The original asahi code assumed a subgroup size of 32 and a workgroup
size of 32 * 32 = 1024. This makes doing ctz(ballot(b)) across an
entire workgroup an almost trivial operation. On panfrost, we won't be
so blessed unless we choose a workgroup size of 16 * 16 = 256. It's
also not clear that we want to use workgroups at all and we may better
off sticking to just subgroup parallelism and cutting out memory
bandwidth by more than half. With the new code, the only requirement
should be that the subgroup size is a power of two (this is always true)
and that the workgroup size is an even multiple of the subgroup size.
Even though the new code looks way more complicated, thanks to the magic
of NIR constant folding, it should all fold down to the original code on
asahi and something even smaller if one opts to go for a single subgroup.
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Reviewed-by: Mary Guillemard <mary@mary.zone>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38404>
ed0998ca98