As discussed in https://llvm.discourse.group/t/mlir-support-for-sparse-tensors/2020
this CL is the start of sparse tensor compiler support in MLIR. Starting with a
"dense" kernel expressed in the Linalg dialect together with per-dimension
sparsity annotations on the tensors, the compiler automatically lowers the
kernel to sparse code using the methods described in Fredrik Kjolstad's thesis.
Many details are still TBD. For example, the sparse "bufferization" is purely
done locally since we don't have a global solution for propagating sparsity
yet. Furthermore, code to input and output the sparse tensors is missing.
Nevertheless, with some hand modifications, the generated MLIR can be
easily converted into runnable code already.
Reviewed By: nicolasvasilache, ftynse
Differential Revision: https://reviews.llvm.org/D90994
std.alloc only supports memrefs with identity layout, which means we can simplify the lowering to LLVM and compute strides only from (static and dynamic) sizes.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D91549
This commit does the renaming mentioned in the title in order to bring
`spv` dialect closer to the MLIR naming conventions.
Reviewed By: antiagainst
Differential Revision: https://reviews.llvm.org/D91609
This utility function is helpful for dialect-specific builders that need
to access the context through location, and the location itself may be
either provided as an argument or expected to be recovered from the
implicit location stack.
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D91623
The shape of the result of a dynamic_tensor_from_elements is defined via its
result type and operands. We already fold dim operations when they reference
one of the statically sized dimensions. Now, also fold dim on the dynamically
sized dimensions by picking the corresponding operand.
Differential Revision: https://reviews.llvm.org/D91616
It may be necessary for interface methods to process or return variables with
the interface class type, in particular for attribute and type interfaces that
can return modified attributes and types that implement the same interface.
However, the code generated by ODS in this case would not compile because the
signature (and the body if provided) appear in the definition of the Model
class and before the interface class, which derives from the Model. Change the ODS
interface method generator to emit only method declarations in the Model class
itself, and emit method definitions after the interface class. Mark as "inline"
since their definitions are still emitted in the header and are no longer
implicitly inline. Add a forward declaration of the interface class before the
Concept+Model classes to make the class name usable in declarations.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D91499
In ODS, attributes of an operation can be provided as a part of the "arguments"
field, together with operands. Such attributes are accepted by the op builder
and have accessors generated.
Implement similar functionality for ODS-generated op-specific Python bindings:
the `__init__` method now accepts arguments together with operands, in the same
order as in the ODS `arguments` field; the instance properties are introduced
to OpView classes to access the attributes.
This initial implementation accepts and returns instances of the corresponding
attribute class, and not the underlying values since the mapping scheme of the
value types between C++, C and Python is not yet clear. Default-valued
attributes are not supported as that would require Python to be able to parse
C++ literals.
Since attributes in ODS are tightely related to the actual C++ type system,
provide a separate Tablegen file with the mapping between ODS storage type for
attributes (typically, the underlying C++ attribute class), and the
corresponding class name. So far, this might look unnecessary since all names
match exactly, but this is not necessarily the cases for non-standard,
out-of-tree attributes, which may also be placed in non-default namespaces or
Python modules. This also allows out-of-tree users to generate Python bindings
without having to modify the bindings generator itself. Storage type was
preferred over the Tablegen "def" of the attribute class because ODS
essentially encodes attribute _constraints_ rather than classes, e.g. there may
be many Tablegen "def"s in the ODS that correspond to the same attribute type
with additional constraints
The presence of the explicit mapping requires the change in the .td file
structure: instead of just calling the bindings generator directly on the main
ODS file of the dialect, it becomes necessary to create a new file that
includes the main ODS file of the dialect and provides the mapping for
attribute types. Arguably, this approach offers better separability of the
Python bindings in the build system as the main dialect no longer needs to know
that it is being processed by the bindings generator.
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D91542
These includes have been deprecated in favor of BuiltinDialect.h, which contains the definitions of ModuleOp and FuncOp.
Differential Revision: https://reviews.llvm.org/D91572
This replaces the old type decomposition logic that was previously mixed
into bufferization, and makes it easily accessible.
