Enable constant folding of ops within the math dialect, and introduce constant folders for ceil and log2
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D117085
This change makes it possible to use a different buffer deallocation strategy. E.g., `-buffer-deallocation` can be used, which also works for allocations that are not in destination-passing style.
Differential Revision: https://reviews.llvm.org/D117096
This op is an example for how to deal with ops who's OpResult may aliasing with one of multiple OpOperands.
Differential Revision: https://reviews.llvm.org/D116868
Enable constant folding of ops within the math dialect, and introduce constant folders for ceil and log2
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D117085
Given a while loop whose condition is given by a cmp, don't recomputed the comparison (or its inverse) in the after region, instead use a constant since the original condition must be true if we branched to the after region.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D117047
Most convolution operations need explicit padding of the input to
ensure all accesses are inbounds. In such cases, having a pad
operation can be a significant overhead. One way to reduce that
overhead is to try to fuse the pad operation with the producer of its
source.
A sequence
```
linalg.generic -> linalg.pad_tensor
```
can be replaced with
```
linalg.fill -> tensor.extract_slice -> linalg.generic ->
tensor.insert_slice.
```
if the `linalg.generic` has all parallel iterator types.
Differential Revision: https://reviews.llvm.org/D116418
This patch adds corresponding memory effects to mlir llvm-dialect load/store/addressof ops, which thus enables canonicalizations of those ops (like dead code elimination) that rely on the effect interface
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D117041
Moved all TOSA decomposition patterns so that they can be optionally populated
and used by external rewrites. This avoids decomposing TOSa operations when
backends may benefit from the non-decomposed version.
Reviewed By: rsuderman, mehdi_amini
Differential Revision: https://reviews.llvm.org/D116526
Given an if of the form, simplify it by eliminating the not and swapping the regions
scf.if not(c) {
yield origTrue
} else {
yield origFalse
}
becomes
scf.if c {
yield origFalse
} else {
yield origTrue
}
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D116990
Previously, CallOps did not have any aliasing OpResult/OpOperand pairs. Therefore, CallOps were mostly ignored by the analysis and buffer copies were not inserted when necessary.
This commit introduces the following changes:
* Function bbArgs writable by default. A function can now be bufferized without inspecting its callers.
* Callers must introduce buffer copies of function arguments when necessary. If a function is external, the caller must conservatively assume that a function argument is modified by the callee after bufferization. If the function is not external, the caller inspects the callee to determine if a function argument is modified.
Differential Revision: https://reviews.llvm.org/D116457
D115722 added a DL spec to GPU modules. It happens that the DL
default interface implementation is sensitive to the name of the
DL spec attribute. This patch is fixing the name of the attribute
to be the expected one.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D116956
Given a select whose result is an i1, we can eliminate the conditional in the select completely by adding a few arithmetic operations.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D116839
This revision fixes SubviewOp, InsertSliceOp, ExtractSliceOp construction during bufferization
where not all offset/size/stride operands were properly specified.
A test that exhibited problematic behaviors related to incorrect memref casts is introduced.
Init tensor optimization is disabled in teh testing func bufferize pass.
Differential Revision: https://reviews.llvm.org/D116899
init_tensor elimination is arguably a pre-optimization that should be separated from comprehensive bufferization.
In any case it is still experimental and easily results in wrong IR with violated SSA def-use orderings.
Isolate the optimization behind a flag, separate the test cases and add a test case that would results in wrong IR.
Differential Revision: https://reviews.llvm.org/D116936
We allow the omission of a map in memref.reinterpret_cast under the assumption,
that the cast might cast to an identity layout. This change adds verification
that the static knowledge that is present in the reinterpret_cast supports
this assumption.
Differential Revision: https://reviews.llvm.org/D116601
This patch changes the syntax of omp.atomic.read to take the address of
destination, instead of having the value in a result. This will allow
using omp.atomic.read operation within an omp.atomic.capture operation
thus making its implementation less complex.
Reviewed By: peixin
Differential Revision: https://reviews.llvm.org/D116396
Fix crash in the presence of yield values. Multiple fixes to affine loop
tiling pre-condition checks and return status. Do not signal pass
failure on a failure to tile since the IR is still valid. Detect index
set computation failure in checkIfHyperrectangular and return failure.
Replace assertions with proper status return. Move checks to an
appropriate place earlier in the utility before mutation happens.
Differential Revision: https://reviews.llvm.org/D116738
With this change, the analysis takes a look at OpOperands instead of OpResults. OpOperands can bufferize out-of-place (even if they have no aliasing OpResults). The analysis does no longer care about OpResults.
Previously, only OpResults could bufferize out-of-place, so OpOperands that have no aliasing OpResults were never copied by Comprehensive Bufferize. This does not fit wwell with the new CallOp bufferization that is introduced in a subsequent change. In essence, called FuncOps can then be treated as "black boxes" that may read/write to any bbArg, even if they do not return anything.
Differential Revision: https://reviews.llvm.org/D115706
This is in preparation of fixing CallOp bufferization. Add explicit linalg.inplaceable attrs to all bbArgs, except for the ones where inplaceability should be decided by the analysis.
Differential Revision: https://reviews.llvm.org/D115840
Recent commits added a possibility for indices in LLVM dialect GEP operations
to be supplied directly as constant attributes to ensure they remain such until
translation to LLVM IR happens. Make this required for indexing into LLVM
struct types to match LLVM IR requirements, otherwise the translation would
assert on constructing such IR.
For better compatibility with MLIR-style operation construction interface,
allow GEP operations to be constructed programmatically using Values pointing
to known constant operations as struct indices.
Depends On D116758
Reviewed By: wsmoses
Differential Revision: https://reviews.llvm.org/D116759
This make LLVM dialect constants to work with `m_constant` matches. Implement
the folding hook for this operation as required by the trait. This in turn
allows LLVM::ConstantOp to properly participate in constant-folding.
Depends On D116757
Reviewed By: wsmoses
Differential Revision: https://reviews.llvm.org/D116758
In LLVM IR, the GEP indices that correspond to structures are required to be
i32 constants. MLIR models constants as just values defined by special
operations, and there is no verification that it is the case for structure
indices in GEP. Furthermore, some common transformations such as control flow
simplification may lead to the operands becoming non-constant. Make it possible
to directly supply constant values to LLVM GEPOp to guarantee they remain
constant until the translation to LLVM IR. This is not yet a requirement and
the verifier is not modified, this will be introduced separately.
Reviewed By: wsmoses
Differential Revision: https://reviews.llvm.org/D116757
This patch adds two folds. One for a repeated xor (e.g. xor(xor(x, a), a)) and one for a repeated trunc (e.g. trunc(trunc(x))).
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D116383
This patch adds omp.atomic.write lowering to LLVM IR.
Also, changed the syntax to have equal symbol instead of the comma to
make it more intuitive.
Reviewed By: kiranchandramohan, peixin
Differential Revision: https://reviews.llvm.org/D116416
`tensor.collapse_shape` op when fused with a consumer elementwise
`linalg.generic` operation results in creation of tensor.expand_shape
ops. In purely dynamic cases this can end up with a dynamic dimensions
being expanded to more than one dynamic dimension. This is disallowed
by the semantics of `tensor.expand_shape` operation. (While the
transformation is itself correct, its a gap in the specification of
`tensor.expand_shape` that is the issue). So disallow fusions which
result in such a pattern.
Differential Revision: https://reviews.llvm.org/D116703
This change simplifies BufferizableOpInterface and other functions. Overall, the API will get smaller: Functions related to custom IR traversal are deleted entirely. This will makes it easier to write BufferizableOpInterface implementations.
This is also in preparation of unifying Comprehensive Bufferize and core bufferization. While Comprehensive Bufferize could theoretically maintain its own IR traversal, there is no reason to do so, because all bufferize implementations in BufferizableOpInterface have to support partial bufferization anyway. And we can share a larger part of the code base between the two bufferizations.
Differential Revision: https://reviews.llvm.org/D116448
Until now, bufferization assumed that the yieleded tensor of a linalg.tiled_loop is an output tensor. This is not necessarily the case.
Differential Revision: https://reviews.llvm.org/D116685
Verify only the outer type being LLVM-compatible, the elemental types if
present are already checked by the type verifiers. This makes some LLVM dialect
operations compatible with mixed-dialect types that appear during progressive
lowering.
Reviewed By: wsmoses
Differential Revision: https://reviews.llvm.org/D116671
There is no need to inspect the ReturnOp of the called function.
This change also refactors the bufferization of CallOps in such a way that `lookupBuffer` is called only a single time. This is important for a later change that fixes CallOp bufferization. (There is currently a TODO among the test cases.)
Note: This change modifies a test case but is marked as NFC. There is no change of functionality, but FuncOps with empty bodies are now reported with a different error message.
Differential Revision: https://reviews.llvm.org/D116446
Such CallOps were not handled properly. When computing the new result types (and replacement values) of a CallOp, non-tensor return values were not accounted for.
Differential Revision: https://reviews.llvm.org/D116445
Depends On D115008
This change opens the way for D115012, and removes some corner cases in `CodegenUtils.cpp`. The `SparseTensorAttrDefs.td` already specifies that we allow `0` bitwidth for the two overhead types and that it is interpreted to mean the architecture's native width.
Reviewed By: aartbik
Differential Revision: https://reviews.llvm.org/D115010
Presently the result type verification checks if the type is used by a `pdl::OperationOp` inside the matcher. This is unnecessarily restrictive; the type could come from a `pdl::OperandOp or `pdl::OperandsOp` and still be inferrable.
Reviewed By: rriddle, Mogball
Differential Revision: https://reviews.llvm.org/D116083
After https://reviews.llvm.org/D115821 it became possible to create
`tensor<elem_type>` with a single `tensor.from_elements` operation without
collapsing tensor shape from `tensor<1xelem_type>` to `tensor<elem_type>`
Differential Revision: https://reviews.llvm.org/D115891
Two canonicalizations for select %x, 1, 0
If the return type is i1, return simply the condition %x, otherwise extui %x to the return type.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D116517
Replace and(ext(a),ext(b)) with ext(and(a,b)). This both reduces one instruction, and results in the computation (and/or) being done on a smaller type.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D116519
This patch creates folds for cmpi( ext(%x : i1, iN) != 0) -> %x
In essence this matches patterns matching an extension of a boolean, that != 0, which is equivalent to the original condition.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D116504
MLIR already exposes landingpads, the invokeop and the personality function on LLVM functions. With this intrinsic it should be possible to implement exception handling via the exception handling mechanisms provided by the Itanium ABI.
Differential Revision: https://reviews.llvm.org/D116436
Per the discussion in https://reviews.llvm.org/D116345 it makes sense
to move AtomicRMWOp out of the standard dialect. This was accentuated by the
need to add a fold op with a memref::cast. The only dialect
that would permit this is the memref dialect (keeping it in the standard dialect
or moving it to the arithmetic dialect would require those dialects to have a
dependency on the memref dialect, which breaks linking).
As the AtomicRMWKind enum is used throughout, this has been moved to Arith.
Reviewed By: Mogball
Differential Revision: https://reviews.llvm.org/D116392
vector.transfer operations do not have rank-reducing semantics.
Bail on illegal rank-reduction: we need to check that the rank-reduced
dims are exactly the leading dims. I.e. the following is illegal:
```
%0 = vector.transfer_write %v, %t[0,0], %cst :
vector<2x4xf32>, tensor<2x4xf32>
%1 = tensor.insert_slice %0 into %tt[0,0,0][2,1,4][1,1,1] :
tensor<2x4xf32> into tensor<2x1x4xf32>
```
Cannot fold into:
```
%0 = vector.transfer_write %v, %t[0,0,0], %cst :
vector<2x4xf32>, tensor<2x1x4xf32>
```
For this, check the trailing `vectorRank` dims of the insert_slice result
tensor match the trailing dims of the inferred result tensor.
Differential Revision: https://reviews.llvm.org/D116409
The semantics of the ops that implement the
`OffsetSizeAndStrideOpInterface` is that if the number of offsets,
sizes or strides are less than the rank of the source, then some
default values are filled along the trailing dimensions (0 for offset,
source dimension of sizes, and 1 for strides). This is confusing,
especially with rank-reducing semantics. Immediate issue here is that
the methods of `OffsetSizeAndStridesOpInterface` assumes that the
number of values is same as the source rank. This cause out-of-bounds
errors.
So simplifying the specification of `OffsetSizeAndStridesOpInterface`
to make it invalid to specify number of offsets/sizes/strides not
equal to the source rank.
Differential Revision: https://reviews.llvm.org/D115677
* There is no reason to forbid that case
* Also, user will get very unfriendly error like `expected result type with offset = -9223372036854775808 instead of 1`
Differential Revision: https://reviews.llvm.org/D114678
These conversions are better suited to be applied at whole tensor
level. Applying these as canonicalizations end up triggering such
canonicalizations at all levels of the stack which might be
undesirable. For example some of the resulting code patterns wont
bufferize in-place and need additional stack buffers. Best is to be
more deliberate in when these canonicalizations apply.
Differential Revision: https://reviews.llvm.org/D115912
TOSA's canonicalizers that change dense operations should be moved to a
seperate optimization pass to avoid canonicalizing to operations not supported
for relevant backends.
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D115890
When the input and output of a pool2d op are both 1x1, it can be canonicalized to a no-op
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D115908
Slight rename and better variable type usage in tosa.conv2d to
tosa.fully_connected lowering. Included disabling pass for padded
convolutions.
Reviewed By: not-jenni
Differential Revision: https://reviews.llvm.org/D115776
This patch extends the GPU kernel outlining pass so that it can take in
an optional data layout specification that will be attached to the GPU
module operation generated. If the data layout specification is not provided
the default data layout is used instead.
Reviewed By: herhut, mehdi_amini
Differential Revision: https://reviews.llvm.org/D115722
If all the dims are reduction dims, it is already in inner-most/outer-most
reduction form.
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D115820
* Call `replaceOp` instead of `mapBuffer`.
* Remove bvm and all helper functions around bvm.
* Simplify FuncOp bufferization and rely on existing functionality to generate ToMemrefOps for function BlockArguments.
