mirror of https://github.com/llvm/torch-mlir
360 lines
12 KiB
Python
360 lines
12 KiB
Python
# Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
# See https://llvm.org/LICENSE.txt for license information.
|
|
# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
# Also available under a BSD-style license. See LICENSE.
|
|
"""
|
|
Translator from torch.jit.ScriptFunction to MLIR.
|
|
|
|
|
|
The following defines a set of classes for converting types used by Python and PyTorch into MLIR types from the
|
|
`torch` dialect.
|
|
|
|
The expected use of this module is to create an instance of one of the classes below, and then calling the
|
|
`to_mlir` method to generate the MLIR representation of the type.
|
|
|
|
Information about what types are supported by each class can be found in docstrings of each of the classes.
|
|
|
|
In addition this module defines a function that take a torch.jit.ScriptFunction and converts it into an MLIR module.
|
|
|
|
The expected use for this module is to use the function
|
|
`build_module(jit_function: torch.jit.ScriptFunction annotation: Annotation) -> ir.Module`
|
|
to convert the TorchScript function into MLIR using the `torch` dialect.
|
|
"""
|
|
|
|
import abc
|
|
import re
|
|
from typing import Any, Optional, Iterable, Dict
|
|
from typing import Union
|
|
|
|
import numpy as np
|
|
import torch
|
|
import torch._C
|
|
import torch.jit
|
|
from torch._ops import OpOverload
|
|
|
|
from torch_mlir import ir
|
|
from torch_mlir.dialects.func import FuncOp
|
|
from torch_mlir.dialects.torch.importer.jit_ir import ModuleBuilder
|
|
|
|
|
|
class TorchMlirType(abc.ABC):
|
|
"""
|
|
A `TorchMlirType` is an object that produces MLIR
|
|
types in the `torch` dialect. The only requirement
|
|
for a class to be a subclass of `TorchMlirType` is
|
|
to define a `to_mlir(self, ir.Context) -> ir.Type`.
|
|
Each class is allowed to have different types of
|
|
__init__ methods depending on the information they
|
|
require to produce the given MLIR representation.
|
|
"""
|
|
|
|
@abc.abstractmethod
|
|
def to_mlir(self, context: ir.Context) -> ir.Type:
|
|
pass
|
|
|
|
|
|
class TorchTensorTypeError(Exception):
|
|
def __init__(self, value: str):
|
|
super().__init__()
|
|
self.value = value
|
|
|
|
def __str__(self) -> str:
|
|
return self.value
|
|
|
|
|
|
class TorchTensorType(TorchMlirType):
|
|
"""
|
|
This class is used to generate types of the form
|
|
!torch.tensor and !torch.vtensor<SHAPE, DTYPE>,
|
|
where SHAPE is a list representing the shape of the tensor,
|
|
and DTYPE is an MLIR data type.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
*,
|
|
shape: Optional[Iterable[Optional[int]]] = None,
|
|
dtype: Optional[torch.dtype] = None,
|
|
):
|
|
self.shape = shape
|
|
self.dtype = dtype
|
|
|
|
if dtype is None and shape is not None:
|
|
err = "If shape is specified, dtype must also be specified"
|
|
raise TorchTensorTypeError(err)
|
|
|
|
def __str__(self):
|
|
return f"Torch Tensor (shape={self.shape}, dtype={self.dtype})"
|
|
|
|
def to_mlir(self, context: ir.Context) -> ir.Type:
|
|
if self.dtype is None:
|
|
return ir.Type.parse("!torch.tensor", context=context)
|
|
|
|
shape_asm = self._shape_to_mlir_asm()
|
|
dtype_asm = self._dtype_to_mlir_asm()
|
|
return ir.Type.parse(
|
|
f"!torch.vtensor<{shape_asm},{dtype_asm}>", context=context
|
|
)
|
|
|
|
def _shape_to_mlir_asm(self) -> str:
|
|
if self.shape is None:
|
|
return "*"
|
|
|
|
str_sizes = map(lambda x: "?" if x is None else str(x), self.shape)
|
|
return f'[{",".join(str_sizes)}]'
|
|
|
|
def _dtype_to_mlir_asm(self) -> str:
|
|
if self.dtype in [torch.float64]:
|
|
return "f64"
|
|
if self.dtype in [torch.float, torch.float32]:
|
|
return "f32"
|
|
if self.dtype in [torch.int, torch.int32]:
|
|
return "si32"
|
|
if self.dtype in [torch.int64]:
|
|
return "si64"
|
|
if self.dtype in [torch.bool]:
|
|
return "i1"
|
|
|
|
raise NotImplementedError(f"Unsupported dtype: {self.dtype}")
|
|
|
|
|
|
class TorchNnModuleType(TorchMlirType):
|
|
"""This class is used to generate types for `!torch.nn.Module`s."""
|
|
|
|
def __init__(self, module_name: str):
|
|
self.module_name = module_name
|
|
|
|
def __str__(self):
|
|
return "torch.nn.Module"
|
|
|
|
def to_mlir(self, context: ir.Context) -> ir.Type:
|
|
return ir.Type.parse(f'!torch.nn.Module<"{self.module_name}">', context=context)
|
|
|
|
|
|
class PythonType(TorchMlirType):
|
|
"""
|
|
This class is used to convert regular Python types
|
|
into their corresponding `torch` dialect representation.
