User Guide

Contents

Window: GPU Roofline Regions

After running
GPU Roofline Insights
perspective,
Intel® Advisor
generates a GPU Roofline report that shows detailed metrics for parts of your application executed on a GPU and for parts executed on a CPU. To access this report, click
GPU Roofline Regions
button on top of the results window.

Controls

Review the summary metrics for parts of your application executed on an accelerator.
  • Create a snapshot for the current project results. For details, see Create a Read-only Result Snapshot .
  • Click a
    +
    button to open previously closed panes. With this button, you can add the following panes:
    • CPU Roofline
      pane that enables you to view the actual performance of functions/loops executed on a CPU against hardware-imposed performance ceilings visualized on a Roofline chart.
      For details about interpretation, see Examine Bottlenecks on CPU Roofline Chart .
    • CPU
      pane that enables you to review performance metrics of your application performance on a CPU and compare tem with performance metrics on an accelerator. For details, see CPU Metrics .
Review the actual performance of GPU kernels in your application against hardware-imposed performance ceilings using the
GPU Roofline
chart.
For details about interpretation, see Examine Bottlenecks on GPU Roofline Chart .
See detailed description of GPU Roofline chart controls in Pane: GPU Roofline Chart.
Use the
GPU Details
tab to view the detailed information about the execution of a selected kernel:
  • View program metrics for a selected kernel in the
    Summary
    pane.
  • Identify the memory level your selected kernel is bounded by using the
    Roofline Guidance
    pane.
  • Explore the compute operations count and memory level utilization metrics in the
    OP/S and Bandwidth
    pane. Use the drop-downs to view the operations count, memory traffic, and arithmetic intensity (AI) for floating-point and integer operations at different memory levels.
  • View how the selected kernel impacts each memory level and explore the amount of data passed through each memory level using the
    Memory Metrics
    pane.
  • Explore the ratio of compute, memory and other instructions grouped by types in the
    Instruction Mix
    pane.
  • Get detailed overview of instruction types used during the execution of your application using the
    Instruction Mix Details
    pane. Use the drop-downs to expand each instruction category and view the included instruction types and instruction count. For compute category,
    Intel Advisor
    determines the data type. The dominating data type in th entire kernel is highlighted blue. Filter instructions by type and dominating data type using a filter button.
  • View how th loops in a selected kernel utilize the execution unit (EU) in the
    Performance Characteristics
    pane.
Switch between
GPU Source
and
GPU Assembly
tabs to:
  • Examine the source code and offload details for each source line. Select a loop in the
    GPU
    table or a dot in the
    GPU Roofline
    to focus on the corresponding parts of source and assembly code.
  • Review GPU assembly representation for a selected kernel. Select a code line to highlight the corresponding part in source code.
For details about interpreting
GPU Roofline Insights
perspective results, see Explore GPU Roofline Results
Review performance metrics of your application performance on a GPU accelerator. For details about metrics, see Accelerator Metrics .

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.