The approach for handling reductions in the outer most
dimension follows that for inner most dimensions, outlined
below
First, transpose to move reduction dims, if needed
Convert reduction from n-d to 2-d canonical form
Then, for outer reductions, we emit the appropriate op
(add/mul/min/max/or/and/xor) and combine the results.
Differential Revision: https://reviews.llvm.org/D107675
The existing vector transforms reduce the dimension of transfer_read
ops. However, beyond a certain point, the vector op actually has
to be reduced to a scalar load, since we can't load a zero-dimension
vector. This handles this case.
Note that in the longer term, it may be preferaby to support
zero-dimension vectors. see
https://llvm.discourse.group/t/should-we-have-0-d-vectors/3097.
Differential Revision: https://reviews.llvm.org/D103432
When the output indexing map has a permutation we need to consider in
the contraction vector type.
Differential Revision: https://reviews.llvm.org/D106469
This simplifies the vector to LLVM lowering. Previously, both vector.load/store and vector.transfer_read/write lowered directly to LLVM. With this commit, there is a single path to LLVM vector load/store instructions and vector.transfer_read/write ops must first be lowered to vector.load/store ops.
* Remove vector.transfer_read/write to LLVM lowering.
* Allow non-unit memref strides on all but the most minor dimension for vector.load/store ops.
* Add maxTransferRank option to populateVectorTransferLoweringPatterns.
* vector.transfer_reads with changing element type can no longer be lowered to LLVM. (This functionality is needed only for SPIRV.)
Differential Revision: https://reviews.llvm.org/D106118
Simplify vector unrolling pattern to be more aligned with rest of the
patterns and be closer to vector distribution.
The new implementation uses ExtractStridedSlice/InsertStridedSlice
instead of the Tuple ops. After this change the ops based on Tuple don't
have any more used so they can be removed.
This allows removing signifcant amount of dead code and will allow
extending the unrolling code going forward.
Differential Revision: https://reviews.llvm.org/D105381
Remove `getDynOperands` and `createOrFoldDimOp` from MemRef.h to decouple MemRef a bit from Tensor. These two functions are used in other dialects/transforms.
Differential Revision: https://reviews.llvm.org/D105260
The implementation has become too unwieldy and cognitive overhead wins.
Instead compress the implementation in preparation for additional lowering paths.
This is a resubmit of https://reviews.llvm.org/D105359 without ordering ambiguities.
Differential Revision: https://reviews.llvm.org/D105367
The implementation has become too unwieldy and cognitive overhead wins.
Instead compress the implementation in preparation for additional lowering paths.
Differential Revision: https://reviews.llvm.org/D105359
* Split memref.dim into two operations: memref.dim and tensor.dim. Both ops have the same builder interface and op argument names, so that they can be used with templates in patterns that apply to both tensors and memrefs (e.g., some patterns in Linalg).
* Add constant materializer to TensorDialect (needed for folding in affine.apply etc.).
* Remove some MemRefDialect dependencies, make some explicit.
Differential Revision: https://reviews.llvm.org/D105165
Uses elementwise interface to generalize canonicalization pattern and add a new
pattern for vector.contract case.
Differential Revision: https://reviews.llvm.org/D104343
* Previously, we were only generating .h.inc files. We foresee the need to also generate implementations and this is a step towards that.
* Discussed in https://llvm.discourse.group/t/generating-cpp-inc-files-for-dialects/3732/2
* Deviates from the discussion above by generating a default constructor in the .cpp.inc file (and adding a tablegen bit that disables this in case if this is user provided).
* Generating the destructor started as a way to flush out the missing includes (produces a link error), but it is a strict improvement on its own that is worth doing (i.e. by emitting key methods in the .cpp file, we root vtables in one translation unit, which is a non-controversial improvement).
Differential Revision: https://reviews.llvm.org/D105070
The current code does not preserve the order of the parallel
dimensions when doing multi-reductions and thus we can end
up in scenarios where the result shape does not match the
desired shape after reduction.
This patch fixes that by ensuring that the parallel indices
are in order and then concatenates them to the reduction dimensions
so that the reduction dimensions are innermost.
