New Decimal <--> Floating conversion

[pmattione-nvidia]

This PR contains the main algorithm for the new decimal <--> floating conversion code. This algorithm was written to address the precision issues described here.

Summary

• The new algorithm is more accurate than the previous code, but it is also far more complex.
• It can perform conversions that were not even possible in the old code due to overflow (decimal32/64/128 conversions only worked for scale factors up to 10^9/18/38, respectively). Now the entire floating-point range is convertible, including denormals.
• This new algorithm is significantly faster in some parts of the conversion phase-space, and in some parts slightly slower.

Previous PR's

These contain the supporting parts of this work:

• Explicit conversion PR
• Benchmarking PR
• Powers-of-10 PR
• Utilities PR. These utilities are updated here to support denormals.

Algorithm Outline

We convert floating -> (integer) decimal by:

• Extract the floating-point mantissa (converted to integer) and power-of-2
• For float we use a uint64 to contain our data during the below shifting/scaling, for double uint128_t
• In this shifting integer, we alternately apply the extracted powers-of-2 (bit-shifts, until they're all used) and scale-factor powers-of-10 (multiply/divide) as needed to reach the desired scale factor.

Decimal -> floating is just the reverse operation.

Supplemental Changes

• Testing: Add decimal128, add precise-conversion tests. Remove kludges due to inaccurate conversions. Add test for zeroes.
• Benchmarking: Enable regions of conversion phase-space for benchmarking that were not possible in the old algorithm.
• Unary: Cleanup by using CUDF_ENABLE_IF. Call new conversion code for base-10 fixed-point.

Performance for various conversions/input-ranges

• Note: F32/F64 is float/double

New algorithm is FASTER by:

• F64 --> decimal64: 60% for E8 --> E15
• F64 --> decimal128: 13% for E-8 --> E-15
• F64 --> decimal128: 22% for E8 --> E15
• F64 --> decimal128: 27% for E31 --> E38
• decimal32 --> F64: 18% for E-3 --> E4
• decimal64 --> F64: 27% for E-14 --> E-7
• decimal64 --> F64: 17% for E-3 --> E4
• decimal128 --> F64: 21% for E-14 --> E-7
• decimal128 --> F64: 11% for E-3 --> E4
• decimal128 --> F64: 13% for E31 --> E38

New algorithm is SLOWER by:

• F32 --> decimal32: 3% for E-3 --> E4
• F32 --> decimal64: 2% for E-14 --> E14
• F64 --> decimal32: 3% for E-3 --> E4
• decimal32 --> F32: 5% for E-3 --> E4
• decimal128 --> F64: 36% for E-37 --> E-30

Other kernels:

• The PYMOD binary-op benchmark is 7% slower.

Performance discussion

• Many conversions have identical speed, indicating these algorithms are often fast and we are instead bottlenecked on overheads such as getting the input to the gpu in the first place.
• F64 conversions are often much faster than the old algorithm as the new algorithm completely avoids the FP64 pipeline. Other than the cast to double itself, all of the operations are on integers. Thus we don't have threads competing with each other and taking turns for access to the floating-point cores.
• The conversions are slightly slower for floats with powers-of-10 near zero. Presumably this is due to code overhead for e.g., handling a large range of inputs, UB-checks for bit shifts, branches for denormals, etc.
• The conversion is slower for decimal128 conversions with very small exponents, which requires several large divisions (128bit divided by 64bit).
• The PYMOD kernel is slower due to register pressure from the introduction of the new division routines in the earlier PR. Even though this benchmark does not perform decimal <--> floating conversions, it gets hit because of inlined template code in the kernel increasing the code/register pressure.

Checklist

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

[pmattione-nvidia]

NOTE: Nevermind, the below is now working, the algorithm was tweaked accordingly.

Note that currently the python tests are failing because of 97938.2f -> decimal64 with scale-factor -2. Note that the tests this was failing on before (in the original linked issue) now pass (higher levels of precision), this failure is different. This failure due to the increment_on_truncation() function, where we increment when we should not. Specifically we are triggering the "// The only way that this produces the incorrect result is if ..." condition.

I could change the xfail in the python test to fail this case instead of the others, but I'm trying to see if I can think of a solution ... I'm betting arrow is able to figure it out BECAUSE of this precision argument (7 decimal digits, which is the max precision of float) which we don't have ... perhaps we should have this as an optional argument?

[wence-]

Some comments on the comments

[jlowe]

Java approval

[davidwendt]

The entire source of this file is in the detail namespace but the file is in a public header directory.
Are these meant to be called only internally then? We are evaluating public/internal functions/source recently.
Perhaps this could be moved to detail directory instead -- cudf/fixed_point/detail would probably be ok.

Now I'm seeing this is called from public unary.hpp which has its own issue.
So this is probably out of scope for this PR.

[hyperbolic2346]

Is there an issue for this change?

[pmattione-nvidia]

Some of this functionality needs to be exposed for the spark-rapids-jni repository to wrap around for it to perform conversions with the spark-specific rounding. I'm still finalizing the details for the sr-jni code, when that's done I'll have a better idea of what the minimal amount of code we can expose is.

[pmattione-nvidia]

Note that cuDF necessarily will NOT match pyarrow, due to choosing to truncate where pyarrow wants to round.

[hyperbolic2346]

The comments in this code are really good. I appreciate the time spent adding them to explain the dense conversions going on in here.

[hyperbolic2346]

Is there an issue for this change?

[pmattione-nvidia]

/merge