Move FFI linking to the new crate torch-build #447
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
jerry73204
pushed a commit
to jerry73204/tch-nms
that referenced
this pull request
Jan 20, 2022
|
I added a demo crate tch-nms that integrates C++/CUDA code with tch-rs. |
jerry73204
force-pushed
the
torch-build
branch
from
dc418c0
to
f579eec
Compare
jerry73204
pushed a commit
to jerry73204/tch-nms
that referenced
this pull request
Jan 20, 2022
|
I made
|
The patch move most FFI linking code in `build.rs` to a separate
torch-build crate. It paves the way to support 3rd party C++/CUDA
extensions directly linking to libtorch. The patch refactors the
code into this hierarchy.
tch-rs
`-- src <-- torch-sys
`-- build.rs <-- torch-build
In this way, potential contributors can setup an extension crate in
this layout, where FFI linkage to custom C++/CUDA code is supported by
torch-build.
3rd_party_ext
|-- build.rs <-- torch-build
|-- src <-- torch-sys
`-- csrc <-- C++/CUDA code
The torch-build is designed to be idempotent. If there are more than one
crates depend on torch-build. The probing is always done at most once,
so that dependent crates see identical probing information.
The following improvements are done in torch-build compare to original torch-sys.
- Allow more control on building process through "download-libtorch" and
"parallel-build" Cargo features.
- Automatic CXX11 ABI probing.
- Rewrite CUDA shared library probing.
- Probe system device to compile correct arches.
- Probe AMD HIP, CUDA and separated CPU/CUDA libraries and link
accordingly.
- Expose public API to let users to obtain detailed probing information.
jerry73204
force-pushed
the
torch-build
branch
from
f579eec
to
e9928d7
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The work is an attempt to support third-party C++/CUDA extensions in compile time. The current
torch-sys/src/build.rsdesign runs all linking jobs fortorch-syscrate only, making it difficult to link extra C++/CUDA code. The patch moves FFI linking and probing utilities to a standalonetorch-buildcrate, supporting C++ ABI probing and CUDA so file probing. Thetorch-sysusestorch-buildto link libtorch. The patch refactors the dependency tree into this hierarchy.This way enables third party extensions to share the FFI linking code with
tch-rs. In compile time, the libtorch probing is done only once intorch-build. The downstreamtorch-sysand extension code links to libtorch a common interface.The torch-build is designed to be idempotent. The probing is always done at most once, so that dependent crates see identical probing information. Further improvements are done in this work.
accordingly.
The remaining work is Windows support. The PR is marked as a draft to enable discussion and see how Windows support can be done.