[Stream] Update executable functions in encoding specialization pass.

[hanhanW]

For the stream bindings, the encoding types of duplicated executables need to be updated with resolved layouts. In the revision, it is done by type interface. The revision introduces EncodingTypeInterface in the IREE::Encoding dialect. There are two interface methods:

• getEncodingType: returns the tensor type with the encoding. E.g., the bounded type is returned in the Flow::DispatchTensorType implementation.
• updateEncoding: returns the same type but with the new encoding. E.g., the encoding in the bounded type is updated to the new encoding in the Flow::DispatchTensorType implementation.

The revision implements the interface methods for
Flow::DispatchTensorType and uses them to update the bindings.

Codegen already looks at flow.dispatch.tensor.load/store and will have the incoming layout information available. If there is a layout transferring need, codegen should be able to generate relayout ops when they materialize the encodings. E.g., the incoming layout is attached and codegen knows the target layout. It can generate relayout ops to bring the incoming layout to the target layout.

#encoding3 =  ...operand_index = 0 : index, op_type =  matmul, element_types = [f32, f32, f32],
  layouts = [#iree_cpu.cpu_encoding_layout<configuration = {
    encoding_info = {innerDimsPos = [0, 1],
                     innerTileSizes = [16, 1],
                     outerDimsPerm = [0, 1]}}>]
>
#encoding6 =  ...operand_index = 0 : index, op_type =  matmul, element_types = [f32, f32, f32],
  user_indexing_maps = [#map, #map1, #map2]
>
%7 = flow.dispatch.tensor.load %4,
  offsets = [0, 0], sizes = [%2, %0], strides = [1, 1]
  : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #encoding3>>{%2, %0}
  -> tensor<?x?xf32, #encoding6>

[hanhanW]

It depends on #19527, and it is ready for review.

[hanhanW]

cc @bjacob @Max191 Heads-up. In the future, I expect to see something like below. In my encoding specialization pass, we resolve layouts and attach them to bindings. I think here would be where the layout transfer happens. If the request layout is different (i.e., the load shape is different from the source shape), we'll need to generate relayout ops in codegen.

  operand_index = 0 : index, op_type =  matmul, element_types = [f32, f32, f32],
  layouts = [#iree_cpu.cpu_encoding_layout<configuration = {
    encoding_info = {innerDimsPos = [0, 1],
                     innerTileSizes = [16, 1],
                     outerDimsPerm = [0, 1]}}>]
>
  operand_index = 0 : index, op_type =  matmul, element_types = [f32, f32, f32],
  user_indexing_maps = [#map, #map1, #map2]
>
%7 = flow.dispatch.tensor.load %4,
  offsets = [0, 0], sizes = [%2, %0], strides = [1, 1]
  : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #encoding3>>{%2, %0}
  -> tensor<?x?xf32, #encoding6>

[hanhanW]

@benvanik this is ready for review. In the meantime, I'm working on the testing encoding attribute, so we do not need to rely on VMVX/CPU ones.

[hanhanW]

@benvanik this is ready for review. In the meantime, I'm working on the testing encoding attribute, so we do not need to rely on VMVX/CPU ones.

#19879 is up for review. I'll rebase and update the tests once it's landed.

[benvanik]

looks great, and your new test encoding attrs will clean up the tests nicely!

[hanhanW]

looks great, and your new test encoding attrs will clean up the tests nicely!

Thank you, I'm going to merge the cleanup and update the tests. Thanks for the quick review!

[hanhanW]

looks great, and your new test encoding attrs will clean up the tests nicely!

I really like the test encoding attrs. It looks beautiful to me, see 387057d :)