Hello,
I am a trying to lower a basic TensorFlow model to one of the MLIR dialects such as Linalg or affine dialects. I am following the pipeline
TensorFlow model(graphdef format) → MLIR TF (Standard CF) → MLIR xla_hlo → MLIR xla_lhlo → linalg / affine .
I assume that is the correct order of lowering if I am not missing anything.
I used ./tf-mlir-translate --graphdef-to-mlir to convert the graphdef to temp.mlir.
I also used ./tf-opt --tf-standard-pipeline ./temp.mlir -o temp2.mlir to bring it to standard.
Now I am trying to lower it to lhlo using
./tf-opt --xla-hlo-to-lhlo-with-xla temp2.mlir -o lhlo.mlir but facing an issue here. I am getting the following error message.
%constant.10 = f32[200,10]{1,0} constant({…}) xla.mlir:3:1: error: Internal: LHLO opcode constant is not supported.
converting HLO to LHLO
module attributes {tf.versions = {bad_consumers = , min_consumer = 0 : i32, producer = 0 : i32}} {
^
xla.mlir:3:1: note: see current operation: “module”() ( {
“func”() ( {
^bb0(%arg0: memref<8000xi8>, %arg1: memref<8000xi8>): // no predecessors
“std.return”() : () → ()
}) {arg0 = {lmhlo.alloc = 0 : index, lmhlo.liveout = true}, arg1 = {lmhlo.alloc = 1 : index}, sym_name = “main.11”, type = (memref<8000xi8>, memref<8000xi8>) → ()} : () → ()
“module_terminator”() : () → ()
}) {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 0 : i32}} : () → ()
I assume it has something to do with kConstant Hloopcode not implemented. Am I missing a pass somewhere while lowering or does it have something to do with the model itself? I am a beginner at this and trying to learn more about MLIR.If I am missing something or doing something wrong it would be helpful if you can point me at the right direction.
Thanks,
J