It is time to compose Linalg related optimizations with SparseTensor
related optimizations. This is a careful first start by adding some
general Linalg optimizations "upstream" of the sparse compiler in the
full sparse compiler pipeline. Some minor changes were needed to make
those optimizations aware of sparsity.
Note that after this, we will add a sparse specific fusion rule,
just to demonstrate the power of the new composition.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D119971
In SPIR-V, resources are represented as global variables that
are bound to certain descriptor. SPIR-V requires those global
variables to be declared as aliased if multiple ones are bound
to the same slot. Such aliased decorations can cause issues
for transcompilers like SPIRV-Cross when converting to source
shading languages like MSL.
So this commit adds a pass to perform analysis of aliased
resources and see if we can unify them into one.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D119872
This would create a double free when a memref is passed twice to the
same op. This wasn't a problem at the time the pass was written but is
common since the introduction of scf.while.
There's a latent non-determinism that's triggered by the test, but this
change is messy enough as-is so I'll leave that for later.
Differential Revision: https://reviews.llvm.org/D120044
During dialect conversion, target materialization is triggered to create
cast-like operations when a type mismatch occurs between the value that
replaces a rewritten operation and the type that another operations expects as
operands processed by the type conversion. First, a dummy cast is inserted to
make sure the pattern application can proceed. The decision to trigger the
user-provided materialization hook is taken later based on the result of the
dummy cast having uses. However, it only has uses if other patterns constructed
new operations using the casted value as operand. If existing (legal)
operations use the replaced value, they may have not been updated to use the
casted value yet. The conversion infra would then delete the dummy cast first,
and then would replace the uses with now-invalid (null in the bast case) value.
When deciding whether to trigger cast materialization, check for liveness the
uses not only of the casted value, but also of all the values that it replaces.
This was discovered in the finalizing bufferize pass that cleans up
mutually-cancelling casts without touching other operations. It is not
impossible that there are other scenarios where the dialect converison infra
could produce invalid operand uses because of dummy casts erased too eagerly.
Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D119937
Previously `gpu-kernel-outlining` pass was also doing index computation sinking into gpu.launch before actual outlining.
Split ops sinking from `gpu-kernel-outlining` pass into separate pass, so users can use theirs own sinking pass before outlining.
To achieve old behavior users will need to call both passes: `-gpu-launch-sink-index-computations -gpu-kernel-outlining`.
Differential Revision: https://reviews.llvm.org/D119932
Currently some of the nested IR building inconsistently uses `nb` and `b`, it's very easy to call wrong builder outside of the current scope, so for simplicity all builders are always called `b`, and in nested IR building regions they just shadow the "parent" builder.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D120003
A very small refactoring, but a big impact on tests that expect an exact order.
This revision fixes the tests, but also makes them less brittle for similar
minor changes in the future!
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D119992
This commit adds a pattern to wrap a tensor.pad op with
an scf.if op to separate the cases where we don't need padding
(all pad sizes are actually zeros) and where we indeed need
padding.
This pattern is meant to handle padding inside tiled loops.
Under such cases the padding sizes typically depend on the
loop induction variables. Splitting them would allow treating
perfect tiles and edge tiles separately.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D117018
The pad-slice swap pattern generates `scf.if` and `tensor.generate`
to guard against zero-sized slices if it cannot prove the slice is
always non-zero. This is safe but quite conservative. It can be
unnecessary for cases where we know by problem definition such cases
does not exist, even if with dynamic shaped ops or unknown tile/slice
sizes, e.g., convolution padding size = 1 with kernel dim size = 3.
So this commit introduces a control to the pattern to specify
whether to generate the if constructs to handle such cases better,
given that once the if constructs is materialized, it's very hard
to analyze and simplify.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D117017
Fusion of `linalg.generic` with
`tensor.expand_shape/tensor.collapse_shape` currently handles fusion
with reshape by expanding the dimensionality of the `linalg.generic`
operation. This helps fuse elementwise operations better since they
are fused at the highest dimensionality while keeping all indexing
maps involved projected permutations. The intent of these is to push
the reshape to the boundaries of functions.
The presence of named ops (or other ops across which the reshape
cannot be propagated) stops the propagation to the edges of the
function. At this stage, the converse patterns that fold the reshapes
with generic ops by collapsing the dimensions of the generic op can
push the reshape towards edges. In particular it helps the case where
reshapes exist in between named ops and generic ops.
`linalg.named_op` -> `tensor.expand_shape` -> `linalg.generic`
Pushing the reshape down will help fusion of `linalg.named_op` ->
`linalg.generic` using tile + fuse transformations.
