Convert subtensor and subtensor_insert operations to use their
rank-reduced versions to drop unit dimensions.
Differential Revision: https://reviews.llvm.org/D101495
The current implementation had a bug as it was relying on the target vector
dimension sizes to calculate where to insert broadcast. If several dimensions
have the same size we may insert the broadcast on the wrong dimension. The
correct broadcast cannot be inferred from the type of the source and
destination vector.
Instead when we want to extend transfer ops we calculate an "inverse" map to the
projected permutation and insert broadcast in place of the projected dimensions.
Differential Revision: https://reviews.llvm.org/D101738
This is the very first step toward removing the glue and clutter from linalg and
replace it with proper sparse tensor types. This revision migrates the LinalgSparseOps
into SparseTensorOps of a sparse tensor dialect. This also provides a new home for
sparse tensor related transformation.
NOTE: the actual replacement with sparse tensor types (and removal of linalg glue/clutter)
will follow but I am trying to keep the amount of changes per revision manageable.
Differential Revision: https://reviews.llvm.org/D101573
This is the very first step toward removing the glue and clutter from linalg and
replace it with proper sparse tensor types. This revision migrates the LinalgSparseOps
into SparseTensorOps of a sparse tensor dialect. This also provides a new home for
sparse tensor related transformation.
NOTE: the actual replacement with sparse tensor types (and removal of linalg glue/clutter)
will follow but I am trying to keep the amount of changes per revision manageable.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D101488
FillOp allows complex ops, and filling a properly sized buffer with
a default zero complex number is implemented.
Differential Revision: https://reviews.llvm.org/D99939
This revision adds support for vectorizing more general linalg operations with projected permutation maps.
This is achieved by eagerly broadcasting the intermediate vector to the common size
of the iteration domain of the linalg op. This allows a much more natural expression of
generalized vectorization but may introduce additional computations until all the
proper canonicalizations are implemented.
This generalization modifies the vector.transfer_read/write permutation logic and
exposes the fact that the logic employed in vector.contract was too ad-hoc.
As a consequence, changes occur in the permutation / transposition logic for contraction. In turn this prompts supporting more cases in the lowering of contract
to matrix intrinsics, which is required to make the corresponding tests pass.
Differential revision: https://reviews.llvm.org/D101165
Splat constant folding was limited to `std.constant` operations. Instead, use
the constant matcher and apply splat constant folding to any constant-like
operation that holds a splat attribute.
Differential Revision: https://reviews.llvm.org/D101301
The interchange option attached to the linalg to loop lowering affects only the loops and does not update the memory accesses generated in to body of the operation. Instead of performing the interchange during the loop lowering use the interchange pattern.
Differential Revision: https://reviews.llvm.org/D100758
Example:
```
%0 = linalg.init_tensor : tensor<...>
%1 = linalg.generic ... outs(%0: tensor<...>)
%2 = linalg.generic ... outs(%0: tensor<...>)
```
Memref allocated as a result of `init_tensor` bufferization can be incorrectly overwritten by the second linalg.generic operation
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D100921
This will prevent fusion that spains all dims and generates
(d0, d1, ...) -> () reshape that isn't legal
Differential Revision: https://reviews.llvm.org/D100805
Break up the dependency between SCF ops and substituteMin helper and make a
more generic version of AffineMinSCFCanonicalization. This reduce dependencies
between linalg and SCF and will allow the logic to be used with other kind of
ops. (Like ID ops).
Differential Revision: https://reviews.llvm.org/D100321
Instead of always running the region builder check if the generalized op has a region attached. If yes inline the existing region instead of calling the region builder. This change circumvents a problem with named operations that have a region builder taking captures and the generalization pass not knowing about this captures.
Differential Revision: https://reviews.llvm.org/D100880
The patch extends the vectorization pass to lower linalg index operations to vector code. It allocates constant 1d vectors that enumerate the indexes along the iteration dimensions and broadcasts/transposes these 1d vectors to the iteration space.
Differential Revision: https://reviews.llvm.org/D100373
This patch extends the control-flow cost-model for detensoring by
implementing a forward-looking pass on block arguments that should be
detensored. This makes sure that if a (to-be-detensored) block argument
"escapes" its block through the terminator, then the successor arguments
are also detensored.
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D100457
The patch replaces the index operations in the body of fused producers and linearizes the indices after expansion.
Differential Revision: https://reviews.llvm.org/D100479
Update the dimensions of the index operations to account for dropped dimensions and replace the index operations of dropped dimensions by zero.
Differential Revision: https://reviews.llvm.org/D100395
Instead of interchanging loops during the loop lowering this pass performs the interchange by permuting the indexing maps. It also updates the iterator types and the index accesses in the body of the operation.
Differential Revision: https://reviews.llvm.org/D100627
Rationale:
Now that vector<?xindex> is allowed, the restriction on vectorization
of index types in the sparse compiler can be removed. Also needs
generalization of scatter/gather index types.
Reviewed By: gysit
Differential Revision: https://reviews.llvm.org/D100522
The patch updates the tiling pass to add the tile offsets to the indices returned by the linalg operations.
Differential Revision: https://reviews.llvm.org/D100379
The patch extends the linalg to loop lowering pass to replace all linalg index operations by the induction variables of the generated loop nests.
Differential Revision: https://reviews.llvm.org/D100364
This patch introduces the neccessary infrastructure changes to implement
cost-modelling for detensoring. In particular, it introduces the
following changes:
- An extension to the dialect conversion framework to selectively
convert sub-set of non-entry BB arguments.