Differential Revision: https://reviews.llvm.org/D104884
The patch changes the pretty printed FillOp operand order from output, value to value, output. The change is a follow up to https://reviews.llvm.org/D104121 that passes the fill value using a scalar input instead of the former capture semantics.
Differential Revision: https://reviews.llvm.org/D104356
VectorTransferPermutationMapLoweringPatterns can be enabled via a pass option. These additional patterns lower permutation maps to minor identity maps with broadcasting, if possible, allowing for more efficient vector load/stores. The option is deactivated by default.
Differential Revision: https://reviews.llvm.org/D102593
Broadcast dimensions of vector transfer ops are always in-bounds. This is consistent with the fact that the starting position of a transfer is always in-bounds.
Differential Revision: https://reviews.llvm.org/D102566
Add TransferWritePermutationLowering, which replaces permutation maps of TransferWriteOps with vector.transpose.
Differential Revision: https://reviews.llvm.org/D102548
Broadcast dimensions of a vector transfer op have no corresponding dimension in the mask vector. E.g., a 2-D TransferReadOp, where one dimension is a broadcast, can have a 1-D `mask` attribute.
This commit also adds a few additional transfer op integration tests for various combinations of broadcasts, masking, dim transposes, etc.
Differential Revision: https://reviews.llvm.org/D101745
Broadcast dimensions of a vector transfer op have no corresponding dimension in the mask vector. E.g., a 2-D TransferReadOp, where one dimension is a broadcast, can have a 1-D `mask` attribute.
This commit also adds a few additional transfer op integration tests for various combinations of broadcasts, masking, dim transposes, etc.
Differential Revision: https://reviews.llvm.org/D101745
VectorTransfer split previously only split read xfer ops. This adds
the same logic to write ops. The resulting code involves 2
conditionals for write ops while read ops only needed 1, but the created
ops are built upon the same patterns, so pattern matching/expectations
are all consistent other than in regards to the if/else ops.
Differential Revision: https://reviews.llvm.org/D102157
This patch adds support for vectorizing loops with 'iter_args'
implementing known reductions along the vector dimension. Comparing to
the non-vector-dimension case, two additional things are done during
vectorization of such loops:
- The resulting vector returned from the loop is reduced to a scalar
using `vector.reduce`.
- In some cases a mask is applied to the vector yielded at the end of
the loop to prevent garbage values from being written to the
accumulator.
Vectorization of reduction loops is disabled by default. To enable it, a
map from loops to array of reduction descriptors should be explicitly passed to
`vectorizeAffineLoops`, or `vectorize-reductions=true` should be passed
to the SuperVectorize pass.
Current limitations:
- Loops with a non-unit step size are not supported.
- n-D vectorization with n > 1 is not supported.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D100694
TransferReadOps that are a scalar read + broadcast are handled by TransferReadToVectorLoadLowering.
Differential Revision: https://reviews.llvm.org/D101808
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
Move TransposeOp lowering in its own populate function as in some cases
it is better to keep it during ContractOp lowering to better
canonicalize it rather than emiting scalar insert/extract.
Differential Revision: https://reviews.llvm.org/D101647
Three patterns are added to convert into vector.multi_reduction into a
sequence of vector.reduction as the following:
- Transpose the inputs so inner most dimensions are always reduction.
- Reduce rank of vector.multi_reduction into 2d with inner most
reduction dim (get the 2d canical form)
- 2D canonical form is converted into a sequence of vector.reduction.
There are two things we might worth in a follow up diff:
- An scf.for (maybe optionally) around vector.reduction instead of unrolling it.
- Breakdown the vector.reduction into a sequence of vector.reduction
(e.g tree-based reduction) instead of relying on how downstream dialects
handle it.
Note: this will requires passing target-vector-length
Differential Revision: https://reviews.llvm.org/D101570
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
The current implementation allows for TransferWriteOps with broadcasts that do not make sense. E.g., a broadcast could write a vector into a single (scalar) memory location, which is effectively the same as writing only the last element of the vector.
Differential Revision: https://reviews.llvm.org/D100842
Move the existing optimization for transfer op on tensor to folder and
canonicalization. This handles the write after write case and read after write
and also add write after read case.
Differential Revision: https://reviews.llvm.org/D100597
harsh-nod