There was a bug in scf.for loop bufferization that could lead to a missing buffer copy (alloc was there, but not the copy).
Differential Revision: https://reviews.llvm.org/D135053
`getDestinationOperands` was almost a duplicate of `DestinationStyleOpInterface::getOutputOperands`. Now that the interface has been moved to mlir/Interfaces, it is no longer needed.
Differential Revision: https://reviews.llvm.org/D136240
scf.for used to only support 1:1 type conversion, this patch add support for 1:N type conversion.
Reviewed By: aartbik
Differential Revision: https://reviews.llvm.org/D136314
Context: https://discourse.llvm.org/t/psa-retire-linalg-filter-based-patterns/63785
In the process, also retire `tileConsumerAndFuseProducers` that is now replaced by `tileConsumerAndFuseProducerGreedilyUsingSCFForOp`.
Context: https://discourse.llvm.org/t/psa-retire-tileandfuselinalgops-method/63850
When performing this replacement, a change of behavior appeared: the older `tileConsumerAndFuseProducers` would split the parallel
and non-parallel dimensions automatically and perform a first level of tile-and-fuse on parallel dimensions only and then introduce a
second level of tiling-only on the reduction dimensions. The newer `tileConsumerAndFuseProducerGreedilyUsingSCFForOp` on the other hand
does not perform this breakdown. As a consequence, the transform specification is evolved to produce the same output.
Additionally, replace some uses of `unsigned` by `int64_t` where possible without pulling in larger interface changes (left for a future PR).
Context: https://www.youtube.com/watch?v=Puio5dly9N8
Lastly, tests that were performing tile and fuse and distribute on tensors are retired: the generated IR mixing scf.for, tensors and
distributed processor ids was racy at best ..
Differential Revision: https://reviews.llvm.org/D135559
This patch refactors the tiling and tile + fuse implementation using
`TilingInterface`. Primarily, it exposes the functionality as simple
utility functions instead of as a Pattern to allow calling it from a
pattern as it is done in the test today or from within the transform
dialect (in the future). This is a step towards deprecating similar
methods in Linalg dialect.
- The utility methods do not erase the root operations.
- The return value provides the values to use for replacements.
Differential Revision: https://reviews.llvm.org/D134144
The current approach for handling `iter_args` was to replace all uses
of the value that is used as `init` value with the corresponding
region block argument within the `scf.for`. This is not always
correct. Instead a more deliberate approach needs to be taken to
handle these. If the slice being fused represents a slice of the
destination operand of the untiled op, then
- Make the destination of the fused producer the `init` value of the
loop nest
- For the tiled and fused producer op created, replace the slice of
the destination operand with a slice of the corresponding region
iter arg of the innermost loop of the generated loop nest
Differential Revision: https://reviews.llvm.org/D134411
This change allows the SCF LoopPipelining transform to handle ops with
nested regions within the pipelined `scf.for` body. The op and nested
regions are treated as a single unit from the transform's perspective.
This change also makes explicit the requirement that only ops whose
parent Block is the loop body Block are allowed to be scheduled by the
caller.
Reviewed By: ThomasRaoux, nicolasvasilache
Differential Revision: https://reviews.llvm.org/D133965
Currently, if the `before` and `after` regions of a while op have
tensor args in different indices, this leads to a crash.
This moves the pass-through check for args to the handling of the
condition block, since that is where the results are produced, so
it's also where copies must be made.
Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D133477
Currently, one-shot-bufferize crashes as soon as there's
a mixture of tensor and non-tensor arguments. This seems
to happen for no good reason.
Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D133419
This change refines the semantics of scf.foreach_thread. Tensors that are inserted into in the terminator must now be passed to the region explicitly via `shared_outs`. Inside of the body of the op, those tensors are then accessed via block arguments.
The body of a scf.foreach_thread is now treated as a repetitive region. I.e., op dominance can no longer be used in conflict detection when using a value that is defined outside of the body. Such uses may now be considered as conflicts (if there is at least one read and one write in the body), effectively privatizing the tensor. Shared outputs are not privatized when they are used via their corresponding block arguments.
As part of this change, it was also necessary to update the "tiling to scf.foreach_thread", such that the generated tensor.extract_slice ops use the scf.foreach_thread's block arguments. This is implemented by cloning the TilingInterface op inside the scf.foreach_thread, rewriting all of its outputs with block arguments and then calling the tiling implementation. Afterwards, the cloned op is deleted again.
Differential Revision: https://reviews.llvm.org/D133114
`getTiledImplementation`/`generateResultTileValue` only computes the tiled operation, but does not insert the result into any tensor.
Differential Revision: https://reviews.llvm.org/D133015
The patch introduces the required changes to update the pass declarations and definitions to use the new autogenerated files and allow dropping the old infrastructure.
