[BUG] libcudf.round doesn't fully work on certain values

[galipremsagar]

Describe the bug
round method seems to be returning the actual value for the following inputs.

Steps/Code to reproduce bug

In [8]: import cudf

In [9]: s = cudf.Series([1.1320000000000001, 1.234234])

In [10]: s._column.round(decimals=3).element_indexing(0)
Out[10]: np.float64(1.1320000000000001)

In [11]: s._column.round(decimals=3).element_indexing(1)
Out[11]: np.float64(1.234)

Expected behavior

In [10]: s._column.round(decimals=3).element_indexing(0)
Out[10]: np.float64(1.132)

Environment overview (please complete the following information)

• Environment location: [Bare-metal]
• Method of cuDF install: [from source]

[vyasr]

I'm going to guess that this is a floating point precision issue and the best representation of 1.132 is in fact 1.1320000000000001, so the latter is actually the rounded value. Worth looking into though.

[bdice]

This seems like an expected behavior due to how floating point representation works in binary. There is no exact binary representation for the decimal 1.132, it seems. Try it in https://www.h-schmidt.net/FloatConverter/IEEE754.html.

There are also some known differences between cudf and pandas in how we print floats.

[bdice]

Using a decimal column with appropriate scale is the correct way to get exact decimal representations.