tree debd6d52b70ac108fc23797704e9bc6b300429f5
parent f1b71b19b7d9722dfa85f71e743fb9dc07f70c78
author Lei Zhang <antiagainst@google.com> 1681329342 -0400
committer GitHub <noreply@github.com> 1681329342 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJkNwy+CRBK7hj4Ov3rIwAAVnMIABbXsZqciSnjSnZN2ChbZQM0
 0SWri4hEXS+RkI7eCgVUuHzF8nBMa/11PAhyOxDsBPQH2SOay/NOO4goezpcvXD1
 Z6aeEMzbjBtCqLSGLi/atkhv+78a/NJHts1saSt/O2bNMjqOVrCiDFJh/CewCEZP
 ocQcOI7SJvw46xkeUR+ThABLnqgbNgtrZ4jzMPCKIXi7VDgNRIKP5eM4xrSw1L4F
 WI4dkd/Z8P8C5fjHHzluBTGdNBUlSojp4xbIic2r+QI7e0b8F8Ej0wIQY+MEbqzj
 /mwTUgcY2UYiksxgsZda6yM73i0+ctC3O+K6AcBt6wOZiu9P1XQAhTVsK3nFFeE=
 =ZKCZ
 -----END PGP SIGNATURE-----
 

[spirv] Use convolution output type bitwidth for tiling (#13049)

In the case of int8 convolutions, we can have int8 input/filter but
int32 accumulator. So we need to use the output type bitwidth for
deducing the tiling scheme.

This should address the perf regression for int8 EfficientNet.