Cory Bloor

@yxsamliu might be able to answer your LLVM questions. Most of my knowledge of the ROCm LLVM fork comes from him.

> 1. obviously `amdfftw` doesn't build neatly, so maybe it's worth pinging @amd-toolchain-support and @cgmb about that The AMD CPU and GPU math library teams are entirely separate. I'm a...

I think this question has been answered. The Debian packaging for ROCm is fairly mature now, and it does not use the ROCm LLVM fork.

This was fixed in 0c443d12011da16a036057e0472ae59c68bc901f.

How about rocm-cmake only creates the check target if `PROJECT_IS_TOP_LEVEL` is true?

@kzhuravl, it's really comforting to hear that the mirror will be maintained for a while. I'm thrilled with the effort going into upstreaming the components that are tightly coupled to...

I appreciate the work you've done to enable gfx1031, but this is not the approach we're going to take for adding support. It just doesn't scale very well if we...

@littlewu2508, I've prepared some (experimental) patches that may help. There would also need to be a patch for rocBLAS to treat these ISAs as being equivalent (due to Tensile having...

@littlewu2508, I've improved the logic to handle ISAs that are supersets of other ISAs (e.g. gfx1011 is gfx1010 with additional instructions) and tested the patch set with gfx1030 code objects...

I've put together [a patch implementing a workaround that enables rocBLAS on more GPUs](https://salsa.debian.org/rocm-team/rocblas/-/blob/master/debian/patches/0012-expand-isa-compatibility.patch). This one is a bit uglier than the patches for the runtime. It is essentially the...