Reviewed By: mehdi_amini, rriddle
Differential Review: https://reviews.llvm.org/D132838
The patch introduces the required changes to update the pass declarations and definitions to use the new autogenerated files and allow dropping the old infrastructure.
Reviewed By: mehdi_amini, rriddle
Differential Review: https://reviews.llvm.org/D132838
A part of the functionality of `bufferize` is extracted into `getBufferType`. Also, bufferized scf.yields inside scf.if are now created with the correct bufferized type from the get-to.
Differential Revision: https://reviews.llvm.org/D132862
Even though iter_arg and init_arg of an scf.for loop may have the same tensor type, their bufferized memref types are not necessarily equal. It is sometimes necessary to insert a cast in case of differing layout maps.
Differential Revision: https://reviews.llvm.org/D132860
This change generalizes getBufferType. This function can be used to predict the buffer type of any tensor value (not just BlockArguments) without changing any IR. It also subsumes getMemorySpace. This is useful for loop bufferization, where the precise buffer type of an iter_arg cannot be known without examining the loop body.
Differential Revision: https://reviews.llvm.org/D132859
This patch adds a an `eraseArguments` function that erases a subrange of
a block's arguments. This can be used inplace of the terrible pattern
```
block->eraseArguments(llvm::to_vector(llvm::seq(...)));
```
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D132890
While The tiling interface provides a mechanism for operations to be
tiled into tiled version of the op (or another op at the same level of
abstraction), the `generateScalarImplementation` method added here is
the "exit point" after all transformations have been done. Ops that
implement this method are expected to generate IR that are directly
lowerable to backend dialects like LLVM or SPIR-V dialects.
Differential Revision: https://reviews.llvm.org/D130612
While most of methods in ViewLikeInterface accept an `OpFoldResult` for
the offset/size/stride that may be static, represented as `Attribute`,
or dynamic, represented as `Value`, the `Range` abstraction only
accepted `Values`. This can often lead to known-constant
offset/size/strides being materialized into constant operations and
hinder further constant propagation without explicitly running the
constant folding pass. This often leads to a more complicated than
necessary addressing code being emitted. Switch `Range` to use
`OpFoldResult`. Code that uses `Range` currently keeps materializing the
constants to minimize the effect of this change on the IR. Further
commits will make use of this.
Reviewed By: nicolasvasilache, mravishankar
Differential Revision: https://reviews.llvm.org/D129633
Replace iterators of the outermost loop with region arguments of the innermost
one. The changes avoid later `bufferization` passes to insert allocation within
the body of the innermost loop.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D130083
Replace iterators of the outermost loop with region arguments of the innermost
one. The changes avoid later `bufferization` passes to insert allocation within
the body of the innermost loop.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D130083
This is moslty NFC and will allow tensor.parallel_insert_slice to gain
rank-reducing semantics by reusing the vast majority of the tensor.insert_slice impl.
Depends on D128857
Differential Revision: https://reviews.llvm.org/D128920
This allows purging references of scf.ForeachThreadOp and scf.PerformConcurrentlyOp from
ParallelInsertSliceOp.
This will allowmoving the op closer to tensor::InsertSliceOp with which it should share much more
code.
In the future, the decoupling will also allow extending the type of ops that can be used in the
parallel combinator as well as semantics related to multiple concurrent inserts to the same
result.
Differential Revision: https://reviews.llvm.org/D128857
This was previous implemented as part of the BufferizableOpInterface of ForEachThreadOp. Moving the implementation to ParallelInsertSliceOp to be consistent with the remaining ops and to have a nice example op that can serve as a blueprint for other ops.
Differential Revision: https://reviews.llvm.org/D128666
This change updates all remaining bufferization patterns (except for scf.while) and the remaining bufferization infrastructure to infer the memory space whenever possible instead of falling back to "0". (If a default memory space is set in the bufferization options, we still fall back to that value if the memory space could not be inferred.)
Differential Revision: https://reviews.llvm.org/D128423
Add a failure return value and bufferization options argument. This is to keep a subsequent change smaller.
Differential Revision: https://reviews.llvm.org/D128278
An optional thread_dim_mapping index array attribute specifies for each
virtual thread dimension, how it remaps 1-1 to a set of concrete processing
element resources (e.g. a CUDA grid dimension or a level of concrete nested
async parallelism). At this time, the specification is backend-dependent and
is not verified by the op, beyond being an index array attribute.
It is the reponsibility of the lowering to interpret the index array in the
context of the concrete target the op is lowered to, or to ignore it when
the specification is ill-formed or unsupported for a particular target.
Differential Revision: https://reviews.llvm.org/D128633
This allows for better type inference during bufferization and is in preparation of supporting memory spaces.
Differential Revision: https://reviews.llvm.org/D128422
