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 3 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).
We form a graph over the specified roots, provided in `pdl.rewrite`, where two roots are connected by a directed edge if the target root can be connected (via a chain of operations) in the underlying pattern to the source root. We place a restriction that the path connecting the two candidate roots must only contain the nodes in the subgraphs underneath these two roots. The cost of an edge is the smallest number of upward traversals (edges) required to go from the source to the target root, and the connector is a `Value` in the intersection of the two subtrees rooted at the source and target root that results in that smallest number of such upward traversals. Optimal root ordering is then formulated as the problem of finding a spanning arborescence (i.e., a directed spanning tree) of minimal weight.
In order to determine the spanning arborescence (directed spanning tree) of minimum weight, we use the [Edmonds' algorithm](https://en.wikipedia.org/wiki/Edmonds%27_algorithm). The worst-case computational complexity of this algorithm is O(_N_^3) for a single root, where _N_ is the number of specified roots. The `pdl`-to-`pdl_interp` lowering calls this algorithm as a subroutine _N_ times (once for each candidate root), so the overall complexity of root ordering is O(_N_^4). If needed, this complexity could be reduced to O(_N_^3) with a more efficient algorithm. However, note that the underlying implementation is very efficient, and _N_ in our instances tends to be very small (<10). Therefore, we believe that the proposed (asymptotically suboptimal) implementation will suffice for now.
Testing: a unit test of the algorithm
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D108549
Transpose convolution decomposition is now performed in a separate pass. This
allows padding / constant propagation to be performed at the TOSA level. It
also adds support for striding when there is no dilation.
Differential Revision: https://reviews.llvm.org/D114409
This revision makes concrete use of 0-d vectors to extend the semantics of
InsertElementOp.
Reviewed By: dcaballe, pifon2a
Differential Revision: https://reviews.llvm.org/D114388
This revision starts making concrete use of 0-d vectors to extend the semantics of
ExtractElementOp.
In the process a new VectorOfAnyRank Tablegen OpBase.td is added to allow progressive transition to supporting 0-d vectors by gradually opting in.
Differential Revision: https://reviews.llvm.org/D114387
`vector::InsertElementOp` and `vector::ExtractElementOp` have had their `position`
operand changed to accept `AnySignlessIntegerOrIndex` for better operability with
operations that use `index`, such as affine loops.
LLVM's `extractelement` and `insertelement` can also accept `i64`, so lowering
directly to these operations without explicitly inserting casts is allowed. SPIRV's
equivalent ops can also accept `i64`.
Reviewed By: nicolasvasilache, jpienaar
Differential Revision: https://reviews.llvm.org/D114139
NamedAttribute is currently represented as an std::pair, but this
creates an extremely clunky .first/.second API. This commit
converts it to a class, with better accessors (getName/getValue)
and also opens the door for more convenient API in the future.
Differential Revision: https://reviews.llvm.org/D113956
* It works similar to scf.for coversion, but convert condition and yield ops as part of scf.whille pattern so it don't need to maintain external state
Differential Revision: https://reviews.llvm.org/D113007
Split tosa.reshape into three individual lowerings: collapse, expand and a
combination of both. Add simple dynamic shape support.
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D113936
FMAOp -> LLVM conversion is done progressively by peeling off 1 dimension from FMAOp at each pattern iteration. Add the recursively bounded property declaration to the pattern so that the rewriter can apply it multiple times.
Without this, FMAOps with 3+D do not lower to LLVM.
Differential Revision: https://reviews.llvm.org/D113886
Names should be consistent across all operations otherwise painful bugs will surface.
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D113762
Support load with broadcast, elementwise divf op and remove the
hardcoded restriction on the vector size. Picking the right size should
be enfored by user and will fail conversion to llvm/spirv if it is not
supported.
Differential Revision: https://reviews.llvm.org/D113618
New TOSA pad operation can support explicitly specifying the pad value. Added
lowering to linalg that uses the explicit value.
Differential Revision: https://reviews.llvm.org/D113515
Use existing helper instead of handling only a subset of indices lowering
arithmetic. Also relax the restriction on the memref rank for the GPU mma ops
as we can now support any rank.
Differential Revision: https://reviews.llvm.org/D113383
Given that LLVM dialect types may now optionally contain types from other
dialects, which itself is motivated by dialect interoperability and progressive
lowering, the conversion should no longer assume that the outermost LLVM
dialect type can be left as is. Instead, it should inspect the types it
contains and attempt to convert them to the LLVM dialect. Introduce this
capability for LLVM array, pointer and structure types. Only literal structures
are currently supported as handling identified structures requires the
converison infrastructure to have a mechanism for avoiding infite recursion in
case of recursive types.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D112550
This predates the templated variant, and has been simply forwarding
to getSplatValue<Attribute> for some time. Removing this makes the
API a bit more uniform, and also helps prevent users from thinking
it is "cheap".
There are several aspects of the API that either aren't easy to use, or are
deceptively easy to do the wrong thing. The main change of this commit
is to remove all of the `getValue<T>`/`getFlatValue<T>` from ElementsAttr
and instead provide operator[] methods on the ranges returned by
`getValues<T>`. This provides a much more convenient API for the value
ranges. It also removes the easy-to-be-inefficient nature of
getValue/getFlatValue, which under the hood would construct a new range for
the type `T`. Constructing a range is not necessarily cheap in all cases, and
could lead to very poor performance if used within a loop; i.e. if you were to
naively write something like:
```
DenseElementsAttr attr = ...;
for (int i = 0; i < size; ++i) {
// We are internally rebuilding the APFloat value range on each iteration!!
APFloat it = attr.getFlatValue<APFloat>(i);
}
```
Differential Revision: https://reviews.llvm.org/D113229
In order to support fusion with mma matrix type we need to be able to
execute elementwise operations on them. This add an op to be able to
support some basic elementwise operations. This is a is not a full
solution as it only supports a limited scope or operations. Ideally we would
want to be able to fuse with more kind of operations.
Differential Revision: https://reviews.llvm.org/D112857
wmma intrinsics have a large number of combinations, ideally we want to be able
to target all the different variants. To avoid a combinatorial explosion in the
number of mlir op we use attributes to represent the different variation of
load/store/mma ops. We also can generate with tablegen helpers to know which
combinations are available. Using this we can avoid having too hardcode a path
for specific shapes and can support more types.
This patch also adds boiler plates for tf32 op support.
Differential Revision: https://reviews.llvm.org/D112689
Add the shufflevector conversion. It only handles the static, i.e., IntegerAttr, index.
Co-authored: Xinyi Liu <xyliuhelen@gmail.com>
Reviewed by: antiagainst
Differential revision: https://reviews.llvm.org/D112161
Allow lowering of wmma ops with 64bits indexes. Change the default
version of the test to use default layout.
Differential Revision: https://reviews.llvm.org/D112479
The current implementation invokes materializations
whenever an input operand does not have a mapping for the
desired type, i.e. it requires materialization at the earliest possible
point. This conflicts with goal of dialect conversion (and also the
current documentation) which states that a materialization is only
required if the materialization is supposed to persist after the
conversion process has finished.
