lock the mutex around the condvar destruction (iree-org#8703)

Unify dispatch region formation + Conversion to Flow passes (iree-org…

…#8662)

There is a weird dependence between the
ConvertToFlowBefore/AfterDispatchRegionFormationPass and the dispatch
region formation itself. Its represents a weird separation of what
cant be moved into dispatches (due to bufferization issue), what needs
to be converted to flow operations and what cannot be converted to
flow operation but still need to be in dispatches.

The only way to resolve these dependencies is to unify these passes.
It also cleans up the dispatch region formation works, allow the
formation to either work on root + fused op or moving individual ops
into the dispatch.

Since the converion to flow happens within dispatch region the
following passes are unnecessary
- `ConvertToFlowAfterDispatchFormation`
- `ConvertToFlowBeforeDisaptchFormation`.
A new pass `ConvertToFlow` is added just for testing the conversion
patterns

Switch to prefixed accessors (NFC) (iree-org#8666)

Switch to prefixed accessors preparing for upstream switch.

[CUDA] Add pattern to collapse vector.transfer_read to 2D when possib…

…le (iree-org#8636)

This is needed to lower these ops in the tensorcore pipeline

[CUDA] Add pattern to collapse vector.transfer_read to 2D when possib…

…le (iree-org#8636)

This is needed to lower these ops in the tensorcore pipeline

[CUDA] Add pattern to collapse vector.transfer_read to 2D when possib…

…le (iree-org#8636)

This is needed to lower these ops in the tensorcore pipeline

[NFC] Add StrategyTilingLevel enum class to clear tiling level meanin…

…g. (iree-org#8633)

The codegen used to use TilingLevel::L1Tiles and
TilingLevel::VectorTiles. This is confusing in strategy approach because
they are not targeting tiling sizes for L1 and vector. They are actually
the tiling sizes for parallel dims and reduction dims. Correct the
variable name to make confusion less.

Connect transform dialect passes into IREE. (iree-org#8602)

* Add a hook to invoke passes from Sandbox.

This commit adds a flag `-iree-codegen-use-sandbox-passes` that can be
used to transformations from sandbox within IREE in an e2e manner.
For now the pipeline just
- Sets the number of workgroups to {1, 1, 1} for sequential execution.
- Bufferizes the operations and lowers them to loops and LLVM IR.

* Connect transform dialect passes into IREE.

This revision provides a first connection between IREE and the transform dialect by
allowing the orthogonal specification of a linalg_transform file (policy) that is
parsed and applied on the fly.

Co-authored-by: Mahesh Ravishankar <[email protected]>

Fix Lambda layers test and bump Swiftshader (iree-org#8605)

Changes to swiftshader head caused an accuracy regression. This is
becase x**2 is numerically computationally complex. Changed the
computation to be x*x which is computationally identical while not
depending on an exponential operator.

Bumped swiftshader version and rebuilt docker images.

Return error instead of asserting in bufferization copy callback. (ir…

…ee-org#8528)

Current callback just asserts that the copy callback can insert a
valid copy. Return a nullptr, and throw an error when copy cannot be
generated. Also update all call sites to handle the error.