- An extension to branch conversion pattern to selectively convert
sub-set of a branche's operands.
- An interface for detensoring cost-modelling.
- 2 simple implementations of 2 different cost models.
This sets the stage to explose cost-modelling for detessoring in an
easier way. We still need to come up with better cost models.
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D99945
Fusing a constant with a linalg.generic operation can result in the
fused operation being illegal since the loop bound computation
fails. Avoid such fusions.
Differential Revision: https://reviews.llvm.org/D100272
The `linalg.index` operation provides access to the iteration indexes of immediately enclosing linalg operations. It takes a dimension `dim` attribute and returns the iteration index in the given dimension. Having `linalg.index` allows us to unify `linalg.generic` and `linalg.indexed_generic` and also enables index access in named operations.
Differential Revision: https://reviews.llvm.org/D100292
Recent change enable dropping unit-trip loops of "reduction" iterator
type as well. This is fine as long as there is one other "reduction"
iterator in the operation. Without this the initialized value (value
of `out`) is not read which leads to a correctness issue.
Also fix a bug in the `fill` -> `tensor_reshape` folding. The `out`
operand of the `fill` needs to be reshaped to get the `out` operand of
the generated `fill` operation.
Differential Revision: https://reviews.llvm.org/D100145
Some sparse matrices operate on integral values (in contrast with the common
f32 and f64 values). This CL expands the compiler and runtime support to deal
with several common type combinations.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D99999
Linalg fusion on tensors has mismatching assumptions on the operand side than on the region bbArg side.
Relax the behavior on the operand/indexing map side so that we better support output operands that may also be read from.
Differential revision: https://reviews.llvm.org/D99499
Right now Elementwise operations fusion in Linalg fuses everything it
can. This can run up against resource limits of the target hardware
without some checks. This patch adds a callback function that clients
can use to implement a cost function. When two elementwise operations
are deemed structurally fusable, the callback can be used to control
if the fusion applies.
Differential Revision: https://reviews.llvm.org/D99820
The moved `populate` methods are only relevant to Linalg
operations. So they are better of in `linalg` namespace. Also rename
`populateLinalgTensorOpsFusionPatterns` to
`populateElementwiseOpsFusionPatterns`. This makes the scope of these
patterns explicit and disambiguates it with fusion on tensors using
tile + fuse.
Differential Revision: https://reviews.llvm.org/D99819
Rationale:
Small indices and values, when allowed by the required range of the
input tensors, can reduce the memory footprint of sparse tensors
even more. Note, however, that we must be careful zero extending
the values (since sparse tensors never use negatives for indexing),
but LLVM treats the index type as signed in most memory operations
(like the scatter and gather). This CL dots all the i's in this regard.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D99777
Subtensor operations that are taking a slice out of a tensor that is
unit-extent along a dimension can be rewritten to drop that dimension.
Differential Revision: https://reviews.llvm.org/D99226
Drop usage of `emitRemark` and use `notifyMatchFailure` instead to
avoid unnecessary spew during compilation.
Differential Revision: https://reviews.llvm.org/D99485
init tensor operands also has indexing map and generally follow
the same constraints we expect for non-init-tensor operands.
Differential Revision: https://reviews.llvm.org/D99115
This commit exposes an option to the pattern
FoldWithProducerReshapeOpByExpansion to allow
folding unit dim reshapes. This gives callers
more fine-grained controls.
Differential Revision: https://reviews.llvm.org/D99114
Until now Linalg fusion only allow fusing producers whose operands
are all permutation indexing maps. It's easier to deduce the
subtensor/subview but it is an unnecessary constraint, as in tiling
we have more advanced logic to deduce the subranges even when the
operand is not of permutation indexing maps, e.g., the input operand
for convolution ops.
This patch uses the logic on tiling side to deduce subranges for
fusion. This enables fusing convolution with its consumer ops
when possible.
Along the way, we are now generating proper affine.min ops to guard
against size boundaries, if we cannot be certain they won't be
out of bounds.
Differential Revision: https://reviews.llvm.org/D99014
This is a preparation step to reuse makeTiledShapes in tensor
fusion. Along the way, did some lightweight cleanups.
Differential Revision: https://reviews.llvm.org/D99013
All linalg operations having a region builder shall call it during op creation. Calling it during vectorization is obsolete.
Differential Revision: https://reviews.llvm.org/D99168
Fix the BlockAndValueMapping update that was missing entries for scf.for op's blockIterArgs.
Skip cloning subtensors of the padded tensor as the logic for these is separate.
Add a filter to drop side-effecting ops.
Tests are beefed up to verify the IR is sound in all hoisting configurations for 2-level 3-D tiled matmul.
Differential Revision: https://reviews.llvm.org/D99255
To match an interface or trait, users currently have to use the `MatchAny` tag. This tag can be quite problematic for compile time for things like the canonicalizer, as the `MatchAny` patterns may get applied to *every* operation. This revision adds better support by bucketing interface/trait patterns based on which registered operations have them registered. This means that moving forward we will only attempt to match these patterns to operations that have this interface registered. Two simplify defining patterns that match traits and interfaces, two new utility classes have been added: OpTraitRewritePattern and OpInterfaceRewritePattern.
Differential Revision: https://reviews.llvm.org/D98986
This revision introduces proper backward slice computation during the hoisting of
PadTensorOp. This allows hoisting padding even across multiple levels of tiling.
Such hoisting requires the proper handling of loop bounds that may depend on enclosing
loop variables.