This revision refactors this such that whenever a target
materialization "might" be necessary, we insert an
unrealized_conversion_cast to act as a temporary materialization.
This allows for deferring the invocation of the user
materialization hooks until the end of the conversion process,
where we actually have a better sense if it's actually
necessary. This has several benefits:
* In some cases a target materialization hook is no longer
necessary
When performing a full conversion, there are some situations
where a temporary materialization is necessary. Moving forward,
these users won't need to provide any target materializations,
as the temporary materializations do not require the user to
provide materialization hooks.
* getRemappedValue can now handle values that haven't been
converted yet
Before this commit, it wasn't well supported to get the remapped
value of a value that hadn't been converted yet (making it
difficult/impossible to convert multiple operations in many
situations). This commit updates getRemappedValue to properly
handle this case by inserting temporary materializations when
necessary.
Another code-health related benefit is that with this change we
can move a majority of the complexity related to materializations
to the end of the conversion process, instead of handling adhoc
while conversion is happening.
Differential Revision: https://reviews.llvm.org/D111620
Specification specified the output type for quantized average pool should be
an i32. Only accumulator should be an i32, result type should match the input
type.
Caused in https://reviews.llvm.org/D111590
Reviewed By: sjarus, GMNGeoffrey
Differential Revision: https://reviews.llvm.org/D112484
The change is based on the proposal from the following discussion:
https://llvm.discourse.group/t/rfc-memreftype-affine-maps-list-vs-single-item/3968
* Introduce `MemRefLayoutAttr` interface to get `AffineMap` from an `Attribute`
(`AffineMapAttr` implements this interface).
* Store layout as a single generic `MemRefLayoutAttr`.
This change removes the affine map composition feature and related API.
Actually, while the `MemRefType` itself supported it, almost none of the upstream
can work with more than 1 affine map in `MemRefType`.
The introduced `MemRefLayoutAttr` allows to re-implement this feature
in a more stable way - via separate attribute class.
Also the interface allows to use different layout representations rather than affine maps.
For example, the described "stride + offset" form, which is currently supported in ASM parser only,
can now be expressed as separate attribute.
Reviewed By: ftynse, bondhugula
Differential Revision: https://reviews.llvm.org/D111553
This is the only lowering to Linalg Tosa has, so it's needlessly
verbose. Likely this was a carry over from IREE's usage where we
originally lowered to linalg on buffers (the only linalg that existed at
the time), so the everything on tensors needed the suffix. We're dropping
it in IREE also, having transitioned entirely to using Linalg on
tensors.
Reviewed By: sjarus
Differential Revision: https://reviews.llvm.org/D111911
Part of the arith update broke UiToFp32. Fixed the lowering and included a new
test to detect a regression.
Differential Revision: https://reviews.llvm.org/D111772
Precursor: https://reviews.llvm.org/D110200
Removed redundant ops from the standard dialect that were moved to the
`arith` or `math` dialects.
Renamed all instances of operations in the codebase and in tests.
Reviewed By: rriddle, jpienaar
Differential Revision: https://reviews.llvm.org/D110797
1. To avoid two ExecutionModeOp using the same name, adding the value of execution mode in name when converting to LLVM dialect.
2. To avoid syntax error in spv.OpLoad, add OpTypeSampledImage into SPV_Type.
Reviewed by:antiagainst
Differential revision:https://reviews.llvm.org/D111193
Average pool assumed the same input/output type. Result type for integers
is always an i32, should be updated appropriately.
Reviewed By: GMNGeoffrey
Differential Revision: https://reviews.llvm.org/D111590
It was bundling quite a lot of patterns that convert high-D
vector ops into low-D elementary ops. It might not be good
for all of the patterns to happen for a particular downstream
user. For example, `ShapeCastOpRewritePattern` rewrites
`vector.shape_cast` into data movement extract/insert ops.
Instead, split the entry point into multiple ones so users
can pull in patterns on demand.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D111225
This patch extends Linalg core vectorization with support for min/max reductions
in linalg.generic ops. It enables the reduction detection for min/max combiner ops.
It also renames MIN/MAX combining kinds to MINS/MAXS to make the sign explicit for
floating point and signed integer types. MINU/MAXU should be introduce din the future
for unsigned integer types.
Reviewed By: pifon2a, ThomasRaoux
Differential Revision: https://reviews.llvm.org/D110854
The discussion in https://reviews.llvm.org/D110425 demonstrated that "packing"
may be a confusing term to define the behavior of this op in presence of the
attribute. Instead, indicate the intended effect of preventing the folder from
being applied.
Reviewed By: nicolasvasilache, silvas
Differential Revision: https://reviews.llvm.org/D111046
This patch is mainly to propogate location attribute from spv.GlobalVariable to llvm.mlir.global.
It also contains three small changes.
1. Remove the restriction on UniformConstant In SPIRVToLLVM.cpp;
2. Remove the errorCheck on relaxedPrecision when deserializering SPIR-V in Deserializer.cpp
3. In SPIRVOps.cpp, let ConstantOp take signedInteger too.
Co-authered: Alan Liu <alanliu.yf@gmail.com> and Xinyi Liu <xyliuhelen@gmail.com>
Reviewed by:antiagainst
Differential revision: https://reviews.llvm.org/D110207
Add support for dynamic shared memory for GPU launch ops: add an
optional operand to gpu.launch and gpu.launch_func ops to specify the
amount of "dynamic" shared memory to use. Update lowerings to connect
this operand to the GPU runtime.
Differential Revision: https://reviews.llvm.org/D110800
This patch introduces a generic reduction detection utility that works
across different dialecs. It is mostly a generalization of the reduction
detection algorithm in Affine. The reduction detection logic in Affine,
Linalg and SCFToOpenMP have been replaced with this new generic utility.
The utility takes some basic components of the potential reduction and
returns: 1) the reduced value, and 2) a list with the combiner operations.
The logic to match reductions involving multiple combiner operations disabled
until we can properly test it.
Reviewed By: ftynse, bondhugula, nicolasvasilache, pifon2a
Differential Revision: https://reviews.llvm.org/D110303
This commits updates the remaining usages of the ArrayRef<Value> based
matchAndRewrite/rewrite methods in favor of the new OpAdaptor
overload.
Differential Revision: https://reviews.llvm.org/D110360
This has been a TODO for a long time, and it brings about many advantages (namely nice accessors, and less fragile code). The existing overloads that accept ArrayRef are now treated as deprecated and will be removed in a followup (after a small grace period). Most of the upstream MLIR usages have been fixed by this commit, the rest will be handled in a followup.
Differential Revision: https://reviews.llvm.org/D110293
Initially, the padding transformation and the related operation were only used
to guarantee static shapes of subtensors in tiled operations. The
transformation would not insert the padding operation if the shapes were
already static, and the overall code generation would actively remove such
"noop" pads. However, this transformation can be also used to pack data into
smaller tensors and marshall them into faster memory, regardless of the size
mismatches. In context of expert-driven transformation, we should assume that,
if padding is requested, a potentially padded tensor must be always created.
