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
Currently we only data-tile matmuls. When we start data-tiling convs, we will need to track in the encoding the type of op using the encoded tensors. This doesn't map 1:1 to linalg named ops. This could be a new enumeration attribute, EncodingOpType, that could for now have a single value MATMUL (used for all of the flavors of matmuls that we currently handle, including batch matmuls, vecmat, matvec etc). Then when we start data-tiling convs, there could be a new value CONV.
Note: in the past we used to have such an enum, EncodingUser. I don't think that was a great name.
The text was updated successfully, but these errors were encountered:
Currently we only data-tile matmuls. When we start data-tiling convs, we will need to track in the encoding the type of op using the encoded tensors. This doesn't map 1:1 to linalg named ops. This could be a new enumeration attribute,
EncodingOpType
, that could for now have a single valueMATMUL
(used for all of the flavors of matmuls that we currently handle, including batch matmuls, vecmat, matvec etc). Then when we start data-tiling convs, there could be a new valueCONV
.Note: in the past we used to have such an enum,
EncodingUser
. I don't think that was a great name.The text was updated successfully, but these errors were encountered: