MLIR

Multi-Level IR Compiler Framework

MLIR Python Bindings

Current status: Under development and not enabled by default

Building 

Pre-requisites 

  • A relatively recent Python3 installation
  • Installation of python dependencies as specified in mlir/python/requirements.txt

CMake variables 

  • MLIR_ENABLE_BINDINGS_PYTHON:BOOL

    Enables building the Python bindings. Defaults to OFF.

  • Python3_EXECUTABLE:STRING

    Specifies the python executable used for the LLVM build, including for determining header/link flags for the Python bindings. On systems with multiple Python implementations, setting this explicitly to the preferred python3 executable is strongly recommended.

  • MLIR_BINDINGS_PYTHON_LOCK_VERSION:BOOL

    Links the native extension against the Python runtime library, which is optional on some platforms. While setting this to OFF can yield some greater deployment flexibility, linking in this way allows the linker to report compile time errors for unresolved symbols on all platforms, which makes for a smoother development workflow. Defaults to ON.

It is recommended to use a python virtual environment. Many ways exist for this, but the following is the simplest:

# Make sure your 'python' is what you expect. Note that on multi-python
# systems, this may have a version suffix, and on many Linuxes and MacOS where
# python2 and python3 co-exist, you may also want to use `python3`.
which python
python -m venv ~/.venv/mlirdev
source ~/.venv/mlirdev/bin/activate

# Note that many LTS distros will bundle a version of pip itself that is too
# old to download all of the latest binaries for certain platforms.
# The pip version can be obtained with `python -m pip --version`, and for
# Linux specifically, this should be cross checked with minimum versions
# here: https://github.com/pypa/manylinux
# It is recommended to upgrade pip:
python -m pip install --upgrade pip


# Now the `python` command will resolve to your virtual environment and
# packages will be installed there.
python -m pip install -r mlir/python/requirements.txt

# Now run `cmake`, `ninja`, et al.

For interactive use, it is sufficient to add the tools/mlir/python_packages/mlir_core/ directory in your build/ directory to the PYTHONPATH. Typically:

export PYTHONPATH=$(cd build && pwd)/tools/mlir/python_packages/mlir_core

Note that if you have installed (i.e. via ninja install, et al), then python packages for all enabled projects will be in your install tree under python_packages/ (i.e. python_packages/mlir_core). Official distributions are built with a more specialized setup.

Design 

Use cases 

There are likely two primary use cases for the MLIR python bindings:

  1. Support users who expect that an installed version of LLVM/MLIR will yield the ability to import mlir and use the API in a pure way out of the box.

  2. Downstream integrations will likely want to include parts of the API in their private namespace or specially built libraries, probably mixing it with other python native bits.

Composable modules 

In order to support use case #2, the Python bindings are organized into composable modules that downstream integrators can include and re-export into their own namespace if desired. This forces several design points:

  • Separate the construction/populating of a py::module from PYBIND11_MODULE global constructor.

  • Introduce headers for C++-only wrapper classes as other related C++ modules will need to interop with it.

  • Separate any initialization routines that depend on optional components into its own module/dependency (currently, things like registerAllDialects fall into this category).

There are a lot of co-related issues of shared library linkage, distribution concerns, etc that affect such things. Organizing the code into composable modules (versus a monolithic cpp file) allows the flexibility to address many of these as needed over time. Also, compilation time for all of the template meta-programming in pybind scales with the number of things you define in a translation unit. Breaking into multiple translation units can significantly aid compile times for APIs with a large surface area.

Submodules 

Generally, the C++ codebase namespaces most things into the mlir namespace. However, in order to modularize and make the Python bindings easier to understand, sub-packages are defined that map roughly to the directory structure of functional units in MLIR.

Examples:

  • mlir.ir
  • mlir.passes (pass is a reserved word :( )
  • mlir.dialect
  • mlir.execution_engine (aside from namespacing, it is important that “bulky”/optional parts like this are isolated)

In addition, initialization functions that imply optional dependencies should be in underscored (notionally private) modules such as _init and linked separately. This allows downstream integrators to completely customize what is included “in the box” and covers things like dialect registration, pass registration, etc.

