Composite pass allows to run sequence of passes in the loop until fixed
point or maximum number of iterations is reached. The usual candidates
are canonicalize+CSE as canonicalize can open more opportunities for CSE
and vice-versa.
This patch adds a the `LowerVectorToArmNeonPattern` patterns to the
ArmNeon.
This pattern inspects `vector.contract` ops that can be 1-1 mapped to an
`arm.neon.smmla` intrinsic. The contract ops must be separated into
tiles who's inputs must fit that of a single smmla op (`2x8xi32` inputs
and `2x2xi32` output). The `vector.contract` inputs must be sign
extended from narrow types (<=i8) to be converted. If all conditions are
met, an smmla op is inserted with additional `vector.shape_casts` to
handle linearizing the input and output dimension.
This is another follow-up of #83004, which made the same change for
`MLIR_CUDA_CONVERSIONS_ENABLED`. As the previous PR, this PR commit
exposes mentioned CMake variable through `mlir-config.h` and uses the
macro that is introduced with the same name. This replaces the macro
`MLIR_ENABLE_DEPRECATED_GPU_SERIALIZATION`, which the CMake files
previously defined manually.
`isContiguousAccess` is an important affine analysis utility but is only
tested very indirectly via passes like vectorization and is not exposed.
Expose it and add a test pass for it that'll make it easier/feasible to
write test cases. This is especially needed since the utility can be
significantly enhanced in power, and we need a test pass to exercise it
directly.
This pass can in the future be used to test the utility of invariant
accesses as well.
The ArmSME compilation pipeline has evolved significantly and is now
sufficiently complex enough that it warrants a proper lowering pipeline
that encapsulates the various passes and orderings. Currently the
pipeline is loosely defined in our integration tests, but these have
diverged and are not using the same passes or ordering everywhere.
This patch introduces a test-lower-to-arm-sme pipeline mirroring
test-lower-to-llvm that provides some sanity when running e2e examples
and can be used a reference for targeting ArmSME in MLIR.
All the integration tests are updated to use this pipeline. The
intention is to productize the pipeline once it becomes more mature.
This op is the inverse of all-gather. It is useful to have an explicit
concise representation instead of having a blob of slicing logic.
Add lowering for the op that slices from the tensor based on the
in-group process index.
Make resharding generate an all-slice instead of inserting the slicing
logic directly.
This removes 3 dead function decls from mlir-opt:
- registerTestLowerToNVVM - recently removed in #75775 when NVVM was
productized.
- registerTestPreparationPassWithAllowedMemrefResults - removed in
D90778 (f7bc568266).
- registerTestGenericIRVisitorsInterruptPass - added in D116230
(8067ced144) but never existed. Pass is registered by
registerTestGenericIRVisitorsPass.
In the process a couple of test transform dialect ops are added just
for testing. These operations are not intended to use as full flushed
out of transformation ops, but are rather operations added for testing.
A separate operation is added to `LinalgTransformOps.td` to convert a
`TilingInterface` operation to loops using the
`generateScalarImplementation` method implemented by the
operation. Eventually this and other operations related to tiling
using the `TilingInterface` need to move to a better place (i.e. out
of `Linalg` dialect)
* Rename mesh.process_index -> mesh.process_multi_index.
* Add mesh.process_linear_index op.
* Add lowering of mesh.process_multi_index into an expression using
mesh.process_linear_index, mesh.cluster_shape and
affine.delinearize_index.
This is useful to lower mesh ops and prepare them for further lowering
where the runtime may have only the linear index of a device/process.
For example in MPI we have a rank (linear index) in a communicator.
Support WalkResult for AffineExpr walk and support interrupting walks
along the lines of Operation::walk. This allows interrupted walks when a
condition is met. Also, switch from std::function to llvm::function_ref
for the walk function.
Make it so that PDL in pattern rewrites can be optionally disabled.
PDL is still enabled by default and not optional bazel. So this should
be a NOP for most folks, while enabling other to disable.
This only works with tests disabled. With tests enabled this still
compiles but tests fail as there is no lit config to disable tests that
depend on PDL rewrites yet.
Make it so that PDL in pattern rewrites can be optionally disabled.
PDL is still enabled by default and not optional bazel. So this should
be a NOP for most folks, while enabling other to disable.
This is piped through mlir-tblgen invocation and that could be
changed/avoided by splitting up the passes file instead.
This only works with tests disabled. With tests enabled this still
compiles but tests fail as there is no lit config to disable tests that
depend on PDL rewrites yet.