This pattern is intended to replace the following patterns
1) FoldReshapeByLinearization : These patterns create indexing maps
that are not projected permutations that affect future
transformations. They are only useful for folding unit-dimensions.
2) PushReshapeByExpansion : This pattern has the same functionality
but has some restrictions
a) It tries to avoid creating new reshapes that limits its
applicability. The pattern added here can achieve the same
functionality through use of the `controlFn` that allows clients
of the pattern freedom to make this decision.
b) It does not work for ops with indexing semantics.
These patterns will be deprecated in a future patch.
Differential Revision: https://reviews.llvm.org/D119365
This change exposes printer flags in AsmState and AsmStateImpl. All functions
receiving AsmState as a parameter now use the flags from the AsmState instead of
taking an additional OpPrintingFlags parameter.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D119870
This allow user to register a callback that can annotate operations
during software pipelining. This allows user potential annotate op to
know what part of the pipeline they correspond to.
Differential Revision: https://reviews.llvm.org/D119866
Without results, there is no getType injected and so generating one in prefixed form doesn't result in any failures during C++ compilation.
Differential Revision: https://reviews.llvm.org/D119871
Optional parameters with `defaultValue` set will be populated with that value if they aren't encountered during parsing. Moreover, parameters equal to their default values are elided when printing.
Depends on D118210
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D118544
Casting between scalable vectors and fixed-length vectors doesn't make
sense. If one of the operands is scalable, the other has to be scalable
to be able to guarantee they have the same shape at runtime.
Differential Revision: https://reviews.llvm.org/D119568
This patch changes the syntax of omp.atomic.update to allow the other
dialects to modify the variable with appropriate operations in the
region.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D119522
Add verifier for gpu.alloc op to verify if the dimension operand counts
and symbol operand counts are same as their memref counterparts.
Differential Revision: https://reviews.llvm.org/D117427
Support ALLOW filters and DENY filters. This is needed for compatibility with existing code that specifies more complex op filters.
Differential Revision: https://reviews.llvm.org/D119820
Also, it seems Khronos has changed html spec format so small adjustment to script was needed.
Base op parsing is also probably broken.
Differential Revision: https://reviews.llvm.org/D119678
This patch changes the argument from template-IRBuilder to IRBuilderBase
thus allowing us to write less code while getting the location from a
builder.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D119717
* While annoying, this is the only way to get C++ exception handling out of the happy path for normal iteration.
* Implements sq_length and sq_item for the sequence protocol (used for iteration, including list() construction).
* Implements mp_subscript for general use (i.e. foo[1] and foo[1:1]).
* For constructing a `list(op.results)`, this reduces the time from ~4-5us to ~1.5us on my machine (give or take measurement overhead) and eliminates C++ exceptions, which is a worthy goal in itself.
* Compared to a baseline of similar construction of a three-integer list, which takes 450ns (might just be measuring function call overhead).
* See issue discussed on the pybind side: https://github.com/pybind/pybind11/issues/2842
Differential Revision: https://reviews.llvm.org/D119691
Rationale:
empty line between main include for this file
moved include that actually defines code into right section
Note that this revision started as breaking up ops/attrs even more
(for bug https://github.com/llvm/llvm-project/issues/52748), but due
the the connection in Dialect.initalize(), this cannot be split further).
All heavy lifting refactoring was already done by River in previous cleanup.
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D119617
Adds a pointer type to EmitC. The emission of pointers is so far only
possible by using the `emitc.opaque` type
Co-authored-by: Simon Camphausen <simon.camphausen@iml.fraunhofer.de>
Reviewed By: jpienaar
Differential Revision: https://reviews.llvm.org/D119337
Adapt the region builder signature to hand in the attributes of the created ops. The revision is a preparation step the support named ops that need access to the operation attributes during op creation.
Depends On D119692
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D119693
The module flag to indicate use of hostcall is insufficient to catch
all cases where hostcall might be in use by a kernel. This is now
replaced by a function attribute that gets propagated to top-level
kernel functions via their respective call-graph.
If the attribute "amdgpu-no-hostcall-ptr" is absent on a kernel, the
default behaviour is to emit kernel metadata indicating that the
kernel uses the hostcall buffer pointer passed as an implicit
argument.
The attribute may be placed explicitly by the user, or inferred by the
AMDGPU attributor by examining the call-graph. The attribute is
inferred only if the function is not being sanitized, and the
implictarg_ptr does not result in a load of any byte in the hostcall
pointer argument.
Reviewed By: jdoerfert, arsenm, kpyzhov
Differential Revision: https://reviews.llvm.org/D119216
When lowering to memrefCopy call, the size for i1 type was calculated as 0.
