tree e10ed61081f11208f8db6bca52d2938635f724fd
parent 33ca9005c9d11c861a7419d8159d37469a039863
author Stanley Winata <68087699+raikonenfnu@users.noreply.github.com> 1667945918 -0800
committer GitHub <noreply@github.com> 1667945918 -0800
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJjatW+CRBK7hj4Ov3rIwAAVgoIACoUa2ZQFU7Ytri6e0gSkLIj
 WXbLCobL/BsgwyxLiRMBr22Afq0mo7FmhOw3O3EQfIrpOOPFCIcXfXREqSlz5Em8
 IO34qXwO6psmkiKrBDM8Qwdafo1k/lSZnwXkhDJvphoiNOVu1+o37MK3k6AVtrI7
 FNq4Gxz7uJCvrdxU6rIv6SFWDgODpkYmQ7KxOvywn5og9Th8leZpdB0QYu/viqKn
 QeblRhj89kpViCmtza2hZOm/E7sfldERJsm4SqizNdCZs7H3rAtxMa5uP0JQipzS
 SE8D0LYpEV59YaU+3HsoGxgqFTN9MpJLq/R/y+zsGCORXDcwKWUeugegBriSHxw=
 =9UTz
 -----END PGP SIGNATURE-----
 

[GPU] Apply warp reduce to intermediate warp reduce results. (#11082)

Instead of doing vector.reduction after obtaining warp reduction
results, we apply another warp reduce to obtain the final reduction
results.

This shouldn't give much perf boost on it's own, but will give
flexibility to enable warp reduction on fp16 and int8 down
the line.