User Guide

  • 2021.2
  • 06/24/2021
  • Public Content
Contents

Graphics Trace Analyzer Window: Platform View

Use the
Platform View
window to analyze concurrent execution of all your application threads, as well as correlate CPU and GPU activity.
To access
Platform View
window, select a trace for analysis in the Open Trace Capture window, and then click
Open
, or double-click the trace preview. The
Platform View
window graphically represents a distribution of performance data over time.
Trace title
bar
Displays the opened trace file name. The file name consists of the application name, capture date and time in the following format: YYYY_MM_DD_HH_MM_SS.
Main toolbar
Displays the opened trace data, notifications, opens the Intel GPA User Guide, restores the default view, and changes the Intel GPA color scheme.
Time Line Viewer toolbar
Filters tracks, shows event duration, zooms the selection in and out.
Time Line Viewer pane
Visualizes the opened trace content.
Trace Content pane
Lists the opened trace content.
Synchronization Flow Graph
The arrows reveal event dependencies on the time line. Identify inefficient events synchronization leading to incomplete GPU load. To see the graphs, click one of the objects. Red arrows demonstrate parent-child relationships between events. Green arrows connect synchronization objects in UMD Driver Queues (Device Context Pane) and CPU threads. They also show synchronization points between objects. If you want to see several dependencies at a time, press and hold CTRL and pick objects by clicking.
This feature is only available on Windows* hosts.
Selection Statistics pane
Shows detailed information about the selection in the
Trace Content
pane.
The pane is only visible when any event or group of events is selected, unless the
Aggregated selection summary
is enabled in the
Settings
pane.

Main Toolbar

Back
button
Returns to the
Open Trace Capture
window.
Trace Overview
pane
Provides the following statistical information for the analyzed trace:
  • Events count
    :
    • Total number of events in the trace
    • Number of events in the visible area of the
      Time Line Viewer
  • Selected count
    :
    • Total number of selected events in the trace
    • Number of selected events in the visible area of the
      Time Line Viewer
View Options
pane
Provides controls for trace display customization:
  • Highlight selection
    button - highlight selected events and their relative position on to other events the
    Trace Content
    pane using vertical bars.
  • Cross-Track dependency selection
    - enable/disable automatic highlighting of all related events from multiple tracks, as opposed to only highlighting the selected event.
  • Track utilization duration in time units/percents
    button - switch between percents and time units in track utilization duration.
Message
pane
Displays messages and non-critical errors.
If more than one message is open, use the Up or Down arrows to navigate between messages.
Trace Info
button
Provides the following properties of the analyzed trace:
  • Intel® GPA version
  • Name of the system on which the trace was captured
  • Hardware and software of the system on which the trace was captured
Settings
button
  • Enables the following:
    • Highlight selection - highlights the selected events
    • Cross-track dependency - simultaneous selection of dependent events
    • Bottom-up queues - changes order of GPU/CPU queues visualization
    • Fully-qualified function names - shows the full function name along with the name of the interface the function belongs to
    • Aggregated selection summary - pins the
      Selection Statistics
      pane so that it remains open after deselecting events
  • Adjusts UI scale
  • Restores the default layout and track colors
  • Changes the color scheme of the
    Platform View
    interface
Send feedback
button
Opens a form to send feedback to the Intel GPA technical support.
Help
button
Displays information on navigation controls and opens the Intel GPA online documentation.

Time Line Viewer Toolbar

The
Time Line Viewer
toolbar enables you to search and filter events, assess event position and duration on the time line, and adjust the manner of displaying the trace in the
Time Line Viewer
pane. The toolbar includes the following options:
  • Type Filter Expression
    field searches and filters events based on lexicographical comparison.
  • Match Whole Word
    button enables searching and filtering events by whole names.
  • Select Filtered Tasks
    button selects filtered events in the
    Time Line Viewer
    pane.
  • Time line ruler correlates all the displayed events with the time line. The range of the time line corresponds to the length of the trace in the current zoom level. Depending on the zoom-in level, the time intervals are displayed in seconds, milliseconds, microseconds, on nanoseconds. The default trace length is 5s.
    To mark a specific timestamp, you can create a bookmark on the time line ruler by double-clicking the necessary point on the time ruler.
  • Zoom to Selection
    button zooms into the time range of selected events.
  • Full Zoom Out
    button restores the default zoom level.

