wzh99 commented on issue #11704:
URL: https://github.com/apache/tvm/issues/11704#issuecomment-1157628320
I have found a more surprising case of operator fusion. I exchange the two
inputs of both `multiply` and `add` in the original test case:
```
def @main(%x2: Tensor[(1, 2), float32] /* ty=Tensor[(1, 2), float32] */, %w:
Tensor[(2, 2), float32] /* ty=Tensor[(2, 2), float32] */, %x0: Tensor[(1, 2),
float32] /* ty=Tensor[(1, 2), float32] */, %x1: Tensor[(1, 2), float32] /*
ty=Tensor[(1, 2), float32] */) -> Tensor[(1, 1, 2), float32] {
%0 = nn.dense(%x2, %w, units=None) /* ty=Tensor[(1, 2), float32] */;
%1 = maximum(%x0, %x1) /* ty=Tensor[(1, 2), float32] */;
%2 = multiply(%0, %1) /* ty=Tensor[(1, 2), float32] */;
%3 = expand_dims(%1, axis=1) /* ty=Tensor[(1, 1, 2), float32] */;
add(%2, %3) /* ty=Tensor[(1, 1, 2), float32] */
}
```
This program can be successfully compiled! No error is reported. I also
check the fused version of this program:
```
def @main(%x2: Tensor[(1, 2), float32] /* ty=Tensor[(1, 2), float32] */, %w:
Tensor[(2, 2), float32] /* ty=Tensor[(2, 2), float32] */, %x0: Tensor[(1, 2),
float32] /* ty=Tensor[(1, 2), float32] */, %x1: Tensor[(1, 2), float32] /*
ty=Tensor[(1, 2), float32] */) -> Tensor[(1, 1, 2), float32] {
%1 = fn (%p01: Tensor[(1, 2), float32] /* ty=Tensor[(1, 2), float32] */,
%p11: Tensor[(1, 2), float32] /* ty=Tensor[(1, 2), float32] */, Primitive=1) ->
Tensor[(1, 2), float32] {
maximum(%p01, %p11) /* ty=Tensor[(1, 2), float32] */
} /* ty=fn (Tensor[(1, 2), float32], Tensor[(1, 2), float32]) ->
Tensor[(1, 2), float32] */;
%3 = %1(%x0, %x1) /* ty=Tensor[(1, 2), float32] */;
%4 = fn (%p02: Tensor[(1, 2), float32] /* ty=Tensor[(1, 2), float32] */,
%p12: Tensor[(2, 2), float32] /* ty=Tensor[(2, 2), float32] */, %p2: Tensor[(1,
2), float32] /* ty=Tensor[(1, 2), float32] */, Primitive=1) -> Tensor[(1, 2),
float32] {
%2 = nn.dense(%p02, %p12, units=None) /* ty=Tensor[(1, 2), float32] */;
multiply(%2, %p2) /* ty=Tensor[(1, 2), float32] */
} /* ty=fn (Tensor[(1, 2), float32], Tensor[(2, 2), float32], Tensor[(1,
2), float32]) -> Tensor[(1, 2), float32] */;
%5 = %4(%x2, %w, %3) /* ty=Tensor[(1, 2), float32] */;
%6 = fn (%p0: Tensor[(1, 2), float32] /* ty=Tensor[(1, 2), float32] */,
%p1: Tensor[(1, 2), float32] /* ty=Tensor[(1, 2), float32] */, Primitive=1) ->
Tensor[(1, 1, 2), float32] {
%0 = expand_dims(%p0, axis=1) /* ty=Tensor[(1, 1, 2), float32] */;
add(%p1, %0) /* ty=Tensor[(1, 1, 2), float32] */
} /* ty=fn (Tensor[(1, 2), float32], Tensor[(1, 2), float32]) ->
Tensor[(1, 1, 2), float32] */;
%6(%3, %5) /* ty=Tensor[(1, 1, 2), float32] */
}
```
This time the computation graph is not fused into one single group, but
three instead. Note that I only exchange the predecessors of two vertices, but
the results of operator fusion are so different.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
