Prefer IVF-PQ compute_similarity blockDim=512 on AMD (gfx1100)

[konkolchin]

Summary

• Prefer blockDim=512 for IVF-PQ compute_similarity launches on HIP/AMD after stock occupancy auto-shrink, which under-sizes blocks on RDNA3 (gfx1100).
• Measured on RX 7900 XTX (library IVF_PQ m=32, SIFT-1M): ~+21–26% search QPS at nprobe 8/16/32 vs stock heuristic, recall@10 unchanged; optional env HIPVS_IVF_PQ_BLOCK_THREADS still overrides when set.
• Mid-grid hipVS/cuVS library ratio improves from ~0.48–0.50× to ~0.60–0.62× on peer GPUs (not a claim of closing the full gap).

Context

• Validated in DXC/AMD PoC harness: https://github.com/konkolchin/ann-harness-amd (results/lib_bench/LAUNCH_KNOBS.md, FILLED_pq_m32.md).
• Change is HIP-only (__HIP_PLATFORM_AMD__) inside compute_similarity_select.

Test plan

• Library bench on gfx1100: INDEX_TYPE=IVF_PQ M=32 — stderr shows block_threads=512; nprobe=32 QPS ~270k vs ~217k stock
• Recall@10 ladder unchanged (~0.35 → ~0.73)
• CI / reviewers: build libcuvs for target AMD arch and run existing IVF-PQ gtests / neighbors benches if available

[kevinjosephamd]

Thanks for this contribution and for sharing the performance results, those improvements look promising. Would you mind removing the HIPVS_IVF_PQ_BLOCK_THREADS environment-variable handling? We’d prefer not to expose an environment variable for configuring this launch parameter. Instead, the preferred block size should be derived from the device’s reported wavefront size.
Could you also update the comments to describe the device-derived policy and remove the reference to the “patch 0005” override? Thanks again for working on this optimization.