Intel® Advisor User Guide

ID 766448
Date 12/16/2022
Public

A newer version of this document is available.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

Document Table of Contents

Compare GPU Roofline Results

Use the Roofline Compare functionality to display Roofline chart data from other Intel® Advisor results or non-archived snapshots for comparison purposes to track optimization progress.

Prerequisites

To compare the GPU Roofline results, make sure to get the following:

  • A baseline GPU Roofline result or snapshot
  • One or more GPU Roofline results or snapshots of the same application with an optimization applied
IMPORTANT:
You can only compare Roofline results of the same type: CPU Roofline or GPU Roofline.

To compare the results:

  1. Open a baseline GPU Roofline result/snapshot.
  2. From the Compare drop-down toolbar, click + to load a comparison result/snapshot. You can load multiple results/snapshots for comparison one by one.

When the comparison is uploaded:

  • The filenames for uploaded results/snapshots are displayed in the Compared Results region.
  • Similar kernels from all compared results are recognized automatically. They are connected with a dashed arrow line. The performance improvement between the kernels is shown above the line, in per cent. The improvement is calculated as the difference in FLOPS, INTOPS, or OPS and Total Time.
    NOTE:
    The arrows showing the relationship among kernels do not reappear if you upload a new comparison file.
  • Kernels from different compared results are shown as different icons on the chart. For example, for three , the baseline kernels can be shown as circles and comparison kernels can be triangles and diamonds.
  • To highlight all dots from a specific compared result, open the Compare drop-down and hover over the result name.
  • Each time you change the Roofline configuration or filter the dots on the chart, the comparison is updated automatically.
  • You can remove a selected result from Compared Results by hovering over it and clicking the X icon. The result is removed from the chart and appears in the Ready for comparison region. Click a name in the Ready for comparison region to reload the result back to the chart.
  • You can save the comparison itself to a file using the export feature.
NOTE:
To find the same kernels among the results, Intel Advisor compares several kernel features, such as their type, nesting level, source code file name and line, and function name. When a certain threshold of similar or equal features is reached, the two kernels are considered a match and connected with a dashed line.

However, this method still has few limitations. Sometimes, there can be no match for the same kernel if one is optimized or moved in the source code to four or more lines from the original place. Intel Advisor tries to ensure some balance between matching source code changes and false positives.