Page MenuHomePhabricator

[mlir][linalg] Prepare fusion on tensors for scalar operands.
ClosedPublic

Authored by gysit on Jun 8 2021, 5:50 AM.

Details

Summary

Adapt fusion on tensors to support structured ops taking scalar operands.

Diff Detail

Event Timeline

gysit created this revision.Jun 8 2021, 5:50 AM
gysit requested review of this revision.Jun 8 2021, 5:50 AM
mravishankar accepted this revision.Jun 8 2021, 10:32 AM

So not related to this particular change, but does the fusion of operations where one of the operands was previously say a tensor<f32> and is now just f32 type work?

This revision is now accepted and ready to land.Jun 8 2021, 10:32 AM
gysit added a comment.Jun 8 2021, 11:16 PM

Using the new interface it should since there is an indexing map etc. The producer of the scalar will never be fused though since it cannot be another linalg operation.