tree b2a7affdddc7cb2e296f8ec9fb377c3f1b8f02ec
parent 575509fcc692d0544d6709033c64419621b343b4
author Han-Chung Wang <hanchung@google.com> 1636834983 -0800
committer GitHub <noreply@github.com> 1636834983 -0800
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJhkB6oCRBK7hj4Ov3rIwAADc4IADedltmgQxbJt95LvOaGALaJ
 pTCazacW8dYZYgqbl5FoMESZ+qyXlOzzxzK32ogHsVFM4sz+DJ9c6q3m2lhaT3yh
 DGVTVsrpr00iT3TMk1aj5lwZ1If6W7Hjz7DwQ7D4np09WglfvkCQnWR1hsnZ7jEz
 JvP587bdJK/vD9V9WeeOyFGECHWVPFqhWyvpU8mVOi1AbZvc1obHAw4/T9VSc8id
 bH07ZSs4axRZZPMxGwsLyo3rBQ0UKbBFo5TfLTAG3l78IbUQ2DpyXi++WBWrXSsy
 rkCDJ5vV2DnfqvkqJb0WdlgYEWupa2StNi8eACAVgR0mK+XLVsfxKHkZJ3zgSo8=
 =1u0s
 -----END PGP SIGNATURE-----
 

Enable more vectorizations in LLVMCPUTileFuseAndVectorizePass (#7652)

Vectorization will turn Linalg ops to vector ops and arith ops. Since we
don't propagate information through arith ops during bufferization. We
can't unconditionally vectorize all the ops. To prevent creating extra
memref.alloc ops, we can't tile along reduction dims.

The next step is to get unroll vector pass in, so we can vectorize more
ops.

This PR improves the performance of transformer-benchmark from 58 ms to
33 ms.