Update the transformation accordingly. To do this, introduce an optional
`packing` attribute to the `pad_tensor` op that serves as an indication that
the padding is an intentional choice (as opposed to side effect of type
normalization) and should be left alone by cleanups.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D110425
Currently DenseElementsAttr only exposes the ability to get the full range of values for a given type T, but there are many situations where we just want the beginning/end iterator. This revision adds proper value_begin/value_end methods for all of the supported T types, and also cleans up a bit of the interface.
Differential Revision: https://reviews.llvm.org/D104173
For `memref.subview` operations, when there are more than one
unit-dimensions, the strides need to be used to figure out which of
the unit-dims are actually dropped.
Differential Revision: https://reviews.llvm.org/D109418
The discussion on forum:
https://llvm.discourse.group/t/bug-in-partial-dialect-conversion/4115
The `applyPartialConversion` didn't handle the operations, that were
marked as illegal inside dynamic legality callback.
Instead of reporting error, if such operation was not converted to legal set,
the method just added it to `unconvertedSet` in the same way as unknown operations.
This patch fixes that and handle dynamically illegal operations as well.
The patch includes 2 fixes for existing passes:
* `tensor-bufferize` - explicitly mark `std.return` as legal.
* `convert-parallel-loops-to-gpu` - ugly fix with marking visited operations
to avoid recursive legality checks.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D108505
The conversion pattern is particularly useful for conversion of
block arguments in the master op.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D109610
This renames the primary methods for creating a zero value to `getZero`
instead of `getNullValue` and renames predicates like `isAllOnesValue`
to simply `isAllOnes`. This achieves two things:
1) This starts standardizing predicates across the LLVM codebase,
following (in this case) ConstantInt. The word "Value" doesn't
convey anything of merit, and is missing in some of the other things.
2) Calling an integer "null" doesn't make any sense. The original sin
here is mine and I've regretted it for years. This moves us to calling
it "zero" instead, which is correct!
APInt is widely used and I don't think anyone is keen to take massive source
breakage on anything so core, at least not all in one go. As such, this
doesn't actually delete any entrypoints, it "soft deprecates" them with a
comment.
Included in this patch are changes to a bunch of the codebase, but there are
more. We should normalize SelectionDAG and other APIs as well, which would
make the API change more mechanical.
Differential Revision: https://reviews.llvm.org/D109483
Conversion to the LLVM dialect is being refactored to be more progressive and
is now performed as a series of independent passes converting different
dialects. These passes may produce `unrealized_conversion_cast` operations that
represent pending conversions between built-in and LLVM dialect types.
Historically, a more monolithic Standard-to-LLVM conversion pass did not need
these casts as all operations were converted in one shot. Previous refactorings
have led to the requirement of running the Standard-to-LLVM conversion pass to
clean up `unrealized_conversion_cast`s even though the IR had no standard
operations in it. The pass must have been also run the last among all to-LLVM
passes, in contradiction with the partial conversion logic. Additionally, the
way it was set up could produce invalid operations by removing casts between
LLVM and built-in types even when the consumer did not accept the uncasted
type, or could lead to cryptic conversion errors (recursive application of the
rewrite pattern on `unrealized_conversion_cast` as a means to indicate failure
to eliminate casts).
In fact, the need to eliminate A->B->A `unrealized_conversion_cast`s is not
specific to to-LLVM conversions and can be factored out into a separate type
reconciliation pass, which is achieved in this commit. While the cast operation
itself has a folder pattern, it is insufficient in most conversion passes as
the folder only applies to the second cast. Without complex legality setup in
the conversion target, the conversion infra will either consider the cast
operations valid and not fold them (a separate canonicalization would be
necessary to trigger the folding), or consider the first cast invalid upon
generation and stop with error. The pattern provided by the reconciliation pass
applies to the first cast operation instead. Furthermore, having a separate
pass makes it clear when `unrealized_conversion_cast`s could not have been
eliminated since it is the only reason why this pass can fail.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D109507
OpenMP reductions need a neutral element, so we match some known reduction
kinds (integer add/mul/or/and/xor, float add/mul, integer and float min/max) to
define the neutral element and the atomic version when possible to express
using atomicrmw (everything except float mul). The SCF-to-OpenMP pass becomes a
module pass because it now needs to introduce new symbols for reduction
declarations in the module.
Reviewed By: chelini
Differential Revision: https://reviews.llvm.org/D107549
The lowering has been incorrectly using the operands of the original op instead
of rewritten operands provided to matchAndRewrite call. This may lead to
spurious materializations and generally invalid IR.
Reviewed By: aartbik
Differential Revision: https://reviews.llvm.org/D109355
The conversion has been incorrectly using the operands of the original
operation instead of the converted operands provided to the matchAndRewrite
call. This may lead to spurious materializations and generally invalid IR if
the producer of the original operands is deleted in the process of conversion.
Reviewed By: csigg
Differential Revision: https://reviews.llvm.org/D109356
Create a gpu memset op and corresponding CUDA and ROCm wrappers.
Reviewed By: herhut, lorenrose1013
Differential Revision: https://reviews.llvm.org/D107548
FuncOp always lowers to an LLVM external linkage presently. This makes it impossible to define functions in mlir which are local to the current module. Until MLIR FuncOps have a more formal linkage specification, this commit allows funcop's to have an optionally specified llvm.linkage attribute, whose value will be used as the linkage of the llvm funcop when lowered.
Differential Revision: https://reviews.llvm.org/D108524
Support LLVM linkage
The StringAttr version doesn't need a context, so we can just use the
existing `SymbolRefAttr::get` form. The StringRef version isn't preferred
so we want to encourage people to use StringAttr.
There is an additional form of getSymbolRefAttr that takes a (SymbolTrait
implementing) operation. This should also be moved, but I'll do that as
a separate patch.
Differential Revision: https://reviews.llvm.org/D108922
SymbolRefAttr is fundamentally a base string plus a sequence
of nested references. Instead of storing the string data as
a copies StringRef, store it as an already-uniqued StringAttr.
This makes a lot of things simpler and more efficient because:
1) references to the symbol are already stored as StringAttr's:
there is no need to copy the string data into MLIRContext
multiple times.
2) This allows pointer comparisons instead of string
comparisons (or redundant uniquing) within SymbolTable.cpp.
3) This allows SymbolTable to hold a DenseMap instead of a
StringMap (which again copies the string data and slows
lookup).
This is a moderately invasive patch, so I kept a lot of
compatibility APIs around. It would be nice to explore changing
getName() to return a StringAttr for example (right now you have
to use getNameAttr()), and eliminate things like the StringRef
version of getSymbol.
Differential Revision: https://reviews.llvm.org/D108899
Needed to switch to extract to support tosa.reverse using dynamic shapes.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D108744
Includes the quantized version of average pool lowering to linalg dialect.
