Skip to content

Latest commit

 

History

History
82 lines (55 loc) · 3.22 KB

chrome_tracing.md

File metadata and controls

82 lines (55 loc) · 3.22 KB
Raw

Using the Intercept Layer for OpenCL Applications with Chrome Tracing

The Intercept Layer for OpenCL Applications includes support for generating JSON files compatible with the Google Chrome built-in profiler front end. This document describes how to use the Intercept Layer for OpenCL Applications to visualize how an OpenCL application executes using Chrome Tracing.

Background and Setup

The Chrome Tracing Format is described here.

To start up the Chrome Tracing front end, open Chrome and enter chrome://tracing as your "URL". This should give you a Window that looks like this one:

Empty Chrome Tracing Window

Configuring Chrome Tracing

There are (currently) three Chrome Tracing-related controls for the Intercept Layer for OpenCL Applications:

  • ChromeCallLogging: This is the control for tracing OpenCL host APIs. It will plot OpenCL calls for each thread of the host application, similar to those dumped when CallLogging is enabled.
  • ChromePerformanceTiming: This is the control for tracing OpenCL device commands. It will plot OpenCL commands for each command queue created by the application, similar to those dumped when DevicePerformanceTiming is enabled.
  • ChromePerformanceTimingInStages: This is a further control that sits on top of ChromePerformanceTiming(it does nothing when this flag is not on). When it is on it splits up the events into "Queued", "Submitted", and "Execution" stages, and reorders the calls approximately by start time.

Collecting Chrome Tracing Data

After setting some combination of these controls, run your application, and you should see a "CLIntercept_trace.json" file in your CLIntercept_Dump directory.

Visualizing Chrome Tracing Data

After collecting a "CLIntercept_Trace.json" file, simply click the "load" button in the "chrome::tracing" UI, or drag your file into Chrome.

If all goes well you will see a timegraph like this one:

Chrome Tracing Example

You can navigate around the timegraph with 'wasd' controls similar to many popular games: w zooms in, s zooms out, a goes backwards in time, and d moves forwards in time.

Let's zoom in a bit and look at the what's in the timegraph:

Chrome Tracing Detail

When the ChromePerformanceTimingInStages flag is also set, the calls are further split into stages and reorganized; here is what that looks like on a timegraph:

Chrome Tracing with Stages Detail

Overhead

Empirically, the overhead of Chrome Tracing is very low. The difference in scores between a run of LuxMark without Chrome Tracing vs. enabling ChromeCallLogging and ChromePerformanceTiming was less than 1%, and the trace file size for the 2+ minutes of execution was 26MB. Similarly, ChromePerformanceTimingInStages does not create a noticeable performance hit, however the trace file size can be up to 3x larger than without it.

* Other names and brands may be claimed as the property of others.

Copyright (c) 2018-2019, Intel(R) Corporation