Skip to content
This repository was archived by the owner on Dec 17, 2025. It is now read-only.

Conversation

@shinaoka
Copy link
Member

Summary

This PR updates to use instead of hardcoded for automatic precision selection.

Changes

  • Changed parameter from to
  • Automatically selects appropriate precision based on epsilon:
    • (f64) for epsilon >= 1e-8
    • (Df64) for epsilon < 1e-8

Benefits

  1. Performance improvement: Avoids unnecessary extended precision overhead when f64 is sufficient
  2. Better accuracy: Automatically uses extended precision when needed (epsilon < 1e-8)
  3. Matches actual usage: Benchmark now reflects real-world usage patterns where precision is selected based on epsilon

Testing

  • Benchmark runs successfully with automatic precision selection
  • Performance improved for epsilon >= 1e-8 cases (no xprec overhead)
  • Extended precision still used when needed (epsilon < 1e-8)

- Change from SPIR_TWORK_FLOAT64X2 to SPIR_TWORK_AUTO
- Automatically selects FLOAT64 for epsilon >= 1e-8, FLOAT64X2 for epsilon < 1e-8
- Improves performance by avoiding unnecessary extended precision overhead
- Better matches actual usage patterns where precision is selected based on epsilon
@terasakisatoshi terasakisatoshi merged commit 4e761cb into main Nov 16, 2025
9 checks passed
@terasakisatoshi terasakisatoshi deleted the use-auto-twork-precision branch November 16, 2025 02:46
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants