Previously we had to utilize the 3D path for float16 formats since
the hw would implicitly convert f16->f32 canonicalizing NaNs
resulting in copies that were not bit exact.
The `HALF_PRECISION` bit was discovered which avoids this
conversion, so we can go back to using the 2D path.
Using the 2D path is faster than the 3D path. Results of Crucible
bench.cast-image show an improvement of slightly above a 50% average
for the 1 MiB image->image copy of {4,6,16,32,64)B chunks, and for
the 16 GiB there's a slight improvement.
The affected formats are {R16,R16G16,R16G16A16}_SFLOAT.
Signed-off-by: Karmjit Mahil <karmjit.mahil@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35709>
Karmjit Mahil
65861e81af