Add Python DecimalColumn

[shwina]

Resolves #6657.

[GPUtester]

Please update the changelog in order to start CI tests.

View the gpuCI docs here.

[codecov]

Codecov Report

Merging #6715 (870e040) into branch-0.18 (8860baf) will increase coverage by 0.12%.
The diff coverage is 88.53%.

@@               Coverage Diff               @@
##           branch-0.18    #6715      +/-   ##
===============================================
+ Coverage        82.09%   82.22%   +0.12%     
===============================================
  Files               97       98       +1     
  Lines            16474    16655     +181     
===============================================
+ Hits             13524    13694     +170     
- Misses            2950     2961      +11     

Impacted Files	Coverage Δ
python/cudf/cudf/_lib/__init__.py	100.00% <ø> (ø)
python/cudf/cudf/core/column/lists.py	91.66% <ø> (-0.09%)	⬇️
python/cudf/cudf/utils/ioutils.py	78.71% <ø> (ø)
python/cudf/cudf/io/orc.py	86.89% <63.63%> (-1.51%)	⬇️
python/cudf/cudf/core/dataframe.py	90.49% <72.41%> (-0.22%)	⬇️
python/cudf/cudf/core/series.py	91.10% <80.00%> (-0.06%)	⬇️
python/cudf/cudf/core/column/numerical.py	94.08% <83.33%> (-0.33%)	⬇️
python/cudf/cudf/core/dtypes.py	89.94% <84.00%> (-0.45%)	⬇️
python/cudf/cudf/core/column/column.py	88.12% <85.71%> (-0.02%)	⬇️
python/cudf/cudf/io/csv.py	91.66% <86.66%> (-1.67%)	⬇️
... and 18 more

---

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 6c116e3...870e040. Read the comment docs.

[harrism]

This PR needs a description, and a decision on whether to retarget to 0.18.

[shwina]

We plan to have this ready for review before code freeze (EOD today).

[karthikeyann]

cpp approval.
Update Copyright year.

[codereport]

@shwina @kkraus14 CNL's behavior is that it can't represent numbers outside the bounds of the underlying type regrardless of what the positive scale (negative scale for Python) is. CNL also has a from_rep facility to avoid shifting, but that doesn't seem make a difference. Godbolt: https://godbolt.org/z/rd1xoe

libcudf on the other hand actually does work with values greater than the bounds of the underlying type, as long as you construct from the already shifted values (which is what fixed_point_column_wrapper does). For example, the numbers stored below with scale=6 (read scale=-6 in the equivalent Python code), are outside the bounds of an int32_t, but the operations all still work, and our explicit conversion operator to std::string is able to print the numbers correctly as well.

  using fp_wrapper = cudf::test::fixed_point_column_wrapper<int32_t>;

  auto const a = fp_wrapper{{100000000}, scale_type{6}};
  auto const b = fp_wrapper{{5000000},   scale_type{7}};
  auto const c = fp_wrapper{{2},         scale_type{0}};

  auto const expected1 = fp_wrapper{{150000000}, scale_type{6}};
  auto const expected2 = fp_wrapper{{50000000},  scale_type{6}};

  auto const result1 = cudf::binary_operation(a, b, cudf::binary_operator::ADD, {});
  auto const result2 = cudf::binary_operation(a, c, cudf::binary_operator::DIV, {});

  CUDF_TEST_EXPECT_COLUMNS_EQUAL(expected1, result1->view());
  CUDF_TEST_EXPECT_COLUMNS_EQUAL(expected2, result2->view());

Update:
Just realized to make it even more clear, you can set the precision as high as you like:

  using fp_wrapper = cudf::test::fixed_point_column_wrapper<int64_t>;

  auto a = fp_wrapper{{100000000}, scale_type{100}};
  auto b = fp_wrapper{{5000000}, scale_type{101}};
  auto c = fp_wrapper{{2}, scale_type{0}};

  auto expected1 = fp_wrapper{{150000000}, scale_type{100}};
  auto expected2 = fp_wrapper{{50000000}, scale_type{100}};

  auto result1 = cudf::binary_operation(a, b, cudf::binary_operator::ADD, {});
  auto result2 = cudf::binary_operation(a, c, cudf::binary_operator::DIV, {});

  CUDF_TEST_EXPECT_COLUMNS_EQUAL(expected1, result1->view());
  CUDF_TEST_EXPECT_COLUMNS_EQUAL(expected2, result2->view());

[vuule]

Is this just a missing binop test? I can't tell if it's testing something more specific/related to this PR.

[shwina]

cc: @codereport; IIRC it's just a test we wrote because this case wasn't working in Python. Eventually, we traced that to a Python bug, but we decided to keep this test because why would you delete a test? :D

[vuule]

Any test that ever failed is a good test 👍
I asked just to make sure it covers what it's supposed to.

[kkraus14]

@gpucibot merge