Enable end2end affine-to-neura lowering

[ShangkunLi]

In this pr:

• Add lowering from memref and builtin dialects
• Add neura.load_indexed/store_indexed operation for memref like memory access
• Fix some typos in files

Now, we can write code in cpp and lower it to affine dialect using polygeist. And more high-level transforms can be implemented in affine level, like polyhedral-based optimization, loop-unroll/fusion/fission/interchange/tiling/vectorize, etc.

[ShangkunLi]

Sry, some errors occurred when I try to transform bert_nodex.mlir into dataflow mlir.

Will fix it soon.

[ShangkunLi]

Sry, some errors occurred when I try to transform bert_nodex.mlir into dataflow mlir.

Will fix it soon.

The problem arises from the TransformCtrlToDataFlowPass.cpp incorrectly handling nested loops. I may try to solve it.

• Current pass cannot handle the cond_br that do not pass arguments, like

^bb3(%8: !neura.data<i64, i1>):  // 2 preds: ^bb2, ^bb4
    %9 = "neura.cast"(%8) <{cast_type = "int_to_index"}> : (!neura.data<i64, i1>) -> !neura.data<index, i1>
    %10 = "neura.icmp"(%9, %1) <{cmpType = "slt"}> : (!neura.data<index, i1>, !neura.data<index, i1>) -> !neura.data<i1, i1>
    neura.cond_br %10 : !neura.data<i1, i1> then to ^bb4 else to ^bb5
^bb4:  // pred: ^bb3
    %11 = neura.load_indexed %arg0[%2, %2, %2, %2, %2, %9 : !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>] memref<?x1x1x1x1x128xi8> : !neura.data<i8, i1>
    neura.store_indexed %11 to %arg1[%2, %2, %5, %2, %2, %9 : !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>] memref<?x1x128x1x1x128xi8> : !neura.data<i8, i1>
    %12 = "neura.add"(%9, %0) : (!neura.data<index, i1>, !neura.data<index, i1>) -> !neura.data<index, i1>
    %13 = "neura.cast"(%12) <{cast_type = "index_to_int"}> : (!neura.data<index, i1>) -> !neura.data<i64, i1>
    neura.br %13 : !neura.data<i64, i1> to ^bb3

• Errors when inserting reserve & ctrl_mov for nested loop.

[tancheng]

Sry, some errors occurred when I try to transform bert_nodex.mlir into dataflow mlir.
Will fix it soon.

The problem arises from the TransformCtrlToDataFlowPass.cpp incorrectly handling nested loops. I may try to solve it.

• Current pass cannot handle the cond_br that do not pass arguments, like

^bb3(%8: !neura.data<i64, i1>):  // 2 preds: ^bb2, ^bb4
    %9 = "neura.cast"(%8) <{cast_type = "int_to_index"}> : (!neura.data<i64, i1>) -> !neura.data<index, i1>
    %10 = "neura.icmp"(%9, %1) <{cmpType = "slt"}> : (!neura.data<index, i1>, !neura.data<index, i1>) -> !neura.data<i1, i1>
    neura.cond_br %10 : !neura.data<i1, i1> then to ^bb4 else to ^bb5
^bb4:  // pred: ^bb3
    %11 = neura.load_indexed %arg0[%2, %2, %2, %2, %2, %9 : !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>] memref<?x1x1x1x1x128xi8> : !neura.data<i8, i1>
    neura.store_indexed %11 to %arg1[%2, %2, %5, %2, %2, %9 : !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>] memref<?x1x128x1x1x128xi8> : !neura.data<i8, i1>
    %12 = "neura.add"(%9, %0) : (!neura.data<index, i1>, !neura.data<index, i1>) -> !neura.data<index, i1>
    %13 = "neura.cast"(%12) <{cast_type = "index_to_int"}> : (!neura.data<index, i1>) -> !neura.data<i64, i1>
    neura.br %13 : !neura.data<i64, i1> to ^bb3

• Errors when inserting reserve & ctrl_mov for nested loop.

Thanks @ShangkunLi, I didn't handle such case as GPT/Gemini told me MLIR has the rule that all basic block would only have arguments as live-ins (rather than directly use previously existing variable in other blocks), but it seems not always correct. Do you wanna fix this in this PR or later? I didn't see the --transform-ctrl-to-data is used in bert_node0.mlir?

[ShangkunLi]

Sry, some errors occurred when I try to transform bert_nodex.mlir into dataflow mlir.
Will fix it soon.

The problem arises from the TransformCtrlToDataFlowPass.cpp incorrectly handling nested loops. I may try to solve it.

• Current pass cannot handle the cond_br that do not pass arguments, like

^bb3(%8: !neura.data<i64, i1>):  // 2 preds: ^bb2, ^bb4
    %9 = "neura.cast"(%8) <{cast_type = "int_to_index"}> : (!neura.data<i64, i1>) -> !neura.data<index, i1>
    %10 = "neura.icmp"(%9, %1) <{cmpType = "slt"}> : (!neura.data<index, i1>, !neura.data<index, i1>) -> !neura.data<i1, i1>
    neura.cond_br %10 : !neura.data<i1, i1> then to ^bb4 else to ^bb5
^bb4:  // pred: ^bb3
    %11 = neura.load_indexed %arg0[%2, %2, %2, %2, %2, %9 : !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>] memref<?x1x1x1x1x128xi8> : !neura.data<i8, i1>
    neura.store_indexed %11 to %arg1[%2, %2, %5, %2, %2, %9 : !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>, !neura.data<index, i1>] memref<?x1x128x1x1x128xi8> : !neura.data<i8, i1>
    %12 = "neura.add"(%9, %0) : (!neura.data<index, i1>, !neura.data<index, i1>) -> !neura.data<index, i1>
    %13 = "neura.cast"(%12) <{cast_type = "index_to_int"}> : (!neura.data<index, i1>) -> !neura.data<i64, i1>
    neura.br %13 : !neura.data<i64, i1> to ^bb3

• Errors when inserting reserve & ctrl_mov for nested loop.

Thanks @ShangkunLi, I didn't handle such case as GPT/Gemini told me MLIR has the rule that all basic block would only have arguments as live-ins (rather than directly use previously existing variable in other blocks), but it seems not always correct. Do you wanna fix this in this PR or later? I didn't see the --transform-ctrl-to-data is used in bert_node0.mlir?

Filed an issue #54. I may try to fix this in the next pr. For this pr, I just tested these lowering patterns.

[tancheng]

Thanks for the PR, let's wait for the pass on github action before merging :-)