[P1] Clean the predicate attribute or operands in neura operations

[ShangkunLi]

We may not need a predicate attribute or operands in neura operations any more.

We need clean them up.

[tancheng]

Hi @MeowMJ, can you help with this cleanup? it is simple and could help you ramp up.

Basically, just remove all the Optional<AnyType>:$predicate from each op's declaration, as we now embed the predicate bit into the data, we don't need provide a predicate operand to an operation

dataflow/include/NeuraDialect/NeuraOps.td

Line 21 in 3bc45aa

let arguments = (ins AnyType:$lhs, AnyType:$rhs, Optional<AnyType>:$predicate);

[tancheng]

Hi @MeowMJ, any update on this?

[MeowMJ]

Hi @MeowMJ, any update on this?

No. I lost my weekend dealing with the evil landlord :( Will finish it quickly.

[MeowMJ]

How is the predicate bit embedded into the data? I have finished the code, and need this info to modify the .mlir to pass the check.

[tancheng]

This pass convert conventional value to predicated value:

dataflow/test/neura/ctrl/branch_for.mlir

Line 17 in 7f845d6

// RUN: --leverage-predicated-value \

after this pass, the IR would be changed from

%11 = "neura.add"(%5, %8) : (i64, i64) -> i64

to

%21 = "neura.add"(%19, %13) : (!neura.data<i64, i1>, !neura.data<i64, i1>) -> !neura.data<i64, i1>

Here is the pass' implementation.

---

How is the predicate bit embedded into the data

In terms of functionality, it is the grant_predicate or grant_once embed/update the predicate bit into the data during acceleration:

dataflow/test/neura/ctrl/branch_for.mlir

Lines 138 to 139 in 7f845d6

	// CTRL2DATA-NEXT: %22 = "neura.icmp"(%21, %11) <{cmpType = "slt"}> : (!neura.data<i64, i1>, !neura.data<i64, i1>) -> !neura.data<i1, i1>
	// CTRL2DATA-NEXT: %23 = neura.grant_predicate %21, %22 : !neura.data<i64, i1>, !neura.data<i1, i1> -> !neura.data<i64, i1>

The grant_predicate op is generated by this pass:

dataflow/test/neura/ctrl/branch_for.mlir

Line 35 in 7f845d6

// RUN: --transform-ctrl-to-data-flow \

---

and need this info to modify the .mlir to pass the check.

I thought you don't need to modify the test, as your task is just a cleanup, as we discussed before. I am a bit confused.

[MeowMJ]

I thought you don't need to modify the test, as your task is just a cleanup, as we discussed before. I am a bit confused.

After the cleanup, some tests are failing, and I need to address them.. For example, fdiv.mlir has funcs like test_fdiv_with_predicate_true:

dataflow/test/neura/interpreter/basic_operation/fdiv.mlir

Lines 51 to 58 in 7f845d6

	func.func @test_fdiv_with_predicate_true() -> f32 {
	%a = arith.constant 15.0 : f32
	%b = arith.constant 3.0 : f32
	%pred = arith.constant 1 : i32
	%res = "neura.fdiv"(%a, %b, %pred) : (f32, f32, i32) -> f32
	// CHECK: [neura-interpreter] → Output: 5.000000
	return %res : f32
	}

[tancheng]

Ah, good catch. The test was originally born for the predicate operand you just deleted. I think it is okay to remove this test. Is there any other test failed?

[tancheng]

I think all

dataflow/tools/neura-interpreter/neura-interpreter.cpp

Line 647 in 7f845d6

if (op.getNumOperands() > 2) {

are not needed then. @ShangkunLi WDYT? We shouldn't provide pred as operand for non-predicated-value ops.

But maybe it is okay to leave them there now, as in the future, we still want to support hybrid mode?

[ShangkunLi]

I think all

dataflow/tools/neura-interpreter/neura-interpreter.cpp

Line 647 in 7f845d6

if (op.getNumOperands() > 2) {
are not needed then. @ShangkunLi WDYT? We shouldn't provide pred as operand for non-predicated-value ops.
But maybe it is okay to leave them there now, as in the future, we still want to support hybrid mode?

