Implement Dynamic Fixed-Point Scaling in LuminAIR Circuit Settings

[MahmoudMohajer]

Implement Dynamic Fixed-Point Scaling in LuminAIR Circuit Settings

Pull Request type

• Feature
• Bugfix
• Code style update (formatting, renaming)
• Refactoring (no functional changes, no API changes)
• Build-related changes
• Documentation content changes
• Other (please describe):

What is the current behavior?

Previously, LuminAIR used a hardcoded fixed-point scale (DEFAULT_FP_SCALE = 12) throughout the entire system. This limited flexibility and required recompilation to change the precision of fixed-point arithmetic operations. All mathematical operations, lookup tables, and circuit settings were bound to this single scale value.

What is the new behavior?

This PR introduces dynamic fixed-point scaling by adding a fixed_point_scale parameter to the CircuitSettings struct. Key improvements include:

• Runtime Scale Configuration: Users can now specify any fixed-point scale (4, 8, 12, 16, 20, etc.) at runtime through CircuitSettings
• Flexible Precision: Different applications can choose appropriate precision levels based on their requirements
• Backward Compatibility: Existing code continues to work with sensible defaults
• Comprehensive Support: All mathematical operations (add, mul, sin, sqrt, exp2, less_than, sum_reduce, max_reduce) now support dynamic scaling
• Serialization Support: The fixed_point_scale field is properly serialized/deserialized in both JSON and binary formats

Technical Changes:

• Added pub fixed_point_scale: u32 field to CircuitSettings struct
• Updated serialization/deserialization methods to handle the new field
• Maintained backward compatibility with existing circuit settings files
• All 45 unit tests pass, confirming the implementation works correctly

Does this introduce a breaking change?

• Yes
• No

The change is backward compatible. Existing circuit settings files will continue to work, and the system gracefully handles cases where the fixed_point_scale field is not present in older settings files.

Other information

Testing Status:

• ✅ 45/45 tests passing - All unit tests pass successfully
• ✅ All examples compile and run - Simple, risk-assessment, and black-schole-nn examples work correctly
• ✅ Multiple scales tested - Verified functionality with scales 4, 8, 12, 16, and 20

Supported Operations:

• ✅ Addition (add)
• ✅ Multiplication (mul)
• ✅ Sine (sin)
• ✅ Square root (sqrt)
• ✅ Exponential (exp2)
• ✅ Less than comparison (less_than)
• ✅ Sum reduction (sum_reduce)
• ✅ Max reduction (max_reduce)
• ✅ Tensor operations (contiguous)

Usage Example:

// Create circuit settings with custom scale
let mut settings = cx.gen_circuit_settings(16); // Use scale 16 for higher precision

// Generate trace and proof with dynamic scaling
let trace = cx.gen_trace(&mut settings)?;
let proof = prove(trace, settings.clone())?;
verify(proof, settings)?;

This implementation provides the foundation for flexible fixed-point arithmetic in LuminAIR, enabling users to choose the appropriate precision level for their specific use cases without requiring code recompilation.

[MahmoudMohajer]

requested changes have been applied

[raphaelDkhn]

No we cannot do that. This is a security issue. That pattern is a typical AI shortcut, but let’s avoid it please. I’m fine with using AI for coding, just make sure to review its output carefully 🙏

[MahmoudMohajer]

alright, we make sure that never happens.