This also deletes TestFinalizingBufferize, because after we remove the type
decomposition, it doesn't do anything that is not already provided by
func-bufferize.
Differential Revision: https://reviews.llvm.org/D90899
The current code allows strided layouts, but the number of elements allocated is ambiguous. It could be either the number of elements in the shape (the current implementation), or the amount of elements required to not index out-of-bounds with the given maps (which would require evaluating the layout map).
If we require the canonical layouts, the two will be the same.
Reviewed By: nicolasvasilache, ftynse
Differential Revision: https://reviews.llvm.org/D91523
This adds a simple definition of a "workshare loop" operation for
the OpenMP MLIR dialect, excluding the "reduction" and "allocate"
clauses and without a custom parser and pretty printer.
The schedule clause also does not yet accept the modifiers that are
permitted in OpenMP 5.0.
Co-authored-by: Kiran Chandramohan <kiran.chandramohan@arm.com>
Reviewed By: ftynse, clementval
Differential Revision: https://reviews.llvm.org/D86071
The logic of vector on boolean was missed. This patch adds the logic and test on
it.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D91403
scf.parallel is currently not a good fit for tiling on tensors.
Instead provide a path to parallelism directly through scf.for.
For now, this transformation ignores the distribution scheme and always does a block-cyclic mapping (where block is the tile size).
Differential revision: https://reviews.llvm.org/D90475
Support multi-dimension vector for InsertMap/ExtractMap op and update the
transformations. Currently the relation between IDs and dimension is implicitly
deduced from the types. We can then calculate an AffineMap based on it. In the
future the AffineMap could be part of the operation itself.
Differential Revision: https://reviews.llvm.org/D90995
Depends On D89958
1. Adds `async.group`/`async.awaitall` to group together multiple async tokens/values
2. Rewrite scf.parallel operation into multiple concurrent async.execute operations over non overlapping subranges of the original loop.
Example:
```
scf.for (%i, %j) = (%lbi, %lbj) to (%ubi, %ubj) step (%si, %sj) {
"do_some_compute"(%i, %j): () -> ()
}
```
Converted to:
```
%c0 = constant 0 : index
%c1 = constant 1 : index
// Compute blocks sizes for each induction variable.
%num_blocks_i = ... : index
%num_blocks_j = ... : index
%block_size_i = ... : index
%block_size_j = ... : index
// Create an async group to track async execute ops.
%group = async.create_group
scf.for %bi = %c0 to %num_blocks_i step %c1 {
%block_start_i = ... : index
%block_end_i = ... : index
scf.for %bj = %c0 t0 %num_blocks_j step %c1 {
%block_start_j = ... : index
%block_end_j = ... : index
// Execute the body of original parallel operation for the current
// block.
%token = async.execute {
scf.for %i = %block_start_i to %block_end_i step %si {
scf.for %j = %block_start_j to %block_end_j step %sj {
"do_some_compute"(%i, %j): () -> ()
}
}
}
// Add produced async token to the group.
async.add_to_group %token, %group
}
}
// Await completion of all async.execute operations.
async.await_all %group
```
In this example outer loop launches inner block level loops as separate async
execute operations which will be executed concurrently.
At the end it waits for the completiom of all async execute operations.
Reviewed By: ftynse, mehdi_amini
Differential Revision: https://reviews.llvm.org/D89963
The index type does not have a bitsize and hence the size of corresponding allocations cannot be computed. Instead, the promotion pass now has an explicit option to specify the size of index.
Differential Revision: https://reviews.llvm.org/D91360
This exposes a hook to configure legality of operations such that only
`scf.parallel` operations that have mapping attributes are marked as
illegal. Consequently, the transformation can now also be applied to
mixed forms.
Differential Revision: https://reviews.llvm.org/D91340
This revision adds support in the parser/printer for "deferrable" aliases, i.e. those that can be resolved after printing has finished. This allows for printing aliases for operation locations after the module instead of before, i.e. this is now supported:
```
"foo.op"() : () -> () loc(#loc)
#loc = loc("some_location")
```
Differential Revision: https://reviews.llvm.org/D91227
07f1047f41 changed the CMake detection to use find_package(Python3 ...
but didn't update the lit configuration to use the expected Python3_EXECUTABLE
cmake variable to point to the interpreter path.
