Implement cudf::group_by (hash) for decimal32 and decimal64

[codereport]

Follow up PR to #7169

This PR resolves a part of #3556.

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (branch-0.19@53d7ad2). Click here to learn what that means.
The diff coverage is n/a.

@@              Coverage Diff               @@
##             branch-0.19    #7190   +/-   ##
==============================================
  Coverage               ?   82.19%           
==============================================
  Files                  ?      100           
  Lines                  ?    16955           
  Branches               ?        0           
==============================================
  Hits                   ?    13937           
  Misses                 ?     3018           
  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 53d7ad2...d6028c1. Read the comment docs.

[harrism]

Moving to P0 for 0.19.

[davidwendt]

Looks like you put all the fixed-point groupby tests into tests/group_by/group_count_test.cpp . There are individual test source files for count, min, max, sum, mean, etc. in tests/group_by/. I would recommend splitting the tests into the matching test source files.

Also, once that is done you can declare those tests that are not yet supported as disabled instead of commenting them out. Here are some examples I used for disabling some groupby dictionary tests:

cudf/cpp/tests/groupby/group_sum_test.cpp

Lines 139 to 141 in 4f87a59

	// These tests will not work until the following ptxas bug is fixed in 10.2
	// https://nvbugswb.nvidia.com/NvBugs5/SWBug.aspx?bugid=3186317&cp=
	TYPED_TEST(groupby_sum_test, DISABLED_dictionary)

cudf/cpp/tests/groupby/group_mean_test.cpp

Lines 144 to 146 in 4f87a59

	// This tests will not work until the following ptxas bug is fixed in 10.2
	// https://nvbugswb.nvidia.com/NvBugs5/SWBug.aspx?bugid=3186317&cp=
	TEST_F(groupby_dictionary_mean_test, DISABLED_basic)

[davidwendt]

Another place I added a check for dictionary columns is in the cudf::groupby::verify_valid_requests() internal function. I think that is because the dictionary code path would crash at runtime. The fixed-point aggs may just give incorrect results so it may not be necessary.

cudf/cpp/src/groupby/groupby.cu

Lines 122 to 139 in 568df5b

	// The aggregations listed in the lambda below will not work with a values column of type
	// dictionary if this is compiled with nvcc/ptxas 10.2.
	// https://nvbugswb.nvidia.com/NvBugs5/SWBug.aspx?bugid=3186317&cp=
	#if (__CUDACC_VER_MAJOR__ == 10) and (__CUDACC_VER_MINOR__ == 2)
	CUDF_EXPECTS(
	std::all_of(
	requests.begin(),
	requests.end(),
	[](auto const& request) {
	return std::all_of(
	request.aggregations.begin(), request.aggregations.end(), [&request](auto const& agg) {
	return (!cudf::is_dictionary(request.values.type()) ||
	!(agg->kind == aggregation::SUM or agg->kind == aggregation::MEAN or
	agg->kind == aggregation::STD or agg->kind == aggregation::VARIANCE));
	});
	}),
	"dictionary type not supported for this aggregation");
	#endif

[codereport]

Another place I added a check for dictionary columns is in the cudf::groupby::verify_valid_requests() internal function. I think that is because the dictionary code path would crash at runtime. The fixed-point aggs may just give incorrect results so it may not be necessary.

I came across this in my code exploring but I don't think it is necessary.

[kkraus14]

@gpucibot merge