This includes a lit test for the transform. It is not 100% correct as the
multiplier / shift should be done in i64 however this is negligable rounding
difference.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D108676
Lowering to table was incorrect as it did not apply a 128 offset before
extracting the value from the table. Fixed and correct tensor length on input
table.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D108436
When padding quantized operations, the padding needs to equal the zero point
of the input value. Corrected the pass to change the padding value if quantized.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D108440
This revision adds native ODS support for VariadicOfVariadic operand
groups. An example of this is the SwitchOp, which has a variadic number
of nested operand ranges for each of the case statements, where the
number of case statements is variadic. Builtin ODS support allows for
generating proper accessors for the nested operand ranges, builder
support, and declarative format support. VariadicOfVariadic operands
are supported by providing a segment attribute to use to store the
operand groups, mapping similarly to the AttrSizedOperand trait
(but with a user defined attribute name).
`build` methods for VariadicOfVariadic operand expect inputs of the
form `ArrayRef<ValueRange>`. Accessors for the variadic ranges
return a new `OperandRangeRange` type, which represents a
contiguous range of `OperandRange`. In the declarative assembly
format, VariadicOfVariadic operands and types are by default
formatted as a comma delimited list of value lists:
`(<value>, <value>), (), (<value>)`.
Differential Revision: https://reviews.llvm.org/D107774
Presently, the lowering of nested scf.parallel loops to OpenMP creates one omp.parallel region, with two (nested) OpenMP worksharing loops on the inside. When lowered to LLVM and executed, this results in incorrect results. The reason for this is as follows:
An OpenMP parallel region results in the code being run with whatever number of threads available to OpenMP. Within a parallel region a worksharing loop divides up the total number of requested iterations by the available number of threads, and distributes accordingly. For a single ws loop in a parallel region, this works as intended.
Now consider nested ws loops as follows:
omp.parallel {
A: omp.ws %i = 0...10 {
B: omp.ws %j = 0...10 {
code(%i, %j)
}
}
}
Suppose we ran this on two threads. The first workshare loop would decide to execute iterations 0, 1, 2, 3, 4 on thread 0, and iterations 5, 6, 7, 8, 9 on thread 1. The second workshare loop would decide the same for its iteration. This means thread 0 would execute i \in [0, 5) and j \in [0, 5). Thread 1 would execute i \in [5, 10) and j \in [5, 10). This means that iterations i in [5, 10), j in [0, 5) and i in [0, 5), j in [5, 10) never get executed, which is clearly wrong.
This permits two options for a remedy:
1) Change the semantics of the omp.wsloop to be distinct from that of the OpenMP runtime call or equivalently #pragma omp for. This could then allow some lowering transformation to remedy the aforementioned issue. I don't think this is desirable for an abstraction standpoint.
2) When lowering an scf.parallel always surround the wsloop with a new parallel region (thereby causing the innermost wsloop to use the number of threads available only to it).
This PR implements the latter change.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D108426
Tosa rescale can contain uint8 types. Added support for these types
using an unrealized conversion cast. Optimistically it would be better to
use bitcast however it does not support unsigned integers.
Differential Revision: https://reviews.llvm.org/D108427
Its possible for the clamp to have invalid min/max values on its range. To fix
this we validate the range of the min/max and clamp to a valid range.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D108256
LLVM considers global variables marked as externals to be defined within the module if it is initialized (including to an undef). Other external globals are considered as being defined externally and imported into the current translation unit. Lowering of MLIR Global Ops does not properly propagate undefined initializers, resulting in a global which is expected to be defined within the current TU, not being defined.
Differential Revision: https://reviews.llvm.org/D108252
Existing linalg.conv2d is not well optimized for performance. Changed to a
version that is more aligned for optimziation. Include the corresponding
transposes to use this optimized version.
This also splits the conv and depthwise conv into separate implementations
to avoid overly complex lowerings.
Reviewed By: antiagainst
Differential Revision: https://reviews.llvm.org/D107504
The conversion is a straightforward one-to-one mapping with optional unrolling
for nD vectors, similarly to other cast operations.
Depends On D107889
Reviewed By: cota, akuegel
Differential Revision: https://reviews.llvm.org/D107891
Dilation only requires increasing the padding on the left/right side of the
input, and including dilation in the convolution. This implementation still
lacks support for strided convolutions.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D107680
These ops were not ported to the nD vector conversion when it was introduced
and nobody needed them so far.
Reviewed By: gysit
Differential Revision: https://reviews.llvm.org/D107750
If the source value to load is bool, and we have native storage
capability support for the source bitwidth, we still cannot directly
rewrite uses; we need to perform casting to bool first.
Reviewed By: hanchung
Differential Revision: https://reviews.llvm.org/D107119
If the source value to store is bool, and we have native storage
capability support for the target bitwidth, we still cannot directly
store; we need to perform casting to match the target memref
element's bitwidth.
Reviewed By: hanchung
Differential Revision: https://reviews.llvm.org/D107114
Make broadcastable needs the output shape to determine whether the operation
includes additional broadcasting. Include some canonicalizations for TOSA
to remove unneeded reshape.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D106846
The verifier of the llvm.call operation was not checking for mismatches between
the number of operation results and the number of results in the signature of
the callee. Furthermore, it was possible to construct an llvm.call operation
producing an SSA value of !llvm.void type, which should not exist. Add the
verification and treat !llvm.void result type as absence of call results.
Update the GPU conversions to LLVM that were mistakenly assuming that it was
fine for llvm.call to produce values of !llvm.void type and ensure these calls
do not produce results.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D106937
- Fixed symbol insertion into `symNameToModuleMap`. Insertion
needs to happen whether symbols are renamed or not.
- Added check for the VCE triple and avoid dropping it.
- Disabled function deduplication. It requires more careful
rules. Right now it can remove different functions.
- Added tests for symbol rename listener.
- And some other code/comment cleanups.
Reviewed By: ergawy
Differential Revision: https://reviews.llvm.org/D106886
Includes a version of a quantized conv2D operations with a lowering from TOSA
to linalg with corresponding test. We keep the quantized and quantized variants
as separate named ops to avoid the additional operations for non-quantized
convolutions.
Differential Revision: https://reviews.llvm.org/D106407
Type conversion and argument materialization are context-free: there is no available information on which op / branch is currently being converted.
As a consequence, bare ptr convention cannot be handled as an argument materialization: it would apply irrespectively of the parent op.
This doesn't typecheck in the case of non-funcOp and we would see cases where a memref descriptor would be inserted in place of the pointer in another memref descriptor.
For now the proper behavior is to revert to a specific BarePtrFunc implementation and drop the blanket argument materialization logic.
This reverts the relevant piece of the conversion to LLVM to what it was before https://reviews.llvm.org/D105880 and adds a relevant test and documentation to avoid the mistake by whomever attempts this again in the future.
Reviewed By: arpith-jacob
Differential Revision: https://reviews.llvm.org/D106495
The unstrided transposed conv can be represented as a regular convolution.
Lower to this variant to handle the basic case. This includes transitioning from
the TC defined convolution operation and a yaml defined one.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D106389
Added the named op variants for quantized matmul and quantized batch matmul
with the necessary lowerings/tests from tosa's matmul/fully connected ops.
