Support Customized DFG Sorting Strategy

[ShangkunLi]

Due to the time limit, I just enabled the customized sorting strategy for steering-based dataflow IR mapping.

Maybe we can properly design the mixed sorting later. For now, I just use topologically sorted ops for evaluation.

[tancheng]

Can you please remind me what's wrong with "mixed"? You plan to fix "mixed" if no "due to time limit"?

[ShangkunLi]

For this case:

module {
  func.func @simple_add_loop() -> i64 attributes {accelerator = "neura", dataflow_mode = "steering"} {
    %0 = neura.reserve : i64
    %1 = neura.reserve : i64
    %2 = neura.reserve : i1
    %3 = "neura.constant"() <{value = 16 : i64}> : () -> i64
    %4 = "neura.constant"() <{value = 1 : i64}> : () -> i64
    %5 = "neura.constant"() <{value = 1 : i64}> : () -> i64
    %6 = "neura.constant"() <{value = 0 : i64}> : () -> i64
    %7 = neura.invariant %4, %2 : i64, i1 -> i64
    %8 = neura.invariant %3, %2 : i64, i1 -> i64
    %9 = neura.carry %5, %2, %0 : i64, i1, i64 -> i64
    %10 = neura.carry %6, %2, %1 : i64, i1, i64 -> i64
    %11 = "neura.icmp"(%10, %8) <{cmpType = "slt"}> : (i64, i64) -> i1
    neura.ctrl_mov %11 -> %2 : i1 i1
    %12 = neura.false_steer %9, %11 : i64, i1 -> i64
    %13 = "neura.add"(%9, %9) : (i64, i64) -> i64
    neura.ctrl_mov %13 -> %0 : i64 i64
    %14 = "neura.add"(%10, %7) : (i64, i64) -> i64
    neura.ctrl_mov %14 -> %1 : i64 i64
    "neura.return"(%12) : (i64) -> ()
  }
}

Its topological order is:

[MapToAcceleratorPass] Topologically sorted op: %0 = neura.reserve : i64
[MapToAcceleratorPass] Topologically sorted op: %1 = neura.reserve : i64
[MapToAcceleratorPass] Topologically sorted op: %2 = neura.reserve : i1
[MapToAcceleratorPass] Topologically sorted op: %3 = "neura.constant"() <{value = 16 : i64}> : () -> i64
[MapToAcceleratorPass] Topologically sorted op: %4 = "neura.constant"() <{value = 1 : i64}> : () -> i64
[MapToAcceleratorPass] Topologically sorted op: %5 = "neura.constant"() <{value = 1 : i64}> : () -> i64
[MapToAcceleratorPass] Topologically sorted op: %6 = "neura.constant"() <{value = 0 : i64}> : () -> i64
[MapToAcceleratorPass] Topologically sorted op: %9 = "neura.data_mov"(%3) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %7 = "neura.data_mov"(%4) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %11 = "neura.data_mov"(%5) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %13 = "neura.data_mov"(%6) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %10 = neura.invariant %9, %2 : i64, i1 -> i64
[MapToAcceleratorPass] Topologically sorted op: %8 = neura.invariant %7, %2 : i64, i1 -> i64
[MapToAcceleratorPass] Topologically sorted op: %12 = neura.carry %11, %2, %0 : i64, i1, i64 -> i64
[MapToAcceleratorPass] Topologically sorted op: %14 = neura.carry %13, %2, %1 : i64, i1, i64 -> i64
[MapToAcceleratorPass] Topologically sorted op: %16 = "neura.data_mov"(%10) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %25 = "neura.data_mov"(%8) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %21 = "neura.data_mov"(%12) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %22 = "neura.data_mov"(%12) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %18 = "neura.data_mov"(%12) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %24 = "neura.data_mov"(%14) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %15 = "neura.data_mov"(%14) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %23 = "neura.add"(%21, %22) : (i64, i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %26 = "neura.add"(%24, %25) : (i64, i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: %17 = "neura.icmp"(%15, %16) <{cmpType = "slt"}> : (i64, i64) -> i1
[MapToAcceleratorPass] Topologically sorted op: neura.ctrl_mov %23 -> %0 : i64 i64
[MapToAcceleratorPass] Topologically sorted op: neura.ctrl_mov %26 -> %1 : i64 i64
[MapToAcceleratorPass] Topologically sorted op: neura.ctrl_mov %17 -> %2 : i1 i1
[MapToAcceleratorPass] Topologically sorted op: %19 = "neura.data_mov"(%17) : (i1) -> i1
[MapToAcceleratorPass] Topologically sorted op: %20 = neura.false_steer %18, %19 : i64, i1 -> i64
[MapToAcceleratorPass] Topologically sorted op: %27 = "neura.data_mov"(%20) : (i64) -> i64
[MapToAcceleratorPass] Topologically sorted op: "neura.return"(%27) : (i64) -> ()

