This commit introduced a cyclic dependency:
Memref dialect depends on Standard because it used ConstantIndexOp.
Std depends on the MemRef dialect in its EDSC/Intrinsics.h
Working on a fix.
This reverts commit 8aa6c3765b924d86f623d452777eb76b83bf2787.
Create the memref dialect and move several dialect-specific ops without
dependencies to other ops from std dialect to this dialect.
Moved ops:
AllocOp -> MemRef_AllocOp
AllocaOp -> MemRef_AllocaOp
DeallocOp -> MemRef_DeallocOp
MemRefCastOp -> MemRef_CastOp
GetGlobalMemRefOp -> MemRef_GetGlobalOp
GlobalMemRefOp -> MemRef_GlobalOp
PrefetchOp -> MemRef_PrefetchOp
ReshapeOp -> MemRef_ReshapeOp
StoreOp -> MemRef_StoreOp
TransposeOp -> MemRef_TransposeOp
ViewOp -> MemRef_ViewOp
The roadmap to split the memref dialect from std is discussed here:
https://llvm.discourse.group/t/rfc-split-the-memref-dialect-from-std/2667
Differential Revision: https://reviews.llvm.org/D96425
Some of the lowering of vector.contract didn't support integer case. Since
reduction of integer cannot accumulate we always break up the reduction op, it
should be merged by a separate canonicalization if possible.
Differential Revision: https://reviews.llvm.org/D96461
These patterns unrolls transfer read/write ops if the vector consumers/
producers are extract/insert slices op. Transfer ops can map to hardware
load/store functionalities, where the vector size matters for bandwidth
considerations. So these patterns should be collected separately, instead
of being generic canonicalization patterns.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D96782
Currently, vector.contract joins the intermediate result and the accumulator
argument (of ranks K) using summation. We desire more joining operations ---
such as max --- to help vector.contract express reductions. This change extends
Vector_ContractionOp to take an optional attribute (called "kind", of enum type
CombiningKind) specifying the joining operation to be add/mul/min/max for int/fp
, and and/or/xor for int only. By default this attribute has value "add".
To implement this we also need to extend vector.outerproduct, since
vector.contract gets transformed to vector.outerproduct (and that to
vector.fma). The extension for vector.outerproduct is also an optional kind
attribute that uses the same enum type and possible values. The default is
"add". In case of max/min we transform vector.outerproduct to a combination of
compare and select.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D93280
These patterns move vector.bitcast ops to be before
insert ops or after extract ops where suitable.
With them, bitcast will happen on smaller vectors
and there are more chances to share extract/insert
ops.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D96040
This patch adds patterns to use vector.shape_cast to cast
away leading 1-dimensions from a few vector operations.
It allows exposing more canonical forms of vector.transfer_read,
vector.transfer_write, vector_extract_strided_slice, and
vector.insert_strided_slice. With this, we can have more
opportunity to cancelling extract/insert ops or forwarding
write/read ops.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D95873
This revision starts evolving the APIs to manipulate ops with offsets, sizes and operands towards a ValueOrAttr abstraction that is already used in folding under the name OpFoldResult.
The objective, in the future, is to allow such manipulations all the way to the level of ODS to avoid all the genuflexions involved in distinguishing between values and attributes for generic constant foldings.
Once this evolution is accepted, the next step will be a mechanical OpFoldResult -> ValueOrAttr.
Differential Revision: https://reviews.llvm.org/D95310
Implement Bug 46698, making ODS synthesize a getType() method that returns a
specific C++ class for OneResult methods where we know that class. This eliminates
a common source of casts in things like:
myOp.getType().cast<FIRRTLType>().getPassive()
because we know that myOp always returns a FIRRTLType. This also encourages
op authors to type their results more tightly (which is also good for
verification).
I chose to implement this by splitting the OneResult trait into itself plus a
OneTypedResult trait, given that many things are using `hasTrait<OneResult>`
to conditionalize various logic.
While this changes makes many many ops get more specific getType() results, it
is generally drop-in compatible with the previous behavior because 'x.cast<T>()'
is allowed when x is already known to be a T. The one exception to this is that
we need declarations of the types used by ops, which is why a couple headers
needed additional #includes.
I updated a few things in tree to remove the now-redundant `.cast<>`'s, but there
are probably many more than can be removed.
Differential Revision: https://reviews.llvm.org/D93790
Extend unroll to support all element-wise ops and allow unrolling for ops with
vector operands of with the same shape as the destination but different element
type (like Cmp or Select).
Differential Revision: https://reviews.llvm.org/D93121
Transfer_ops can now work on both buffers and tensor. Right now, lowering of
the tensor case is not supported yet.
Differential Revision: https://reviews.llvm.org/D93500
This better matches the rest of the infrastructure, is much simpler, and makes it easier to move these types to being declaratively specified.
Differential Revision: https://reviews.llvm.org/D93432
Given that OpState already implicit converts to Operator*, this seems reasonable.
The alternative would be to add more functions to OpState which forward to Operation.
Reviewed By: rriddle, ftynse
Differential Revision: https://reviews.llvm.org/D92266
These includes have been deprecated in favor of BuiltinDialect.h, which contains the definitions of ModuleOp and FuncOp.
Differential Revision: https://reviews.llvm.org/D91572
motivated by a refactoring in the new sparse code (yet to be merged), this avoids some lengthy code dup
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D91465
Support multi-dimension vector for InsertMap/ExtractMap op and update the
transformations. Currently the relation between IDs and dimension is implicitly
deduced from the types. We can then calculate an AffineMap based on it. In the
future the AffineMap could be part of the operation itself.
Differential Revision: https://reviews.llvm.org/D90995
Fix semantic in the distribute integration test based on offline feedback. This
exposed a bug in block distribution, we need to make sure the id is multiplied
by the stride of the vector. Fix the transformation and unit test.
Differential Revision: https://reviews.llvm.org/D89291
Based on discourse discussion, fix the doc string and remove examples with
wrong semantic. Also fix insert_map semantic by adding missing operand for
vector we are inserting into.
Differential Revision: https://reviews.llvm.org/D89563
Adding unroll support for transfer read and transfer write operation. This
allows to pick the ideal size for the memory access for a given target.
Differential Revision: https://reviews.llvm.org/D89289
When distributing a vector larger than the given multiplicity, we can
distribute it by block where each id gets a chunk of consecutive element
along the dimension distributed. This adds a test for this case and adds extra
checks to make sure we don't distribute for cases not multiple of multiplicity.
Differential Revision: https://reviews.llvm.org/D89061
Add basic canonicalization patterns for the extractMap/insertMap to allow them
to be folded into Transfer ops.
Also mark transferRead as memory read so that it can be removed by dead code.
Differential Revision: https://reviews.llvm.org/D88622
This is the first of several steps to support distributing large vectors. This
adds instructions extract_map and insert_map that allow us to do incremental
lowering. Right now the transformation only apply to simple pointwise operation
with a vector size matching the multiplicity of the IDs used to distribute the
vector.
This can be used to distribute large vectors to loops or SPMD.
Differential Revision: https://reviews.llvm.org/D88341
This commit adds support for subviews which enable to reduce resulting rank
by dropping static dimensions of size 1.
Differential Revision: https://reviews.llvm.org/D88534
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
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 `splitFullAndPartialTransferPrecondition` has a restrictive condition to
prevent the pattern to be applied recursively if it is nested under an scf.IfOp.
Relaxing the condition to the immediate parent op must not be an scf.IfOp lets
the pattern be applied more generally while still preventing recursion.
Differential Revision: https://reviews.llvm.org/D85209
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
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
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
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
