Prior to this change, the "ExtractSliceFromReshape" pattern would transform
```
%collapsed = tensor.collapse_shape %input [[0, 1], [2]]
: tensor<1x11x100xf32> into tensor<11x100xf32>
%slice = tensor.extract_slice %collapsed [%offt, 0] [%size, 100] [1, 1]
: tensor<11x100xf32> to tensor<?x100xf32>
```
into a loop that iterated over the range `%size - %offt`, that pieces
together multiple sub-slices of `%input` along the first dimension. This
is correct but obviously inefficient. The technical condition is that
collapsing at-most-one non-unit dimension of `%src` will not result in a
subsequent slice along the corresponding dimension of `%collapsed`
mapping across discontinuities in the index space of `%src`. Thus, the
definition of a "linearized dimension" (from the perspective of
`tensor.collapse_shape`) is updated to reflect this condition.
The transform will now generate
```
%slice = tensor.extract_slice %input [0, %offt, 0][1, %size, 100] [1, 1]
: tensor<1x11x100xf32> to tensor<1x?x100xf32>
%result = tensor.collapse_shape [[0, 1], [2]]
: tensor<1x?x100xf32> to tensor<?x100xf32>
```
which can be further canonicalized.
Additional tests are added to check this family of edge cases.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D135726
Inserting a tensor into an equivalent tensor is a no-op after bufferization. No alloc is needed.
Differential Revision: https://reviews.llvm.org/D132662
tensor.empty/linalg.init_tensor produces an uninititalized tensor that can be used as a destination operand for destination-style ops (ops that implement `DestinationStyleOpInterface`).
This change makes it possible to implement `TilingInterface` for non-destination-style ops without depending on the Linalg dialect.
RFC: https://discourse.llvm.org/t/rfc-add-tensor-from-shape-operation/65101
Differential Revision: https://reviews.llvm.org/D135129
This relands commit 5d4603a02d0c3e0106b10d245322b1d2072c0c3d.
It cludes fixes to GCC test failures and simplification to
the implementation.
Co-authored-by: Mahesh Ravishankar <ravishankarm@google.com>
Co-authored-by: Christopher Bate <cbate@nvidia.com>
This reverts commit 5d4603a02d0c3e0106b10d245322b1d2072c0c3d.
The Dialect/Tensor/fold-consecutive-insert-extract-slice.mlir test is
failing when built with GCC
This commit adds utility functions to perform general merging of
OffsetSizeAndStrideOpInterface by supporting producer rank
reducing and non-unit strides.
With it we can extend MergeConsecutiveExtractSlice to support
more cases.
Co-authored-by: Mahesh Ravishankar <ravishankarm@google.com>
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D134294
Consecutive tensor.insert_slice/tensor.extract_slice can be
created for the case like tiling convolution and then downsizing
2-D convolutions into 1-D ones. It hinders further transformations.
So adding these patterns to clean it up.
Given that bufferization is sensitive and have requirements over
the IR structure (see https://reviews.llvm.org/D132666),
these patterns are put in Transforms/ with separate entry points
for explicit collection.
Reviewed By: ThomasRaoux, mravishankar
Differential Revision: https://reviews.llvm.org/D133871
The transformation would fail if none of the sliced dimensions were
linearized by the producing `tensor.collapse_shape`. This is a trivial
edge case but it wasn't correctly tested. Fixes the issue and adds a test.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D134088
Currently, there's an optimization that claims dimensions of size 1 are always
contiguous. This is not necessarily the case for subviews.
```
Input:
[
[
[0, 1],
[2, 3]
],
[
[4, 5]
[6, 7]
]
]
Subview:
[
[
[0, 1],
],
[
[4, 5]
]
]
```
The old logic treats this subview as contiguous, when it is not.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D134026
Bufferization already makes the assumption that buffers pass function
boundaries in the strided form and uses the corresponding affine map layouts.
Switch it to use the recently introduced strided layout instead to avoid
unnecessary casts when bufferizing further operations to the memref dialect
counterparts that now largely rely on the strided layout attribute.
Depends On D133947
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D133951
The three following ops in the memref dialect: transpose, expand_shape,
collapse_shape, have been originally designed to operate on memrefs with
strided layouts but had to go through the affine map representation as the type
did not support anything else. Make these ops produce memref values with
StridedLayoutAttr instead now that it is available.