This resulted in an empty path on MacOS.
We lower them to a std.global_memref (uniqued by constant value) + a
std.get_global_memref to produce the corresponding memref value.
This allows removing Linalg's somewhat hacky lowering of tensor
constants, now that std properly supports this.
Differential Revision: https://reviews.llvm.org/D91306
It was incorrect in the presence of a tensor argument with multiple
uses.
The bufferization of subtensor_insert was writing into a converted
memref operand, but there is no guarantee that the converted memref for
that operand is safe to write into. In this case, the same converted
memref is written to in-place by the subtensor_insert bufferization,
violating the tensor-level semantics.
I left some comments in a TODO about ways forward on this. I will be
working actively on this problem in the coming days.
Differential Revision: https://reviews.llvm.org/D91371
The tokens are already handled by the lexer. This revision exposes them
through the parser interface.
This revision also adds missing functions for question mark parsing and
completes the list of valid punctuation tokens in the documentation.
Differential Revision: https://reviews.llvm.org/D90907
Add an ODS-backed generator of default builders. This currently does not
support operation with attribute arguments, for which the builder is
just ignored. Attribute support will be introduced separately for
builders and accessors.
Default builders are always generated with the same number of result and
operand groups as the ODS specification, i.e. one group per each operand
or result. Optional elements accept None but cannot be omitted. Variadic
groups accept iterable objects and cannot be replaced with a single
object.
For some operations, it is possible to infer the result type given the
traits, but most traits rely on inline pieces of C++ that we cannot
(yet) forward to Python bindings. Since the Ops where the inference is
possible (having the `SameOperandAndResultTypes` trait or
`TypeMatchesWith` without transform field) are a small minority, they
also require the result type to make the builder syntax more consistent.
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D91190
This change does two main things
1) An operation might have multiple dependences to the same
producer. Not tracking them correctly can result in incorrect code
generation with fusion. To rectify this the dependence tracking
needs to also have the operand number in the consumer.
2) Improve the logic used to find the fused loops making it easier to
follow. The only constraint for fusion is that linalg ops (on
buffers) have update semantics for the result. Fusion should be
such that only one iteration of the fused loop (which is also a
tiled loop) must touch only one (disjoint) tile of the output. This
could be relaxed by allowing for recomputation that is the default
when oeprands are tensors, or can be made legal with promotion of
the fused view (in future).
Differential Revision: https://reviews.llvm.org/D90579
This CL integrates the new sparse annotations (hereto merely added as fully
transparent attributes) more tightly to the generic linalg op in order to add
verification of the annotations' consistency as well as to make make other
passes more aware of their presence (in the long run, rewriting rules must
preserve the integrity of the annotations).
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D91224
Previous the textual form of the pass pipeline would implicitly nest,
instead we opt for the explicit form here: this has less surprise.
This also avoids asserting in the bindings when passing a pass pipeline
with incorrect nesting.
Differential Revision: https://reviews.llvm.org/D91233
If block A and B are in different regions and region of A is not an ancestor of
B, either A is included in region of B or the two regions are disjoint. In both
case A doesn't post-dominate B.
Differential Revision: https://reviews.llvm.org/D91225
The previous logic for inlining a region A with N blocks into region B
would produce incorrect results on rollback for N greater than 1. This
rollback logic would leave blocks 1..N in region B and only move block 0
to region A.
The new inlining action recording stores the block move actions from N-1
to 0. Now on roll back, block 0 is moved to region A and then 1..N is
appended to the list of blocks in region A.
Differential Revision: https://reviews.llvm.org/D91185
I would like to use this for D90589 to switch std.alloc to assemblyFormat.
Hopefully it will be useful in other places as well.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D91068
This patch converts elementwise ops on tensors to linalg.generic ops
with the same elementwise op in the payload (except rewritten to
operate on scalars, obviously). This is a great form for later fusion to
clean up.
E.g.
