[FEA] Add libcudf any() and all() reductions

[harrism]

Is your feature request related to a problem? Please describe.

Currently cuDF Python implements any() and all() for a column by calling libcudf's min() and max() (see NVIDIA/thrust#1863). libcudf should provide any() and all() instead. They can be made faster and also give us better control over semantics.

Describe the solution you'd like
It would be faster to provide an optimized any() and all() that first check that the values are non-zero to get a boolean per element and then use bitwise reductions in CUDA for a faster reduction.

Describe alternatives you've considered
It currently works, so this is just a potential optimization.

[harrism]

@kovaltan FYI in case you have thoughts on this.

[kovaltan]

I agree to have any and all function in cudf for better control over semantics.
But I suspect about the potential optimization.

It would be faster to provide an optimized any() and all() that first check that the values are non-zero to get a boolean per element and then use bitwise reductions in CUDA for a faster reduction.

The current implementation of reduction has already output dtype option.
For non boolean column, if we set output_dtype = bool8 and do reduction min (for all), or max (for any),
the reducntion kernel cast a element into cudf::bool and then do reduction for rows.
The current implementation uses 1 byte atomicMin for cudf::bool reduction, there would be a room for optimization.
But in future, I'm thinking to re-implement reduction using null supported iterator given by PR NVIDIA/thrust#1833, then the implementation will use cub and it would be well optimised.

Btw, I've found a bug about all and any of integer column, and fixed it, I will file it later.

[harrism]

In my experience even CUB doesn't have an optimized 1-bit-per-boolean reduction (i.e. using __popc() or __syncthreads_and()/__syncthreads_or().

[rgsl888prabhu]

In my experience even CUB doesn't have an optimized 1-bit-per-boolean reduction (i.e. using __popc() or __syncthreads_and()/__syncthreads_or().

@harrism Why not use thrust::all_of and thrust::any_of ?

[jrhemstad]

In my experience even CUB doesn't have an optimized 1-bit-per-boolean reduction (i.e. using __popc() or __syncthreads_and()/__syncthreads_or().

@harrism Why not use thrust::all_of and thrust::any_of ?

thrust::all_of/any_of are slow and should not be used. See: https://github.com/thrust/thrust/issues/1016

[karthikeyann]

faster early out for any_of and all_of
https://github.com/thrust/thrust/issues/1016#issuecomment-542867313

[jrhemstad]

faster early out for any_of and all_of
thrust/thrust#1016 (comment)

@karthikeyann Your results in that issue are fishy/impossible.

[karthikeyann]

Yes. looked into it. it had a bug. updating it now.