Depends On D133938
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D133947
Memref subview operation has been initially designed to work on memrefs with
strided layouts only and has never supported anything else. Port it to use the
recently added StridedLayoutAttr instead of extracting the strided from
implicitly from affine maps.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D133938
This change adds a set of utilities to replace the result of a
`tensor.collapse_shape -> tensor.extract_slice` chain with the
equivalent result formed by aggregating slices of the
`tensor.collapse_shape` source. In general, it is not possible to
commute `extract_slice` and `collapse_shape` if linearized dimensions
are sliced. The i-th dimension of the `tensor.collapse_shape`
result is a "linearized sliced dimension" if:
1) Reassociation indices of tensor.collapse_shape in the i'th position
is greater than size 1 (multiple dimensions of the input are collapsed)
2) The i-th dimension is sliced by `tensor.extract_slice`.
We can work around this by stitching together the result of
`tensor.extract_slice` by iterating over any linearized sliced dimensions.
This is equivalent to "tiling" the linearized-and-sliced dimensions of
the `tensor.collapse_shape` operation in order to manifest the result
tile (the result of the `tensor.extract_slice`). The user of the
utilities must provide the mechanism to create the tiling (e.g. a loop).
In the tests, it is demonstrated how to apply the utilities using either
`scf.for` or `scf.foreach_thread`.
The below example illustrates the pattern using `scf.for`:
```
%0 = linalg.generic ... -> tensor<3x7x11x10xf32>
%1 = tensor.collapse_shape %0 [[0, 1, 2], [3]] : ... to tensor<341x10xf32>
%2 = tensor.extract_slice %1 [13, 0] [10, 10] [2, 1] : .... tensor<10x10xf32>
```
We can construct %2 by generating the following IR:
```
%dest = linalg.init_tensor() : tensor<10x10xf32>
%2 = scf.for %iv = %c0 to %c10 step %c1 iter_args(%arg0) -> tensor<10x10xf32> {
// Step 1: Map this output idx (%iv) to a multi-index for the input (%3):
%linear_index = affine.apply affine_map<(d0)[]->(d0*2 + 11)>(%iv)
%3:3 = arith.delinearize_index %iv into (3, 7, 11)
// Step 2: Extract the slice from the input
%4 = tensor.extract_slice %0 [%3#0, %3#1, %3#2, 0] [1, 1, 1, 10] [1, 1, 1, 1] :
tensor<3x7x11x10xf32> to tensor<1x1x1x10xf32>
%5 = tensor.collapse_shape %4 [[0, 1, 2], [3]] :
tensor<1x1x1x10xf32> into tensor<1x10xf32>
// Step 3: Insert the slice into the destination
%6 = tensor.insert_slice %5 into %arg0 [%iv, 0] [1, 10] [1, 1] :
tensor<1x10xf32> into tensor<10x10xf32>
scf.yield %6 : tensor<10x10xf32>
}
```
The pattern was discussed in the RFC here: https://discourse.llvm.org/t/rfc-tensor-extracting-slices-from-tensor-collapse-shape/64034
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D129699
This reverts commit 5711957875738c1318f89afd7bf4be388f85a087.
A circular dependency is introduced here from Dialect/Utils/ to the
ViewLikeInterface, but it already depends on Dialect/Utils.
Also this introduces a dependency from lib/Dialect/Tensor to Linalg,
which isn't obviously correct from a layering point of view.
This change adds a set of utilities to replace the result of a
`tensor.collapse_shape -> tensor.extract_slice` chain with the
equivalent result formed by aggregating slices of the
`tensor.collapse_shape` source. In general, it is not possible to
commute `extract_slice` and `collapse_shape` if linearized dimensions
are sliced. The i-th dimension of the `tensor.collapse_shape`
result is a "linearized sliced dimension" if:
1) Reassociation indices of tensor.collapse_shape in the i'th position
is greater than size 1 (multiple dimensions of the input are collapsed)
2) The i-th dimension is sliced by `tensor.extract_slice`.
We can work around this by stitching together the result of
`tensor.extract_slice` by iterating over any linearized sliced dimensions.
This is equivalent to "tiling" the linearized-and-sliced dimensions of
the `tensor.collapse_shape` operation in order to manifest the result
tile (the result of the `tensor.extract_slice`). The user of the
utilities must provide the mechanism to create the tiling (e.g. a loop).
In the tests, it is demonstrated how to apply the utilities using either
`scf.for` or `scf.foreach_thread`.
