[graphite] Make nanobench measure the right amount of work in SKPs
SKPBench tiles the SkPicture into multiple surfaces, with tile WH
differing between GPU and CPU backends. Graphite was being incorrectly
classified as a CPU backend and ended up using smaller tile sizes, so
the SkPicture would be played back many more times relative to Ganesh.
In addition, each surface's contents is a subset of the total picture,
so batching was artificially limited compared to Ganesh.
Added a call to Device::flushPendingWorkToRecorder() in
Surface_Graphite::onFlush(). This ensures DrawPass::Make() is called
when nanobench and viewer are measuring the bulk of the work
(viewer's "flush" time was always 0 for Graphite since the
DrawPass::Make was only being counted in the total time when it
was executed for swapBuffers()). Flushing in this manner also prevents
batching across loops in nanobench, or resetting/clearing prior loops
recorded draws when the benchmark starts with a fullscreen clear.
The SKPBench change should make all graphite benchmarks report lower
times compared to what's in perf.skia.org. The flush change should
increase their reported times for benchmarks that required multiple
loops to get an accurate time measurement (for expensive SKPs with
loops == 1, it shouldn't be affected).
Change-Id: I9256dbfc4c7c021377be8f5137b48036cc67e4a2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/548157
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Greg Daniel <egdaniel@google.com>
3 files changed