John Pennycook

I agree with @tomdeakin that you should open an issue against DPC++. It looks to me like [the implementation](https://github.com/intel/llvm/blob/sycl/sycl/plugins/cuda/pi_cuda.cpp#L2279) of `parallel_for` there could use more sophisticated heuristics to determine the...

> The extra SYCLUSM instructions are due to worse codegen and also from converting the global index to 64-bit (arguably semantically necessary in SYCL but not Cuda). If you know...

Sorry for the late response here. There was a lot of discussion about how to expose this functionality, and whether additional guarantees were required to make it useful. Proposed changes...

@rolandschulz and I just encountered this independently, and despite knowing about this issue it still took us a very long time to find my mistake. ```c++ #include #include #include using...

> Using preload("res://path/to/class.gd") feels very hack like to me. I agree, and I think recommending this style of programming in a warning message is not a good idea. > Maybe...

> Does the code actually needs fixing in the first place? The code you posted in your original comment works correctly. The engine correctly identifies the static function, and calls...

> I'd speculate that `sycl::ndi_item` and `sycl::sub_group` only make sense in the context of using `sycl::nd_range` submission. The extension (semantically) just eases the access to them but doesn't create them...

> I would still be very careful here. If range size is a prime number, I'd expect an `if` around the kernel body making the control flow non-convergent. Calling sub-group...

> @Pennycook, could you please comment on this? I'm wondering if it's possible to implement fused cooperative kernels by defining some limitation(s) for interactions between the two features. I don't...

Thanks for taking a look at this so quickly. I believe the best way to get icx and icpx is from https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html. There are a few different options there (direct...