ROCm 5.0 Segmentation fault

[sampie]

Hi

/opt/rocm/bin/rocminfo works just fine and shows GPUs, but when I try to make and run a simple helloworld program, I get a segmentation fault.

I have RDNA2 GPU (6800m). What could be the problem? I am not running kernel module from ROCm 5.0 package, but instead I am running vanilla Linux kernel 5.16.9. Has ROCm 5.0 kernel module been already incorporated to upstream Linux kernel?

--- The simple helloworld.cpp ---

#include
#include <hip/hip_runtime.h>

int main(int argc, char* argv[])
{

std::cout << "Before." << std::endl;

hipDeviceProp_t devProp;
hipGetDeviceProperties(&devProp, 0);

std::cout << "After." << std::endl;

return 0;
}

---

--- GDB showing the crash dump ---

Core was generated by `./helloworld'.
Program terminated with signal SIGSEGV, Segmentation fault.

warning: Section `.reg-xstate/887' in core file too small.
#0 0x00007f79e0257688 in rocr::image::ImageRuntime::GetImageInfoMaxDimension(hsa_agent_s, hsa_agent_info_t, void*) () from /opt/rocm-5.0.0/hip/lib/../../lib/libhsa-runtime64.so.1
[Current thread is 1 (Thread 0x7f79e00fdec0 (LWP 887))]
(gdb) bt
#0 0x00007f79e0257688 in rocr::image::ImageRuntime::GetImageInfoMaxDimension(hsa_agent_s, hsa_agent_info_t, void*) () from /opt/rocm-5.0.0/hip/lib/../../lib/libhsa-runtime64.so.1
#1 0x00007f79e02567b1 in rocr::image::hsa_amd_image_get_info_max_dim(hsa_agent_s, hsa_agent_info_t, void*) () from /opt/rocm-5.0.0/hip/lib/../../lib/libhsa-runtime64.so.1
#2 0x00007f79e0194679 in rocr::AMD::GpuAgent::GetInfo(hsa_agent_info_t, void*) const () from /opt/rocm-5.0.0/hip/lib/../../lib/libhsa-runtime64.so.1
#3 0x00007f79e01ba0ff in rocr::HSA::hsa_agent_get_info(hsa_agent_s, hsa_agent_info_t, void*) () from /opt/rocm-5.0.0/hip/lib/../../lib/libhsa-runtime64.so.1
#4 0x00007f79e0c45bb6 in ?? () from /opt/rocm-5.0.0/hip/lib/libamdhip64.so.5
#5 0x00007f79e0c4738d in ?? () from /opt/rocm-5.0.0/hip/lib/libamdhip64.so.5
#6 0x00007f79e0c47c79 in ?? () from /opt/rocm-5.0.0/hip/lib/libamdhip64.so.5
#7 0x00007f79e0c02fde in ?? () from /opt/rocm-5.0.0/hip/lib/libamdhip64.so.5
#8 0x00007f79e0c3af56 in ?? () from /opt/rocm-5.0.0/hip/lib/libamdhip64.so.5
#9 0x00007f79e0a7bdfe in ?? () from /opt/rocm-5.0.0/hip/lib/libamdhip64.so.5
#10 0x00007f79e0a9c365 in hipGetDeviceProperties () from /opt/rocm-5.0.0/hip/lib/libamdhip64.so.5
#11 0x0000000000201b09 in main (argc=, argv=) at helloworld.cpp:10

[langyuxf]

Is your configuration APU(Cezanne or Renoir) + RX6800M?
Cezanne and Renoir series APU have problems in ROCm 5.0 release.

For your issue, if you want to run hip application on RX6800M.
First, identify RX6800M deviceIndex, either 0 or 1.

Assume it is 0. Then, run like this

hipDeviceProp_t devProp;
hipGetDeviceProperties(&devProp, 0);

$ ROCR_VISIBLE_DEVICES=0 ./helloworld

[sampie]

Hi,

Yes, I have a laptop with 5900HX + 6800M.

Thank you! Running with ROCR_VISIBLE_DEVICES=0 worked!

BTW: Apparently it is 0, but how can I identify 6800M deviceIndex? Is it somewhere visible in rocminfo output (I did not see such field)?

[langyuxf]

You can use following fields in rocminfo.

Node: 1
Device Type: GPU

If Device Type is GPU, then device index should be Node-1.
You can confirm it with following code.
Check its name/gcnArchName/pciDeviceID, etc.

hipDeviceProp_t props;
hipGetDeviceProperties(&props, deviceIndex);

std::cout << props.name << std::endl;
std::cout << props.gcnArchName<< std::endl;
std::cout << props.pciDeviceID<< std::endl;

[ROCmSupport]

Thanks @sampie for reaching out.
Looks like your issue is resolved with the suggestions given by @xfyucg.
Please note that ROCm does not work on integrated GPUs and APUs. So always try on discrete GPUs by mapping the corresponding node.
Thank you.

[sampie]

Yes, the main issue is resolved and 6800M seems to work under rocm with all the code that I have run with it so far.

However, I guess, if possible, it would be good to try to fix rocm in such a way that it would not crash if ROCR_VISIBLE_DEVICES is not set. It looks like rocminfo works just fine, so I wonder why hipGetDeviceProperties does crash.