4511c7b7fb
gazelle ended up being more liability than asset for our C++ rules. It required devs to manually run the command frequently (and was easy to forget until the CQ failed). The fact that we still had to edit the source files (e.g. the "srcs" cc_libraries) meant that the mixture between generated and hand-written caused some tension (see include/third_party/vulkan for a good example). The combination of gazelle and our IWYU enforcement added several bits of churn without any real benefit. The generated rules also didn't help identify cases where we were not keeping tight boundaries (e.g. non-gpu code and gpu code). Identifying third_party deps automatically ended up being trickier than anticipated (see the deleted //third_party/file_map_for_bazel.json) Using the "maximum set of dependencies" worked ok, but ended up increasing build time unnecessarily. For example, compiling CanvasKit for WebGL always needed to compile Dawn because SkSLCompiler.cpp sometimes needs to include tint/tint.h. Follow-up CLs will rebuild the BUILD.bazel rules without gazelle. Note to Reviewers: - The only file worth manually reviewing here is bazel/Makefile. Change-Id: I36d6fc3747487fabaf699690780c95f1f6765770 Bug: skia:12541 Reviewed-on: https://skia-review.googlesource.com/c/skia/+/543976 Reviewed-by: Leandro Lovisolo <lovisolo@google.com> Reviewed-by: Ben Wagner <bungeman@google.com> |
||
|---|---|---|
| .. | ||
| __init__.py | ||
| _adb_path.py | ||
| _adb.py | ||
| _benchresult.py | ||
| _hardware_android.py | ||
| _hardware_nexus_6p.py | ||
| _hardware_pixel2.py | ||
| _hardware_pixel_c.py | ||
| _hardware_pixel.py | ||
| _hardware.py | ||
| _os_path.py | ||
| README.md | ||
| sheet.py | ||
| skiaperf.py | ||
| skpbench.cpp | ||
| skpbench.py | ||
skpbench
skpbench is a benchmarking tool for replaying skp or mksp files on android devices. it achieves a lower variance in framerate by controlling the clock speed and stopping all other processes that could cause interference.
Build
skpbench consists of the skpbench binary which must be built for the phone you intend to run on, and skpbench.py which runs on the machine the phone is connected to via ADB and is the entry point.
The to build skia for android are at https://skia.org/user/build#android and reproduced here.
Download the Android NDK
./bin/sk asset download android_ndk_linux /tmp/ndk
After this is set up once, build skpbench for your target cpu (assumed to be arm64 here for a Pixel 3)
bin/gn gen out/arm64 --args='ndk="/tmp/ndk" target_cpu="arm64" is_debug=false'
ninja -C out/arm64 skpbench
Benchmark an SKP on a connected device.
First, copy the built skpbench binary and an example skp file to the device. (or pull a skp corpus using instructions in the section below)
adb push out/arm64/skpbench /data/local/tmp
adb push /home/nifong/Downloads/foo.skp /data/local/tmp/skps/
Run skpbench.py
python tools/skpbench/skpbench.py \
--adb \
--config gles \
/data/local/tmp/skpbench \
/data/local/tmp/skps/foo.skp
--adb specifies that it should use adb to the only connected device and run skpbench there.
--force is necessary because we don't yet have a configuration to monitor vitals on the Pixel 3.
--config gles specifies Open GL ES is the backend GPU config to use.
Additional documentation of arguments is printed by python tools/skpbench/skpbench.py --help
Output appears in the following format
accum median max min stddev samples sample_ms clock metric config bench
0.1834 0.1832 0.1897 0.1707 1.59% 101 50 cpu ms gles foo.skp
accum is the time taken to draw all frames, divided by the number of frames.
metric specifies that the unit is ms (milliseconds per frame)
MSKP corpus
A manually collected corpus of MSKPs from around 30 top apps (using skia via HWUI) and of about 20 actions in RenderEngine exists in a google cloud storage folder managed by skia/infra/bots/assets/mskp/upload.py
To download the fileset, first determine the highest current version of the fileset
gsutil ls gs://skia-assets/assets/mskp/
Download the latest version.
gsutil cp gs://skia-assets/assets/mskp/5.zip ~/Downloads
Unzip the archive and adb push it to the device.
To upload a new version of the corpus, use the steps above to download and unzip the last version, change the content however you need, then Use the upload tool, passing the directory of the altered archive (not a zip file). Note that you must provide it as an absolute path.
python upload.py --target_dir=/home/nifong/scratch/new_mskps
The upload script should print a version number. Finally, submit something like https://skia-review.googlesource.com/c/skia/+/304376 to point jobs at the new version.
Production
skpbench is run as a tryjob from gerrit, where it uploads the results to perf.skia.org.
Once such job name is Perf-Android-Clang-Pixel4XL-GPU-Adreno640-arm64-Release-All-Android_Skpbench
Perf results are available by querying with this or similar. extra_config = Android_Skpbench sub_result = accum_cpu_ms
Example perf query https://perf.skia.org/e/?queries=extra_config%3DAndroid_Skpbench%26sub_result%3Daccum_cpu_ms