Differential revision: https://reviews.llvm.org/D98965
This nicely aligns the naming with RewritePatternSet. This type isn't
as widely used, but we keep a using declaration in to help with
downstream consumption of this change.
Differential Revision: https://reviews.llvm.org/D99131
This doesn't change APIs, this just cleans up the many in-tree uses of these
names to use the new preferred names. We'll keep the old names around for a
couple weeks to help transitions.
Differential Revision: https://reviews.llvm.org/D99127
- Drop unnecessary occurrences of rewriter.eraseOp: dead linalg ops on tensors should be cleaned up by DCE.
- reimplement the part of Linalg on fusion that constructs the body and block arguments: the previous implementation had too much magic. Instead this spells out all cases explicitly and asserts / introduces TODOs for incorrect cases.
As a consequence, we can use the default traversal order for this pattern.
Differential Revision: https://reviews.llvm.org/D99070
GreedyPatternRewriteDriver was changed from bottom-up traversal to top-down traversal. Not all passes work yet with that change for traversal order. To give some time for fixing, add an option to allow to switch back to bottom-up traversal. Use this option in FusionOfTensorOpsPass which fails otherwise.
Differential Revision: https://reviews.llvm.org/D99059
This updates the codebase to pass the context when creating an instance of
OwningRewritePatternList, and starts removing extraneous MLIRContext
parameters. There are many many more to be removed.
Differential Revision: https://reviews.llvm.org/D99028
This reverts commit 32a744ab20.
CI is broken:
test/Dialect/Linalg/bufferize.mlir:274:12: error: CHECK: expected string not found in input
// CHECK: %[[MEMREF:.*]] = tensor_to_memref %[[IN]] : memref<?xf32>
^
`BufferizeAnyLinalgOp` fails because `FillOp` is not a `LinalgGenericOp` and it fails while reading operand sizes attribute.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D98671
This is a temporary work-around to get our all-annotations-all-flags
stress testing effort run clean. In the long run, we want to provide
efficient implementations of strided loads and stores though
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D98563
It is to use the methods in LinalgInterfaces.cpp for additional static shape verification to match the shaped operands and loop on linalgOps. If I used the existing methods, I would face circular dependency linking issue. Now we can use them as methods of LinalgOp.
Reviewed By: hanchung
Differential Revision: https://reviews.llvm.org/D98163
Return the vectorization results using a vector passed by reference instead of returning them embedded in a structure.
Differential Revision: https://reviews.llvm.org/D98182
Reduction updates should be masked, just like the load and stores.
Note that alternatively, we could use the fact that masked values are
zero of += updates and mask invariants to get this working but that
would not work for *= updates. Masking the update itself is cleanest.
This change also replaces the constant mask with a broadcast of "true"
since this constant folds much better for various folding patterns.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D98000
Found with exhaustive testing, it is possible that a while loop
appears in between chainable for loops. As long as we don't
scalarize reductions in while loops, this means we need to
terminate the chain at the while. This also refactors the
reduction code into more readable helper methods.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D97886
Some elementwise operations are not scalarizable, vectorizable, or tensorizable.
Split `ElementwiseMappable` trait into the following, more precise traits.
- `Elementwise`
- `Scalarizable`
- `Vectorizable`
- `Tensorizable`
This allows for reuse of `Elementwise` in dialects like HLO.
Differential Revision: https://reviews.llvm.org/D97674
This patch continues detensorizing implementation by detensoring
internal control flow in functions.
In order to detensorize functions, all the non-entry block's arguments
are detensored and branches between such blocks are properly updated to
reflect the detensored types as well. Function entry block (signature)
is left intact.
This continues work towards handling github/google/iree#1159.
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D97148
The universal index was maintained if dense indices were still
in place, and lattice points followed. However, it should only
be kept if any of those following lattice points actually
consumes the universal index. This change also fixes an
inaccuracy with a missing broadcast around vector invariant.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D97594
Similar to mask-load/store and compress/expand, the gather and
scatter operation now allow for higher dimension uses. Note that
to support the mixed-type index, the new syntax is:
vector.gather %base [%i,%j] [%kvector] ....
The first client of this generalization is the sparse compiler,
which needs to define scatter and gathers on dense operands
of higher dimensions too.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D97422
When computing dense address, a vectorized index must be accounted
for properly. This bug was formerly undetected because we get 0 * prev + i
in most cases, which folds away the scalar part. Now it works for all cases.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D97317
This transformation was only used for quick experimentation and is not general enough.
Retire it.
Differential Revision: https://reviews.llvm.org/D97266
This commit is the first baby step towards detensoring in
linalg-on-tensors.
Detensoring is the process through which a tensor value is convereted to one
or potentially more primitive value(s). During this process, operations with
such detensored operands are also converted to an equivalen form that works
on primitives.
The detensoring process is driven by linalg-on-tensor ops. In particular, a
linalg-on-tensor op is checked to see whether *all* its operands can be
detensored. If so, those operands are converted to thier primitive
counterparts and the linalg op is replaced by an equivalent op that takes
those new primitive values as operands.
This works towards handling github/google/iree#1159.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D96271
Simplifies the way lattices are optimized with less, but more
powerful rules. This also fixes an inaccuracy where too many
lattices resulted (expecting a non-existing universal index).
Also puts no-side-effects on all proper getters and unifies
bufferization flags order in integration tests (for future,
more complex use cases).
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D97134
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 8aa6c3765b.
