commit 9b97d80cb2c70fcad9e4a5a5b4f0c97822c28f00
author bjacob <benoitjacob@google.com> Tue Sep 20 13:25:34 2022 -0400
committer GitHub <noreply@github.com> Tue Sep 20 17:25:34 2022 +0000
tree 7e766dc866e80c973a223bf5f6849f1b7ea515cb
parent 8ad7a71b9e81276066a8636ed9947fb432c4829e

mmt4d i8mm kernel register fix (#10476)

This comes on top of #10475, the only actual diff here is in
runtime/src/iree/builtins/ukernel/arch/arm_64/mmt4d_tile_arm_64_i8mm.S.

I noticed that google benchmark was incorrectly running only 1000
iterations of this benchmark. Debugging, its floating point code
computing elapsed time got wrong arithmetic in determining whether it
was time to cut the benchmark short. Eventually I realized my mistake:
this kernel uses all 32 SIMD registers, including v8--v15 whose bottom
64bit are callee-saved. We were overwriting the google-benchmark code's
timestamp value stored in one of those!

The standard fix would be to push these 8 x 64bit to stack an pop at the
end, which could be done in a total of 4 store-pair + 4 load-pair = 8
instructions, but we save those instructions by instead just not using
v8--v15 at all; we didn't really need them.