Intel® Advisor User Guide

ID 766448
Date 11/07/2023
Public

A newer version of this document is available.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

Document Table of Contents

Problem: Broken Call Tree

Symptoms

After executing the Offload Modeling perspective, in the Accelerated Regions tab, you see one of the following:

  • A code region is duplicated.

  • A code region is located at a wrong place.

  • A code region has incorrect number of trip counts reported in any column of the Trip Counts column group.

  • A code region with your code has a System Modulediagnostics message and Cannot be modeled: System Modulereason for not offloading.

Any of these symptoms mean that the Intel® Advisor detected the application call tree incorrectly during Survey.

Details

A broken call tree often happens if you use a program model with SYCL or Intel® oneAPI Threading Building Blocks. These program models run code in many threads using a complicated scheduler, and the Intel Advisor sometimes cannot correctly detect their call stacks. As a result, some code instances might have no metrics or incorrect metrics in a report and a call tree is broken.

Cause

This can happen due to the following reasons:

  • Call stacks were detected incorrectly.

  • A heavy optimization was used.

  • Debug information has issues.

Possible Solution

NOTE:

This is not an issue if all hotspots and code you are interested in are outside of the broken part of the call tree. You can ignore it in this case.

To fix a broken call tree, do the following:

  • Make sure you compiled binary with -g option.

    NOTE:

    You can recompile it with the -debug inline-debug-info option to get enhanced debug information.

  • Recompile the binary with a lower optimization level: use -O2.

  • If you collect performance metrics with advisor CLI: When running the Survey analysis, try the following:

    • Remove --stackwalk-mode=online option if you used it when running the Survey analysis.

    • Add --no-stack-stitching option.

  • Offload only specific code regions if their estimated execution time on a target device is greater than or equal to the original execution time. Rerun the performance modeling with --select-loops to specify loops of interest and --enforce-offloads to make sure all of them are offloaded. For example:

    advisor-python <APM>/analyze.py <project-dir> --select-loops=[<file-name1>:<line-number1>,<file-name1>:<line-number2>,<file-name2>:<line-number3>] --enforce-offloads
    NOTE:
    Replace <APM> with $APM on Linux* OS or %APM% on Windows* OS.

    For details, see Enforce Offloading for Specific Loops

  • If you model a multithreaded code that runs with a complicated scheduler, you might see a code region with suspiciously low trip counts and multiple instances of the same region loop present in the scheduler. This means that the Offload Modeling could not correctly detect the call stacks. Use the --enable-batching option to artificially increase the number of trip counts by using total number of executions instead of average number trip counts.

See Also