Time Line Viewer Pane

Use the Time Line Viewer pane to view and configure time line tracks of the traced processes. Each track comprises the following areas:
  • Track Description
    consists of a track name and name of the process or platform, which the track belongs to, as well as the percentage or time when the device was utilized for the trace area visible in the current zoom level.
  • Track Data
    is graphic representation of performance events and other data, such as metrics and queue packages.
  • Track Toolbar
    closes tracks and shows metric ranges.
    • CPU tracks
      Visualize aggregated CPU activity reflecting thread execution on CPU cores. Use to analyze thread execution order, execution duration, and distribution between CPU cores. Blocks of the same color represent the same thread. If you zoom into the trace, you can see the process name the thread belongs to and thread execution duration.
    • GPU queues
      Show GPU queues for all active video adapters generating graphics content. GPU queues can be of different types depending on the application. The Time Line Viewer pane visualizes GPU utilization over time: execution of the DMA packages on GPU. DMA package color corresponds to the color of the thread, from which the DMA package was submitted. All DMA packages have names visible on mouse hover. Additionally, DMA packages, which are essential for analysis, are marked with different icons depending on their type:
      • Render package. Render package with a present call is crosshatched.
      • Signal package
      • Wait package
      • Paging package
      Selecting any of these packages shows an arrow that reveals calls in CPU threads related to that package. For example, you can trace the origin of a Render package from the CPU thread, to the User-Mode Driver, and up to the hardware queue.
    • Flip queues
      Shows flip queues for all active video adapters. Flip queue reflects the relationship between the application present calls, present packages of GPU/CPU queues, and Vertical Synchronization (VSync) event of the monitor. Flip queue package consists of two blocks: solid-color and crosshatched. Solid-color block shows the time when Desktop Window Manager (DWM) is generating the content that needs to be displayed. The crosshatched block shows the time during which the content waits for the moment to be displayed (VSync).
      For applications utilizing layered flip queues, multiple layers can be displayed as sub-tracks of the general Flip Queue track.
    • CPU queues
      Shows CPU queues for all processes generating graphics content. CPU queue represents ordered command packages to be executed, but not yet submitted to GPU. CPU queues can be of different types depending on the application. CPU queue package color corresponds to the color of the thread, from which the CPU queue package was submitted. All CPU packages have names visible on mouse hover. Additionally, CPU packages, which are essential for analysis, are marked with different icons depending on their type:
      • Render package. Render package with a present call is hatched for packages named Present Token and crosshatched for packages named Present Render.
      • Signal package
      • Wait package
      • Paging package
      CPU synchronization is represented as
      WaitForSingleObject
      ,
      WaitForMultipleObjects
      , and
      SetEvent
      function calls on the time line track of the threads. You can use these captured events to profile different synchronization issues.
      Click on any of these events to visualize dependencies between synchronization events in the form of arrows. For example, an arrow pointing from a
      SetEvent
      call in one thread to a
      WaitForSingleObject
      call in a different thread indicates that a thread was unblocked by a
      SetEvent
      call in another thread.
      On CPU queues tracks, areas where a thread was active are highlighted as green bars. Areas where a thread was idle are shown as gray bars.
      A thread is considered active and is highlighted accordingly if this thread was actively executed on a CPU core at this time.
    • Events
      Shows events as markers that have a timestamp, but do not have duration. Markers scope can be global or process-defined. Markers are visualized as colored triangles. Global markers are placed on the Time Line ruler; process-defined markers are shown on executed threads.
    • Metrics
      Shows any GPU/CPU metrics, which you enabled in System Analyzer or System Analyzer HUD.
    • Regions
      Shows regions, which are logical application execution blocks, defined with Intel® ITT API or graphics debug API. By default, block color corresponds to the color of the time line track. Each block has a name and duration.
    • Threads
      track
      Shows executed threads of profiled processes. Use this data to analyze your application performance and behavior based on the Intel® ITT API and system events. Each colored block represents a logical block of application execution marked up by the user or generated by the system. Each block has a name and duration, and can have nested blocks. By default, block color corresponds to the color of the time line track.
    • Parallel Execution
      track
      Parallel Execution track visualizes how the driver parallelizes execution of submitted render events.

