The pooling ops are among the last remaining hard coded Linalg operations that have no region attached. They got obsolete due to the OpDSL pooling operations. Removing them allows us to delete specialized code and tests that are not needed for the OpDSL counterparts that rely on the standard code paths.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D110909
Add support for dynamic shared memory for GPU launch ops: add an
optional operand to gpu.launch and gpu.launch_func ops to specify the
amount of "dynamic" shared memory to use. Update lowerings to connect
this operand to the GPU runtime.
Differential Revision: https://reviews.llvm.org/D110800
For convolution, the input window dimension's access affine map
is of the form `(d0 * s0 + d1)`, where `d0`/`d1` is the output/
filter window dimension, and `s0` is the stride.
When tiling, https://reviews.llvm.org/D109267 changed how the
way dimensions are acquired. Instead of directly querying using
`*.dim` ops on the original convolution op, we now get it by
applying the access affine map to the loop upper bounds. This
is fine for dimensions having single-dimension affine maps,
like matmul, but not for convolution input. It will cause
incorrect compuation and out of bound. A concrete example, say
we have 1x225x225x3 (NHWC) input, 3x3x3x32 (HWCF) filter, and
1x112x112x3 (NHWC) output with stride 2, (112 * 2 + 3) would be
227, which is different from the correct input window dimension
size 225.
Instead, we should first calculate the max indices for each loop,
and apply the affine map to them, and then plus one to get the
dimension size. Note this makes no difference for matmul-like
ops given they will have `d0 - 1 + 1` effectively.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D110849
* This could have been removed some time ago as it only had one op left in it, which is redundant with the new approach.
* `matmul_i8_i8_i32` (the remaining op) can be trivially replaced by `matmul`, which natively supports mixed precision.
Differential Revision: https://reviews.llvm.org/D110792
This revision retires a good portion of the complexity of the codegen strategy and puts the logic behind pass logic.
Differential revision: https://reviews.llvm.org/D110678
Unroll-and-jam currently doesn't work when the loop being unroll-and-jammed
or any of its inner loops has iter_args. This patch modifies the
unroll-and-jam utility to support loops with iter_args.
Reviewed By: bondhugula
Differential Revision: https://reviews.llvm.org/D110085
Adapt the signature of the PaddingValueComputationFunction callback to either return the padding value or failure to signal padding is not desired.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D110572
This revision makes sure that when the output buffer materializes locally
(in contrast with the passing in of output tensors either in-place or not
in-place), the zero initialization assumption is preserved. This also adds
a bit more documentation on our sparse kernel assumption (viz. TACO
assumptions).
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D110442
The sparse constant provides a constant tensor in coordinate format. We first split the sparse constant into a constant tensor for indices and a constant tensor for values. We then generate a loop to fill a sparse tensor in coordinate format using the tensors for the indices and the values. Finally, we convert the sparse tensor in coordinate format to the destination sparse tensor format.
Add tests.
Reviewed By: aartbik
Differential Revision: https://reviews.llvm.org/D110373
When splitting with linalg.copy, cannot write into the destination alloc directly. Instead, write into a subview of the alloc.
Differential Revision: https://reviews.llvm.org/D110512
For such cases, the type of the constant DenseElementsAttr is
different from the transpose op return type.
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D110446
These are among the last operations still defined explicitly in C++. I've
tried to keep this commit as NFC as possible, but these ops
definitely need a non-NFC cleanup at some point.
Differential Revision: https://reviews.llvm.org/D110440
* If the input is a constant splat value, we just
need to reshape it.
* If the input is a general constant with one user,
we can also constant fold it, without bloating
the IR.
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D110439
Initially, the padding transformation and the related operation were only used
to guarantee static shapes of subtensors in tiled operations. The
transformation would not insert the padding operation if the shapes were
already static, and the overall code generation would actively remove such
"noop" pads. However, this transformation can be also used to pack data into
smaller tensors and marshall them into faster memory, regardless of the size
mismatches. In context of expert-driven transformation, we should assume that,
if padding is requested, a potentially padded tensor must be always created.
Update the transformation accordingly. To do this, introduce an optional
`packing` attribute to the `pad_tensor` op that serves as an indication that
the padding is an intentional choice (as opposed to side effect of type
normalization) and should be left alone by cleanups.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D110425
This canonicalization pattern complements the tensor.cast(pad_tensor) one in
propagating constant type information when possible. It contributes to the
feasibility of pad hoisting.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D110343
* Do not discard static result type information that cannot be inferred from lower/upper padding.
* Add optional argument to `PadTensorOp::inferResultType` for specifying known result dimensions.
Differential Revision: https://reviews.llvm.org/D110380
This test makes sure kernels map to efficient sparse code, i.e. all
compressed for-loops, no co-iterating while loops. In addition, this
revision removes the special constant folding inside the sparse
compiler in favor of Mahesh' new generic linalg folding. Thanks!
NOTE: relies on Mahesh fix, which needs to be rebased first
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D110001
The current folder of constant -> generic op only handles splat
constants. The same logic holds for scalar constants. Teach the
pattern to handle such cases.
Differential Revision: https://reviews.llvm.org/D109982
Now not just SUM, but also PRODUCT, AND, OR, XOR. The reductions
MIN and MAX are still to be done (also depends on recognizing
these operations in cmp-select constructs).
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D110203
Compute the tiled producer slice dimensions directly starting from the consumer not using the producer at all.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D110147
It was previously assumed that tensor.insert_slice should be bufferized first in a greedy fashion to avoid out-of-place bufferization of the large tensor. This heuristic does not hold upon further inspection.
