As far as I know, TVM tries to fuse a matmul with a following element-wise operator, but what about the following two cases:
- Matmul followed by a reduction op, such as softmax
- Matmul followed by another matmul
Will TVM try to fuse that? Thanks!
As far as I know, TVM tries to fuse a matmul with a following element-wise operator, but what about the following two cases:
Will TVM try to fuse that? Thanks!
These cases are not fused. The fusion itself would be trivial, the challenge is scheduling (to generate efficient code).