Antoine Pitrou

@mapleFU Can you update to the latest git main?

Here are the benchmark results here (Ubuntu 22.04, AMD Zen 2 CPU): ``` --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Non-regressions: (24) --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- benchmark baseline contender change % counters CompressionInputNonZeroCopyBenchmarkDirectRead/InputBytes:65536/PerReadBytes:65536 1.102 GiB/sec 1.186 GiB/sec 7.628 {'family_index':...

Slightly faster (see the percentages).

For the Docker-based builds (i.e. all Linux builds), you can take a look at https://arrow.apache.org/docs/developers/continuous_integration/docker.html

@ursabot please benchmark lang=R,Python

@ursabot please benchmark lang=Python

@ursabot please benchmark lang=R

Ran the benchmarks locally (gcc 12.3.0, AMD Ryzen 9 3900X). There are a number of improved benchmarks and also a couple sizable regressions: Details ``` -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Non-regressions: (368) -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- benchmark...

Ok, there's a lot of C++ to decipher here but I'd rather see an explanation of how this is different from what already exists. Apart from calling the primitive "gather"...

My intuition is that this is primarily avoiding the cost of `bit_util::SetBitsTo` for the no-nulls case, which is good, but does not require a wholesale refactor.