Obtain cuda kernel time from IR

[TUMSchieben]

Hi there,

is there any way to obtain the execution time of cuda kernels from IR side?
Eg. use a handwritten mlir containing 'mhlo.dot' op and some timing primitives?
I saw similar solutions in iree-llvm-sandbox, which is for CPU, code fragment is like:

func.func private @nanoTime() -> i64 attributes {llvm.emit_c_interface}
func.func public @main(...) {
    %c0 = arith.constant 0 : index
    %0 = func.call @nanoTime() : () -> i64
    %1 = func.call @matmul(...)
    %2 = func.call @nanoTime() : () -> i64
    %3 = arith.subi %2, %0 : i64
    memref.store %3, %arg2[%c0] : memref<?xi64>
    return %1 : ...
  }

thx