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
Rationale:
Narrower types for overhead storage yield a smaller memory footprint for
sparse tensors and thus needs to be supported. Also, more value types
need to be supported to deal with all kinds of kernels. Since the
"one-size-fits-all" sparse storage scheme implementation is used
instead of actual codegen, the library needs to be able to support
all combinations of desired types. With some crafty templating and
overloading, the actual code for this is kept reasonably sized though.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D96819
This revision adds support for hoisting "subtensor + vector.transfer_read" / "subtensor_insert + vector.transfer_write pairs" across scf.for.
The unit of hoisting becomes a HoistableRead / HoistableWrite struct which contains a pair of "vector.transfer_read + optional subtensor" / "vector.transfer_write + optional subtensor_insert".
scf::ForOp canonicalization patterns are applied greedily on the successful application of the transformation to cleanup the IR more eagerly and potentially expose more transformation opportunities.
Differential revision: https://reviews.llvm.org/D96731
SliceAnalysis originally was developed in the context of affine.for within mlfunc.
It predates the notion of region.
This revision updates it to not hardcode specific ops like scf::ForOp.
When rooted at an op, the behavior of the slice computation changes as it recurses into the regions of the op. This does not support gathering all values transitively depending on a loop induction variable anymore.
Additional variants rooted at a Value are added to also support the existing behavior.
Differential revision: https://reviews.llvm.org/D96702
This revision takes advantage of the newly extended `ref` directive in assembly format
to allow better region handling for LinalgOps. Specifically, FillOp and CopyOp now build their regions explicitly which allows retiring older behavior that relied on specific op knowledge in both lowering to loops and vectorization.
This reverts commit 3f22547fd1 and reland 973e133b76 with a workaround for
a gcc bug that does not accept lambda default parameters:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59949
Differential Revision: https://reviews.llvm.org/D96598
This reverts commit 973e133b76.
It triggers an issue in gcc5 that require investigation, the build is
broken with:
/tmp/ccdpj3B9.s: Assembler messages:
/tmp/ccdpj3B9.s:5821: Error: symbol `_ZNSt17_Function_handlerIFvjjEUljjE2_E9_M_invokeERKSt9_Any_dataOjS6_' is already defined
/tmp/ccdpj3B9.s:5860: Error: symbol `_ZNSt14_Function_base13_Base_managerIUljjE2_E10_M_managerERSt9_Any_dataRKS3_St18_Manager_operation' is already defined
This revision takes advantage of the newly extended `ref` directive in assembly format
to allow better region handling for LinalgOps. Specifically, FillOp and CopyOp now build their regions explicitly which allows retiring older behavior that relied on specific op knowledge in both lowering to loops and vectorization.
Differential Revision: https://reviews.llvm.org/D96598
The AffineMap in the MemRef inferred by SubViewOp may have uncompressed symbols which result in type mismatch on otherwise unused symbols. Make the computation of the AffineMap compress those unused symbols which results in better canonical types.
Additionally, improve the error message to report which inferred type was expected.
Differential Revision: https://reviews.llvm.org/D96551
The dimension order of a filter in tensorflow is
[filter_height, filter_width, in_channels, out_channels], which is different
from current definition. The current definition follows TOSA spec. Add TF
version conv ops to .tc, so we do not have to insert a transpose op around a
conv op.
Reviewed By: antiagainst
Differential Revision: https://reviews.llvm.org/D96038
This revision connects the generated sparse code with an actual
sparse storage scheme, which can be initialized from a test file.
Lacking a first-class citizen SparseTensor type (with buffer),
the storage is hidden behind an opaque pointer with some "glue"
to bring the pointer back to tensor land. Rather than generating
sparse setup code for each different annotated tensor (viz. the
"pack" methods in TACO), a single "one-size-fits-all" implementation
has been added to the runtime support library. Many details and
abstractions need to be refined in the future, but this revision
allows full end-to-end integration testing and performance
benchmarking (with on one end, an annotated Lingalg
op and, on the other end, a JIT/AOT executable).
Reviewed By: nicolasvasilache, bixia
Differential Revision: https://reviews.llvm.org/D95847
This revision fixes the indexing logic into the packed tensor that result from hoisting padding. Previously, the index was incorrectly set to the loop induction variable when in fact we need to compute the iteration count (i.e. `(iv - lb).ceilDiv(step)`).
Differential Revision: https://reviews.llvm.org/D96417
This revision fixes the fact that the padding transformation did not have enough information to set the proper type for the padding value.
Additionally, the verifier for Yield in the presence of PadTensorOp is fixed to properly report incorrect number of results or operands. Previously, the error would be silently ignored which made the core issue difficult to debug.
Differential Revision: https://reviews.llvm.org/D96264
This reverts commit 511dd4f438 along with
a couple fixes.
Original message:
Now the context is the first, rather than the last input.
This better matches the rest of the infrastructure and makes
it easier to move these types to being declaratively specified.
Phabricator: https://reviews.llvm.org/D96111
Now the context is the first, rather than the last input.
This better matches the rest of the infrastructure and makes
it easier to move these types to being declaratively specified.
Differential Revision: https://reviews.llvm.org/D96111
This makes ignoring a result explicit by the user, and helps to prevent accidental errors with dropped results. Marking LogicalResult as no discard was always the intention from the beginning, but got lost along the way.
Differential Revision: https://reviews.llvm.org/D95841
This revision takes advantage of recent extensions to vectorization to refactor contraction detection into a bona fide Linalg interface.
The mlit-linalg-ods-gen parser is extended to support adding such interfaces.
The detection that was originally enabling vectorization is refactored to serve as both a test on a generic LinalgOp as well as to verify ops that declare to conform to that interface.
This is plugged through Linalg transforms and strategies but it quickly becomes evident that the complexity and rigidity of the C++ class based templating does not pay for itself.