The new patterns break down subgroup reduce ops with vector values into
a sequence of subgroup reductions that fit the native shuffle size. The
maximum/native shuffle size is parametrized.
The overall goal is to be able to perform multi-element reductions with
a sequence of `gpu.shuffle` ops.
The `test-lower-to-nvvm` pipeline serves as the common and proper
pipeline for nvvm+host compilation, and it's used across our CUDA
integration tests.
This PR updates the `test-lower-to-nvvm` pipeline to `gpu-lower-to-nvvm`
and moves it within `InitAllPasses.h`. The aim is to call it from
Python, also having a standardize compilation process for nvvm.
I was not able to fully triage why this just started failing on one of
our bots as it seems that the use was added 4 months ago. I would assume
that it was accidentally coming in transitively in some way as the dep
was definitely missing.
For context, this started failing in [our
byo_llvm](https://github.com/openxla/iree/blob/main/build_tools/llvm/byo_llvm.sh)
build on a stock build of MLIR on top of an existing LLVM. We were
getting:
```
ld.lld: error: undefined symbol: mlir::registerSPIRVDialectTranslation(mlir::DialectRegistry&) >>> referenced by mlir-opt.cpp
>>> tools/mlir-opt/CMakeFiles/mlir-opt.dir/mlir-opt.cpp.o:(main)
```
Does transformations like
all_reduce(x) + all_reduce(y) -> all_reduce(x + y)
max(all_reduce(x), all_reduce(y)) -> all_reduce(max(x, y))
when the all_reduce element-wise op is max.
Added general rewrite pattern HomomorphismSimplification and
EndomorphismSimplification that encapsulate the general algorithm.
Made specialization for all-reduce with respect to
addf, addi, minsi, maxsi, minimumf and maximumf
in the Arithmetic dialect.
When emitting bytecode, clients can specify a target dialect version to
emit in `BytecodeWriterConfig`. This exposes a target dialect version to
the DialectBytecodeWriter, which can be queried by name and used to
back-deploy attributes, types, and properties.
Deprecate the `gpu-to-cubin` & `gpu-to-hsaco` passes in favor of the
`TargetAttr` workflow. This patch removes remaining upstream uses of the
aforementioned passes, including the option to use them in `mlir-opt`. A
future patch will remove these passes entirely.
The passes can be re-enabled in `mlir-opt` by adding the CMake flag: `-DMLIR_ENABLE_DEPRECATED_GPU_SERIALIZATION=1`.
This revision avoids the registration of dialect extensions in Pass::getDependentDialects.
Such registration of extensions can be dangerous because `DialectRegistry::isSubsetOf` is
always guaranteed to return false for extensions (i.e. there is no mechanism to track
whether a lambda is already in the list of already registered extensions).
When the context is already in a multi-threaded mode, this is guaranteed to assert.
Arguably a more structured registration mechanism for extensions with a unique ExtensionID
could be envisioned in the future.
In the process of cleaning this up, multiple usage inconsistencies surfaced around the
registration of translation extensions that this revision also cleans up.
Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D157703
[mlir] Expose a mechanism to provide a callback for encoding types and attributes in MLIR bytecode.
Two callbacks are exposed, respectively, to the BytecodeWriterConfig and to the ParserConfig. At bytecode parsing/printing, clients have the ability to specify a callback to be used to optionally read/write the encoding. On failure, fallback path will execute the default parsers and printers for the dialect.
Testing shows how to leverage this functionality to support back-deployment and backward-compatibility usecases when roundtripping to bytecode a client dialect with type/attributes dependencies on upstream.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D153383
[mlir] Expose a mechanism to provide a callback for encoding types and attributes in MLIR bytecode.
Two callbacks are exposed, respectively, to the BytecodeWriterConfig and to the ParserConfig. At bytecode parsing/printing, clients have the ability to specify a callback to be used to optionally read/write the encoding. On failure, fallback path will execute the default parsers and printers for the dialect.
Testing shows how to leverage this functionality to support back-deployment and backward-compatibility usecases when roundtripping to bytecode a client dialect with type/attributes dependencies on upstream.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D153383
This commit adds a utility to implement liveness analysis using the
sparse backward data-flow analysis framework. Theoretically, liveness
analysis assigns liveness to each (value, program point) pair in the
program and it is thus a dense analysis. However, since values are
immutable in MLIR, a sparse analysis, which will assign liveness to
each value in the program, suffices here.
