HCC STATUS_CHECK Error: HSA_STATUS_ERROR_INCOMPATIBLE_ARGUMENTS about hcc HOT 10 CLOSED

Could you let me know your hardware configuration? By default HCC produces GCN ISA for Fiji ASIC with AMD:AMDGPU:8:0:3 ISA version string. The error message you saw are mostly related to incompatible GCN ISA version. You can use:

--amdgpu-target=AMD:AMDGPU:7:0:0 for Kaveri devices
--amdgpu-target=AMD:AMDGPU:7:0:1 for Hawaii devices
--amdgpu-target=AMD:AMDGPU:8:0:1 for Carrizo devices
--amdgpu-target=AMD:AMDGPU:8:0:3 for Fiji / Polaris devices

from hcc.

That's probably it, when I build it hcc/etc from source I use "AMD:AMDGPU:7:0:1" and it was fine with 1.3.1, haven't been able to finish building 1.4 from source due to #196

from hcc.

Linking with "--amdgpu-target=AMD:AMDGPU:7:0:1" has resolved this.

from hcc.

I'm running into the same issues while trying to use a machine with a Fiji and a Hawaii GPU. I tried to compile passing either or both --amdgpu-target=AMD:AMDGPU:7:0:1 --amdgpu-target=AMD:AMDGPU:8:0:3, but I still get the HSA_STATUS_ERROR_INCOMPATIBLE_ARGUMENTS error.

from hcc.

@pszi1ard may I understand the output of hcc --version in your installation? If it shows clang-3.5, then it does not support one single binary with multiple ISAs. It's only supported in Clang 4.0-based HCC, which should be included in the upcoming ROCm 1.5.

In case you are not afraid of building the compiler from source, you can find a useful script at:
https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/releases/tag/preview_1.0.17066
which helps you build HCC, HIP for that particular developer preview release.

from hcc.

Indeed, it's based on clang 3.5.

$ /opt/rocm/hcc/bin/hcc --version | head -n1
HCC clang version 3.5.0 (based on HCC 0.10.16501-81f0a2f-02246a0 LLVM 3.5.0svn)

However I still can't run any hcc program, even if I specify only one of the target. Is that because the binary ends up with code for only one target arch, but by default the code will execute on both detected devices? How can I restrict which devices can the runtime access? Does GPU_DEVICE_ORDINAL work?

from hcc.

@pszi1ard Given your system configuration ( Fiji + Hawaii ), I would recommend you to try build HIP & HCC from source, before we officially publish Clang 4.0 (or 5.0)-based HCC in the next ROCm release.

could you open up a new ticket and post relevant information there? we can continue the discussion over there.

from hcc.

I have a source build here change line 67 to target different GPU by default.

Personally I'd never mix GPUs in a system. didn't even know it was possible.

from hcc.

@pszi1ard Given your system configuration ( Fiji + Hawaii ), I would recommend you to try build HIP & HCC from source, before we officially publish Clang 4.0 (or 5.0)-based HCC in the next ROCm release.

I'm not sure I'll have the time, but I will report back if I do. Not sure what to open a ticket for, but here's an RFE for selecting devices using and env var: #249)/

I have a source build here change line 67 to target different GPU by default.

Thanks, I might try that!

Personally I'd never mix GPUs in a system. didn't even know it was possible.

Why? I have dev boxes with different gen devices exactly in order to facilitate testing. Not supporting such a setup is just counter-intuitive.

from hcc.

@pszi1ard You could also get a newer hcc installer package from the preview HIP release so you won't have to build hcc from source
https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/releases

from hcc.