Anton Smirnov
Anton Smirnov
Tests fail because of incompatibility with GPUCompiler 1.3. Are you on latests commit? Otherwise there's already a fix, but it will be merged in a bit
MI300x requires LLVM 18, so you can try with Julia 1.12, but I haven't tested AMDGPU with 1.12 at all yet.
Hm... Interesting. I cannot reproduce the error with MWE that you provided (my machine also runs as part of the CI). Can you maybe output the array before `cumsum` to...
Ah... My bad, the error is not related to `cumsum` because they are asynchronous and are checked before every kernel launch, so the fact that it errors before calling `cumsum`...
Still cannot reproduce... If you can print out the values maybe that will help
I see latest KomaMRI AMDGPU tests pass, so I assume this is no longer a problem. Feel free to re-open if needed.
What OS are you on? Is this an official build of ROCm? I'd take a look at output of [libtree](https://github.com/haampie/libtree) on `libMIOpen.so`: ``` pxl-th@Tower:~$ libtree /opt/rocm-6.2.2/lib/libMIOpen.so libMIOpen.so.1 ├── libhiprtc.so.6 [runpath]...
Hi, @laochailan. Can you try moving: ```julia global libMIOpen_path = get_library(lib_prefix * "MIOpen"; rocm_path) ``` before line: ```julia global libhsaruntime = if Sys.islinux() get_library("libhsa-runtime64"; rocm_path, ext="so.1") else "" end ```...
Are these segfaults reproducible with C code? E.g. creating rocBLAS handle and doing matmul
Can you try running these scripts with `HSA_OVERRIDE_GFX_VERSION=11.0.0` env variable set?