runtime error when running Cuda backend for SYCL - returns CL_INVALID_VALUE -30 error

[varsha-madananth]

I built sycl toolchain with support for nvidia cuda. The app builds find on -fsycl-targets=nvptx64-nvidia-cuda-sycldevice. I get the following error when i run the exe. I have attached the compile commands and also gdb output. Any help debugging would be appreciated.

clinfo -l
Platform #0: Intel(R) FPGA Emulation Platform for OpenCL(TM)
-- Device #0: Intel(R) FPGA Emulation Device Platform #1: Intel(R) OpenCL -- Device #0: Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz
Platform #2: NVIDIA CUDA
`-- Device #0: GeForce GTX 1080

simplesyclapplogs.txt

Thanks,
Varsha

[bader]

@varsha-madananth, according to the log, your application used OpenCL back-end, which seems unintentional.
I suggest checking if SYCL_BE environment variable is set like this SYCL_BE=PI_CUDA.
Also it would be great if queue is created with device selector argument to make sure that application is using correct device.
More information how to build/run SYCL applications using CUDA back-end you can find here: https://github.com/codeplaysoftware/sycl-for-cuda/blob/cuda/sycl/doc/GetStartedWithSYCLCompiler.md.
Tagging @Ruyk, who can correct me if I got it wrong.

[varsha-madananth]

@bader : I have followed the instructions from https://github.com/codeplaysoftware/sycl-for-cuda/blob/cuda/sycl/doc/GetStartedWithSYCLCompiler.md.
Currently, the executable does not work with SYCL_BE=PI_CUDA not set as per the documentation.
If i do not set the variable, i get the following. This is how i know the env was set in the logs which was attached.

Build program log for 'GeForce GTX 1080':
:1:2: error: source file is not valid UTF-8
PU<U+0001>
^
:1:1: error: unknown type name 'P'
PU<U+0001>
^
:1:4: error: source file is not valid UTF-8
PU<U+0001>
^
:1:3: error: variable has address space that is not supported in program scope declaration
PU<U+0001>
^
:1:4: error: expected ';' after top level declarator
PU<U+0001>
^
;
0 (CL_SUCCESS)

However, there has been changes to the codebase yesterday which gives build errors while doing a make on sycl-for-cuda. I probably will have to open another issue for that. I will keep you posted.

Thanks,
Varsha

[Ruyk]

There is an issue on the default device selection introduced by the latest PI plugin changes. Basically, SYCL_BE is ignored when using default selection, and can return an opencl device when there is no SPIRV binary or viceversa. We are working on a fix for that on the cuda-dev branch. This also affects a number of LIT tests, in case you see failures there.

[Ruyk]

The CUDA backend has been merged, this problem should be solved now. Could you try again and let us know?

[jinz2014]

I also got the same error (sycl branch). Could you add the change to the sycl branch ?

https://github.com/codeplaysoftware/sycl-for-cuda/blob/cuda/sycl/doc/GetStartedWithSYCLCompiler.md

People follow the get started guide. Could you please sort out or clean up the guide so that people can focus on building the CUDA backend ? Currently, they are mixed together.

Thanks

[Ruyk]

#1293

[bader]

@varsha-madananth, does #1293 with clarified instructions fix your problem?

[jinz2014]

Yes. Thanks for improving the guide.

…

On Tue, Mar 17, 2020 at 5:54 AM Alexey Bader ***@***.***> wrote: @varsha-madananth <https://github.com/varsha-madananth>, does #1293 <#1293> with clarified instructions fix your problem? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1119 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB3QI4H3ZJZIJMMPFLILS73RH5JGNANCNFSM4KTOPRNA> .

[bader]

@Ruyk, 👍.