Liveness analysis has many applications. It can be used to avoid the
computation of extraneous operations that have no effect on the memory
or the final output of a program. It can also be used to optimize
register allocation. Both of these applications help achieve one very
important goal: reducing runtime.
A value is considered "live" iff it:
(1) has memory effects OR
(2) is returned by a public function OR
(3) is used to compute a value of type (1) or (2).
It is also to be noted that a value could be of multiple types (1/2/3) at
the same time.
A value "has memory effects" iff it:
(1.a) is an operand of an op with memory effects OR
(1.b) is a non-forwarded branch operand and a block where its op could
take the control has an op with memory effects.
A value `A` is said to be "used to compute" value `B` iff `B` cannot be
computed in the absence of `A`. Thus, in this implementation, we say that
value `A` is used to compute value `B` iff:
(3.a) `B` is a result of an op with operand `A` OR
(3.b) `A` is used to compute some value `C` and `C` is used to compute
`B`.
---
It is important to note that there already exists an MLIR liveness
utility here: llvm-project/mlir/include/mlir/Analysis/Liveness.h. So,
what is the need for this new liveness analysis utility being added by
this commit? That need is explained as follows:-
The similarities between these two utilities is that both use the
fixpoint iteration method to converge to the final result of liveness.
And, both have the same theoretical understanding of liveness as well.
However, the main difference between (a) the existing utility and (b)
the added utility is the "scope of the analysis". (a) is restricted to
analysing each block independently while (b) analyses blocks together,
i.e., it looks at how the control flows from one block to the other,
how a caller calls a callee, etc. The restriction in the former implies
that some potentially non-live values could be marked live and thus the
full potential of liveness analysis will not be realised.
This can be understood using the example below:
```
1 func.func private @private_dead_return_value_removal_0() -> (i32, i32) {
2 %0 = arith.constant 0 : i32
3 %1 = arith.addi %0, %0 : i32
4 return %0, %1 : i32, i32
5 }
6 func.func @public_dead_return_value_removal_0() -> (i32) {
7 %0:2 = func.call @private_dead_return_value_removal_0() : () -> (i32, i32)
8 return %0#0 : i32
9 }
```
Here, if we just restrict our analysis to a per-block basis like (a), we
will say that the %1 on line 3 is live because it is computed and then
returned outside its block by the function. But, if we perform a
backward data-flow analysis like (b) does, we will say that %0#1 of line
7 is not live because it isn't returned by the public function and thus,
%1 of line 3 is also not live. So, while (a) will be unable to suggest
any IR optimizations, (b) can enable this IR to convert to:-
```
1 func.func private @private_dead_return_value_removal_0() -> i32 {
2 %0 = arith.constant 0 : i32
3 return %0 : i32
4 }
5 func.func @public_dead_return_value_removal_0() -> i32 {
6 %0 = call @private_dead_return_value_removal_0() : () -> i32
7 return %0 : i32
8 }
```
One operation was removed and one unnecessary return value of the
function was removed and the function signature was modified. This is an
optimization that (b) can enable but (a) cannot. Such optimizations can
help remove a lot of extraneous computations that are currently being
done.
Signed-off-by: Srishti Srivastava <srishtisrivastava.ai@gmail.com>
Reviewed By: matthiaskramm, jcai19
Differential Revision: https://reviews.llvm.org/D153779
TL;DR the following API functions have been merged
```
void populateFoldUnitExtentDimsViaReshapesPatterns(RewritePatternSet &patterns);
void populateFoldUnitExtentDimsViaSlicesPatterns(RewritePatternSet &patterns);
```
into
```
void populateFoldUnitExtentDimsPatterns(RewritePatternSet &patterns,
ControlDropUnitDims &options);
```
To use the previous functionality use
```
ControlDropUnitDims options;
// By default options.rankReductionStrategy is
// ControlDropUnitDims::RankReductionStrategy::ReassociativeReshape.
populateFoldUnitExtentDimsPatterns(patterns, options);
```
and
```
ControlDropUnitDims options;
options.rankReductionStrategy = ControlDropUnitDims::RankReductionStrategy::ExtractInsertSlice
populateFoldUnitExtentDimsPatterns(patterns, options);
```
This pass is quite old and needed to be updated based on the current
approach to transformations in Linalg
- Instead of two patterns, one to just remove loop dimensions that are
unit extent (and using 0 in the indexing maps), and another to drop
the unit-extents in the operand shapes, combine into a single
transformation. This avoid creating an intermediate step with
indexing maps having 0's in the domains exp ressions.
