Fold the operation if the source is a scalar constant or splat constant.
Update transform-patterns-matmul-to-vector.mlir because the broadcast ops are folded in the conversion.
Reviewed By: aartbik
Differential Revision: https://reviews.llvm.org/D87703
When allowed, use 32-bit indices rather than 64-bit indices in the
SIMD computation of masks. This runs up to 2x and 4x faster on
a number of AVX2 and AVX512 microbenchmarks.
Reviewed By: bkramer
Differential Revision: https://reviews.llvm.org/D87116
Masked loading/storing in various forms can be optimized
into simpler memory operations when the mask is all true
or all false. Note that the backend does similar optimizations
but doing this early may expose more opportunities for further
optimizations. This further prepares progressively lowering
transfer read and write into 1-D memory operations.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D85769
This new pattern mixes vector.transpose and direct lowering to vector.reduce.
This allows more progressive lowering than immediately going to insert/extract and
composes more nicely with other canonicalizations.
This has 2 use cases:
1. for very wide vectors the generated IR may be much smaller
2. when we have a custom lowering for transpose ops we can target it directly
rather than rely LLVM
Differential Revision: https://reviews.llvm.org/D85428
The intrinsics were already supported and vector.transfer_read/write lowered
direclty into these operations. By providing them as individual ops, however,
clients can used them directly, and it opens up progressively lowering transfer
operations at higher levels (rather than direct lowering to LLVM IR as done now).
Reviewed By: bkramer
Differential Revision: https://reviews.llvm.org/D85357
Introduces the expand and compress operations to the Vector dialect
(important memory operations for sparse computations), together
with a first reference implementation that lowers to the LLVM IR
dialect to enable running on CPU (and other targets that support
the corresponding LLVM IR intrinsics).
Reviewed By: reidtatge
Differential Revision: https://reviews.llvm.org/D84888
This revision adds a transformation and a pattern that rewrites a "maybe masked" `vector.transfer_read %view[...], %pad `into a pattern resembling:
```
%1:3 = scf.if (%inBounds) {
scf.yield %view : memref<A...>, index, index
} else {
%2 = linalg.fill(%extra_alloc, %pad)
%3 = subview %view [...][...][...]
linalg.copy(%3, %alloc)
memref_cast %extra_alloc: memref<B...> to memref<A...>
scf.yield %4 : memref<A...>, index, index
}
%res= vector.transfer_read %1#0[%1#1, %1#2] {masked = [false ... false]}
```
where `extra_alloc` is a top of the function alloca'ed buffer of one vector.
This rewrite makes it possible to realize the "always full tile" abstraction where vector.transfer_read operations are guaranteed to read from a padded full buffer.
The extra work only occurs on the boundary tiles.
This revision adds a transformation and a pattern that rewrites a "maybe masked" `vector.transfer_read %view[...], %pad `into a pattern resembling:
```
%1:3 = scf.if (%inBounds) {
scf.yield %view : memref<A...>, index, index
} else {
%2 = vector.transfer_read %view[...], %pad : memref<A...>, vector<...>
%3 = vector.type_cast %extra_alloc : memref<...> to
memref<vector<...>> store %2, %3[] : memref<vector<...>> %4 =
memref_cast %extra_alloc: memref<B...> to memref<A...> scf.yield %4 :
memref<A...>, index, index
}
%res= vector.transfer_read %1#0[%1#1, %1#2] {masked = [false ... false]}
```
where `extra_alloc` is a top of the function alloca'ed buffer of one vector.
This rewrite makes it possible to realize the "always full tile" abstraction where vector.transfer_read operations are guaranteed to read from a padded full buffer.
The extra work only occurs on the boundary tiles.
Differential Revision: https://reviews.llvm.org/D84631
This reverts commit 35b65be041127db9fe23d3128a004c888893cbae.
Build is broken with -DBUILD_SHARED_LIBS=ON with some undefined
references like:
VectorTransforms.cpp:(.text._ZN4llvm12function_refIFvllEE11callback_fnIZL24createScopedInBoundsCondN4mlir25VectorTransferOpInterfaceEE3$_8EEvlll+0xa5): undefined reference to `mlir::edsc::op::operator+(mlir::Value, mlir::Value)'
This revision adds a transformation and a pattern that rewrites a "maybe masked" `vector.transfer_read %view[...], %pad `into a pattern resembling:
```
%1:3 = scf.if (%inBounds) {
scf.yield %view : memref<A...>, index, index
} else {
%2 = vector.transfer_read %view[...], %pad : memref<A...>, vector<...>
%3 = vector.type_cast %extra_alloc : memref<...> to
memref<vector<...>> store %2, %3[] : memref<vector<...>> %4 =
memref_cast %extra_alloc: memref<B...> to memref<A...> scf.yield %4 :
memref<A...>, index, index
}
%res= vector.transfer_read %1#0[%1#1, %1#2] {masked = [false ... false]}
```
where `extra_alloc` is a top of the function alloca'ed buffer of one vector.
