Performance improvement for some libcudf regex functions for long strings

[davidwendt]

Description

Changes the internal regex logic to minimize character counting to help performance with longer strings. The improvement applies mainly to libcudf regex functions that return strings (i.e. extract, replace, split). The changes here also improve the internal device APIs for clarity to improve maintenance. The most significant change makes the position variables input-only and returning an optional pair to indicate a successful match.

There are some more optimizations that are possible here where character positions are passed back and forth that could be replaced with byte positions to further reduce counting. Initial measurements showed this noticeably slowed down small strings so more analysis is required before continuing this optimization.

Reference: #13048

More Detail

First, there is a change to some internal regex function signatures. Notable the reprog_device::find() and reprog_device::extract() member functions declared in cpp/src/strings/regex/regex.cuh that are used by all the libcudf regex functions. The in/out parameters are now input-only parameters (pass by value) and the return is an optional pair that includes the match result. Also, the begin parameter is now an iterator and the end parameter now has a default. This change requires updating all the definitions and uses of the find and extract member functions.

Using an iterator as the begin parameter allows for some optimizations in the calling code to minimize character counting that may be needed for processing multi-byte UTF-8 characters. Rather than using the cudf::string_view::byte_offset() member function to convert character positions to byte positions, an iterator can be incremented as we traverse through the string which helps reduce some character counting. So the changes here involve removing some calls to byte_offset() and incrementing (really moving) iterators with a pattern like itr += (new_pos - itr.position()); There is another PR #13428 to make a move_to iterator member function.

It is possible to reduce the character counting even more as mentioned above but further optimization requires some deeper analysis.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[harrism]

There are a lot of disparate changes here. I can't review them in detail, but I wonder what is the relationship between them all? Is there something common between the optimizations?

[davidwendt]

There are a lot of disparate changes here. I can't review them in detail, but I wonder what is the relationship between them all? Is there something common between the optimizations?

Yes. There are two main changes here. One is the change to the signatures of the two reprog_device member functions find() and extract() and this allows for the 2nd main change which is using the string_view::const_iterator more instead of the string_view::byte_offset() to help minimize character counting. I've also added more detail about both of these changes in the PR description under the More Detail heading.

[harrism]

Nice speedups!

[mythrocks]

I'm reviewing this right now. Please pardon my prior absence.

[mythrocks]

A couple of clarifications and nitpicks. Looks good to me.

[davidwendt]

/merge