Back to PBT run

input.mlir

dataflow.thread private @t_0 domain(#dataflow.thread_domain<dense>)(%arg_0: f32, %arg_1: i64, %arg_2: index) ctrl (%thread_ctrl: none) iv (%coord_0: index) {
  dataflow.thread.yield
}

dataflow.thread private @t_1 domain(#dataflow.thread_domain<dense>)() ctrl (%thread_ctrl: none) {
  dataflow.thread.yield %thread_ctrl : none
}

func.func @host(%m_ref: memref<?xf32>) {
  %c_i32 = arith.constant 1 : i32
  %c_f32 = arith.constant 1.000000e+00 : f32
  %c_i64 = arith.constant 1 : i64
  %c_idx = arith.constant 4 : index
  %token_0 = dataflow.thread.launch @t_0(%c_f32, %c_i64, %c_idx) grid(%c_idx) : (f32, i64, index) -> !dataflow.thread_token
  dataflow.thread.wait %token_0 : !dataflow.thread_token
  %token_1 = dataflow.thread.launch @t_1() : () -> !dataflow.thread_token
  return
}