This rewrite makes it possible to realize the "always full tile" abstraction where vector.transfer_read operations are guaranteed to read from a padded full buffer.
The extra work only occurs on the boundary tiles.
Differential Revision: https://reviews.llvm.org/D84631
Introduces the scatter/gather operations to the Vector dialect
(important memory operations for sparse computations), together
with a first reference implementation that lowers to the LLVM IR
dialect to enable running on CPU (and other targets that support
the corresponding LLVM IR intrinsics).
The operations can be used directly where applicable, or can be used
during progressively lowering to bring other memory operations closer to
hardware ISA support for a gather/scatter. The semantics of the operation
closely correspond to those of the corresponding llvm intrinsics.
Note that the operation allows for a dynamic index vector (which is
important for sparse computations). However, this first reference
lowering implementation "serializes" the address computation when
base + index_vector is converted to a vector of pointers. Exploring
how to use SIMD properly during these step is TBD. More general
memrefs and idiomatic versions of striding are also TBD.
Reviewed By: arpith-jacob
Differential Revision: https://reviews.llvm.org/D84039
Summary: Vector contract patterns were only parameterized by a `vectorTransformsOptions`. As a result, even if an mlir file was containing several occurrences of `vector.contract`, all of them would be lowered in the same way. More granularity might be required . This Diff adds a `constraint` argument to each of these patterns which allows the user to specify with more precision on which `vector.contract` should each of the lowering apply.
Differential Revision: https://reviews.llvm.org/D83960
This revision folds vector.transfer operations by updating the `masked` bool array attribute when more unmasked dimensions can be discovered.
Differential revision: https://reviews.llvm.org/D83586
We temporarily had separate OUTER lowering (for matmat flavors) and
AXPY lowering (for matvec flavors). With the new generalized
"vector.outerproduct" semantics, these cases can be merged into
a single lowering method. This refactoring will simplify future
decisions on cost models and lowering heuristics.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D83585
This specialization allows sharing more code where an AXPY follows naturally
in cases where an OUTERPRODUCT on a scalar would be generated.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D83453
TransposeOp are often followed by ExtractOp.
In certain cases however, it is unnecessary (and even detrimental) to lower a TransposeOp to either a flat transpose (llvm.matrix intrinsics) or to unrolled scalar insert / extract chains.
Providing foldings of ExtractOp mitigates some of the unnecessary complexity.
Differential revision: https://reviews.llvm.org/D83487
This revision adds foldings for ExtractOp operations that come from previous InsertOp.
InsertOp have cumulative semantic where multiple chained inserts are necessary to produce the final value from which the extracts are obtained.
Additionally, TransposeOp may be interleaved and need to be tracked in order to follow the producer consumer relationships and properly compute positions.
Differential revision: https://reviews.llvm.org/D83150
The UnrollVectorPattern is can be used in a programmable fashion by:
```
OwningRewritePatternList patterns;
patterns.insert<UnrollVectorPattern<AddFOp>>(ArrayRef<int64_t>{2, 2}, ctx);
patterns.insert<UnrollVectorPattern<vector::ContractionOp>>(
ArrayRef<int64_t>{2, 2, 2}, ctx);
...
applyPatternsAndFoldGreedily(getFunction(), patterns);
```
Differential revision: https://reviews.llvm.org/D83064
Default vector.contract lowering essentially yields a series of sdot/ddot
operations. However, for some layouts a series of saxpy/daxpy operations,
chained through fma are more efficient. This CL introduces a choice between
the two lowering paths. A default heuristic is to follow.
Some preliminary avx2 performance numbers for matrix-times-vector.
Here, dot performs best for 64x64 A x b and saxpy for 64x64 A^T x b.
```
------------------------------------------------------------
A x b A^T x b
------------------------------------------------------------
GFLOPS sdot (reassoc) saxpy sdot (reassoc) saxpy
------------------------------------------------------------
1x1 0.6 0.9 0.6 0.9
2x2 2.5 3.2 2.4 3.5
4x4 6.4 8.4 4.9 11.8
8x8 11.7 6.1 5.0 29.6
16x16 20.7 10.8 7.3 43.3
32x32 29.3 7.9 6.4 51.8
64x64 38.9 79.3
128x128 32.4 40.7
------------------------------------------------------------
```
Reviewed By: nicolasvasilache, ftynse
Differential Revision: https://reviews.llvm.org/D83012
More efficient implementation of the multiply-reduce pair,
no need to add in a zero vector. Microbenchmarking on AVX2
yields the following difference in vector.contract speedup
(over strict-order scalar reduction).
