Daniel Jünger

> So pattern_bits is equal to words_per_block which is equal to 8 in Parquet/Impala/etc, right? Maybe in the example that parameter should be on default_filter_policy::pattern_word() instead of parquet_filter_policy::pattern_word()? Whups, yep,...

The PR is ready for another round of reviews (thanks in advance!). I addressed most of the previous suggestions and implemented the discussed policy concept.

Forgot to post some final benchmark results: ``` # Benchmark Results ## bloom_filter_add_unique_size ### [0] NVIDIA H100 80GB HBM3 | Key | Hash | Word | WordsPerBlock | Distribution |...

> Convince CCCL to expose cuda::atomic_ref::compare_exchange_* for 16B types Discussion thread (NVIDIA internal): https://nvidia.slack.com/archives/CCP05T27R/p1721095033011529

Few more implementation details come to my mind: - When doing an `insert` with perfect hashing, we don't need to check the content of the slot first. Instead, we can...

Hi @amiralirz, Thank you for bringing this up! This indeed is a solid bug and a tricky one to solve, too. When the second k2 is inserted at the tombstone...

I just clarified the bug description a bit. Hope that's ok @amiralirz

Dissecting the issue a bit more, I can come up with two possible solutions that involve changing the way how `insert` handles tombstones. We currently try to reclaim previously erased...

I like the idea. We'll incorporate it into the test.

Hi @kevkrist, Thanks for raising this issue. I'm looking into it. I suspect two possible reasons where this bug could originate from: 1. I made some recent changes to the...