Enable rounding for Decimal32 and Decimal64 in cuDF

[a-hirota]

Description

Checklist

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

Example of Behavior After Modification

The following example demonstrates the behavior of the round function for both float and Decimal32Dtype before and after the modification.

import cudf

def check(t):
    a = cudf.DataFrame({'a': [1.00, 2.054, 3.01]}).astype(t)
    print(type(t), a)
    a = a.round(decimals=1, how="half_up")
    print(type(t), a)

print('float')
check(float)
print('decimal')
check(cudf.Decimal32Dtype(precision=3, scale=2))

Output

After Modification

For Decimal32Dtype, rounding works as expected:

float
<class 'type'>        a
0  1.000
1  2.054
2  3.010
<class 'type'>      a
0  1.0
1  2.1
2  3.0

decimal
<class 'cudf.core.dtypes.Decimal32Dtype'>       a
0  1.00
1  2.05
2  3.01
<class 'cudf.core.dtypes.Decimal32Dtype'>      a
0  1.0
1  2.1
2  3.0

Before Modification

For Decimal32Dtype, rounding did not modify the values:

float
<class 'type'>        a
0  1.000
1  2.054
2  3.010
<class 'type'>      a
0  1.0
1  2.1
2  3.0

decimal
<class 'cudf.core.dtypes.Decimal32Dtype'>       a
0  1.00
1  2.05
2  3.01
<class 'cudf.core.dtypes.Decimal32Dtype'>       a
0  1.00
1  2.05
2  3.01

This example shows that, prior to this modification, rounding had no effect on Decimal32Dtype columns, while after the change, Decimal32Dtype columns round as expected.

Additional Information

• Problem:
  cuDF currently does not support rounding for Decimal types, limiting its functionality compared to libcudf, which does support it.
• Solution:
  Updated the cols definition in cudf/python/cudf/cudf/core/indexed_frame.py to allow rounding for Decimal32 and Decimal64 types. The modified code now checks if col.dtype is of Decimal32 or Decimal64 type, allowing rounding directly on these columns without making a copy.
• Alternatives Considered:
  Using apply in Pandas can achieve similar functionality, but it is not as efficient as a native rounding method in cuDF.
• Additional Context:
  • libcudf documentation indicates that Decimal32 and Decimal64 are supported for rounding.

This enhancement improves consistency between cudf and libcudf by enabling direct rounding of Decimal types in cudf.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[vyasr]

Thanks @a-hirota! Could you please add a test that would have failed with the old code that succeeds now? You should be able to construct the expected decimal column and then compare it to what you get out of rounding.

[a-hirota]

Thank you @vyasr for your feedback!

I’ve added test cases that validate the rounding behavior for both Decimal32Dtype and Decimal64Dtype with the half_up and half_even rounding modes. These include scenarios where the previous code would have produced incorrect results, such as rounding .5 values inconsistently in the half_even mode. The expected output is constructed explicitly for each test case, and assertions ensure the correctness of the implementation.

Let me know if there's anything else you'd like me to adjust or add!

[vyasr]

/ok to test

[vyasr]

@a-hirota looks like you need to fix the style, could you please run pre-commit?

[a-hirota]

@vyasr
Apologies, I have run pre-commit.

[vyasr]

No worries! Thank you.

[vyasr]

/ok to test

[vyasr]

/ok to test

[vyasr]

@a-hirota it looks like pre-commit is still making changes to the test that you added.

[a-hirota]

@a-hirota it looks like pre-commit is still making changes to the test that you added.
@vyasr
Sorry, I forgot to push the test file...

[vyasr]

/ok to test

[vyasr]

/ok to test

[mroeschke]

Looks good. Thanks @a-hirota

[mroeschke]

/merge