Re: ci.rocm: invalid device function

[Cordell Bloor <cgmb@slerp.xyz>]

Hi Samuel,

On 2025-07-22 08:52, Samuel Thibault wrote:
> [starpu][7ed4aa7b79e4][starpu_hip_report_error] Error: oops in hip_codelet (../../examples/basic_examples/block_hip.hip:46)... 98: invalid device function
>
> Usually this error appears when the code is compiled for a too recent
> architecture and run on an older one. [....]
>
> /usr/bin/hipcc ../../examples/basic_examples/block_hip.hip -c -o 
> basic_examples/block_hip.o -D__HIP_PLATFORM_HCC__= 
> -D__HIP_PLATFORM_AMD__= -I/usr/include -I/usr/lib/llvm-17/lib/clang/17 
> -L/usr/lib -DSTARPU_HIP_PLATFORM_AMD -g -I../include -I../../include/ 
> -I../src -I../../src/
>
> Do you have an idea of what could be wrong?

I believe the problem is that you have not specified your target GPU 
architectures. It is probably building for a the default of gfx906. My 
gfx906 test machine was sleeping, so I booted it up. You can see that it 
passes [1]. You will need to add `--offload-arch <gfxid1> --offload-arch 
<gfx2> ...`.

The values of `<gfx1> <gfx2> ...` can be provided by rocm-target-arch 
from pkg-rocm-tools (0.9.0~exp2).

Sincerely,
Cory Bloor

[1]: 
https://ci.rocm.debian.net/packages/s/starpu/unstable/amd64+gfx906/80699/