- Expose the core transformation as a utility function and add a
pattern that calls this transformation.
This is a mostly NFC change, apart from the API change and dropping
the patterns/test that only dropped the loops that are unit extents.
Differential Revision: https://reviews.llvm.org/D155518
This mirrors the test-lower-to-llvm pass pipeline that provides some sanity when running e2e examples.
One peculiarity of the GPU pipeline is that we want to allow 32b indexing in kernels.
This is currently not straightforward as there are dependencies between passes.
This new test pass orders passes in a way that connects end-to-end.
Differential Revision: https://reviews.llvm.org/D155463
This is the counterpart to the forward dense dataflow analysis and
integrates into the dataflow framework. The implementation follows the
structure of existing dataflow analyses.
Reviewed By: Mogball, phisiart
Differential Revision: https://reviews.llvm.org/D154713
A distinct attribute associates a referenced attribute with a unique
identifier. Every call to its create function allocates a new
distinct attribute instance. The address of the attribute instance
temporarily serves as its unique identifier. Similar to the names
of SSA values, the final unique identifiers are generated during
pretty printing.
Examples:
#distinct = distinct[0]<42.0 : f32>
#distinct1 = distinct[1]<42.0 : f32>
#distinct2 = distinct[2]<array<i32: 10, 42>>
This mechanism is meant to generate attributes with a unique
identifier, which can be used to mark groups of operations
that share a common properties such as if they are aliasing.
The design of the distinct attribute ensures minimal memory
footprint per distinct attribute since it only contains a reference
to another attribute. All distinct attributes are stored outside of
the storage uniquer in a thread local store that is part of the
context. It uses one bump pointer allocator per thread to ensure
distinct attributes can be created in-parallel.
Reviewed By: rriddle, Dinistro, zero9178
Differential Revision: https://reviews.llvm.org/D153360
This patch adds support for narrow bitwidth storage emulation. The goal is to support sub-byte type
codegen for LLVM CPU. Specifically, a type converter is added to convert memref of narrow bitwidth
(e.g., i4) into supported wider bitwidth (e.g., i8). Another focus of this patch is to populate the
pattern for int4 memref.load. memref.store pattern should be added in a seperate patch.
Reviewed By: hanchung, mravishankar
Differential Revision: https://reviews.llvm.org/D151519
Promised interfaces allow for a dialect to "promise" the implementation of an interface, i.e.
declare that it supports an interface, but have the interface defined in an extension in a library
separate from the dialect itself. A promised interface is powerful in that it alerts the user when
the interface is attempted to be used (e.g. via cast/dyn_cast/etc.) and the implementation has
not yet been provided. This makes the system much more robust against misconfiguration,
and ensures that we do not lose the benefit we currently have of defining the interface in
the dialect library.
Differential Revision: https://reviews.llvm.org/D120368
This patch implements a mechanism to read/write use-list orders from/to the mlir bytecode format. When producing bytecode, use-list orders are appended to each value of the IR. When reading bytecode, use-lists orders are loaded in memory and used at the end of parsing to sort the existing use-list chains.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D149755
IsolatedRegions are emitted in sections in order for the reader to be
able to skip over them. A new class is exposed to manage the state and
allow the readers to load these IsolatedRegions on-demand.
Differential Revision: https://reviews.llvm.org/D149515
The utility functions takes a region and makes it isolated from above
by appending to the entry block arguments that represent the captured
values and replacing all uses of the captured values within the region
with the newly added arguments. The captures values are returned.
The utility function also takes an optional callback that allows
cloning operations that define the captured values into the region
during the process of making it isolated from above. The cloned value
is no longer a captured values. The operands of the operation are then
captured values. This is done transitively allow cloning of a DAG of
operations into the region based on the callback.
Reviewed By: jpienaar
Differential Revision: https://reviews.llvm.org/D148684
Ops can implement this interface to specify lower/upper bounds for their result values and block arguments. Bounds can be specified for:
* Index-type values
* Dimension sizes of shapes values
The bounds are added to a constraint set. Users can query this constraint set to compute bounds wrt. to a user-specified set of values. Only EQ bounds are supported at the moment.
This revision also contains interface implementations for various tensor dialect ops, which illustrates how to implement this interface.
