Reland #118503: [Offload] Introduce offload-tblgen and initial new API implementation #118614
Conversation
callumfare
force-pushed
the
offload_new_api
branch
from
1f08f40 to
b539898
Compare
|
Hi @jplehr, I've added a fix for the shared libs build, see the last commit. It's unfortunate that there's a dependency on I've also verified a standalone build, including the Let me know if there are any issues |
breaks build OMPT related? ld.lld: error: undefined symbol: llvm::omp::target::ompt::enableDeviceTracing(int) b877a533a6d1 (HEAD -> amd-staging) Revert "Reland llvm#118503: [Offload] Introduce offload-tblgen and initial new API implementation (llvm#118614)" Change-Id: I1f4cb8e06ca625c2a539348d897e5436d314fc0a
…w API implementation (llvm#118614) Reland llvm#118503. Added a fix for builds with `-DBUILD_SHARED_LIBS=ON` (see last commit). Otherwise the changes are identical. Disables until OMPT adapts to this. Previous discussions at the LLVM/Offload meeting have brought up the need for a new API for exposing the functionality of the plugins. This change introduces a very small subset of a new API, which is primarily for testing the offload tooling and demonstrating how a new API can fit into the existing code base without being too disruptive. Exact designs for these entry points and future additions can be worked out over time. The new API does however introduce the bare minimum functionality to implement device discovery for Unified Runtime and SYCL. This means that the `urinfo` and `sycl-ls` tools can be used on top of Offload. A (rough) implementation of a Unified Runtime adapter (aka plugin) for Offload is available [here](https://github.com/callumfare/unified-runtime/tree/offload_adapter). Our intention is to maintain this and use it to implement and test Offload API changes with SYCL. ```sh $ ninja LibomptUnitTests $ OFFLOAD_TRACE=1 ./offload/unittests/OffloadAPI/offload.unittests ``` * Only some of the available device info is exposed, and not all the possible device queries needed for SYCL are implemented by the plugins. A sensible next step would be to refactor and extend the existing device info queries in the plugins. The existing info queries are all strings, but the new API introduces the ability to return any arbitrary type. * It may be sensible at some point for the plugins to implement the new API directly, and the higher level code on top of it could be made generic, but this is more of a long-term possibility. Change-Id: I355ac2cc4122d8023a88bc7f427174e16dfa76b9
|
Something problematic about how offload-tblgen is set up is that all of offload will (typically) be built as a runtime, while offload-tblgen is a host binary and depends on a host library. The compiler configuration used for the host compilation and the runtime build may be substantially different. In our case it breaks because both compilers use different libstdc++ versions (more by accident than intention), but I'd expect this to also break if the built clang used for runtime builds is only configured for cross-compilation. |
|
Hi @nikic, for offload-tblgen I mostly copied how mlir-tblgen is set up but I didn't account for the differences in building the project as a runtime. I'll look into this asap but if you any pointers on how to resolve it I'd appreciate the help |
|
@nikic Do you see the same problem with building libomptarget (also part of the offload runtime)? I've had a look and it seems to also pull in the host-built I don't think the cross-compilation case is broken because |
Interesting question! It does seem to be in a similar situation (though not quite, as this is at least an actual runtime). I think the relevant difference ends up being that libomptarget can link libLLVM.so, while offload-tblgen (like any tblgen) links against libLLVMSupport.a statically.
Hm, I don't really see anything simple that can be done here. As far as I understand the entire offload build is going to be in a nested cmake invocation, so we can't really split it into a runtime and non-runtime part. Maybe there's just nothing that can be done here. (I've worked around the specific problem that we hit already...) |
Package the new libLLVMOffload.so and /usr/include/offload headers introduced in llvm/llvm-project#118614.
Reland #118503. Added a fix for builds with
-DBUILD_SHARED_LIBS=ON(see last commit). Otherwise the changes are identical.New API
Previous discussions at the LLVM/Offload meeting have brought up the need for a new API for exposing the functionality of the plugins. This change introduces a very small subset of a new API, which is primarily for testing the offload tooling and demonstrating how a new API can fit into the existing code base without being too disruptive. Exact designs for these entry points and future additions can be worked out over time.
The new API does however introduce the bare minimum functionality to implement device discovery for Unified Runtime and SYCL. This means that the
urinfoandsycl-lstools can be used on top of Offload. A (rough) implementation of a Unified Runtime adapter (aka plugin) for Offload is available here. Our intention is to maintain this and use it to implement and test Offload API changes with SYCL.Demoing the new API
# From the runtime build directory $ ninja LibomptUnitTests $ OFFLOAD_TRACE=1 ./offload/unittests/OffloadAPI/offload.unittestsOpen questions and future work