Andrey Alekseenko

> This is not currently possible. What do you do with those values once you have obtained them? > I'm optimizing register comsumption of SYCL code for higher performance running...

CMake's [`GNUInstallDirs`](https://cmake.org/cmake/help/latest/module/GNUInstallDirs.html) module seems to be a good solution here, but the `acpp` wrapper would also need to be modified to use the correct path for libraries (and, optionally, includes)....

> Could you please explain the reason of change from get_pointer() to get_multi_ptr in SYCL 2020? I was not involved in the decision, but my understanding is that `multi_ptr` interface...

> Just taking a shot in the dark, I wonder if export AMD_SERIALIZE_COPY=3 would prevent this. Thanks for the suggestion! We don't do any copies here, so `AMD_SERIALIZE_COPY=3` does not...

> Have you built AdaptiveCpp itself (i.e. the runtime library) with amdclang too With amdclang itself; on MI250X with libstdc++ from GCC 7, and on MI50 with libstd++ from GCC...

> This might not be easily possible. On CUDA or HIP, we can just assume that all discovered devices are NVIDIA or AMD respectively. Is `HIP_PLATFORM=nvidia` not supported? :thinking:

Cannot reproduce anymore with 19cd6144aaed64719475c4f4a0ee626bdbc428dc / Compute Runtime 23.35.

Nice progress for SSCP! As I understand, no work is planned for the multipass flow? > The conversions are still there (to remove those I think we'd have to take...

> Yes, I guess we could always add some macro that, if set, potentially adds llvm.assume calls, or changes datatypes inside nd_item. The `assume` calls, I think, can be added...

> The assume calls, I think, can be added unconditionally, no? Unlike changing the return type of API, they don't enforce anything, but, per "as-if" rule, the LLVM optimizer should...