This is the same as the existing multiplier-1 variant of DepthwiseConv2D, but in PyTorch dim order.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D128575
This is to improve consistency within the SPIR-V dialect and make these ops a bit shorter.
Reviewed By: antiagainst
Differential Revision: https://reviews.llvm.org/D130280
This change modifies `structured.tile_to_foreach_thread_op` so that
it accepts either `tile_sizes` or `num_threads` parameters. If
`tile_sizes` are specified, then the number of threads required is
derived the tile sizes rather than the other way around. In both cases,
more aggressive folding of loop parameters is enabled during the
transformation, allowing for the potential elimination of `affine.min`
and `affine.max` operations in the static shape case when calculating
the final adjusted tile size.
Differential Revision: https://reviews.llvm.org/D130139
This operation is a NavigationOp that simplifies the writing of transform IR.
Since there is no way of refering to an interface by name, the current implementation uses
an EnumAttr and depends on the interfaces it supports.
In the future, it would be worthwhile to remove this dependence and generalize.
Differential Revision: https://reviews.llvm.org/D130267
The `tileAndFuseLinalgOps` is a legacy approach for tiling + fusion of
Linalg operations. Since it was also intended to work on operations
with buffer operands, this method had fairly complex logic to make
sure tile and fuse was correct even with side-effecting linalg ops.
While complex, it still wasnt robust enough. This patch deprecates
this method and thereby deprecating the tiling + fusion method for ops
with buffer semantics. Note that the core transformation to do fusion
of a producer with a tiled consumer still exists. The deprecation here
only removes methods that auto-magically tried to tile and fuse
correctly in presence of side-effects.
The `tileAndFuseLinalgOps` also works with operations with tensor
semantics. There are at least two other ways the same functionality
exists.
1) The `tileConsumerAndFuseProducers` method. This does a similar
transformation, but using a slightly different logic to
automatically figure out the legal tile + fuse code. Note that this
is also to be deprecated soon.
2) The prefered way uses the `TilingInterface` for tile + fuse, and
relies on the caller to set the tiling options correctly to ensure
that the generated code is correct.
As proof that (2) is equivalent to the functionality provided by
`tileAndFuseLinalgOps`, relevant tests have been moved to use the
interface, where the test driver sets the tile sizes appropriately to
generate the expected code.
Differential Revision: https://reviews.llvm.org/D129901
This patch adds constant folder for LogOp which only supports single and double precision floating-point.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D130148
This is to improve the consistency within the SPIR-V dialect and to make op names a bit shorter.
Reviewed By: antiagainst
Differential Revision: https://reviews.llvm.org/D130194
This patch adds constant folder for Log1pOp which only supports single and double precision floating-point.
Differential Revision: https://reviews.llvm.org/D129979
This file contains a huge number of tests that should really be in
different dialect/files. It is monolothic because of the legacy
surrounding the old standard dialect, affine operations, etc. Splitting
this up makes the tests much more maintainable given that they are now
group with other similar tests.
This revision adds a new transformation to tile a TilingInterface `op` to a tiled `scf.foreach_thread`, applying
tiling by `num_threads`.
If non-empty, the `threadDimMapping` is added as an attribute to the resulting `scf.foreach_thread`.
0-tile sizes (i.e. tile by the full size of the data) are used to encode
that a dimension is not tiled.
Differential Revision: https://reviews.llvm.org/D129577
This op used to belong to the sparse dialect, but there are use cases for dense bufferization as well. (E.g., when a tensor alloc is returned from a function and should be deallocated at the call site.) This change moves the op to the bufferization dialect, which now has an `alloc_tensor` and a `dealloc_tensor` op.
Differential Revision: https://reviews.llvm.org/D129985
This patch adds constant folder for Log10Op which only support single and double precision floating-point.
Reviewed By: Mogball
Differential Revision: https://reviews.llvm.org/D129740
This patch adds a pattern to decompose a `linalg.generic` operations
that
- has only parallel iterator types
- has more than 2 statements (including the yield)
into multiple `linalg.generic` operation such that each operation has
a single statement and a yield.
The pattern added here just splits the matching `linalg.generic` into
two `linalg.generic`s, one containing the first statement, and the
other containing the remaining. The same pattern can be applied
repeatedly on the second op to ultimately fully decompose the generic
op.
Differential Revision: https://reviews.llvm.org/D129704
A new sparse_tensor operation allows for
custom reduction code to be injected during
linalg.generic lowering for sparse tensors.
An identity value is provided to indicate
the starting value of the reduction. A single
block region is required to contain the
custom reduce computation.