It's mixed sorted order is:

[MapToAcceleratorPass] ALAP Bucket Level 0: 6 ops
  %1 = neura.reserve : i64
  %2 = neura.reserve : i1
  %3 = "neura.constant"() <{value = 16 : i64}> : () -> i64
  %6 = "neura.constant"() <{value = 0 : i64}> : () -> i64
  %9 = "neura.data_mov"(%3) : (i64) -> i64
  %13 = "neura.data_mov"(%6) : (i64) -> i64
[MapToAcceleratorPass] ALAP Bucket Level 1: 8 ops
  %0 = neura.reserve : i64
  %5 = "neura.constant"() <{value = 1 : i64}> : () -> i64
  %11 = "neura.data_mov"(%5) : (i64) -> i64
  %10 = neura.invariant %9, %2 : i64, i1 -> i64
  %14 = neura.carry %13, %2, %1 : i64, i1, i64 -> i64
  %16 = "neura.data_mov"(%10) : (i64) -> i64
  %24 = "neura.data_mov"(%14) : (i64) -> i64
  %15 = "neura.data_mov"(%14) : (i64) -> i64
[MapToAcceleratorPass] ALAP Bucket Level 2: 9 ops
  %4 = "neura.constant"() <{value = 1 : i64}> : () -> i64
  %7 = "neura.data_mov"(%4) : (i64) -> i64
  %12 = neura.carry %11, %2, %0 : i64, i1, i64 -> i64
  %21 = "neura.data_mov"(%12) : (i64) -> i64
  %22 = "neura.data_mov"(%12) : (i64) -> i64
  %18 = "neura.data_mov"(%12) : (i64) -> i64
  %17 = "neura.icmp"(%15, %16) <{cmpType = "slt"}> : (i64, i64) -> i1
  neura.ctrl_mov %17 -> %2 : i1 i1
  %19 = "neura.data_mov"(%17) : (i1) -> i1
[MapToAcceleratorPass] ALAP Bucket Level 3: 6 ops
  %8 = neura.invariant %7, %2 : i64, i1 -> i64
  %25 = "neura.data_mov"(%8) : (i64) -> i64
  %23 = "neura.add"(%21, %22) : (i64, i64) -> i64
  neura.ctrl_mov %23 -> %0 : i64 i64
  %20 = neura.false_steer %18, %19 : i64, i1 -> i64
  %27 = "neura.data_mov"(%20) : (i64) -> i64
[MapToAcceleratorPass] ALAP Bucket Level 4: 3 ops
  %26 = "neura.add"(%24, %25) : (i64, i64) -> i64
  neura.ctrl_mov %26 -> %1 : i64 i64
  "neura.return"(%27) : (i64) -> ()

Here %8 = neura.invariant %7, %2 : i64, i1 -> i64 is a backward user of %17 = "neura.icmp"(%15, %16) <{cmpType = "slt"}> : (i64, i64) -> i1, but it's ALAP level is higher than %17. Making it unable to satisfy the producer-consumer dependency check in

dataflow/lib/NeuraDialect/Mapping/mapping_util.cpp

Line 845 in 81d782e

assert(!user_locs.empty() && "No locations found for backward user");

[tancheng]

Maybe it is not correctly recognized as a critical op?

dataflow/lib/NeuraDialect/Mapping/mapping_util.cpp

Lines 306 to 311 in 81d782e

	// Step 3: Overwrites critical ops with ASAP schedule: shortest path from
	// source.
	for (Operation *op : sorted_ops) {
	if (!critical_ops.count(op)) {
	continue;
	}

File an issue and resolve it later?