For some computation-intensive kernels, we may manually design a deterministic dataflow for them and map them directly.
For example:
We are supposed to configure a matmul as a systolic array-like dataflow directly, avoiding bottlenecked by the NP-hard mapping.