Skip to content

nir: intel-fs: Optimize subgroup intrinsics in uniform control flow

Ian Romanick requested to merge idr/mesa:review/uniform-reduction into main

This MR primarily adds the optimizations described in #3731. As part of this, there are also a couple bug fixes for Intel driver.

I have not enabled the new optimization pass in ACO, but that should be pretty easy. There is some (untested) code to emit the mbcnt instruction, but a flag needs to be added to nir_lower_subgroups_options. Since Intel doesn't have write_invocation, I didn't implement the min/max/and/or exclusive_scan optimizations.

Merge request reports

Loading