Loader 

LLVM/MLIR is a non-trivial python-native project that is likely to co-exist with other non-trivial native extensions. As such, the native extension (i.e. the .so/.pyd/.dylib) is exported as a notionally private top-level symbol (_mlir), while a small set of Python code is provided in mlir/_cext_loader.py and siblings which loads and re-exports it. This split provides a place to stage code that needs to prepare the environment before the shared library is loaded into the Python runtime, and also provides a place that one-time initialization code can be invoked apart from module constructors.

It is recommended to avoid using __init__.py files to the extent possible, until reaching a leaf package that represents a discrete component. The rule to keep in mind is that the presence of an __init__.py file prevents the ability to split anything at that level or below in the namespace into different directories, deployment packages, wheels, etc.

See the documentation for more information and advice: https://packaging.python.org/guides/packaging-namespace-packages/

Use the C-API 

The Python APIs should seek to layer on top of the C-API to the degree possible. Especially for the core, dialect-independent parts, such a binding enables packaging decisions that would be difficult or impossible if spanning a C++ ABI boundary. In addition, factoring in this way side-steps some very difficult issues that arise when combining RTTI-based modules (which pybind derived things are) with non-RTTI polymorphic C++ code (the default compilation mode of LLVM).

Ownership in the Core IR 

There are several top-level types in the core IR that are strongly owned by their python-side reference:

  • PyContext (mlir.ir.Context)
  • PyModule (mlir.ir.Module)
  • PyOperation (mlir.ir.Operation) - but with caveats

All other objects are dependent. All objects maintain a back-reference (keep-alive) to their closest containing top-level object. Further, dependent objects fall into two categories: a) uniqued (which live for the life-time of the context) and b) mutable. Mutable objects need additional machinery for keeping track of when the C++ instance that backs their Python object is no longer valid (typically due to some specific mutation of the IR, deletion, or bulk operation).

Optionality and argument ordering in the Core IR 

The following types support being bound to the current thread as a context manager:

  • PyLocation (loc: mlir.ir.Location = None)
  • PyInsertionPoint (ip: mlir.ir.InsertionPoint = None)
  • PyMlirContext (context: mlir.ir.Context = None)

In order to support composability of function arguments, when these types appear as arguments, they should always be the last and appear in the above order and with the given names (which is generally the order in which they are expected to need to be expressed explicitly in special cases) as necessary. Each should carry a default value of py::none() and use either a manual or automatic conversion for resolving either with the explicit value or a value from the thread context manager (i.e. DefaultingPyMlirContext or DefaultingPyLocation).

The rationale for this is that in Python, trailing keyword arguments to the right are the most composable, enabling a variety of strategies such as kwarg passthrough, default values, etc. Keeping function signatures composable increases the chances that interesting DSLs and higher level APIs can be constructed without a lot of exotic boilerplate.

Used consistently, this enables a style of IR construction that rarely needs to use explicit contexts, locations, or insertion points but is free to do so when extra control is needed.

Operation hierarchy 

As mentioned above, PyOperation is special because it can exist in either a top-level or dependent state. The life-cycle is unidirectional: operations can be created detached (top-level) and once added to another operation, they are then dependent for the remainder of their lifetime. The situation is more complicated when considering construction scenarios where an operation is added to a transitive parent that is still detached, necessitating further accounting at such transition points (i.e. all such added children are initially added to the IR with a parent of their outer-most detached operation, but then once it is added to an attached operation, they need to be re-parented to the containing module).

Due to the validity and parenting accounting needs, PyOperation is the owner for regions and blocks and needs to be a top-level type that we can count on not aliasing. This let’s us do things like selectively invalidating instances when mutations occur without worrying that there is some alias to the same operation in the hierarchy. Operations are also the only entity that are allowed to be in a detached state, and they are interned at the context level so that there is never more than one Python mlir.ir.Operation object for a unique MlirOperation, regardless of how it is obtained.

