[SYCL][E2E] Re-enable copy_dynamic_size on CUDA

[npmiller]

I don't know what fixed it, but this is fixed on the sycl branch and in the latest nightlies.

Fixes #16968

[ayylol]

I don't know what fixed it, but this is fixed on the sycl branch and in the latest nightlies.

The issue manifested specifically when the tests were built for multiple triples, rather than just nvptx triple. I'm guessing those passes are likely only building nvptx triple?

If it still fails here we could do a more fine grained unsupported statement like UNSUPPORTED: cuda && !build-mode so they still run outside of the pre-commit where they pass. Or XFAIL: cuda && run-mode && !build-mode

[npmiller]

No, I did try it building with multiple triples, and I can reproduce the failure with an older nightly build, see:

March 3rd nightly:

$ clang++ -fsycl -fsycl-targets=nvidia_gpu_sm_80,spir64 copy_dynamic_size.cpp -o c
$ ./c
<CUDA>[ERROR]: 
UR CUDA ERROR:
	Value:           700
	Name:            CUDA_ERROR_ILLEGAL_ADDRESS
	Description:     an illegal memory access was encountered
	Function:        wait
	Source Location: /home/llvm/unified-runtime/source/adapters/cuda/event.cpp:134

terminate called after throwing an instance of 'sycl::_V1::exception'
  what():  Native API failed. Native API returns: 2147483646 (UR_RESULT_ERROR_UNKNOWN)
Aborted
$

March 6 nightly:

$ clang++ -fsycl -fsycl-targets=nvidia_gpu_sm_80,spir64 copy_dynamic_size.cpp -o c
$ ./c
$

So I do think it's actually fixed, thanks for checking!

It might have gotten fixed in upstream LLVM, since there was a pulldown merged recently.

[npmiller]

@intel/llvm-gatekeepers This is ready to merge

Only failure disabled in #17344 and tracked in #17212

Failed Tests (1):
  SYCL :: bindless_images/sampling_2D.cpp