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
Add a folder for maps whose body returns only one of the arguments. When this arises the fold replaces the map output with one of the operand tensors.
PiperOrigin-RevId: 369304322
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
Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/47315
Lowering of `concatenateOp` is added from lmhlo to Affine. The lowering
has been added as a part of `lhlo-legalize-to-affine` pass.
Signed-off-by: Prashant Kumar <prashantk@polymagelabs.com>
Copybara import of the project:
--
15314e4579f7a6901cf3475eff25962a34772eaf by Prashant Kumar <prashantk@polymagelabs.com>:
[MLIR] Add concatenateOp lowering from lmhlo to Affine.
Lowering of `concatenateOp` is added from lmhlo to Affine. The lowering
has been added as a part of `lhlo-legalize-to-affine` pass.
Signed-off-by: Prashant Kumar <prashantk@polymagelabs.com>
PiperOrigin-RevId: 368465992
We now use the same special cases for all ops with arity >= 2.
For binary ops, we now have only one special case if at least one of the
operands has exactly one element. In that case, we reshape both operands to
rank 1. Before, we had separate special cases whether the left-hand side
or the right-hand side have a scalar shape.
PiperOrigin-RevId: 366005835
When an op is moved out of an assuming region we already know statically that it
is independent of the assuming region. Hence, there is no need to yield its
results.
PiperOrigin-RevId: 366001405
Add pattern to move operations out of assuming op. This only valid for
constraint-independent ops, like `cstr_broadcastable` and `shape_of`. It will
eventually allow to make assuming regions' constraints independent from each
other so that they can be merged.
PiperOrigin-RevId: 365993145
This matches the behavior of mhlo.case. Additionally, fix the verification of CaseOp in the case of nested ops with mhlo.return-containing regions.
PiperOrigin-RevId: 365936672
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
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
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
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
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
- 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
For now, the pass only reifies the required shape computations. Moving
broadcasts will follow to allow for fusion across them.
PiperOrigin-RevId: 362033715
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
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
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
The linalg named ops are now type polymorphic, so the type-monomorphic
varieties are redundant (and will be deleted soon).
PiperOrigin-RevId: 360509010
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
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
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
- 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
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
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
A shape that contains exactly one element is effectively a scalar. This leads
to a speedup in cases where we have a binary op with one operand that is
effectively a scalar, because we can use the fast path.
PiperOrigin-RevId: 357515552
Also remove BitcastOp. XLA bitcast requires the input buffer to alias the output buffer, which makes bitcast always a no-op.
PiperOrigin-RevId: 356884383
This is being done by just removing the approximation and lowering to atan2 lib calls later to make the implementation the same as XLA. Note that if the approximation is brought back later, it can be fixed by changing the IR checking `less-than(X, 0)` to `less-than(copysign(X, 1), 0)`
PiperOrigin-RevId: 356253941
- Use a common base class to for AllReduce, AllGather, and AllToAll in the ODS spec.
- Add basic verification for replica groups attribute.
PiperOrigin-RevId: 354969654
In IREE, we use indexed generic op to handle the initial value. However, we
lower it to a generic op that carries an init_tensor here, and leave the handle
of initialization problem to later passes.
PiperOrigin-RevId: 354294807
If mhlo.reshape is not purely collapsing some consecutive operand
dimensions into result dimensions, we will generate two linalg
reshape op for it: the first one collapses all operand dimensions
into one dimension, and the second one expands it to all result
dimensions. For this case, the number of collapsed/expanded dimensions
should be coming strictly from the operand/result. It is different
from the case where we can generate one linalg reshape. For that case,
the reassociation map should have rank equal to the largest among
operand/result shape.
PiperOrigin-RevId: 354293826
This is the right behavior for TF and JAX and matches what TF does on GPU. It
doesn't match TF on CPU, but that's really a TF bug.
PiperOrigin-RevId: 353657779
This is the right behavior for TF and JAX and matches what TF does on GPU. It
doesn't match TF on CPU, but that's really a TF bug.
