HIP error: invalid device function

[klausbu]

I am implementing rocalution into an application.

It works fine using the OpenMP backend but when I set the device number for the GPU using: export HIP_VISIBLE_DEVICES=0; export ROCR_VISIBLE_DEVICES=0 , I get the following HIP error message:

Number of CPU cores: 16
Host thread affinity policy - thread mapping on every core
Number of HIP devices in the system: 1
Matrix+Vectors loading:5.7e-05 sec
HIP error: invalid device function
File: /long_pathname_so_that_rpms_can_package_the_debug_info/src/extlibs/rocALUTION/src/base/hip/hip_vector.cpp; line: 845

My small helloGPU device query program works fine:

#include <hip/hip_runtime.h>
#include <cstdio>

__global__ void helloGpu(int gpuId) {
    printf("device %d Gpu thread %d\n", gpuId, (int)threadIdx.x);
}

int main(int agrc, const char * argv[]){
    int device = 0, gpuId = device;

    hipSetDevice(device);

    hipLaunchKernelGGL((helloGpu), dim3(1), dim3(64), 0, 0, gpuId);

    hipDeviceSynchronize();

    return 0;
}

What could be the problem?

EDIT: export HIP_VISIBLE_DEVICES=0; export ROCR_VISIBLE_DEVICES=0; HSA_OVERRIDE_GFX_VERSION=10.3.0

adding HSA_OVERRIDE_GFX_VERSION=10.3.0 fixes the problem.

Maybe rocalution (I am using 5.7.1 RHEL version) is not compiled to support my gfx1031 card?

[ntrost57]

Yes, officially, we do not support gfx1031. You can, however, modify the CMakeLists.txt in root directory and add gfx1031 to the build targets.