AArch64: f16 matmul performance degradation #4476
Description
The performance issue has been reproduced on Apple M2 Pro.
Benchdnn reproducer:
--matmul --reset --allow-enum-tags-only=0 --engine=cpu --runtime_dims_masks= --dt=f16:f16:f16 --stag=abcd --wtag=abcd --dtag=abcd --strides=:: --attr-scales= --attr-zero-points= --attr-precomputed-reductions= --attr-scratchpad=user 1x8x1500x1500:1x8x1500x64
oneDNN 3.8 with ACL 52.6.0 gives 3.3 ms. on Apple M2 Pro.
oneDNN 3.10 with ACL 52.6.0 gives 8.1 ms. on Apple M2 Pro.