ROCm-CompilerSupport
ROCm-CompilerSupport copied to clipboard
The compiler support repository provides various Lightning Compiler related services.
Hello I am writing an FFT algorithm in OpenCL and have found a pretty nasty bug in the ROCm OpenCL implementation. The problem resolves around the following kernel it's `l2`...
Currently tests are always built, this PR allows the user to turn off tests through `cmake -DBUILD_TESTING=OFF`, which is pretty much a standardized cmake variable from CTest. The reason to...
### System information ``` ❯ inxi -GSC -xx System: Host: ernie Kernel: 5.7.7 x86_64 bits: 64 compiler: gcc v: 10.1.0 Desktop: N/A wm: kwin_x11 dm: SDDM Distro: Gentoo Base System...
Otherwise this wont find llvm when built independently.
Without these additional required dependencies, linking fails with errors such as: `undefined reference to llvm::errs()'` Originally reported at: https://bugs.gentoo.org/711006
I'm afraid this is going to be a bit vague. Essentially, I had the `comgr` tooling working as a compilation pipeline for Numba. I was forced to upgrade to ROCm...
I can't seem to find anything in the header file/docs about the layout of the metadata that can be queried on a relocatable object. From the test code there's bits...
I think exit status `AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT` ends up being aliased in the API logic. For example, specifying a nonsense action kind to `amd_comgr_do_action` returns `AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT`, but so does specifying a valid...
Seems like the `amd_comgr_language_t` enum has no entry for LLVM IR as a source language, I'm guessing an action of kind `AMD_COMGR_ACTION_COMPILE_SOURCE_TO_BC` fails due to this?