Differential Revision: https://reviews.llvm.org/D145681
This commit introduces an instantiation of LLVM's LoopInfo for CFGs in
MLIR. To test the LoopInfo, a test pass is added the checks the analysis
results for a set of CFGs.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D147323
The current dialect conversion does not support 1:N type conversions.
This commit implements a (poor-man's) dialect conversion pass that does
just that. To keep the pass independent of the "real" dialect conversion
infrastructure, it provides a specialization of the TypeConverter class
that allows for N:1 target materializations, a specialization of the
RewritePattern and PatternRewriter classes that automatically add
appropriate unrealized casts supporting 1:N type conversions and provide
converted operands for implementing subclasses, and a conversion driver
that applies the provided patterns and replaces the unrealized casts
that haven't folded away with user-provided materializations.
The current pass is powerful enough to express many existing manual
solutions for 1:N type conversions or extend transforms that previously
didn't support them, out of which this patch implements call graph type
decomposition (which is currently implemented with a ValueDecomposer
that is only used there).
The goal of this pass is to illustrate the effect that 1:N type
conversions could have, gain experience in how patterns should be
written that achieve that effect, and get feedback on how the APIs of
the dialect conversion should be extended or changed to support such
patterns. The hope is that the "real" dialect conversion eventually
supports such patterns, at which point, this pass could be removed
again.
Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D144469
The current dialect conversion does not support 1:N type conversions.
This commit implements a (poor-man's) dialect conversion pass that does
just that. To keep the pass independent of the "real" dialect conversion
infrastructure, it provides a specialization of the TypeConverter class
that allows for N:1 target materializations, a specialization of the
RewritePattern and PatternRewriter classes that automatically add
appropriate unrealized casts supporting 1:N type conversions and provide
converted operands for implementing subclasses, and a conversion driver
that applies the provided patterns and replaces the unrealized casts
that haven't folded away with user-provided materializations.
The current pass is powerful enough to express many existing manual
solutions for 1:N type conversions or extend transforms that previously
didn't support them, out of which this patch implements call graph type
decomposition (which is currently implemented with a ValueDecomposer
that is only used there).
The goal of this pass is to illustrate the effect that 1:N type
conversions could have, gain experience in how patterns should be
written that achieve that effect, and get feedback on how the APIs of
the dialect conversion should be extended or changed to support such
patterns. The hope is that the "real" dialect conversion eventually
supports such patterns, at which point, this pass could be removed
again.
Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D144469
Care is taken to order operands from least hoistable to most hoistable and to process subexpressions in the same
order.
This allows exposing more oppportunities for licm, cse and strength reduction.
Such a step should typically be applied while we still have loops in the IR and just before lowering affine ops to arith.
This is because the affine.apply canonicalization currently tries to maximally compose chains of affine.apply operations
and could undo the effects of these decompositions.
Depends on: D145784
Differential Revision: https://reviews.llvm.org/D145685
This converts a specific form of `vector.reduction` to SPIR-V integer
dot product ops.
Add a new test pass to excercise this outside of the main vector to
spirv conversion pass.
Reviewed By: antiagainst
Differential Revision: https://reviews.llvm.org/D145760
This function returns whether a block is nested inside of a loop. There
can be three kinds of loop:
1) The block is nested inside of a LoopLikeOpInterface
2) The block is nested inside another block which is in a loop
3) There is a cycle in the control flow graph
This will be useful for Flang's stack arrays pass, which moves array
allocations from the heap to the stack. Special handling is needed when
allocations occur inside of loops to ensure additional stack space is
not allocated on each loop iteration.
Differential Revision: https://reviews.llvm.org/D141401
This function returns whether a block is nested inside of a loop. There
can be three kinds of loop:
1) The block is nested inside of a LoopLikeOpInterface
2) The block is nested inside another block which is in a loop
3) There is a cycle in the control flow graph
This will be useful for Flang's stack arrays pass, which moves array
allocations from the heap to the stack. Special handling is needed when
allocations occur inside of loops to ensure additional stack space is
not allocated on each loop iteration.
Differential Revision: https://reviews.llvm.org/D141401
The overload of WhileOp::build with arguments for builder functions for
the regions of the op was broken: It did not compute correctly the types
(and locations) of the region arguments, which lead to failed assertions
when the result types were different from the operand types.
Specifically, it used the result types (and operand locations) for *both*
regions, instead of the operand types (and locations) for the 'before'
region and the result types (and loecations) for the 'after' region.
Reviewed By: Mogball, mehdi_amini
Differential Revision: https://reviews.llvm.org/D142952
