CMake: When rocminfo is present, ask users to explicitly enable or disable ROCm testing.
#22478
Conversation
Signed-off-by: Benoit Jacob <[email protected]>
Signed-off-by: Benoit Jacob <[email protected]>
| # But silently continuing with ROCm testing disabled by default has also been | ||
| # surprising. | ||
| if ((IREE_HAL_DRIVER_HIP OR IREE_HAL_DRIVER_AMDGPU) AND NOT DEFINED CACHE{IREE_HIP_TEST_TARGET_CHIP}) | ||
| find_program(_ROCMINFO_COMMAND rocminfo) |
There was a problem hiding this comment.
You could try using offload-arch (formerly amdgpu-arch) instead, as that is more generic, available on Windows too, and produces more machine-friendly output
There was a problem hiding this comment.
I don't have offload-arch in my installation (rocm 7.0), and amdgpu-arch have llvm version suffixes, which would make it a bit tricky to select...
There was a problem hiding this comment.
Wait this is great but the only place I can see offload-arch is in my LLVM directory from a tarball. AFAICS, if I hadn't switched to a llvm tarball as my toolchain (e.g. if I were still using the system toolchain) I wouldn't have this program at all?
There was a problem hiding this comment.
There is a amdgpu-arch under /opt/rocm-7.0.x but not in PATH - it's in some LLVM toolchain subdir.
There was a problem hiding this comment.
Normally I'd be thrilled to switch to building offload-arch ourselves from the LLVM tree but here I need this information early in configure.
There was a problem hiding this comment.
Available on Windows too
Is rocminfo not available on Windows?
There was a problem hiding this comment.
Both have offload-arch / amdgpu-arch.
Windows has hipinfo.
Linux has rocminfo and rocm_agent_enumerator.
I'm still figuring out how to eventually turn that into something sane.
There was a problem hiding this comment.
No urgency on merging this .... this PR is just cherries on cakes.
One safe but kinda suboptimal resolution is to just build the offload-arch target from our llvm-project tree. Then give up on producing a configure-time diagnostic and instead produce a ctest-time diagnostic, which also removes the assumption that build host == test host.
|
@kuhar and I decided to just merge this as-is because however imperfect, it's better than the current state. When Anyway, can't merge at the moment due to CI breakage. |
3 participants
This is a leading cause of surprises and caveats for new users - it's something that is not naturally fitting in generic CMake instructions, but is important for people setting up for ROCm development.