[SYCL][CUDA] Make piextKernelSetArgMemObj setting an error message

[pgorlani]

This patch makes cuda_piextKernelSetArgMemObj setting an error message instead of std::terminate in case of the image format is not supported. This error message is encapsulated in an exception thrown by the RT.

This allows to continue the SYCL-CTS execution in case of tests using unsupported channel types, see #2119 (comment).

[bader]

I would like to hear what others think about pushing messages to std::cerr here instead of attaching it to the exception.

[pgorlani]

Yes, me too. I followed the same way of check_error, i.e.,

llvm/sycl/plugins/cuda/pi_cuda.cpp

Lines 140 to 157 in 1a8bb53

	cuGetErrorString(result, &errorString);
	std::stringstream ss;
	ss << "\nPI CUDA ERROR:"
	<< "\n\tValue: " << result
	<< "\n\tName: " << errorName
	<< "\n\tDescription: " << errorString
	<< "\n\tFunction: " << function << "\n\tSource Location: " << file
	<< ":" << line << "\n"
	<< std::endl;
	std::cerr << ss.str();
	}
	if (std::getenv("PI_CUDA_ABORT") != nullptr) {
	std::abort();
	}
	throw map_error(result);
	}

Note: In case, I could add the check of PI_CUDA_ABORT to pie::die.

[sergey-semenov]

Since we don't call std::terminate here anymore, I would definitely prefer attaching the message to the exception and leave it to the caller to handle however they see fit.

[pgorlani]

I realized I took the wrong approach modifying the pi::die function. Basically a function that std::terminate could be needed, also the other PIs have it.

In order to solve #2119 (comment), I made cuda_piextKernelSetArgMemObj returning an error and setting a message.
In such a way the RT will throw an exception with the specified message allowing SYCL-CTS to continue the test execution, i.e.,

../util/sycl_exceptions.h:252: warning:
  SYCL exception
  with category name: 'sycl'
  with code: 'runtime'
  with code raw value: '1'
  with code message: 'SYCL Error'
  with what: 'Native API failed. Native API returns: -996 (Function exists but
  address is not available)
  PI CUDA kernels only support images with channel types int32, uint32, float,
  and half.
   -996 (Function exists but address is not available)'

../util/test_base.cpp:28: FAILED:
explicitly with message:
  a SYCL exception was caught: Native API failed. Native API returns: -996
  (Function exists but address is not available)
  PI CUDA kernels only support images with channel types int32, uint32, float,
  and half.
   -996 (Function exists but address is not available)

[pgorlani]

Hi @bader, are you happy with the proposed solution?

[bader]

Hi @bader, are you happy with the proposed solution?

I'm okay with it.
@sergey-semenov, please, take a look.

[smaslov-intel]

[SYCL][CUDA] Make piextKernelSetArgMemObj throwing an exception

Please update comment as this is not throwing an exception now

[pgorlani]

[SYCL][CUDA] Make piextKernelSetArgMemObj throwing an exception

Please update comment as this is not throwing an exception now

@smaslov-intel Done.