GNU C Library Sees Up to 12.9x Improvement with New Generic FMA Implementation
14 days ago
- #FMA
- #GNU C Library
- #performance
- Adhemerval Zanella contributed a new generic FMA implementation to the GNU C Library.
- The new implementation uses mostly integer arithmetic, avoiding the performance overhead of setting rounding modes.
- Fixes include signaling NaN issues, use of math_uint128.h for 64-bit multiplication, and arm32 rounding mode fixes.
- The implementation shows significant performance improvements, up to 12.9x in some cases.
- Glibc 2.43 will include additional optimized math functions from the CORE-MATH project.
- Optimized implementations for functions like remainder, frexpf, and others have been added.
- SVID handling for various functions was moved to compat symbols, improving performance.
- Glibc 2.43 is scheduled for release in February.