When any of the memrefs in a structured linalg op has a zero dimension, it becomes dead.
This is consistent with the fact that linalg ops deduce their loop bounds from their operands.
Note however that this is not the case for the `tensor<0xelt_type>` which is a special convention
that must be lowered away into either `memref<elt_type>` or just `elt_type` before this
canonicalization can kick in.
Differential Revision: https://reviews.llvm.org/D85413
A new first-party modeling for LLVM IR types in the LLVM dialect has been
developed in parallel to the existing modeling based on wrapping LLVM `Type *`
instances. It resolves the long-standing problem of modeling identified
structure types, including recursive structures, and enables future removal of
LLVMContext and related locking mechanisms from LLVMDialect.
This commit only switches the modeling by (a) renaming LLVMTypeNew to LLVMType,
(b) removing the old implementaiton of LLVMType, and (c) updating the tests. It
is intentionally minimal. Separate commits will remove the infrastructure built
for the transition and update API uses where appropriate.
Depends On D85020
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D85021
Replaced definition of named ND ConvOps with tensor comprehension
syntax which reduces boilerplate code significantly. Furthermore,
new ops to support TF convolutions added (without strides and dilations).
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D84628
This commit is part of a greater project which aims to add
full end-to-end support for convolutions inside mlir. The
reason behind having conv ops for each rank rather than
having one generic ConvOp is to enable better optimizations
for every N-D case which reflects memory layout of input/kernel
buffers better and simplifies code as well. We expect plain linalg.conv
to be progressively retired.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D83879
- replace DotOp, now that DRR rules have been dropped.
- Capture arguments mismatch in the parser. The number of parsed arguments must
equal the number of expected arguments.
Reviewed By: ftynse, nicolasvasilache
Differential Revision: https://reviews.llvm.org/D82952
linalg.indexed_generic (consumer) with tensor arguments.
The implementation of fusing std.constant producer with a
linalg.indexed_generic consumer was already in place. It is exposed
with this change. Also cleaning up some of the patterns that implement
the fusion to not be templated, thereby avoiding lot of conditional
checks for calling the right instantiation.
Differential Revision: https://reviews.llvm.org/D84566
The `makeTiledViews` did not use the sizes of the tiled views based on
the result of the loop bound inference computation. This manifested as
an error in computing tile sizes with convolution where not all the
result expression of concatenated affine maps are simple
AffineDimExpr.
Differential Revision: https://reviews.llvm.org/D84366
linalg.conv does not support memrefs with rank smaller than 3 as stated here:
https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/convolution
However it does not verify it and thus crashes with "LLVM ERROR: out of memory"
error for 1D case and "nWin > 0 && "expected at least one window dimension"" assertion
for 2D case. This commit adds check for that in the verification method.
Differential Revision: https://reviews.llvm.org/D84317
Loop bound inference is right now very limited as it supports only permutation maps and thus
it is impossible to implement convolution with linalg.generic as it requires more advanced
loop bound inference. This commits solves it for the convolution case.
Depends On D83158
Differential Revision: https://reviews.llvm.org/D83191
The underlying infrastructure supports this already, just add the
pattern matching for linalg.generic.
Differential Revision: https://reviews.llvm.org/D84335
This commit adds functionality needed for implementation of convolutions with
linalg.generic op. Since linalg.generic right now expects indexing maps to be
just permutations, offset indexing needed in convolutions is not possible.
Therefore in this commit we address the issue by adding support for symbols inside
indexing maps which enables more advanced indexing. The upcoming commit will
solve the problem of computing loop bounds from such maps.
Differential Revision: https://reviews.llvm.org/D83158
Summary:
linalg.copy + linalg.fill can be used to create a padded local buffer.
The `masked` attribute is only valid on this padded buffer.
When forwarding to vector.transfer ops, the attribute must be reset
conservatively.
