Bubble up extract_slice above Linalg operation.
A sequence of operations
%0 = linalg.<op> ... arg0, arg1, ...
%1 = tensor.extract_slice %0 ...
can be replaced with
%0 = tensor.extract_slice %arg0
%1 = tensor.extract_slice %arg1
%2 = linalg.<op> ... %0, %1, ...
This results in the reduce computation of the linalg operation.
The implementation uses the tiling utility functions. One difference
from the tiling process is that we don't need to insert the checking
code for the out-of-bound accesses. The use of the slice itself
represents that the code writer is sure about the boundary condition.
To avoid adding the boundary condtion check code, `omitPartialTileCheck`
is introduced for the tiling utility functions.
Differential Revision: https://reviews.llvm.org/D122437
This has been on _Both for a couple of weeks. Flip usages in core with
intention to flip flag to _Prefixed in follow up. Needed to add a couple
of helper methods in AffineOps and Linalg to facilitate a pure flag flip
in follow up as some of these classes are used in templates and so
sensitive to Vector dialect changes.
Differential Revision: https://reviews.llvm.org/D122151
- Adds default implementations of `isDefinedOutsideOfLoop` and `moveOutOfLoop` since 99% of all implementations of these functions were identical
- `moveOutOfLoop` takes one operation and doesn't return anything anymore. 100% of all implementations of this function would always return `success` and uses would either respond with a pass failure or an `llvm_unreachable`.
This revision supports padding only a subset of the iteration dimensions via an additional padding-dimensions parameter. This control allows us to pad an operation in multiple steps. For example, one may want to pad only the output dimensions of a producer matmul fused into a consumer loop nest, before tiling and padding its reduction dimension.
Depends On D122309
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D122560
Pass the padding options using arrays instead of lambdas. In particular pass the padding value as string and use the argument parser to create the padding value. Arrays are a more natural choice that matches the current use cases and avoids converting arrays to lambdas.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D122309
This transformation allow to break up a reduction dimension in a
parallel and a reduction dimension. This is followed by a separate
reduction op. This allows to generate tree reduction which is beneficial
on target allowing to take advantage parallelism.
Differential Revision: https://reviews.llvm.org/D122045
Previously, only LinalgOps whose operands are defined by an ExtractSliceOp could be padded. The revision supports walking a use-def chain of LinalgOps to find an ExtractSliceOp.
Reviewed By: hanchung
Differential Revision: https://reviews.llvm.org/D122116
This revision introduces a heuristic to stop fusion for shape-only tensors. A shape-only tensor only defines the shape of the consumer computation while the data is not used. Pure producer consumer fusion thus shall not fuse the producer of a shape-only tensor. In particular, since the shape-only tensor will have other uses that actually consume the data.
The revision enables fusion for consumers that have two uses of the same tensor. One as input operand and one as shape-only output operand. In these cases, we want to fuse only the input operand and avoid output fusion via iteration argument.
Reviewed By: hanchung
Differential Revision: https://reviews.llvm.org/D120981
This provides a way to create an operation without manipulating
OperationState directly. This is useful for creating unregistered ops.
Reviewed By: rriddle, mehdi_amini
Differential Revision: https://reviews.llvm.org/D120787
ExpandShapeOp builder cannot infer the result type since it doesn't know
how the dimension needs to be split. Remove this builder so that it
doesn't get used accidently. Also remove one potential path using it in
generic fusion.
Differential Revision: https://reviews.llvm.org/D122019
The current dialect registry allows for attaching delayed interfaces, that are added to attrs/dialects/ops/etc.
when the owning dialect gets loaded. This is clunky for quite a few reasons, e.g. each interface type has a
separate tracking structure, and is also quite limiting. This commit refactors this delayed mutation of
dialect constructs into a more general DialectExtension mechanism. This mechanism is essentially a registration
callback that is invoked when a set of dialects have been loaded. This allows for attaching interfaces directly
on the loaded constructs, and also allows for loading new dependent dialects. The latter of which is
extremely useful as it will now enable dependent dialects to only apply in the contexts in which they
are necessary. For example, a dialect dependency can now be conditional on if a user actually needs the
interface that relies on it.
Differential Revision: https://reviews.llvm.org/D120367
This commit moves FuncOp out of the builtin dialect, and into the Func
dialect. This move has been planned in some capacity from the moment
we made FuncOp an operation (years ago). This commit handles the
functional aspects of the move, but various aspects are left untouched
to ease migration: func::FuncOp is re-exported into mlir to reduce
the actual API churn, the assembly format still accepts the unqualified
`func`. These temporary measures will remain for a little while to
simplify migration before being removed.
