This is a proof of concept recognition of the most basic forms of ReLu
operations, used to show-case sparsification of end-to-end PyTorch
models. In the long run, we must avoid lowering such constructs too
early (with this need for raising them back).
See discussion at
https://discourse.llvm.org/t/min-max-abs-relu-recognition-starter-project/78918
A general question is: is it possible to support hooks here to infer the
encoding? E.g., when the extracted tensor slice is rank-reduced, the
encoding need to be updated accordingly as well.
1. Verify that the type of explicit/implicit values should be the same
as the tensor element type.
2. Verify that implicit value could only be zero.
3. Verify that explicit/implicit values should be numeric.
4. Fix the type change issue caused by SparseTensorType(enc).
Codegen "vectors" for pos/crd/val use the capacity as memref size, not
the actual used size. Although the sparsifier itself always uses just
the defined pos/crd/val parts, printing these and passing them back to a
runtime environment could benefit from wrapping the basic pos/crd/val
getters into a proper memref view that sets the right size.
This patch generalizes tensor.expand_shape and memref.expand_shape to
consume the output shape as a list of SSA values. This enables us to
implement generic reshape operations with dynamic shapes using
collapse_shape/expand_shape pairs.
The output_shape input to expand_shape follows the static/dynamic
representation that's also used in `tensor.extract_slice`.
Differential Revision: https://reviews.llvm.org/D140821
---------
Signed-off-by: Gaurav Shukla<gaurav.shukla@amd.com>
Signed-off-by: Gaurav Shukla <gaurav.shukla@amd.com>
Co-authored-by: Ramiro Leal-Cavazos <ramiroleal050@gmail.com>
This ensures the explicit value is generated (and not a load into the
values array). Note that actually not storing values array at all is
still TBD, this is just the very first step.
1. Explicit value means the non-zero value in a sparse tensor. If
explicitVal is set, then all the non-zero values in the tensor have the
same explicit value. The default value Attribute() indicates that it is
not set.
2. Implicit value means the "zero" value in a sparse tensor. If
implicitVal is set, then the "zero" value in the tensor is equal to the
implicit value. For now, we only support `0` as the implicit value but
it could be extended in the future. The default value Attribute()
indicates that the implicit value is `0` (same type as the tensor
element type).
Example:
```
#CSR = #sparse_tensor.encoding<{
map = (d0, d1) -> (d0 : dense, d1 : compressed),
posWidth = 64,
crdWidth = 64,
explicitVal = 1 : i64,
implicitVal = 0 : i64
}>
```
Note: this PR tests that implicitVal could be set to other values as
well. The following PR will add verifier and reject any value that's not
zero for implicitVal.
This patch generalizes tensor.expand_shape and memref.expand_shape to
consume the output shape as a list of SSA values. This enables us to
implement generic reshape operations with dynamic shapes using
collapse_shape/expand_shape pairs.
The output_shape input to expand_shape follows the static/dynamic
representation that's also used in `tensor.extract_slice`.
Differential Revision: https://reviews.llvm.org/D140821
Co-authored-by: Ramiro Leal-Cavazos <ramiroleal050@gmail.com>
A `sparse_tensor.extract_space %tensor at %iterator` extracts a *sparse*
iteration space defined `%tensor`, the operation to traverse the
iteration space will be introduced in following PRs.
Tensor copy insertion currently uses memory_space = 0 when creating a
tensor copy using alloc_tensor. This memory space should instead be the
default memory space provided in bufferization options.
In order to support various external frameworks (JAX vs PyTorch) we need
a bit more flexibility in [dis]assembling external buffers to and from
sparse tensors in MLIR land. This PR adds a direct-out option that
avoids the rigid pre-allocated for copy-out semantics.
Note that over time, we expect the [dis]assemble operations to converge
into something that supports all sorts of external frameworks. Until
then, this option helps in experimenting with different options.
`linalg.generic` ops with sparse tensors do not necessarily bufferize to
element-wise access, because insertions into a sparse tensor may change
the layout of (or reallocate) the underlying sparse data structures.
This fixes an "infinite" loop bug, where the incoming IR was repeatedly
rewritten while adding identical cast operations. The test for
compatible types should include the notion of an encoding. If it
differs, then a
naive fusion into the consumer is invalid.
This change lifts the restriction that purely allocated empty sparse
tensors cannot escape the method. Instead it makes a best effort to add
a finalizing operation before the escape.
