Implement `num_traits::Float` trait for `Df64` by sakuraba07 · Pull Request #15 · tuwien-cms/xprec-rs

sakuraba07 · 2025-11-05T06:02:26Z

This PR implements #2.

Summary

This PR implements the num_traits::Float trait for Df64, enabling compatibility with generic numeric code that relies on this standard trait. Additionally, the NumCast trait is implemented to support generic numeric type conversions.

Problem

As requested in #2, the Df64 type lacked the num_traits::Float trait implementation, which prevented access to essential floating-point inspection methods like is_infinite(), epsilon(), is_nan(), and classify(). This limited interoperability with Rust's numeric ecosystem and generic code that depends on these trait bounds.

Solution

`traits.rs` - Float trait implementation

Implemented all 58 required num_traits::Float methods for Df64
Adopted a delegation design pattern for consistency:
- Constants (8 methods): Return existing Df64 constants (NAN, INFINITY, EPSILON, etc.)
- Checks (6 methods): Delegate to checks module functions
- Basic operations (5 methods): Delegate to Signed trait
- Rounding (7 methods): Delegate to ComplexField trait
- Comparison (3 methods): Delegate to RealField trait
- Exponential/logarithmic (11 methods): Delegate to ComplexField trait
- Trigonometric (7 methods): Delegate to ComplexField trait
- Hyperbolic (6 methods): Delegate to ComplexField trait
- Conversion (2 methods): to_degrees() and to_radians() implemented directly using PI constant
- TODO (2 methods): cbrt() and integer_decode() marked as todo!() with explanatory comments
Updated 4 locations to use explicit type syntax <Df64 as From<f64>>::from() instead of Df64::from() for consistency
Added 5 new tests:
- test_float_constants() - Verify Float constant methods
- test_float_neg_zero() - Verify neg_zero() implementation
- test_float_checks() - Verify classification methods (is_nan(), is_infinite(), etc.)
- test_float_classify() - Verify classify() method
- test_float_conversions() - Verify to_degrees() and to_radians()

`convert.rs` - NumCast trait implementation

Implemented num_traits::NumCast trait to enable generic numeric type conversions
Leverages existing FromPrimitive::from_f64() implementation

Changes

src/traits.rs:
- Add complete Float trait implementation (+326 lines)
- Add 5 focused tests for Float trait methods
- Update 4 locations to use explicit From<f64> syntax
src/convert.rs:
- Add NumCast trait implementation (+6 lines)

Testing

All existing tests pass
New tests in traits.rs verify:
- Float constant methods (nan(), infinity(), epsilon(), etc.)
- Float classification methods (is_nan(), is_infinite(), is_finite(), etc.)
- Special value handling (neg_zero(), classify())
- Angle conversions (to_degrees(), to_radians())

Future Work

Implement ComplexField::cbrt() and subsequently Float::cbrt() using Newton iteration or other stable algorithms
Design and implement Df64::integer_decode() with appropriate representation for double-double mantissa (possibly using extended mantissa format or custom representation)

Implement the num_traits::Float trait to provide standard floating point operations interface for Df64 type. Changes: - traits.rs: Add Float trait implementation with 58 methods using delegation design pattern. Add 5 new tests for Float trait methods. Update 4 locations to use explicit From<f64> type syntax. - convert.rs: Add NumCast trait implementation. Note: cbrt() and integer_decode() are marked as todo!() due to requiring additional design for double-double representation. Closes tuwien-cms#2

shinaoka · 2025-11-22T05:50:28Z

Thank you for the contribution. I have a few comments:

Float trait is more basic than ComplexField trait. It is better not to use ComplexField to define Float trait.
Could you add unit tests for the new trait?