The C/C++ API allows for Region/Block to also be detached, but it simplifies the ownership model a lot to eliminate that possibility in this API, allowing the Region/Block to be completely dependent on its owning operation for accounting. The aliasing of Python Region/Block instances to underlying MlirRegion/MlirBlock is considered benign and these objects are not interned in the context (unlike operations).

If we ever want to re-introduce detached regions/blocks, we could do so with new “DetachedRegion” class or similar and also avoid the complexity of accounting. With the way it is now, we can avoid having a global live list for regions and blocks. We may end up needing an op-local one at some point TBD, depending on how hard it is to guarantee how mutations interact with their Python peer objects. We can cross that bridge easily when we get there.

Module, when used purely from the Python API, can’t alias anyway, so we can use it as a top-level ref type without a live-list for interning. If the API ever changes such that this cannot be guaranteed (i.e. by letting you marshal a native-defined Module in), then there would need to be a live table for it too.

Style 

In general, for the core parts of MLIR, the Python bindings should be largely isomorphic with the underlying C++ structures. However, concessions are made either for practicality or to give the resulting library an appropriately “Pythonic” flavor.

Properties vs get*() methods 

Generally favor converting trivial methods like getContext(), getName(), isEntryBlock(), etc to read-only Python properties (i.e. context). It is primarily a matter of calling def_property_readonly vs def in binding code, and makes things feel much nicer to the Python side.

For example, prefer:

m.def_property_readonly("context", ...)

Over:

m.def("getContext", ...)

repr methods 

Things that have nice printed representations are really great :) If there is a reasonable printed form, it can be a significant productivity boost to wire that to the __repr__ method (and verify it with a doctest ).

CamelCase vs snake_case 

Name functions/methods/properties in snake_case and classes in CamelCase. As a mechanical concession to Python style, this can go a long way to making the API feel like it fits in with its peers in the Python landscape.

If in doubt, choose names that will flow properly with other PEP 8 style names .

Prefer pseudo-containers 

Many core IR constructs provide methods directly on the instance to query count and begin/end iterators. Prefer hoisting these to dedicated pseudo containers.

For example, a direct mapping of blocks within regions could be done this way:

region = ...

for block in region:

  pass

However, this way is preferred:

region = ...

for block in region.blocks:

  pass

print(len(region.blocks))
print(region.blocks[0])
print(region.blocks[-1])

Instead of leaking STL-derived identifiers (front, back, etc), translate them to appropriate __dunder__ methods and iterator wrappers in the bindings.

Note that this can be taken too far, so use good judgment. For example, block arguments may appear container-like but have defined methods for lookup and mutation that would be hard to model properly without making semantics complicated. If running into these, just mirror the C/C++ API.

Provide one stop helpers for common things 

One stop helpers that aggregate over multiple low level entities can be incredibly helpful and are encouraged within reason. For example, making Context have a parse_asm or equivalent that avoids needing to explicitly construct a SourceMgr can be quite nice. One stop helpers do not have to be mutually exclusive with a more complete mapping of the backing constructs.

Testing 

Tests should be added in the test/Bindings/Python directory and should typically be .py files that have a lit run line.

We use lit and FileCheck based tests:

  • For generative tests (those that produce IR), define a Python module that constructs/prints the IR and pipe it through FileCheck.
  • Parsing should be kept self-contained within the module under test by use of raw constants and an appropriate parse_asm call.
  • Any file I/O code should be staged through a tempfile vs relying on file artifacts/paths outside of the test module.
  • For convenience, we also test non-generative API interactions with the same mechanisms, printing and CHECKing as needed.

