Expose conditional join size calculation

[vyasr]

Resolves #8918 by providing a new API for getting the output size for conditional joins (except full joins). This PR removes the unnecessary conditional_join.cuh header and inlines the logic into the conditional_join.cu file where it is used, and adds the new logic into that file as well. The public APIs are now exposed in conditional_join.hpp.

Adding the join size calculation also revealed a couple of bugs in the conditional join tests that were hiding a real bug in the conditional join implementation. The main test bug was the use of std::equal with the actual result as the first iterator, so if the actual result was empty it was never compared against the expected result (even if it was nonempty). This bug masked a couple of minor errors in the expected outputs encoded in the test. These are now fixed. This bug was also hiding a deeper issue where the AST device code was always using the left row index to pull data for the left hand operand to binary operations, even when the lhs was actually a column from the right table. That bug is now fixed as well.

Contributes to #8145.

[vyasr]

@jlowe it looks like the hash join APIs for the output size are all detail APIs, so I assume that the same is fine here but let me know if that's not the case and I can plumb this up higher in the hierarchy.

[jlowe]

it looks like the hash join APIs for the output size are all detail APIs

The APIs are public, part of the hash_join class. We will be calling this from an application and providing bindings to it from Java, so it should not be hidden in a detail header.

[vyasr]

it looks like the hash join APIs for the output size are all detail APIs

The APIs are public, part of the hash_join class. We will be calling this from an application and providing bindings to it from Java, so it should not be hidden in a detail header.

I forgot that the hash join class itself was public, and none of free functions in cuDF for joins had this parameter added so I was confused (such as inner_join vs hash_join::inner_join). This is fixed now.

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (branch-21.10@2b92220). Click here to learn what that means.
The diff coverage is n/a.

@@               Coverage Diff               @@
##             branch-21.10    #8928   +/-   ##
===============================================
  Coverage                ?   10.65%           
===============================================
  Files                   ?      114           
  Lines                   ?    19080           
  Branches                ?        0           
===============================================
  Hits                    ?     2033           
  Misses                  ?    17047           
  Partials                ?        0           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2b92220...2e4e94a. Read the comment docs.

[jlowe]

The failing test is intended to test the cudf behavior when an invalid table reference is specified. It's asking for a RIGHT table column on an expression that only takes one table. In the past it would silently ignore table references and act like it was referencing the LEFT table, but that doesn't seem to be the case anymore. If the behavior is explicitly undefined for this case we can remove the test, but it would be interesting to know why the behavior changed.

[vyasr]

Discussed offline, I'm changing the behavior to always throw an exception in compute_column. This also necessitates updating the test, but the new (defined) behavior is preferable to the old UB.

[mythrocks]

Generally +1. A couple of clarification questions, really.

This was an enjoyable, and informative read. I'd like to look more closely at the kernels themselves, later on.

[mythrocks]

LGTM!

[vyasr]

@gpucibot merge