Correctly set the index value for __shf_up.

[jchlanda]

Please see https://www.khronos.org/registry/OpenCL/extensions/intel/cl_intel_subgroups.html for the details of the shuffles.

This was uncovered when writing libclc's Intel subgroup shuffles, which use the same built-in bpermute (https://github.com/intel/llvm/pull/4664/files) and was failing tests from llvm-test-suite (among others: https://github.com/intel/llvm-test-suite/blob/intel/SYCL/SubGroup/shuffle.hpp#L88).

[mangupta]

@jchlanda : Can you please rebase this change against develop branch?

[jchlanda]

@jchlanda : Can you please rebase this change against develop branch?

Sorry @mangupta , this might be a silly question, but where would this go on develop, I've seen https://github.com/ROCm-Developer-Tools/HIP/blob/develop/include/hip/device_functions.h#L29 dispatching to amd_device_funtions.h but that files doesn't seem to be checked in. I've grepped through the branch and can't find the definitions for the shuffles anywhere else.

[mangupta]

@jchlanda : https://github.com/ROCm-Developer-Tools/hipamd/blob/develop/include/hip/amd_detail/amd_device_functions.h is the file that you are looking for. Since the change needs to go to the HIPAMD repo, closing the PR here.

[jchlanda]

Thank you, added it in here: ROCm/hipamd#2