This reverts commit ee2aba0. Rolling back the previous Float trait implementation to address PR review feedback and implement a better design approach. Reasons for revert: - Float trait should not depend on ComplexField trait (more basic trait should not depend on more complex trait) - Need to refactor implementation to use only basic operations Will re-implement Float trait without ComplexField dependency in the next commit. Related to PR tuwien-cms#15

@shinaoka

Implement the num_traits::Float trait to provide standard floating-point operations for Df64 without depending on higher-level traits. Changes: - traits.rs: Add Float trait implementation with 58 methods delegating directly to low-level modules (checks, round, exp, circular, hyperbolic, funcs, arith, roots, consts) instead of ComplexField - convert.rs: Add NumCast trait implementation to support generic numeric type conversions Design: - Float and ComplexField are now independent, both delegating to the same low-level modules - All methods use #[inline(always)] for performance - cbrt() and integer_decode() marked as todo!() pending double-double specific design Addresses PR tuwien-cms#15 review feedback from @shinaoka

Add 56 unit tests covering all 58 methods of the num_traits::Float trait implementation for Df64, ensuring API correctness and mathematical properties. Test coverage: - Constants (8 methods): nan, infinity, epsilon, min/max values - Classification (7 methods): is_nan, is_infinite, is_finite, is_normal, classify - Basic arithmetic (3 methods): abs, signum, recip - Rounding (5 methods): floor, ceil, round, trunc, fract - Comparison (5 methods): min, max, abs_sub, mul_add, copysign - Powers & roots (3 methods): powi, powf, sqrt, hypot - Exponential & logarithmic (8 methods): exp, exp2, ln, log, log2, log10, exp_m1, ln_1p - Trigonometric (8 methods): sin, cos, tan, asin, acos, atan, atan2, sin_cos - Hyperbolic (6 methods): sinh, cosh, tanh, asinh, acosh, atanh - Angle conversion (2 methods): to_degrees, to_radians - Not yet implemented (2 methods): cbrt, integer_decode (panic tests) Test characteristics: - Each test covers positive, negative, and edge cases (zero, infinity, NaN) - Validates mathematical properties (e.g., odd/even functions, identities) - Uses high precision tolerance (1e-30 typical, relaxed to 1e-20~1e-29 where needed) - Special validations: tan(π/2) divergence, atan(∞) convergence All 94 tests pass (57 Float trait tests + 37 existing implementation tests). Related to PR tuwien-cms#15

sakuraba07 · 2025-11-25T05:50:56Z

@shinaoka Thank you for the review. I've addressed both of your comments:

1. Removed ComplexField dependency

The Float trait implementation now delegates directly to low-level modules instead of ComplexField:

Before: Float::sin() → ComplexField::sin() → circular::sin()
After: Float::sin() → circular::sin()

See commit: b0db211

2. Added comprehensive unit tests

Added 56 unit tests covering all 58 Float trait methods:

Each test validates basic functionality, edge cases (zero, infinity, NaN), and mathematical properties
Tests cover positive and negative values, odd/even function properties, special cases like tan(π/2) divergence and atan(∞) convergence
High precision validation (1e-30 typical tolerance, relaxed to 1e-20~1e-29 where needed)

All 94 tests pass (57 new Float trait tests + 37 existing tests).

See commit: 2163c11

Summary of changes

traits.rs: Refactored Float implementation to be independent of ComplexField (+871 lines with tests)
convert.rs: Added NumCast trait implementation
All methods use #[inline(always)] for performance
cbrt() and integer_decode() remain as todo!() (will be addressed in future work)

Ready for re-review. Please let me know if you have any other concerns.

Integrate +0.0/-0.0 behavior tests into existing Float trait tests and add dedicated tests for signed zero arithmetic and copysign operations.

sakuraba07 · 2025-11-26T02:51:41Z

@shinaoka I've added two more commits addressing additional issues found during review:

1. Fix `abs_sub` implementation (`6fd2583`)

The previous implementation returned abs(self - other), but the correct behavior per num_traits::Float specification is max(self - other, 0). This has been fixed along with the corresponding tests.

