We are planning to introduce lookahead/warpspeed based implementations to a few more algorithms. We already did this for cub::DeviceScan and cub::DeviceRunLengthEncode is on its way. These changes have shown to require significant modifications to the current tuning policies.
Since we just put all tuning policies out into the public for CCCL 3.5, but the release branch was not created yet, we should try to prepare the tuning policies for the lookahead extensions of all affected algorithms before CCCL 3.5 branches. This concerns:
We are planning to introduce lookahead/warpspeed based implementations to a few more algorithms. We already did this for
cub::DeviceScanandcub::DeviceRunLengthEncodeis on its way. These changes have shown to require significant modifications to the current tuning policies.Since we just put all tuning policies out into the public for CCCL 3.5, but the release branch was not created yet, we should try to prepare the tuning policies for the lookahead extensions of all affected algorithms before CCCL 3.5 branches. This concerns:
DeviceScan(already done during initial development)DeviceScan::ByKey: Prepare scan-by-key tuning policy for lookahead #10079DeviceReduce::ReduceByKey: Prepare reduce-by-key tuning policy for lookahead #10078DeviceSelect::If / Flagged / Unique: Prepare select and partition tuning policies for lookahead #10076DeviceSelect::UniqueByKey: Prepare select and partition tuning policies for lookahead #10076DevicePartition::If / Flagged(two-way): Prepare select and partition tuning policies for lookahead #10076DevicePartition::If(three-way): Prepare three-way partition tuning policies for lookahead #10077DeviceRunLengthEncode::Encode(already done in Restructure the RLE encode tuning policy #10028)DeviceRunLengthEncode::NonTrivialRuns: Prepare non-trivial-runs tuning policy for lookahead #10080DeviceMemcpy::BatchedandDeviceCopy::Batched: Prepare batched-copy tuning policy for lookahead #10081