Repository navigation
Experiment with explicit aggregate input ownership - #10179
connortsui20 wants to merge 6 commits into
Conversation
Merging this PR will degrade performance by 21.11%
|
| Mode | Benchmark | BASE |
HEAD |
Efficiency | |
|---|---|---|---|---|---|
| ❌ | Simulation | null_count_run_end[(10000, 1024, 0.01)] |
6.4 µs | 13.4 µs | -52.61% |
| ❌ | Simulation | null_count_run_end[(32000, 256, 0.01)] |
6.4 µs | 13.4 µs | -52.61% |
| ❌ | Simulation | null_count_run_end[(10000, 256, 0.01)] |
6.4 µs | 13.4 µs | -52.21% |
| ❌ | Simulation | null_count_run_end[(32000, 1024, 0.01)] |
6.4 µs | 13.4 µs | -52.01% |
| ❌ | Simulation | sparse_null_count |
62.6 µs | 99.8 µs | -37.3% |
| ❌ | Simulation | bitpacked_compress_u32 |
36.7 µs | 58.2 µs | -36.9% |
| ❌ | WallTime | mul_u64_nonnull_neon |
28.9 µs | 40 µs | -27.85% |
| ❌ | Simulation | for_compress_i32 |
95.5 µs | 130 µs | -26.56% |
| ❌ | Simulation | new_raw_prim_test_between[i32, 2048] |
67.6 µs | 91.3 µs | -25.94% |
| ❌ | Simulation | compress[(1000, 256)] |
98.1 µs | 123 µs | -20.2% |
| ❌ | Simulation | compress[(1000, 16)] |
104.4 µs | 127.3 µs | -17.95% |
| ❌ | Simulation | bitpacked_decompress_u32 |
51.1 µs | 62.1 µs | -17.73% |
| ❌ | Simulation | compress[(1000, 4)] |
118 µs | 140.4 µs | -16% |
| ❌ | Simulation | sequence_decompress_u32 |
35.8 µs | 42.6 µs | -15.93% |
| ❌ | WallTime | mul_i64_nonnull_neon |
32.8 µs | 39 µs | -15.9% |
| ❌ | WallTime | multiply_shapes_neon[(32768, PerRowPerRow)] |
32.8 µs | 39 µs | -15.85% |
| ❌ | Simulation | compress[(4000, 256)] |
131 µs | 154.8 µs | -15.4% |
| ❌ | Simulation | canonicalize_sparse_list[(1024, 17, 8)] |
459.2 µs | 539.1 µs | -14.82% |
| ❌ | Simulation | canonicalize_sparse_list[(512, 7, 4)] |
534.3 µs | 622.6 µs | -14.17% |
| ❌ | Simulation | compress[(4000, 16)] |
147.8 µs | 171 µs | -13.57% |
| ... | ... | ... | ... | ... | ... |
ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.
Tip
Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.
Comparing ct/aggregate-input-owner (34b84b9) with ct/aggregate-array-stats (142bbbc)2
Footnotes
-
503 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩
-
No successful run was found on
ct/aggregate-array-stats(f69c359) during the generation of this report, so e7865fd was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩
b732851 to
7b210bd
Compare
7d17dc1 to
6de2d57
Compare
Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
6de2d57 to
34b84b9
Compare
Archived ownership prototype while the experiment is simplified. The original implementation and benchmark evidence remain preserved.
Original PR description
Progress towards #10175
Summary
Repeated computation needs an owner that ties retained information to the input that produced it. This experiment introduces that lifetime explicitly so kernels can reuse aggregate results and producer guarantees while cache ownership remains an open question.
Depends on #10178. Array caching remains available for comparison.
Changes
ArrayInputowns an immutable array and aggregate store. Scoped execution supplies its facts to existing cast and take kernels, while producers and stable selections establish or transfer eligible bounds and ordering. Private constructor proofs remain separate from generic aggregate results.Typed retention keeps rich aggregate states available without scalar conversion, which the compressor experiment builds on. Array, input, and disabled modes exercise the same compute paths. Both stores remain allocated, so lookup bypass alone does not establish a memory saving.
API Changes
Adds the experimental input owner, eager execution methods, and scoped aggregate lookup. The owner has no encoding ID or file representation.