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 removes the implementation of TypedAttr and ElementsAttr
from DenseArrayAttr and, in doing so, removes the need store a shaped
type. The attribute now stores a size (number of elements), an MLIR type
as a discriminator, and a raw byte array.
The intent of DenseArrayAttr was not to be a drop-in replacement for DenseElementsAttr. It was meant to be a simple container of integers or floats that map to C++ types. The ElementsAttr implementation on DenseArrayAttr had many holes in it, and fixing those holes would require evolving DenseArrayAttr in a way that is incompatible with its original purpose.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D137606
Calculating the position of the region trailing objects isn't free,
given that it's the last trailing object, and inlining the size check
removes the need for users to explicitly add size checks for
micro-optimization.
This commit refactors attribute/type alias generation to be similar to how
we do it for operations, i.e. we generate aliases determined on what is
actually necessary when printing the IR (using a dummy printer for alias
collection). This allows for generating aliases only when necessary, and
also allows for proper propagation of when a nested alias can be deferred.
This also necessitated a fix for location parsing to actually parse aliases
instead of ignoring them.
Fixes#59041
Differential Revision: https://reviews.llvm.org/D138886
This adds an `enable` flag to OpPrintingFlags::enableDebugInfo
that allows for overriding any command line flags for debug printing,
and matches the format that we use for other `enableBlah` API.
We properly order dependencies between attribute/type aliases,
but we currently always print attribute aliases separately from type
aliases. This creates problems if an attribute wants to use a type
alias during printing.
This commit refactors alias collection such that attribute/type aliases
are collected together and printed together.
Differential Revision: https://reviews.llvm.org/D138162
We currently only support one level of aliases, which isn't great
in situations where an attribute/type can have multiple duplicated
components nested within it(e.g. debuginfo metadata). This commit
refactors alias generation to support nested aliases, which requires
changing alias grouping to take into account the depth of child
aliases, to ensure that attributes/types aren't printed before the
aliases they use.
The only real user facing change here was that we no longer print
0 as an alias suffix, which would be unnecessarily expensive to keep
in the new alias generation method (and isn't that valuable of a
behavior to preserve).
Differential Revision: https://reviews.llvm.org/D136541
This allows for using the llvm namespace cast methods instead of the ones on the Value class. The Value class method are kept for now, but we'll want to remove these eventually (with a really long lead time).
Related change: https://reviews.llvm.org/D134327
Differential Revision: https://reviews.llvm.org/D135870
This patch moves the 'printOp' functionality to the public API of
AsmPrinter and rename it to 'printCustomOrGenericOp'. No 'parseOp'
is needed at this time as existing APIs are able to parse operations
producing results where results are omitted in the textual form
(the LHS of an operation is redundant when it comes to building the
operation itself as it only contains the result names).
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D135006
(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
This is the corresponding method to
`OpAsmParser::parseOptionalLocationSpecifier` that prints a location
`loc(...)` based on the op printing flags. Together, these two functions
allow propagating user-level location info outside of their usual spots.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D134910
The three following ops in the memref dialect: transpose, expand_shape,
collapse_shape, have been originally designed to operate on memrefs with
strided layouts but had to go through the affine map representation as the type
did not support anything else. Make these ops produce memref values with
StridedLayoutAttr instead now that it is available.
Depends On D133938
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D133947
This patch adds a flag to `Attribute::print` that prints the attribute
without its type.
Fixes#57689
Reviewed By: rriddle, lattner
Differential Revision: https://reviews.llvm.org/D133822
This is necessary/useful for building generic tooling that can roundtrip external
resources without needing to explicitly handle them. For example, this allows
for viewing the resources encoded within a bytecode file without having to
explicitly know how to process them (e.g. making it easier to interact with a
reproducer encoded in bytecode).
Differential Revision: https://reviews.llvm.org/D133460
Resources are encoded in two separate sections similarly to
attributes/types, one for the actual data and one for the data
offsets. Unlike other sections, the resource sections are optional
given that in many cases they won't be present. For testing,
bytecode serialization is added for DenseResourceElementsAttr.
Differential Revision: https://reviews.llvm.org/D132729
This allows for extracting assembly information when printing an attribute
or type, such as the dialect resources referenced. This functionality is used in
a followup that adds resource support to the bytecode. This change also results
in a nice cleanup of AsmPrinter now that we don't need to awkwardly workaround
optional AsmStates.
Differential Revision: https://reviews.llvm.org/D132728
This patch makes parsing dense arrays with type elision work properly.
If a ranked tensor type is supplied to `parseAttribute` on a dense
array, the element type is skipped. Moreover, if type elision is set to
`AttrTypeElision::Must`, the element type is elided.
For example, this allows
```
memref.global @z : memref<3xi32> = array<1, 2, 3>
```
Fixes#57433
Depends on D132758
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D132964
This patch turns `DenseArrayBaseAttr` into a fully-functional attribute by
adding a generic parser and printer, supporting bool or integer and floating
point element types with bitwidths divisible by 8. It has been renamed
to `DenseArrayAttr`. The patch maintains the specialized subclasses,
e.g. `DenseI32ArrayAttr`, which remain the preferred API for accessing
elements in C++.
