Rounding in floating point arithmetic

The integration of sticky-bit preservation circuitry in chained multiply-add operations addresses the issue of mantissa shifting and sticky bit exclusion, ensuring accurate floating point arithmetic results and reducing hardware needs, thus enhancing the efficiency of data processing apparatuses.

GB2629812AActive Publication Date: 2024-11-13ARM LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
GB2023006940
Authority / Receiving Office
GB · GB
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-11
Publication Date
2024-11-13
Estimated Expiration
2043-05-11

AI Technical Summary

Technical Problem

Existing data processing apparatuses face challenges in performing floating point arithmetic operations efficiently, particularly in fused multiply-add operations, due to area and power consumption concerns, and may produce slightly different results when using chained multiply-add units compared to fused operations, especially when shifting mantissas leads to exclusion of sticky bits.

Method used

Incorporating sticky-bit preservation circuitry to adjust zero-value shifted mantissas back to non-zero during exponent alignment in chained multiply-add operations, allowing for equivalent results to fused operations without additional shift-clamping steps and dedicated fused circuitry.

Benefits of technology

This approach ensures accurate rounding and reduces hardware requirements, achieving results equivalent to fused operations while conserving area and power, without the need for additional shift-clamping steps.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

Disclosed is a method and circuitry 300 for performing a combined floating point arithmetic operation on three input values. The arithmetic operation starts by executing a first operation 301 with rou
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Apparatus and method for controlling rounding when performing a floating point operation

    US20170010863A1

  • Floating-Point Multiply-Add with Down-Conversion

    US20170293470A1

  • Multi-input floating-point adder

    US20200371748A1