DPCPP-SYCL cannot find GPU target from Python (Pybind11)

[Jonas5]

I have a simple Pybind11-binding:

#include <CL/sycl.hpp>
#include <pybind11/pybind11.h>

int foo(int a)
{
sycl::queue q(sycl::gpu_selector_v);
sycl::buffer<int, 1> b(&a, sycl::range<1>(1));
q.submit([&](sycl::handler& cgh){
auto acc = b.get_access<sycl::access::mode::read_write>(cgh);
cgh.single_task<class foo>([=](){
acc[0] +=1;
});
});
//accessor
auto acc = b.get_access<sycl::access::mode::read>();
return acc[0];
}

PYBIND11_MODULE(test_binder, m) {
m.def("foo", &foo);
}

And a simple script to invoke:

import sys

binder_path = "..."
sys.path.append(binder_path)

from test_binder import *

print(foo(1))

The binder is built in a oneapi-sourced environment, and linked with:

-fsycl -sycl-std=2020 -std=c++20 -fsycl-unnamed-lambda -fPIC

(Cpp-executables built with the same flags are able to execute fine on the GPU.)

When running the `.py` script in the sourced environment, it works fine on the CPU. However, when openCL-devices are forced:

export ONEAPI_DEVICE_SELECTOR=opencl:*

I get the following error:

Traceback (most recent call last):
File "...py", line 8, in <module>
print(foo(1))
RuntimeError: No device of requested type available. Please check https://software.intel.com/content/www/us/en/develop/articles/intel-oneapi-dpcpp-system-requirements.html -1 (PI_ERROR_DEVICE_NOT_FOUND)

Is it possible to keep the OpenCL-device exposed through the compiled Python-binder?

(AOT-compiling with -fsycl-targets did not resolve the issue)

[NoorjahanSk_Intel]

Hi,

Thanks for posting in Intel Communities.

Could you please provide the detailed steps you have followed to build your code?

Also please provide the below details:

OS & Hardware details

The compiler version being used

Could you please let us know why you are using OpenCL devices?

Also please try with a level_zero device and let us know if you observe the same issue.

ONEAPI_DEVICE_SELECTOR=level_zero:gpu

Thanks & Regards,

Noorjahan.

[NoorjahanSk_Intel]

Hi,

As you want to close this thread as mentioned in internal notes, we are going ahead and closing this issue.

If you need any additional information, please submit a new question as this thread will no longer be monitored.

Thanks & Regards,

Noorjahan.