PiperOrigin-RevId: 353628258
This is the right behavior for TF and JAX and matches what TF does on GPU. It
doesn't match TF on CPU, but that's really a TF bug.
PiperOrigin-RevId: 353624935
We prototyped the lowering from mhlo.dot to linalg.matmul in IREE. Since Linalg
now supports matmul in tensors world, we can move the lowering logic to tensors
world, and upstream to legalize_to_linalg.cc. The patch lowers the mhlo.dot to
the linalg.matmul/matvec/dot in tensors world.
PiperOrigin-RevId: 351184911
This updates the tests to no longer rely on tensor_store. Once all users of this behavior have adopted, the tensor_store support will be removed.
PiperOrigin-RevId: 348624899
For floating point operations, this uses std.pow.
For integer operations, this lowers to a loop.
This adds a dependency on scf.
PiperOrigin-RevId: 348537232
- MLIR MemRefs do not preserve layout information correctly when unit dimensions
are involved. Operations like convolution that use cuDNN however need the correct
layout to be preserved so that we do not end up creating an incompatible combination
of input/filter/output layout that is not supported by cuDNN.
- Add these layouts to convolution attributes in the form of I32ArrayAttr for representing
the layout in "minor_to_major" form similar to XLA.
PiperOrigin-RevId: 348034757
It didn't return 0 for 0.0 and -0.0.
Currently we emit -0.0 for -0.0 which is correct according to the HLO dialect.
For the TF_SignOp we should emit 0.0 in that case, we will leave that as a TODO.
Enable the tests which work now, and add another one for Int64.
Also improve the registration code, we should not register the Int32 kernel.
PiperOrigin-RevId: 347981124
It didn't return 0 for 0.0 and -0.0.
Currently we emit -0.0 for -0.0 which is correct according to the HLO dialect.
For the TF_SignOp we should emit 0.0 in that case, we will leave that as a TODO.
Enable the tests which work now, and add another one for Int64.
Also improve the registration code, we should not register the Int32 kernel.
PiperOrigin-RevId: 347602378
It didn't return 0 for 0.0 and -0.0.
Currently we emit -0.0 for -0.0 which is correct according to the HLO dialect.
For the TF_SignOp we should emit 0.0 in that case, we will leave that as a TODO.
Enable the tests which work now, and add another one for Int64.
Also improve the registration code, we should not register the Int32 kernel.
PiperOrigin-RevId: 347590340
- Add this attribute to match the corresponding XLA HLO attribute on convolution
operations.
- A true value indicates a reversal of the corresponding kernel spatial dimension.
- Since XLA builder does not support this attribute, use a custom HLO converted to map
from mlir::mhlo::ConvOp to XLA.
PiperOrigin-RevId: 346891737
Two tensors having the same SSA-value isn't sufficient for equality for floating types, as `NaN != NaN`. As written this causes `tf.IsNan` to [miscompile](https://github.com/google/iree/issues/4061).
PiperOrigin-RevId: 345730640
- Add a variant of the fused convolution that does not need a side input and side input scale.
- Rename the existing one to `ConvForwardFusedSideInputOp`.
- Update tests to exercise all variants of the convolution ops in the GPU dialect.
- Eliminate unused `LHLO_ExtentBuffer` and changed LHLO_Buffer to allow any integer element
type to match what XLA can generate sometimes for scratch buffers.
PiperOrigin-RevId: 345701569
A. Unique TensorFlower
Hanhan Wang
Benjamin Kramer
Adrian Kuegel
Prashant Kumar
Jacques Pienaar
Alexander Belyaev
Rahul Joshi
Geoffrey Martin-Noble
Stella Laurenzo
Tim Shen
Stephan Herhut
Richard Uhler
Tres Popp
Mahesh Ravishankar
Lei Zhang
Smit Hinsu
Phoenix Meadowlark