This PR implements "automatic" location inference in the bindings. The
way it works is it walks the frame stack collecting source locations
(Python captures these in the frame itself). It is inspired by JAX's
[implementation](523ddcfbca/jax/_src/interpreters/mlir.py (L462))
but moves the frame stack traversal into the bindings for better
performance.
The system supports registering "included" and "excluded" filenames;
frames originating from functions in included filenames **will not** be
filtered and frames originating from functions in excluded filenames
**will** be filtered (in that order). This allows excluding all the
generated `*_ops_gen.py` files.
The system is also "toggleable" and off by default to save people who
have their own systems (such as JAX) from the added cost.
Note, the system stores the entire stacktrace (subject to
`locTracebackFramesLimit`) in the `Location` using specifically a
`CallSiteLoc`. This can be useful for profiling tools (flamegraphs
etc.).
Shoutout to the folks at JAX for coming up with a good system.
---------
Co-authored-by: Jacques Pienaar <jpienaar@google.com>
This is a companion to #118583, although it can be landed independently
because since #117922 dialects do not have to use the same Python
binding framework as the Python core code.
This PR ports all of the in-tree dialect and pass extensions to
nanobind, with the exception of those that remain for testing pybind11
support.
This PR also:
* removes CollectDiagnosticsToStringScope from NanobindAdaptors.h. This
was overlooked in a previous PR and it is duplicated in Diagnostics.h.
---------
Co-authored-by: Jacques Pienaar <jpienaar@google.com>
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.
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
bfb1ba7526
and
9566ee2806.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
<img
src="https://github.com/llvm/llvm-project/assets/5657668/443852b6-ac25-45bb-a38b-5dfbda09d5a7"
height="400" />
<p></p>
So turns out that none of the `replace=True` things actually work
because of the map caches (except for
`register_attribute_builder(replace=True)`, which doesn't use such a
cache). This was hidden by a series of unfortunate events:
1. `register_type_caster` failure was hidden because it was the same
`TestIntegerRankedTensorType` being replaced with itself (d'oh).
2. `register_operation` failure was hidden behind the "order of events"
in the lifecycle of typical extension import/use. Since extensions are
loaded/registered almost immediately after generated builders are
registered, there is no opportunity for the `operationClassMapCache` to
be populated (through e.g., `module.body.operations[2]` or
`module.body.operations[2].opview` or something). Of course as soon as
you as actually do "late-bind/late-register" the extension, you see it's
not successfully replacing the stale one in `operationClassMapCache`.
I'll take this opportunity to propose we ditch the caches all together.
I've been cargo-culting them but I really don't understand how they
work. There's this comment above `operationClassMapCache`
```cpp
/// Cache of operation name to external operation class object. This is
/// maintained on lookup as a shadow of operationClassMap in order for repeat
/// lookups of the classes to only incur the cost of one hashtable lookup.
llvm::StringMap<pybind11::object> operationClassMapCache;
```
But I don't understand how that's true given that the canonical thing
`operationClassMap` is already a map:
```cpp
/// Map of full operation name to external operation class object.
llvm::StringMap<pybind11::object> operationClassMap;
```
Maybe it wasn't always the case? Anyway things work now but it seems
like an unnecessary layer of complexity for not much gain? But maybe I'm
wrong.
This PR replaces the mixin `OpView` extension mechanism with the
standard inheritance mechanism.
Why? Firstly, mixins are not very pythonic (inheritance is usually used
for this), a little convoluted, and too "tight" (can only be used in the
immediately adjacent `_ext.py`). Secondly, it (mixins) are now blocking
are correct implementation of "value builders" (see
[here](https://github.com/llvm/llvm-project/pull/68764)) where the
problem becomes how to choose the correct base class that the value
builder should call.
