Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

There are substantial SASS diffs, for example in

Function : void cub::_V_300400_SM_750_800_860_900_1000::detail::rle::DeviceRleSweepKernel<cub::_V_300400_SM_750_800_860_900_1000::detail::rle::non_trivial_runs::policy_selector_from_types<long, cuda::std::__4::complex<float>>, cuda::std::__4::complex<float> const*, long*, long*, long*, cub::_V_300400_SM_750_800_860_900_1000::ReduceByKeyScanTileState<long, int, true>, cuda::std::__4::equal_to<void>, int, long, cub::_V_300400_SM_750_800_860_900_1000::detail::rle::streaming_context<cuda::std::__4::complex<float> const*, long, long>>(cuda::std::__4::complex<float> const*, long*, long*, long*, cub::_V_300400_SM_750_800_860_900_1000::ReduceByKeyScanTileState<long, int, true>, cuda::std::__4::equal_to<void>, int, int, cub::_V_300400_SM_750_800_860_900_1000::detail::rle::streaming_context<cuda::std::__4::complex<float> const*, long, long>)

on SM75.

There are substantial SASS diffs, for example in [...] on SM75.

With the change from #7733, I asked Cursor:

The last two commits lead to SASS differences on the benchmark cub.bench.run_length_encode.non_trivial_runs.base for SM75. This was likely caused by the rewrite of the tuning information and dispatch layer of DeviceRleDispatch. Please find out why the SASS changed and fix the new tuning and dispatch code so the SASS change is gone. You are not allowed to revert to the old code.

and it indeed found the root cause. I am impressed.

🥳 CI Workflow Results

🟩 Finished in 21h 20m: Pass: 100%/249 | Total: 9d 05h | Max: 3h 57m | Hits: 71%/153868

See results here.