Current version does not use the contraction op interface as its verifiers
are not compatible with scalar operations.
Differential Revision: https://reviews.llvm.org/D105063
This deletes all the pooling ops in LinalgNamedStructuredOpsSpec.tc. All the
uses are replaced with the yaml pooling ops.
Reviewed By: gysit, rsuderman
Differential Revision: https://reviews.llvm.org/D106181
This simplifies the vector to LLVM lowering. Previously, both vector.load/store and vector.transfer_read/write lowered directly to LLVM. With this commit, there is a single path to LLVM vector load/store instructions and vector.transfer_read/write ops must first be lowered to vector.load/store ops.
* Remove vector.transfer_read/write to LLVM lowering.
* Allow non-unit memref strides on all but the most minor dimension for vector.load/store ops.
* Add maxTransferRank option to populateVectorTransferLoweringPatterns.
* vector.transfer_reads with changing element type can no longer be lowered to LLVM. (This functionality is needed only for SPIRV.)
Differential Revision: https://reviews.llvm.org/D106118
The dialect-specific cast between builtin (ex-standard) types and LLVM
dialect types was introduced long time before built-in support for
unrealized_conversion_cast. It has a similar purpose, but is restricted
to compatible builtin and LLVM dialect types, which may hamper
progressive lowering and composition with types from other dialects.
Replace llvm.mlir.cast with unrealized_conversion_cast, and drop the
operation that became unnecessary.
Also make unrealized_conversion_cast legal by default in
LLVMConversionTarget as the majority of convesions using it are partial
conversions that actually want the casts to persist in the IR. The
standard-to-llvm conversion, which is still expected to run last, cleans
up the remaining casts standard-to-llvm conversion, which is still
expected to run last, cleans up the remaining casts
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D105880
After the Math has been split out of the Standard dialect, the
conversion to the LLVM dialect remained as a huge monolithic pass.
This is undesirable for the same complexity management reasons as having
a huge Standard dialect itself, and is even more confusing given the
existence of a separate dialect. Extract the conversion of the Math
dialect operations to LLVM into a separate library and a separate
conversion pass.
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D105702
After the MemRef has been split out of the Standard dialect, the
conversion to the LLVM dialect remained as a huge monolithic pass.
This is undesirable for the same complexity management reasons as having
a huge Standard dialect itself, and is even more confusing given the
existence of a separate dialect. Extract the conversion of the MemRef
dialect operations to LLVM into a separate library and a separate
conversion pass.
Reviewed By: herhut, silvas
Differential Revision: https://reviews.llvm.org/D105625
This class and classes that extend it are general utilities for any dialect
that is being converted into the LLVM dialect. They are in no way specific to
Standard-to-LLVM conversion and should not make their users depend on it.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D105542
Simplify vector unrolling pattern to be more aligned with rest of the
patterns and be closer to vector distribution.
The new implementation uses ExtractStridedSlice/InsertStridedSlice
instead of the Tuple ops. After this change the ops based on Tuple don't
have any more used so they can be removed.
This allows removing signifcant amount of dead code and will allow
extending the unrolling code going forward.
Differential Revision: https://reviews.llvm.org/D105381
"Standard-to-LLVM" conversion is one of the oldest passes in existence. It has
become quite large due to the size of the Standard dialect itself, which is
being split into multiple smaller dialects. Furthermore, several conversion
features are useful for any dialect that is being converted to the LLVM
dialect, which, without this refactoring, creates a dependency from those
conversions to the "standard-to-llvm" one.
Put several of the reusable utilities from this conversion to a separate
library, namely:
- type converter from builtin to LLVM dialect types;
- utility for building and accessing values of LLVM structure type;
- utility for building and accessing values that represent memref in the LLVM
dialect;
- lowering options applicable everywhere.
Additionally, remove the type wrapping/unwrapping notion from the type
converter that is no longer relevant since LLVM types has been reimplemented as
first-class MLIR types.
Reviewed By: pifon2a
Differential Revision: https://reviews.llvm.org/D105534
Split out GPU ops library from GPU transforms. This allows libraries to
depend on GPU Ops without needing/building its transforms.
Differential Revision: https://reviews.llvm.org/D105472
Remove `getDynOperands` and `createOrFoldDimOp` from MemRef.h to decouple MemRef a bit from Tensor. These two functions are used in other dialects/transforms.
Differential Revision: https://reviews.llvm.org/D105260
* Split memref.dim into two operations: memref.dim and tensor.dim. Both ops have the same builder interface and op argument names, so that they can be used with templates in patterns that apply to both tensors and memrefs (e.g., some patterns in Linalg).
* Add constant materializer to TensorDialect (needed for folding in affine.apply etc.).
* Remove some MemRefDialect dependencies, make some explicit.
Differential Revision: https://reviews.llvm.org/D105165
This patch brings support for setting runtime preemption specifiers of
LLVM's GlobalValues. In LLVM semantics, if the `dso_local` attribute
is not explicitly requested, then it is inferred based on linkage and
visibility. We model this same behavior with a UnitAttribute: if it is
present, then we explicitly request the GlobalValue to marked as
`dso_local`, otherwise we rely on the GlobalValue itself to make this
decision.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D104983
Reduce code duplication: Move various helper functions, that are duplicated in TensorDialect, MemRefDialect, LinalgDialect, StandardDialect, into a new StaticValueUtils.cpp.
Differential Revision: https://reviews.llvm.org/D104687
Specify the `!async.group` size (the number of tokens that will be added to it) at construction time. `async.await_all` operation can potentially race with `async.execute` operations that keep updating the group, for this reason it is required to know upfront how many tokens will be added to the group.
Reviewed By: ftynse, herhut
Differential Revision: https://reviews.llvm.org/D104780
This commit moves the type translator from LLVM to MLIR to a public header for use by external projects or other code.
Unlike a previous attempt (https://reviews.llvm.org/D104726), this patch moves the type conversion into separate files which remedies the linker error which was only caught by CI.
Differential Revision: https://reviews.llvm.org/D104834
The main goal of this commit is to remove the dependency of Standard dialect on the Tensor dialect.
* Rename SubTensorOp -> tensor.extract_slice, SubTensorInsertOp -> tensor.insert_slice.
* Some helper functions are (already) duplicated between the Tensor dialect and the MemRef dialect. To keep this commit smaller, this will be cleaned up in a separate commit.
* Additional dialect dependencies: Shape --> Tensor, Tensor --> Standard
* Remove dialect dependencies: Standard --> Tensor
* Move canonicalization test cases to correct dialect (Tensor/MemRef).
Note: This is a fixed version of https://reviews.llvm.org/D104499, which was reverted due to a missing update to two CMakeFile.txt.
Differential Revision: https://reviews.llvm.org/D104676
Adapt the FillOp definition to use a scalar operand instead of a capture. This patch is a follow up to https://reviews.llvm.org/D104109. As the input operands are in front of the output operands the patch changes the internal operand order of the FillOp. The pretty printed version of the operation remains unchanged though. The patch also adapts the linalg to standard lowering to ensure the c signature of the FillOp remains unchanged as well.