Differential Revision: https://reviews.llvm.org/D115515
Instead of modifying the existing scf.if op, create a new op with memref OpOperands/OpResults and delete the old op.
New allocations / other memrefs can now be yielded from the op. This functionality is deactivated by default and guarded against by AssertDestinationPassingStyle.
Differential Revision: https://reviews.llvm.org/D115491
With VectorType supporting scalable dimensions, we don't need many of
the operations currently present in ArmSVE, like mask generation and
basic arithmetic instructions. Therefore, this patch also gets
rid of those.
Having built-in scalable vector support also simplifies the lowering of
scalable vector dialects down to LLVMIR.
Scalable dimensions are indicated with the scalable dimensions
between square brackets:
vector<[4]xf32>
Is a scalable vector of 4 single precission floating point elements.
More generally, a VectorType can have a set of fixed-length dimensions
followed by a set of scalable dimensions:
vector<2x[4x4]xf32>
Is a vector with 2 scalable 4x4 vectors of single precission floating
point elements.
The scale of the scalable dimensions can be obtained with the Vector
operation:
%vs = vector.vscale
This change is being discussed in the discourse RFC:
https://llvm.discourse.group/t/rfc-add-built-in-support-for-scalable-vector-types/4484
Differential Revision: https://reviews.llvm.org/D111819
Instead of modifying the existing scf.for op, create a new op with memref OpOperands/OpResults and delete the old op.
New allocations / other memrefs can now be yielded from the loop. This functionality is deactivated by default and guarded against by AssertDestinationPassingStyle.
This change also introduces `replaceOp`, which will be utilized by all other `bufferize` implementations in future commits. Bufferization will then no longer rely on old (pre-bufferize) ops to DCE away. Instead old ops are deleted on the spot. This improves debuggability because there won't be any duplicate ops anymore (bufferized + not-yet-bufferized) when dumping IR during bufferization. It is also less fragile because unbufferized IR can no longer silently "hang around" due to an implementation bug.
Differential Revision: https://reviews.llvm.org/D114926
Remove the RangeOp and the RangeType that are not actively used anymore. After removing RangeType, the LinalgTypes header only includes the generated dialect header.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D115727
Implementation of the interface allows querying the size and alignments of an LLVMArrayType as well as query the size and alignment of a struct containing an LLVMArrayType.
The implementation should yield the same results as llvm::DataLayout, including support for over aligned element types.
There is no customization point for adjusting an arrays alignment; it is simply taken from the element type.
Differential Revision: https://reviews.llvm.org/D115704
This is the second part of https://reviews.llvm.org/D114993 after slicing
into 2 independent commits.
This is needed at the moment to get good codegen from 2d vector.transfer
ops that aim to compile to SIMD load/store instructions but that can
only do so if the whole 2d transfer shape is handled in one piece, in
particular taking advantage of the memref being contiguous rowmajor.
For instance, if the target architecture has 128bit SIMD then we would
expect that contiguous row-major transfers of <4x4xi8> map to one SIMD
load/store instruction each.
The current generic lowering of multi-dimensional vector.transfer ops
can't achieve that because it peels dimensions one by one, so a transfer
of <4x4xi8> becomes 4 transfers of <4xi8>.
The new patterns here are only enabled for now by
-test-vector-transfer-flatten-patterns.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114993
This is the first part of https://reviews.llvm.org/D114993 which has been
split into small independent commits.
This is needed at the moment to get good codegen from 2d vector.transfer
ops that aim to compile to SIMD load/store instructions but that can
only do so if the whole 2d transfer shape is handled in one piece, in
particular taking advantage of the memref being contiguous rowmajor.
For instance, if the target architecture has 128bit SIMD then we would
expect that contiguous row-major transfers of <4x4xi8> map to one SIMD
load/store instruction each.
The current generic lowering of multi-dimensional vector.transfer ops
can't achieve that because it peels dimensions one by one, so a transfer
of <4x4xi8> becomes 4 transfers of <4xi8>.
The new patterns here are only enabled for now by
-test-vector-transfer-flatten-patterns.
Reviewed By: nicolasvasilache
* Generalizes passes linalg-detensorize, linalg-fold-unit-extent-dims, convert-elementwise-to-linalg.
* I feel that more work could be done in the future (i.e. make FunctionLike into a proper OpInterface and extend actions in dialect conversion to be trait based), and this patch would be a good record of why that is useful.
* Note for downstreams:
* Since these passes are now generic, they do not automatically nest with pass managers set up for implicit nesting.
* The Detensorize pass must run on a FunctionLike, and this requires explicit nesting.
* Addressed missed comments from the original and per-suggestion removed the assert on FunctionLike in ElementwiseToLinalg and DropUnitDims.cpp, which also is what was causing the integration test to fail.
This reverts commit aa8815e42e.
Differential Revision: https://reviews.llvm.org/D115671
* Generalizes passes linalg-detensorize, linalg-fold-unit-extent-dims, convert-elementwise-to-linalg.
