Commit Graph

250 Commits

Author SHA1 Message Date
A. Unique TensorFlower 9ebadc4c4d Integrate LLVM at llvm/llvm-project@482283042f
Updates LLVM usage to match
[482283042f79](https://github.com/llvm/llvm-project/commit/482283042f79)

PiperOrigin-RevId: 365710568
2021-03-29 18:29:48 -07:00
A. Unique TensorFlower 85a306d356 [MLIR][MHLO] Add pattern to inline broadcasted shapes
Simplify reasoning about `cstr_broadcastable` ops in the
`mhlo-move-up-dynamic-broadcasts-for-fusion` pass.

PiperOrigin-RevId: 365560893
2021-03-29 06:32:32 -07:00
A. Unique TensorFlower fb819c1de8 [MLIR][MHLO] Apply patterns in MoveUpDynamicBroadcastsForFusionPass greedily
PiperOrigin-RevId: 365556488
2021-03-29 06:02:06 -07:00
Geoffrey Martin-Noble a2b6060c0c Add folder for HLO NotOp
PiperOrigin-RevId: 364989658
2021-03-25 02:08:38 -07:00
Adrian Kuegel a34aa699f8 Fix tanh lowering for NaN input.
If the input is NaN, the result should be NaN, too.

PiperOrigin-RevId: 364788902
2021-03-24 06:34:36 -07:00
Stella Laurenzo 7f2bf48b8b Integrate LLVM at llvm/llvm-project@b24436ac96
Updates LLVM usage to match
[b24436ac96bd](https://github.com/llvm/llvm-project/commit/b24436ac96bd)

PiperOrigin-RevId: 364615807
2021-03-23 12:20:17 -07:00
A. Unique TensorFlower 8987dfd1d6 [MLIR][HLO] Move broadcasts over n-ary shape-preserving ops
This will open up more fusion opportunities.

PiperOrigin-RevId: 364577231
2021-03-23 09:38:39 -07:00
A. Unique TensorFlower 618223778d Integrate LLVM at llvm/llvm-project@5657f93e78
Updates LLVM usage to match
[5657f93e788f](https://github.com/llvm/llvm-project/commit/5657f93e788f)

PiperOrigin-RevId: 364541987
2021-03-23 06:15:46 -07:00
A. Unique TensorFlower 54f37abc28 [MHLO] Move broadcasts over elementwise ops
Move up dynamic broadcasts and shape computations to allow for more fusion
opportunities.

PiperOrigin-RevId: 364514158
2021-03-23 02:34:41 -07:00
Benjamin Kramer 59fa7c0ef7 [MHLO:linalg] Lower all dynamic broadcasts of static shapes to linalg.generic
We only need the memref_reinterpret_cast if we don't know whether a dimension
gets expanded or not. With static shapes we know that a dimension can only be
expanded if it's a static 1, so lower it in the same way we lower fully
static broadcasts.

PiperOrigin-RevId: 363859181
2021-03-19 03:52:02 -07:00
Hanhan Wang 2e0ee7759b Add support for lowering mhlo.torch_index_select to Linalg on tensors.
The change upstreams the pattern from IREE repo to MHLO repo.

PiperOrigin-RevId: 363406294
2021-03-17 06:33:41 -07:00
A. Unique TensorFlower c54527fe88 Integrate LLVM at llvm/llvm-project@678241795c
Updates LLVM usage to match
[678241795c95](https://github.com/llvm/llvm-project/commit/678241795c95)

PiperOrigin-RevId: 363257913
2021-03-16 13:33:00 -07:00
A. Unique TensorFlower 2be112a603 [MLIR][MHLO] Approximate `tf.Tanh` as constant +/-1 for small/large values
Fix issue raised in https://github.com/tensorflow/tensorflow/issues/47724

PiperOrigin-RevId: 363210296
2021-03-16 10:14:30 -07:00
Jacques Pienaar 3de2024a9b Avoid creating tuple type only for verification
Make the error message a bit more verbose & it is cheaper to verify the elements rather than creating a (potentially) new type.

PiperOrigin-RevId: 363073909
2021-03-15 17:58:19 -07:00
Tim Shen d16860d26d [MLIR] Change LMHLO Conditional and While to capture needed buffers, instead of passing them by operands.
This is consistent with the design of LMHLO FusionOp, and it simplifies the
usage. Before the change, those redundant operands ended up unused as all sub-regions can already capture needed buffers.

PiperOrigin-RevId: 362381155
2021-03-11 14:42:41 -08:00
Hanhan Wang 4f5e1c51dd Add support for lowering NHWC pooling mhlo.reduce_window to Linalg on tensors.
The change upstreams the pattern from IREE repo to MHLO repo.

PiperOrigin-RevId: 362312573
2021-03-11 09:41:34 -08:00
Hanhan Wang 630cabefb0 Add support for lowering 2D depthwise mhlo.conv to Linalg on tensors.
The change upstreams the pattern from IREE repo to MHLO repo.

