tree 56373c3e348ab23a6700c3af800c0caeaf8bbbbf
parent 807f4573562472b1a58ce63a9d91b006cd1fb6cb
author MaheshRavishankar <1663364+MaheshRavishankar@users.noreply.github.com> 1620865558 -0700
committer GitHub <noreply@github.com> 1620865558 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJgnHIWCRBK7hj4Ov3rIwAADPsIAGDpRfl6KcRnTeowaZw7o5Co
 Bhht788wbj/pCjiDj9Q+WvS/902s8U1Xq8AETAhLXgtE1Z63Hdwdxal9bPoevFTd
 jBsEmrlq17GrJiVWFWStGKm/aZe9WDvPZp6QeqZjW9o/mJFTKA5Knr/byi/06u8K
 3W2lzhSQLbXcHgG3R/hyqmS2JQ9HswevFJPCLZI1fWrl3e0MBBaEavsuR76AcnxX
 xdjbGthW/2MsxJpKLcV//Y7MFlUOQ7ii8RBQYIWMXTVc9cC0LzqraCycqibTZv/m
 fb2biNobKtwSXE+OGS8f0EUQKAfTobyAANsDzHYu+LcbAt/xrG77WtubqybYK4Y=
 =+/4W
 -----END PGP SIGNATURE-----
 

Inline IndexCastOp and int/float ops into the dispatch region. (#5865)

Allowing folding index_cast operation and other integer/float
operations into the dispatch region avoids round-tripping to host to
get values from a tensor on device required for some index
computation.

Note: This only works for MHLO lowering path. A more nuanced solution
might be needed in general with explicit white-listing of operations,
but that requires more use cases.

Fixes #5620