[Unity] Fix FuseOpsByPattern when a subgraph can be matched by multiple residual patterns

[masahi]

Consider this subgraph:

lv_ = R.matmul(x2, y2)
lv = R.matmul(x, y)
lv1 = R.add(lv, bias)
lv2 = R.add(lv1, lv_)
out = R.add(lv2, residual)

This subgraph matches to two different residual patterns matmul -> add -> add starting from each matmul and matching two out of three adds. Currently we end up generating the following partition, since the second add matches to both residual patterns and the two groups are merged incorrectly after the second matching.

@R.function
def gv(x2_1: R.Tensor((32, 8), dtype="float16"), y2_1: R.Tensor((8, 8), dtype="float16"), x_1: R.Tensor((32, 8), dtype="float16"), y_1: R.Tensor((8, 8), dtype="float16"), bias_1: R.Tensor((8,), dtype="float16"), residual_1: R.Tensor((32, 8), dtype="float16")) -> R.Tensor((32, 8), dtype="float16"):
    R.func_attr({"Composite": "cutlass.matmul_bias_residual_add", "Primitive": 1})
    with R.dataflow():
        lv_: R.Tensor((32, 8), dtype="float16") = R.matmul(x2_1, y2_1, out_dtype="float16")
        lv: R.Tensor((32, 8), dtype="float16") = R.matmul(x_1, y_1, out_dtype="float16")
        lv1: R.Tensor((32, 8), dtype="float16") = R.add(lv, bias_1)
        lv2: R.Tensor((32, 8), dtype="float16") = R.add(lv1, lv_)
        gv_1: R.Tensor((32, 8), dtype="float16") = R.add(lv2, residual_1)
        R.output(gv_1)
    return gv_1

gv1: R.Tensor((32, 8), dtype="float16") = gv(x2, y2, x, y, bias, residual)

This bug has been fixed by checking, on successful pattern matching, whether there exists a matched expression that is part of a previously matched group.

@yelite @vinx13

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @quic-sanirudh _{See #10317 for details}

_{Generated by tvm-bot}