# What
This PR adds a shared library `MLIRPythonSupport` which contains all of
the CRTP classes ike `PyConcreteValue`, `PyConcreteType`,
`PyConcreteAttribute`, as well as other useful code like `Defaulting*`
and etc enabling their reuse in downstream projects. Downstream projects
can now do
```c++
struct PyTestType : mlir::python::MLIR_BINDINGS_PYTHON_DOMAIN::PyConcreteType<PyTestType> {
...
};
class PyTestAttr : public mlir::python::MLIR_BINDINGS_PYTHON_DOMAIN::PyConcreteAttribute<PyTestAttr> {
...
}
NB_MODULE(_mlirPythonTestNanobind, m) {
PyTestType::bind(m);
PyTestAttr::bind(m);
}
```
instead of using the discordant alternative
`mlir_type_subclass`/`mlir_attr_subclass` (same goes for
`PyConcreteValue`/`mlir_value_subclass`).
# Why
This PR is mostly code motion (along with CMake) but before I describe
the changes I want to state the goals/benefits:
1. Currently upstream "core" extensions and "dialect" extensions ([all
of the `Dialect*` extensions
here](d7c734b5a1/mlir/lib/Bindings/Python))
are a two-tier system;
**a**. [core
extensions](https://github.com/llvm/llvm-project/blob/main/mlir/lib/Bindings/Python/IRTypes.cpp#L361)
enjoy first class support as far as type inference[^3], type stub
generation, and ease of implementation, while dialect extensions [have
poorer support](https://reviews.llvm.org/D150927), incorrect type stub
generation much more tedious (boilerplate) implementation;
**b**. Crucially, this two-tiered system is reflected in the fact that
**the two sets of types/attributes are not in the same Python object
hierarchy**. To wit: `isinstance(..., Type)` and `isinstance(...,
Attribute)` are not supported for the dialect extensions[^2];
**c**. Since these types are not exposed in public headers, downstream
users (dialect extensions or not) cannot write functions that overload
on e.g. `PyFloat8*Type` - that's quite a [useful
feature](fdbee98df8/cpp_ext/TorchOps.cpp (L29-L69))!
2. The dialect extensions incur a sizeable performance penalty relative
to the core extensions in that every single trip across the wire (either
`python->cpp` or `cpp->python`) requires work in addition to nanobind's
own casting/construction pipeline;
**a**. When going from `python->cpp`, [we extract the capsule object
from the Python
object](https://github.com/llvm/llvm-project/blob/main/mlir/include/mlir/Bindings/Python/NanobindAdaptors.h#L219C24-L219C46)
and then extract from the capsule the `Mlir*` opaque struct/ptr. This
side isn't so onerous;
**b**. When going from `cpp->python` we call long-hand call Python
`import` APIs and construct the Python object using `_CAPICreate`. Note,
there at least 2 `attr` calls incurred in addition to `_CAPICreate`;
this is already much more [efficiently handled by nanobind
itself](4ba51fcf79/src/nb_internals.h (L381-L382))!
3. This division blocks various features: in some configurations[^1] we
trigger a circular import bug because "dialect" types and attributes
perform an [import of the root `_mlir`
module](bd9651bf78/mlir/include/mlir/Bindings/Python/NanobindAdaptors.h (L585))
when they are created (the types themselves, not even instances of those
types). This blocks type stub generation for dialect extensions (i.e.,
the reason we currently only generate type stubs for `_mlir`).
# How
Prior this was not done/possible because of "ODR" issues but I have
resolved those issues; the basic idea for how we solve this is "move
things we want to share into shared libraries":
1. Move IRCore (stuff like `PyConcreteValue`, `PyConcreteType`,
`PyConcreteAttribute`) into `MLIRPythonSupport`;
- Note, we move the rest of the things in `IRModule.h` (renamed to
`IRCore.h`) because `PyConcreteValue`, `PyConcreteType`,
`PyConcreteAttribute` depend on them. This makes for a bigger PR than
one would hope for but ultimately I think we should give people access
to these classes to use as they see fit (specifically inherit from, but
also liberally use in bindings signatures instead of the opaque `Mlir*`
struct wrappers).
2. Put all of this code into a nested namespace
`MLIR_BINDINGS_PYTHON_DOMAIN` which is determined by a compile time
define (and tied to `MLIR_BINDINGS_PYTHON_NB_DOMAIN`). This is necessary
in order to prevent conflicts on both symbol name **and** typeid
(necessary for nanobind to not double register binded types) between
multiple bindings libraries (e.g., `torch-mlir`, and `jax`). Note
[nanobind doesn't support `module_local` like
pybind11](https://nanobind.readthedocs.io/en/latest/porting.html#removed-features).
It does support `NB_DOMAIN` but that is not sufficient for
disambiguating typeids across projects (to wit: we currently define
`NB_DOMAIN` and it was still necessary to move everything to a nested
namespace);
3. Build the [nanobind library itself as a shared
object](https://github.com/wjakob/nanobind/blob/master/cmake/nanobind-config.cmake#L127)
(and link it to both the extensions and `MLIRPythonSupport`).
4. CMake to make this work, in-tree, out-of-tree, downstream, upstream,
etc.
# Testing
Three tests are added here
1. `PythonTestModuleNanobind` is ported to use
`PyConcreteType<PyTestType>` instead of `mlir_type_subclass` and
`PyConcreteAttribute<PyTestAttr>` instead of `mlir_atrr_subclass`,
verifying this works for non-core extensions in-tree;
2. `StandaloneExtensionNanobind` is ported to use `struct PyCustomType :
mlir::python::MLIR_BINDINGS_PYTHON_DOMAIN::PyConcreteType<PyCustomType>`
instead of `mlir_type_subclass` verifying this works for non-core
extensions out-of-tree;
3. `StandaloneExtensionNanobind`'s `smoketest` is extended to also load
another bindings package (namely `mlir`) verifying
`MLIR_BINDINGS_PYTHON_DOMAIN` successfully disambiguates symbols and
typeids.
I have also tested this downstream:
https://github.com/llvm/eudsl/pull/287 as well run the following builder
bots:
mlir-nvidia-gcc7:
https://lab.llvm.org/buildbot/#/buildrequests/6654424?redirect_to_build=true
I have also tested against IREE:
https://github.com/iree-org/iree/pull/21916
# Integration
It is highly recommended to set the CMake var
`MLIR_BINDINGS_PYTHON_NB_DOMAIN` (which will also determine
`MLIR_BINDINGS_PYTHON_DOMAIN`) to something unique for each downstream.
This can also be passed explicitly to `add_mlir_python_modules` if your
project builds multiple bindings packages. I added a `WARNING` to this
effect in `AddMLIRPython.cmake`.
[^3]: Python values being typed correctly when exiting from cpp;
[^1]: Specifically when the modules are imported using `importlib`,
which occurs with nanobind's
[stubgen](https://github.com/wjakob/nanobind/blob/master/src/stubgen.py#L965);
[^2]: The workaround we implemented was a class method for the dialect
bindings called `Class.isinstance(...)`;
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::`
to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
Following a rather direct approach to expose PDL usage from C and then
Python. This doesn't yes plumb through adding support for custom
matchers through this interface, so constrained to basics initially.
This also exposes greedy rewrite driver. Only way currently to define
patterns is via PDL (just to keep small). The creation of the PDL
pattern module could be improved to avoid folks potentially accessing
the module used to construct it post construction. No ergonomic work
done yet.
---------
Signed-off-by: Jacques Pienaar <jpienaar@google.com>
Maksim Levental