[Relay][VM] Fix code generation for packed functions + tuples #3287
Conversation
src/relay/op/tensor/transform.cc
Outdated
| @@ -292,7 +292,7 @@ RELAY_REGISTER_OP("concatenate") | |||
| .add_type_rel("Concatenate", ConcatenateRel) | |||
| .set_attr<FInferCorrectLayout>("FInferCorrectLayout", ConcatenateLayout) | |||
| .set_attr<FTVMCompute>("FTVMCompute", ConcatenateCompute) | |||
| .set_attr<TOpPattern>("TOpPattern", kInjective); | |||
| .set_attr<TOpPattern>("TOpPattern", kOpaque); | |||
There was a problem hiding this comment.
Maybe leave this line out of the PR?
|
Please remove concat opaque property for now, and consider manually construct a function that needs to deal with such a case |
|
I don't think it is possible to write such a function because anything that uses tuples gets fused currently. Used |
| } else if (auto tuple_ty = ty.as<TupleTypeNode>()) { | ||
| for (size_t f = 0; f < tuple_ty->fields.size(); f++) { | ||
| const auto& field = tuple_ty->fields[f]; | ||
| CHECK(field.as<TensorTypeNode>()) |
There was a problem hiding this comment.
can you extend it and allow recursive tuple? it is more uniform this way.
There was a problem hiding this comment.
I don't really want to do it right now, recursive is a pain given the way the code generator is written right now. This is mostly to fix CI breakage, I would like to make some updates to compiler likely after FCRC tutorial.
|
Thanks, @jroesch @MarisaKirisame @wweic , this PR is now merged |
Disabling fusion for concat in #3268 uncovered a bug in the VM compiler, this fixes it.
We were previously not unpacking tuples, this changes code generation to unpack the tuples into registers before invoking the function.
cc @wweic @zhiics @icemelon9 @MarisaKirisame