Differential Revision: https://reviews.llvm.org/D83782
Improve the logic deciding if it is safe to hoist vector transfer read/write
out of the loop. Change the logic to prevent hoisting operations if there are
any unknown access to the memref in the loop no matter where the operation is.
For other transfer read/write in the loop check if we can prove that they
access disjoint memory and ignore them in this case.
Differential Revision: https://reviews.llvm.org/D83538
This revision adds support for vectorizing named and generic contraction ops to vector.contract. Cases in which the memref is 0-D are special cased to emit std.load/std.store instead of vector.transfer. Relevant tests are added.
Differential revision: https://reviews.llvm.org/D83307
While lowering min/max pooling ops to loops, generate the right kind of
load/stores (std or affine) instead of always generating std
load/stores.
Differential Revision: https://reviews.llvm.org/D83080
Summary:
This revision replaces MatmulOp, now that DRR rules have been dropped.
This revision also fixes minor parsing bugs and a plugs a few holes to get e2e paths working (e.g. library call emission).
During the replacement the i32 version had to be dropped because only the EDSC operators +, *, etc support type inference.
Deciding on a type-polymorphic behavior, and implementing it, is left for future work.
Reviewers: aartbik
Subscribers: mehdi_amini, rriddle, jpienaar, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, aartbik, liufengdb, stephenneuendorffer, Joonsoo, grosul1, frgossen, Kayjukh, jurahul, msifontes
Tags: #mlir
Differential Revision: https://reviews.llvm.org/D81935
This revision replaces MatmulOp, now that DRR rules have been dropped.
This revision also fixes minor parsing bugs and a plugs a few holes to get e2e paths working (e.g. library call emission).
During the replacement the i32 version had to be dropped because only the EDSC operators +, *, etc support type inference.
Deciding on a type-polymorphic behavior, and implementing it, is left for future work.
Differential Revision: https://reviews.llvm.org/D79762
This option avoids to accidentally reuse variable across -LABEL match,
it can be explicitly opted-in by prefixing the variable name with $
Differential Revision: https://reviews.llvm.org/D81531
Allow for dynamic indices in the `dim` operation.
Rather than an attribute, the index is now an operand of type `index`.
This allows to apply the operation to dynamically ranked tensors.
The correct lowering of dynamic indices remains to be implemented.
Differential Revision: https://reviews.llvm.org/D81551
Having the input dumped on failure seems like a better
default: I debugged FileCheck tests for a while without knowing
about this option, which really helps to understand failures.
Remove `-dump-input-on-failure` and the environment variable
FILECHECK_DUMP_INPUT_ON_FAILURE which are now obsolete.
Differential Revision: https://reviews.llvm.org/D81422
This revision adds a helper function to hoist vector.transfer_read /
vector.transfer_write pairs out of immediately enclosing scf::ForOp
iteratively, if the following conditions are true:
1. The 2 ops access the same memref with the same indices.
2. All operands are invariant under the enclosing scf::ForOp.
3. No uses of the memref either dominate the transfer_read or are
dominated by the transfer_write (i.e. no aliasing between the write and
the read across the loop)
To improve hoisting opportunities, call the `moveLoopInvariantCode` helper
function on the candidate loop above which to hoist. Hoisting the transfers
results in scf::ForOp yielding the value that originally transited through
memory.
This revision additionally exposes `moveLoopInvariantCode` as a helper in
LoopUtils.h and updates SliceAnalysis to support return scf::For values and
allow hoisting across multiple scf::ForOps.
Differential Revision: https://reviews.llvm.org/D81199
Update linalg to affine lowering for convop to use affine load for input
whenever there is no padding. It had always been using std.loads because
max in index functions (needed for non-zero padding if not materializing
zeros) couldn't be represented in the non-zero padding cases.
In the future, the non-zero padding case could also be made to use
affine - either by materializing or using affine.execute_region. The
latter approach will not impact the scf/std output obtained after
lowering out affine.
