Adds the command:
make skps_release_and_SIMD
for perfing builds against a set of SKPs in ~/skps for release and
simd builds of CanvasKit. Also outputs a summary of the perf results
in a table format.
See the document "SIMD CanvasKit Build Performance Testing"
for more details:
https://docs.google.com/document/d/114kdSGPMnOSQCZ7pFgd3MGMn5mIW562RMoXVmD13e0M/edit#
Bug: skia:10453
Change-Id: I311629a1420301dda41f7ec57ce1403b05fd949b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/301982
Reviewed-by: Elliot Evans <elliotevans@google.com>
Reviewed-by: Kevin Lubick <kjlubick@google.com>
For each skp in the corpus, we start a fresh instance of
Chromium (via puppeteer), draw the skp and measure that time.
This process is repeated a fixed amount of repetitions
and the median, the average, and the std deviation is reported
to perf (as well as the individual datapoints as an FYI).
Importantly (and something we'll need to change about
SkottieFrames), we measure the average time between frames
after unlocking the framerate. This ensures we account for
the time needed by the GPU to actually draw (flush() returns
after the GPU has all the instructions, but not necessarily
has been able to draw).
This implementation is very similar to the SkottieFrames
code; a notable deviation is the repetitions are handled
outside of the html, i.e. a new chrome window per run.
I explored using content_shell, but noticed that requires
building Chromium, which our infrastructure is not set up
to do well.
Change-Id: I14fdbdc951604d3fdf06e81a4be7e614d0e53c03
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/295079
Commit-Queue: Kevin Lubick <kjlubick@google.com>
Reviewed-by: Nathaniel Nifong <nifong@google.com>
Reviewed-by: Eric Boren <borenet@google.com>
and rotations both have an effect on cache usage. Snapping translations to integer coordinations
reduces cache usage. Opacity of path painting does not have an effect on cache usage.
Bug: skia:10272
Change-Id: Id5d5f08cb43645c9ec44b9d8e5e96643041727c3
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/292280
Reviewed-by: Kevin Lubick <kjlubick@google.com>
IMPORTANT LESSON: when bringing in node (and possibly other
executables) via CIPD, add them to the path in gen_tasks_logic
so the parent executable (the task driver itself) has the right
PATH set. Otherwise, the subprocesses it spawns might grab the
wrong version because of how golang handles environments of
subprocesses.
This is starting as a fork of Skottie WASM. I hope to have a more unified
system for creating and running benchmarks.
Overall overview:
gen_tasks_logic.go creates a task in task.json that compiles
CanvasKit and the task drivers and then executes our task
(i.e. perf_puppeteer.go)
perf_puppeteer runs a node program (perf-with-puppeteer.js)
that uses puppeteer to execute benchmarking code on an
html page (canvaskit-skottie-frames-load.html).
I needed to update the node package so npm could be updated from
3.x to 6.14.4 so it knew about `npm ci`. This may not have been
entirely necessary, given the problems of executing the correct
npm (see important lesson above), but it hasn't broken things
further, so more up-to-date is probably a good thing.
Suggested Review Order:
- canvaskit-skottie-frames-load.html (note it is similar to
skottie-wasm-perf.html, but it waits for a button click
to start animating and records times from the main JS thread
itself)
- perf-with-puppeteer.js (similar to skottie-wasm-perf.js, but
has some things made optional [e.g. tracing])
- perf_puppeteer_test.go (shows the inputs/outputs of various steps)
- perf_puppeteer.go
- Everything else.
Change-Id: I380e81b825f36682c257664d488267edaf36369e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/285783
Commit-Queue: Kevin Lubick <kjlubick@google.com>
Reviewed-by: Eric Boren <borenet@google.com>
