JavaFX Direct3D 12 - Call for performance testing help

[Lukasz Kostyra]

Hello openjfx-dev,

Because Windows is very open hardware-wise, it is difficult to prepare the backend for all possible hardware combinations available, especially from performance perspective. To make sure JavaFX performance does not degrade compared to old D3D backend we would like to call for volunteers to help performance-test the backend. These tests will let us get a general idea on how the backend behaves on different hardware and which areas of the backend to focus on while moving forward with optimization effort for Direct3D 12.

At this point we tested the backend quite extensively on Intel-based integrated GPUs and did some testing on a machine running a recent discrete Nvidia GPU. We are primarily looking for testing Direct3D 12 backend on a system running an AMD discrete GPU, but any hardware combinations are welcome - the more the merrier :). Also note that these tests require a Windows machine, as D3D12 backend is Windows-only.

We run performance testing using RenderPerfTest JavaFX app located in the JavaFX repository under "tests/performance/animation/RenderPerfTest". I wrote a bash script to use with this app for performance-testing the backend. The script will run all available demos on RenderPerfTest with a set amount of objects per test, average the FPS results from each run and output the results into a CSV file.


How to run perf tests:

  1.  Download the test script - https://gist.github.com/lukostyra/bc354a5fd845b82805ffb3380caebe9a
  2.  Get JavaFX Direct3D 12 build - either download the EA2 SDK from [ https://jdk.java.net/javafxdirect3d12/ ] or build it from scratch from direct3d12 [ https://github.com/openjdk/jfx-sandbox/tree/direct3d12 ] branch (make sure to build with -PCONF=Release; at the time of writing this email there is no functional difference between the sandbox repo and the EA2 build).
  3.  Put the script in the same directory as renderperf/RenderPerfTest.java - I usually copy the contents of tests/performance/animation/RenderPerfTest/src into a separate directory where the script is located. The script will look in the current directory specifically for "renderperf" directory containing a "RenderPerfTest.java" source file.
  4.  Ensure nothing else is running on your system.
  5.  Perform a baseline test run with D3D backend:
   ./run_renderperf_all.sh -j <path_to_jfx_sdk> -b d3d -r 3
When the script is done, this should result in "RenderPerf_results_d3d-<date>-<time>.csv" file. This will let us establish how your machine performs on RenderPerf’s test cases on the old D3D backend.
  6.  Perform a perf test run for D3D12 backend:
   ./run_renderperf_all.sh -j <path_to_jfx_sdk> -b d3d12 -r 3
This should result in "RenderPerf_results_d3d12-<date>-<time>.csv" file. The results will be used as a comparison to D3D backend.

Once you are done, reply to this email thread with details what hardware you ran the tests on and with contents of both CSV files specifying which came from which test run - I’m pretty sure attachments will be stripped when sending an email to the mailing list, so to make sure the results get to us you should paste them inline. Also, include the console output of running any JavaFX app on D3D12 backend with "-Dprism.verbose=true".

Note that these tests will run for quite a bit of time. Each test case's single run will take 15 seconds - 5 seconds of warm-up and 10 seconds of actual test run where FPS is measured. Testing every single case, 3 runs each, for a single backend takes approximately 40 minutes - ensure your computer won't lock or turn off the screens during that process.


Test script reference:
    ./run_renderperf_all.sh -j/--jfx <path_to_jfx_sdk> [-b/--backend <backend>] [-r/--runs <runs>]

Where:
    -j, --jfx  -  required; path to directory where JavaFX SDK is located (has to be where JavaFX bin and lib folders are located)
    -b, --backend  -  optional, defaults to "d3d12"; short-hand for which Prism backend to use. Whatever is set here will be forwarded to "-Dprism.order" property.
    -r, --runs  -  optional, defaults to 3; how many times each test case should run. Script will average FPS results from these runs.

Notes:
   * Closing currently running test during the warm-up stage (first 5 seconds since starting the execution of the test case) should stop the test script early.
  * Script starts by declaring an array of test cases and object counts to run. We recommend leaving those as they are, however if there are tests which do not load because of a "Timeout" message (the script will then exit early) or perform significantly worse on D3D backend (below 20 FPS) you can lower the object count and retry.


Let me know if you have any problems with running the tests. And, in advance, thanks for your help!

Regards,
Lukasz
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/openjfx-dev/attachments/20251113/71fdfd81/attachment-0001.htm>