Differential Revision: https://reviews.llvm.org/D81191
This revision adds a helper function to hoist alloc/dealloc pairs and
alloca op out of immediately enclosing scf::ForOp if both conditions are true:
1. all operands are defined outside the loop.
2. all uses are ViewLikeOp or DeallocOp.
This is now considered Linalg-specific and will be generalized on a per-need basis.
Differential Revision: https://reviews.llvm.org/D81152
Summary:
The fusion for tensor_reshape is embedding the information to indexing maps,
thus the exising pattenr also works for indexed_generic ops.
Depends On D80347
Differential Revision: https://reviews.llvm.org/D80348
Summary:
Different from the fusion between generic ops, indices are involved. In this
context, we need to re-map the indices for producer since the fused op is built
on consumer's perspective. This patch supports all combination of the fusion
between indexed_generic ops and generic ops, which includes tests case:
1) generic op as producer and indexed_generic op as consumer.
2) indexed_generic op as producer and generic op as consumer.
3) indexed_generic op as producer and indexed_generic op as consumer.
Differential Revision: https://reviews.llvm.org/D80347
This revision replaces the load + vector.type_cast by appropriate vector transfer
operations. These play more nicely with other vector abstractions and canonicalization
patterns and lower to load/store with or without masks when appropriate.
Differential Revision: https://reviews.llvm.org/D80809
This revision adds custom rewrites for patterns that arise during linalg structured
ops vectorization. These patterns allow the composition of linalg promotion,
vectorization and removal of redundant copies.
The patterns are voluntarily limited and restrictive atm.
More robust behavior will be implemented once more powerful side effect modeling and analyses are available on view/subview.
On the transfer_read side, the following pattern is rewritten:
```
%alloc = ...
[optional] %view = std.view %alloc ...
%subView = subview %allocOrView ...
[optional] linalg.fill(%allocOrView, %cst) ...
...
linalg.copy(%in, %subView) ...
vector.transfer_read %allocOrView[...], %cst ...
```
into
```
[unchanged] %alloc = ...
[unchanged] [optional] %view = std.view %alloc ...
[unchanged] [unchanged] %subView = subview %allocOrView ...
...
vector.transfer_read %in[...], %cst ...
```
On the transfer_write side, the following pattern is rewriten:
```
%alloc = ...
[optional] %view = std.view %alloc ...
%subView = subview %allocOrView...
...
vector.transfer_write %..., %allocOrView[...]
linalg.copy(%subView, %out)
```
Differential Revision: https://reviews.llvm.org/D80728
operands of Generic ops.
Unit-extent dimensions are typically used for achieving broadcasting
behavior. The pattern added (along with canonicalization patterns
added previously) removes the use of unit-extent dimensions, and
instead uses a more canonical representation of the computation. This
new pattern is not added as a canonicalization for now since it
entails adding additional reshape operations. A pass is added to
exercise these patterns, along with an API entry to populate a
patterns list with these patterns.
Differential Revision: https://reviews.llvm.org/D79766
alloc/dealloc/copies.
Add options to LinalgPromotion to use callbacks for implementating the
allocation, deallocation of buffers used for the promoted subviews,
and to copy data into and from the original subviews to the allocated
buffers.
Also some misc. cleanup of the code.
Differential Revision: https://reviews.llvm.org/D80365
Modifying the loop nest builder for generating scf.parallel loops to
not generate scf.parallel loops for non-parallel iterator types in
Linalg operations. The existing implementation incorrectly generated
scf.parallel for all tiled loops. It is rectified by refactoring logic
used while lowering to loops that accounted for this.
Differential Revision: https://reviews.llvm.org/D80188
Summary:
This revision refactors the Linalg tiling pass to be written as pattern applications and retires the use of the folder in Linalg tiling.
In the early days, tiling was written as a pass that would create (partially) folded and canonicalized operations on the fly for better composability.
As this evolves towards composition of patterns, the pass-specific folder is counter-productive and is retired.
