IntelPython / dpnp / 41197227906

83%

push

github

Feature/Add radix select based top k kernels (#3079)

This PR improves the implementation of `top_k` in libtensor by adding
several kernels that use a radix select approach, instead of a merge
sort approach

There are variants to the kernels:
* Multiple rows processed per work group (best with short rows)
* One row per work-group
* Multiple work-groups per row

Local benchmarking shows a significant performance improvement (results:
https://github.com/IntelPython/dpnp/pull/3079#issuecomment-6505914307)

Also factors a number of radix-based utilities into a separate header
file, which can be reused for future implementations such as median and
partition functions

1604 of 2994 branches covered (53.57%)

Branch coverage included in aggregate %.

29435 of 34567 relevant lines covered (85.15%)

17687.2 hits per line