The below example illustrates the pattern using `scf.for`:
```
%0 = linalg.generic ... -> tensor<3x7x11x10xf32>
%1 = tensor.collapse_shape %0 [[0, 1, 2], [3]] : ... to tensor<341x10xf32>
%2 = tensor.extract_slice %1 [13, 0] [10, 10] [2, 1] : .... tensor<10x10xf32>
```
We can construct %2 by generating the following IR:
```
%dest = linalg.init_tensor() : tensor<10x10xf32>
%2 = scf.for %iv = %c0 to %c10 step %c1 iter_args(%arg0) -> tensor<10x10xf32> {
// Step 1: Map this output idx (%iv) to a multi-index for the input (%3):
%linear_index = affine.apply affine_map<(d0)[]->(d0*2 + 11)>(%iv)
%3:3 = arith.delinearize_index %iv into (3, 7, 11)
// Step 2: Extract the slice from the input
%4 = tensor.extract_slice %0 [%3#0, %3#1, %3#2, 0] [1, 1, 1, 10] [1, 1, 1, 1] :
tensor<3x7x11x10xf32> to tensor<1x1x1x10xf32>
%5 = tensor.collapse_shape %4 [[0, 1, 2], [3]] :
tensor<1x1x1x10xf32> into tensor<1x10xf32>
// Step 3: Insert the slice into the destination
%6 = tensor.insert_slice %5 into %arg0 [%iv, 0] [1, 10] [1, 1] :
tensor<1x10xf32> into tensor<10x10xf32>
scf.yield %6 : tensor<10x10xf32>
}
```
The pattern was discussed in the RFC here: https://discourse.llvm.org/t/rfc-tensor-extracting-slices-from-tensor-collapse-shape/64034
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D129699
This change refines the semantics of scf.foreach_thread. Tensors that are inserted into in the terminator must now be passed to the region explicitly via `shared_outs`. Inside of the body of the op, those tensors are then accessed via block arguments.
The body of a scf.foreach_thread is now treated as a repetitive region. I.e., op dominance can no longer be used in conflict detection when using a value that is defined outside of the body. Such uses may now be considered as conflicts (if there is at least one read and one write in the body), effectively privatizing the tensor. Shared outputs are not privatized when they are used via their corresponding block arguments.
As part of this change, it was also necessary to update the "tiling to scf.foreach_thread", such that the generated tensor.extract_slice ops use the scf.foreach_thread's block arguments. This is implemented by cloning the TilingInterface op inside the scf.foreach_thread, rewriting all of its outputs with block arguments and then calling the tiling implementation. Afterwards, the cloned op is deleted again.
Differential Revision: https://reviews.llvm.org/D133114
parallel_insert_slice doesn't return a value therefore we shouldn't try
to fold the result. The insert folding don't apply to this op.
The current folding would cause pattern rewrite to not be able to
converge.
Differential Revision: https://reviews.llvm.org/D132668
tensor.pad is lowered to tensor.generate + tensor.insert_slice during bufferization. For best performance with constant padding values, users should vectorize the IR before bufferizing it.
This change also relaxes tje restriction that no new ops that bufferize to a memory write should be added during bufferization. Since bufferization has been split into two steps a while ago (tensor copy insertion + bufferization), it is reasonable to allow this now.
Differential Revision: https://reviews.llvm.org/D132355
This commit folds a `tensor.cast` op into a `tensor.collapse_shape` op
when following two conditions meet:
1. the `tensor.collapse_shape` op consumes result of the `tensor.cast` op.
2. `tensor.cast` op casts to a more dynamic version of the source tensor.
This is added as a canonicalization pattern in `tensor.collapse_shape` op.
Signed-Off-By: Gaurav Shukla <gaurav@nod-labs.com>
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D130650
The result shape of a rank-reducing subview cannot be inferred in the general case. Just the result rank is not enough. The only thing that we can infer is the layout map.
This change also improves the bufferization patterns of tensor.extract_slice and tensor.insert_slice to fully support rank-reducing operations.
Differential Revision: https://reviews.llvm.org/D129144
This revision updates the op semantics to also allow rank-reducing behavior as well
as updates the implementation to reuse code between the sequential and the parallel
version of the op.
Depends on D128920
Differential Revision: https://reviews.llvm.org/D128985
This is moslty NFC and will allow tensor.parallel_insert_slice to gain
rank-reducing semantics by reusing the vast majority of the tensor.insert_slice impl.
Depends on D128857
Differential Revision: https://reviews.llvm.org/D128920
All bufferizable ops that bufferize to an allocation receive a `bufferization.escape` attribute during TensorCopyInsertion.
Differential Revision: https://reviews.llvm.org/D128137
This change changes the bufferization so that it utilizes the new TensorCopyInsertion pass. One-Shot Bufferize no longer calls the One-Shot Analysis. Instead, it relies on the TensorCopyInsertion pass to make the entire IR fully inplacable. The `bufferize` implementations of all ops are simplified; they no longer have to account for out-of-place bufferization decisions. These were already materialized in the IR in the form of `bufferization.alloc_tensor` ops during the TensorCopyInsertion pass.