We want to support hybrid mode. But even in hybrid mode, there should not be any predicate operands. We only want pure data operands for BB switch mode, and embed the predicate bit in data for dataflow mode.

[tancheng]

I think all
dataflow/tools/neura-interpreter/neura-interpreter.cpp
Line 647 in 7f845d6
if (op.getNumOperands() > 2) {
are not needed then. @ShangkunLi WDYT? We shouldn't provide pred as operand for non-predicated-value ops.
But maybe it is okay to leave them there now, as in the future, we still want to support hybrid mode?

We want to support hybrid mode. But even in hybrid mode, there should not be any predicate operands. We only want pure data operands for BB switch mode, and embed the predicate bit in data for dataflow mode.

Makes sense. But @MeowMJ no need to delete the interpreter's code now (unless you already done it). We can do the cleanup later. @itemkelvin excellent robust implementation :-)

[MeowMJ]

Ah, good catch. The test was originally born for the predicate operand you just deleted. I think it is okay to remove this test. Is there any other test failed?

35 tests failed. The reasons for failing can be divided into the following:

• Tests designed for the predicate. They are in neura/interpreter/basic_operation/ folder. These can be addressed by deleting them.
• Tests with unmatched output using // RUN: --leverage-predicated-value \. They are in other folders. Apparently, dataflow/lib/NeuraDialect/Transforms^CeveragePredicatedValuePass.cpp didn't convert all conventional value to predicated value. These can be addressed by editing the //CHECK ... lines in my understanding.

The test log:

Failed Tests (35):
  Neura Dialect Tests :: affine2neura/bert/bert_node0/bert_node0.mlir
  Neura Dialect Tests :: affine2neura/bert/bert_node1/bert_node1.mlir
  Neura Dialect Tests :: affine2neura/bert/bert_node2/bert_node2.mlir
  Neura Dialect Tests :: affine2neura/bert/bert_node28/bert_node28.mlir
  Neura Dialect Tests :: affine2neura/bert/bert_node3/bert_node3.mlir
  Neura Dialect Tests :: affine2neura/bert/bert_node8/bert_node8.mlir
  Neura Dialect Tests :: affine2neura/bert/bert_node9/bert_node9.mlir
  Neura Dialect Tests :: code_gen/test_code_generate.mlir
  Neura Dialect Tests :: controflow_fuse/complex_nested/complex_nested.mlir
  Neura Dialect Tests :: controflow_fuse/non_perfect_nested/non_perfect_nested.mlir
  Neura Dialect Tests :: controflow_fuse/perfect_nested/perfect_nested.mlir
  Neura Dialect Tests :: controflow_fuse/perfect_reduction/perfect_reduction.mlir
  Neura Dialect Tests :: controflow_fuse/simple_loop/simple_loop.mlir
  Neura Dialect Tests :: controflow_fuse/simple_loop_reduction/simple_loop_reduction.mlir
  Neura Dialect Tests :: mapping_quality/branch_for.mlir
  Neura Dialect Tests :: mapping_quality/tiny_loop.mlir
  Neura Dialect Tests :: neura/ctrl/branch.mlir
  Neura Dialect Tests :: neura/ctrl/branch_for.mlir
  Neura Dialect Tests :: neura/ctrl/branch_with_and_without_arg.mlir
  Neura Dialect Tests :: neura/ctrl/branch_without_arg.mlir
  Neura Dialect Tests :: neura/ctrl/for_with_if.mlir
  Neura Dialect Tests :: neura/ctrl/nested_branch.mlir
  Neura Dialect Tests :: neura/for_loop/test.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/add.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/cast.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/cond_br.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/fadd.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/fadd_fadd.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/fdiv.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/fmul.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/fmul_add.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/phi.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/predicated_data.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/sub.mlir
  Neura Dialect Tests :: neura/interpreter/basic_operation/vfmul.mlir

[tancheng]

what changes you need to make for https://github.com/coredac/dataflow/blob/main/test/neura/ctrl/branch_for.mlir? Remove predicate = true from const? This is a reasonable change needs to be done. Any other type of change across the 30+ files?