You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The current fusion implementation of the Index Tree dialect does not work for MTTKRP. MTTKRP's Index Tree dialect after fusion does not change and still has 2 roots (itree nodes).
The correct fused index tree should only have 1 root and 3 compute nodes.
// MTTKRP:
Aij = Bikl * Clj * Dkj
// -> Current index tree has 2 roots and each one has 1 compute node
Tikj = Bikl * Clj
Aij = Tikj * Dkj
// -> Fused index tree should have 1 root with 3 compute nodes
Tj = 0
Tj = Bikl * Clj
Aij = Tj * Dkj
The text was updated successfully, but these errors were encountered:
We need to a cost model to enable fusion. The preliminary experiments demonstrate that fusion of dense MTTKRP is not efficient due to low data locality.
The current fusion implementation of the Index Tree dialect does not work for MTTKRP. MTTKRP's Index Tree dialect after fusion does not change and still has 2 roots (
itree
nodes).The correct fused index tree should only have 1 root and 3 compute nodes.
The text was updated successfully, but these errors were encountered: