Hi Trupti, On Wed, 14 Jan 2026 01:22:16 +0530 Trupti <[email protected]> wrote: > The build failures of rocfft on ppc64el are caused by a crash in > clang/LLVM when compiling HIP code. > Specifically, compiling HIP client test sources triggers an internal > compiler error in the LLVM PowerPC backend during code generation. > This is an upstream toolchain issue and not a functional problem in > rocfft itself. > > The compiler crash has been reported upstream and is tracked in the LLVM > issue tracker: https://github.com/llvm/llvm-project/issues/168727 > > I tested a targeted workaround on ppc64el by disabling HIP client tests, > while keeping the core rocfft library unchanged. With this change, > rocfft builds and installs successfully on ppc64el. > > Given this, would it be acceptable to apply a targeted workaround that > disables HIP client tests on ppc64el, instead of removing the ppc64el > binaries at this stage?
The rocfft library uses run-time compilation for all its functionality. As such, problems in the compiler toolchain are likely to appear at run-time. I'd seen similar bugs long ago with LLVM 15 and they went away with LLVM 17 before returning with LLVM 21. We're also seeing compilation problems with rocblas on ppc64el. The compiler bugs on ppc64el seem to come and go, which makes supporting these big mathlibs on ppc64el quite difficult.
I will be retaining ppc64el in the HIP Runtime, so it shouldn't be very hard to build rocfft on ppc64el from source. There seems to be promising efforts for getting ROCm on ppc64el working [1], but I don't have trust in the compiler right now and I don't want to maintain a big library like rocfft for ppc64el until I can reasonably expect for compiler bugs to be fixed in a timely manner.
Sincerely, Cory Bloor [1]: https://www.fitzsim.org/blog/?p=797