Therefore, this revision changes the API for vectorization patterns to get rid of templates as much as possible.
Variadic templates are relegated to the internals of LinalgTransformationFilter as much as possible and away from the user-facing APIs.
It is expected other patterns / transformations will follow the same path and drop as much C++ templating as possible from the class definition.
Differential revision: https://reviews.llvm.org/D95973
This revision defines a Linalg contraction in general terms:
1. Has 2 input and 1 output shapes.
2. Has at least one reduction dimension.
3. Has only projected permutation indexing maps.
4. its body computes `u5(u1(c) + u2(u3(a) * u4(b)))` on some field
(AddOpType, MulOpType), where u1, u2, u3, u4 and u5 represent scalar unary
operations that may change the type (e.g. for mixed-precision).
As a consequence, when vectorization of such an op occurs, the only special
behavior is that the (unique) MulOpType is vectorized into a
`vector.contract`. All other ops are handled in a generic fashion.
In the future, we may wish to allow more input arguments and elementwise and
constant operations that do not involve the reduction dimension(s).
A test is added to demonstrate the proper vectorization of matmul_i8_i8_i32.
Differential revision: https://reviews.llvm.org/D95939
This revision unifies Linalg vectorization and paves the way for vectorization of Linalg ops with mixed-precision operations.
The new algorithm traverses the ops in the linalg block in order and avoids recursion.
It uses a BlockAndValueMapping to keep track of vectorized operations.
The revision makes the following modifications but is otherwise NFC:
1. vector.transfer_read are created eagerly and may appear in a different order than the original order.
2. a more progressive vectorization to vector.contract results in only the multiply operation being converted to `vector.contract %a, %b, %zero`, where `%zero` is a
constant of the proper type. Later vector canonicalizations are assumed to rewrite vector.contract %a, %b, %zero + add to a proper accumulate form.
Differential revision: https://reviews.llvm.org/D95797
This is the last revision to migrate using SimplePadOp to PadTensorOp, and the
SimplePadOp is removed in the patch. Update a bit in SliceAnalysis because the
PadTensorOp takes a region different from SimplePadOp. This is not covered by
LinalgOp because it is not a structured op.
Also, remove a duplicated comment from cpp file, which is already described in a
header file. And update the pseudo-mlir in the comment.
This is as same as D95615 but fixing one dep in CMakeLists.txt
Different from D95671, the fix was applied to run target.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D95785
This reverts commit d9b953d84b.
This commit resulted in build bot failures and the author is away from a
computer, so I am reverting on their behalf until they have a chance to
look into this.
This is the last revision to migrate using SimplePadOp to PadTensorOp, and the
SimplePadOp is removed in the patch. Update a bit in SliceAnalysis because the
PadTensorOp takes a region different from SimplePadOp. This is not covered by
LinalgOp because it is not a structured op.
Also, remove a duplicated comment from cpp file, which is already described in a
header file. And update the pseudo-mlir in the comment.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D95671
This is the last revision to migrate using SimplePadOp to PadTensorOp, and the
SimplePadOp is removed in the patch. Update a bit in SliceAnalysis because the
PadTensorOp takes a region different from SimplePadOp. This is not covered by
LinalgOp because it is not a structured op.
Also, remove a duplicated comment from cpp file, which is already described in a
header file. And update the pseudo-mlir in the comment.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D95615
This revision creates a build method of PadTensorOp which can be mapped to
SimplePad op. The verifier is updated to accept a static custom result type,
which has the same semantic as SimplePadOp.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D95555
This revision adds a layer of SFINAE to the composable codegen strategy so it does
not have to require statically defined ops but instead can also be used with OpInterfaces, Operation* and an op name string.
A linalg.matmul_i8_i8_i32 is added to the .tc spec to demonstrate how all this works end to end.
Differential Revision: https://reviews.llvm.org/D95600
This revision improves the usage of the codegen strategy by adding a few flags that
make it easier to control for the CLI.
Usage of ModuleOp is replaced by FuncOp as this created issues in multi-threaded mode.
A simple benchmarking capability is added for linalg.matmul as well as linalg.matmul_column_major.
This latter op is also added to linalg.
Now obsolete linalg integration tests that also take too long are deleted.
Correctness checks are still missing at this point.
Differential revision: https://reviews.llvm.org/D95531
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
This revision addresses a remaining comment that was overlooked in https://reviews.llvm.org/D95243:
the pad hoisting transformation is made to additionally bail out on side effecting ops other than LoopLikeOps.
This transformation anchors on a padding op whose result is only used as an input
to a Linalg op and pulls it out of a given number of loops.
The result is a packing of padded tailes of ops that is amortized just before
the outermost loop from which the pad operation is hoisted.
Differential revision: https://reviews.llvm.org/D95243
This revision allows the base Linalg tiling pattern to optionally require padding to
a constant bounding shape.
When requested, a simple analysis is performed, similar to buffer promotion.
A temporary `linalg.simple_pad` op is added to model padding for the purpose of
connecting the dots. This will be replaced by a more fleshed out `linalg.pad_tensor`
op when it is available.
In the meantime, this temporary op serves the purpose of exhibiting the necessary
properties required from a more fleshed out pad op, to compose with transformations
properly.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D95149
Fusion of generic/indexed_generic operations with tensor_reshape by
expansion when the latter just adds/removes unit-dimensions is
disabled since it just adds unit-trip count loops.
Differential Revision: https://reviews.llvm.org/D94626
representing dependence from producer result to consumer.
