_version(&major, &minor, &patch);

  printf("version: %d.%d.%d\n", major, minor, patch);

  printf("CUDA: %d\n", c_torch_cuda_is_available());

  c_at_Tensor *tp = c_torch_ones_1d(9, c_torch_kFloat32);

  c_at_Tensor *ft = c_torch_fft_fft(tp, /* n */-1, C_TORCH_DEFAULT_DIM, c_torch_fft_norm_none);

  delete_c_at_Tensor(tp);
  delete_c_at_Tensor(ft);

  return 0;
}

Memory management

Currently c-libtorch does not have a policy for memory management. C struct simply wraps C++ object(Assume no detach(), clone() operation for Tensor).

Variable

C API has c_ prefix. C++ namespaces are concatenated with _

Example: at::Tensor -> c_at_Tensor

Function signature

C API has c_ prefix. C++ namespaces are concatenated with _

Example: torch::cuda::is_available -> c_torch_cuda_is_available

Implemented APIs

• [x] torch::version
• [x] torch::cuda::is_available
• [x] torch::ones
• [x] torch::zeros
• [ ] torch::eye
• [ ] torch::jit::load
• [x] torch::fft, torch::ifft

TODO

• [ ] TensorFlow C API style memory management.
• [ ] Better error handling.

License

c-libtorch itself is licensed under MIT license.

Third-party licenses

• libtorch, pytorch: BSD-like license.
• acutest: MIT license.