MAC Pipeline Filter Weight Conversion for Wider Dynamic Range
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing pipelines face limitations in dynamic range and memory efficiency when using filter coefficients for image filtering, as they often require increased memory allocation and precision to maintain accuracy, which can lead to higher costs and reduced performance.
Innovation Solution
The implementation of data format conversion circuitry that converts filter coefficients from integer or floating-point formats to block-scaled fraction formats, allowing for a wider dynamic range without significantly increasing memory footprint, and further conversion to floating-point or Winograd formats for enhanced processing capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If filter coefficients are converted to higher precision data formats to maintain filtering accuracy, then measurement precision is improved, but device complexity and memory allocation increase
Solution Approach 1:
The patent changes the data format parameter of filter coefficients from standard floating-point or integer formats to a custom block-scaled fraction format. This parameter change allows the system to represent filter coefficients with enhanced dynamic range using the same memory allocation, thereby maintaining filtering accuracy without increasing device complexity or memory requirements
Solution Approach 2:
The patent introduces a block-scaling dimension to the data representation by grouping multiple filter coefficients into blocks and applying a common scale factor to each block. This dimensional approach allows efficient representation of many coefficients with reduced storage requirements while maintaining the precision needed for accurate filtering operations
2Measurement precision
If memory allocation is increased to store more precise filter coefficients, then measurement precision is improved, but loss of substance increases
Solution Approach 1:
By changing the data format parameter to block-scaled fraction format, the system achieves higher effective precision without proportionally increasing memory allocation. The block-scaling mechanism allows multiple coefficients to share common representation elements, reducing overall memory consumption while maintaining the precision required for accurate filtering
Solution Approach 2:
The patent applies partial precision enhancement selectively through block-scaling, where only the necessary portion of the coefficient representation is enhanced beyond standard formats. This selective approach provides sufficient precision for filtering accuracy while avoiding excessive memory allocation that would occur with uniform high-precision representation of all coefficients
Data Source
AI summary
A system and/or an integrated circuit including: (a) a multiplier-accumulator execution pipeline including multiplier-accumulator circuits to process image data, using associated filter weights, via a plurality of multiply and accumulate operations and (b) first data format conversion circuitry including (i) inputs to receive filter weights of a plurality of sets of filter weights, wherein each set includes a plurality of filter weights each having a block-scaled fraction data format, (ii) conversion circuitry, coupled to the inputs, to convert the filter weights of each set from the block-scaled fraction data format to a floating point data format, and (iii) outputs to output the filter weights having the floating point data format. In operation, the multiplier-accumulator circuits of the pipeline are configured to perform the plurality of multiply and accumulate operations using (a) the image data and (b) the filter weights having the floating point data format.