GH-44072: [C++][Parquet] Add Float16 reading benchmarks

[pitrou]

Local benchmark numbers:

---------------------------------------------------------------------------------------------------------------------------
Benchmark                                                                 Time             CPU   Iterations UserCounters...
---------------------------------------------------------------------------------------------------------------------------
BM_ReadColumnPlain<false,Int32Type>/null_probability:-1            20038480 ns     20019703 ns           36 bytes_per_second=1.9512Gi/s items_per_second=523.772M/s
BM_ReadColumnPlain<true,Int32Type>/null_probability:0              37114403 ns     36766588 ns           19 bytes_per_second=1.06245Gi/s items_per_second=285.198M/s
BM_ReadColumnPlain<true,Int32Type>/null_probability:1              44589582 ns     44371707 ns           16 bytes_per_second=901.475Mi/s items_per_second=236.316M/s
BM_ReadColumnPlain<true,Int32Type>/null_probability:50             65624754 ns     65322683 ns           11 bytes_per_second=612.345Mi/s items_per_second=160.522M/s
BM_ReadColumnPlain<true,Int32Type>/null_probability:99             43072631 ns     42932582 ns           16 bytes_per_second=931.693Mi/s items_per_second=244.238M/s
BM_ReadColumnPlain<true,Int32Type>/null_probability:100            36710045 ns     36475141 ns           19 bytes_per_second=1.07093Gi/s items_per_second=287.477M/s

BM_ReadColumnPlain<false,Float16LogicalType>/null_probability:-1   52718868 ns     52616204 ns           12 bytes_per_second=380.111Mi/s items_per_second=199.288M/s
BM_ReadColumnPlain<true,Float16LogicalType>/null_probability:0     71273144 ns     71093105 ns           10 bytes_per_second=281.321Mi/s items_per_second=147.493M/s
BM_ReadColumnPlain<true,Float16LogicalType>/null_probability:1     80674727 ns     80358048 ns            8 bytes_per_second=248.886Mi/s items_per_second=130.488M/s
BM_ReadColumnPlain<true,Float16LogicalType>/null_probability:50   138249159 ns    137922632 ns            5 bytes_per_second=145.009Mi/s items_per_second=76.0264M/s
BM_ReadColumnPlain<true,Float16LogicalType>/null_probability:99    86938382 ns     86576176 ns            8 bytes_per_second=231.01Mi/s items_per_second=121.116M/s
BM_ReadColumnPlain<true,Float16LogicalType>/null_probability:100   74154244 ns     73984356 ns            9 bytes_per_second=270.327Mi/s items_per_second=141.729M/s

• GitHub Issue: [C++][Parquet] Add Float16 reading benchmarks #44072

[github-actions]

⚠️ GitHub issue #44072 has been automatically assigned in GitHub to PR creator.

[mapleFU]

I'm a bit curiousity why benchmarking int32, it's to comparing with Float32?

[pitrou]

Yes, and because the other benchmarks are dict-encoded.

[mapleFU]

From the benchmark result, the Float16 can be well optimized, I think this can be multiple times(maybe 4x) faster with some optimization ( maybe detect the FLBA length, enhance the builder, etc)

[pitrou]

Yes, I plan to work on #44072

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 4 benchmarking runs that have been run so far on merge-commit 9986b7b.

There were no benchmark performance regressions. 🎉

The full Conbench report has more details. It also includes information about 5 possible false positives for unstable benchmarks that are known to sometimes produce them.