|
|
The list of supported types can be found in the dictionary
|
|
`_type_to_asm_dict`.
|
|
"""
|
|
|
|
_type_to_asm_dict = {
|
|
bool: "!torch.bool",
|
|
int: "!torch.int",
|
|
type(None): "!torch.none",
|
|
}
|
|
|
|
def __init__(self, type_: Any):
|
|
self.type_ = type_
|
|
|
|
def __str__(self):
|
|
return str(self.type_)
|
|
|
|
def to_mlir(self, context: ir.Context) -> ir.Type:
|
|
asm = self._type_to_asm_dict.get(self.type_)
|
|
if asm is None:
|
|
raise NotImplementedError(f"Unsupported type: {self.type_}")
|
|
return ir.Type.parse(asm, context=context)
|
|
|
|
|
|
# TODO: This functionality should be incorporated into ModuleBuilder.import_function.
|
|
class Annotation:
|
|
def __init__(self, types: Iterable[Union[TorchTensorType, type]]):
|
|
self.types = list(
|
|
map(lambda t: PythonType(t) if isinstance(t, type) else t, types)
|
|
)
|
|
|
|
def __str__(self):
|
|
result = f"Annotation instance with {len(self.types)} types\n"
|
|
for e, type_ in enumerate(self.types):
|
|
result += f" Type of argument {e + 1}: {str(type_)}\n"
|
|
return result
|
|
|
|
def __iter__(self):
|
|
return iter(self.types)
|
|
|
|
|
|
class AnnotationConverter:
|
|
@staticmethod
|
|
def to_mlir_array_attr(annotation: Annotation, context: ir.Context) -> ir.ArrayAttr:
|
|
dict_attrs = []
|
|
for type_ in annotation:
|
|
if not isinstance(type_, TorchTensorType):
|
|
dict_attrs.append(ir.DictAttr.get({}, context=context))
|
|
continue
|
|
|
|
ir_type = type_.to_mlir(context)
|
|
with context:
|
|
type_attr = ir.TypeAttr.get(ir_type)
|
|
dict_attr = ir.DictAttr.get({"torch.type_bound": type_attr})
|
|
dict_attrs.append(dict_attr)
|
|
|
|
return ir.ArrayAttr.get(dict_attrs, context=context)
|
|
|
|
|
|
def get_func_op_with_name(module: ir.Module, name: str) -> Optional[FuncOp]:
|
|
with module.context:
|
|
name_attr = ir.StringAttr.get(name)
|
|
for op in module.body.operations:
|
|
if isinstance(op, FuncOp) and op.name == name_attr:
|
|
# Add name of torch op as debug_module_name so that
|
|
# run_pipeline_with_repro_report can use it.
|
|
module.operation.attributes["torch.debug_module_name"] = name_attr
|
|
return op
|
|
|
|
return None
|
|
|
|
|
|
def is_tensor_type(typ: torch._C.Type):
|
|
return typ.isSubtypeOf(torch.TensorType.get()) or (
|
|
isinstance(typ, torch.OptionalType)
|
|
and typ.getElementType().isSubtypeOf(torch._C.TensorType.get())
|
|
)
|
|
|
|
|
|
def is_list_of_tensors_type(typ: torch._C.Type):
|
|
return isinstance(typ, torch.ListType) and is_tensor_type(typ.getElementType())
|
|
|
|
|
|
name_mangle_regex = re.compile("[^a-zA-Z0-9]")
|
|
|
|
|
|
def build_ts_script_function(
|
|
schema: torch._C.FunctionSchema, kwargs: Dict[str, Any]
|
|
) -> torch.jit.ScriptFunction:
|
|
"""Build a torch.jit.ScriptFunction that corresponds to the schema.
|
|
|
|
Constants are inlined for the purposes of invalidating the compile cache when they change.
|
|
|
|
Parameters
|
|
----------
|
|
schema: torch._C.FunctionSchema
|
|
PyTorch's representation for ops, contains type information needed for inlining constants into the TS graph.
|
|
kwargs: Dict
|
|
A dictionary with all arguments passed in through __torch_dispatch__ (including int/float/bool params).
|
|
|
|
Returns
|
|
-------
|
|
torch.jit.ScriptFunction
|
|
Fully specialized (all constants) TS graph whose only arguments are tensors.
