Visualizing MLIR codes to generate DFGs through graphviz

[HobbitQia]

The document and example can be found in test/visualize.

[HobbitQia]

It seems that this PR should be resolved first then I can keep pushing new commits to my GitHub repo...

[tancheng]

It seems that this PR should be resolved first then I can keep pushing new commits to my GitHub repo...

It would be perfect if a PR is this small :-)
But why this PR should be merged first, you are always able to push stuff into your branch, right? What am I missing?

[HobbitQia]

It seems that this PR should be resolved first then I can keep pushing new commits to my GitHub repo...

It would be perfect if a PR is this small :-) But why this PR should be merged first, you are always able to push stuff into your branch, right? What am I missing?

Yes. currently this PR only contains the basic steps to visualize DFGs without something about assemblyFormat. I mean if I push my further commits to my repos, these commits will automatically become a part of this PR, which is not what we want to see so I revert them in e7c1903.

[HobbitQia]

Thanks for Shangkun's help! Now we can solely rely onmlir-neura-opt to generate dot files and I have modified test.mlir so it can pass tests locally.

[HobbitQia]

Hi @tancheng

When I try to enable multiple createPrintOpGraph passes in neuraPasses.cpp, I found that there will be much output information in stderr (not only dot file). e.g., [ctrl2data] Asserting live-out values dominated by block arguments. It will influence the generation of the png since graphviz cannot neglect this information to proceed its generation.

[ShangkunLi]

Hi @tancheng

When I try to enable multiple createPrintOpGraph passes in neuraPasses.cpp, I found that there will be much output information in stderr (not only dot file). e.g., [ctrl2data] Asserting live-out values dominated by block arguments. It will influence the generation of the png since graphviz cannot neglect this information to proceed its generation.

The problem comes from trying to convert stderr into content in a file by using 2>&1. This will convert all the debugging info printed by llvm::errs() into a file.

I think a better way would be to explicitly assign an os for the --view-op-graph pass when we create this pass. Or it will use the llvm::errs() as the os by default.
https://github.com/llvm/llvm-project/blob/78b45b489559cd44437af4eac6c10188eb0cec3c/mlir/lib/Transforms/ViewOpGraph.cpp#L102

For example,

std::string filename = "opgraph.dot";
llvm::raw_fd_ostream os(filename, ec);
pm.addPass(mlir::createPrintOpGraphPass(os));