This allows `DenseArrayAttr` to hold signed and unsigned integer elements:
```
array<si8: -128, 127>
array<ui8: 255>
```
"Exotic" floating point elements:
```
array<bf16: 1.2, 3.4>
```
And integers of other bitwidths:
```
array<i24: 8388607>
```
Reviewed By: rriddle, lattner
Differential Revision: https://reviews.llvm.org/D132758
Introduce a new attribute to represent the strided memref layout. Strided
layouts are omnipresent in code generation flows and are the only kind of
layouts produced and supported by a half of operation in the memref dialect
(view-related, shape-related). However, they are internally represented as
affine maps that require a somewhat fragile extraction of the strides from the
linear form that also comes with an overhead. Furthermore, textual
representation of strided layouts as affine maps is difficult to read: compare
`affine_map<(d0, d1, d2)[s0, s1] -> (d0*32 + d1*s0 + s1 + d2)>` with
`strides: [32, ?, 1], offset: ?`. While a rudimentary support for parsing a
syntactically sugared version of the strided layout has existed in the codebase
for a long time, it does not go as far as this commit to make the strided
layout a first-class attribute in the IR.
This introduces the attribute and updates the tests that using the pre-existing
sugared form to use the new attribute instead. Most memref created
programmatically, e.g., in passes, still use the affine form with further
extraction of strides and will be updated separately.
Update and clean-up the memref type documentation that has gotten stale and has
been referring to the details of affine map composition that are long gone.
See https://discourse.llvm.org/t/rfc-materialize-strided-memref-layout-as-an-attribute/64211.
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D132864
Follow-up to D123774, where the syntax of dense arrays was discussed. It
was included that the syntax should be changed to `array<i32: 1, 2>`.
This patch changes the syntax but importantly preserves the `[1, 2]`
syntax when embedding these attributes in assembly formats through ODS.
Reviewed By: mehdi_amini, jpienaar
Differential Revision: https://reviews.llvm.org/D131738
This patch adds a DenseI1ArrayAttr to support arrays of i1. Importantly,
the implementation is as a simple `ArrayRef<bool>` instead of using bit
compression, which was problematic in DenseElementsAttr.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D130957
Previously, DenseArrayAttr used VectorType for its shaped type.
VectorType is problematic for arrays because it doesn't support zero
dimensions, meaning that an empty array would have `vector<i32>` as its
type. ElementsAttr would think that an empty dense array is size 1, not
0. This patch switches over to TensorType, which does support zero
dimensions.
Fixes#56860
Reviewed By: mehdi_amini
Differential Revision: https://reviews.llvm.org/D130921
This attribute is technical debt from the early stages of MLIR, before
ElementsAttr was an interface and when it was more difficult for
dialects to define their own types of attributes. At present it isn't
used at all in tree (aside from being convenient for eliding other
ElementsAttr), and has had little to no evolution in the past three years.
Differential Revision: https://reviews.llvm.org/D129917
This attributes is intended cover the current set of use cases that abuse
DenseElementsAttr, e.g. when the data is large. Using resources for large
data is one of the major reasons why they were added; e.g. they can be
deallocated mid-compilation, they support a wide variety of data origins
(e.g, heap allocated, mmap'd, etc.), they can support mutation, etc.
I considered at length not having a builtin variant of this, and instead
having multiple versions of this attribute for dialects that are interested,
but they all boiled down to the exact same attribute definition. Given the
generality of this attribute, it feels more aligned to keep it next to DenseArrayAttr
(given that DenseArrayAttr covers the "small" case, and DenseResourcesElementsAttr
covers the "large" case). The underlying infra used to build this attribute is
general, and having a builtin attribute doesn't preclude users from defining
their own when it makes sense (they can even share a blob manager with the
builtin dialect to avoid data duplication).
Differential Revision: https://reviews.llvm.org/D130022
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
https://reviews.llvm.org/D126446 added support for encoding
binary blobs in MLIR assembly. To enable cross-architecture
compatibility, these need to be encoded in little-endian format.
This patch is a first step in that direction by reading and
writing the alignment tag that those blobs are prefixed by
in little-endian format. This fixes assertion failures in
several test cases on big-endian platforms.
The actual content of the blob is not yet handled here.
Differential Revision: https://reviews.llvm.org/D129483
This commit refactors the syntax of "ugly" attribute/type formats to not use
strings for wrapping. This means that moving forward attirbutes and type formats
will always need to be in some recognizable form, i.e. if they use incompatible
characters they will need to manually wrap those in a string, the framework will
no longer do it automatically.
This has the benefit of greatly simplifying how parsing attributes/types work, given
that we currently rely on some extremely complicated nested parser logic which is
quite problematic for a myriad of reasons; unecessary complexity(we create a nested
source manager/lexer/etc.), diagnostic locations can be off/wrong given string escaping,
etc.