PiperOrigin-RevId: 362300550
2021-03-11 08:41:38 -08:00
Benjamin Kramer 94f9740c67 [MLIR][HLO:Linalg] Lower mhlo.dynamic_iota to indexed_generic
This is the same as iota, but instead of taking the dimensions from the result
tensor we use the supplied shape extents tensor.

PiperOrigin-RevId: 362298548
2021-03-11 08:31:29 -08:00
Benjamin Kramer 09f8046816 [MLIR:HLO:LINALG] Fix codegen for mhlo.reshape when one side is rank 0
This is an annoying edge case because the collapse->expand lowering expects at
least R1 or it will produce invalid linalg reshapes. Using the direct lowering
works fine.

PiperOrigin-RevId: 362269199
2021-03-11 05:29:56 -08:00
Benjamin Kramer d77b556822 [MLIR][MHLO] Allow recursion in the shape_of mover
This allows it to push shape_of over a chain of ops all the way to the top.

PiperOrigin-RevId: 362249009
2021-03-11 02:52:21 -08:00
Benjamin Kramer 67a770e4e0 [HLO:MLIR] Make binary op type reification emit shape_of instead of tensor ops
This gives cleaner code and allows shape optimizations to happen on the result.

PiperOrigin-RevId: 362242975
2021-03-11 02:01:35 -08:00
Rahul Joshi 9902e6ee32 [HLO] Add LMHLO CollectivePermute verification.
- Extract verification of source target pairs attached to collective permute into a common
  helper function and use that to verify both MHLO and LMHLO variants.
- Change MlirGpuTestBase::ParseMlirModule to allow returning back a failure, and use
  that to update the mlir_gpu_compile_test to check the new behavior.

PiperOrigin-RevId: 362156962
2021-03-10 15:37:12 -08:00
A. Unique TensorFlower c217a6ef61 [MHLO] Add pass to move up dynamic broadcasts for fusion
For now, the pass only reifies the required shape computations. Moving
broadcasts will follow to allow for fusion across them.

PiperOrigin-RevId: 362033715
2021-03-10 06:21:57 -08:00
Stephan Herhut cabd4d9a06 Canonicalize dynamic_broadcast_in_dim to own shape with rank narrowing on the shape to a corresponding tensor.cast.
PiperOrigin-RevId: 362028291
2021-03-10 05:43:54 -08:00
A. Unique TensorFlower 218476128e [MLIR][KernelGen] Fix zeta lowering at poles
Return nan at zeta poles or inf where the limit is defined. Also test the kernel
based on the series representation of zeta.

PiperOrigin-RevId: 361993482
2021-03-10 01:09:10 -08:00
A. Unique TensorFlower 55eda81407 [MLIR][HLO] Reify shape extents as `index` values
PiperOrigin-RevId: 361519167
2021-03-08 02:42:47 -08:00
A. Unique TensorFlower 39650a5d5a Remove rank 1 specialization from TransformUnrankedHloPass.
For binary ops, we already special-case rank 0 vs rank 1, and same shape. So we
don't need to special-case a maximum rank of 1.

PiperOrigin-RevId: 360891955
2021-03-04 05:24:53 -08:00
Adrian Kuegel 62b357b601 Remove rank 1 specialization from TransformUnrankedHloPass.
For binary ops, we already special-case rank 0 vs rank 1, and same shape. So we
don't need to special-case a maximum rank of 1.

PiperOrigin-RevId: 360881387
2021-03-04 04:04:11 -08:00
Geoffrey Martin-Noble 8687f3e4cf Lower MHLO Dot to type-polymorphic linalg named ops
The linalg named ops are now type polymorphic, so the type-monomorphic
varieties are redundant (and will be deleted soon).

PiperOrigin-RevId: 360509010
2021-03-02 14:00:58 -08:00
Adrian Kuegel 0683db3b24 Legalize MinimumBroadcastShapes op.
Use it in TransformUnrankedHloPass, which allows to reduce the maximum
rank for rank specialized broadcast from 6 to 5.

PiperOrigin-RevId: 360415743
2021-03-02 06:39:01 -08:00
Jacques Pienaar 329b1fd071 Verify compatible shapes in unpack verification rather than exact
Previously this would be too strict and fail if dynamic and static dims were
compared. Dynamic/unknown are treated as "maybe equal" to a static value without further info, so at this layer don't flag as invalid unless truly are.

PiperOrigin-RevId: 360189086
2021-03-01 08:00:16 -08:00
Benjamin Kramer e19ccf975e Filter static dimensions from dynamic_broadcast_in_dim's init_tensor
Otherwise we'd generate invalid IR for those cases.

PiperOrigin-RevId: 360144122
2021-03-01 03:03:54 -08:00
Adrian Kuegel e6a1f5f0f9 Add MinimumBroadcastShapesOp to chlo dialect.
This op is useful for rank specialization of broadcasts. Kernel Generator
needs to generate one kernel for each rank, so if we can minimize the rank
of the broadcast shape, we can support more cases with the same number of
special-cased kernels.

