Segmentation fault when emitting gpu dialect with python binding

MLIR
1

Hi folks,

I am encountered some issues when trying to emit gpu dialect with python binding.

This snippet

with Context(), Location.unknown():
    module = Module.create()
    with InsertionPoint(module.body):
        # gpu.GPUModuleOp(operation: object)
        # Cannot find any clue about what should be passed to above API.
        # Use general operation creation API instead
        gpu_module = Operation.create(
            name="gpu.module",
            results=[], operands=[],
            attributes={"sym_name": StringAttr.get("ops")},
            successors=None,
            regions=1
        )

        gpu_module_block = Block.create_at_start(gpu_module.body)
        with InsertionPoint(gpu_module_block):
            gpu.ModuleEndOp()

        with InsertionPoint.at_block_terminator(gpu_module_block):
            # gpu.GPUFuncOp(operation: object)
            # Cannot find any clue about what should be passed to above API.
            # Use general operation creation API instead
            f = Operation.create(
                name="gpu.func",
                results=[], operands=[],
                attributes={
                    "sym_name": StringAttr.get("kernel"),
                    "function_type": TypeAttr.get(FunctionType.get([], [])),
                },
                successors=None,
                regions=1
            )
            with InsertionPoint(Block.create_at_start(f.regions[0])):
                gpu.ReturnOp([])

print(module)

gives me a segmentation fault, but follows are workable.

with Context(), Location.unknown():
    module = Module.create()
    with InsertionPoint(module.body):
        # gpu.GPUModuleOp(operation: object)
        # Cannot find any clue about what should be passed to above API.
        # Use general operation creation API instead
        gpu_module = Operation.create(
            name="gpu.module",
            results=[], operands=[],
            attributes={"sym_name": StringAttr.get("ops")},
            successors=None,
            regions=1
        )

        gpu_module_block = Block.create_at_start(gpu_module.body)
        with InsertionPoint(gpu_module_block):
            gpu.ModuleEndOp()

print(module)

gives

module {
  gpu.module @ops {
  }
}

and

with Context(), Location.unknown():
    module = Module.create()
    with InsertionPoint(module.body):
        # gpu.GPUModuleOp(operation: object)
        # Cannot find any clue about which operation object should be passed to above API.
        # Use general operation creation API instead
        gpu_module = Operation.create(
            name="gpu.module",
            results=[], operands=[],
            attributes={"sym_name": StringAttr.get("ops")},
            successors=None,
            regions=1
        )

        gpu_module_block = Block.create_at_start(gpu_module.body)
        with InsertionPoint(gpu_module_block):
            gpu.ModuleEndOp()

        with InsertionPoint.at_block_terminator(gpu_module_block):
            f = func.FuncOp("kernel", ([], []))
            with InsertionPoint(f.add_entry_block()):
                func.ReturnOp([])

print(module)

gives

module {
  gpu.module @ops {
    func.func @kernel() {
      return
    }
  }
}

and

with Context(), Location.unknown():
    module = Module.create()
    with InsertionPoint(module.body):
        # gpu.GPUModuleOp(operation: object)
        # Cannot find any clue about what should be passed to above API.
        # Use general operation creation API instead
        gpu_module = Operation.create(
            name="gpu.module",
            results=[], operands=[],
            attributes={"sym_name": StringAttr.get("ops")},
            successors=None,
            regions=1
        )

        gpu_module_block = Block.create_at_start(gpu_module.body)
        with InsertionPoint(gpu_module_block):
            gpu.ModuleEndOp()

        with InsertionPoint.at_block_terminator(gpu_module_block):
            f = Operation.create(
                name="func.func",
                results=[], operands=[],
                attributes={
                    "sym_name": StringAttr.get("kernel"),
                    "function_type": TypeAttr.get(FunctionType.get([], [])),
                },
                successors=None,
                regions=1
            )
            with InsertionPoint(Block.create_at_start(f.regions[0])):
                func.ReturnOp([])

print(module)

also gives

module {
  gpu.module @ops {
    func.func @kernel() {
      return
    }
  }
}

Ddi I miss anything?

2

You can run python under gdb to debug segfaults. Here it seems like the issue is in gpu.func’s verifier, which crashes at this line because there’s no workgroup_attributions attribute on the op: llvm-project/GPUDialect.cpp at 77b202f974faa82679d9984aeef58355dc80ba80 · llvm/llvm-project · GitHub

3

Thanks for the kind help! It works. BTW, do you have any suggestions about using GPUModuleOp and GPUFuncOP to emit gpu dialect? General Operation.create seems not easy to use, and a little bit tricky for beginners like me.

4

I suspect that the people doing real work with the GPU dialect are using the c++ interface. To be more usable, it needs extensions apis added. If you want to contribute, look at what was done for the func dialect. Using Operation.create is a backdoor for directly creating the low level representation, and if you are finding that you need to use it, it is a sign that work had not been done to define the API for the dialect properly – contributions welcome, but there are not general answers on how to do that (it varies based on judgment and understanding of the core data structures).

5

This is still there at HEAD, can someone file a bug so we fix this?
(@ftynse this is your code I think)

6

filed here: [mlir][gpu] Segfault in gpu.func verifier on invalid input · Issue #58045 · llvm/llvm-project · GitHub