Instead of using getTypeSizeInBits() and dividing by 8, we should just use getTypeSize().
Differential Revision: https://reviews.llvm.org/D119540
The lowering creates llvm.insertvalue with the rank value, so it needs to use
index type instead of 64 bit integer type. Otherwise, we get an error:
llvm.insertvalue' op Type mismatch: cannot insert 'i64' into '!llvm.struct<(i32, ptr<i8>)>'
Differential Revision: https://reviews.llvm.org/D119534
This patch simply adds an optional garbage collector attribute to LLVMFuncOp which maps 1:1 to the "gc" property of functions in LLVM.
Differential Revision: https://reviews.llvm.org/D119492
This allows operations to control the block ids used by the printer in nested regions.
Reviewed By: Mogball
Differential Revision: https://reviews.llvm.org/D115849
Add new operations to the gpu dialect to represent device side
asynchronous copies. This also add the lowering of those operations to
nvvm dialect.
Those ops are meant to be low level and map directly to llvm dialects
like nvvm or rocdl.
We can further add higher level of abstraction by building on top of
those operations.
This has been discuss here:
https://discourse.llvm.org/t/modeling-gpu-async-copy-ampere-feature/4924
Differential Revision: https://reviews.llvm.org/D119191
These functions allow for defining pattern fragments usable within the `match` and `rewrite` sections of a pattern. The main structure of Constraints and Rewrites functions are the same, and are similar to functions in other languages; they contain a signature (i.e. name, argument list, result list) and a body:
```pdll
// Constraint that takes a value as an input, and produces a value:
Constraint Cst(arg: Value) -> Value { ... }
// Constraint that returns multiple values:
Constraint Cst() -> (result1: Value, result2: ValueRange);
```
When returning multiple results, each result can be optionally be named (the result of a Constraint/Rewrite in the case of multiple results is a tuple).
These body of a Constraint/Rewrite functions can be specified in several ways:
* Externally
In this case we are importing an external function (registered by the user outside of PDLL):
```pdll
Constraint Foo(op: Op);
Rewrite Bar();
```
* In PDLL (using PDLL constructs)
In this case, the body is defined using PDLL constructs:
```pdll
Rewrite BuildFooOp() {
// The result type of the Rewrite is inferred from the return.
return op<my_dialect.foo>;
}
// Constraints/Rewrites can also implement a lambda/expression
// body for simple one line bodies.
Rewrite BuildFooOp() => op<my_dialect.foo>;
```
* In PDLL (using a native/C++ code block)
In this case the body is specified using a C++(or potentially other language at some point) code block. When building PDLL in AOT mode this will generate a native constraint/rewrite and register it with the PDL bytecode.
```pdll
Rewrite BuildFooOp() -> Op<my_dialect.foo> [{
return rewriter.create<my_dialect::FooOp>(...);
}];
```
Differential Revision: https://reviews.llvm.org/D115836
This allows for defining simple patterns in a single line. The lambda
body of a Pattern expects a single operation rewrite statement:
```
Pattern => replace op<my_dialect.foo>(operands: ValueRange) with operands;
```
Differential Revision: https://reviews.llvm.org/D115835
Having clarified that executing the SerializeToHsaco pass can
depend on a ROCm installation, switch from calling lld as a library to
using the copy of lld guaranteed to be included in a ROCm install.
This removes the workaround introduced in D119277
Reviewed By: whchung
Differential Revision: https://reviews.llvm.org/D119463
If the result operand has a unit leading dim it is removed from all operands.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D119206
This patch factors out space information from IntegerPolyhedron, PresburgerSet
and PWMAFunction to PresburgerSpace and its extension with local variables,
PresburgerLocalSpace.
Generally any new data structure additions in Presburger library will require
space information. This patch removes the need to duplicate the space
information.
Reviewed By: arjunp
Differential Revision: https://reviews.llvm.org/D119280
Fix fold-memref-subview-ops for affine.load/store. We need to expand out
the affine apply on its operands.
Differential Revision: https://reviews.llvm.org/D119402
Move expandAffineMap and expandAffineApplyExpr out to AffineUtils. This
is a useful method. The child revision uses it. NFC.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D119401
While testing LLVM 14.0.0 rc1 on Solaris, I ran into a compile failure:
from /var/llvm/llvm-14.0.0-rc1/rc1/llvm-project/mlir/lib/ExecutionEngine/SparseTensorUtils.cpp:22:
/usr/include/sys/types.h:103:16: error: conflicting declaration ‘typedef short int index_t’
103 | typedef short index_t;
| ^~~~~~~
In file included from
/var/llvm/llvm-14.0.0-rc1/rc1/llvm-project/mlir/lib/ExecutionEngine/SparseTensorUtils.cpp:17:
/var/llvm/llvm-14.0.0-rc1/rc1/llvm-project/mlir/include/mlir/ExecutionEngine/SparseTensorUtils.h:26:7:
note: previous declaration as ‘using index_t = uint64_t’
26 | using index_t = uint64_t;
| ^~~~~~~
The same issue had already occured in the past and fixed in D72619
<https://reviews.llvm.org/D72619>. More detailed explanation can also be
found there.
