commit b0512e2e6c00583f20b6f3b78c1ca52327ed745a
author Stanley Winata <68087699+raikonenfnu@users.noreply.github.com> Fri Jul 19 20:50:08 2024 -0700
committer GitHub <noreply@github.com> Fri Jul 19 20:50:08 2024 -0700
tree ef5d69113aec10a17f3b0a2f08b702226f965da4
parent 5b112cbc21046aaa028a05da62da1e3f8d19bc20

[GPU][VectDist] Refactor multiReducOp lowering to reduce acc at the end. (#17974)

The main motivation behind this commit is to fix the numerics issue we
find in Attention GPU C++ pipeline. Where our output seems to be some
scale off from reference (i.e out = k * ref). Through experiments we
determine that the cause of the issue is the multiReducOp distribution,
which is required for scaling/merge computation.

Through analyzing of IR and experimentation, reducing the srcVector and
non constant accumulator at the same time with another multiDimReduction
seems to output numerically wrong values.

This commit fixes numerical issue by refactoring the distribution of
multireductionOp to 3 steps. First every thread locally reduce the
srcVector it holds with combiningIdentity as localInit. Second, it does
a subgroup/warp reduce amongs other threads. Finally, each thread does a
local reduction of the intermediate reduced data it has with the
accumulator it holds.

---------

Signed-off-by: Stanley Winata <stanley.winata@amd.com>