Differential Revision: https://reviews.llvm.org/D104121
The main goal of this commit is to remove the dependency of Standard dialect on the Tensor dialect.
* Rename ops: SubTensorOp --> ExtractTensorOp, SubTensorInsertOp --> InsertTensorOp
* Some helper functions are (already) duplicated between the Tensor dialect and the MemRef dialect. To keep this commit smaller, this will be cleaned up in a separate commit.
* Additional dialect dependencies: Shape --> Tensor, Tensor --> Standard
* Remove dialect dependencies: Standard --> Tensor
* Move canonicalization test cases to correct dialect (Tensor/MemRef).
Differential Revision: https://reviews.llvm.org/D104499
* Remove dependency: Standard --> MemRef
* Add dependencies: GPUToNVVMTransforms --> MemRef, Linalg --> MemRef, MemRef --> Tensor
* Note: The `subtensor_insert_propagate_dest_cast` test case in MemRef/canonicalize.mlir will be moved to Tensor/canonicalize.mlir in a subsequent commit, which moves over the remaining Tensor ops from the Standard dialect to the Tensor dialect.
Differential Revision: https://reviews.llvm.org/D104506
The index cast operation accepts vector types. Implement its lowering in this patch.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D104280
There's no need for `toSmallVector()` as `SmallVector.h` already provides a `to_vector` free function that takes a range.
Reviewed By: Quuxplusone
Differential Revision: https://reviews.llvm.org/D104024
## Introduction
This proposal describes the new op to be added to the `std` (and later moved `memref`)
dialect called `alloca_scope`.
## Motivation
Alloca operations are easy to misuse, especially if one relies on it while doing
rewriting/conversion passes. For example let's consider a simple example of two
independent dialects, one defines an op that wants to allocate on-stack and
another defines a construct that corresponds to some form of looping:
```
dialect1.looping_op {
%x = dialect2.stack_allocating_op
}
```
Since the dialects might not know about each other they are going to define a
lowering to std/scf/etc independently:
```
scf.for … {
%x_temp = std.alloca …
… // do some domain-specific work using %x_temp buffer
… // and store the result into %result
%x = %result
}
```
Later on the scf and `std.alloca` is going to be lowered to llvm using a
combination of `llvm.alloca` and unstructured control flow.
At this point the use of `%x_temp` is bound to either be either optimized by
llvm (for example using mem2reg) or in the worst case: perform an independent
stack allocation on each iteration of the loop. While the llvm optimizations are
likely to succeed they are not guaranteed to do so, and they provide
opportunities for surprising issues with unexpected use of stack size.
## Proposal
We propose a new operation that defines a finer-grain allocation scope for the
alloca-allocated memory called `alloca_scope`:
```
alloca_scope {
%x_temp = alloca …
...
}
```
Here the lifetime of `%x_temp` is going to be bound to the narrow annotated
region within `alloca_scope`. Moreover, one can also return values out of the
alloca_scope with an accompanying `alloca_scope.return` op (that behaves
similarly to `scf.yield`):
```
%result = alloca_scope {
%x_temp = alloca …
…
alloca_scope.return %myvalue
}
```
Under the hood the `alloca_scope` is going to lowered to a combination of
`llvm.intr.stacksave` and `llvm.intr.strackrestore` that are going to be invoked
automatically as control-flow enters and leaves the body of the `alloca_scope`.
The key value of the new op is to allow deterministic guaranteed stack use
through an explicit annotation in the code which is finer-grain than the
function-level scope of `AutomaticAllocationScope` interface. `alloca_scope`
can be inserted at arbitrary locations and doesn’t require non-trivial
transformations such as outlining.
## Which dialect
Before memref dialect is split, `alloca_scope` can temporarily reside in `std`
dialect, and later on be moved to `memref` together with the rest of
memory-related operations.
## Implementation
An implementation of the op is available [here](https://reviews.llvm.org/D97768).
Original commits:
* Add initial scaffolding for alloca_scope op
* Add alloca_scope.return op
* Add no region arguments and variadic results
* Add op descriptions
* Add failing test case
* Add another failing test
* Initial implementation of lowering for std.alloca_scope
* Fix backticks
* Fix getSuccessorRegions implementation
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D97768
This is the first step to convert vector ops to MMA operations in order to
target GPUs tensor core ops. This currently only support simple cases,
transpose and element-wise operation will be added later.
Differential Revision: https://reviews.llvm.org/D102962
This adds Sdot2d op, which is similar to the usual Neon
intrinsic except that it takes 2d vector operands, reflecting the
structure of the arithmetic that it's performing: 4 separate
4-dimensional dot products, whence the vector<4x4xi8> shape.
This also adds a new pass, arm-neon-2d-to-intr, lowering
this new 2d op to the 1d intrinsic.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D102504
This allow creating a matrix with all elements set to a given value. This is
needed to be able to implement a simple dot op.
Differential Revision: https://reviews.llvm.org/D103870
This is a roll forward of D102679.
This patch simplifies the implementation of Sequence and makes it compatible with llvm::reverse.
It exposes the reverse iterators through rbegin/rend which prevents a dangling reference in std::reverse_iterator::operator++().
Note: Compared to D102679, this patch introduces a `asSmallVector()` member function and fixes compilation issue with GCC 5.
Differential Revision: https://reviews.llvm.org/D103948
tosa.matmul is a batched matmul, update the lowering for linalg
with the tests.
Reviewed By: sjarus
Differential Revision: https://reviews.llvm.org/D103937
This allows us to remove the `spv.mlir.endmodule` op and
all the code associated with it.
Along the way, tightened the APIs for `spv.module` a bit
by removing some aliases. Now we use `getRegion` to get
the only region, and `getBody` to get the region's only
block.
Reviewed By: mravishankar, hanchung
Differential Revision: https://reviews.llvm.org/D103265
Consolidate the type conversion in a single function to make it simpler
to use. This allow to re-use the type conversion for up coming ops.
Differential Revision: https://reviews.llvm.org/D103868
This reverts commit e772216e70
(and fixup 7f6c878a2c).
The build is broken with gcc5 host compiler:
In file included from
from mlir/lib/Dialect/Utils/StructuredOpsUtils.cpp:9:
tools/mlir/include/mlir/IR/BuiltinAttributes.h.inc:424:57: error: type/value mismatch at argument 1 in template parameter list for 'template<class ItTy, class FuncTy, class FuncReturnTy> class llvm::mapped_iterator'
std::function<T(ptrdiff_t)>>;
^
tools/mlir/include/mlir/IR/BuiltinAttributes.h.inc:424:57: note: expected a type, got 'decltype (seq<ptrdiff_t>(0, 0))::const_iterator'
This is both more efficient and more ergonomic than going
through an std::string, e.g. when using llvm::utostr and
in string concat cases.
Unfortunately we can't just overload ::get(). This causes an
ambiguity because both twine and stringref implicitly convert
from std::string.
