Intel® FPGA SDK for OpenCL™ Pro Edition: Best Practices Guide

ID 683521
Date 3/28/2022
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

5.4. Reducing Area Resource Use While Profiling

Due to various performance counters being added to the pipeline, introducing profiling into your design can result in a large amount of area resource use. This may be inconvenient for particularly large designs as adding profiling performance counters might result in no fit errors.
To reduce the amount of area resources that profiling takes up, you can choose to profile with shared performance counters. This profiling mode allows counters to be shared by various signals over multiple design runs to reduce the number of performance counters added to the design. During runtime, the Profiler Runtime Wrapper runs the host application four times, where, for each run, the counters count a different signal.
Note: You must invoke the Profiler Runtime Wrapper only once.

To turn on the shared performance counters profiling mode, perform these steps:

  1. Include the -profile-shared-counters flag along with the -profile flag during your aoc compile.
  2. Include the -sc flag when running your design with the Profiler Runtime Wrapper.

    Without the -sc flag, your design runs only once so, you lack data for everything after the first shared signal.

    CAUTION:
    The shared performance counters profiling mode works well only for kernels and designs that are deterministic. Because the host application and design are run multiple times to collect all of the data, non-deterministic designs result in shared data that is difficult to combine, and it may be difficult to determine where design problems occur temporally.