[sleeepyjack]

Description

Adds an opt-in FastL2Add flag (last template parameter of cuco::bloom_filter_policy, default false) that speeds up add for filters that fit in L2.

• Each key leaves one hash-selected word of its block's first 32-byte sector untouched (no atomic for it). Partial-sector atomic updates are processed more efficiently by recent GPUs.
• Under the flag, add_coop shuffles all keys of the tile before issuing the now-conditional atomics, avoiding a convergence check per shuffle.

Trade-offs: the bit pattern changes (filters with and without the flag are not interchangeable) and the false-positive rate rises slightly (0.1315% → 0.1374% at 16 bits/key, k = 8). With the flag off, the bloom kernels compile to SASS identical to dev.

Performance

RTX PRO 6000 Blackwell Max-Q (sm_120, 128 MB L2), CUDA 13.4, default add/contains benchmarks (int64 keys, 256-bit blocks, k = 8, 1e9 keys), with the policy's FastL2Add set locally for the comparison:

Filter size	add	contains
≤ 32 MB	+51% (52.5 → 79.6 G keys/s at 32 MB)	−1 to −2%
≥ 64 MB	same	same

Diffs below count changes within ±1% as SAME.

nvbench_compare: dev vs this PR with FastL2Add=0

bloom_filter_add_unique_size

FilterSizeMB	Ref Time	Cmp Time	%Diff	Status
1	18.835 ms	18.859 ms	0.12%	🔵 SAME
2	18.856 ms	18.897 ms	0.22%	🔵 SAME
4	18.912 ms	18.950 ms	0.20%	🔵 SAME
8	18.961 ms	18.987 ms	0.14%	🔵 SAME
16	18.988 ms	19.020 ms	0.17%	🔵 SAME
32	19.046 ms	19.046 ms	-0.00%	🔵 SAME
64	25.067 ms	25.070 ms	0.01%	🔵 SAME
128	113.420 ms	113.402 ms	-0.02%	🔵 SAME
256	185.779 ms	185.752 ms	-0.01%	🔵 SAME
512	209.515 ms	209.520 ms	0.00%	🔵 SAME
1024	217.887 ms	217.876 ms	-0.00%	🔵 SAME
2048	220.609 ms	220.610 ms	0.00%	🔵 SAME

bloom_filter_contains_unique_size

FilterSizeMB	Ref Time	Cmp Time	%Diff	Status
1	11.679 ms	11.633 ms	-0.40%	🔵 SAME
2	11.825 ms	11.811 ms	-0.11%	🔵 SAME
4	11.896 ms	11.900 ms	0.03%	🔵 SAME
8	11.990 ms	11.969 ms	-0.17%	🔵 SAME
16	12.011 ms	12.007 ms	-0.03%	🔵 SAME
32	12.021 ms	12.055 ms	0.28%	🔵 SAME
64	12.096 ms	12.121 ms	0.21%	🔵 SAME
128	25.799 ms	25.805 ms	0.02%	🔵 SAME
256	53.155 ms	53.157 ms	0.00%	🔵 SAME
512	60.922 ms	60.928 ms	0.01%	🔵 SAME
1024	64.044 ms	64.037 ms	-0.01%	🔵 SAME
2048	65.464 ms	65.458 ms	-0.01%	🔵 SAME

nvbench_compare: FastL2Add=0 vs FastL2Add=1

bloom_filter_add_unique_size

FilterSizeMB	Ref Time	Cmp Time	%Diff	Status
1	18.859 ms	12.306 ms	-34.75%	🟢 FAST
2	18.897 ms	12.363 ms	-34.58%	🟢 FAST
4	18.950 ms	12.398 ms	-34.57%	🟢 FAST
8	18.987 ms	12.464 ms	-34.36%	🟢 FAST
16	19.020 ms	12.500 ms	-34.28%	🟢 FAST
32	19.046 ms	12.561 ms	-34.05%	🟢 FAST
64	25.070 ms	25.028 ms	-0.17%	🔵 SAME
128	113.402 ms	113.312 ms	-0.08%	🔵 SAME
256	185.752 ms	185.902 ms	0.08%	🔵 SAME
512	209.520 ms	209.701 ms	0.09%	🔵 SAME
1024	217.876 ms	218.160 ms	0.13%	🔵 SAME
2048	220.610 ms	221.159 ms	0.25%	🔵 SAME

bloom_filter_contains_unique_size

FilterSizeMB	Ref Time	Cmp Time	%Diff	Status
1	11.633 ms	11.855 ms	1.91%	🔴 SLOW
2	11.811 ms	11.965 ms	1.30%	🔴 SLOW
4	11.900 ms	12.024 ms	1.04%	🔴 SLOW
8	11.969 ms	12.093 ms	1.04%	🔴 SLOW
16	12.007 ms	12.110 ms	0.86%	🔵 SAME
32	12.055 ms	12.148 ms	0.77%	🔵 SAME
64	12.121 ms	12.226 ms	0.86%	🔵 SAME
128	25.805 ms	25.800 ms	-0.02%	🔵 SAME
256	53.157 ms	53.154 ms	-0.01%	🔵 SAME
512	60.928 ms	60.922 ms	-0.01%	🔵 SAME
1024	64.037 ms	64.040 ms	0.00%	🔵 SAME
2048	65.458 ms	65.460 ms	0.00%	🔵 SAME

[sleeepyjack]