Sample FileCheck test 

# RUN: %PYTHON %s | mlir-opt -split-input-file | FileCheck

# TODO: Move to a test utility class once any of this actually exists.
def print_module(f):
  m = f()
  print("// -----")
  print("// TEST_FUNCTION:", f.__name__)
  print(m.to_asm())
  return f

# CHECK-LABEL: TEST_FUNCTION: create_my_op
@print_module
def create_my_op():
  m = mlir.ir.Module()
  builder = m.new_op_builder()
  # CHECK: mydialect.my_operation ...
  builder.my_op()
  return m

Integration with ODS 

The MLIR Python bindings integrate with the tablegen-based ODS system for providing user-friendly wrappers around MLIR dialects and operations. There are multiple parts to this integration, outlined below. Most details have been elided: refer to the build rules and python sources under mlir.dialects for the canonical way to use this facility.

Users are responsible for providing a {DIALECT_NAMESPACE}.py (or an equivalent directory with __init__.py file) as the entrypoint.

Generating _{DIALECT_NAMESPACE}_ops_gen.py wrapper modules 

Each dialect with a mapping to python requires that an appropriate _{DIALECT_NAMESPACE}_ops_gen.py wrapper module is created. This is done by invoking mlir-tblgen on a python-bindings specific tablegen wrapper that includes the boilerplate and actual dialect specific td file. An example, for the StandardOps (which is assigned the namespace std as a special case):

#ifndef PYTHON_BINDINGS_STANDARD_OPS
#define PYTHON_BINDINGS_STANDARD_OPS

include "mlir/Bindings/Python/Attributes.td"
include "mlir/Dialect/StandardOps/IR/Ops.td"

#endif

In the main repository, building the wrapper is done via the CMake function add_mlir_dialect_python_bindings, which invokes:

mlir-tblgen -gen-python-op-bindings -bind-dialect={DIALECT_NAMESPACE} \
    {PYTHON_BINDING_TD_FILE}

The generates op classes must be included in the {DIALECT_NAMESPACE}.py file in a similar way that generated headers are included for C++ generated code:

from ._my_dialect_ops_gen import *

Extending the search path for wrapper modules 

When the python bindings need to locate a wrapper module, they consult the dialect_search_path and use it to find an appropriately named module. For the main repository, this search path is hard-coded to include the mlir.dialects module, which is where wrappers are emitted by the abobe build rule. Out of tree dialects and add their modules to the search path by calling:

mlir._cext.append_dialect_search_prefix("myproject.mlir.dialects")

Wrapper module code organization 

The wrapper module tablegen emitter outputs:

  • A _Dialect class (extending mlir.ir.Dialect) with a DIALECT_NAMESPACE attribute.
  • An {OpName} class for each operation (extending mlir.ir.OpView).
  • Decorators for each of the above to register with the system.

Note: In order to avoid naming conflicts, all internal names used by the wrapper module are prefixed by _ods_.

Each concrete OpView subclass further defines several public-intended attributes:

  • OPERATION_NAME attribute with the str fully qualified operation name (i.e. std.absf).
  • An __init__ method for the default builder if one is defined or inferred for the operation.
  • @property getter for each operand or result (using an auto-generated name for unnamed of each).
  • @property getter, setter and deleter for each declared attribute.

It further emits additional private-intended attributes meant for subclassing and customization (default cases omit these attributes in favor of the defaults on OpView):

  • _ODS_REGIONS: A specification on the number and types of regions. Currently a tuple of (min_region_count, has_no_variadic_regions). Note that the API does some light validation on this but the primary purpose is to capture sufficient information to perform other default building and region accessor generation.
  • _ODS_OPERAND_SEGMENTS and _ODS_RESULT_SEGMENTS: Black-box value which indicates the structure of either the operand or results with respect to variadics. Used by OpView._ods_build_default to decode operand and result lists that contain lists.

Default Builder 

Presently, only a single, default builder is mapped to the __init__ method. The intent is that this __init__ method represents the most specific of the builders typically generated for C++; however currently it is just the generic form below.