This assumes that
(1) we never build sparse tensors across method boundaries
(e.g. allocate in one, insert in other method)
(2) if we have other uses of the empty allocation in the
same method, we assume that either that op will fail
or will do the finalization for us.
This is best-effort, but fixes some very obvious missing cases.
1. Add python test for n out of m
2. Add more methods for python binding
3. Add verification for n:m and invalid encoding tests
4. Add e2e test for n:m
Previous PRs for n:m #80501#79935
Because the sparse vectorizer relies on the code coming out of the
sparsifier, the "patterns" are not always made very general. However, a
recent change in the generated code revealed an obvious situation where
the subscript analysis could be made a bit more robust.
Fixes:
https://github.com/llvm/llvm-project/issues/79897
1. C++ enum is set through enum class LevelType : uint_64.
2. C enum is set through typedef uint_64 level_type. It is due to the
limitations in Windows build: setting enum width to ui64 is not
supported in C.
Rewrite *all* public methods, making original internal, private methods,
and exposing wrappers under the original name. This works a bit better
in practice (when combined with c-interface mechanism of torch-mlir for
example).
Similar to the emit_c_interface, this pull request adds a pass that
converts public entry methods that use sparse tensors as input
parameters and/or output return values into wrapper functions that
[dis]assemble the individual tensors that constitute the actual storage
used externally into MLIR sparse tensors. This pass can be used to
prepare the public entry methods of a program that is compiled by the
MLIR sparsifier to interface with an external runtime, e.g., when
passing sparse tensors as numpy arrays from and to Python. Note that
eventual bufferization decisions (e.g. who [de]allocates the underlying
memory) should be resolved in agreement with the external runtime
(Python, PyTorch, JAX, etc.)
Introduce a new extension for simple print-debugging of the transform
dialect scripts. The initial version of this extension consists of two
ops that are printing the payload objects associated with transform
dialect values. Similar ops were already available in the test extenion
and several downstream projects, and were extensively used for testing.
The `GreedyPatternRewriteDriver` tries to iteratively fold ops and apply
rewrite patterns to ops. It has special handling for constants: they are
CSE'd and sometimes moved to parent regions to allow for additional
CSE'ing. This happens in `OperationFolder`.
To allow for efficient CSE'ing, `OperationFolder` maintains an internal
lookup data structure to find the existing constant ops with the same
value for each `IsolatedFromAbove` region:
```c++
/// A mapping between an insertion region and the constants that have been
/// created within it.
DenseMap<Region *, ConstantMap> foldScopes;
```
Rewrite patterns are allowed to modify operations. In particular, they
may move operations (including constants) from one region to another
one. Such an IR rewrite can make the above lookup data structure
inconsistent.
We encountered such a bug in a downstream project. This bug materialized
in the form of an op that uses the result of a constant op from a
different `IsolatedFromAbove` region (that is not accessible).
This commit changes the behavior of the `GreedyPatternRewriteDriver`
such that `OperationFolder` is used to CSE constants at the beginning of
each iteration (as the worklist is populated), but no longer during an
iteration. `OperationFolder` is no longer used after populating the
worklist, so we do not have to care about inconsistent state in the
`OperationFolder` due to IR rewrites. The `GreedyPatternRewriteDriver`
now performs the op folding by itself instead of calling
`OperationFolder::tryToFold`.
This change changes the order of constant ops in test cases, but not the
region in which they appear. All broken test cases were fixed by turning
`CHECK` into `CHECK-DAG`.
Alternatives considered: The state of `OperationFolder` could be
partially invalidated with every `notifyOperationModified` notification.
That is more fragile than the solution in this commit because incorrect
rewriter API usage can lead to missing notifications and hard-to-debug
`IsolatedFromAbove` violations. (It did not fix the above mention bug in
a downstream project, which could be due to incorrect rewriter API usage
or due to another conceptual problem that I missed.) Moreover, ops are
frequently getting modified during a greedy pattern rewrite, so we would
likely keep invalidating large parts of the state of `OperationFolder`
over and over.
Migration guide: Turn `CHECK` into `CHECK-DAG` in test cases. Constant
ops are no longer folded during a greedy pattern rewrite. If you rely on
folding (and rematerialization) of constant ops during a greedy pattern
rewrite, turn the folder into a pattern.
This removes the temporary DENSE24 attribute and replaces it with proper
recognition of dense to 24 conversion. The compressionh will be
performed on the device prior to performing the matrix mult. Note that
we no longer need to start with the linalg version, we can lift this to
the proper named linalg op. Also renames some files into more consistent
names.