* I feel that more work could be done in the future (i.e. make FunctionLike into a proper OpInterface and extend actions in dialect conversion to be trait based), and this patch would be a good record of why that is useful.
* Note for downstreams:
* Since these passes are now generic, they do not automatically nest with pass managers set up for that.
* If running them over nested functions, you must nest explicitly. Upstream has adopted this style but *-opt still has some uses of implicit pipelines via args. See tests for argument changes needed.
Differential Revision: https://reviews.llvm.org/D115645
Adapt the LinalgStrategyVectorizationPattern pass to apply the vectorization patterns in two stages. The change ensures the generic pad tensor op vectorization pattern does not run too early. Additionally, the revision adds the transfer op canonicalization patterns to the set of applied patterns, since they are needed to enable efficient vectorization for rank-reduced convolutions.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D115627
Use the current instead of the new source type to compute the rank-reduction map in getCanonicalSubViewResultType. Otherwise, the computation of the rank-reduction map fails when folding a cast into a subview since the strides of the new source type cannot be related to the strides of the current result type.
Depends On D115428
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D115446
Using this implementation of the interface it is possible to query the size, ABI alignment as well as the preferred alignment of a struct. It should yield the same results as LLVMs `llvm::DataLayout` on an equivalent `llvm::StructType`, including for packed structs.
Additionally it is also possible to increase the ABI and preferred alignment using a data layout entry with the type `llvm.struct<()>, which serves the same functionality as the `a:` component in LLVMs data layout string.
Differential Revision: https://reviews.llvm.org/D115600
Do not compose pad tensor operations if the extract slice of the outer pad tensor operation is rank reducing. The inner extract slice op cannot be rank-reducing since it source type must match the desired type of the padding.
Depends On D115359
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D115428
Tighten the matcher of the PadTensorOpVectorizationWithInsertSlicePattern pattern. Only match if the PadOp result is used by the InsertSliceOp source. Fail if the result is used by the InsertSliceOp dest.
Depends On D115336
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D115359
Adapt the computation of a static bounding box to take rank-reducing slice operations into account by filtering out reduced size one dimensions. The revision is needed to make padding work for decomposed convolution operations. The decomposition introduces rank reducing extract slice operations that previously let padding fail.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D115336
Flags some potential cases where splitting isn't happening and so could result
in confusing results. Also update some test files where there were near misses
in splitting that seemed unintentional.
Differential Revision: https://reviews.llvm.org/D109636
The 0-D case gets lowered in almost the same way that the 1-D case does
in VectorCreateMaskOpConversion. I also had to slightly update the
verifier for the op to always require exactly 1 operand in the 0-D case.
Depends On D115220
Reviewed by: ftynse
Differential revision: https://reviews.llvm.org/D115221
Following the example of `VectorOfAnyRankOf`, I've done a few changes in the
`.td` files to help with adding the support for the 0-D case gradually.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D115220
InsertSliceOp may have subprefix semantics where missing trailing dimensions
are automatically inferred directly from the operand shape.
This revision fixes an overflow that occurs in such cases when the impl is based on the op rank.
Differential Revision: https://reviews.llvm.org/D115549
Switch the attribute creation operations to use attr-dict-with-
keyword to avoid conflicts (in the case of pdl.attribute) and
confusion(in the case of pdl_interp.create_attribute) with
having a DictionaryAttr as a value and specifying the
attributes of the operation itself (as a dictionary).
Differential Revision: https://reviews.llvm.org/D114815
The results of a rewrite are optional, but we currently require
them to be present in the assembly format. This commit
makes the results component in the format optional.
Differential Revision: https://reviews.llvm.org/D114814
Wrong type was used for the result type in the tosa.conv_2d canonicalization.
The type should match the result element type should match the result type
not the input element type.
Differential Revision: https://reviews.llvm.org/D115463
This patterns tries to convert an inner (outer) dim reduction to an
outer (inner) dim reduction. Doing this on a 1D or 0D vector results
in an infinite loop since the converted op is same as the original
operation. Just returning failure when source rank <= 1 fixes the
issue.
Differential Revision: https://reviews.llvm.org/D115426
- Define a gpu.printf op, which can be lowered to any GPU printf() support (which is present in CUDA, HIP, and OpenCL). This op only supports constant format strings and scalar arguments
- Define the lowering of gpu.pirntf to a call to printf() (which is what is required for AMD GPUs when using OpenCL) as well as to the hostcall interface present in the AMD Open Compute device library, which is the interface present when kernels are running under HIP.
- Add a "runtime" enum that allows specifying which of the possible runtimes a ROCDL kernel will be executed under or that the runtime is unknown. This enum controls how gpu.printf is lowered
This change does not enable lowering for Nvidia GPUs, but such a lowering should be possible in principle.