The tiling options struct evolves to take a tile size creation function which allows materializing tile sizes on the fly (in particular constant tile sizes). This plays better with folding and DCE.
With the folder going away in Tiling, the check on whether subviews are the same in linalg fusion needs to be more robust. This revision also implements such a check.
In the current form, there are still some canonicalizations missing due to AffineMin/Max ops fed by scf::ForOp. These will be improved at a later time.
Differential Revision: https://reviews.llvm.org/D80267
For now the promoted buffer is indexed using the `full view`. The full view might be
slightly bigger than the partial view (which is accounting for boundaries).
Unfortunately this does not compose easily with other transformations when multiple buffers
with shapes related to each other are involved.
Take `linalg.matmul A B C` (with A of size MxK, B of size KxN and C of size MxN) and suppose we are:
- Tiling over M by 100
- Promoting A only
This is producing a `linalg.matmul promoted_A B subview_C` where `promoted_A` is a promoted buffer
of `A` of size (100xK) and `subview_C` is a subview of size mxK where m could be smaller than 100 due
to boundaries thus leading to a possible incorrect behavior.
We propose to:
- Add a new parameter to the tiling promotion allowing to enable the use of the full tile buffer.
- By default all promoted buffers will be indexed by the partial view.
Note that this could be considered as a breaking change in comparison to the way the tiling promotion
was working.
Differential Revision: https://reviews.llvm.org/D79927
This revision starts decoupling the include the kitchen sink behavior of Linalg to LLVM lowering by inserting a -convert-linalg-to-std pass.
The lowering of linalg ops to function calls was previously lowering to memref descriptors by having both linalg -> std and std -> LLVM patterns in the same rewrite.
When separating this step, a new issue occurred: the layout is automatically type-erased by this process. This revision therefore introduces memref casts to perform these type erasures explicitly. To connect everything end-to-end, the LLVM lowering of MemRefCastOp is relaxed because it is artificially more restricted than the op semantics. The op semantics already guarantee that source and target MemRefTypes are cast-compatible. An invalid lowering test now becomes valid and is removed.
Differential Revision: https://reviews.llvm.org/D79468
All ops of the SCF dialect now use the `scf.` prefix instead of `loop.`. This
is a part of dialect renaming.
Differential Revision: https://reviews.llvm.org/D79844
The existing implementation of SubViewOp::getRanges relies on all
offsets/sizes/strides to be dynamic values and does not work in
combination with canonicalization. This revision adds a
SubViewOp::getOrCreateRanges to create the missing constants in the
canonicalized case.
This allows reactivating the fused pass with staged pattern
applications.
However another issue surfaces that the SubViewOp verifier is now too
strict to allow folding. The existing folding pattern is turned into a
canonicalization pattern which rewrites memref_cast + subview into
subview + memref_cast.
The transform-patterns-matmul-to-vector can then be reactivated.
Differential Revision: https://reviews.llvm.org/D79759
This is only valid if the source tensors (result tensor) is static
shaped with all unit-extents when the reshape is collapsing
(expanding) dimensions.
Differential Revision: https://reviews.llvm.org/D79764
The main objective of this revision is to change the way static information is represented, propagated and canonicalized in the SubViewOp.
In the current implementation the issue is that canonicalization may strictly lose information because static offsets are combined in irrecoverable ways into the result type, in order to fit the strided memref representation.
The core semantics of the op do not change but the parser and printer do: the op always requires `rank` offsets, sizes and strides. These quantities can now be either SSA values or static integer attributes.
The result type is automatically deduced from the static information and more powerful canonicalizations (as powerful as the representation with sentinel `?` values allows). Previously static information was inferred on a best-effort basis from looking at the source and destination type.
Relevant tests are rewritten to use the idiomatic `offset: x, strides : [...]`-form. Bugs are corrected along the way that were not trivially visible in flattened strided memref form.
Lowering to LLVM is updated, simplified and now supports all cases.
