[SYCL][COMPAT] Add new file group_utils.hpp

[intwanghao]

Helper functions in this file is used to migrate CUB functions. The motivation of this PR is to reduce the gap between syclcompat namespace and dpct namespace.

[ziranzha]

LGTM

[JackAKirk]

Thanks for the contribution.

Please in general can you provide a commit message here #17263 (comment) which will aid reviews and provide git history for future development.

An appropriate message will depend on the content of the contribution.
For example in this particular PR I think it would be good to explain to reviewers:

• What motivates this contribution in general
• Is this targeting a specific device: It looks like all of these algorithms are completely shared memory based, whereas for most modern GPUs such group algorithms have optimal implementations involving subgroup level shuffles, irrespective of whether the group algorithms are in subgroup or workgroup scope (For example see the implementations beginning in https://github.com/intel/llvm/blob/sycl/sycl/include/sycl/group_algorithm.hpp). Presumably the target device is one which does not support shuffles natively? If not, why aren't shuffles used?

Also, all new functionality to intel/llvm requires in repo testing. For this case the appropriate location to add a test is sycl/test-e2e/syclcompat. Please can you add an appropriate test for these algorithms.

The above applies generally to #17264

[intwanghao]

Hi @JackAKirk, thanks for review. Helper functions in this file is used to migrate CUB functions. The motivation of this PR is to reduce the gap between syclcompat namespace and dpct namespace. The shared memory based implementation is initial version that ensure there are functionality mapping in sycl side can be used for migration. Also, the e2e test is added.

[JackAKirk]

Again, this e.g. maps to

llvm/sycl/include/sycl/group_algorithm.hpp

Line 552 in 68cacbf

shift_group_right(Group g, T x, typename Group::linear_id_type delta = 1) {

I won't go through all the cases. But I think most of these functions are already implemented in a better way that doesn't just rely on shared memory.

[intwanghao]

Same as above.

[JackAKirk]

Hi @JackAKirk, thanks for review. Helper functions in this file is used to migrate CUB functions. The motivation of this PR is to reduce the gap between syclcompat namespace and dpct namespace. The shared memory based implementation is initial version that ensure there are functionality mapping in sycl side can be used for migration. Also, the e2e test is added.

I see thanks.
We have been told to try to be more strict on contributions regarding performance. I've made some suggestions for how you could improve the performance of these functions.

[intwanghao]

Hi, @JackAKirk. Could you please help to review again? I have addressed the comments.

[JackAKirk]

Hi, @JackAKirk. Could you please help to review again? I have addressed the comments.

Thanks, this just needs docs added to https://github.com/intel/llvm/blob/sycl/sycl/doc/syclcompat/README.md then it should be good to go.

[intwanghao]

All passing, ready to merge. @intel/llvm-gatekeepers