Differential Revision: https://reviews.llvm.org/D103754
This patch simplifies the implementation of Sequence and makes it compatible with llvm::reverse.
It exposes the reverse iterators through rbegin/rend which prevents a dangling reference in std::reverse_iterator::operator++().
Differential Revision: https://reviews.llvm.org/D102679
Now that memref supports arbitrary element types, add support for memref of
memref and make sure it is properly converted to the LLVM dialect. The type
support itself avoids adding the interface to the memref type itself similarly
to other built-in types. This allows the shape, and therefore byte size, of the
memref descriptor to remain a lowering aspect that is easier to customize and
evolve as opposed to sanctifying it in the data layout specification for the
memref type itself.
Factor out the code previously in a testing pass to live in a dedicated data
layout analysis and use that analysis in the conversion to compute the
allocation size for memref of memref. Other conversions will be ported
separately.
Depends On D103827
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D103828
Historically, MemRef only supported a restricted list of element types that
were known to be storable in memory. This is unnecessarily restrictive given
the open nature of MLIR's type system. Allow types to opt into being used as
MemRef elements by implementing a type interface. For now, the interface is
merely a declaration with no methods. Later, methods to query, e.g., the type
size or whether a type can alias elements of another type may be added.
Harden the "standard"-to-LLVM conversion against memrefs with non-builtin
types.
See https://llvm.discourse.group/t/rfc-memref-of-custom-types/3558.
Depends On D103826
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D103827
Some places in the alloc-like op conversion use the converted index type
whereas other places use the pointer-sized integer type, which may not be the
same. Consistently use the converted index type, similarly to other address
calculations.
Reviewed By: pifon2a
Differential Revision: https://reviews.llvm.org/D103826
This patch convert the if condition on standalone data operation such as acc.update,
acc.enter_data and acc.exit_data to a scf.if with the operation in the if region.
It removes the operation when the if condition is constant and false. It removes the
the condition if it is contant and true.
Conversion to scf.if is done in order to use the translation to LLVM IR dialect out of the box.
Not sure this is the best approach or we should perform this during the translation from OpenACC
to LLVM IR dialect. Any thoughts welcome.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D103325
Convert data operands from the acc.parallel operation using the same conversion pattern than D102170.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D103337
Convert data operands from the acc.data operation using the same conversion pattern than D102170.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D103332
Support for tensor types in the unrolled version will follow in a separate commit.
Add a new pass option to activate lowering of transfer ops with tensor types (default: deactivated).
Differential Revision: https://reviews.llvm.org/D102666
Depthwise convolution should support kernel dilation and non-dilation should
not be a special case. Updated op definition to include a dilation attribute.
This also adds a tosa.depthwise_conv2d lowering to linalg to support the new
linalg behavior.
Differential Revision: https://reviews.llvm.org/D103219
Previously, this assumed use of ModuleOp and FuncOp. There is no need to
restrict this, and using interfaces allows these patterns to be used
during dialect conversion to LLVM.
Some assertions were removed due to inconsistent implementation of
FunctionLikeOps.
Differential Revision: https://reviews.llvm.org/D103447
No tests fail and this seems to be technical debt from when the math
dialect was created. This should not be there as it prevents users from
configuring their converion target freely and results in unexpected
behavior on seemingly unrelated ops.
Differential Revision: https://reviews.llvm.org/D103388
Depends On D103109
If any of the tokens/values added to the `!async.group` switches to the error state, than the group itself switches to the error state.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D103203
Depends On D103102
Not yet implemented:
1. Error handling after synchronous await
2. Error handling for async groups
Will be addressed in the followup PRs
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D103109
In order to allow large matmul operations using the MMA ops we need to chain
operations this is not possible unless "DOp" and "COp" type have matching
layout so remove the "DOp" layout and force accumulator and result type to
match.
Added a test for the case where the MMA value is accumulated.
Differential Revision: https://reviews.llvm.org/D103023
The casting ops (sitofp, uitofp, fptosi, fptoui) lowering currently does
not handle n-D vectors. This patch fixes that.
Differential Revision: https://reviews.llvm.org/D103207
Indexed Generic should be going away in the future. Migrate to linalg.index.
Reviewed By: NatashaKnk, nicolasvasilache
Differential Revision: https://reviews.llvm.org/D103110
Disallow transfer ops that change the element type of the transfer. Such transfers could be supported in the future, if needed.
Differential Revision: https://reviews.llvm.org/D102746
Lower a 1D vector transfer op to LLVM if the last dim stride is 1. Also fixes a bug in the original unit stride computation.
Differential Revision: https://reviews.llvm.org/D102897
Fix inconsistent MLIR CMake variable names. Consistently name them as
MLIR_ENABLE_<feature>.
Eg: MLIR_CUDA_RUNNER_ENABLED -> MLIR_ENABLE_CUDA_RUNNER
MLIR follows (or has mostly followed) the convention of naming
cmake enabling variables in the from MLIR_ENABLE_... etc. Using a
convention here is easy and also important for convenience. A counter
pattern was started with variables named MLIR_..._ENABLED. This led to a
sequence of related counter patterns: MLIR_CUDA_RUNNER_ENABLED,
MLIR_ROCM_RUNNER_ENABLED, etc.. From a naming standpoint, the imperative
form is more meaningful. Additional discussion at:
https://llvm.discourse.group/t/mlir-cmake-enable-variable-naming-convention/3520
Switch all inconsistent ones to the ENABLE form. Keep the couple of old
mappings needed until buildbot config is migrated.
Differential Revision: https://reviews.llvm.org/D102976
This adds the straightforward conversion for EqualOp
(two complex numbers are equal if both the real and the imaginary part are equal).
Differential Revision: https://reviews.llvm.org/D102840
vector.transfer_read and vector.transfer_write operations are converted
to llvm intrinsics with specific alignment information, however there
doesn't seem to be a way in llvm to take information from llvm.assume
intrinsics and change this alignment information. In any
event, due the to the structure of the llvm.assume instrinsic, applying
this information at the llvm level is more cumbersome. Instead, let's
generate the masked vector load and store instrinsic with the right
alignment information from MLIR in the first place. Since
we're bothering to do this, lets just emit the proper alignment for
loads, stores, scatter, and gather ops too.
Differential Revision: https://reviews.llvm.org/D100444
VectorTransferPermutationMapLoweringPatterns can be enabled via a pass option. These additional patterns lower permutation maps to minor identity maps with broadcasting, if possible, allowing for more efficient vector load/stores. The option is deactivated by default.
Differential Revision: https://reviews.llvm.org/D102593
This is a hook that allows for providing custom initialization of the pattern, e.g. if it has bounded recursion, setting the debug name, etc., without needing to define a custom constructor. A non-virtual hook was chosen to avoid polluting the vtable with code that we really just want to be inlined when constructing the pattern. The alternative to this would be to just define a constructor for each pattern, this unfortunately creates a lot of otherwise unnecessary boiler plate for a lot of patterns and a hook provides a much simpler/cleaner interface for the very common case.