  • One argument for each declared result:
    • For single-valued results: Each will accept an mlir.ir.Type.
    • For variadic results: Each will accept a List[mlir.ir.Type].
  • One argument for each declared operand or attribute:
    • For single-valued operands: Each will accept an mlir.ir.Value.
    • For variadic operands: Each will accept a List[mlir.ir.Value].
    • For attributes, it will accept an mlir.ir.Attribute.
  • Trailing usage-specific, optional keyword arguments:
    • loc: An explicit mlir.ir.Location to use. Defaults to the location bound to the thread (i.e. with Location.unknown():) or an error if none is bound nor specified.
    • ip: An explicit mlir.ir.InsertionPoint to use. Default to the insertion point bound to the thread (i.e. with InsertionPoint(...):).

In addition, each OpView inherits a build_generic method which allows construction via a (nested in the case of variadic) sequence of results and operands. This can be used to get some default construction semantics for operations that are otherwise unsupported in Python, at the expense of having a very generic signature.

Extending Generated Op Classes 

Note that this is a rather complex mechanism and this section errs on the side of explicitness. Users are encouraged to find an example and duplicate it if they don’t feel the need to understand the subtlety. The builtin dialect provides some relatively simple examples.

As mentioned above, the build system generates Python sources like _{DIALECT_NAMESPACE}_ops_gen.py for each dialect with Python bindings. It is often desirable to to use these generated classes as a starting point for further customization, so an extension mechanism is provided to make this easy (you are always free to do ad-hoc patching in your {DIALECT_NAMESPACE}.py file but we prefer a more standard mechanism that is applied uniformly).

To provide extensions, add a _{DIALECT_NAMESPACE}_ops_ext.py file to the dialects module (i.e. adjacent to your {DIALECT_NAMESPACE}.py top-level and the *_ops_gen.py file). Using the builtin dialect and FuncOp as an example, the generated code will include an import like this:

try:
  from . import _builtin_ops_ext as _ods_ext_module
except ImportError:
  _ods_ext_module = None

Then for each generated concrete OpView subclass, it will apply a decorator like:

@_ods_cext.register_operation(_Dialect)
@_ods_extend_opview_class(_ods_ext_module)
class FuncOp(_ods_ir.OpView):

See the _ods_common.py extend_opview_class function for details of the mechanism. At a high level:

  • If the extension module exists, locate an extension class for the op (in this example, FuncOp):
    • First by looking for an attribute with the exact name in the extension module.
    • Falling back to calling a select_opview_mixin(parent_opview_cls) function defined in the extension module.
  • If a mixin class is found, a new subclass is dynamically created that multiply inherits from ({_builtin_ops_ext.FuncOp}, _builtin_ops_gen.FuncOp).

The mixin class should not inherit from anything (i.e. directly extends object only). The facility is typically used to define custom __init__ methods, properties, instance methods and static methods. Due to the inheritance ordering, the mixin class can act as though it extends the generated OpView subclass in most contexts (i.e. issubclass(_builtin_ops_ext.FuncOp, OpView) will return False but usage generally allows you treat it as duck typed as an OpView).

There are a couple of recommendations, given how the class hierarchy is defined:

  • For static methods that need to instantiate the actual “leaf” op (which is dynamically generated and would result in circular dependencies to try to reference by name), prefer to use @classmethod and the concrete subclass will be provided as your first cls argument. See _builtin_ops_ext.FuncOp.from_py_func as an example.
  • If seeking to replace the generated __init__ method entirely, you may actually want to invoke the super-super-class mlir.ir.OpView constructor directly, as it takes an mlir.ir.Operation, which is likely what you are constructing (i.e. the generated __init__ method likely adds more API constraints than you want to expose in a custom builder).

A pattern that comes up frequently is wanting to provide a sugared __init__ method which has optional or type-polymorphism/implicit conversions but to otherwise want to invoke the default op building logic. For such cases, it is recommended to use an idiom such as:

  def __init__(self, sugar, spice, *, loc=None, ip=None):
    ... massage into result_type, operands, attributes ...
    OpView.__init__(self, self.build_generic(
        results=[result_type],
        operands=operands,
        attributes=attributes,
        loc=loc,
        ip=ip))

Refer to the documentation for build_generic for more information.