When MatmulTranpose reports incorrect shapes of the arguments
it cannot represent itself as MATMUL, because the reading
of the first argument's shape will be confusing.
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
Thank you for the review, Kiran!
I guess the messages could be changed to report the corresponding MATMUL or TRANSPOSE errors depending on which part of the combined call the error is related to, but I will leave the decision to Tom. This is just a quick change to avoid confusing reports from the runtime.
Sure. Tom is not yet back at work. I was just temporarily standing in for Tom here. Once he is back, he can update the messages if required. CC: @tblah
Thanks for stepping in @kiranchandramohan. I think combined messages for matmul-transpose are fine because the messages apply equally to both operations. I don't think these need to be especially user friendly because these messages shouldn't ever happen.
If somebody really does want to reference them to a real intrinsic in source then I think the TRANSPOSE should be used because conceptually that is evaluated first.