Tested on `amd64-pc-solaris2.11` and `sparcv9-solaris2.11`.
Differential Revision: https://reviews.llvm.org/D119323
This makes getAliasingOpResult symmetric to getAliasingOpOperand. The previous implementation was confusing for users and implemented in such a way only because there are currently no bufferizable ops that have multiple aliasing OpResults.
Differential Revision: https://reviews.llvm.org/D119259
They used to be classes with a virtual `run` function. This was inconvenient because post analysis steps are stored in BufferizationOptions. Because of this design choice, BufferizationOptions were not copyable.
Differential Revision: https://reviews.llvm.org/D119258
Reuse the higher precision F32 approximation for the F16 one (by expanding and
truncating). This is partly RFC as I'm not sure what the expectations are here
(e.g., these are only for F32 and should not be expanded, that reusing
higher-precision ones for lower precision is undesirable due to increased
compute cost and only approximations per exact type is preferred, or this is
appropriate [at least as fallback] but we need to see how to make it more
generic across all the patterns here).
Differential Revision: https://reviews.llvm.org/D118968
Implements optional attribute or type parameters, including support for such parameters in the assembly format `struct` directive. Also implements optional groups.
Depends on D117971
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D118208
For 0-D as well as 1-D vectors, both these patterns should
return a failure as there is no need to collapse the shape
of the source. Currently, only 1-D vectors were handled. This
patch handles the 0-D case as well.
Reviewed By: Benoit, ThomasRaoux
Differential Revision: https://reviews.llvm.org/D119202
The conversion to the new ControlFlow dialect didn't change the
GPUToROCDL pass - this commit fixes this issue.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D119188
Add support for computing an overapproximation of the number of integer points
in a polyhedron. The returned result is actually the number of integer points
one gets by computing the "rational shadow" obtained by projecting out the
local IDs, finding the minimal axis-parallel hyperrectangular approximation
of the shadow, and returning the number of integer points in that. This does
not currently support symbols.
Reviewed By: Groverkss
Differential Revision: https://reviews.llvm.org/D119228
These have generally been replaced by better ODS functionality, and do not
need to be explicitly provided anymore.
Differential Revision: https://reviews.llvm.org/D119065
Fix the verification function of spirv::ConstantOp to allow nesting
array attributes.
Reviewed By: antiagainst
Differential Revision: https://reviews.llvm.org/D118939
Add the class MultiAffineFunction which represents functions whose domain is an
IntegerPolyhedron and which produce an output given by a tuple of affine
expressions in the IntegerPolyhedron's ids.
Also add support for piece-wise MultiAffineFunctions, which are defined on a
union of IntegerPolyhedrons, and may have different output affine expressions
on each IntegerPolyhedron. Thus the function is affine on each individual
IntegerPolyhedron piece in the domain.
This is part of a series of patches leading up to parametric integer programming.
Depends on D118778.
Reviewed By: Groverkss
Differential Revision: https://reviews.llvm.org/D118779
* Implement `FlatAffineConstraints::getConstantBound(EQ)`.
* Inject a simpler constraint for loops that have at most 1 iteration.
* Taking into account constant EQ bounds of FlatAffineConstraint dims/symbols during canonicalization of the resulting affine map in `canonicalizeMinMaxOp`.
Differential Revision: https://reviews.llvm.org/D119153
This is both more efficient and more ergonomic to use, as inverting a
bit vector is trivial while inverting a set is annoying.
Sadly this leaks into a bunch of APIs downstream, so adapt them as well.
This would be NFC, but there is an ordering dependency in MemRefOps's
computeMemRefRankReductionMask. This is now deterministic, previously it
was dependent on SmallDenseSet's unspecified iteration order.
Differential Revision: https://reviews.llvm.org/D119076
This patch makes IntegerPolyhedron and derived classes use of getters to access
IntegerPolyhedron space information (`numIds, numDims, numSymbols`) instead of
directly accessing them.
This patch makes it easier to change the underlying implementation of the way
identifiers are stored, making it easier to extend/modify existing implementation.
Reviewed By: arjunp
Differential Revision: https://reviews.llvm.org/D118888
Benjamin Kramer