Add optional sticky fixed-point overflow tracking (CUDF_TRACK_FIXED_POINT_OVERFLOW)

[Avinash-Raj]

Summary

• Introduces an opt-in CMake flag CUDF_TRACK_FIXED_POINT_OVERFLOW (default OFF). When enabled, defines CUDF_TRACK_FIXED_POINT_OVERFLOW for the cudf target (PUBLIC) so headers and translation units agree on fixed_point layout.
• When the macro is defined, fixed_point carries a sticky _overflow_flag, exposed as overflow_occurred(), updated on +, -, *, / and rescaled() (including shift overflow via detail::shift_overflows).
• Overflow predicates still assert only under __CUDACC_DEBUG__; enabling tracking does not by itself turn overflows into exceptions.
• This flag is required to ensure velox-cudf throws an exception whenever an decimal overflow occurs.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[bdice]

I would favor adding new operators/functions rather than a compile-time flag — we want to be able to ship one build of cuDF that works universally.

[devavret]

@bdice I'd imagine this is not limited to scalar functions like binaryops and such. We'd also need to detect overflow in aggregations like groupby and reduce. And those would require overloading the existing +,* operators.

What's your opinion on adding a new type maybe called safe_decimal.

Spark should also have a concept of arithmetic exception for overflows. How do they support this with libcudf decimal?

[mattgara]

What's your opinion on adding a new type maybe called safe_decimal.

Inline with this thinking, we could extend cudf's fixed_point to implement this:

enum class overflow_tracking { off, on };

template <typename Rep, Radix Rad,
          overflow_tracking Track = overflow_tracking::off>
class fixed_point{ ...

and use if constexpr (Track == overflow_tracking::on) to branch the implementation based on this configuration. Effectively, this should be compile-time semantically equivalent to having a safe_decimal, but with less code duplication.

[ttnghia]

This is related to ANSI support that I've recently filed: #21676
We can use some kind of input enum like mentioned above, and custom return type as my suggestion.

[Avinash-Raj]

Ok, replaced the CMake CUDF_TRACK_FIXED_POINT_OVERFLOW flag with an opt-in overflow_tracking NTTP on numeric::fixed_point and add decimal*_safe / type_id::DECIMAL*_SAFE so velox-cudf gets per-element overflow detection without requiring a separate libcudf build.

[pmattione-nvidia]

You might want to put overflow tracking into the floating <--> decimal conversion code as well (include/cudf/fixed_point/detail/floating_conversion.hpp). E.g. there are the guarded_left_shift() and guarded_right_shift() functions, and there are potential overflows in convert_floating_to_integral_shifting(), shift_to_decimal_pospow(), and shift_to_decimal_negpow(). All potential overflow sites are mentioned explicitly in the code comments in these functions.