2. Add signed zero behavior tests (`070a5b1`)

Added comprehensive tests for +0.0 and -0.0 behavior across Float trait methods. These tests:

Document current implementation behavior for signed zeros
Cover arithmetic, comparison, trigonometric, and other operations
Note where behavior differs from IEEE 754 (e.g., 1/0 returns NaN instead of Infinity)

The signed zero tests are integrated into existing test functions where appropriate, with dedicated tests for arithmetic operations and copysign.

All 94 tests pass.

Ready for re-review.

mwallerb

Thanks for the PR! Looks good, but the tests need work.

Please do not use check against an absolute threshold. Most functions satisfy a relative accuracy goal
Please do not use magic numbers (1e-29 etc.) in the threshold. Use a small multiple of the machine epsilon
Please reuse the functions in test_utils rather rolling your own tests whenever useful. Otherwise, please make use of the macros inside the approx crate, in particular approx::assert_relative_eq and assert_ulps_eq.

In general, most of the tests seem to duplicate the unit tests in the respective modules, for example, exp2 and so forth. Why test again and in a separate place?

mwallerb · 2025-11-26T11:21:07Z

+    #[inline(always)]
+    fn mul_add(self, a: Self, b: Self) -> Self {
+        // There are two requirements that one has with fma: (1) it must be
+        // accurate without intermediate rounding and (2) it must be at least
+        // as fast as (a*b)+c. We have no way of satisfying both, so we go
+        // for performance.
+        (self * a) + b
+    }


This duplicates an existing method, please unify.

mwallerb · 2025-11-26T11:24:43Z

+    #[inline(always)]
+    fn to_degrees(self) -> Self {
+        self * consts::ONE_OVER_PI * <Df64 as From<f64>>::from(180.0)
+    }
+
+    #[inline(always)]
+    fn to_radians(self) -> Self {
+        self * consts::PI / <Df64 as From<f64>>::from(180.0)
+    }


Please use precomputed values for PI/180 and its reciprocal here.

…an conversion - Add mul_add_qq function to arith.rs for fused multiply-add on Df64 - Add RADIANS_PER_DEGREE and DEGREES_PER_RADIAN constants to consts.rs - Update Float::mul_add and ComplexField::mul_add to delegate to arith::mul_add_qq - Update to_degrees/to_radians to use precomputed constants instead of runtime computation Addresses PR review feedback from mwallerb

…agic numbers - Replace absolute thresholds (1e-30, 1e-29, etc.) with assert_ulps_eq! macro - Use Df64::EPSILON-based comparisons for near-zero checks - Tests now use relative precision based on machine epsilon - Add comments clarifying that detailed precision tests are in respective modules - Float trait tests focus on verifying correct delegation to low-level modules Addresses PR review feedback from mwallerb regarding test quality

- Add is_subnormal() method delegating to checks::is_subnormal() - Add clamp() method delegating to funcs::clamp() - Add corresponding unit tests

- Fix Basic arithmetic comment: 5 -> 3 methods - Fix Trigonometric functions comment: 7 -> 8 methods - Fix test section name: "special constants" -> "not yet implemented" - Move test_float_powi and test_float_powf to exponential section

sakuraba07 · 2025-12-01T04:45:14Z

@mwallerb Thank you for the detailed review. I've addressed all your feedback in the following commits:

1. Test improvements (`1c4a7cb`)

Replaced all absolute thresholds (1e-30, 1e-29, etc.) with assert_ulps_eq! macro from the approx crate
Use Df64::EPSILON-based comparisons for near-zero checks
Added comments clarifying that Float trait tests focus on verifying correct delegation, while detailed precision tests remain in their respective modules

Regarding the test duplication concern: you're right that some tests appear similar to those in the underlying modules. The Float trait tests are intentionally minimal now - they verify that delegation works correctly rather than re-testing mathematical accuracy.

