[ROCm 5.0/OpenCL] Immediate crash

[FilipVaverka]

Upgrading to ROCm 5.0 seems to cause any OpenCL application (including clinfo distributed with ROCm) crash (segmentation fault) immediately on startup - see stack trace below.

I'm running latest OpenSUSE Tumbleweed (ROCm 5.0 packages from AMD repository) with

5.16.5-1-default #1 SMP PREEMPT Thu Feb 3 05:26:48 UTC 2022 (1af4009) x86_64 x86_64 x86_64 GNU/Linux

on

Lenovo Legion 5 (AMD Ryzen 7 4800H with Radeon Graphics)

gdb /opt/rocm/opencl/bin/clinfo stack trace:

Thread 1 "clinfo" received signal SIGSEGV, Segmentation fault.
0x00007fffefbe1c08 in rocr::image::ImageRuntime::GetImageInfoMaxDimension(hsa_agent_s, hsa_agent_info_t, void*) () from /opt/rocm-5.0.0/opencl/bin/../lib/../../lib/libhsa-runtime64.so.1
(gdb) bt
#0 0x00007fffefbe1c08 in rocr::image::ImageRuntime::GetImageInfoMaxDimension(hsa_agent_s, hsa_agent_info_t, void*) ()
from /opt/rocm-5.0.0/opencl/bin/../lib/../../lib/libhsa-runtime64.so.1
#1 0x00007fffefbe0d71 in rocr::image::hsa_amd_image_get_info_max_dim(hsa_agent_s, hsa_agent_info_t, void*) ()
from /opt/rocm-5.0.0/opencl/bin/../lib/../../lib/libhsa-runtime64.so.1
#2 0x00007fffefb22cb1 in rocr::AMD::GpuAgent::GetInfo(hsa_agent_info_t, void*) const () from /opt/rocm-5.0.0/opencl/bin/../lib/../../lib/libhsa-runtime64.so.1
#3 0x00007fffefb47d5f in rocr::HSA::hsa_agent_get_info(hsa_agent_s, hsa_agent_info_t, void*) ()
from /opt/rocm-5.0.0/opencl/bin/../lib/../../lib/libhsa-runtime64.so.1
#4 0x00007ffff7594fe7 in ?? ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libamdocl64.so
#5 0x00007ffff7596a08 in ?? ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libamdocl64.so
#6 0x00007ffff75972c1 in ?? ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libamdocl64.so
#7 0x00007ffff7560f0e in ?? ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libamdocl64.so
#8 0x00007ffff758ab46 in ?? ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libamdocl64.so
#9 0x00007ffff7549ab5 in ?? ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libamdocl64.so
#10 0x00007ffff78ea948 in __pthread_once_slow (once_control=0x7ffff78318c0,
init_routine=0x7ffff7c4b6c0 std::__once_proxy()) at pthread_once.c:117
#11 0x00007ffff7549c49 in clIcdGetPlatformIDsKHR ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libamdocl64.so
#12 0x00007ffff7dbae6d in khrIcdVendorAdd ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libOpenCL.so.1
#13 0x00007ffff7dbcce6 in khrIcdOsVendorsEnumerate ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libOpenCL.so.1
#14 0x00007ffff78ea948 in __pthread_once_slow (
once_control=0x7ffff7fbf0e8 ,
init_routine=0x7ffff7dbcb00 )
at pthread_once.c:117
#15 0x00007ffff7dbb3f1 in clGetPlatformIDs ()
from /opt/rocm-5.0.0/opencl/bin/../lib/libOpenCL.so.1
#16 0x000000000040c0bc in cl::Platform::get(std::vector<cl::Platform, std::alloc--Type for more, q to quit, c to continue without paging--
atorcl::Platform >*) ()
#17 0x00000000004022a0 in main ()

[langyuxf]

There is a workaround for this issue. Just override the gfx version.

$ HSA_OVERRIDE_GFX_VERSION=9.0.2 /opt/rocm/opencl/bin/clinfo

[fabiscafe]

@xfyucg Thanks for this. Do you know if this is a known problem and if there will be a fix for it in time? ROCm 5.0 is the first time I'm able to use OpenCL on AMD hardware on linux. That's a huge thing for me. :)

[langyuxf]

@fabiscafe Only Renoir and Cezanne series APU have this issue in ROCm 5.0. It should be fixed in a future release.

[ROCmSupport]

Hi @FilipVaverka
Thanks for reaching out.
Can you please share the GPU details for better understanding. Thank you.

[FilipVaverka]

Here is rocminfo.txt.

The GPU is gfx90c - Renoir iGPU integrated in the AMD Ryzen 7 4800H.
The workaround suggested by @xfyucg is working for me.

[dagrim]

Hello,
Does anyone know what is lacking in ROCm to support Renoir APU ? I mean, which components would I need to rebuild with additional gfx support to allow HSA / TensorFlow / Pytorch on those GPUs ?
Thanks !

[ROCmSupport]

Thanks @FilipVaverka
Integrated GPUs are not supported with ROCm officially. Thank you.
Yes, you can go with workarounds.

[Mushoz]

People that have gotten their iGPU to work with ROCm, are you able to use tensorflow at all? I am trying to use the workarounds presented here, but tensorflow-rocm refuses to run on my 4800u's integrated vega GPU.

[langyuxf]

People that have gotten their iGPU to work with ROCm, are you able to use tensorflow at all? I am trying to use the workarounds presented here, but tensorflow-rocm refuses to run on my 4800u's integrated vega GPU.

#1743