Vector Processor LUT Reuse for Convolution and Function Approximation

Solution Overview

Problem

The existing vector processors face challenges in efficiently utilizing the Look-Up Table (LUT) memory, which occupies a significant space, limiting its use to function approximation and impacting computational performance for other operations.

Innovation Solution

The vector processor incorporates a LUT memory that stores data corresponding to index values, allowing it to perform operations beyond function approximation, including convolution, depthwise convolution, activation, pooling, and normalization, by using a controller to manage data storage and retrieval efficiently.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the LUT memory is used to store coefficients for function approximation, then the accuracy of function approximation is improved, but the LUT memory occupies a large amount of space and cannot be used for other operations

Engineering Contradiction:

Improveaccuracy of function approximationVSAvoidversatility of LUT memory

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The LUT memory is designed to serve multiple functions: it can store coefficients for function approximation (improving accuracy) while simultaneously storing data for other operations such as convolution, depthwise convolution, activation, pooling, and normalization (improving versatility). This multi-functional design resolves the contradiction by allowing the same memory resource to be utilized for diverse computational tasks.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The controller dynamically manages the LUT memory by identifying index values based on instructions and storing appropriate data in the LUT memory using these index values. This dynamic allocation allows the LUT memory to switch between storing function approximation coefficients and other operational data, thereby achieving both high accuracy in function approximation and high versatility in supporting multiple operations.

Inventive Principle:

Principle #15Dynamics

2Measurement precision

If the LUT memory is dedicated to function approximation, then the accuracy of function approximation is improved, but the computational performance for other operations deteriorates

Engineering Contradiction:

Improveaccuracy of function approximationVSAvoidcomputational performance

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The LUT memory is designed to serve multiple functions: it can store coefficients for function approximation (improving accuracy) while simultaneously storing data for other operations such as convolution, depthwise convolution, activation, pooling, and normalization (improving versatility). This multi-functional design resolves the contradiction by allowing the same memory resource to be utilized for diverse computational tasks.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The controller dynamically manages the LUT memory by identifying index values based on instructions and storing appropriate data in the LUT memory using these index values. This dynamic allocation allows the LUT memory to switch between storing function approximation coefficients and other operational data, thereby achieving both high accuracy in function approximation and high versatility in supporting multiple operations.

Inventive Principle:

Principle #15Dynamics

3Productivity

If additional storage space is allocated for other operations, then the computational performance for various operations is improved, but the total device complexity increases

Engineering Contradiction:

Improvecomputational performanceVSAvoiddevice complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The patent merges the function approximation storage and other operation data storage into a single LUT memory structure. By using the same memory resource for multiple purposes through dynamic management, the device avoids the complexity of implementing separate storage structures, thereby improving computational performance without significantly increasing device complexity.

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentUS20260003513A1Vector processor and operation method thereof

Publication Date: 2026.01.01 REBELLIONS INC

AI summary

A vector processor and an operation method of the vector processor are disclosed. Specifically, the vector processor may include a look-up table (LUT) memory in which data corresponding to an index value is stored, a processing unit configured to perform an operation based on the data, and a controller configured to identify a first index value based on an instruction and store first data in the LUT memory using the first index value.