Building ROCm clr for a desktop Debian

[j-jasz]

I'm looking for way to build latest stable release of ROCm HIP runtime for amd64 Debian system.
I'm aware that the runtime is already packaged as rocm-hipamd in Debian repos.
The problem with it is that it is so old (ROCm 5.7) that it wont be enough for GPU rendering to work in most recent stable release of Blender.

I think a lot of people would appreciate if there was a quick way to compile only the needed clr binary: libamdhip64.so.x.x.xxxxx
As far as I know this one executable is all that is needed to Run Cycles rendering in Blender.

I would be very much fine if I could:

• download Docker image
• build stripped version of ROCm with just 1 binary that I mentioned that is compatible with my host machine
• copy the binary to /usr/lib/x86_64-linux-gnu/

Could this be something that is added as an official package?

[stellaraccident]

That is not far from how many within AMD who don't need the whole stack do it. However, there are some additional wrinkles:

1. The major SOVERSION has been increased twice since that vintage and the current version is no longer binary compatible.
2. CLR does depend on both ROCR-Runtime (pretty small), LLVM (big) and a small handful of other low level libraries. LLVM is the big one because CLR builds the RTC compiler and uses it internally for some functions.
3. I think there is still a CMake flag to not build the RTC support, but it is not supported and last I checked -- non functional.

I think what would be needed here, given the various vintages, is a source build of Blender atop a minimal build of ROCm. That minimal build isn't going to be as minimal as it was in 5.7 because of the LLVM dep. The minimal build you are looking for is roughly what -DTHEROCK_ENABLE_ALL=OFF -DTHEROCK_ENABLE_CORE=ON would yield (https://github.com/ROCm/TheRock?tab=readme-ov-file#optional-configuration-flags). That would be fairly easy to wrap a docker image around.

[j-jasz]

Thanks for quick response.

1. Before swapping libamdhip64.so or any other binary I would purge any traces of past ROCm including ROCm-specific shared libraries so that shouldn't be the issue I think.
2. I was aware of that, but thank you for clarification anyway. I assume that by LLVM you mean https://github.com/ROCm/llvm-project and not upstream LLVM? Having dedicated LLVM for ROCm from the perspective of frankensteining newer version of ROCm on top of Stable Debian release look more favorable as Debian often lags behind other distributions and can have outdated LLVM. So as long as the build process is documented well enough and automated where it makes sense I'm fine with compiling LLVM and other smaller dependencies too.

Too clarify - I'm not looking for running either ROCm nor Blender inside Docker. If anything - I'd use Docker only if the ROCm compilation with it is going to be easier than doing it on the host directly. I mentioned it because from what I understand it can provide an isolation - and that would be useful for cleaning purposes. I don't plan on recompiling ROCm frequently - maybe once every couple months when the new stable version is released or in a case that newer Blender version demands it. All of this should be identical or very similar for a lot of Debian desktop users that just wants to install drivers and forget about them.

Building Blender from source on generic Debian Linux should be pretty straight forward. All of the documentation should be here:
https://developer.blender.org/docs/handbook/building_blender/linux/#__tabbed_1_2
You can also use ccache and ninja or turn off in CMake building unnecessary GPU backends (CUDA/Optix, oneAPI, Metal):
https://developer.blender.org/docs/handbook/building_blender/options/
After the compilation the binary should be located in ./build_linux/bin/