Question about "krnl.global"()

[yuucyf]

I used onnx-mlir with the --EmitMLIR parameter to generate mlir build-in dialect op from a model file. However, the generated .mlir file still contains operations like "krnl.global"(). Isn't this a krnl op rather than a built-in MLIR dialect op? Shouldn't this be considered a bug?

eg:
%0 = "krnl.global"() {name = "constant_0", shape = [], value = dense<8.000000e+00> : tensor} : () -> memref

[AlexandreEichenberger]

No, it is not a bug, and as a matter of fact, we lower a few kernel operations late. If you wanted to have a pure mlir dialect, you would have to add a pass that adds the lowering of these krnl ops to mlir equivalent ones.

You can search for the rules that lower these specific ops, and add a pass. We would have no issue if you wanted to upstream that code, and maybe create a new target (e.g. -EmitOnlyMLIR or an option like -only-mlir that would activate that pass when the -EmitMLIR target is on.

Hope this helps. We can probably help you a bit with specific issues you may have if you start this effort and want to upstream it. I suspect other folks might be interested in that too, and we are happy to help more folks getting onboard of onnx-mlir.

[yuucyf]

No, it is not a bug, and as a matter of fact, we lower a few kernel operations late. If you wanted to have a pure mlir dialect, you would have to add a pass that adds the lowering of these krnl ops to mlir equivalent ones.

You can search for the rules that lower these specific ops, and add a pass. We would have no issue if you wanted to upstream that code, and maybe create a new target (e.g. -EmitOnlyMLIR or an option like -only-mlir that would activate that pass when the -EmitMLIR target is on.

Hope this helps. We can probably help you a bit with specific issues you may have if you start this effort and want to upstream it. I suspect other folks might be interested in that too, and we are happy to help more folks getting onboard of onnx-mlir.

However, using the -EmitMLIR option might easily lead others to misunderstand that it generates a purely MLIR dialect. Also, I’m not yet sure how many krnl operations are present in the MLIR generated by the -EmitMLIR option?