Correctly Rounded Dot Products under Round-to-Odd
summary
Presents an algorithm that produces correctly rounded dot products under round-to-odd by reducing products to a summation via error‑free transformations for multiplication and a modified EFT for addition. The key fix makes Fast Two Sum an EFT under RO by using a successor of a power of two sigma so the high part has an odd significand, enabling an Extract Vector/Axom-style recursive summation. Implements the method for IEEE‑style encodings and combines it with EFT multiplication to produce correct rounded results for dot products and matrix multiplies on RO hardware.
AI-written summary · about
related talks
26mGenerating Correctly Rounded Math Libraries for New Floating Point Variants
22mOne Polynomial Approximation to Produce Correctly Rounded Results of an Elementa
58mPortable floating-point calculations
22mCorrectly Rounded Math Libraries without Worrying about the Application’s Rounding Mode
06mLightning Talk: A Compiler Optimization and the 2-Adic Numbers
17mMiniFloat-NN: A RISC-V ISA Extension for Low-Precision NN Training
01h 41mFast Conversion From Cpp Floating Point Numbers
16mWhat About the Integer Numbers?