Near Divide Accumulate Instruction Shift Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing divide and add/accumulate instructions require significant time and energy due to resource-intensive division operations, even when precision is not needed, as they perform accurate division for all operand values.

Innovation Solution

A near divide and add/accumulate instruction that employs a shift operation when an operand is equal to or near a power of two, reducing the need for division and thereby minimizing resource usage while maintaining accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If accurate division operations are performed for all operand values, then computation precision is maintained, but processing time and energy consumption increase significantly

Engineering Contradiction:
Improvecomputation precisionVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent changes the computational approach based on the parameter characteristics of the dividend. When the dividend is detected to be a power of two, the computation method is changed from accurate division to bit shifting, which maintains sufficient precision for such cases while dramatically reducing processing time and resource consumption.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent applies partial action by performing accurate division only when necessary (when the dividend is not a power of two). For power-of-two dividends, a simpler approximation method (bit shifting) is used, which provides sufficient precision for the specific case while reducing overall computational burden across all operations.

Inventive Principle:
Principle #16Partial or excessive action

2Measurement precision

If accurate division operations are performed for all operand values, then computation precision is maintained, but energy consumption increases significantly

Engineering Contradiction:
Improvecomputation precisionVSAvoidenergy consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The computational method is dynamically changed based on the dividend parameter. When the dividend is a power of two, the energy-intensive accurate division is replaced with low-energy bit shifting operations, significantly reducing energy consumption while maintaining adequate precision for power-of-two cases.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

Accurate division is performed only partially - specifically only when the dividend is not a power of two. This selective approach reduces overall energy consumption by avoiding unnecessary high-energy operations for cases where simpler methods suffice.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If division operations are performed for all cases, then accurate results are obtained, but resource usage increases unnecessarily when operands are powers of two

Engineering Contradiction:
Improveresult accuracyVSAvoidprocessing efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent changes the operation type based on the dividend parameter. When the dividend is a power of two, division operations are changed to bit shifting operations, which improves processing efficiency by using simpler, faster hardware operations while maintaining sufficient result accuracy for this specific case.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent applies division operations partially - only when the dividend is not a power of two. For power-of-two dividends, more efficient bit shifting is used, thereby improving overall processing efficiency by reducing the frequency of resource-intensive division operations while maintaining adequate accuracy.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4016289A1Efficient divide and accumulate instruction when an operand is equal to or near a power of two
Publication Date: 2022.06.22 INTEL CORP
  • EP4016289A1 patent drawingFigure 1
  • EP4016289A1 patent drawingFigure 2
  • EP4016289A1 patent drawingFigure 3

AI summary

Techniques and apparatuses for performing a near multiply and accumulate instruction are described. An apparatus includes decoder circuitry to decode an instruction, the instruction to include a field for an identifier of a first source operand, a field for an identifier of a second source operand, and a field for an identifier of a third source operand. The apparatus also includes execution circuitry to execute the decoded instruction to perform a division on a pair of data elements from the first and second source operands to produce a quotient data element via a shift operation when at least one data element in the pair of data elements is equal to a power of two or near a power of two or via a division operation on the pair of data elements when the pair of data elements is neither equal to a power of two or near a power of two.