And:
[MLIR][AMDGPU] Always set amdgpu-implicitarg-num-bytes=56 on kernels
This is something that Clang always sets on both OpenCL and HIP kernels, and failing to include it causes mysterious crashes with printf() support.
In addition, revert the max-flat-work-group-size to (1, 256) to avoid triggering bugs in the AMDGPU backend.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D110448
Depends On D115263
By aligning block size to inner loop iterations parallel_compute_fn LLVM can later unroll and vectorize some of the inner loops with small number of trip counts. Up to 2x speedup in multiple benchmarks.
Reviewed By: bkramer
Differential Revision: https://reviews.llvm.org/D115436
With complex recursive structure of async dispatch function LLVM can't always propagate constants to the parallel_compute_fn and it often prevents optimizations like loop unrolling and vectorization. We help LLVM by pushing known constants into the parallel_compute_fn explicitly.
Reviewed By: bkramer
Differential Revision: https://reviews.llvm.org/D115263
LinalgOp results usually bufferize inplace with output args. With this change, they may buffer inplace with input args if the value of the output arg is not used in the computation.
Differential Revision: https://reviews.llvm.org/D115022
This patch supports the atomic construct (update) following section 2.17.7 of OpenMP 5.0 standard. Also added tests and verifier for the same.
Reviewed By: kiranchandramohan, peixin
Differential Revision: https://reviews.llvm.org/D112982
The region of `linalg.generic` might contain `tensor` operations. For
example, current lowering of `gather` uses a `tensor.extract` in the
body of the `LinalgOp`. Bufferize the ops within a `LinalgOp` region
as well to catch such cases.
Differential Revision: https://reviews.llvm.org/D115322
This adds a new option `dialectFilter` to BufferizationOptions. Only ops from dialects that are allow-listed in the filter are bufferized. Other ops are left unbufferized. Note: This option requires `allowUnknownOps = true`.
To make use of `dialectFilter`, BufferizationOptions or BufferizationState must be passed to various helper functions.
The purpose of this change is to provide a better infrastructure for partial bufferization, which will be fully activated in a subsequent change.
Differential Revision: https://reviews.llvm.org/D114691
The new form of printing attribute in the declarative assembly is eliding the `#dialect.mnemonic` prefix to only keep the `<....>` part.
Differential Revision: https://reviews.llvm.org/D113873
This revision implements sparse outputs (from scratch) in all cases where
the loops can be reordered with all but one parallel loops outer. If the
inner parallel loop appears inside one or more reductions loops, then an
access pattern expansion is required (aka. workspaces in TACO speak).
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D115091
Quantized case needs to include zero-point corrections before the tosa.mul.
Disabled for the quantized use-case.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D115264
For a 1x1 weight and stride of 1, the input/weight can be reshaped and
multiplied elementwise then reshaped back
Reviewed By: rsuderman, KoolJBlack
Differential Revision: https://reviews.llvm.org/D115207
Fixed the tosa.conv2d to tosa.fully_connected canonicalization for incorrect
output channels. Included uptes to tests to include checks for the result
shapes during canonicalization.
This allows conv2d to transform to the simpler fully_connected operation.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D115170
Remove all function calls related to buffer equivalence from bufferize implementations.
Add a new PostAnalysisStep for scf.for that ensures that yielded values are equivalent to the corresponding BBArgs. (This was previously checked in `bufferize`.) This will be relaxed in a subsequent commit.
Note: This commit changes two test cases. These were broken by design
and should not have passed. With the new scf.for PostAnalysisStep, this
bug was fixed.
Differential Revision: https://reviews.llvm.org/D114927
Collect equivalent BBArgs right after the equivalence analysis of the FuncOp and before bufferizing. This is in preparation of decoupling bufferization from aliasInfo.
Also gather equivalence info for CallOps, which was missing in the
previous commit.
Differential Revision: https://reviews.llvm.org/D114847
To support creating both a mask with just a single `true` and `false` values,
I had to relax the restriction in the verifier that the rank is always equal to
the length of the attribute array, in other words, we now allow:
- `vector.constant_mask [0] : vector<i1>` which gets lowered to
`arith.constant dense<false> : vector<i1>`
- `vector.constant_mask [1] : vector<i1>` which gets lowered to
`arith.constant dense<true> : vector<i1>`
(the attribute list for the 0-D case must be a singleton containing
either `0` or `1`)
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D115023
Fix affine.for unroll for multi-result upper bound maps: these can't be
unrolled/unroll-and-jammed in cases where the trip count isn't known to
be a multiple of the unroll factor.
Fix and clean up repeated/unnecessary checks/comments at helper callees.
Also, fix clang-tidy variable naming warnings and redundant includes.
