Support nullable indices in boolean take kernel and some optimizations

[jhorstmann]

Which issue does this PR close?

Closes #2057.

Rationale for this change

What changes are included in this PR?

• Use the PrimitiveIter to iterate over indices, which makes it easier to handle nullable indices.
• Split up the kernel into separate logic for taking the values and the validity bitmap. This reduces branching in the inner loop and thus increases the performance.

Are there any user-facing changes?

no

Benchmark results compared to master

Gnuplot not found, using plotters backend
take bool 512           time:   [1.3921 us 1.3927 us 1.3933 us]                           
                        change: [+3.7963% +3.9074% +3.9940%] (p = 0.00 < 0.05)
                        Performance has regressed.

take bool 1024          time:   [3.6781 us 3.8022 us 3.9126 us]                            
                        change: [-33.173% -31.574% -30.047%] (p = 0.00 < 0.05)
                        Performance has improved.

take bool 4096          time:   [10.833 us 10.878 us 10.924 us]                            
                        change: [-67.171% -67.042% -66.917%] (p = 0.00 < 0.05)
                        Performance has improved.

take bool nulls 512     time:   [1.4058 us 1.4067 us 1.4077 us]                                 
                        change: [+4.1519% +4.3855% +4.6259%] (p = 0.00 < 0.05)
                        Performance has regressed.

take bool nulls 1024    time:   [2.5999 us 2.6011 us 2.6027 us]                                  
                        change: [-31.505% -31.418% -31.333%] (p = 0.00 < 0.05)
                        Performance has improved.

take bool nulls 4096    time:   [9.7339 us 9.7377 us 9.7420 us]                                  
                        change: [-54.586% -54.558% -54.531%] (p = 0.00 < 0.05)
                        Performance has improved.

There is a small regression for very small inputs, which will probably be alleviated by #1857. Larger inputs see a big improvement.

[tustvold]

This seems to lose the special case of no null values, I would have naively thought this would make a non-trivial performance difference for the very common case where there aren't any nulls?

[jhorstmann]

I was also a bit surprised, we are probably trading the bounds check for indices.value(i), which is avoided by the iterator, vs. the added null check. I can try adding another variant without nulls and benchmark that.

[tustvold]

Yeah, I basically wonder if it isn't showing up as a regression as separating out the null mask handling makes up for it. However, perhaps we can make it even faster by adding the special case back 😄

[jhorstmann]

I'm planning to look at iteration in general in #1857 and ideally convince the compiler to do the specialization itself.

[ursabot]

Benchmark runs are scheduled for baseline = 72dada6 and contender = a6379fe. a6379fe is a master commit associated with this PR. Results will be available as each benchmark for each run completes.
Conbench compare runs links:
[Skipped ⚠️ Benchmarking of arrow-rs-commits is not supported on ec2-t3-xlarge-us-east-2] ec2-t3-xlarge-us-east-2
[Skipped ⚠️ Benchmarking of arrow-rs-commits is not supported on test-mac-arm] test-mac-arm
[Skipped ⚠️ Benchmarking of arrow-rs-commits is not supported on ursa-i9-9960x] ursa-i9-9960x
[Skipped ⚠️ Benchmarking of arrow-rs-commits is not supported on ursa-thinkcentre-m75q] ursa-thinkcentre-m75q
Buildkite builds:
Supported benchmarks:
ec2-t3-xlarge-us-east-2: Supported benchmark langs: Python, R. Runs only benchmarks with cloud = True
test-mac-arm: Supported benchmark langs: C++, Python, R
ursa-i9-9960x: Supported benchmark langs: Python, R, JavaScript
ursa-thinkcentre-m75q: Supported benchmark langs: C++, Java