This combines the BruteForceManager and Gridmanager, starting with
brute force when it's fastest for low N and then transitioning to less
accurate grid that scales for larger N.
Also updates the set of benchmarks to run based on more reasonable
configs (e.g. having a reasonable level of accuracy to be worth
considering).
Updates Device to use the HybridBoundsManager with brute force up to
64 draws, and then a grid configured to make 16x16 pixel cells. My
guess is we will see a mix of perf regressions and improvements with
this. The existing use of the NaiveBoundsManager had negligible CPU
cost but disallowed all re-ordering. The brute force and grid
managers will add CPU cost but enable re-ordering, which shows up
as shorter command buffers (e.g. 17k commands vs. 28k commands in the
motionmark suits benchmark). However, because we don't have SSBOs
there still isn't as much batching that would let the GPU take
advantage of this re-ordering so I'm not sure how visible the wins
will be yet.
Bug: skia:13201, skia:12787
Change-Id: Iad58fccab45def5f702a30860e063669424dfcf2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/550518
Reviewed-by: Jim Van Verth <jvanverth@google.com>
Reviewed-by: Nicolette Prevost <nicolettep@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
e9eda5802a