Skip to content

Non-optimal II optimization for Pointwise Conv1D #1223

Open
@Duchstf

Description

@Duchstf

Details

The current hls4ml pointwise conv1d is not pipelined, and when being forced to pipeline it seems that it can perform better than the current standard implementation in hls4ml

New behavior

Better handles on optimizing IIs

Motivation

II is very important in the context of CMS level 1 trigger

Parts of hls4ml being affected

pointwise conv1d hls template.

For detailed comparisons I attached a set of slides here, after a quick discussion with Javier about how the latest hls4ml branch works it might be that my config might not be properly set. If it's the case then it should also be properly documented as well.

hls4ml_pointwise.pdf

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions