This speeds up registered op creation by 10-11% by allowing lookup by
TypeID instead of StringRef.
This can break your build/tests at runtime with an error that you're creating
an unregistered operation that you have registered. If so you are likely using
a class inheriting from the "real" operation. See for example in this patch the
case of:
class ConstantIndexOp : public arith::ConstantOp {
If one is using `builder.create<ConstantIndexOp>()` they actually create an
`arith.constant` operation, but the builder will fetch the TypeID for
the `ConstantIndexOp` class which does not correspond to any registered
operation. To fix it the `ConstantIndexOp` class got this addition:
static ::mlir::TypeID resolveTypeID() { return TypeID::get<ConstantOp>(); }
The base class llvm::ThreadPoolInterface will be renamed
llvm::ThreadPool in a subsequent commit.
This is a breaking change: clients who use to create a ThreadPool must
now create a DefaultThreadPool instead.
This patch expose the type and attribute names in C++ as methods in the
`AbstractType` and `AbstractAttribute` classes, and keep a map of names
to `AbstractType` and `AbstractAttribute` in the `MLIRContext`. Type and
attribute names should be unique.
It adds support in ODS to generate the `getName` methods in
`AbstractType` and `AbstractAttribute`, through the use of two new
variables, `typeName` and `attrName`. It also adds names to C++-defined
type and attributes.
This is a follow-up to 8c2bff1ab9 which lazy-initialized the
diagnostic and removed the need to dynamically abandon() an
InFlightDiagnostic. This further simplifies the code to not needed to
return a reference to an InFlightDiagnostic and instead eagerly emit
errors.
Also use `emitError` as name instead of `getDiag` which seems more
explicit and in-line with the common usage.
Add capabilities for comparing opaque properties. This is useful when
dealing with arbitrary operations which can be compare based on their
OperationName. Now you can furthermore compare their properties without
the need to determine their actual type.
A distinct attribute associates a referenced attribute with a unique
identifier. Every call to its create function allocates a new
distinct attribute instance. The address of the attribute instance
temporarily serves as its unique identifier. Similar to the names
of SSA values, the final unique identifiers are generated during
pretty printing.
Examples:
#distinct = distinct[0]<42.0 : f32>
#distinct1 = distinct[1]<42.0 : f32>
#distinct2 = distinct[2]<array<i32: 10, 42>>
This mechanism is meant to generate attributes with a unique
identifier, which can be used to mark groups of operations
that share a common properties such as if they are aliasing.
The design of the distinct attribute ensures minimal memory
footprint per distinct attribute since it only contains a reference
to another attribute. All distinct attributes are stored outside of
the storage uniquer in a thread local store that is part of the
context. It uses one bump pointer allocator per thread to ensure
distinct attributes can be created in-parallel.
Reviewed By: rriddle, Dinistro, zero9178
Differential Revision: https://reviews.llvm.org/D153360
applyExtensions can load further dialects, invalidating the reference to
the dialect pointer in the dialects DenseMap. Capture the pointer to
prevent that from happening.
Currently, the dialects precede the registered operations in the context object, which means that the latter is destroyed first. At the same time, Operation::~Operation dereferences the registered operation when destroying properties, which can cause use-after-free (e.g. if a dialect owns an op). This patch fixes that by changing the order of the members so that dialects come after registered operations.
Differential Revision: https://reviews.llvm.org/D151440
This new features enabled to dedicate custom storage inline within operations.
This storage can be used as an alternative to attributes to store data that is
specific to an operation. Attribute can also be stored inside the properties
storage if desired, but any kind of data can be present as well. This offers
a way to store and mutate data without uniquing in the Context like Attribute.
See the OpPropertiesTest.cpp for an example where a struct with a
std::vector<> is attached to an operation and mutated in-place:
struct TestProperties {
int a = -1;
float b = -1.;
std::vector<int64_t> array = {-33};
};
More complex scheme (including reference-counting) are also possible.
The only constraint to enable storing a C++ object as "properties" on an
operation is to implement three functions:
- convert from the candidate object to an Attribute
- convert from the Attribute to the candidate object
- hash the object
Optional the parsing and printing can also be customized with 2 extra
functions.
A new options is introduced to ODS to allow dialects to specify:
let usePropertiesForAttributes = 1;
When set to true, the inherent attributes for all the ops in this dialect
will be using properties instead of being stored alongside discardable
attributes.
The TestDialect showcases this feature.
Another change is that we introduce new APIs on the Operation class
to access separately the inherent attributes from the discardable ones.
We envision deprecating and removing the `getAttr()`, `getAttrsDictionary()`,
and other similar method which don't make the distinction explicit, leading
to an entirely separate namespace for discardable attributes.
Recommit d572cd1b06 after fixing python bindings build.
Differential Revision: https://reviews.llvm.org/D141742
This new features enabled to dedicate custom storage inline within operations.
This storage can be used as an alternative to attributes to store data that is
specific to an operation. Attribute can also be stored inside the properties
storage if desired, but any kind of data can be present as well. This offers
a way to store and mutate data without uniquing in the Context like Attribute.
See the OpPropertiesTest.cpp for an example where a struct with a
std::vector<> is attached to an operation and mutated in-place:
struct TestProperties {
int a = -1;
float b = -1.;
std::vector<int64_t> array = {-33};
};
More complex scheme (including reference-counting) are also possible.
The only constraint to enable storing a C++ object as "properties" on an
operation is to implement three functions:
- convert from the candidate object to an Attribute
- convert from the Attribute to the candidate object
- hash the object
Optional the parsing and printing can also be customized with 2 extra
functions.
A new options is introduced to ODS to allow dialects to specify:
let usePropertiesForAttributes = 1;
When set to true, the inherent attributes for all the ops in this dialect
will be using properties instead of being stored alongside discardable
attributes.
The TestDialect showcases this feature.
Another change is that we introduce new APIs on the Operation class
to access separately the inherent attributes from the discardable ones.
We envision deprecating and removing the `getAttr()`, `getAttrsDictionary()`,
and other similar method which don't make the distinction explicit, leading
to an entirely separate namespace for discardable attributes.
Differential Revision: https://reviews.llvm.org/D141742
The way dependent dialects are implemented is by recursively calling
loadDialect in the constructor. This means we have to reload from the
dialect table because the constructor might have rehashed that table.
The steps for loading a dialect are
1. Insert a nullptr into loadedDialects. This indicates the dialect is
loading
2. Call ctor(). This recursively loads dependent dialects
3. Insert the new dialect into the table.
We had a conflict between steps 2 and 3 here. You have to be extremely
unlucky though as rehashing is rare and operator[] does no generation
checking on DenseMap. Changing that to an iterator would've uncovered
this issue immediately.
X. Sun et al. (https://dl.acm.org/doi/10.5555/3454287.3454728) published
a paper showing that an FP format with 4 bits of exponent, 3 bits of
significand and an exponent bias of 11 would work quite well for ML
applications.
Google hardware supports a variant of this format where 0x80 is used to
represent NaN, as in the Float8E4M3FNUZ format. Just like the
Float8E4M3FNUZ format, this format does not support -0 and values which
would map to it will become +0.
This format is proposed for inclusion in OpenXLA's StableHLO dialect: https://github.com/openxla/stablehlo/pull/1308
As part of inclusion in that dialect, APFloat needs to know how to
handle this format.
Differential Revision: https://reviews.llvm.org/D146441
The concept of the ActionManager acts as a sort of "Hub" that can receive
various types of action and dispatch them to a set of registered handlers.
One handler will handle the action or it'll cascade to other handlers.
This model does not really fit the current evolution of the Action tracing
and debugging: we can't foresee a good case where this behavior compose with
the use-case behind the handlers. Instead we simplify it with a single
callback installed on the Context.
Differential Revision: https://reviews.llvm.org/D144811
This is a preparation for adding support for more infrastructure around the concept
of Action and make tracing Action more of a first class concept.
The doc will be updated later in a subsequent revision after the changes are
completed.
Action belongs to IR because of circular dependency: Actions are dispatched through
the MLIRContext but Action will learn to encapsulate IR construct.
Differential Revision: https://reviews.llvm.org/D144809
Float8E5M2FNUZ and Float8E4M3FNUZ have been added to APFloat in D141863.
This change adds these types as MLIR builtin types alongside Float8E5M2
and Float8E4M3FN (added in D133823 and D138075).
Reviewed By: krzysz00
Differential Revision: https://reviews.llvm.org/D143744
This allows for interfaces to define a set of "base classes",
which are interfaces whose methods/extra class decls/etc.
should be inherited by the derived interface. This more
easily enables combining interfaces and their dependencies,
without lots of awkard casting. Additional implicit conversion
operators also greatly simplify the conversion process.
One other aspect of this "inheritance" is that we also implicitly
add the base interfaces to the attr/op/type. The user can still
add them manually if desired, but this should help remove some
of the boiler plate when an interface has dependencies.
See https://discourse.llvm.org/t/interface-inheritance-and-dependencies-interface-method-visibility-interface-composition
Differential Revision: https://reviews.llvm.org/D140198
This streamlines the implementation and makes it so that the virtual
tables are in the binary instead of dynamically assembled during initialization.
The dynamic allocation size of op registration is also smaller with this
change.
This reverts commit 7bf1e441da
and re-introduce e055aad5ff
after fixing the windows crash by making ParseAssemblyFn a
unique_function again
Differential Revision: https://reviews.llvm.org/D141492
This streamlines the implementation and makes it so that the virtual tables are in the binary instead of dynamically assembled during initialization.
The dynamic allocation size of op registration is also smaller with this
change.
Differential Revision: https://reviews.llvm.org/D141492
This patch mechanically replaces None with std::nullopt where the
compiler would warn if None were deprecated. The intent is to reduce
the amount of manual work required in migrating from Optional to
std::optional.
This is part of an effort to migrate from llvm::Optional to
std::optional:
https://discourse.llvm.org/t/deprecating-llvm-optional-x-hasvalue-getvalue-getvalueor/63716
This change fixes a bug where a dialect is initialized multiple times. This triggers an assertion when the ops of the dialect are registered (`error: operation named ... is already registered`).
This bug can be triggered as follows:
1. Dialect A depends on dialect B (as per ADialect.td).
2. Somewhere there is an extension of dialect B that depends on dialect A (e.g., it defines external models create ops from dialect A). E.g.:
```
registry.addExtension(+[](MLIRContext *ctx, BDialect *dialect) {
BDialectOp::attachInterface ...
ctx->loadDialect<ADialect>();
});
```
3. When dialect A is loaded, its `initialize` function is called twice:
```
ADialect::ADialect()
| |
| v
| ADialect::initialize()
v
getOrLoadDialect<BDialect>()
|
v
(load extension of BDialect)
|
v
ctx->loadDialect<ADialect>() // user wrote this in the extension
|
v
getOrLoadDialect<ADialect>() // the dialect is not "fully" loaded yet
|
v
ADialect::ADialect()
|
v
ADialect::initialize()
```
An example of a dialect extension that depends on other dialects is `Dialect/Tensor/Transforms/BufferizableOpInterfaceImpl.cpp`. That particular dialect extension does not trigger this bug. (It would trigger this bug if the SCF dialect would depend on the Tensor dialect.)
This change introduces a new dialect state: dialects that are currently being loaded. Same as dialects that were already fully loaded (and initialized), dialects that are in the process of being loaded are not loaded a second time.
Differential Revision: https://reviews.llvm.org/D136685
(Re-Apply with fixes to clang MicrosoftMangle.cpp)
This is a first step towards high level representation for fp8 types
that have been built in to hardware with near term roadmaps. Like the
BFLOAT16 type, the family of fp8 types are inspired by IEEE-754 binary
floating point formats but, due to the size limits, have been tweaked in
various ways in order to maximally use the range/precision in various
scenarios. The list of variants is small/finite and bounded by real
hardware.
This patch introduces the E5M2 FP8 format as proposed by Nvidia, ARM,
and Intel in the paper: https://arxiv.org/pdf/2209.05433.pdf
As the more conformant of the two implemented datatypes, we are plumbing
it through LLVM's APFloat type and MLIR's type system first as a
template. It will be followed by the range optimized E4M3 FP8 format
described in the paper. Since that format deviates further from the
IEEE-754 norms, it may require more debate and implementation
complexity.
Given that we see two parts of the FP8 implementation space represented
by these cases, we are recommending naming of:
* `F8M<N>` : For FP8 types that can be conceived of as following the
same rules as FP16 but with a smaller number of mantissa/exponent
bits. Including the number of mantissa bits in the type name is enough
to fully specify the type. This naming scheme is used to represent
the E5M2 type described in the paper.
* `F8M<N>F` : For FP8 types such as E4M3 which only support finite
values.
The first of these (this patch) seems fairly non-controversial. The
second is previewed here to illustrate options for extending to the
other known variant (but can be discussed in detail in the patch
which implements it).
Many conversations about these types focus on the Machine-Learning
ecosystem where they are used to represent mixed-datatype computations
at a high level. At that level (which is why we also expose them in
MLIR), it is important to retain the actual type definition so that when
lowering to actual kernels or target specific code, the correct
promotions, casts and rescalings can be done as needed. We expect that
most LLVM backends will only experience these types as opaque `I8`
values that are applicable to some instructions.
MLIR does not make it particularly easy to add new floating point types
(i.e. the FloatType hierarchy is not open). Given the need to fully
model FloatTypes and make them interop with tooling, such types will
always be "heavy-weight" and it is not expected that a highly open type
system will be particularly helpful. There are also a bounded number of
floating point types in use for current and upcoming hardware, and we
can just implement them like this (perhaps looking for some cosmetic
ways to reduce the number of places that need to change). Creating a
more generic mechanism for extending floating point types seems like it
wouldn't be worth it and we should just deal with defining them one by
one on an as-needed basis when real hardware implements a new scheme.
Hopefully, with some additional production use and complete software
stacks, hardware makers will converge on a set of such types that is not
terribly divergent at the level that the compiler cares about.
(I cleaned up some old formatting and sorted some items for this case:
If we converge on landing this in some form, I will NFC commit format
only changes as a separate commit)
Differential Revision: https://reviews.llvm.org/D133823
This is a first step towards high level representation for fp8 types
that have been built in to hardware with near term roadmaps. Like the
BFLOAT16 type, the family of fp8 types are inspired by IEEE-754 binary
floating point formats but, due to the size limits, have been tweaked in
various ways in order to maximally use the range/precision in various
scenarios. The list of variants is small/finite and bounded by real
hardware.
This patch introduces the E5M2 FP8 format as proposed by Nvidia, ARM,
and Intel in the paper: https://arxiv.org/pdf/2209.05433.pdf
As the more conformant of the two implemented datatypes, we are plumbing
it through LLVM's APFloat type and MLIR's type system first as a
template. It will be followed by the range optimized E4M3 FP8 format
described in the paper. Since that format deviates further from the
IEEE-754 norms, it may require more debate and implementation
complexity.
Given that we see two parts of the FP8 implementation space represented
by these cases, we are recommending naming of:
* `F8M<N>` : For FP8 types that can be conceived of as following the
same rules as FP16 but with a smaller number of mantissa/exponent
bits. Including the number of mantissa bits in the type name is enough
to fully specify the type. This naming scheme is used to represent
the E5M2 type described in the paper.
* `F8M<N>F` : For FP8 types such as E4M3 which only support finite
values.
The first of these (this patch) seems fairly non-controversial. The
second is previewed here to illustrate options for extending to the
other known variant (but can be discussed in detail in the patch
which implements it).
Many conversations about these types focus on the Machine-Learning
ecosystem where they are used to represent mixed-datatype computations
at a high level. At that level (which is why we also expose them in
MLIR), it is important to retain the actual type definition so that when
lowering to actual kernels or target specific code, the correct
promotions, casts and rescalings can be done as needed. We expect that
most LLVM backends will only experience these types as opaque `I8`
values that are applicable to some instructions.
MLIR does not make it particularly easy to add new floating point types
(i.e. the FloatType hierarchy is not open). Given the need to fully
model FloatTypes and make them interop with tooling, such types will
always be "heavy-weight" and it is not expected that a highly open type
system will be particularly helpful. There are also a bounded number of
floating point types in use for current and upcoming hardware, and we
can just implement them like this (perhaps looking for some cosmetic
ways to reduce the number of places that need to change). Creating a
more generic mechanism for extending floating point types seems like it
wouldn't be worth it and we should just deal with defining them one by
one on an as-needed basis when real hardware implements a new scheme.
Hopefully, with some additional production use and complete software
stacks, hardware makers will converge on a set of such types that is not
terribly divergent at the level that the compiler cares about.
(I cleaned up some old formatting and sorted some items for this case:
If we converge on landing this in some form, I will NFC commit format
only changes as a separate commit)
Differential Revision: https://reviews.llvm.org/D133823
Dynamic dialects are dialects that can be defined at runtime.
Dynamic dialects are extensible by new operations, types, and
attributes at runtime.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D125201
This patch removes the `type` field from `Attribute` along with the
`Attribute::getType` accessor.
Going forward, this means that attributes in MLIR will no longer have
types as a first-class concept. This patch lays the groundwork to
incrementally remove or refactor code that relies on generic attributes
being typed. The immediate impact will be on attributes that rely on
`Attribute` containing a type, such as `IntegerAttr`,
`DenseElementsAttr`, and `ml_program::ExternAttr`, which will now need
to define a type parameter on their storage classes. This will save
memory as all other attribute kinds will no longer contain a type.
Moreover, it will not be possible to generically query the type of an
attribute directly. This patch provides an attribute interface
`TypedAttr` that implements only one method, `getType`, which can be
used to generically query the types of attributes that implement the
interface. This interface can be used to retain the concept of a "typed
attribute". The ODS-generated accessor for a `type` parameter
automatically implements this method.
Next steps will be to refactor the assembly formats of certain operations
that rely on `parseAttribute(type)` and `printAttributeWithoutType` to
remove special handling of type elision until `type` can be removed from
the dialect parsing hook entirely; and incrementally remove uses of
`TypedAttr`.
Reviewed By: lattner, rriddle, jpienaar
Differential Revision: https://reviews.llvm.org/D130092
Previously default attributes were only usable by way of the ODS generated
accessors, but this was undesirable as
1. The ODS getters could construct Attribute each get request;
2. For non-C++ uses this would require either duplicating some of tee default
attribute generating or generating additional bindings to generate methods;
3. Accessing op.getAttr("foo") and op.getFoo() would return different results;
Generate method to populate default attributes that can be used to address
these.
This merely adds this facility but does not employ by default on any path.
Differential Revision: https://reviews.llvm.org/D128962
This helps to prevent tsan failures when users inadvertantly mutate the
context in a non-safe way.
Differential Revision: https://reviews.llvm.org/D112021
The current dialect registry allows for attaching delayed interfaces, that are added to attrs/dialects/ops/etc.
when the owning dialect gets loaded. This is clunky for quite a few reasons, e.g. each interface type has a
separate tracking structure, and is also quite limiting. This commit refactors this delayed mutation of
dialect constructs into a more general DialectExtension mechanism. This mechanism is essentially a registration
callback that is invoked when a set of dialects have been loaded. This allows for attaching interfaces directly
on the loaded constructs, and also allows for loading new dependent dialects. The latter of which is
extremely useful as it will now enable dependent dialects to only apply in the contexts in which they
are necessary. For example, a dialect dependency can now be conditional on if a user actually needs the
interface that relies on it.
Differential Revision: https://reviews.llvm.org/D120367
This change gives explicit order of verifier execution and adds
`hasRegionVerifier` and `verifyWithRegions` to increase the granularity
of verifier classification. The orders are as below,
1. InternalOpTrait will be verified first, they can be run independently.
2. `verifyInvariants` which is constructed by ODS, it verifies the type,
attributes, .etc.
3. Other Traits/Interfaces that have marked their verifier as
`verifyTrait` or `verifyWithRegions=0`.
4. Custom verifier which is defined in the op and has marked
`hasVerifier=1`
If an operation has regions, then it may have the second phase,
5. Traits/Interfaces that have marked their verifier as
`verifyRegionTrait` or
`verifyWithRegions=1`. This implies the verifier needs to access the
operations in its regions.
6. Custom verifier which is defined in the op and has marked
`hasRegionVerifier=1`
Note that the second phase will be run after the operations in the
region are verified. Based on the verification order, you will be able to
avoid verifying duplicate things.
Reviewed By: Mogball
Differential Revision: https://reviews.llvm.org/D116789
I see a lot of array sorting in stack traces of our compiler, canonicalizer traverses this list every time it builds a pattern set, and it gets expensive very quickly.
Reviewed By: rriddle, mehdi_amini
Differential Revision: https://reviews.llvm.org/D118937
When constructing an OperationName, the overwhelming majority of
cases are from registered operations. This revision adds a non-locked
lookup into the currently registered operations, which prevents locking
in the common case. This revision also optimizes several uses of
RegisteredOperationName that expect the operation to be registered,
e.g. such as in OpBuilder.
These changes provides a reasonable speedup (5-10%) in some
compilations, especially on platforms where locking is expensive.
Differential Revision: https://reviews.llvm.org/D117187
Querying threads directly from the thread pool fails if there is no thread pool or if multithreading is not enabled. Returns 1 by default.
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D116259
