tree 02579709c6f30e0065b37be96eab2519ee908f59
parent 7011e9f3f579cbd904badb5ee3ab747d996c0ba8
author Lei Zhang <antiagainst@google.com> 1670270054 -0800
committer GitHub <noreply@github.com> 1670270054 +0000
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJjjkxmCRBK7hj4Ov3rIwAAT54IAFEGrhfpfeRmNhC/cmXzjOym
 iZQmM6w9uE7mJuekBMw6uZ+e7f9bMy9Q3BCx5Hmb6MmrBoNBuHzsg1WiSWa+hkye
 nzCsluLigLpV9OQBKepI5BeguWA6IHhktTq3XcntscONfztWBccKoqFNA5Od7TGi
 OYigVDKBQSgJxH4EBjTT+3SaGXlcilUE2/y72VJh4vhYmC/aAvksvZDKwjWSJt/y
 2/FbVe5ZgB7Sn+11Dyh1RQu4y+VpoOO7SUCMWEU37H5x3DbYLy2caQBegVvVkFdC
 q0rf9u1+44nCDo2AznfAjmDSaJJKMqEdh6tx0QuQpa1PXIT+etYEEOJ2P5uTncc=
 =364N
 -----END PGP SIGNATURE-----
 

[spirv] Fix C matrix promotion with bufferization allocations (#11418)

For cases like linalg.matmul + arith.extf, bufferization inserts an
allocation because of the different element type of those two ops. Here
we already have the C matrix in workgroup memory; there is no need to
promote it again. Though we need to make sure to use the proper marker
so that we can insert barriers and enable following steps to distribute
to SIMT threads.

Co-authored-by: Thomas <thomasraoux@google.com>