Reviewed by: aartbik
Differential Revision: https://reviews.llvm.org/D128004
In `linalg::tileConsumerAndFuseProducers`, there are two levels of
tiling and fusion; we partition the tile sizes and only use one
half for each of them. The partition is using the first non-parallel
dimension *after* interchange as the boundary. However, concrete
tiling happens *together with* loop interchange, so we still need
to provide the partial tile sizes *before* the interchange.
Otherwise, there will be inconsistency, which is what this patch
is to fix.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D129804
The lds_barrier op allows workgroups to wait at a barrier for
operations to/from their local data store (LDS) to complete without
incurring the performance penalties of a full memory fence.
Reviewed By: nirvedhmeshram
Differential Revision: https://reviews.llvm.org/D129522
bufferization.writable is used in most cases instead. All remaining test cases are updated. Some code that is no longer needed is deleted.
Differential Revision: https://reviews.llvm.org/D129739
This revision removes the LinalgPromotion pattern and adds a `transform.structured.promotion` op.
Since the LinalgPromotion transform allows the injection of arbitrary C++ via lambdas, the current
transform op does not handle it.
It is left for future work to decide what the right transform op control is for those cases.
Note the underlying implementation remains unchanged and the mechanism is still controllable by
lambdas from the API.
During this refactoring it was also determined that the `dynamicBuffers` option does not actually
connect to a change of behavior in the algorithm.
This also exhibits that the related test is wrong (and dangerous).
Both the option and the test are therefore removed.
Lastly, a test that connects patterns using the filter-based mechanism is removed: all the independent
pieces are already tested separately.
Context: https://discourse.llvm.org/t/psa-retire-linalg-filter-based-patterns/63785
Differential Revision: https://reviews.llvm.org/D129649
This change removes the partial bufferization passes from the sparse compilation pipeline and replaces them with One-Shot Bufferize. One-Shot Analysis (and TensorCopyInsertion) is used to resolve all out-of-place bufferizations, dense and sparse. Dense ops are then bufferized with BufferizableOpInterface. Sparse ops are still bufferized in the Sparsification pass.
Details:
* Dense allocations are automatically deallocated, unless they are yielded from a block. (In that case the alloc would leak.) All test cases are modified accordingly. E.g., some funcs now have an "out" tensor argument that is returned from the function. (That way, the allocation happens at the call site.)
* Sparse allocations are *not* automatically deallocated. They must be "released" manually. (No change, this will be addressed in a future change.)
* Sparse tensor copies are not supported yet. (Future change)
* Sparsification no longer has to consider inplacability. If necessary, allocations and/or copies are inserted during TensorCopyInsertion. All tensors are inplaceable by the time Sparsification is running. Instead of marking a tensor as "not inplaceable", it can be marked as "not writable", which will trigger an allocation and/or copy during TensorCopyInsertion.
Differential Revision: https://reviews.llvm.org/D129356
- Adds verification for `nvgpu.mma.sync` op
- Adds tests to `mlir/test/Dialect/NVGPU/invalid.mlir`
- `nvgpu.mma.sync` verifier caught a bug and triggered a failure in m16n8k4_tf32_f32 variant in `mlir/test/Conversion/NVGPUToNVVM/nvgpu-to-nvvm.mlir`
- The output shape of vector holding thread-level accumulators was inconsistent and fixed in this change
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D129400
This pass tests patterns that are already tested elsewhere by applying them in a semi-targeted
fashion using anchor function and op names.
From now on, targeted tests should use the transform dialect interpreter.
Differential Revision: https://reviews.llvm.org/D129627
since vector.reduce support accumulator in all the cases remove the
assert assuming old definition.
Differential Revision: https://reviews.llvm.org/D129602
Existing implementation of structured op splitting creates several
affine.apply and affine.min operations in its subshape computation.
As these shapes are further used in data slice extraction, this may lead
to slice shapes being dynamic even when the original shapes and the
splitting point are static. This is particularly visible when splitting
is combined with further subsetting transformations such as tiling. Use
composition and folding more aggressively in splitting to avoid this.
In particular, introduce a `createComposedAffineMin` function that the
affine map used in "min" with the maps used by any `affine.apply` that
may be feeding the operands to the "min". This enables production of
more static shapes. Also introduce a `createComposedFoldedAffineApply`
function that combines the existing `createComposedAffineApply` with
in-place folding to propagate constants produced by zero-input affine
maps. Using these when splitting allows the subsequent canonicalizer
pass to recover static shapes for structured ops.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D129379
This allows vectorizing linalg reductions without changing the operation
order. Therefore this produce a valid vectorization even if operations
are not associative.
Differential Revision: https://reviews.llvm.org/D129535
Introduce a structured transform op that emits IR computing the multi-tile
sizes with requested parameters (target size and divisor) for the given
structured op. The sizes may fold to arithmetic constant operations when the
shape is constant. These operations may then be used to call the existing
tiling transformation with a single non-zero dynamic size (i.e. perform
strip-mining) for each of the dimensions separately, thus achieving multi-size
tiling with optional loop interchange. A separate test exercises the entire
script.