A mixed static-dynamic mode test that wouldn't previously lower is added.
It is an open question, and a longer discussion, whether a better result type representation would be a nicer alternative. For now, the subview op carries the required semantic.
Differential Revision: https://reviews.llvm.org/D79662
This reverts commit 80d133b24f.
Per Stephan Herhut: The canonicalizer pattern that was added creates
forms of the subview op that cannot be lowered.
This is shown by failing Tensorflow XLA tests such as:
tensorflow/compiler/xla/service/mlir_gpu/tests:abs.hlo.test
Will provide more details offline, they rely on logs from private CI.
Summary:
The main objective of this revision is to change the way static information is represented, propagated and canonicalized in the SubViewOp.
In the current implementation the issue is that canonicalization may strictly lose information because static offsets are combined in irrecoverable ways into the result type, in order to fit the strided memref representation.
The core semantics of the op do not change but the parser and printer do: the op always requires `rank` offsets, sizes and strides. These quantities can now be either SSA values or static integer attributes.
The result type is automatically deduced from the static information and more powerful canonicalizations (as powerful as the representation with sentinel `?` values allows). Previously static information was inferred on a best-effort basis from looking at the source and destination type.
Relevant tests are rewritten to use the idiomatic `offset: x, strides : [...]`-form. Bugs are corrected along the way that were not trivially visible in flattened strided memref form.
It is an open question, and a longer discussion, whether a better result type representation would be a nicer alternative. For now, the subview op carries the required semantic.
Reviewers: ftynse, mravishankar, antiagainst, rriddle!, andydavis1, timshen, asaadaldien, stellaraccident
Reviewed By: mravishankar
Subscribers: aartbik, bondhugula, mehdi_amini, rriddle, jpienaar, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, liufengdb, stephenneuendorffer, Joonsoo, bader, grosul1, frgossen, Kayjukh, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79662
Summary:
This revision introduces a helper function to allow applying rewrite patterns, interleaved with more global transformations, in a staged fashion:
1. the first stage consists of an OwningRewritePatternList. The RewritePattern in this list are applied once and in order.
2. the second stage consists of a single OwningRewritePattern that is applied greedily until convergence.
3. the third stage consists of applying a lambda, generally used for non-local transformation effects.
This allows creating custom fused transformations where patterns can be ordered and applied at a finer granularity than a sequence of traditional compiler passes.
A test that exercises these behaviors is added.
Differential Revision: https://reviews.llvm.org/D79518
This [discussion](https://llvm.discourse.group/t/viewop-isnt-expressive-enough/991/2) raised some concerns with ViewOp.
In particular, the handling of offsets is incorrect and does not match the op description.
Note that with an elemental type change, offsets cannot be part of the type in general because sizeof(srcType) != sizeof(dstType).
Howerver, offset is a poorly chosen term for this purpose and is renamed to byte_shift.
Additionally, for all intended purposes, trying to support non-identity layouts for this op does not bring expressive power but rather increases code complexity.
This revision simplifies the existing semantics and implementation.
This simplification effort is voluntarily restrictive and acts as a stepping stone towards supporting richer semantics: treat the non-common cases as YAGNI for now and reevaluate based on concrete use cases once a round of simplification occurred.
Differential revision: https://reviews.llvm.org/D79541
Originally, these operations were folded only if all expressions in their
affine maps could be folded to a constant expression that can be then subject
to numeric min/max computation. This introduces a more advanced version that
partially folds the affine map by lifting individual constant expression in it
even if some of the expressions remain variable. The folding can update the
operation in place to use a simpler map. Note that this is not as powerful as
canonicalization, in particular this does not remove dimensions or symbols that
became useless. This allows for better composition of Linalg tiling and
promotion transformation, where the latter can handle some canonical forms of
affine.min that the folding can now produce.
Differential Revision: https://reviews.llvm.org/D79502
Nicolas Vasilache