This PR looks big/complicated but appearances are deceiving; 4 things
were needed to make this work:
1. Drop `skipDefaultBuilders` in
`OpPythonBindingGen::emitDefaultOpBuilders`
2. Former mixin extension classes are converted to inherit from the
generated `OpView` instead of being "mixins"
a. extension classes that simply were calling into an already generated
`super().__init__` continue to do so
b. (almost all) extension classes that were calling `self.build_generic`
because of a lack of default builder being generated can now also just
call `super().__init__`
3. To handle the [lone single
use-case](https://sourcegraph.com/search?q=context%3Aglobal+select_opview_mixin&patternType=standard&sm=1&groupBy=repo)
of `select_opview_mixin`, namely
[linalg](https://github.com/llvm/llvm-project/blob/main/mlir/python/mlir/dialects/_linalg_ops_ext.py#L38),
only a small change was necessary in `opdsl/lang/emitter.py` (thanks to
the emission/generation of default builders/`__init__`s)
4. since the `extend_opview_class` decorator is removed, we need a way
to register extension classes as the desired `OpView` that `op.opview`
conjures into existence; so we do the standard thing and just enable
replacing the existing registered `OpView` i.e.,
`register_operation(_Dialect, replace=True)`.
Note, the upgrade path for the common case is to change an extension to
inherit from the generated builder and decorate it with
`register_operation(_Dialect, replace=True)`. In the slightly more
complicated case where `super().__init(self.build_generic(...))` is
called in the extension's `__init__`, this needs to be updated to call
`__init__` in `OpView`, i.e., the grandparent (see updated docs).
Note, also `<DIALECT>_ext.py` files/modules will no longer be automatically loaded.
Note, the PR has 3 base commits that look funny but this was done for
the purpose of tracking the line history of moving the
`<DIALECT>_ops_ext.py` class into `<DIALECT>.py` and updating (commit
labeled "fix").
This PR implements python enum bindings for *all* the enums - this includes `I*Attrs` (including positional/bit) and `Dialect/EnumAttr`.
There are a few parts to this:
1. CMake: a small addition to `declare_mlir_dialect_python_bindings` and `declare_mlir_dialect_extension_python_bindings` to generate the enum, a boolean arg `GEN_ENUM_BINDINGS` to make it opt-in (even though it works for basically all of the dialects), and an optional `GEN_ENUM_BINDINGS_TD_FILE` for handling corner cases.
2. EnumPythonBindingGen.cpp: there are two weedy aspects here that took investigation:
1. If an enum attribute is not a `Dialect/EnumAttr` then the `EnumAttrInfo` record is canonical, as far as both the cases of the enum **and the `AttrDefName`**. On the otherhand, if an enum is a `Dialect/EnumAttr` then the `EnumAttr` record has the correct `AttrDefName` ("load bearing", i.e., populates `ods.ir.AttributeBuilder('<NAME>')`) but its `enum` field contains the cases, which is an instance of `EnumAttrInfo`. The solution is to generate an one enum class for both `Dialect/EnumAttr` and "independent" `EnumAttrInfo` but to make that class interopable with two builder registrations that both do the right thing (see next sub-bullet).
2. Because we don't have a good connection to cpp `EnumAttr`, i.e., only the `enum class` getters are exposed (like `DimensionAttr::get(Dimension value)`), we have to resort to parsing e.g., `Attribute.parse(f'#gpu<dim {x}>')`. This means that the set of supported `assemblyFormat`s (for the enum) is fixed at compile of MLIR (currently 2, the only 2 I saw). There might be some things that could be done here but they would require quite a bit more C API work to support generically (e.g., casting ints to enum cases and binding all the getters or going generically through the `symbolize*` methods, like `symbolizeDimension(uint32_t)` or `symbolizeDimension(StringRef)`).
A few small changes:
1. In addition, since this patch registers default builders for attributes where people might've had their own builders already written, I added a `replace` param to `AttributeBuilder.insert` (`False` by default).
2. `makePythonEnumCaseName` can't handle all the different ways in which people write their enum cases, e.g., `llvm.CConv.Intel_OCL_BI`, which gets turned into `INTEL_O_C_L_B_I` (because `llvm::convertToSnakeFromCamelCase` doesn't look for runs of caps). So I dropped it. On the otherhand regularization does need to done because some enums have `None` as a case (and others might have other python keywords).
3. I turned on `llvm` dialect generation here in order to test `nvvm.WGMMAScaleIn`, which is an enum with [[ d7e26b5620/mlir/include/mlir/IR/EnumAttr.td (L22-L25) | no explicit discriminator ]] for the `neg` case.
