Add fltflt rounding and fmod functions

[tbensonatl]

Add support for the following float-float (fltflt) functions:

• Round toward nearest, with ties toward even
• Truncate toward zero
• Truncate toward negative infinity
• fmod (floating point remainder)

Also includes are new unit tests and benchmarks for the newly introduced functions.

[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.

[greptile-apps]

Greptile Summary

Adds comprehensive float-float arithmetic rounding and modulo functions with proper IEEE-754 semantics. The implementation includes fltflt_round_to_nearest() with banker's rounding (round half to even), fltflt_round_toward_zero() for truncation, fltflt_floor() for rounding toward negative infinity, and fltflt_fmod() for floating-point remainder.

Key implementation details:

• Proper handling of the fast path (|hi| < 2^23) where hi may have fractional parts
• Correct treatment of the slow path (|hi| >= 2^23) where hi is already an integer
• Edge case handling for when hi is exactly an integer but lo has opposite sign
• Round-to-nearest uses tie-breaking to even when exactly at 0.5
• Constructors modernized with constexpr and = default
• Added fltflt_add_same_sign() optimization (11 FLOPs vs 20 FLOPs)

Testing and benchmarking:

• Comprehensive unit tests covering edge cases, tie-breaking, boundary conditions, and large values
• Performance benchmarks added for all new functions comparing fltflt against float/double baselines

Confidence Score: 5/5

• Safe to merge - well-implemented mathematical functions with comprehensive tests
• Implementation is mathematically sound with proper IEEE-754 semantics, comprehensive unit tests covering edge cases, and follows existing code patterns
• No files require special attention

Important Files Changed

Filename	Overview
include/matx/kernels/fltflt.h	Adds fltflt rounding functions (round_to_nearest, round_toward_zero, floor) and fmod with proper tie-breaking and edge case handling
test/00_misc/FloatFloatTests.cu	Comprehensive unit tests for new rounding functions covering edge cases, tie-breaking, and boundary conditions
bench/00_misc/fltflt_arithmetic.cu	Adds benchmark kernels for round, fmod, trunc, floor, and cast operations with proper ILP and timing instrumentation
bench/scripts/run_fltflt_benchmarks.py	Updates benchmark list to include new functions in test execution and reporting

_{Last reviewed commit: 0651eb6}

[greptile-apps]

_{4 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[cliffburdick]

constexpr?

[cliffburdick]

How would someone use this function?

[tbensonatl]

Currently you would likely call it directly in a kernel. I plan on switching the sar_bp operator to use this function when computing a distance because in that case we are adding squared values and thus we know they are non-negative. I will also benchmark if it is worth adding the conditionals required to call this function when appropriate in operator+(), but I suspect the answer is no.

[cliffburdick]

/build