Differential Revision: https://reviews.llvm.org/D114662
This reverts commit 13bdb7ab4a. The commit introduced/uncovered an unintended bug in models containing Conv2D.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D115079
Allow ops that are not bufferizable in the input IR. (Deactivated by default.)
bufferization::ToMemrefOp and bufferization::ToTensorOp are generated at the bufferization boundaries.
Differential Revision: https://reviews.llvm.org/D114669
The implementation only allows to bit-cast between two 0-D vectors. We could
probably support casting from/to vectors like `vector<1xf32>`, but I wasn't
convinced that this would be important and it would require breaking the
invariant that `BitCastOp` works only on vectors with equal rank.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114854
This change provides `BufferizableOpInterface` implementations for ops from the Bufferization dialects. These ops are needed at the bufferization boundaries for partial bufferization.
Differential Revision: https://reviews.llvm.org/D114618
This is a lightweight operation, useful for writing unit tests. It will be utilized for testing in subsequent commits.
Differential Revision: https://reviews.llvm.org/D114693
This revision adds 0-d vector support to vector.transfer ops.
In the process, numerous cleanups are applied, in particular around normalizing
and reducing the number of builders.
Reviewed By: ThomasRaoux, springerm
Differential Revision: https://reviews.llvm.org/D114803
However, since CallOps have no aliasing OpResults, their OpOperands always bufferize out-of-place.
This change removes `bufferizesToMemoryWrite` from `CallOpInterface`. This method was called, but its return value did not matter.
Differential Revision: https://reviews.llvm.org/D114616
The new affine map generated by linearizeCollapsedDims should not drop
dimensions. We need to make sure we create a map with at least as many
dimensions as the source map. This prevents
FoldProducerReshapeOpByLinearization from generating invalid IR.
This solves regression in IREE due to e4e4da86af
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D114838
This reverts commit 9a844c2a9b.
The new affine map generated by linearizeCollapsedDims should not drop
dimensions. We need to make sure we create a map with at least as many
dimensions as the source map. This prevents
FoldProducerReshapeOpByLinearization from generating invalid IR.
This solves regression in IREE due to e4e4da86af
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D114838
Proper test for sparse tensor outputs is a single condition throughout
the whole tensor index expression (not a general conjunction, since this
may include other conditions that cause cancellation).
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D114810
This revision reintroduces tensor.insert_slice verification which seems
to have vanished over time: a verifier was initially introduced in cf9503c1b7
but for some reason the invalid.mlir was not properly updated; as time passed the verifier was not called anymore and later the code was deleted.
As a consequence, a non-negligible portion of tests has run astray using invalid
tensor.insert_slice semantics and needed to be fixed.
Also, extract isRankReducedType from TensorOps for better reuse
Originally, this facility was used by both tensor and memref forms but
it got copied around as dialects were split.
Differential Revision: https://reviews.llvm.org/D114715
The canonical type of the result of the `memref.subview` needs to make
sure that the previously dropped unit-dimensions are the ones dropped
for the canonicalized type as well. This means the generic
`inferRankReducedResultType` cannot be used. Instead the current
dropped dimensions need to be querried and the same need to be dropped.
Reviewed By: nicolasvasilache, ThomasRaoux
Differential Revision: https://reviews.llvm.org/D114751
For a 1x1 weight and stride of 1, the input/weight can be reshaped and passed into a fully connected op then reshaped back
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D114757
Add the decompose patterns that lower higher dimensional convolutions to lower dimensional ones to CodegenStrategy and use CodegenStrategy to test the decompose patterns. Additionally, remove the assertion that checks the anchor op name is set in the CodegenStrategyTest pass. Removing the assertion allows us to simplify the pipelines used in the interchange and decompose tests.
Depends On D114797
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114798
Add support for an empty anchor op string in vectorization. An empty anchor op string is useful after fusion when there are multiple different operations to vectorize.
Depends On D114689
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114690
Pad the operation using a top down traversal. The top down traversal unlocks folding opportunities and dim op canonicalizations due to the introduced extract slice operation after the padded operation.
Depends On D114585
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114689
Add CSE after every transformation. Transformations such as tiling introduce redundant computation, for example, one AffineMinOp for every operand dimension pair. Follow up transformations such as Padding and Hoisting benefit from CSE since comparing slice sizes simplifies to comparing SSA values instead of analyzing affine expressions.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114585
Moves sparse tensor output support forward by generalizing from injective
insertions only to include reductions. This revision accepts the case with all
parallel outer and all reduction inner loops, since that can be handled with
an injective insertion still. Next revision will allow the inner parallel loop
to move inward (but that will require "access pattern expansion" aka "workspace").
Reviewed By: bixia
Differential Revision: https://reviews.llvm.org/D114399
The verifier computed an illegal type with negative dimension size when collapsing partially static memrefs.
Differential Revision: https://reviews.llvm.org/D114702
This is commit 4 of 4 for the multi-root matching in PDL, discussed in https://llvm.discourse.group/t/rfc-multi-root-pdl-patterns-for-kernel-matching/4148 (topic flagged for review).