```
// Compute: %arg0 + %arg1 - %arg2
func @f(%arg0: tensor<?xf32>, %arg1: tensor<?xf32>, %arg2: tensor<?xf32>) -> tensor<?xf32> {
%0 = addf %arg0, %arg1 : tensor<?xf32>
%1 = subf %0, %arg2 : tensor<?xf32>
return %1 : tensor<?xf32>
}
```
Running this through
`mlir-opt -convert-std-to-linalg -linalg-fusion-for-tensor-ops` we get:
```
func @f(%arg0: tensor<?xf32>, %arg1: tensor<?xf32>, %arg2: tensor<?xf32>) -> tensor<?xf32> {
%0 = linalg.generic {indexing_maps = [#map0, #map0, #map0, #map0], iterator_types = ["parallel"]} ins(%arg0, %arg1, %arg2 : tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) {
^bb0(%arg3: f32, %arg4: f32, %arg5: f32): // no predecessors
%1 = addf %arg3, %arg4 : f32
%2 = subf %1, %arg5 : f32
linalg.yield %2 : f32
} -> tensor<?xf32>
return %0 : tensor<?xf32>
}
```
So the elementwise ops on tensors have nicely collapsed into a single
linalg.generic, which is the form we want for further transformations.
Differential Revision: https://reviews.llvm.org/D90354
This patch adds an `ElementwiseMappable` trait as discussed in the RFC
here:
https://llvm.discourse.group/t/rfc-std-elementwise-ops-on-tensors/2113/23
This trait can power a number of transformations and analyses.
A subsequent patch adds a convert-elementwise-to-linalg pass exhibits
how this trait allows writing generic transformations.
See https://reviews.llvm.org/D90354 for that patch.
This trait slightly changes some verifier messages, but the diagnostics
are usually about as good. I fiddled with the ordering of the trait in
the .td file trait lists to minimize the changes here.
Differential Revision: https://reviews.llvm.org/D90731
This only exposes the ability to round-trip a textual pipeline at the
moment.
To exercise it, we also bind the libTransforms in a new Python extension. This
does not include any interesting bindings, but it includes all the
mechanism to add separate native extensions and load them dynamically.
As such passes in libTransforms are only registered after `import
mlir.transforms`.
To support this global registration, the TableGen backend is also
extended to bind to the C API the group registration for passes.
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D90819
This patch introduces a new conversion pattern for `spv.ExecutionMode`.
`spv.ExecutionMode` may contain important information about the entry
point, which we want to preserve. For example, `LocalSize` provides
information about the work-group size that can be reused. Hence, the
pattern for entry-point ops changes to the following:
- `spv.EntryPoint` is still simply removed
- Info from `spv.ExecutionMode` is used to create a global struct variable,
which looks like:
```
struct {
int32_t executionMode;
int32_t values[]; // optional values
};
```
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D89989
Introduce an ODS/Tablegen backend producing Op wrappers for Python bindings
based on the ODS operation definition. Usage:
mlir-tblgen -gen-python-op-bindings -Iinclude <path/to/Ops.td> \
-bind-dialect=<dialect-name>
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D90960
Slicing, that is element access with `[being🔚step]` structure, is
a common Python idiom for sequence-like containers. It is also necessary
to support custom accessor for operations with variadic operands and
results (an operation an return a slice of its operands that correspond
to the given variadic group).
Add generic utility to support slicing in Python bindings and use it
for operation operands and results.
Depends On D90923
Reviewed By: stellaraccident, mehdi_amini
Differential Revision: https://reviews.llvm.org/D90936
VectorInsertDynamicOp in SPIRV dialect
conversion from vector.insertelement to spirv VectorInsertDynamicOp
Differential Revision: https://reviews.llvm.org/D90927
Locations often get very long and clutter up operations when printed inline with them. This revision adds support for using aliases with trailing operation locations, and makes printing with aliases the default behavior. Aliases in the trailing location take the form `loc(<alias>)`, such as `loc(#loc0)`. As with all aliases, using `mlir-print-local-scope` can be used to disable them and get the inline behavior.
Differential Revision: https://reviews.llvm.org/D90652
This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`.
This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation.
To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter.
Differential Revision: https://reviews.llvm.org/D90512
Tei Jeong