tree 03eed9da834bf0b9636a0677d9a8b761b5574a03
parent faff7e0cc5ff9e25303bb7e9f6ac87ee88c5bef4
author Lei Zhang <antiagainst@google.com> 1691778440 -0400
committer GitHub <noreply@github.com> 1691778440 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJk1n2ICRBK7hj4Ov3rIwAADxUIAHUbB1WH/iCy7gLpPWGn24Zv
 vIeVgWzy451p6RX9XIaPk62J8ICS8Mt5x3g0Ukp6csVRZyVdZiKp4OnPN02xm7Xw
 S+gUDGxr91OguThxfGGwBmvPsmho7PCLTm+9uiOKRyRHWYn0mqPnmjzlZOXPQ+9w
 cR4186ZJUXnMzG/fVXkSIkzSJ843yU66wS40UK57T5oyNSNguTe4ssftMYkUyeLf
 nIhX8xMPOPmm5hac4Pwt+r8pogkbxuU+Uvd9PMcyVy21JgvKPkGVjNhWzO6rA4Vp
 +HJg+PqJ9OrN2jRA2aYn3g1rGr88+F7p1dUagcpo7rmte6ZqO9Qd00F5Vr26zIQ=
 =u+SY
 -----END PGP SIGNATURE-----
 

[gpu] Distribute fused producer elementwise ops in SIMT pipeline (#13908)

The fused producer ops have different dimensions than the consumer
matmul op. Tiling and distribution the matmul op's configuration may not
the the best.

This commit postpones the producer op tiling distribution to where copy
ops are handled. There we use all threads in the workgroup to distribute
in a flat manner.