Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/49970
1, add hlo-to-lhlo support for DynamicReshape and DynamicBroadcastInDim
2, add a flag `convert-to-lmhlo-only` to seperate following two case:
- hlo-to-lhlo only. Simply lowers all mhlo ops to their lmhlo
counterparts, do not apply any optimization (e.g. elide any
buffer copy). Buffer optimization is not easy in dynamic
shape world especially when involving control flow, thus we
leave this to another dedicated pass.
- hlo-to-lhlo-or-memref-directly. Lowers some metadata-only mhlo
ops (e.g. reshape) to memref dialect directly and Lowers others
to their lmhlo counterparts.
Copybara import of the project:
--
562bd65a368f6194405c4ae6900e3b4388a5ec03 by Wenyi Zhao <reyizero@gmail.com>:
[MLIR][DISC] bufferize DynamicReshape and DynamicBroadcastInDim
1, add hlo-to-lhlo support for DynamicReshape and DynamicBroadcastInDim
2, add a flag `convert-to-lmhlo-only` to seperate following two case:
- hlo-to-lhlo only. Simply lowers all mhlo ops to their lmhlo
counterparts, do not apply any optimization (e.g. elide any
buffer copy). Buffer optimization is not easy in dynamic
shape world especially when involving control flow, thus we
leave this to another dedicated pass.
- hlo-to-lhlo-or-memref-directly. Lowers some metadata-only mhlo
ops (e.g. reshape) to memref dialect directly and Lowers others
to their lmhlo counterparts.
PiperOrigin-RevId: 377603395
Fix usage of default constructor. Instead, always use the parameterized
constructor and make the maximum supported rank explicit.
PiperOrigin-RevId: 377037155
Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/49598
This PR implements logic for lowering memref.tensor_load ops that are
inserted during `mhlo-legalize-to-lmhlo`
Copybara import of the project:
--
80eb377af4e02182e1aecc943a41ca5d7d1c2100 by Wenyi Zhao <reyizero@gmail.com>:
[MLIR][DISC] legalize tensor_load inserted during hlo-to-lhlo conversion
This PR implements logic for lowering memref.tensor_load ops that are
inserted during `mhlo-legalize-to-lmhlo`.
--
ac452fe3dcd591211cd5c59be9189fe2f7153b41 by Wenyi Zhao <reyizero@gmail.com>:
minor fix
--
6b36017f8632a06adbc3e05a62975fa641d0260f by Wenyi Zhao <reyizero@gmail.com>:
minor refine
--
846005cc76d0033112e47825c2e9a97790b6925f by Wenyi Zhao <reyizero@gmail.com>:
minor fix
--
f6a4becaa287d5ca323b2d152a4d0ae053730fd9 by Wenyi Zhao <reyizero@gmail.com>:
fix
--
5555749f60f7fce8f57962860ef65efccf0362ba by Wenyi Zhao <reyizero@gmail.com>:
fix
--
8873b9b6d9315c1199ca9f7c133ecf377ecd2fa6 by Wenyi Zhao <reyizero@gmail.com>:
fix
PiperOrigin-RevId: 376942547
The maximum supported target rank of 5 is sufficient for all operations but
`select`. Make the maximum target rank configurable in the rank specialization.
This reduces the number of generated kernels for operations that don't require
it.
PiperOrigin-RevId: 376822496
Add the first MLIR-generated kernel that relies on an in-TF lowering. Fusion for
this kernel relies on the generalized rank specialization for operation groups.
PiperOrigin-RevId: 376805435
Take advantage of the fact that scalars are already ranked and that they are
neutral elements to broadcasting. Do not reshape scalars, do not consider them
for broadcasting, and materialize ranked operations on scalars accordingly.
PiperOrigin-RevId: 375968371
Rank specialization cases can be applied to all argument tensors of smaller
ranks than the expected maximum rank. This is crucial if all operands are
effectively scalars and the maximum reduced rank is 0.
PiperOrigin-RevId: 375712020
Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/49454
The new interface is more safe to be used during dialect conversion
(e.g. converting from tensor world to buffer world).
Copybara import of the project:
--
a6968072d59bec3c3bbaef0121d297e807c37c91 by Wenyi Zhao <reyizero@gmail.com>:
[MLIR][DISC] Upgrade to use the new `reifyReturnTypeShapes` interface.
The new interface is more safe to be used during dialect conversion
(e.g. converting from tensor world to buffer world).
--
55e7c6b7f2f99b99e226645a57e2433fae3e90ed by Wenyi Zhao <reyizero@gmail.com>:
minor fix
PiperOrigin-RevId: 375500273
This only works for updating tensors, not add/min/max computations. It requires
the index depth to be 1 because of the limitation in Linalg. We can not compare
multiple indices without packing indices.
PiperOrigin-RevId: 375137721
For rank specialization clusters that have only two operands, we can materialize
two extra cases in which either of them is a scalar. This avoids redundant index
computations in these cases.
PiperOrigin-RevId: 375037390
* The op defines this to be index, any integer, or pred (i1).
* Many TensorFlow legalizations produce integers for the shape.
PiperOrigin-RevId: 374566113
* The former is typically invariant regardless of backend.
* The latter may need to be done differently depending on capabilities of the lowering target.
PiperOrigin-RevId: 374492924
Add lowering pattern for rank specialization clusters with more than one
non-scalar operand. The lowering resembles that of the `TransformUnrankedHlo`
pass and switches cases for maximal ranks from 1 through 8.
PiperOrigin-RevId: 374377002
The pattern can be generalized to also rank specialize operations with a single
non-scalar operand. Also extract helper functions that can be reused in
following specializations.
PiperOrigin-RevId: 374198381
Also cluster operations that operate on same shape operands. These implicitly
satisfy the broadcasting semantics requirement. Also, add test cases for some
cases that appear in the current MLIR-generated kernels.
PiperOrigin-RevId: 374191950
The ReduceRegion* patterns are matching on the same ops as the PointwiseToLinalg*
patterns and on certain toolchains (MSVC) the order can be wrong. If the pointwise
runs first then it converts the op *within* the reduction before the reduction one
runs, leading to nested linalg op weirdness.
PiperOrigin-RevId: 373848269
Add a pass to cluster unranked C/HLO operations in one
`chlo.rank_specialization_cluster` op. The C/HLO operations are moved to the
body of the operation. Later passes can use this to rank-specialize all these
operations together.
PiperOrigin-RevId: 373336725
This strips away the signedness with a type converter, using unrealized
conversion casts. The rest is mostly mechanically pushing the original op down
the pipeline so lowerings can see the original types.
Signed types stay signless for now. This can be changed in the HLO bridge later.
I did a pass over all ops and added unsigned lowerings where they were missing.
There may be more.
Currently the lowering will die at a later stage because it doesn't understand
the unrealized casts.
PiperOrigin-RevId: 371077494
This uses a indexed linalg.generic, which is rather awkward standalone but
allows fusing into the output of the concatenate and avoid to ever materialize
it in memory. I think this is the only way to get that with the current linalg
stack, fusion across a concatenate would require more infrastructure.
PiperOrigin-RevId: 369677652
Assuming ops can only be merged if their witnesses will dominate the merged
assuming op. This is not the case if the second op's witness is a result of the
first.
PiperOrigin-RevId: 369192868
Wenyi Zhao
A. Unique TensorFlower
Hanhan Wang
Adrian Kuegel
Robert Suderman
Stella Laurenzo
Ben Vanik
Geoffrey Martin-Noble
Benjamin Kramer