Hi guys, I have question on if the value of box size could be dynamic.
What confuses me is the when I look at the ods of this op, it takes a instance of ‘Variadic’ type as its operand(which means it could be a dynamic value), while in its lowering implementation, this op is lowered to call cuda API which is wrapped by a thin wrapper func ‘mgpuTensorMapEncodeTiledMemref’(in llvm-project/mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp at main · llvm/llvm-project · GitHub).
In this wrapper function the global tensor strides is infered from the box size, but the type of these strides is uint64_t.
Based on my understanding, it indicates that if the box size value is of dynamic value, though the ods verify can accept but the lowering could fail anyway. Am I right?
A follow-up question is: if so, Does the expectation of this op only handle static value box size?