Differential Revision: https://reviews.llvm.org/D102440
Comment was poorly written. Changed to bail on contradictory information in
the double round.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D102651
Initial version of pooling assumed normalization was accross all elements
equally. TOSA actually requires the noramalization is perform by how
many elements were summed (edges are not artifically dimmer). Updated
the lowering to reflect this change with corresponding tests.
Reviewed By: NatashaKnk
Differential Revision: https://reviews.llvm.org/D102540
Splitting the memref dialect lead to an introduction of several dependencies
to avoid compilation issues. The canonicalize pass also depends on the
memref dialect, but it shouldn't. This patch resolves the dependencies
and the unintuitive includes are removed. However, the dependency moves
to the constructor of the std dialect.
Differential Revision: https://reviews.llvm.org/D102060
Group functions/structs in namespaces for better code readability.
Depends On D102123
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D102124
Make "target rank" a pass option of VectorToSCF.
Depends On D102101
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D102123
Lowering div elementwise op to the linalg dialect. Since tosa only supports integer division, that is the only version that is currently implemented.
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D102430
Create a copy of vector-to-loops.mlir and adapt the test for
ProgressiveVectorToSCF. Fix a small bug in getExtractOp() triggered by
this test.
Differential Revision: https://reviews.llvm.org/D102388
Do not rely on pass labels to detect if the pattern was already applied in the past (which allows for more some extra optimizations to avoid extra InsertOps and ExtractOps). Instead, check if these optimizations can be applied on-the-fly.
This also fixes a bug, where vector.insert and vector.extract ops sometimes disappeared in the middle of the pass because they get folded away, but the next application of the pattern expected them to be there.
Differential Revision: https://reviews.llvm.org/D102206
Rounding to integers requires rounding (for floating points) and clipping
to the min/max values of the destination range. Added this behavior and
updated tests appropriately.
Reviewed By: sjarus, silvas
Differential Revision: https://reviews.llvm.org/D102375
Instead of an SCF for loop, these pattern generate fully unrolled loops with no temporary buffer allocations.
Differential Revision: https://reviews.llvm.org/D101981
Broadcast dimensions of a vector transfer op have no corresponding dimension in the mask vector. E.g., a 2-D TransferReadOp, where one dimension is a broadcast, can have a 1-D `mask` attribute.
This commit also adds a few additional transfer op integration tests for various combinations of broadcasts, masking, dim transposes, etc.
Differential Revision: https://reviews.llvm.org/D101745
Broadcast dimensions of a vector transfer op have no corresponding dimension in the mask vector. E.g., a 2-D TransferReadOp, where one dimension is a broadcast, can have a 1-D `mask` attribute.
This commit also adds a few additional transfer op integration tests for various combinations of broadcasts, masking, dim transposes, etc.
Differential Revision: https://reviews.llvm.org/D101745
Add a conversion pass to convert higher-level type before translation.
This conversion extract meangingful information and pack it into a struct that
the translation (D101504) will be able to understand.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D102170
First step in adding alignment as an attribute to MLIR global definitions. Alignment can be specified for global objects in LLVM IR. It can also be specified as a named attribute in the LLVMIR dialect of MLIR. However, this attribute has no standing and is discarded during translation from MLIR to LLVM IR. This patch does two things: First, it adds the attribute to the syntax of the llvm.mlir.global operation, and by doing this it also adds accessors and verifications. The syntax is "align=XX" (with XX being an integer), placed right after the value of the operation. Second, it allows transforming this operation to and from LLVM IR. It is checked whether the value is an integer power of 2.
Reviewed By: ftynse, mehdi_amini
Differential Revision: https://reviews.llvm.org/D101492
Updated tests to include broadcast of left and right. Includes
bypass if in-type and out-type match shape (no broadcasting).
Differential Revision: https://reviews.llvm.org/D102276
The current design uses a unique entry for each argument/result attribute, with the name of the entry being something like "arg0". This provides for a somewhat sparse design, but ends up being much more expensive (from a runtime perspective) in-practice. The design requires building a string every time we lookup the dictionary for a specific arg/result, and also requires N attribute lookups when collecting all of the arg/result attribute dictionaries.
This revision restructures the design to instead have an ArrayAttr that contains all of the attribute dictionaries for arguments and another for results. This design reduces the number of attribute name lookups to 1, and allows for O(1) lookup for individual element dictionaries. The major downside is that we can end up with larger memory usage, as the ArrayAttr contains an entry for each element even if that element has no attributes. If the memory usage becomes too problematic, we can experiment with a more sparse structure that still provides a lot of the wins in this revision.
This dropped the compilation time of a somewhat large TensorFlow model from ~650 seconds to ~400 seconds.
Differential Revision: https://reviews.llvm.org/D102035
Implements support for undialated depthwise convolution using the existing
depthwise convolution operation. Once convolutions migrate to yaml defined
versions we can rewrite for cleaner implementation.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D101579
This patch adds support for vectorizing loops with 'iter_args'
implementing known reductions along the vector dimension. Comparing to
the non-vector-dimension case, two additional things are done during
vectorization of such loops:
- The resulting vector returned from the loop is reduced to a scalar
using `vector.reduce`.
- In some cases a mask is applied to the vector yielded at the end of
the loop to prevent garbage values from being written to the
accumulator.
Vectorization of reduction loops is disabled by default. To enable it, a
map from loops to array of reduction descriptors should be explicitly passed to
`vectorizeAffineLoops`, or `vectorize-reductions=true` should be passed
to the SuperVectorize pass.
Current limitations:
- Loops with a non-unit step size are not supported.
- n-D vectorization with n > 1 is not supported.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D100694
All linalg.init operations must be fed into a linalg operation before
subtensor. The inserted linalg.fill guarantees it executes correctly.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D101848
Lowerings equal and arithmetic_right_shift for elementwise ops to linalg dialect using linalg.generic
Reviewed By: rsuderman
Differential Revision: https://reviews.llvm.org/D101804
Move TransposeOp lowering in its own populate function as in some cases
it is better to keep it during ContractOp lowering to better
canonicalize it rather than emiting scalar insert/extract.
Differential Revision: https://reviews.llvm.org/D101647
Constant-0 dim expr values should be avoided for linalg as it can prevent
fusion. This includes adding support for rank-0 reshapes.
Differential Revision: https://reviews.llvm.org/D101418
This enables to express more complex parallel loops in the affine framework,
for example, in cases of tiling by sizes not dividing loop trip counts perfectly
or inner wavefront parallelism, among others. One can't use affine.max/min
and supply values to the nested loop bounds since the results of such
affine.max/min operations aren't valid symbols. Making them valid symbols
isn't an option since they would introduce selection trees into memref
subscript arithmetic as an unintended and undesired consequence. Also
add support for converting such loops to SCF. Drop some API that isn't used in
the core repo from AffineParallelOp since its semantics becomes ambiguous in
presence of max/min bounds. Loop normalization is currently unavailable for
such loops.
Depends On D101171
Reviewed By: bondhugula
Differential Revision: https://reviews.llvm.org/D101172
Benjamin Kramer