[BUG] cudf::strings::replace_re can be 20X slower than a single thread on the CPU for some cases

[ttnghia]

Similar to #12694, we discovered a similar issue with cudf::strings::replace_re for some input data.

In particular, using the same data file as in #12694, Spark CPU (regexp_replace) can complete its job in around 600ms while the GPU version (which calls to cudf::strings::replace_re) on the same input takes around 11 seconds.

[ttnghia]

Example of reproducible code:

{
    ScopeTimer timer("replace_re");

    auto const pattern    = std::string("\\[\\{|\\}\\]");
    auto const regex_prog = cudf::strings::regex_program::create(
      pattern, cudf::strings::regex_flags::DEFAULT, cudf::strings::capture_groups::NON_CAPTURE);
    auto const replace = cudf::string_scalar("");
    cudf::strings::replace_re(strs_input, *regex_prog, replace);
}

Output:

replace_re: 5002.102684ms

[davidwendt]

Yes, long strings for any regex is a known issue. This is likely not going to be fixed.

[GregoryKimball]

I'll close this in favor of #13048, where we can track progress on long string performance.