Bump torch-mlir to bff2a99fd596766a8d85f6ebb9ef80044dcc6b81 (#20632)

iree-org/torch-mlir@bff2a99

Signed-off-by: jinchen62 <[email protected]>

[Preprocessing] Use value bounds for dim multiple checking (#20583)

This also adds support for udiv to util.assume value bounds constraints.

In the future we can add more constraints to represent udiv, but for now this
implements at least one working option.

[GPU] Use upstream barrier elimination pass (#20534)

Reduce register pressure - and, for some platforms, potentially improve
performance through removing actual barriers - by using the upstream
buffer elimination pass.

Fixes #20029

---------

Co-authored-by: Jakub Kuderski <[email protected]>

[Codegen] Use GPUPadLayoutAttr to resolve layouts. (#20565)

Instead of using the nop layout resolver, the revision explicitly uses
the GPUPadLayoutAttr to resolve all the layouts. If the resolver is
present in an executable target, it is prioritized.

The revision exposes a "hidden" dependency between padding layout
resolver and nop layout resolver. Because the current type converter is
not decoupled enough from data-tiling usage.

It is a step towards #20160

---------

Signed-off-by: hanhanW <[email protected]>

[Dispatch Creation] Disable broken reshape fusion by collapsing (#20576)

This reverts "[Dispatch Creation] Enable more reshape movement (#20320)"
until the upstream pattern can be fixed. This is a short term fix for
the invalid IR generated in
#20548. The entire commit is
reverted to prevent regressions.

Signed-off-by: Ian Wood <[email protected]>

[Dispatch Creation] Disable broken reshape fusion by collapsing (#20576)

This reverts "[Dispatch Creation] Enable more reshape movement (#20320)"
until the upstream pattern can be fixed. This is a short term fix for
the invalid IR generated in
#20548. The entire commit is
reverted to prevent regressions.

Signed-off-by: Ian Wood <[email protected]>

[Dispatch Creation] Disable broken reshape fusion by collapsing (#20576)

This reverts "[Dispatch Creation] Enable more reshape movement (#20320)"
until the upstream pattern can be fixed. This is a short term fix for
the invalid IR generated in
#20548. The entire commit is
reverted to prevent regressions.

Signed-off-by: Ian Wood <[email protected]>

Integrate llvm/llvm-project@52e10e6 (#20572)

This patch carries revert of
llvm/llvm-project#121389. Torch-MLIR needs to be
updated with that PR's change for us to be able to integrate.

Signed-off-by: nithinsubbiah <[email protected]>

[GPU] Extend Vector distribute to multiple reductions (#20310)

Extends vector distribution in a dispatch with single reduction kernel
to mulitple reduction kernels.

Add e2e tests for gather (4/5) (#20465)

Adds e2e tests for `iree_linalg_ext.gather` and adds two interfaces
needed for codegen.

---------

Signed-off-by: Ian Wood <[email protected]>