Depends On D129217
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D129287
Extend the definition of the Tile structured transform op to enable it
accepting handles to operations that produce tile sizes at runtime. This is
useful by itself and prepares for more advanced tiling strategies. Note that
the changes are relevant only to the transform dialect, the tiling
transformation itself already supports dynamic sizes.
Depends On D129216
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D129217
This handle manipulation operation allows one to define a new handle that is
associated with a the same payload IR operations N times, where N can be driven
by the size of payload IR operation list associated with another handle. This
can be seen as a sort of broadcast that can be used to ensure the lists
associated with two handles have equal numbers of payload IR ops as expected by
many pairwise transform operations.
Introduce an additional "expensive" check that guards against consuming a
handle that is assocaited with the same payload IR operation more than once as
this is likely to lead to double-free or other undesired effects.
Depends On D129110
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D129216
This patch uses constFoldUnaryOpConditional to replace current folder in order to support constant dense.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D129459
This can be enabled by using a `-re` suffix when defining the expected line,
e.g. `expected-error-re`. This support is similar to what clang provides in its "expected"
diagnostic framework(e.g. the `-re` is also the same). The regex definitions themselves are
similar to FileCheck in that regex blocks are specified within `{{` `}}` blocks.
Differential Revision: https://reviews.llvm.org/D129343
A previous revision implemented expand/collapse reshaping between
dense and sparse tensors for sparse2dense and dense2sparse since those
could use the "cheap" view reshape on the already materialized
dense tensor (at either the input or output side), and do some
reshuffling from or to sparse. The dense2dense case, as always,
is handled with a "cheap" view change.
This revision implements the sparse2sparse cases. Lacking any "view"
support on sparse tensors this operation necessarily has to perform
data reshuffling on both ends.
Tracker for improving this:
https://github.com/llvm/llvm-project/issues/56477
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D129416
Prevent creating multiple output for the same Value when distributing
operations out of WarpExecuteOnLane0Op. This avoid creating combinatory
explosion of outputs.
Differential Revision: https://reviews.llvm.org/D129465
This patch is similar to D129108, it adds a conditional unary constant folder which allow to exit when the constants not meet the fold condition. And use it for Log2Op to make it able to fold the constant dense.
Differential Revision: https://reviews.llvm.org/D129251
This is a partial revert of D128615.
to_memref(to_tensor(x)) always be folded to x. But to_tensor(to_memref(x)) cannot be folded in the general case because writes to the intermediary memref may go unnoticed.
Differential Revision: https://reviews.llvm.org/D129354
Because the buffer descriptor structure (the V#) has no backwards-compatibility
guarentees, and since said guarantees have been violated in practice
(see https://github.com/llvm/llvm-project/issues/56323 ), and since
the `targetIsRDNA` attribute isn't something that higher-level clients can set
in general, make the lowering of the amdgpu dialect to rocdl take a --chipset
option.
Note that this option is a string because adding a parser for the Chipset
struct to llvm::cl wasn't working out.
Reviewed By: herhut
Differential Revision: https://reviews.llvm.org/D129228
This revision revisits the implementation of applyToOne and its handling
of recoverable errors as well as propagation of null handles.
The implementation is simplified to always require passing a vector<Operation*>
in which the results are returned, resulting in less template instantiation magic.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D129185
This Transform dialect op allows one to merge the lists of Payload IR
operations pointed to by several handles into a single list associated with one
handle. This is an important Transform dialect usability improvement for cases
where transformations may temporarily diverge for different groups of Payload
IR ops before converging back to the same script. Without this op, several
copies of the trailing transformations would have to be present in the
transformation script.
Depends On D129090
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D129110
Introduce a new transformation on structured ops that splits the iteration
space into two parts along the specified dimension. The index at which the
splitting happens may be static or dynamic. This transformation can be seen as
a rudimentary form of index-set splitting that only supports the splitting
along hyperplanes parallel to the iteration space hyperplanes, and is therefore
decomposable into per-dimension application.
It is a key low-level transformation that enables independent scheduling for
different parts of the iteration space of the same op, which hasn't been
possible previously. It may be used to implement, e.g., multi-sized tiling. In
future, peeling can be implemented as a combination of split-off amount
computation and splitting.
The transformation is conceptually close to tiling in its separation of the
iteration and data spaces, but cannot be currently implemented on top of
TilingInterface as the latter does not properly support `linalg.index`
offsetting.
Note that the transformation intentionally bypasses folding of
`tensor.extract_slice` operations when creating them as this folding was found
to prevent repeated splitting of the same operation because due to internal
assumptions about extract/insert_slice combination in dialect utilities.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D129090
