So far I got it to compile but I don't know how to test it.

I don't have a desktop system with AMD to test right now, my cards are in headless boxes for ML.

Maybe the reporter of #407646 can help test?

So far I got it to compile but I don't know how to test it.

I don't have a desktop system with AMD to test right now, my cards are in headless boxes for ML.

Maybe the reporter of #407646 can help test?

Blender-hip is still broken (#410748) for me so I can't check it with that.

So I thought maybe I should try running some of the hiprt examples with this: https://github.com/GPUOpen-LibrariesAndSDKs/HIPRTSDK

HIPRT build puts the library files in $out/bin. I moved them to $out/lib in postInstall.
With that change, I can compile the HIPRTSDK examples.
But I get a runtime error:

❯ ./dist/bin/Debug/03_custom_intersection64D
Orochi error: '�E' on line 66  in '../common/TutorialBase.cpp'.

❯ ./dist/bin/Debug/00_context_creation64D 
Orochi error: 'mv��' on line 29  in '../00_context_creation/main.cpp'

I'm seeing issues with the compiled kernels, too, have not tried the examples yet, but it seems like a lot of kernels aren't compiled (the filesizes produced of the hipfb files are way too small).

if you add the /lib subdirectory of the rocmPackages.clr derivation to LD_LIBRARY_PATH, the examples seem to work. Still working to see what the deal is with the sizes of those hipfb and bc files, how big are they for you?

Still working to see what the deal is with the sizes of those hipfb and bc files, how big are they for you?

Thanks for pointing out the file names. The cmake install didn't copy the bc file. It was next to some python scripts. copying it to the build output fixed the runtime issues that I had.

File sizes are around 1.1MBs with BITCODE=ON.
I also tried BAKE_COMPILED_KERNEL=ON. With that the .so file was around 1.5MBs but I couldn't get the examples to run.

./lib:
total 2800
-r--r--r-- 12 root root  609895 Jan  1  1970 hiprt02005_6.3_amd.hipfb
-r--r--r--  2 root root 1111019 Jan  1  1970 hiprt02005_6.3_amd_lib_linux.bc
-r-xr-xr-x 10 root root 1050176 Jan  1  1970 libhiprt0200564.so*
-r--r--r-- 12 root root   89922 Jan  1  1970 oro_compiled_kernels.hipfb

if you add the /lib subdirectory of the rocmPackages.clr derivation to LD_LIBRARY_PATH, the examples seem to work.

Adding clr/lib to library path solved that issue. 👍

hiprt ver.02005
Executing on 'AMD Radeon RX 6800 XT'
In file included from /tmp/nix-shell-47955-0/comgr-67b6cf/input/../common/TutorialKernels.h:23:
../common/Aabb.h:24:10: fatal error: 'hiprt/hiprt_types.h' file not found
   24 | #include <hiprt/hiprt_types.h>
      |          ^~~~~~~~~~~~~~~~~~~~~
1 error generated when compiling for gfx1030.

After adding the hiprt include directory:

hiprt ver.02005
Executing on 'AMD Radeon RX 6800 XT'
orortcLinkAddFile FAILED (error=2) loading file: /nix/store/xjn8gydp3nm8pg57lp9435xdgbd0i8q9-hiprt-2.5.a21e075.3/lib/hiprt02005_6.3_amd_lib_linux.bc
Orortc error: 'HIPRTC_ERROR_PROGRAM_CREATION_FAILURE' [ 2 ] on line 407 in '/build/source/hiprt/impl/Compiler.cpp'.HIPRT error: '2' on line 224  in '../common/TutorialBase.cpp'.

Now I just realized that bc file isn't there. 😅 /nix/store/xjn8gydp3nm8pg57lp9435xdgbd0i8q9-hiprt-2.5.a21e075.3/lib/hiprt02005_6.3_amd_lib_linux.bc

I ran the first 5 examples and the images are the same as the demo one, so I think it's working now.

I think our bitcode files are much smaller than for example arch and fedora because we're using some kind of newer offload bundler compression in our tree. Not sure. I got all examples working too. I just passed -DNO_ENCRYPT=ON to cmake to avoid rebuilding that tool (fedora also does this) needed for encryption, and to avoid the gcc nativeBuildInput. I also added a gpuTargets parameter, and a postPatch section that patches the python scripts to honor that setting. I can submit PR for that separately once this is merged though

I think it's probably better not to hardcode that filename in the install command in postInstall, since if we move to rocm 6.4.0, then this will start failing without anything having changed with hiprt itself. Also I think you should add a command to symlink libhiprt0200564.so as libhiprt64.so (probably also taking care to futureproof it so that renaming step doesn't need to be updated everytime they end up changing whatever "02005" signifies in that filename. Also I think you should put rocmPackages.clr in buildInputs, then you can remove the HIP_PATH cmake flag, since it'll pick it up automatically from the environment variable that rocmPackages.clr exports

I think it's probably better not to hardcode that filename in the install command in postInstall, since if we move to rocm 6.4.0, then this will start failing without anything having changed with hiprt itself. Also I think you should add a command to symlink libhiprt0200564.so as libhiprt64.so (probably also taking care to futureproof it so that renaming step doesn't need to be updated everytime they end up changing whatever "02005" signifies in that filename.

Sure 👍
Do you think a simple wildcard like this is enough?

    ln -sr $out/lib/libhiprt*64.so $out/lib/libhiprt64.so
    install -v -Dm644 ../scripts/bitcodes/hiprt*_amd_lib_linux.bc $out/lib/

Also I think you should put rocmPackages.clr in buildInputs, then you can remove the HIP_PATH cmake flag, since it'll pick it up automatically from the environment variable that rocmPackages.clr exports

I removed the HIP_PATH while keeping clr in nativeBuildInputs and it still worked. 🤔

Hi @justinkb
Did you try blender with hip-rt? I built it with hip-rt but it crashes when you switch the render engine.
but I suppose we should discuss this in a separate PR?

Hi.
Is there anything needed to merge this?

I think more properly clr should be listed in buildInputs as opposed to nativeBuildInputs, if I look at how other similar derivations specify their dependencies

I think more properly clr should be listed in buildInputs as opposed to nativeBuildInputs

I agree, it should be a runtime dependency :)

nixpkgs-review result

Generated using nixpkgs-review.

Command: nixpkgs-review pr 411736
Commit: 6efc5ad5e164368b6572443bda8a53cc1a2c4dbe

x86_64-linux

Hi.
Is there anything needed to merge this?

Approved, I don't have merge perms.

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/prs-ready-for-review/3032/5723

I don't have merge perms

You can apply to be a commiter or ping one you know/have interacted with.