This PR integrates the various components (root ordering algorithm, nondeterministic execution of PDL bytecode) to implement multi-root PDL matching. The main idea is for the pattern to specify mulitple candidate roots. The PDL-to-PDLInterp lowering selects one of these roots and "hangs" the pattern from this root, traversing the edges downwards (from operation to its operands) when possible and upwards (from values to its uses) when needed. The root is selected by invoking the optimal matching multiple times, once for each candidate root, and the connectors are determined form the optimal matching. The costs in the directed graph are equal to the number of upward edges that need to be traversed when connecting the given two candidate roots. It can be shown that, for this choice of the cost function, "hanging" the pattern an inner node is no better than from the optimal root.
The following three main additions were implemented as a part of this PR:
1. OperationPos predicate has been extended to allow tracing the operation accepting a value (the opposite of operation defining a value).
2. Predicate checking if two values are not equal - this is useful to ensure that we do not traverse the edge back downwards after we traversed it upwards.
3. Function for for building the cost graph among the candidate roots.
4. Updated buildPredicateList, building the predicates optimal branching has been determined.
Testing: unit tests (an integration test to follow once the stack of commits has landed)
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D108550
This is commit 1 of 4 for the multi-root matching in PDL, discussed in https://llvm.discourse.group/t/rfc-multi-root-pdl-patterns-for-kernel-matching/4148 (topic flagged for review).
These operations are:
* pdl.get_accepting_ops: Returns a list of operations accepting the given value or a range of values at the specified position. Thus if there are two operations `%op1 = "foo"(%val)` and `%op2 = "bar"(%val)` accepting a value at position 0, `%ops = pdl_interp.get_accepting_ops of %val : !pdl.value at 0` will return both of them. This allows us to traverse upwards from a value to operations accepting the value.
* pdl.choose_op: Iteratively chooses one operation from a range of operations. Therefore, writing `%op = pdl_interp.choose_op from %ops` in the example above will select either `%op1`or `%op2`.
Testing: Added the corresponding test cases to mlir/test/Dialect/PDLInterp/ops.mlir.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D108543
Use primarily matvec instead of matmul to test hoist padding. Test the hoisting only starting from already padded IR. Use one-dimensional tiling only except for the tile_and_fuse test that exercises hoisting on a larger loop nest with fill and pad tensor operations in the backward slice.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114608
If `allowReturnMemref` is set to true, arbitrary memrefs may be returned from FuncOps. Also remove allocation hoisting code, which is only partly implemented at the moment.
The purpose of this commit is to untangle `bufferize` from `aliasInfo`. (Even with this change, they are not fully untangled yet.)
Differential Revision: https://reviews.llvm.org/D114507
Rename the check prefixes to HOIST21 and HOIST32 to clarify the different flag configurations.
Depends On D114438
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114442
Instead of checking for unexpected operations (any operation with a region except for scf::For and `padTensorOp` or operations with a memory effect) while cloning the packing loop nest perform the checks early. Update `dropNonIndexDependencies` to check for unexpected operations. Additionally, check all of these operations have index type operands only.
Depends On D114428
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114438
Limit hoist padding to pad tensor ops that depend only on a constant value. Supporting arbitrary padding values that depend on computations part of the backward slice to hoist require complex analysis to ensure the computation can be hoisted.
Depends On D114420
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114428
Adapt hoist padding to filter the backward slice before cloning the packing loop nest. The filtering removes all operations that are not used to index the hoisted pad tensor op and its extract slice op. The filtering is needed to support the more complex loop nests created after fusion. For example, fusing the producer of an output operand can added linalg ops and pad tensor ops to the backward slice. These operations have regions and currently prevent hoisting.
The following example demonstrates the effect of the newly introduced `dropNonIndexDependencies` method that filters the backward slice:
```
%source = linalg.fill(%cst, %arg0)
scf.for %i
%unrelated = linalg.fill(%cst, %arg1) // not used to index %source!
scf.for %j (%arg2 = %unrelated)
scf.for %k // not used to index %source!
%ubi = affine.min #map(%i)
%ubj = affine.min #map(%j)
%slice = tensor.extract_slice %source [%i, %j] [%ubi, %ubj]
%padded_slice = linalg.pad_tensor %slice
```
dropNonIndexDependencies(%padded_slice, %slice)
removes [scf.for %k, linalg.fill(%cst, %arg1)] from backwardSlice.
Depends On D114175
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114420
* 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/D114138
The padding tests previously contained the tile loops. This revision removes the tile loops since padding itself does not consider the loops. Instead the induction variables are passed in as function arguments which promotes them to symbols in the affine expressions. Note that the pad-and-hoist.mlir test still exercises padding in the context of the full loop nest.
Depends On D114175
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D114227
William S. Moses