Lowering stablehlo.custom_call to linalg

[endaiHW]

Request description

I am trying to lower stablehlo to linalg, but the --stablehlo-legalize-to-linalg pass does not seem to convert stablehlo.custom_call. Is there a suitable pass available to handle this?

[sdasgup3]

You're right, the --stablehlo-legalize-to-linalg pass in MLIR doesn't directly handle stablehlo.custom_call operations. My best guess for this is because custom_call ops are inherently implementation-defined and can represent arbitrary functions or external library calls. There's no general way to convert them to standard Linalg operations without knowing their specific implemetation.

In any case, can you please let me know a bit more about your specific scenario. For example, is it always possible to identify target function via call_target_name?

[endaiHW]

I'm lowering StableHLO exported from AF2 to Linalg and encountered an external call to lapack_ssyevd_ffi. Can this be directly lowered to func.func, or are there additional steps needed to handle such external calls properly?
%4:3 = stablehlo.custom_call @lapack_ssyevd_ffi(%3) {mhlo.backend_config = {mode = 86 : ui8, uplo = 76 : ui8}, operand_layouts = [dense<[2, 3, 1, 0]> : tensor<4xindex>], output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [0], operand_index = 0, operand_tuple_indices = []>], result_layouts = [dense<[2, 3, 1, 0]> : tensor<4xindex>, dense<[2, 1, 0]> : tensor<3xindex>, dense<[1, 0]> : tensor<2xindex>]} : (tensor<1x68x4x4xf32>) -> (tensor<1x68x4x4xf32>, tensor<1x68x4xf32>, tensor<1x68xi32>)