Note, dialects that didn't get a `GEN_ENUM_BINDINGS` don't have any enums to generate.
Let me know if I should add more tests (the three trivial ones I added exercise both the supported `assemblyFormat`s and `replace=True`).
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D157934
depends on D150839
This diff uses `MlirTypeID` to register `TypeCaster`s (i.e., `[](PyType pyType) -> DerivedTy { return pyType; }`) for all concrete types (i.e., `PyConcrete<...>`) that are then queried for (by `MlirTypeID`) and called in `struct type_caster<MlirType>::cast`. The result is that anywhere an `MlirType mlirType` is returned from a python binding, that `mlirType` is automatically cast to the correct concrete type. For example:
```
c0 = arith.ConstantOp(f32, 0.0)
# CHECK: F32Type(f32)
print(repr(c0.result.type))
unranked_tensor_type = UnrankedTensorType.get(f32)
unranked_tensor = tensor.FromElementsOp(unranked_tensor_type, [c0]).result
# CHECK: UnrankedTensorType
print(type(unranked_tensor.type).__name__)
# CHECK: UnrankedTensorType(tensor<*xf32>)
print(repr(unranked_tensor.type))
```
This functionality immediately extends to typed attributes (i.e., `attr.type`).
The diff also implements similar functionality for `mlir_type_subclass`es but in a slightly different way - for such types (which have no cpp corresponding `class` or `struct`) the user must provide a type caster in python (similar to how `AttrBuilder` works) or in cpp as a `py::cpp_function`.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D150927
The raw `OpView` classes are used to bypass the constructors of `OpView`
subclasses, but having a separate class can create some confusing
behaviour, e.g.:
```
op = MyOp(...)
# fails, lhs is 'MyOp', rhs is '_MyOp'
assert type(op) == type(op.operation.opview)
```
Instead we can use `__new__` to achieve the same thing without a
separate class:
```
my_op = MyOp.__new__(MyOp)
OpView.__init__(my_op, op)
```
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143830
Previously this was incorrectly assigning py::none to where function was
expected which resulted in failure if one used a non-attribute for
attribute without registered builder.
Fix Python 3.6.9 issue encountered due to type checking here. Will
add back in follow up.
This reverts commit 1f47fee2948ef48781084afe0426171d000d7997.
For cases where we can automatically construct the Attribute allow for more
user-friendly input. This is consistent with C++ builder generation as well
choice of which single builder to generate here (most
specialized/user-friendly).
Registration of attribute builders from more pythonic input is all Python side.
The downside is that
* extra checking to see if user provided a custom builder in op builders,
* the ODS attribute name is load bearing
upside is that
* easily change these/register dialect specific ones in downstream projects,
* adding support/changing to different convenience builders are all along with
the rest of the convenience functions in Python (and no additional changes
to tablegen file or recompilation needed);
Allow for both building with Attributes as well as raw inputs. This change
should therefore be backwards compatible as well as allow for avoiding
recreating Attribute where already available.
Differential Revision: https://reviews.llvm.org/D139568
I'm not sure what is the right fix here, but adding a name to all these
lead to many segfaults.
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D116506
* Now that packaging has stabilized, removes old mechanisms for loading extensions, preferring direct importing.
* Removes _cext_loader.py, _dlloader.py as unnecessary.
* Fixes the path where the CAPI dll is written on Windows. This enables that path of least resistance loading behavior to work with no further drama (see: https://bugs.python.org/issue36085).
* With this patch, `ninja check-mlir` on Windows with Python bindings works for me, modulo some failures that are actually due to a couple of pre-existing Windows bugs. I think this is the first time the Windows Python bindings have worked upstream.
* Downstream changes needed:
* If downstreams are using the now removed `load_extension`, `reexport_cext`, etc, then those should be replaced with normal import statements as done in this patch.
Reviewed By: jdd, aartbik
Differential Revision: https://reviews.llvm.org/D108489
This will allow the bindings to be built as a library and reused in out-of-tree
projects that want to provide bindings on top of MLIR bindings.
Reviewed By: stellaraccident, mikeurbach
Differential Revision: https://reviews.llvm.org/D101075