With Linalg on tensors the dependence between operations can be from
the result of the producer to the consumer. This change just does a
NFC refactoring of the LinalgDependenceGraphElem to allow representing
both OpResult and OpOperand*.
Differential Revision: https://reviews.llvm.org/D95208
Use cases with 16- or even 8-bit pointer/index structures have been identified.
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D95015
This is a very minor improvement during iteration graph construction.
If the first attempt considering the dimension order of all tensors fails,
a second attempt is made using the constraints of sparse tensors only.
Dense tensors prefer dimension order (locality) but provide random access
if needed, enabling the compilation of more sparse kernels.
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D94709
With the recent changes to linalg on tensor semantics, the tiling
operations works out-of-the-box for generic operations. Add a test to
verify that and some minor refactoring.
Differential Revision: https://reviews.llvm.org/D93077
Similar to the parallelization strategies, the vectorization strategies
provide control on what loops should be vectorize. Unlike the parallel
strategies, only innermost loops are considered, but including reductions,
with the control of vectorizing dense loops only or dense and sparse loops.
The vectorized loops are always controlled by a vector mask to avoid
overrunning the iterations, but subsequent vector operation folding removes
redundant masks and replaces the operations with more efficient counterparts.
Similarly, we will rely on subsequent loop optimizations to further optimize
masking, e.g. using an unconditional full vector loop and scalar cleanup loop.
The current strategy already demonstrates a nice interaction between the
sparse compiler and all prior optimizations that went into the vector dialect.
Ongoing discussion at:
https://llvm.discourse.group/t/mlir-support-for-sparse-tensors/2020/10
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D94551
This revision uniformizes fusion APIs to allow passing OpOperand, OpResult and adds a finer level of control fusion.
Differential Revision: https://reviews.llvm.org/D94493
getDynOperands behavior is commonly used in a number of passes. Refactored to
use a helper function and avoid code reuse.
Differential Revision: https://reviews.llvm.org/D94340
When fusing tensor_reshape ops with generic/indexed_Generic op, new
linalg.init_tensor operations were created for the `outs` of the fused
op. While correct (technically) it is better to just reshape the
original `outs` operands and rely on canonicalization of init_tensor
-> tensor_reshape to achieve the same effect.
Differential Revision: https://reviews.llvm.org/D93774
Linalg ops are perfect loop nests. When materializing the concrete
loop nest, the default order specified by the Linalg op's iterators
may not be the best for further CodeGen: targets frequently need
to plan the loop order in order to gain better data access. And
different targets can have different preferences. So there should
exist a way to control the order.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D91795
Change the implementation of LinalgOp with TensorReshapeOp by
expansion to be more modular and easier to follow.
Differential Revision: https://reviews.llvm.org/D93748
Add same hoisting transformation existing for transfer ops on buffers for
transfer_ops on tensor. The logic is significantly different so this is done as
a separate transformation and it is expect that user would know which
transformation to use based on the flow.
Differential Revision: https://reviews.llvm.org/D94115
Nicolas changed the tensor abstraction so that every output has
its own shape definition. This simplifies the "inference" that
was used in the sparse compiler.
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D94119
This revision drops init_tensor arguments from Linalg on tensors and instead uniformizes the output buffers and output tensors to be consistent.
This significantly simplifies the usage of Linalg on tensors and is a stepping stone for
its evolution towards a mixed tensor and shape abstraction discussed in https://llvm.discourse.group/t/linalg-and-shapes/2421/19.
Differential Revision: https://reviews.llvm.org/D93469
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
Reductions in innermost loops become harder for the backend to disambiguate
after bufferization into memrefs, resulting in less efficient load-update-store
cycles. By scalarizing innermost reductions, the backend is more likely to assign
a register to perform the reduction (also prepares vectorization). Even though
we could scalarize reductions for more outer loops and while-loops as well,
currently scalarization is only done for chains of innermost for-loops, where
it matters most, to avoid complicating codegen unnecessary (viz. adding lots
of yield instructions).
This CL also refactors condition simplification into the merger class,
where it belongs, so that conditions are simplified only once per loop
nest and not repeatedly as was currently done. This CL also fixes a few
minor bugs, some layout issues, and comments.
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D93143
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
This is useful for scalar code that uses for/while loops.
This has also been confirmed to work for representing std.pow as an
scf.for loop on gpus.
Differential Revision: https://reviews.llvm.org/D93308
Fix a bug causing to pick the wrong vector size to broadcast to when the source
vectors have different ranks.
Differential Revision: https://reviews.llvm.org/D93118
After bufferization, the backend has much more trouble hoisting loop invariant
loads from the loops generated by the sparse compiler. Therefore, this is done
during sparse code generation. Note that we don't bother hoisting derived
invariant expressions on SSA values, since the backend does that very well.
Still TBD: scalarize reductions to avoid load-add-store cycles
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D92534
In the past, the reshape op can be folded only if the indexing map is
permutation in consumer's usage. We can relax to condition to be projected
permutation.
This patch still limits the fusion for scalar cases. Scalar case is a corner
case, because we need to decide where to put extra dims.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D92466
Add support for vectorization for linalg.generic representing element-wise ops.
Those are converted to transfer_read + vector ops + transfer_write.
Also re-organize the vectorization tests to be together.
Implementation derived from the work of @burmako, @agrue and
@fedelebron.
Differential Revision: https://reviews.llvm.org/D92540
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
This change gives sparse compiler clients more control over selecting
individual types for the pointers and indices in the sparse storage schemes.