Trace Content Pane

Type Filter Expression
field
Filters trace file tracks by name or parameters. Saves previous search keywords.
Collapse All Groups
button
Closes all unfolded track groups.
Expand All Groups
button
Unfolds all track groups.
Track list
Includes aggregated information on trace file content in the following order:
  • Bookmarks
    enables you to mark any timestamp on the Time Line ruler. You can change the name and color of a Bookmarks Group.
    To add a bookmark:
    1. Select the required Bookmarks Group to enable a bullet selector.
    2. Click on a desired timestamp on the Time Line ruler.
    To delete a bookmark, do the following:
    1. Select the required Bookmarks Group to highlight it and make the corresponding bookmarks visible.
    2. Click on a bookmark to be deleted.
    3. Press
      Delete
      keyboard button.
    On the Time Line ruler, you can see only the bookmarks of the Bookmarks Groups highlighted in the Trace Content pane. You can edit only a Bookmarks Group that has a bullet selector.
  • Aggregated CPU activity
    (CPU context switches)
  • Active GPU adapters
    (GPU frames; GPU queues; Flip queues; VSync events)
  • Target instrumented process
    for which the opened trace was captured
  • Other processes
    in the system contributed to the trace
  • Target instrumented processes and other processes contain CPU queues, executed threads, system events, regions, and metrics.
  • To change track color or event color in the Trace Viewer, use the
    Track Color Modification
    button or
    Marker Color Modification
    button, respectively.
  • Track from the Track list can be dragged and dropped to the Trace Viewer.

Selection Statistics Pane

Use the
Selection Statistics
pane to view summary information for the selected objects in the Trace Viewer. You can select objects in the Trace Viewer in several ways:
  • Single Selection
    Left-click any object in the track.
  • Group Selection
    Left-click any objects in tracks while keeping the
    CTRL
    keyboard button pressed down.
  • Rectangular Selection
    Press the left mouse button and drag the mouse in any desired direction to select the required region.
  • Range Selection
    Press the right mouse button and drag the mouse horizontally to select events in all tracks within the selected range. You can change the selection by dragging the right or left boundary. The selected range includes all the intercepted tasks.
Intel GPA: Selection Statistics Pane
Aggregated Summary table
All the selected events are grouped by event types:
  • Packages
  • Regions
  • Markers
  • Metrics
For all the selected events you can view the following aggregated information:
  • Name - event name.
  • Track Visibility Status:
    • Track is visible on working area - track is opened in the Trace Viewer.
    • Track is not visible on working area - track is not opened in the Trace Viewer. You can open the track by clicking the button.
  • Track - track name.
  • Count - number of events with the same name in one track.
  • Inclusive Time (ns) - total event duration or sum of total event durations, if the count value exceeds one.
  • Exclusive Time (ns) - total event duration or sum of total event durations, excluding duration of child events, if any.
  • Exclusive % - relation of Exclusive Time to Inclusive Time.
Itemized Summary table
For each selected item in the
Aggregated Summary
table you can view the list of events with the following information:
  • Timestamp (ns) - relative time passed from the beginning of the trace before the task execution started.
  • Inclusive Time (ns) - total event duration.
  • Exclusive Time (ns) - total event duration, excluding duration of child events, if any.
  • Exclusive % - relation of Exclusive Time to Inclusive Time.
Performance Event Details table
Shows arguments for the event selected in the
Itemized Summary
table. Available for GPU queue packages, CPU queue packets, and VSync events.
To customize the displayed columns, click the button and enable/disable the required column.

Product and Performance Information

1

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