commit | 23950469da34d343253845246d0558a0731637d6 | [log] [tgz] |
---|---|---|
author | Stella Laurenzo <stellaraccident@gmail.com> | Fri Mar 15 22:23:56 2024 -0700 |
committer | GitHub <noreply@github.com> | Fri Mar 15 22:23:56 2024 -0700 |
tree | e9a8c9e4c74d3f9a9f9665c52523ef53760caf68 | |
parent | d05b4a16f826ca328f7e737ae473e2b581dc950f [diff] |
[rocm] Excises (almost) dependence on /opt/rocm from the compiler. (#16803) Almost all uses were for bitcode linking, which we have been bundling with the compiler for a long time. I left one ancillary use in the ukernel directory as that depends on some HIP headers. We can sever that dependency too, but will do that in a separate PR as it isn't related. Also takes this chance to: * Remove foot-guns around the use of `--iree-rocm-link-bc` flag by defaulting it to true. Afaict, it was only false because in the early phases it was hard to ensure that everyone had bitcode files available. * Remove all CMake plumbing that was setting `--iree-rocm-link-bc` or `--iree-rocm-bc-dir`, instead making the build system work with defaults for tests. * Revamped error handling logic in ROCMTarget, which was really quite awful. It was spewing errors to stderr and just barreling on in some situations. All failure messages have been moved to the point encountered and handled properly with an MLIR Location. * The above uncovered a bug where we were segfaulting on tests that were missing an ordinal attribute. These now properly error, and I fixed the tests. * Deletes the ROCM bazel lit tests since Bazel lacks the infrastructure to bundle bitcode files and will likely never get it. We prefer to have the build exercise default flag flows which precludes these tests being the same between these systems. I can't say I'm completely satisfied with the code quality of the ROCMTarget yet, and it is clear that this was assembled from a few different sources. However, the sloppy error handling was masking real crash opportunities and has now been fixed. --------- Co-authored-by: Scott Todd <scott.todd0@gmail.com>
IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.
See our website for project details, user guides, and instructions on building from source.
IREE is still in its early phase. We have settled down on the overarching infrastructure and are actively improving various software components as well as project logistics. It is still quite far from ready for everyday use and is made available without any support at the moment. With that said, we welcome any kind of feedback on any communication channels!
See our website for more information.
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.