Vector Memory Array Sorting With Bitonic Compare-and-Swap
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network accelerators (NNAs) lack a method to efficiently sort vector data, relying on non-deterministic algorithms like quicksort, which complicates scheduling and can lead to longer processing times than necessary.
Innovation Solution
A method utilizing the RISC-V vector extension to implement a bitonic sorting algorithm, generating new vectors for comparison and swapping elements based on a mask, allowing for deterministic sorting within a predetermined number of steps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If quicksort algorithm is used for sorting vector data, then sorting can be performed with simple implementation, but the sorting time becomes non-deterministic and may be longer than necessary
Solution Approach 1:
The patent changes the algorithmic parameters from quicksort to bitonic sort, transforming the sorting approach to achieve deterministic time complexity. This parameter change enables predictable performance by ensuring that the sorting operation always completes within a predetermined number of steps, eliminating the variability inherent in quicksort's best-case and worst-case scenarios.
Solution Approach 2:
The bitonic sorting algorithm segments the sorting process into distinct phases: comparison phases where pairs of elements are compared and swapped based on control signals, and permutation phases where elements are reordered. This segmentation allows each phase to be executed in a predetermined number of steps, contributing to overall deterministic timing.
2Device complexity
If quicksort algorithm is used for sorting, then no additional hardware is required, but the processing time varies depending on data order and can exceed worst-case scenarios
Solution Approach 1:
The patent introduces dynamic control signals that selectively enable or disable comparison and swap operations based on the current sorting state. These control signals are generated in a predetermined sequence, allowing the bitonic sort to adapt its operations dynamically while maintaining deterministic overall execution time. The dynamics principle is applied through the use of phase-dependent control logic that directs the sorting network's behavior.
Solution Approach 2:
The patent introduces control signals as intermediaries that mediate between the sorting algorithm and the vector processing units. These control signals orchestrate the comparison and swap operations, ensuring that each phase completes in a predetermined number of steps. The intermediary control layer provides the deterministic timing framework while allowing flexible sorting operations.
3Productivity
If sorting is performed using CPU instead of vector extension, then vector processing capabilities are not utilized, but sorting can be performed with existing resources
Solution Approach 1:
The patent merges the sorting function with the vector processing capabilities of the RISC-V extension. By implementing bitonic sort using vector instructions, the sorting operation leverages the parallel processing power of vector units while maintaining integration with the existing CPU architecture. This merging allows simultaneous utilization of vector hardware resources for sorting operations.
Solution Approach 2:
The patent makes the vector processing unit multi-functional by enabling it to perform both traditional vector operations and sorting operations. The bitonic sorting algorithm is implemented using general vector instructions, allowing the same hardware resources to serve multiple purposes: data processing, data transformation, and data sorting. This universality increases productivity without requiring dedicated sorting hardware.
Data Source
AI summary
A method comprising comparing pairs of elements in an array. The method comprises generating two new vectors from an original array in memory, each comprising one part of each pair to be compared. The two new vectors are compared to generate a mask which indicates which of each pair of elements is less. Based on the mask, elements of the vectors can be swapped as necessary. This could be using a XOR algorithm or a merge algorithm. The vectors are then written back to the original array in memory. This process can be repeated on elements of an array as part of a bitonic sorting algorithm.