Narrower width obviously results in smaller memory footprints, but the
range should always suffice for the maximum number of entries or index value.
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D92126
This CL adds the ability to request different parallelization strategies
for the generate code. Every "parallel" loop is a candidate, and converted
to a parallel op if it is an actual for-loop (not a while) and the strategy
allows dense/sparse outer/inner parallelization.
This will connect directly with the work of @ezhulenev on parallel loops.
Still TBD: vectorization strategy
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D91978
Generalizes invariant handling to anything defined outside the Linalg op
(parameters and SSA computations). Fixes bug that was using parameter number
as tensor number.
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D91985
Print part of an op of the form:
```
<optional-offset-prefix>`[` offset-list `]`
<optional-size-prefix>`[` size-list `]`
<optional-stride-prefix>[` stride-list `]`
```
Also address some leftover nits.
Differential revision: https://reviews.llvm.org/D92031
Exposing some utility functions from Linalg to allow for promotion of
fused views outside of the core tile+fuse logic.
This is an alternative to patch D91322 which adds the promotion logic
to the tileAndFuse method. Downside with that approach is that it is
not easily customizable based on needs.
Differential Revision: https://reviews.llvm.org/D91503
Enhance the tile+fuse logic to allow fusing a sequence of operations.
Make sure the value used to obtain tile shape is a
SubViewOp/SubTensorOp. Current logic used to get the bounds of loop
depends on the use of `getOrCreateRange` method on `SubViewOp` and
`SubTensorOp`. Make sure that the value/dim used to compute the range
is from such ops. This fix is a reasonable WAR, but a btter fix would
be to make `getOrCreateRange` method be a method of `ViewInterface`.
Differential Revision: https://reviews.llvm.org/D90991
This revision refactors code used in various Linalg transformations and makes it a first class citizen to the LinalgStructureOpInterface. This is in preparation to allowing more advanced Linalg behavior but is otherwise NFC.
Differential revision: https://reviews.llvm.org/D91863
Adds tests for full sum reduction (tensors summed up into scalars)
and the well-known sampled-dense-dense-matrix-product. Refines
the optimizations rules slightly to handle the summation better.
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D91818
Add transformation to be able to forward transfer_write into transfer_read
operation and to be able to remove dead transfer_write when a transfer_write is
overwritten before being read.
Differential Revision: https://reviews.llvm.org/D91321
This reverts commit f8284d21a8.
Revert "[mlir][Linalg] NFC: Expose some utility functions used for promotion."
This reverts commit 0c59f51592.
Revert "Remove unused isZero function"
This reverts commit 0f9f0a4046.
Change f8284d21 led to multiple failures in IREE compilation.
Exposing some utility functions from Linalg to allow for promotion of
fused views outside of the core tile+fuse logic.
This is an alternative to patch D91322 which adds the promotion logic
to the tileAndFuse method. Downside with that approach is that it is
not easily customizable based on needs.
Differential Revision: https://reviews.llvm.org/D91503
This commit starts a new pass and patterns for converting Linalg
named ops to generic ops. This enables us to leverage the flexbility
from generic ops during transformations. Right now only linalg.conv
is supported; others will be added when useful.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D91357
Rationale:
Make sure preconditions are tested already during verfication.
Currently, the only way a sparse rewriting rule can fail is if
(1) the linalg op does not have sparse annotations, or
(2) a yet to be handled operation is encounted inside the op
Reviewed By: penpornk
Differential Revision: https://reviews.llvm.org/D91748
As discussed in https://llvm.discourse.group/t/mlir-support-for-sparse-tensors/2020
this CL is the start of sparse tensor compiler support in MLIR. Starting with a
"dense" kernel expressed in the Linalg dialect together with per-dimension
sparsity annotations on the tensors, the compiler automatically lowers the
kernel to sparse code using the methods described in Fredrik Kjolstad's thesis.
Many details are still TBD. For example, the sparse "bufferization" is purely
done locally since we don't have a global solution for propagating sparsity
yet. Furthermore, code to input and output the sparse tensors is missing.
Nevertheless, with some hand modifications, the generated MLIR can be
easily converted into runnable code already.
Reviewed By: nicolasvasilache, ftynse
Differential Revision: https://reviews.llvm.org/D90994
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
This removes the need to have an explicit `cast<>` given that we always know it `isa` instance of the interface.
Differential Revision: https://reviews.llvm.org/D91304
We lower them to a std.global_memref (uniqued by constant value) + a
std.get_global_memref to produce the corresponding memref value.
This allows removing Linalg's somewhat hacky lowering of tensor
constants, now that std properly supports this.
Differential Revision: https://reviews.llvm.org/D91306
It was incorrect in the presence of a tensor argument with multiple
uses.
The bufferization of subtensor_insert was writing into a converted
memref operand, but there is no guarantee that the converted memref for
that operand is safe to write into. In this case, the same converted
memref is written to in-place by the subtensor_insert bufferization,
violating the tensor-level semantics.
I left some comments in a TODO about ways forward on this. I will be
working actively on this problem in the coming days.
Differential Revision: https://reviews.llvm.org/D91371
This change does two main things
1) An operation might have multiple dependences to the same
producer. Not tracking them correctly can result in incorrect code
generation with fusion. To rectify this the dependence tracking
needs to also have the operand number in the consumer.
2) Improve the logic used to find the fused loops making it easier to
follow. The only constraint for fusion is that linalg ops (on
buffers) have update semantics for the result. Fusion should be
such that only one iteration of the fused loop (which is also a
tiled loop) must touch only one (disjoint) tile of the output. This
could be relaxed by allowing for recomputation that is the default
when oeprands are tensors, or can be made legal with promotion of
the fused view (in future).