SPEEDUP SIMD-fma SIMD-mul
4x4 1.45 2.00
8x8 1.40 1.90
32x32 5.32 5.80
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D82833
Use vector compares for the 1-D case. This approach scales much better
than generating insertion operations, and exposes SIMD directly to backend.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D82402
Allow lhs and rhs to have different type than accumulator/destination. Some
hardware like GPUs support natively operations like uint8xuint8xuint32.
Differential Revision: https://reviews.llvm.org/D82069
Use direct vector constants for the 1-D case. This approach
scales much better than generating elaborate insertion operations
that are eventually folded into a constant. We could of course
generalize the 1-D case to higher ranks, but this simplification
already helps in scaling some microbenchmarks that would formerly
crash on the intermediate IR length.
Reviewed By: reidtatge
Differential Revision: https://reviews.llvm.org/D82144
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
Summary:
Even though this operation is intended for 1d/2d conversions currently,
leaving a semantic hole in the lowering prohibits proper testing of this
operation. This CL adds a straightforward reference implementation for the
missing cases.
Reviewers: nicolasvasilache, mehdi_amini, ftynse, reidtatge
Reviewed By: reidtatge
Subscribers: mehdi_amini, rriddle, jpienaar, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, liufengdb, stephenneuendorffer, Joonsoo, grosul1, frgossen, Kayjukh, jurahul, msifontes
Tags: #mlir
Differential Revision: https://reviews.llvm.org/D81503
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
Summary:
This revision adds a common folding pattern that starts appearing on
vector_transfer ops.
Differential Revision: https://reviews.llvm.org/D81281
This allows verifying op-indepent attributes (e.g., attributes that do not require the op to have been created) before constructing an operation. These include checking whether required attributes are defined or constraints on attributes (such as I32 attribute). This is not perfect (e.g., if one had a disjunctive constraint where one part relied on the op and the other doesn't, then this would not try and extract the op independent from the op dependent).
The next step is to move these out to a trait that could be verified earlier than in the generated method. The first use case is for inferring the return type while constructing the op. At that point you don't have an Operation yet and that ends up in one having to duplicate the same checks, e.g., verify that attribute A is defined before querying A in shape function which requires that duplication. Instead this allows one to invoke a method to verify all the traits and, if this is checked first during verification, then all other traits could use attributes knowing they have been verified.
It is a little bit funny to have these on the adaptor, but I see the adaptor as a place to collect information about the op before the op is constructed (e.g., avoiding stringly typed accessors, verifying what is possible to verify before the op is constructed) while being cheap to use even with constructed op (so layer of indirection between the op constructed/being constructed). And from that point of view it made sense to me.
Differential Revision: https://reviews.llvm.org/D80842
This simplifies a lot of handling of BoolAttr/IntegerAttr. For example, a lot of places currently have to handle both IntegerAttr and BoolAttr. In other places, a decision is made to pick one which can lead to surprising results for users. For example, DenseElementsAttr currently uses BoolAttr for i1 even if the user initialized it with an Array of i1 IntegerAttrs.
Differential Revision: https://reviews.llvm.org/D81047
Summary:
Progressive lowering of vector.transpose into an operation that
is closer to an intrinsic, and thus the hardware ISA. Currently
under the common vector transform testing flag, as we prepare
deploying this transformation in the LLVM lowering pipeline.
Reviewers: nicolasvasilache, reidtatge, andydavis1, ftynse
Reviewed By: nicolasvasilache, ftynse
Subscribers: mehdi_amini, rriddle, jpienaar, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, liufengdb, stephenneuendorffer, Joonsoo, grosul1, frgossen, Kayjukh, jurahul, llvm-commits
Tags: #llvm, #mlir
Differential Revision: https://reviews.llvm.org/D80772
This revision expands the types of vector contractions that can be lowered to vector.outerproduct.
All 8 permutation cases are support.
The idiomatic manipulation of AffineMap written declaratively makes this straightforward.
In the process a bug with the vector.contract verifier was uncovered.
The vector shape verification part of the contract op is rewritten to use AffineMap composition.
