Remove grant_always

[ShangkunLi]

In this pr:

• Remove the grant_always operations in --transform-ctrl-to-data-pass
• After removing the grant_always, we can achive CompiledII = 6 : i32, RecMII = 4 : i32, ResMII = 2 : i32 for branch_for.mlir (previous compiledII=7) and CompiledII = 3 : i32, RecMII = 3 : i32, ResMII = 2 : i32 for simpleloop.mlir (previous compiledII=5).

[tancheng]

Hmm, can this constant binding be done in FusePatternsPass.cpp?

[ShangkunLi]

Hmm, can this constant binding be done in FusePatternsPass.cpp?

I didn't add the constant binding feature in this pr yet. And I creates a single --fuse-constant-and-grant-once for the constant binding. Because the ideal flow of fusion is:

Fuse constant and grant_once -> fuse control flow related operations (e.g., loop control related operations) -> fuse computation related operations (e.g., add-mul, add-add, etc.).

Since I need the ir with fused constant and grant_once to identify the loop structures, I think we cannot realize it in FusePatternsPass.cpp

[tancheng]

Can you explain a bit on: "Remove the grant_always operations in --transform-ctrl-to-data-pass"?

What would it look like if grant_always is removed? how is the constant passed in then?

[tancheng]

Can you explain a bit on: "Remove the grant_always operations in --transform-ctrl-to-data-pass"?

What would it look like if grant_always is removed? how is the constant passed in then?

It looks like "replace" instead of "remove".

[ShangkunLi]

Can you explain a bit on: "Remove the grant_always operations in --transform-ctrl-to-data-pass"?
What would it look like if grant_always is removed? how is the constant passed in then?

It looks like "replace" instead of "remove".

We use the grant_once for all the live-out constant ops. And for constants used in other blocks, we use grant_once + grant_predicate to make it valid instead of using grant_always like before.

[tancheng]

Where is --fuse-constant-and-grant-once?

[tancheng]

Since I need the ir with fused constant and grant_once to identify the loop structures, I think we cannot realize it in FusePatternsPass.cpp

I don't get this. Why can't we simply fuse the constant -> grant_once pattern? Why do we need to "identify loop structures"?

[ShangkunLi]

Where is --fuse-constant-and-grant-once?

I haven't made a pr for it yet. And the transformed ir looks like:

module {
  func.func @_Z10simpleloopv() -> i32 attributes {accelerator = "neura", llvm.linkage = #llvm.linkage<external>} {
    %0 = "neura.grant_once"() <{constant_value = 1 : i64}> : () -> !neura.data<i64, i1>
    %1 = "neura.grant_once"() <{constant_value = 128 : i64}> : () -> !neura.data<i64, i1>
    %2 = "neura.grant_once"() <{constant_value = 0 : i32}> : () -> !neura.data<i32, i1>
    %3 = "neura.grant_once"() <{constant_value = 0 : i64}> : () -> !neura.data<i64, i1>
    %4 = neura.reserve : !neura.data<i64, i1>
    %5 = "neura.phi"(%4, %1) : (!neura.data<i64, i1>, !neura.data<i64, i1>) -> !neura.data<i64, i1>
    %6 = neura.reserve : !neura.data<i32, i1>
    %7 = "neura.phi"(%6, %2) : (!neura.data<i32, i1>, !neura.data<i32, i1>) -> !neura.data<i32, i1>
    %8 = neura.reserve : !neura.data<i64, i1>
    %9 = "neura.phi"(%8, %3) : (!neura.data<i64, i1>, !neura.data<i64, i1>) -> !neura.data<i64, i1>
    %10 = "neura.icmp"(%9, %5) <{cmpType = "slt"}> : (!neura.data<i64, i1>, !neura.data<i64, i1>) -> !neura.data<i1, i1>
    %11 = neura.grant_predicate %9, %10 : !neura.data<i64, i1>, !neura.data<i1, i1> -> !neura.data<i64, i1>
    %12 = neura.grant_predicate %7, %10 : !neura.data<i32, i1>, !neura.data<i1, i1> -> !neura.data<i32, i1>
    %13 = neura.grant_predicate %0, %10 : !neura.data<i64, i1>, !neura.data<i1, i1> -> !neura.data<i64, i1>
    %14 = neura.grant_predicate %1, %10 : !neura.data<i64, i1>, !neura.data<i1, i1> -> !neura.data<i64, i1>
    %15 = "neura.not"(%10) : (!neura.data<i1, i1>) -> !neura.data<i1, i1>
    %16 = neura.grant_predicate %7, %15 : !neura.data<i32, i1>, !neura.data<i1, i1> -> !neura.data<i32, i1>
    %17 = "neura.cast"(%11) <{cast_type = "i64_to_i32"}> : (!neura.data<i64, i1>) -> !neura.data<i32, i1>
    %18 = "neura.add"(%12, %17) : (!neura.data<i32, i1>, !neura.data<i32, i1>) -> !neura.data<i32, i1>
    %19 = "neura.add"(%11, %13) : (!neura.data<i64, i1>, !neura.data<i64, i1>) -> !neura.data<i64, i1>
    neura.ctrl_mov %19 -> %8 : !neura.data<i64, i1> !neura.data<i64, i1>
    neura.ctrl_mov %18 -> %6 : !neura.data<i32, i1> !neura.data<i32, i1>
    neura.ctrl_mov %14 -> %4 : !neura.data<i64, i1> !neura.data<i64, i1>
    "neura.return"(%16) : (!neura.data<i32, i1>) -> ()
  }
}

[ShangkunLi]

Since I need the ir with fused constant and grant_once to identify the loop structures, I think we cannot realize it in FusePatternsPass.cpp

I don't get this. Why can't we simply fuse the constant -> grant_once pattern? Why do we need to "identify loop structures"?

I see what you mean. I can modify this.