tree 800bf09b7705bab28165cff3486aadf02d325ef1
parent eee6faf6a1f3107315167618159ff6a85484c01f
author George Steed <george.steed@arm.com> 1701329983 +0000
committer James Zern <jzern@google.com> 1701463103 +0000

av1_fwd_txfm2d_neon.c: Use switch for small square problem sizes

For the smallest problem sizes we have a significant overhead from
needing to load and store the transform intermediate vectors. Avoiding
the kernel lookup and calling the kernels directly significantly
improves performance in these cases.

Benchmarking on a Neoverse N2 machine with Clang 16 and GCC 12, the
speed tests report a geomean ~22% reduction in times reported for 4x4
and a ~4.6% reduction for 8x8.

Change-Id: I5907fab09e40b1cea57c446fd7e604ae911ceae8