Differential Revision: https://reviews.llvm.org/D121266
New buffer allocations can now be returned/yielded from blocks with `allow-return-allocs`. One-Shot Bufferize deallocates all buffers at the end of the block. If this is not possible (because the buffer escapes the block), this is now done by the existing BufferDeallocation pass.
Differential Revision: https://reviews.llvm.org/D121527
This improves the modularity of the bufferization.
From now on, all ops that do not implement BufferizableOpInterface are considered hoisting barriers. Previously, all ops that do not implement the interface were not considered barriers and such ops had to be marked as barriers explicitly. This was unsafe because we could've hoisted across unknown ops where it was not safe to hoist.
As a side effect, this allows for cleaning up AffineBufferizableOpInterfaceImpl. This build unit no longer needed and can be deleted.
Differential Revision: https://reviews.llvm.org/D121519
The revision removes the linalg.fill operation and renames the OpDSL generated linalg.fill_tensor operation to replace it. After the change, all named structured operations are defined via OpDSL and there are no handwritten operations left.
A side-effect of the change is that the pretty printed form changes from:
```
%1 = linalg.fill(%cst, %0) : f32, tensor<?x?xf32> -> tensor<?x?xf32>
```
changes to
```
%1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<?x?xf32>) -> tensor<?x?xf32>
```
Additionally, the builder signature now takes input and output value ranges as it is the case for all other OpDSL operations:
```
rewriter.create<linalg::FillOp>(loc, val, output)
```
changes to
```
rewriter.create<linalg::FillOp>(loc, ValueRange{val}, ValueRange{output})
```
All other changes remain minimal. In particular, the canonicalization patterns are the same and the `value()`, `output()`, and `result()` methods are now implemented by the FillOpInterface.
Depends On D120726
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D120728
This patch removes an old recursive implementation to lower vector.transpose to extract/insert operations
and replaces it with a iterative approach that leverages newer linearization/delinearization utilities.
The patch should be NFC except by the order in which the extract/insert ops are generated.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D121321
Enhance `LinalgTileAndFuseTensorOpsPattern` with an additional rewrite signature that returns the result of the rewrite.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D121212
The current StandardToLLVM conversion patterns only really handle
the Func dialect. The pass itself adds patterns for Arithmetic/CFToLLVM, but
those should be/will be split out in a followup. This commit focuses solely
on being an NFC rename.
Aside from the directory change, the pattern and pass creation API have been renamed:
* populateStdToLLVMFuncOpConversionPattern -> populateFuncToLLVMFuncOpConversionPattern
* populateStdToLLVMConversionPatterns -> populateFuncToLLVMConversionPatterns
* createLowerToLLVMPass -> createConvertFuncToLLVMPass
Differential Revision: https://reviews.llvm.org/D120778
This commit deletes the old dialect conversion-based bufferization patterns, which are now obsolete.
Differential Revision: https://reviews.llvm.org/D120883
The Func has a large number of legacy dependencies carried over from the old
Standard dialect, which was pervasive and contained a large number of varied
operations. With the split of the standard dialect and its demise, a lot of lingering
dead dependencies have survived to the Func dialect. This commit removes a
large majority of then, greatly reducing the dependence surface area of the
Func dialect.
The last remaining operations in the standard dialect all revolve around
FuncOp/function related constructs. This patch simply handles the initial
renaming (which by itself is already huge), but there are a large number
of cleanups unlocked/necessary afterwards:
* Removing a bunch of unnecessary dependencies on Func
* Cleaning up the From/ToStandard conversion passes
* Preparing for the move of FuncOp to the Func dialect
See the discussion at https://discourse.llvm.org/t/standard-dialect-the-final-chapter/6061
Differential Revision: https://reviews.llvm.org/D120624
In D115022, we introduced an optimization where OpResults of a `linalg.generic` may bufferize in-place with an "in" OpOperand if the corresponding "out" OpOperand is not used in the computation.
This optimization can lead to unexpected behavior if the newly chosen OpOperand is in the same alias set as another OpOperand (that is used in the computation). In that case, the newly chosen OpOperand must bufferize out-of-place. This can be confusing to users, as always choosing the "out" OpOperand (regardless of whether it is used) would be expected when having the notion of "destination-passing style" in mind.
