Seeking Guidance on Resolving Issues with Custom MLIR Frontend and Dialect Registration

MLIR
1

I’m working on developing a custom out-of-tree MLIR frontend for a specialized neural network model format. The current goal is to create a program that parses this custom model file and converts it into MLIR code using a custom dialect specific to the model format.

Here’s my current code snippet:

mlir::DialectRegistry registry;
registry.insert<mlir::MyDialect>();
registerAllDialects(registry);
MLIRContext ctx(registry);

mlir::ModuleOp moduleOp = readFromFile(ctx, inputFilename);
moduleOp->print(output->os(), OpPrintingFlags{}.printGenericOpForm(false));

The readFromFile function is responsible for reading the model format and converting it to MLIR. Inside the module, I want to use func.func to represent the model code, and I create it like this:

return builder.create<func::FuncOp>(loc, funcName, funcType);

However, I encountered an issue with the following error message:

LLVM ERROR: Building op `func.func` but it isn't registered in this MLIRContext: the dialect may not be loaded or this operation isn't registered by the dialect. See also https://mlir.llvm.org/getting_started/Faq/#registered-loaded-dependent-whats-up-with-dialects-management

To address this error, I tried adding ctx.loadAllAvailableDialects(); to my code, which resolved the previous issue. However, I now face a new problem: despite specifying that I want non-generic operations, the output still includes generic operations. Here’s an example of the undesired behavior:

"builtin.module"() ({
  "func.func"() <{function_type = (tensor<784x1xf32>) -> (), sym_name = "mnist"}> ({
  }) : () -> ()
}) : () -> ()

I couldn’t find any loadAllAvailableDialects calls in MlirOptMain, so I suspect I might be doing something wrong initially, leading to the ICE. I’m seeking guidance on how to resolve this issue.

1 Like
2

Your func op failed verification and so was printed in generic format. Probably in your tool you are using you aren’t reporting errors. Try passing this output through mlir-opt and it will show

<stdin>:22:3: error: 'func.func' op symbol declaration cannot have public visibility
<stdin>:22:3: note: see current operation:
"func.func"() <{function_type = (tensor<784x1xf32>) -> (), sym_name = "mnist"}> ({
}) : () -> ()
3

Thanks!
Could you please tell how I can fix the problem? Or something to read on the symbol visibility and why functions can’t have the public one.

4

Indeed, you are correct. By adding the following line of code:

funcOp.setPrivate();

I was able to resolve the issue with the printing. Thank you for your assistance!

Now, I have a follow-up question: if I intend to create a JIT engine later and call the function from C++ code, how can I make the function public? Or is this not related to the linking attributes?

Regarding the ICE issue without using loadAllAvailableDialects, is this behavior expected?

Lastly, how can I print the errors effectively? I believe having error messages would be quite useful. Should I register a handler in the Diagnostics Engine to achieve this?

5

See FAQ - MLIR

6

Here you can’t have it public because it does not have a body. If you define a function then you can have it public.

7

Yeah, I couldn’t see the solution staring me in the face. I need to read more carefully next time…

Oh, I see. Seems reasonable, thanks for the help!