Add command to generate two dot files for e2e kernels

[HobbitQia]

For issue 196

[tancheng]

Thanks @HobbitQia, can you paste the two DFGs in png here?

[HobbitQia]

Thanks @HobbitQia, can you paste the two DFGs in png here?

Sure! Here I provided two DFGs before our transformations:

fir_kernel:

relu_kernel:

And I admit that here are some unnecessary nodes outside the core loop and I can try to remove them for cleanup.

[tancheng]

Thanks @HobbitQia, I meant two DFGs for FIR. original and last level. Also, can you please include the FIR's mlir.llvm IR as well? I don't get the const -> br on the right side of your DFG. They may be viewed as "unnecessary nodes" by you, I am just wondering what are them.

[HobbitQia]

Thanks @HobbitQia, I meant two DFGs for FIR. original and last level. Also, can you please include the FIR's mlir.llvm IR as well? I don't get the const -> br on the right side of your DFG. They may be viewed as "unnecessary nodes" by you, I am just wondering what are them.

For FIR, the DFG before transformation:

the DFG after transformation:

And the IR before transformation:

%0 = llvm.mlir.constant(0 : i64) : i64
    %1 = llvm.mlir.constant(0 : i32) : i32
    %2 = llvm.mlir.constant(1 : i64) : i64
    %3 = llvm.mlir.constant(32 : i64) : i64
    llvm.br ^bb1(%0, %1 : i64, i32)
  ^bb1(%4: i64, %5: i32):  // 2 preds: ^bb0, ^bb1
    %6 = llvm.getelementptr inbounds %arg0[%4] : (!llvm.ptr, i64) -> !llvm.ptr, i32
    %7 = llvm.load %6 {alignment = 4 : i64, tbaa = [#tbaa_tag]} : !llvm.ptr -> i32
    %8 = llvm.getelementptr inbounds %arg2[%4] : (!llvm.ptr, i64) -> !llvm.ptr, i32
    %9 = llvm.load %8 {alignment = 4 : i64, tbaa = [#tbaa_tag]} : !llvm.ptr -> i32
    %10 = llvm.mul %9, %7 overflow<nsw> : i32
    %11 = llvm.add %10, %5 overflow<nsw> : i32
    %12 = llvm.add %4, %2 overflow<nsw, nuw> : i64
    %13 = llvm.icmp "eq" %12, %3 : i64
    llvm.cond_br %13, ^bb2, ^bb1(%12, %11 : i64, i32) {loop_annotation = #loop_annotation}
  ^bb2:  // pred: ^bb1
    llvm.return %11 : i32

I think some nodes that are constant or arguments of functions, which make the DFG look redundant. I am not sure whether I can remove them from DFG or not...

[tancheng]

Oh, they are necessary. The issue is the dot or MLIR dot parser cannot represent br -> BB label correctly.

[tancheng]

Hi @syifan, two json files are generated for illustration of DFGs.

[syifan]

Got it. So what does the "transformation" do here?

[tancheng]

Got it. So what does the "transformation" do here?

"transformation" transforms original ctrl flow IRs to dataflow IRs. For example, br is replaced with phi.