Add examples for SYCL Spark

[jasonsewall-intel]

We should consider this a draft, since I'd like your comments. It's useful to have this out here to enable discussion.

Most noteworthy is that the histogram example doesn't compile with the current intel/llvm SYCL implementation because we do not support array reductions (but see intel/llvm#5941)

[Pennycook]

With the change from #16 (comment), the code compiles using the functionality from intel/llvm@863383b.

It doesn't pass the correctness test, but I think that's because nothing here resets the histogram. The test expects histogram[8] == N, but the behavior I see is that histogram[8] == N * iteration. Adding property::reduction::initialize_to_identity{} fixes it, but I don't understand why this is necessary -- some of the other benchmarks have this, but not all of them.

@tomdeakin, could you please take a look? What do you think the correct fix is, here?

[tomdeakin]

@Pennycook All the original benchmarks set property::reduction::initialize_to_identity{} so if it's missing it's likely an oversight in the new benchmarks. That being said, it's up the benchmark to decide how it tests correctness, so if we want to preserve the initial value on each invocation, that's OK too - it just needs a comment in the header.

[Pennycook]

Thanks, @tomdeakin. I don't think I have permissions to modify Jason's branch, and I think he's lost access to his "-intel" GitHub account.... So, I've just added a suggested change to the line that I think needs to be updated.

Do you have permissions to accept the suggested changes?

[tomdeakin]

@Pennycook - think I have those changes in now.

We might want to merge this and #17 into a spark branch in the short term as both these PRs will likely break compiling the other models which don't implement these kernels (yet).

[jasonsewall-intel]

I still have access! I changed the email address before I left. But this looks good and I thank you for making the changes.

[tomdeakin]

All good to merge into a spark branch? I can then merge #17 into the same branch; Im predicting minor conflicts in the driver code but should be simple to fix.

[Pennycook]

All good to merge into a spark branch? I can then merge #17 into the same branch; Im predicting minor conflicts in the driver code but should be simple to fix.

Sounds good to me. Thanks again for doing this. And sorry, @jasonsewall-intel -- I don't know why I thought you'd lost access!