Skip to content

tc: add an option for scaling the batch size

Mike Blumenkrantz requested to merge zmike/mesa:tc-scaling into main

on some platforms, the overhead of triggering new thread jobs is so severe that it's more efficient to record an entire frame with tc and then process it in a single batch than it is to split the frame into multiple batches and have the driver thread incrementally process the workload

to alleviate this, allow drivers to enable scalable batch sizes. when enabled, tc batches will automatically scale to a larger size after they overflow, reducing the total number of batch-jobs that need to be run

improves perf on freedreno by ~5-10% in some cases and on zink+turnip by 40-50% in the same cases

Merge request reports