GPU Floating Point Atomics via Integer Hardware Emulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computing devices lack support for floating point atomic operations, leading to issues with concurrent thread modifications of shared data, where threads may read outdated values due to lack of hardware support for floating point atomic operations.
Innovation Solution
A method is implemented where a graphics processing unit (GPU) selects either a signed or unsigned atomic integer operation based on a floating point number, emulating floating point maximum/minimum operations using integer hardware, to ensure thread-safe data modification and accommodate non-monotonicity across zero.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If floating point atomic operations are implemented in software emulation, then floating point atomic support is achieved on hardware without native support, but processing speed and performance deteriorate
Solution Approach 1:
The patent uses bit pattern manipulation as an intermediary mechanism to translate floating point operations into integer operations. By treating floating point numbers as raw bit patterns and applying integer atomic operations on these patterns, the system achieves floating point atomic functionality through integer hardware, resolving the contradiction between hardware capability and operational versatility
Solution Approach 2:
The patent changes the operational parameter from floating point arithmetic to integer arithmetic while maintaining the same bit-level representation. This parameter change allows the use of faster integer atomic instructions on hardware that lacks floating point atomic support, improving processing speed while maintaining functional capability
2Adaptability or versatility
If integer operations are used to emulate floating point atomics, then hardware compatibility is improved, but precision and accuracy may worsen
Solution Approach 1:
The patent segments the floating point operation into distinct phases: bit pattern extraction, integer atomic operation execution, and result interpretation. This segmentation allows each phase to be optimized independently, maintaining floating point precision through proper bit pattern handling while using integer operations for the atomic update mechanism
Solution Approach 2:
The patent creates a bit-level copy of the floating point number and performs atomic operations on this copy as an integer. The original floating point precision is preserved in the source and destination values, while the integer operation works on a replicated bit pattern, ensuring no precision loss occurs during the atomic update
Data Source
AI summary
This disclosure provides systems, devices, apparatus, and methods, including computer programs encoded on storage media, for floating point min/max atomics using integer hardware. A data processor may obtain a first indication of a floating point number associated with a floating point operation. The data processor may select a signed atomic integer operation or an unsigned atomic integer operation based on at least one of the floating point number or the floating point operation, where the signed atomic integer operation is associated with a condition being met and the unsigned atomic integer operation is associated with a failure of the condition to be met.