This CL removes the special handling of such ops and adds a test that exhibits better behavior and appears in real use cases.
The only test adversely affected is an artificial test which results in a returned memref: this pattern is not allowed by comprehensive bufferization in real scenarios anyway and the offending test is deleted.
Differential Revision: https://reviews.llvm.org/D110072
Previously, comprehensive bufferize would consider all aliasing reads and writes to
the result buffer and matching operand. This resulted in spurious dependences
being considered and resulted in too many unnecessary copies.
Instead, this revision revisits the gathering of read and write alias sets.
This results in fewer alloc and copies.
An exhaustive test cases is added that considers all possible permutations of
`matmul(extract_slice(fill), extract_slice(fill), ...)`.
This pass transforms SCF.ForOp operations to SCF.WhileOp. The For loop condition is placed in the 'before' region of the while operation, and indctuion variable incrementation + the loop body in the 'after' region. The loop carried values of the while op are the induction variable (IV) of the for-loop + any iter_args specified for the for-loop.
Any 'yield' ops in the for-loop are rewritten to additionally yield the (incremented) induction variable.
This transformation is useful for passes where we want to consider structured control flow solely on the basis of a loop body and the computation of a loop condition. As an example, when doing high-level synthesis in CIRCT, the incrementation of an IV in a for-loop is "just another part" of a circuit datapath, and what we really care about is the distinction between our datapath and our control logic (the condition variable).
Differential Revision: https://reviews.llvm.org/D108454
SparseElementsAttr currently does not perform any verfication on construction, with the only verification existing within the parser. This revision moves the parser verification to SparseElementsAttr, and also adds additional verification for when a sparse index is not valid.
Differential Revision: https://reviews.llvm.org/D109189
For `memref.subview` operations, when there are more than one
unit-dimensions, the strides need to be used to figure out which of
the unit-dims are actually dropped.
Differential Revision: https://reviews.llvm.org/D109418
Add an interface that allows grouping together all covolution and
pooling ops within Linalg named ops. The interface currently
- the indexing map used for input/image access is valid
- the filter and output are accessed using projected permutations
- that all loops are charecterizable as one iterating over
- batch dimension,
- output image dimensions,
- filter convolved dimensions,
- output channel dimensions,
- input channel dimensions,
- depth multiplier (for depthwise convolutions)
Differential Revision: https://reviews.llvm.org/D109793
This pass transforms SCF.ForOp operations to SCF.WhileOp. The For loop condition is placed in the 'before' region of the while operation, and indctuion variable incrementation + the loop body in the 'after' region. The loop carried values of the while op are the induction variable (IV) of the for-loop + any iter_args specified for the for-loop.
Any 'yield' ops in the for-loop are rewritten to additionally yield the (incremented) induction variable.
This transformation is useful for passes where we want to consider structured control flow solely on the basis of a loop body and the computation of a loop condition. As an example, when doing high-level synthesis in CIRCT, the incrementation of an IV in a for-loop is "just another part" of a circuit datapath, and what we really care about is the distinction between our datapath and our control logic (the condition variable).
Differential Revision: https://reviews.llvm.org/D108454
Add a new version of fusion on tensors that supports the following scenarios:
- support input and output operand fusion
- fuse a producer result passed in via tile loop iteration arguments (update the tile loop iteration arguments)
- supports only linalg operations on tensors
- supports only scf::for
- cannot add an output to the tile loop nest
The LinalgTileAndFuseOnTensors pass tiles the root operation and fuses its producers.
Reviewed By: nicolasvasilache, mravishankar
Differential Revision: https://reviews.llvm.org/D109766
So far, the CF cost-model for detensoring was limited to discovering
pure CF structures. This means, if while discovering the CF component,
the cost-model found any op that is not detensorable, it gives up on
detensoring altogether. This patch makes it a bit more flexible by
cleaning-up the detensorable component from non-detensorable ops without
giving up entirely.
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D109965
Add a constant propagator for gpu.launch op in cases where the
grid/thread IDs can be trivially determined to take a single constant
value of zero.
Differential Revision: https://reviews.llvm.org/D109994
Even with all parallel loops reading the output value is still allowed so we
don't have to handle reduction loops differently.
Differential Revision: https://reviews.llvm.org/D109851
This enables the sparsification of more kernels, such as convolutions
where there is a x(i+j) subscript. It also enables more tensor invariants
such as x(1) or other affine subscripts such as x(i+1). Currently, we
reject sparsity altogether for such tensors. Despite this restriction,
however, we can already handle a lot more kernels with compound subscripts
for dense access (viz. convolution with dense input and sparse filter).
Some unit tests and an integration test demonstrate new capability.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D109783
There are two main versions of depthwise conv depending whether the multiplier
is 1 or not. In cases where m == 1 we should use the version without the
multiplier channel as it can perform greater optimization.
Add lowering for the quantized/float versions to have a multiplier of one.
Reviewed By: antiagainst
Differential Revision: https://reviews.llvm.org/D108959
This revision fixes a corner case that could appear due to incorrect insertion point behavior in comprehensive bufferization.
Differential Revision: https://reviews.llvm.org/D109830
Add the makeComposedExtractSliceOp method that creates an ExtractSliceOp and folds chains of ExtractSliceOps by computing the sum of their offsets and by multiplying their strides.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D109601
Tobias Gysi