README.md
1# skpbench
2
3skpbench is a benchmarking tool for replaying skp or mksp files on android devices.
4it achieves a lower variance in framerate by controlling the clock speed and stopping
5all other processes that could cause interference.
6
7## Build
8
9skpbench consists of the skpbench binary which must be built for the phone you intend to run on,
10and skpbench.py which runs on the machine the phone is connected to via ADB and is the entry point.
11
12The to build skia for android are at https://skia.org/user/build#android and reproduced here.
13
14Download the Android NDK
15
16```
17./bin/fetch-sk
18./bin/sk asset download android_ndk_linux /tmp/ndk
19```
20
21After this is set up once, build skpbench for your target cpu (assumed to be arm64 here for a Pixel 3)
22
23```
24bin/gn gen out/arm64 --args='ndk="/tmp/ndk" target_cpu="arm64" is_debug=false'
25ninja -C out/arm64 skpbench
26```
27
28## Benchmark an SKP on a connected device.
29
30First, copy the built skpbench binary and an example skp file to the device.
31(or pull a skp corpus using instructions in the section below)
32
33```
34adb push out/arm64/skpbench /data/local/tmp
35adb push /home/nifong/Downloads/foo.skp /data/local/tmp/skps/
36```
37
38Run skpbench.py
39
40```
41python tools/skpbench/skpbench.py \
42 --adb \
43 --config gles \
44 /data/local/tmp/skpbench \
45 /data/local/tmp/skps/foo.skp
46```
47
48`--adb` specifies that it should use adb to the only connected device and run skpbench there.
49`--force` is necessary because we don't yet have a configuration to monitor vitals on the Pixel 3.
50`--config gles` specifies Open GL ES is the backend GPU config to use.
51
52Additional documentation of arguments is printed by `python tools/skpbench/skpbench.py --help`
53
54Output appears in the following format
55```
56 accum median max min stddev samples sample_ms clock metric config bench
57 0.1834 0.1832 0.1897 0.1707 1.59% 101 50 cpu ms gles foo.skp
58```
59
60`accum` is the time taken to draw all frames, divided by the number of frames.
61`metric` specifies that the unit is ms (milliseconds per frame)
62
63## MSKP corpus
64
65A manually collected corpus of MSKPs from around 30 top apps (using skia via HWUI) and of about 20
66actions in RenderEngine exists in a google cloud storage folder managed by skia/infra/bots/assets/mskp/upload.py
67
68To download the fileset, first determine the highest current version of the fileset
69
70```
71gsutil ls gs://skia-assets/assets/mskp/
72```
73
74Download the latest version.
75
76```
77gsutil cp gs://skia-assets/assets/mskp/5.zip ~/Downloads
78```
79
80Unzip the archive and adb push it to the device.
81
82To upload a new version of the corpus, use the steps above to download and unzip the last version, change the
83content however you need, then Use the upload tool, passing the directory of the altered archive (not a zip file).
84Note that you must provide it as an absolute path.
85
86```
87python upload.py --target_dir=/home/nifong/scratch/new_mskps
88```
89
90The upload script should print a version number.
91Finally, submit something like https://skia-review.googlesource.com/c/skia/+/304376
92to point jobs at the new version.
93
94## Production
95
96skpbench is run as a tryjob from gerrit, where it uploads the results to perf.skia.org.
97Once such job name is `Perf-Android-Clang-Pixel4XL-GPU-Adreno640-arm64-Release-All-Android_Skpbench`
98
99Perf results are available by querying with this or similar.
100 extra_config = Android_Skpbench
101 sub_result = accum_cpu_ms
102
103Example perf query
104https://perf.skia.org/e/?queries=extra_config%3DAndroid_Skpbench%26sub_result%3Daccum_cpu_ms
105