One bug in the vector `ops.mlir` test is fixed and a new case not yet captured is added
to the vector`invalid.mlir` test.
Differential Revision: https://reviews.llvm.org/D80393
This revision adds the additional lowering and exposes the patterns at a finer granularity for better programmatic reuse. The unit test makes use of the finer grained pattern for simpler checks.
As the ContractionOpLowering is exposed programmatically, cleanup opportunities appear and static class methods are turned into free functions with static visibility.
Differential Revision: https://reviews.llvm.org/D80375
Summary:
Vector transfer ops semantic is extended to allow specifying a per-dimension `masked`
attribute. When the attribute is false on a particular dimension, lowering to LLVM emits
unmasked load and store operations.
Differential Revision: https://reviews.llvm.org/D80098
Summary:
This revision makes the use of vector transfer operatons more idiomatic by
allowing to omit and inferring the permutation_map.
Differential Revision: https://reviews.llvm.org/D80092
Summary:
First, compact implementation of lowering to LLVM IR. A bit more
challenging than the constant mask due to the dynamic indices, of course.
I like to hear if there are more efficient ways of doing this in LLVM,
but this for now at least gives us a functional reference implementation.
Reviewers: nicolasvasilache, ftynse, bkramer, reidtatge, andydavis1, mehdi_amini
Reviewed By: nicolasvasilache
Subscribers: mehdi_amini, rriddle, jpienaar, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, liufengdb, stephenneuendorffer, Joonsoo, grosul1, frgossen, Kayjukh, jurahul, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79954
Summary:
Makes this operation runnable on CPU by generating MLIR instructions
that are eventually folded into an LLVM IR constant for the mask.
Reviewers: nicolasvasilache, ftynse, reidtatge, bkramer, andydavis1
Reviewed By: nicolasvasilache, ftynse, andydavis1
Subscribers: mehdi_amini, rriddle, jpienaar, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, liufengdb, stephenneuendorffer, Joonsoo, grosul1, frgossen, Kayjukh, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79815
Summary:
Rather than having a full, recursive, lowering of vector.broadcast
to LLVM IR, it is much more elegant to have a progressive lowering
of each vector.broadcast into a lower dimensional vector.broadcast,
until only elementary vector operations remain. This results
in more elegant, step-wise code, that is easier to understand.
Also makes some optimizations in the generated code.
Reviewers: nicolasvasilache, mehdi_amini, andydavis1, grosul1
Reviewed By: nicolasvasilache
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, liufengdb, Joonsoo, grosul1, frgossen, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78071
Summary:
LLVM matrix intrinsics recently introduced an option to support row-major mode.
This matches the MLIR vector model, this revision switches to row-major.
A corner case related to degenerate sizes was also fixed upstream.
This revision removes the guard against this corner case.
A bug was uncovered on the output vector construction which this revision also fixes.
Lastly, this has been tested on a small size and benchmarked independently: no visible performance regression is observed.
In the future, when matrix intrinsics support per op attribute, we can more aggressively translate to that and avoid inserting MLIR-level transposes.
This has been tested independently to work on small matrices.
Differential Revision: https://reviews.llvm.org/D77761
Summary:
Update ShapeCastOp folder to use producer-consumer value forwarding.
Support is added for tracking sub-vectors through trivial shape cast operations,
where the sub-vector shape is preserved across shape cast operations and only
leading ones are added or removed.
Support is preserved for cancelling shape cast operations.
One unit test is added and two are updated.
Reviewers: aartbik, nicolasvasilache
Reviewed By: aartbik, nicolasvasilache
Subscribers: frgossen, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, liufengdb, Joonsoo, grosul1, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77253
Two back-to-back transpose operations are combined into a single transpose, which uses a combination of their permutation vectors.
Differential Revision: https://reviews.llvm.org/D77331
Summary:
Add support for TupleGetOp folding through InsertSlicesOp and ExtractSlicesOp.
Vector-to-vector transformations for unrolling and lowering to hardware vectors
can generate chains of structured vector operations (InsertSlicesOp,
ExtractSlicesOp and ShapeCastOp) between the producer of a hardware vector
value and its consumer. Because InsertSlicesOp, ExtractSlicesOp and ShapeCastOp
are structured, we can track the location (tuple index and vector offsets) of
the consumer vector value through the chain of structured operations to the
producer, enabling a much more powerful producer-consumer fowarding of values
through structured ops and tuple, which in turn enables a more powerful
TupleGetOp folding transformation.
Reviewers: nicolasvasilache, aartbik
Reviewed By: aartbik
Subscribers: grosul1, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, liufengdb, Joonsoo, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76889
