[Feat] Add pipelined pointer-return kernel for find* at high load factor

[rhdong]

Problem

find* (pointer-return find) shows a ~61% throughput cliff at λ=1.0, dropping from ~7.0 to ~2.7 B-KV/s, while value-copy find stays stable at ~3.9 B-KV/s. This is logically wrong — find* does strictly less work than find, so it should never be slower.

Root cause: find* dispatches to tlp_lookup_ptr_kernel_with_filter, a 1-thread-per-key kernel that scans bucket digests serially in 9 iterations of 16 digests each, relying on empty-slot early termination. At λ=1.0, all 128 slots are occupied, so every miss scans all 9 iterations instead of ~2.

Solution

Add lookup_ptr_kernel_with_pipeline — a stripped-down version of the value-copy lookup_kernel_with_io_pipeline_v1 that returns V* pointers instead of copying values.

• Thread model: 32 threads per key (cooperative group), 4 keys/block
• Pipeline: 3-stage double-buffered (vs 4-stage in value-copy):
  1. Digest prefetch (all 128 digests in 1 parallel step)
  2. Digest match + key prefetch
  3. Key verify + write pointer (no value copy)
• Dispatch: load_factor > 0.875 && max_bucket_size == 128 selects the pipelined kernel; otherwise the fast TLP kernel is used as before.

Benchmark (H100 NVL, pure HBM, dim=8, capacity=128M)

λ	find* before	find* after	find (ref)	Δ
0.50	6.977	6.942	3.910	-0.5%
0.75	5.954	5.963	3.881	+0.2%
1.00	2.688	4.366	3.929	+62%

find* at λ=1.0 is now 11% faster than find (was 32% slower). No regression at low load factors.

[github-actions]

Documentation preview

https://nvidia-merlin.github.io/HierarchicalKV/review/pr-251