Differential Revision: https://reviews.llvm.org/D118505
This commit enables support for providing and processing external
resources within MLIR assembly formats. This is a mechanism with which
dialects, and external clients, may attach additional information when
printing IR without that information being encoded in the IR itself.
External resources are not uniqued within the MLIR context, are not
attached directly to any operation, and are solely intended to live and be
processed outside of the immediate IR. There are many potential uses of this
functionality, for example MLIR's pass crash reproducer could utilize this to
attach the pass resource executing when a crash occurs. Other types of
uses may be embedding large amounts of binary data, such as weights in ML
applications, that shouldn't be copied directly into the MLIR context, but
need to be kept adjacent to the IR.
External resources are encoded using a key-value pair nested within a
dictionary anchored by name either on a dialect, or an externally registered
entity. The key is an identifier used to disambiguate the data. The value
may be stored in various limited forms, but general encodings use a string
(human readable) or blob format (binary). Within the textual format, an
example may be of the form:
```mlir
{-#
// The `dialect_resources` section within the file-level metadata
// dictionary is used to contain any dialect resource entries.
dialect_resources: {
// Here is a dictionary anchored on "foo_dialect", which is a dialect
// namespace.
foo_dialect: {
// `some_dialect_resource` is a key to be interpreted by the dialect,
// and used to initialize/configure/etc.
some_dialect_resource: "Some important resource value"
}
},
// The `external_resources` section within the file-level metadata
// dictionary is used to contain any non-dialect resource entries.
external_resources: {
// Here is a dictionary anchored on "mlir_reproducer", which is an
// external entity representing MLIR's crash reproducer functionality.
mlir_reproducer: {
// `pipeline` is an entry that holds a crash reproducer pipeline
// resource.
pipeline: "func.func(canonicalize,cse)"
}
}
```
Differential Revision: https://reviews.llvm.org/D126446
This attribute is similar to DenseElementsAttr but does not support
splat. As such it has a much simpler API and does not need any smart
iterator: it exposes direct ArrayRef access.
A new syntax is introduced so that the generic printing/parsing looks
like:
[:i64 1, -2, 3]
This attribute beings like an ArrayAttr but has a `:` token after the
opening square brace to introduce the element type (supported are I8,
I16, I32, I64, F32, F64) and the comma separated list for the data.
This is particularly convenient for attributes intended to be small,
like those referring to shapes.
For example a `transpose` operation with a `dims` attribute could be
defined as such:
let arguments = (ins AnyTensor:$input, DenseI64ArrayAttr:$dims);
let assemblyFormat = "$input `dims` `=` $dims attr-dict : type($input)";
And printed this way (the element type is elided in this case):
transpose %input dims = [0, 2, 1] : tensor<2x3x4xf32>
The C++ API for dims would just directly return an ArrayRef<int64>
RFC: https://discourse.llvm.org/t/rfc-introduce-a-new-dense-array-attribute/63279
Recommit with a custom DenseArrayBaseAttrStorage class to ensure
over-alignment of the storage to the largest type.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D123774
This attribute is similar to DenseElementsAttr but does not support
splat. As such it has a much simpler API and does not need any smart
iterator: it exposes direct ArrayRef access.
A new syntax is introduced so that the generic printing/parsing looks
like:
[:i64 1, -2, 3]
This attribute beings like an ArrayAttr but has a `:` token after the
opening square brace to introduce the element type (supported are I8,
I16, I32, I64, F32, F64) and the comma separated list for the data.
This is particularly convenient for attributes intended to be small,
like those referring to shapes.
For example a `transpose` operation with a `dims` attribute could be
defined as such:
let arguments = (ins AnyTensor:$input, DenseI64ArrayAttr:$dims);
let assemblyFormat = "$input `dims` `=` $dims attr-dict : type($input)";
And printed this way (the element type is elided in this case):
transpose %input dims = [0, 2, 1] : tensor<2x3x4xf32>
The C++ API for dims would just directly return an ArrayRef<int64>
RFC: https://discourse.llvm.org/t/rfc-introduce-a-new-dense-array-attribute/63279
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D123774
For the hypothetical "a.b.c" op printed within a region that declares "a" as
the default dialect, MLIR would currently elide the "a." prefix and only print
"b.c". However, this becomes ambiguous while parsing as "b.c" may be exist as
the "c" op in the "b" dialect. If it does not, the parsing currently fails. Do
not elide the default dialect if the op name contains further dots to avoid the
ambiguity.
See https://discourse.llvm.org/t/dropping-dialect-prefix-for-ops-with-multiple-dots-in-the-name/62562
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D125975
This was carry over from LLVM IR where the alias definition can
be ambiguous, but MLIR type aliases have no such problems.
Having the `type` keyword is superfluous and doesn't add anything.
This commit drops it, which also nicely aligns with the syntax for
attribute aliases (which doesn't have a keyword).
Differential Revision: https://reviews.llvm.org/D125501
