|
This BuildMlirFromTosaFile API deserializes a tosa fbs or
json file and returns a mlir module as OwningOpRef<ModuleOp>
This also refactors the existing deserialization passes
to use the new API, and then copy the deserialized module's
main function into the existing function, and copy all
attributes of the new function and new module into the
existing function and module.
This allows testing of the new API by running deserialization
passes.
Here is an example showing the attributes on the deserialized
module and functions:
module attributes {tf_saved_model.semantics, tosa.description = "Tosa FBS Converted", tosa.fbs_version = "0.70.0d"} {
func.func @main(%arg0: tensor<1x256x256x3xui8> {tf_saved_model.index_path = ["input_0"]}) -> (tensor<1x1x17x3xf32> {tf_saved_model.index_path = ["output_0"]}) attributes {tf.entry_function = {inputs = "tosa_de
serialized_input_0:0", outputs = "tosa_deserialized_output_0:0"}, tf_saved_model.exported_names = ["tosa_deserialized"]} {
Signed-off-by: Tai Ly <tai.ly@arm.com>
Change-Id: Ia6c0202ef43ce5d37788cd459ed7c3f8424dd619
|