PiperOrigin-RevId: 360137827
2021-03-01 02:23:52 -08:00
Hanhan Wang a8f99ee0f5 Fix the shape of linalg.init_tensor in conv op lowering.
The output spatial dims are not as same as the input spatial dims. Only supports
static output spatial dims for now.

PiperOrigin-RevId: 359775479
2021-02-26 09:34:11 -08:00
Hanhan Wang 90f0d7f935 Add support for lowering mhlo.conv to Linalg on tensors.
This pattern only works for normal convolutions. It does not work for depthwise
convolutions. The Linalg conv ops are defined with static rank, so it only
supports 1d/2d/3d cases, which are the most typical cases.

This also refactors out the same check in lmhlo.conv lowering.

PiperOrigin-RevId: 359503527
2021-02-25 05:59:08 -08:00
Hanhan Wang 45a1249fe2 Add support for lowering mhlo.pad to linalg.pad_tensor
The change upstreams the pattern from IREE repo to MHLO repo.

PiperOrigin-RevId: 359481543
2021-02-25 03:00:39 -08:00
Geoffrey Martin-Noble 89f7f2bd65 Lower integer matmuls to linalg
PiperOrigin-RevId: 359306495
2021-02-24 09:45:07 -08:00
Hanhan Wang 475b4a06a5 Add support for lowering mhlo.slice to subtensor.
PiperOrigin-RevId: 359297978
2021-02-24 09:06:09 -08:00
A. Unique TensorFlower ac0552f127 [MLIR][HLO] Remove duplicate `PopulateTransformUnrankedHloPatterns`
PiperOrigin-RevId: 359046173
2021-02-23 07:50:47 -08:00
Rahul Joshi 5adb7c6e12 [MLIR:LHLO] Add optional call target arg mapping to LMHLO CustomCall operations.
- XLA:HLO -> LMHLO conversion drops all token arguments and return values, however
  custom calls that users write still expect to get buffer pointers for these token types.
- To be able to support this, add an optional call target argument mapping attribute to
  LMHLO custom calls. When this attribute is present, it indicates the number of
  arguments and returns that the custom call expects and also indicates which LMHLO
  arg() or output() maps to which arg or result number of the custom call.

PiperOrigin-RevId: 358826664
2021-02-22 08:43:00 -08:00
Benjamin Kramer a9cc1dcfa0 [mlir][hlo] Add basic rank-specialization for select
This just blows up everything to ranked (up to 6) and is probably quite slow.
This is sufficient to make kernelgen compile SelectV2.

PiperOrigin-RevId: 358777728
2021-02-22 02:41:12 -08:00
Benjamin Kramer b42def4612 [mlir][hlo] Refactor rank specialization to allow an arbitrary number of inputs
This actually simplifies the code a bit.

PiperOrigin-RevId: 358201038
2021-02-18 09:53:03 -08:00
Benjamin Kramer ca4034b56e [mlir][hlo] Make select ready for dynamic shapes (ranked only for now)
Move tf.SelectV2 broadcast lowering to a chlo.broadcast_select op, and lower it
to broadcasts on mhlo from there.

PiperOrigin-RevId: 358179975
2021-02-18 08:08:40 -08:00
Adrian Kuegel 37e31f8b26 Lower Expm1 kernel to math.ExpM1.
PiperOrigin-RevId: 358152908
2021-02-18 04:54:23 -08:00
Richard Uhler b579bd5d9e Support dynamic-shaped operand in verification of BroadcastInDim.
Verification of HLO_BroadcastInDimOp was previously failing or crashing if the
operand had a dynamic shape or was unranked. Update the verification code to
allow the operand to be unranked or have dynamic shape.

PiperOrigin-RevId: 358056793
2021-02-17 16:18:09 -08:00
A. Unique TensorFlower 220deb3709 [MLIR][CHLO] Add legalization for `chlo.polygamma` to MHLO
PiperOrigin-RevId: 357954624
2021-02-17 08:33:01 -08:00
A. Unique TensorFlower 81abaf364d [MLIR][MHLO] Add polygamma op to the CHLO dialect
PiperOrigin-RevId: 357724465
2021-02-16 08:32:33 -08:00
Adrian Kuegel b594254c79 [mhlo] Lower int->bool to a comparison with zero
This matches what TF (and C++) do in this case.

PiperOrigin-RevId: 357566262
2021-02-15 06:38:09 -08:00
Benjamin Kramer 240a44de82 [mhlo] Lower int->int cast to sign extension instead of zero extension
Signless does not mean unsigned here. Currently mhlo only has signed types.

PiperOrigin-RevId: 357561712
2021-02-15 05:58:47 -08:00
Adrian Kuegel 8672735e9a [mhlo] Lower float->bool to a comparison with zero
This matches what TF (and C++) do in this case.

PiperOrigin-RevId: 357553098
2021-02-15 04:36:36 -08:00