|
|
"""
|
|
|
|
# Creates empty TS graph.
|
|
graph = torch._C.Graph()
|
|
# Creates and inserts node with identifier `schema.name`; NB node has no inputs or outputs at this point.
|
|
node = graph.insertNode(graph.create(schema.name, len(schema.returns)))
|
|
# Associate graph inputs/outputs with node inputs/outputs.
|
|
graph_inputs = []
|
|
for arg in schema.arguments:
|
|
arg_name = arg.name if arg.name != "self" else "input"
|
|
|
|
# If arg is a flattened list of tensors, such as in the case of torch.cat
|
|
# then add each element of the list to the graph corresponding to arg
|
|
# and insert a ListConstruct to function as input to the op.
|
|
if is_list_of_tensors_type(arg.type):
|
|
inps = []
|
|
for kwarg in [
|
|
kwarg for kwarg in kwargs if f"{arg_name}_flattened" in kwarg
|
|
]:
|
|
inp = graph.addInput()
|
|
el_typ = arg.type.getElementType()
|
|
if isinstance(el_typ, torch.OptionalType):
|
|
el_typ = el_typ.getElementType()
|
|
inp.setType(el_typ)
|
|
inp.setDebugName(kwarg)
|
|
inps.append(inp)
|
|
graph_inputs.append(kwarg)
|
|
list_cons = graph.insertNode(graph.create("prim::ListConstruct", inps))
|
|
list_cons.moveBefore(node)
|
|
inp = list_cons.output()
|
|
inp.setType(torch.ListType.ofTensors())
|
|
# If arg is a tensor, then add input to the graph corresponding to arg.
|
|
elif is_tensor_type(arg.type) and kwargs[arg_name] is not None:
|
|
inp = graph.addInput()
|
|
if isinstance(arg.type, torch.OptionalType):
|
|
el_typ = arg.type.getElementType()
|
|
else:
|
|
el_typ = arg.type
|
|
inp.setType(el_typ)
|
|
inp.setDebugName(arg_name)
|
|
graph_inputs.append(arg_name)
|
|
# If arg is a constant, inline (at the top of the graph).
|
|
else:
|
|
val = kwargs[arg_name]
|
|
if val == []:
|
|
# Some ops have empty list default values for args
|
|
# (such as aten::max_pool2d_with_indices with int[2] stride=[]
|
|
# but graph.insertConstant doesnt' recognize [] as an empty list IValue.
|
|
# This might be an upstream bug but there doesn't seem to be a way to
|
|
# build a prim::ListConstruct list that's empty.
|
|
val = None
|
|
inp = graph.insertConstant(val)
|
|
inp.node().moveBefore(node)
|
|
|
|
node.addInput(inp)
|
|
|
|
# Reorder graph inputs to match kwargs.
|
|
permutes = [
|
|
{inp: i for i, inp in enumerate(graph_inputs)}[kwarg]
|
|
for kwarg in [kwarg for kwarg in kwargs if kwarg in graph_inputs]
|
|
]
|
|
graph.permuteInputs(permutes)
|
|
|
|
if node.hasMultipleOutputs():
|
|
for outp in node.outputs():
|
|
graph.registerOutput(outp)
|
|
else:
|
|
graph.registerOutput(node.output())
|
|
|
|
fn = torch._C._create_function_from_graph(
|
|
f"{name_mangle_regex.sub('', str(graph))}", graph
|
|
)
|
|
return fn
|
|
|
|
|
|
def build_mlir_module(op: OpOverload, kwargs: Dict[str, Any]) -> ir.Module:
|
|
"""Translate input function into an MLIR module in the `torch` dialect.
|
|
|
|
Parameters
|
|
----------
|
|
op: OpOverload
|
|
Callable from the torch.ops.aten module/namespace that has a _schema field.
|
|
kwargs: Dict
|
|
A dictionary with all arguments passed in through __torch_dispatch__ (including int/float,bool params).
|
|
|
|
Returns
|
|
-------
|
|
ir.Module
|
|
Translation of the input module into an MLIR module.
|
|
"""
|
|
|
|
# The assert here is to catch tensor shapes that have size 0 dimensions, such as those produced in
|
|
# the course of evaluating SliceEndSleStartModule_basic and SliceOutOfLowerBoundEndIndexModule_basic.
|
|
# Such 0 size dimensions fail the assert at mlir/lib/IR/BuiltinTypes.cpp, line 887
|
|
annotations = []
|
|
for arg_name, arg in kwargs.items():
|
|
if isinstance(arg, torch.Tensor):
|
|
assert np.prod(arg.shape) != 0, f"{arg_name} has invalid shape {arg.shape}"
|
|
annotations.append(TorchTensorType(shape=tuple(arg.shape), dtype=arg.dtype))
|
|
annotations = tuple(annotations)
|
|
|
|
script_fun = build_ts_script_function(op._schema, kwargs)
|
|
assert len(annotations) == len(
|
|
list(script_fun.graph.inputs())
|
|
), "Number of annotations and number of graph inputs differs."
|
|
|
|
mb = ModuleBuilder()
|
|
mb.import_function(script_fun)
|
|
|
|
func_op = get_func_op_with_name(mb.module, script_fun.name)
|
|
assert (
|
|
func_op is not None
|
|
), "Unable to find FuncOp in new module. Make sure function was imported correctly into ModuleBuilder"
|
|
|
|
func_annotation = Annotation(annotations)
|
|
arg_attrs = AnnotationConverter.to_mlir_array_attr(func_annotation, mb.context)
|
|
func_op.attributes["arg_attrs"] = arg_attrs
|
|
|
|
return mb.module
|