With this change, we go back to always bufferizing in-place with "out" OpOperands by default, but letting users override the behavior with a bufferization option.
Differential Revision: https://reviews.llvm.org/D120182
Now that sparse tensor types are first-class citizens and the sparse compiler
is taking shape, it is time to make sure other compiler optimizations compose
well with sparse tensors. Mostly, this should be completely transparent (i.e.,
dense and sparse take the same path). However, in some cases, optimizations
only make sense in the context of sparse tensors. This is a first example of
such an optimization, where fusing a sampled elt-wise multiplication only makes
sense when the resulting kernel has a potential lower asymptotic complexity due
to the sparsity.
As an extreme example, running SDDMM with 1024x1024 matrices and a sparse
sampling matrix with only two elements runs in 463.55ms in the unfused
case but just 0.032ms in the fused case, with a speedup of 14485x that
is only possible in the exciting world of sparse computations!
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D120429
Add `BufferizableOpInterface::verifyAnalysis`. Ops can implement this method to check for expected invariants and limitations.
The purpose of this change is to introduce a modular way of checking assertions such as `assertScfForAliasingProperties`.
Differential Revision: https://reviews.llvm.org/D120189
insert is soft deprecated, so remove all references so it's less likely
to be used and can be easily removed in the future.
Differential Revision: https://reviews.llvm.org/D120021
It is time to compose Linalg related optimizations with SparseTensor
related optimizations. This is a careful first start by adding some
general Linalg optimizations "upstream" of the sparse compiler in the
full sparse compiler pipeline. Some minor changes were needed to make
those optimizations aware of sparsity.
Note that after this, we will add a sparse specific fusion rule,
just to demonstrate the power of the new composition.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D119971
The pad-slice swap pattern generates `scf.if` and `tensor.generate`
to guard against zero-sized slices if it cannot prove the slice is
always non-zero. This is safe but quite conservative. It can be
unnecessary for cases where we know by problem definition such cases
does not exist, even if with dynamic shaped ops or unknown tile/slice
sizes, e.g., convolution padding size = 1 with kernel dim size = 3.
So this commit introduces a control to the pattern to specify
whether to generate the if constructs to handle such cases better,
given that once the if constructs is materialized, it's very hard
to analyze and simplify.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D117017
Fusion of `linalg.generic` with
`tensor.expand_shape/tensor.collapse_shape` currently handles fusion
with reshape by expanding the dimensionality of the `linalg.generic`
operation. This helps fuse elementwise operations better since they
are fused at the highest dimensionality while keeping all indexing
maps involved projected permutations. The intent of these is to push
the reshape to the boundaries of functions.
The presence of named ops (or other ops across which the reshape
cannot be propagated) stops the propagation to the edges of the
function. At this stage, the converse patterns that fold the reshapes
with generic ops by collapsing the dimensions of the generic op can
push the reshape towards edges. In particular it helps the case where
reshapes exist in between named ops and generic ops.
`linalg.named_op` -> `tensor.expand_shape` -> `linalg.generic`
Pushing the reshape down will help fusion of `linalg.named_op` ->
`linalg.generic` using tile + fuse transformations.
This pattern is intended to replace the following patterns
1) FoldReshapeByLinearization : These patterns create indexing maps
that are not projected permutations that affect future
transformations. They are only useful for folding unit-dimensions.
2) PushReshapeByExpansion : This pattern has the same functionality
but has some restrictions
a) It tries to avoid creating new reshapes that limits its
applicability. The pattern added here can achieve the same
functionality through use of the `controlFn` that allows clients
of the pattern freedom to make this decision.
b) It does not work for ops with indexing semantics.
These patterns will be deprecated in a future patch.
Differential Revision: https://reviews.llvm.org/D119365
They used to be classes with a virtual `run` function. This was inconvenient because post analysis steps are stored in BufferizationOptions. Because of this design choice, BufferizationOptions were not copyable.
Differential Revision: https://reviews.llvm.org/D119258
This is both more efficient and more ergonomic to use, as inverting a
bit vector is trivial while inverting a set is annoying.
Sadly this leaks into a bunch of APIs downstream, so adapt them as well.
This would be NFC, but there is an ordering dependency in MemRefOps's
computeMemRefRankReductionMask. This is now deterministic, previously it
was dependent on SmallDenseSet's unspecified iteration order.
Differential Revision: https://reviews.llvm.org/D119076
