| commit | 22434802ae922b9f1fab5710323bb934c409bc37 | [log] [tgz] |
|---|---|---|
| author | Femi Adegunloye <[email protected]> | Tue Mar 28 09:23:08 2023 |
| committer | Chromeos LUCI <[email protected]> | Mon Apr 24 15:54:10 2023 |
| tree | 8bedf89ab90747fbc5864b2e9c73a9f212606fa4 | |
| parent | ae7450d4094d22af9ebe32484707830886c526a5 [diff] |
swapbench: compositortest: add width, height, other improvements * Add width and height options to compositortest binary. * Use --display-scale to scale the app resolution. * Fix bug where mod_size wasn't being updated on DecreaseWorkload (fixes visual bug where lower portion of screen would turn gray). * Log num quads after every decrease workload. * Add --sleep-cpu-milliseconds for sleeping for a portion of the frame. * Scale color of vertical bar that runs across window by num quads BUG=b:275508795 TEST=build and deploy to DUT Change-Id: If2621401887bbf65f179f706022374aa632d28d8 Reviewed-on: https://chromium-review.googlesource.com/c/chromiumos/platform/glbench/+/4376236 Reviewed-by: Po-Hsien Wang <[email protected]> Tested-by: Femi Adegunloye <[email protected]> Commit-Queue: Femi Adegunloye <[email protected]>
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.
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
./glbench -notemp [-save [-outdir=<directory>]]
./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 the autotest test_that $DUT graphics_GLBench will \
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 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.
./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.