chore: Swap ByteBuffer for BufferHandle more#5988
Conversation
|
It looks like BoolArray is not preferred for GPU, see e.g. cudf 
But we have ByteBool here |
a10y
force-pushed
the
aduffy/bufferhandle-more
branch
from
703206e to
9719e45
Compare
CodSpeed Performance ReportMerging this PR will degrade performance by 78.85%Comparing
|
| Mode | Benchmark | BASE |
HEAD |
Efficiency | |
|---|---|---|---|---|---|
| ❌ | WallTime | u16_FoR[100K] |
6.8 µs | 9.8 µs | -30.28% |
| ❌ | WallTime | u16_FoR[10K] |
6.2 µs | 29.5 µs | -78.85% |
| ⚡ | WallTime | u64_FoR[10K] |
12.9 µs | 8 µs | +61.25% |
| ❌ | WallTime | u8_FoR[100K] |
5.9 µs | 22.3 µs | -73.48% |
| ❌ | WallTime | u16_FoR[10M] |
7 µs | 10.9 µs | -35.62% |
| ❌ | WallTime | u8_FoR[1K] |
5.9 µs | 7.1 µs | -16.58% |
| ❌ | Simulation | canonical_into_non_nullable[(10000, 1, 0.1)] |
51.8 µs | 60.2 µs | -13.92% |
| ❌ | Simulation | canonical_into_non_nullable[(10000, 1, 0.0)] |
30.5 µs | 39 µs | -21.79% |
| ❌ | Simulation | canonical_into_non_nullable[(10000, 1, 0.01)] |
35.8 µs | 44.2 µs | -19.03% |
| ❌ | Simulation | canonical_into_nullable[(10000, 100, 0.0)] |
4.1 ms | 5.1 ms | -19.94% |
| ⚡ | Simulation | into_canonical_nullable[(10000, 10, 0.0)] |
537.1 µs | 452.3 µs | +18.74% |
| ⚡ | Simulation | into_canonical_nullable[(10000, 10, 0.1)] |
709.6 µs | 632.3 µs | +12.23% |
Footnotes
-
1254 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. ↩
|
Arrow also has a 1-byte boolean extension. I did wonder whether byte bools would be preferred on GPU since bit unpacking would be madness. We had the same conclusion for pipelined CPU code too in fact. 🤷 |
a10y
force-pushed
the
aduffy/bufferhandle-more
branch
from
9719e45 to
9ab0eee
Compare
Codecov Report❌ Patch coverage is ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
a10y
force-pushed
the
aduffy/bufferhandle-more
branch
from
9ab0eee to
6689d6a
Compare
and the only cure is more bufferhandle Signed-off-by: Andrew Duffy <andrew@a10y.dev>
a10y
force-pushed
the
aduffy/bufferhandle-more
branch
from
6689d6a to
db398b5
Compare
Signed-off-by: Andrew Duffy <andrew@a10y.dev>
3 participants
No description provided.