2. Unified `mul_add` implementation (`a1dcb0c`)

Added mul_add_qq function to arith.rs
Both Float::mul_add and ComplexField::mul_add now delegate to this shared implementation

3. Precomputed constants for angle conversion (`a1dcb0c`)

Added RADIANS_PER_DEGREE and DEGREES_PER_RADIAN to consts.rs
to_degrees() and to_radians() now use these precomputed values instead of runtime computation

Additional improvements

4f5b4c9: Implemented is_subnormal() and clamp() methods for Float trait
b802248: Fixed method count comments and reorganized test sections

All 96 tests pass.

Ready for re-review. Please let me know if there are any other concerns.

shinaoka · 2025-12-13T14:43:20Z

@mwallerb
Ready to merge?

sakuraba07 added 4 commits November 24, 2025 13:39

Merge branch 'mainline' into implement-num-traits-float

23f954a

sakuraba07 added 2 commits November 26, 2025 11:04

fix: correct abs_sub to return max(self - other, 0) per Float trait spec

6fd2583

test: add signed zero behavior tests for Float trait

070a5b1

Integrate +0.0/-0.0 behavior tests into existing Float trait tests and add dedicated tests for signed zero arithmetic and copysign operations.

mwallerb reviewed Nov 26, 2025

View reviewed changes

sakuraba07 added 4 commits November 27, 2025 10:07

feat: implement is_subnormal and clamp methods for Float trait

4f5b4c9

- Add is_subnormal() method delegating to checks::is_subnormal() - Add clamp() method delegating to funcs::clamp() - Add corresponding unit tests

mwallerb merged commit 1193025 into tuwien-cms:mainline Dec 13, 2025
2 checks passed

sakuraba07 deleted the implement-num-traits-float branch December 15, 2025 04:07

Navigation Menu

Uh oh!

Implement `num_traits::Float` trait for `Df64` - #15

Implement `num_traits::Float` trait for `Df64`#15
mwallerb merged 11 commits into
tuwien-cms:mainlinefrom
sakuraba07:implement-num-traits-float

sakuraba07 commented Nov 5, 2025

Uh oh!

shinaoka commented Nov 22, 2025

Uh oh!

sakuraba07 commented Nov 25, 2025

Uh oh!

sakuraba07 commented Nov 26, 2025

Uh oh!

mwallerb left a comment •

edited

Loading

Uh oh!

mwallerb Nov 26, 2025

Uh oh!

mwallerb Nov 26, 2025

Uh oh!

sakuraba07 commented Dec 1, 2025

Uh oh!

shinaoka commented Dec 13, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

sakuraba07 commented Nov 5, 2025

Summary

Problem

Solution

traits.rs - Float trait implementation

convert.rs - NumCast trait implementation

Changes

Testing

Future Work

Uh oh!

shinaoka commented Nov 22, 2025

Uh oh!

sakuraba07 commented Nov 25, 2025

1. Removed ComplexField dependency

2. Added comprehensive unit tests

Summary of changes

Uh oh!

sakuraba07 commented Nov 26, 2025

1. Fix abs_sub implementation (6fd2583)

2. Add signed zero behavior tests (070a5b1)

Uh oh!

mwallerb left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mwallerb Nov 26, 2025

Choose a reason for hiding this comment

Uh oh!

mwallerb Nov 26, 2025

Choose a reason for hiding this comment

Uh oh!

sakuraba07 commented Dec 1, 2025

1. Test improvements (1c4a7cb)

2. Unified mul_add implementation (a1dcb0c)

3. Precomputed constants for angle conversion (a1dcb0c)

Additional improvements

Uh oh!

shinaoka commented Dec 13, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

`traits.rs` - Float trait implementation

`convert.rs` - NumCast trait implementation

1. Fix `abs_sub` implementation (`6fd2583`)

2. Add signed zero behavior tests (`070a5b1`)

mwallerb left a comment •

edited

Loading

1. Test improvements (`1c4a7cb`)

2. Unified `mul_add` implementation (`a1dcb0c`)

3. Precomputed constants for angle conversion (`a1dcb0c`)