[C++][Parquet] Optimize column reader by fusing definition level decoding with counting

[tennisleng]

Describe the enhancement requested

There is a TODO in cpp/src/parquet/column_reader.cc to fuse definition level decoding with counting for better performance. Currently, definition levels are decoded and then counted in separate passes.

Component(s)

C++

[Shockp]

I would like to take it. is it fine?. I can contribute again now that I have more time. @tennisleng.

@pitrou Not sure if you track this one as well. I don't want to disturb you.

[Shockp]

Also, I am reading the code to understand the problem and I arrived to this conclusion: Should count perhaps be fused with FindMinMax() instead of with decoding?

Something like this:

for (...) {
    min = ...;
    max = ...;
    matching += levels[i] == target;
}

So we would have:

decode/write
      ↓
validation + count

Rather than:

decode/write
      ↓
validation
      ↓
count

I will investigate both approaches. Not sure about your opinion.