Unable to run cuda backend on Windows 11

[shingen75]

Able to build the clang with cuda backend with CUDA version 11.6 and 12.2 ,however, I am not able to run the compiled sample.
I have reported the same report earlier with a different hardware configuration and different Windows version. The bug report for that bug report is https://github.com/intel/llvm/issues/11058#issue-1881036178

NOTE : Compiling to target CPU works ( clang++ -fsycl simple-sycl-app.cpp -o simple-sycl-app.exe and the ONEAPI_DEVICE_SELECTOR is not set )

To Reproduce The simple-sycl-app.cpp described in the getting started document is used.

Compile and run command:

set PATH=%DPCPP_HOME%\llvm\build_cuda\install\bin;%PATH% set LIB=%DPCPP_HOME%\llvm\build_cuda\install\lib;%LIB% clang++ -fsycl -fsycl-targets=nvptx64-nvidia-cuda simple-sycl-app.cpp -o simple-sycl-app.exe set ONEAPI_DEVICE_SELECTOR=cuda:* set SYCL_PI_TRACE=-1 simple-sycl-app.exe

Result: ---> DLL_PROCESS_ATTACH pi_win_proxy_loader.dll ---> DLL_PROCESS_ATTACH pi_opencl.dll ---> DLL_PROCESS_ATTACH pi_level_zero.dll ---> DLL_PROCESS_ATTACH syclx.dll SYCL_PI_TRACE[all]: Check if plugin is present. Failed to load plugin: pi_cuda.dll SYCL_PI_TRACE[all]: Check if plugin is present. Failed to load plugin: pi_unified_runtime.dll

• OS: Windows 11 Pro version : 22H2
• CPU:13th Gen Intel(R) Core(TM) i7-13700K 3.40 GHz
• Target device and vendor: NVIDIA GeForce RTX 4070
• CUDA version : 11.6 and 12.2 both were attempted
• DPC++ version: clang version 18.0.0 (https://github.com/intel/llvm ed70e2b60c6c0e2db9d7acf4dac0ffed2d3832b8)
• Dependencies version: cmake version 3.20.21032501-MSVC_2 ninja version 1.10.2

Driver Version: 536.67 CUDA : NVIDIA CUDA 12.2.128 and 11.6 both attempted

Windows SDK -Version 10.0.19041.0

[blinkfrog]

I have exactly the same problem.

• OS: Windows 11 Home version : 22H2
• CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz 2.30 GHz
• Target device and vendor: NVIDIA GeForce RTX 3080 Laptop GPU Driver Version: 31.0.15.3713
• CUDA version: 12.2
• DPC++ version: clang version 18.0.0 (https://github.com/intel/llvm https://github.com/intel/llvm/commit/ca56d7900fbcdcfc4ecfa8fd4e2df906caa0cca5)

I also have Windows-version of oneAPI 2023.2 installed and integrated with Visual Studio - in case if there can be some probable conflict with it.

[blinkfrog]

I just wanted to note that I cloned and compiled newer version of DPC++ (https://github.com/intel/llvm/commit/8fe166fb4431645b6c7a26f4d88c0564e8d1d427) and it works with CUDA backend, although many SYCL-programs don't work, as I mentioned there: https://github.com/intel/llvm/issues/11740

[JackAKirk]

Please ensure you have a supported version of windows for cuda: as listed here: https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html#system-requirements

22H2 is not marked as supported but 22H2-SV2 is