Differential Revision: https://reviews.llvm.org/D127652
It is sometimes better to make a copy of the OpResult instead of making a copy of the OpOperand. E.g., when bufferizing tensor.extract_slice.
This implementation will eventually make parts of extract_slice's `bufferize` implementation obsolete (and simplify it). It will only need to handle in-place OpOperands.
Differential Revision: https://reviews.llvm.org/D126819
This simplifies the bufferization itself and is in preparation of connecting with the sparse compiler.
Differential Revision: https://reviews.llvm.org/D126814
No longer pass static dim sizes as an attribute. This was redundant and required extra checks in the verifier. This change also makes the op symmetrical to memref::AllocOp.
Differential Revision: https://reviews.llvm.org/D126178
This change adds a new op `alloc_tensor` to the bufferization dialect. During bufferization, this op is always lowered to a buffer allocation (unless it is "eliminated" by a pre-processing pass). It is useful to have such an op in tensor land, because it allows users to model tensor SSA use-def chains (which drive bufferization decisions) and because tensor SSA use-def chains can be analyzed by One-Shot Bufferize, while memref values cannot.
This change also replaces all uses of linalg.init_tensor in bufferization-related code with bufferization.alloc_tensor.
linalg.init_tensor and bufferization.alloc_tensor are similar, but the purpose of the former one is just to carry a shape. It does not indicate a memory allocation.
linalg.init_tensor is not suitable for modelling SSA use-def chains for bufferization purposes, because linalg.init_tensor is marked as not having side effects (in contrast to alloc_tensor). As such, it is legal to move linalg.init_tensor ops around/CSE them/etc. This is not desirable for alloc_tensor; it represents an explicit buffer allocation while still in tensor land and such allocations should not suddenly disappear or get moved around when running the canonicalizer/CSE/etc.
BEGIN_PUBLIC
No public commit message needed for presubmit.
END_PUBLIC
Differential Revision: https://reviews.llvm.org/D126003
This changes replaces the `fully-dynamic-layout-maps` options (which was badly named) with two new options:
* `unknown-type-conversion` controls the layout maps on buffer types for which no layout map can be inferred.
* `function-boundary-type-conversion` controls the layout maps on buffer types inside of function signatures.
Differential Revision: https://reviews.llvm.org/D125615
This patch augments the `tensor-bufferize` pass by adding a conversion
rule to translate ReshapeOp from the `tensor` dialect to the `memref`
dialect, in addition to adding a unit test to validate the translation.
Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D125031
Run `one-shot-bufferize` instead of `linalg-comprehensive-module-bufferize` and move some test cases to their respective dialects.
Differential Revision: https://reviews.llvm.org/D124323
This change fixes `CollapsedLayoutMap` for cases where the collapsed
dims are size 1. The cases where inner most dims are size 1 and
noncontiguous can be represented by the strided form and therefore can
be allowed. For such cases, the new stride should be of the next entry
in an association whose dimension is not size 1. If the next entry is
dynamic, it's not possible to decide which stride to use at compilation
time and the stride is set to dynamic.
Differential Revision: https://reviews.llvm.org/D124137
The bufferization driver was previously using a GreedyPatternRewriter. This was problematic because bufferization must traverse ops top-to-bottom. The GreedyPatternRewriter was previously configured via `useTopDownTraversal`, but this was a hack; this API was just meant for performance improvements and should not affect the result of the rewrite.
BEGIN_PUBLIC
No public commit message needed.
END_PUBLIC
Differential Revision: https://reviews.llvm.org/D123618
Insert a buffer copy unless the dims are guaranteed to be collapsible. In the verifier, accept collapses unless they are guaranteed to be non-collapsible.
Differential Revision: https://reviews.llvm.org/D123316
Reland Note: Adds a fix to properly mark a commutative operation as folded if we change the order
of its operands. This was uncovered by the fact that we no longer re-process constants.
This avoids accidentally reversing the order of constants during successive
application, e.g. when running the canonicalizer. This helps reduce the number
of iterations, and also avoids unnecessary changes to input IR.
Fixes#51892
Differential Revision: https://reviews.llvm.org/D122692
https://reviews.llvm.org/D122641 introduced fixes to the ExpandShapeOp verifier
but also introduced an artificial layout limitation that prevents the consideration of transposed layouts.
This revision fixes the omissions and reimplements the logic using saturated arithmetic which is more
idiomatic and avoids leaking internal implementation details.
Tests cases are added for transposed layouts.
Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D122845
