You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
**Context:**
We update the llvm version tagged by jax 0.6.0:
```
mhlo=617a9361d186199480c080c9e8c474a5e30c22d1
llvm=179d30f8c3fddd3c85056fd2b8e877a4a8513158
```
We also update Enzyme to the latest version, which is 0.0.180, at commit
`db0181320d6e425ee963bd496ed0d8dbb615be18`
**Description of the Change:**
Firstly, jax recently moved from the Google github organization to its
own jax-ml organization.
This means the urls, and the retrieval method for the underlying llvm
and mhlo git commit tags, needs to be updated. (Thanks @mehrdad2m !)
Now on to the actual changes. I will list the changes in increasing
complexity.
1. The new enzyme cmake target is `EnzymeStatic-21` (from 20)
2. Enzyme works with a later llvm then our target, so it has some llvm
intrinsics unknown to the one we are targeting. We patch them away. They
do not concern us since they are all intrinsics for nvidia backends.
3. `applyPatternsAndFoldGreedily` is removed. Drop-in replacement is
`applyPatternsGreedily`.
llvm/llvm-project#104649,
llvm/llvm-project#126701
4. ops with `CallOpInterface` must have two new optional attributes
`arg_attrs` and `res_attrs`
llvm/llvm-project#123176
5. `CallInterfaceCallable` objects now must be directly casted to the
callee `SymbolRefAttr`, i.e. `callee.get<SymbolRefAttr>()` ->
`cast<SymbolRefAttr>(callee)`
llvm/llvm-project@35e8989
6. The `lookupOrCreateFn` family of functions now return
`FailureOr<funcop>` instead of just `funcop`, so a `.value()` needs to
be used to retrieve the underlying `funcop`.
llvm/llvm-project@e84f6b6
7. The cpp api for `OneShotBufferizePassOptions` no longer needs
complicated lambdas for the type converter options. They can be set with
the `mlir::bufferization::LayoutMapOption::IdentityLayoutMap` options
directly.
8. The individual `match` and `rewrite` methods in pattern rewrites are
removed. Use the two-in-one `matchAndRewrite` instead.
llvm/llvm-project#129861
9. For rewrite patterns with 1-to-N convertions, a new `macthAndRewrite`
overload with `OneToNOpAdaptor` must be used. For us, this is only the
`catalyst.list*` ops. llvm/llvm-project#116470
10. The lowering of `cf::AssertOp` to llvm was split from the
overall`--covert-cf-to-llvm` pass. We need to manually call this
separate pattern for cf.assert duriing quantum to llvm dialect lowering,
where we also convert cf to llvm.
https://github.com/llvm/llvm-project/pull/120431/files
11. The new mhlo depends on a
[shardy](https://github.com/openxla/shardy) dialect. Shardy is built
with bazel, not cmake. Building shardy ourselves would be very difficult
(not having bazel in our build ecosystem is a hard constraint, cc @mlxd
), and also not necessary (we just use mhlo for their "standard"
passes). We thus patch out all shardy components.
12. Three necessary passes were removed in mhlo:
`mhlo-legalize-control-flow`, `mhlo-legalize-to-std`,
`hlo-legalize-sort`
tensorflow/mlir-hlo@4a640be#diff-ef0d7e30da19a396ba036405a9ef636f8b1be194618b0a90f4602671fc2ef34dtensorflow/mlir-hlo@2a5e267#diff-f8c7cb07b43593403e00e0dbf9983f0186b4eb70368cc99af3b924061f1ea46f
- Alongside the removal of `mhlo-legalize-to-std`, the cmake target
`MhloToStandard` was removed too.
We simply patch them back for now.
**For the above two points, note that there will be an overall migration
to the stablehlo repo, as mhlo is sunseting.
Therefore, spending too much time on this isn't necessary, so we just
patch.**
13. The new pattern applicator (`applyPatternsGreedily`) is more
aggressive in dead code elimination,
and is eliminating dead `Value`s in the adjoint gradient method.
The `nodealloc` function we generate for adjoint gradient lowering used
to only return the qreg, not the expval result. This causes the expval
op to be eliminated since it has no users.
This further causes wrong gradient results, since the entire program,
all ops included (regardless
of dead or not), impacts the gradient through chain rule.
To avoid this, we return the expval result as well.
In doing this, we implicitly assume that differentiated qnodes can only
return expval.
Although this assumption is true and also restricted by frontend,
ideally we should not have it hard coded.
We leave this as a TODO for a future feature.
14. The old `--buffer-deallocation` pass is removed. Intended
replacement is `--buffer-deallocation-pipeline`.
This migration is very complicated. We simply add back the old buffer
deallocation pass in the catalyst dialect as a util for now.
We will revisit this in #1778 .
mlir lit test updates:
1. `bufferization.to_tensor/memref` updated assembly format
2. gradient adjoint lowering test returns both qreg and expval
3. Some inverse unrealized conversion cast pairs are canceled by the new
pattern rewriter.
4. `llvm.mlir.undef` is deprecated, use `llvm.mlir.poison` instead.
llvm/llvm-project#125629
**Benefits:**
Up to date with upstream versions.
[sc-92017]
---------
Co-authored-by: Tzung-Han Juang <[email protected]>
Co-authored-by: Ritu Thombre <[email protected]>
Co-authored-by: Mehrdad Malekmohammadi <[email protected]>
Co-authored-by: Mehrdad Malek <[email protected]>
Co-authored-by: David Ittah <[email protected]>
Co-authored-by: Joey Carter <[email protected]>
0 commit comments