Clone this repo:
  1. 0f02b4c compositortest: Decouple platform from API by Matt Turner · 3 weeks ago main release-R125-15853.B-main stabilize-15855.B-main stabilize-15857.B-main
  2. 605b158 Factor out platform checks into platform.h header by Matt Turner · 3 weeks ago
  3. dab8f6c src: Move epoxy and EGL to PLATFORM_LIBS by Matt Turner · 3 weeks ago
  4. fa97b48 .gitignore: Add *.o and compositortest by Matt Turner · 3 weeks ago firmware-geralt-15842.B-main
  5. a365ce6 Normalize Copyright headers by Matt Turner · 3 weeks ago

GLBench runs OpenGL or OpenGL ES microbenchmarks and writes performance numbers to stdout and resulting images to a directory for verification.

For the test to pass the performance numbers have to be better than a predefined threshold, while the resulting images have to be found in a repository of reference images. As the image name encodes the raw pixel MD5 this can be done as a simple file existence check. If we ever get too much pixel variation using a tool like perceptualdiff to waive small differences should be acceptable.

Build for Linux

It might be easier to develop new tests under Linux:

sudo apt-get install libgflags-dev
sudo apt-get install libwaffle-dev
sudo apt-get install waffle-utils (optional)
make

Executable Options

./glbench -notemp [-save [-outdir=<directory>]]

Example

./glbench -save -outdir=img
# board_id: NVIDIA Corporation - Quadro FX 380/PCI/SSE2
swap_swap                    =   214.77 us           [swap_swap.pixmd5-20dbc406b95e214a799a6a7f9c700d2f.png]
clear_color                  =  4448.28 mpixels_sec  [clear_color.pixmd5-e3609de1022a164fe240a562c69367de.png]
clear_depth                  = 10199.76 mpixels_sec  [clear_depth.pixmd5-e3609de1022a164fe240a562c69367de.png]
clear_colordepth             =  3250.57 mpixels_sec  [clear_colordepth.pixmd5-e3609de1022a164fe240a562c69367de.png]
clear_depthstencil           = 26447.22 mpixels_sec  [clear_depthstencil.pixmd5-e3609de1022a164fe240a562c69367de.png]
[...]

ls img
clear_color.pixmd5-e3609de122a164fe240a562c69367de.png
clear_colordepth.pixmd5-e3609de122a164fe240a562c69367de.png
clear_colordepthstencil.pixmd5-e3609de122a164fe240a562c69367de.png
compositing.pixmd5-7d02a16a7ac15cd6cbbc5c786f1.png
[...]

Running from the autotest harness

Running the autotest test_that $DUT graphics_GLBench will \

  1. run glbench -save \
  2. first try to identify known buggy images by searching in deps/glbench/glbench_knownbad_images.txt \
  3. then identify good images by searching in deps/glbench/glbench_reference_images.txt \
  4. TODO(ihf): use perceptualdiff to do a fuzzy compare \
  5. raise an error if the image is completely unknown \
  6. report performance numbers back to the harness \

Handling of reference images

Good reference images themselves are located at ./ref_images/glbench_reference_images/
Images that have outstanding defects and an open bug filed are at ./ref_images/glbench_knownbad_images/chromium-bug-NNNNN/
When that bug is closed the directory should be moved to ./ref_images/glbench_fixedbad_images/chromium-bug-NNNNN/ \

To push out new reference images place them in the appropriate directories (create a new bug if needed) and run ./update_glbench_image_filelists.sh to update the image filelists.

compositortest

CompositorTest is another binary generated in the glbench directory. It runs a small OpenGL or OpenGL ES shader program and optionally sleeps for some time before swapping buffers. The time between swaps is logged to a file and processed for reporting in a tast test. We can see from this simiulated work load how much time can be spent processing a frame before it gets dropped.

example

./compositortest --fullscreen true --gpu_workload_ms 15 Platform OpenGL version: OpenGL ES 3.2 Mesa 22.1.4 (git-d3e5726da1) Update num quads: 144863 calibrating: gpu_elapsed_time: 14.9717 ms Finished calibrating workload. [..] frame 998: cpu_elapsed_time: 16.1118 ms frame 998: gpu_elapsed_time: 14.9748 ms frame 999: cpu_elapsed_time: 16.1824 ms frame 999: gpu_elapsed_time: 14.9726 ms Avg: 62 fps for 16 secs. 1000 frames rendered.