Intel® Advisor User Guide

ID 766448
Date 3/22/2024
Public

A newer version of this document is available.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

Document Table of Contents

Run GPU Roofline Insights Perspective from GUI

Prerequisites:

  1. Configure system to analyze GPU kernels.
  2. In the graphical-user interface (GUI): Create a project and specify an analysis target and target options.

To configure and run the GPU Roofline Insights perspective from the GUI:

  1. Configure the perspective and set analysis properties, depending on desired results:
    • Select a collection accuracy level with analysis properties preset for a specific result:
      • Low: Analyze performance of kernels executed on GPU and plot a GPU Roofline chart for all memory levels. Plot a basic CPU Roofline chart for loops/functions executed on CPU.
      • Medium: Analyze performance of kernels executed on GPU, plot a GPU Roofline chart for all memory levels, and model the application performance to get more optimization recommendations. Plot a basic CPU Roofline chart for loops/functions executed on CPU.
      • High: Analyze performance of kernels executed on GPU, plot a GPU Roofline chart for all memory levels, and model the application performance to get more optimization recommendations. Plot an extended CPU Roofline chart for loops/functions executed on CPU for all memory levels.
    • Select the analyses and properties manually to adjust the perspective flow to your needs. The accuracy level is set to Custom.

    By default, accuracy is set to Low. The higher accuracy value you choose, the higher runtime overhead is added to your application. The Overhead indicator shows the overhead for the selected configuration. For the Custom accuracy, the overhead is calculated automatically for the selected analyses and properties.

    For more information, see GPU Roofline Accuracy Presets.

    NOTE:
    If you want to analyze only code regions executed on GPU, select the Low or Medium accuracy. This decreases analysis overhead.

  2. If you have multiple GPUs connected to your system, select a target GPU to collect data for from the Target GPU drop-down.
    • To run the analysis for all available GPUs, select All devices.
      NOTE:
      To include only certain GPUs in data analysis, run GPU Roofline Insights perspective from a command line using the target-gpu option, as described in Run GPU Roofline Insights Perspective from Command Line.
    • For the multi-tile GPUs, tiles are not displayed in the analysis settings. However, if you select a multi-tile GPU as a target for your analysis, Intel® Advisor will analyze all its tiles and display detailed information about each of them in the GPU Roofline report.
      NOTE:
      Multi-tile GPUs are supported starting with Intel® Advisor 2024.0.
    The drop-down shows an adapter address and a name for each GPU available. The address has the following format: <domain>:<bus>:<device-number>.<function-number>. Here all values are decimal numbers.
  3. From the Analysis Types, select Survey > GPU Profiling or Characterization > FLOP and GPU Profiling.
  4. Run the perspective: click button.

    While the perspective is running, you can do the following in the Analysis Workflow tab:

    • Control the perspective execution:
      • Stop data collection and see the already collected data: Click the button.
      • Pause data collection: Click the button.
      • Cancel data collection and discard the collected data: Click the button.
    • Expand an analysis with to control the analysis execution:
      • Pause the analysis: Click the button.
      • Stop the currently running analysis and start the next analysis selected: Click the button.
      • Interrupt execution of all selected analyses and see the already collected data: Click the button.

To run the GPU Roofline Insights perspective with the Low accuracy from the command line interface:

advisor --collect=roofline --profile-gpu --project-dir=./advi_results -- ./myApplication

See Run GPU Roofline Insights from Command Line for details.

NOTE:
To generate command lines for selected perspective configuration, click the Command Line button.

Once the GPU Roofline Insights perspective collects data, the report opens showing a Summary with performance metrics measured for CPU- and GPU-executed parts of your application and preview Roofline charts. Continue to examine GPU bottlenecks on the Roofline chart to investigate the results.