I noticed that the relay tensor shape type will be converted to int32_t after shape inference, however, if IndexExpr arithmetics is involved, the shape type will be preserved, which breaks the following example
fn (%X: Tensor[(int64(7), int64(16)), float32]) {
%0 = reshape(%X, newshape=[7, -1, 1])
%0
}
After shape inference, it becomes
fn (%X: Tensor[(int64(7), int64(16)), float32])
-> Tensor[(7, int64(16), 1), float32] {
%0 = reshape(%X, newshape=[7, -1, 1]) # ty=Tensor[(7, int64(16), 1), float32]
%0
}
Notice the heterogenous tensor shape type for %0, which breaks the downstream compilation. I looked around the code at
but didn’t find suspicious things. And I would like to inject some debugging code to print out the content of IndexExpr in C++. Anyone knows how to do this? Thanks.