rebase

rebase, and apply functional-type widely as suggested.

rebase and reslove conflict

Is there any remaining issues? could you land this for me?

Thanks for the explanation. Your statement is clear, and I don't have further question.

Hi, this is a great useful feature. I just have a few question on it.

Yes. This demand is reasonable. We loosen the trait to accept those use cases.

update the trait in a relexed way to allow the result type can be other size but not i32 as long as this is a valid size.

missing < in the expected output string of the test.

This is a fix to response the mistake introduced in D150472.
Quote ""tosa.mul"(%0, %1) {shift = 0 : i32} : (tensor<4xi8>, tensor<4xi8>) -> tensor<4xi32> is marked as illegal because result and operand types don't match, even though i8xi8 = i32 is according to the spec".

@mgehre-amd, oh, you're right. I'll send a PR to fix this soon.

put tosa.mul into element-wise binary group.

Fix the typo of Binary to Unary for clamp, sigmoid and tanh.

In D150472#4341193, @mgehre-amd wrote:

Thanks for fixing this!

I think clamp, sigmoid and tanh are also Tosa_ElemWiseUnaryOp.
Other ops like identity, reshape, concat, pad, reverse, slice, tile and transpose should also have same element types on input and output.

Declaring clamp, sigmoid and tanh as Tosa_ElemWiseUnaryOp.

Define the new accumulator type as TypeAttr rather than EnumAttr.

Look like the recent build of mlir repo isn't stable.
Have to be careful to pick a commit to rebase.
Now rebase this PR to a commit that can pass bazel build test locally.

Rebase to include bazel fix for transform dialect (not relevant to this patch)

Good spot. Thanks for the reminder. I missed TosaToLinalg part. Reading from the acctype attr is complemented in newer patch. Do I still miss something necessary else?

Rebase to include OUTSIDE fix for bazel build error.

Read the acc type attr in TosaToLinalg.

As suggested, minimize the change by adding the necessary checkers only.

As suggested, minimize the change by adding the necessary checkers only.

Rebase onto top-of-tree to include the fix that isn't relevant to this patch.

Rebase onto top-of-tree to include the fix that isn't relevant to this patch.

Populate and call checkers (function) directly to avoid virtual function lookup and class
overhead.

Instead of checking a series of if/else operations, adopt operation type look-up method
by implementing polymorphism. I think that may easier to extend to involve more operations
or other checking patterns. And address other review comments.

Define these constant operand checking functions as function template. That might make
code cleaner as it is likely more operations will be added.

Thanks. Feedbacks from you guys are useful and constructive. Let me abandon this PR and go along the suggested way.

Sorry for confusing you. I've not noticed there is a mismatch of attribute meaning between TOSA and MLIR. Could you explain a bit about the difference or why those arguments are better to be MLIR operands rather than MLIR attributes? I thought both attribute types are used to specify constant data on operations.
Alternative way without the pad definition changing, we're trying to add a section in a pass in MLIR to verify if the pad operands are constant in compile time. Any suggestions are welcome.

As suggested, mutate the referenced inputs passed in to avoid doing the extra reassigning and stateful variabled in the lowering.

update the patch title

roll back to use the existing pair-wise broadcasting function. Apply each broadcasting to 3 different pairs of inputs. By chaining them this way as a compound the broadcasting happens all at once.
update mlir tests.

overload reshapeLowerToHigher to support 3 inputs so that the broadcasting happens all at once.
rebase.

• replace failure with notifyMatchFailure
• address review comment
• add more mlir tests
• add comment in code

Remove this pass from the SCF pass as these dialects is independent of each other.

fix clang format error

fix: include header in the wrong place.

fix compilation error.

Thanks. Good to know the enum can be defined in Tblgen. Also change to use walk function as suggested.

Define the enum in tablegen for the pass, and use a walk on the operation instead.

Instead of comparing with -1, change to use ShapedType::isDynamic

Rebase

Add check whether input tensor is dynamic shape.

Rebase and resolve conflicts