Differential Revision: https://reviews.llvm.org/D90579
This CL integrates the new sparse annotations (hereto merely added as fully
transparent attributes) more tightly to the generic linalg op in order to add
verification of the annotations' consistency as well as to make make other
passes more aware of their presence (in the long run, rewriting rules must
preserve the integrity of the annotations).
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D91224
This patch converts elementwise ops on tensors to linalg.generic ops
with the same elementwise op in the payload (except rewritten to
operate on scalars, obviously). This is a great form for later fusion to
clean up.
E.g.
```
// Compute: %arg0 + %arg1 - %arg2
func @f(%arg0: tensor<?xf32>, %arg1: tensor<?xf32>, %arg2: tensor<?xf32>) -> tensor<?xf32> {
%0 = addf %arg0, %arg1 : tensor<?xf32>
%1 = subf %0, %arg2 : tensor<?xf32>
return %1 : tensor<?xf32>
}
```
Running this through
`mlir-opt -convert-std-to-linalg -linalg-fusion-for-tensor-ops` we get:
```
func @f(%arg0: tensor<?xf32>, %arg1: tensor<?xf32>, %arg2: tensor<?xf32>) -> tensor<?xf32> {
%0 = linalg.generic {indexing_maps = [#map0, #map0, #map0, #map0], iterator_types = ["parallel"]} ins(%arg0, %arg1, %arg2 : tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) {
^bb0(%arg3: f32, %arg4: f32, %arg5: f32): // no predecessors
%1 = addf %arg3, %arg4 : f32
%2 = subf %1, %arg5 : f32
linalg.yield %2 : f32
} -> tensor<?xf32>
return %0 : tensor<?xf32>
}
```
So the elementwise ops on tensors have nicely collapsed into a single
linalg.generic, which is the form we want for further transformations.
Differential Revision: https://reviews.llvm.org/D90354
Previously, linalg-bufferize was a "finalizing" bufferization pass (it
did a "full" conversion). This wasn't great because it couldn't be used
composably with other bufferization passes like std-bufferize and
scf-bufferize.
This patch makes linalg-bufferize a composable bufferization pass.
Notice that the integration tests are switched over to using a pipeline
of std-bufferize, linalg-bufferize, and (to finalize the conversion)
func-bufferize. It all "just works" together.
While doing this transition, I ran into a nasty bug in the 1-use special
case logic for forwarding init tensors. That logic, while
well-intentioned, was fundamentally flawed, because it assumed that if
the original tensor value had one use, then the converted memref could
be mutated in place. That assumption is wrong in many cases. For
example:
```
%0 = some_tensor : tensor<4xf32>
br ^bb0(%0, %0: tensor<4xf32>, tensor<4xf32>)
^bb0(%bbarg0: tensor<4xf32>, %bbarg1: tensor<4xf32>)
// %bbarg0 is an alias of %bbarg1. We cannot safely write
// to it without analyzing uses of %bbarg1.
linalg.generic ... init(%bbarg0) {...}
```
A similar example can happen in many scenarios with function arguments.
Even more sinister, if the converted memref is produced by a
`std.get_global_memref` of a constant global memref, then we might
attempt to write into read-only statically allocated storage! Not all
memrefs are writable!
Clearly, this 1-use check is not a local transformation that we can do
on the fly in this pattern, so I removed it.
The test is now drastically shorter and I basically rewrote the CHECK
lines from scratch because:
- the new composable linalg-bufferize just doesn't do as much, so there
is less to test
- a lot of the tests were related to the 1-use check, which is now gone,
so there is less to test
- the `-buffer-hoisting -buffer-deallocation` is no longer mixed in, so
the checks related to that had to be rewritten
Differential Revision: https://reviews.llvm.org/D90657
Previously, they were only defined for `FuncOp`.
To support this, `FunctionLike` needs a way to get an updated type
from the concrete operation. This adds a new hook for that purpose,
called `getTypeWithoutArgsAndResults`.
For now, `FunctionLike` continues to assume the type is
`FunctionType`, and concrete operations that use another type can hide
the `getType`, `setType`, and `getTypeWithoutArgsAndResults` methods.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D90363
The bufferization patterns are moved to the .cpp file, which is
preferred in the codebase when it makes sense.
The LinalgToStandard patterns are kept a header because they are
expected to be used individually. However, they are moved to
LinalgToStandard.h which is the file corresponding to where they are
defined.
This also removes TensorCastOpConverter, which is handled by
populateStdBufferizePatterns now. Eventually, the constant op lowering
will be handled as well, but it there are currently holdups on moving
it (see https://reviews.llvm.org/D89916).
Differential Revision: https://reviews.llvm.org/D90254
Linalg "tile-and-fuse" is currently exposed as a Linalg pass "-linalg-fusion" but only the mechanics of the transformation are currently relevant.
Instead turn it into a "-test-linalg-greedy-fusion" pass which performs canonicalizations to enable more fusions to compose.
This allows dropping the OperationFolder which is not meant to be used with the pattern rewrite infrastructure.
Differential Revision: https://reviews.llvm.org/D90394
This patch adds support for fusing linalg.indexed_generic op with
linalg.tensor_reshape op by expansion, i.e.
- linalg.indexed_generic op -> linalg.tensor_reshape op when the
latter is expanding.
- linalg.tensor_reshape op -> linalg.indexed_generic op when the
former is folding.
Differential Revision: https://reviews.llvm.org/D90082
MaheshRavishankar