def __init__(self, binary_dir):
self.script_path = Path(__file__).resolve()
self.config_path = (
Path(__file__).resolve().parent.joinpath("autogen_ltc_backend.yaml")
)
self.torch_ops_file = TORCH_MLIR_DIR.joinpath(
# fmt: off
"include", "torch-mlir", "Dialect", "Torch", "IR", "GeneratedTorchOps.td",
# fmt: on
)
assert self.torch_ops_file.exists()
self.binary_dir = Path(binary_dir)
assert self.binary_dir.is_dir(), f"Binary directory not found: {self.binary_dir}"
self.source_yaml = self.binary_dir.joinpath("generated_native_functions.yaml")
self.backend_path = TORCH_MLIR_DIR.joinpath(
"python", "torch_mlir", "csrc", "base_lazy_backend"
)
assert self.backend_path.is_dir()
self.generated_path = self.binary_dir.joinpath(
"python", "torch_mlir", "csrc", "base_lazy_backend", "generated"
)
self.generated_path.mkdir(parents=True, exist_ok=True)
# Create symlink to match doc structure
generated_path = self.backend_path.joinpath("generated").resolve()
if not generated_path.exists():
generated_path.symlink_to(
os.path.relpath(self.generated_path, generated_path.parent),
target_is_directory=True,
)
self.tensor_class = "torch::lazy::LazyTensor"
# Set the lazy value class
setValueT(BaseCppType("torch::lazy", "Value"))
def run_gen_lazy_tensor(
aten_path: str,
source_yaml: str,
output_dir: str,
dry_run: bool,
impl_path: Optional[str],
node_base: str = default_args.node_base,
node_base_hdr: Optional[str] = default_args.node_base_hdr,
tensor_class: str = default_args.tensor_class,
tensor_class_hdr: str = default_args.tensor_class_hdr,
shape_inference_hdr: str = default_args.shape_inference_hdr,
lazy_ir_generator: Type[GenLazyIR] = default_args.lazy_ir_generator,
# build_in_tree is true for TS backend and affects include paths
build_in_tree: bool = False,
# per_operator_headers changes whether ATen/Functions.h or individual operator headers are used
# it must match how ATen was built
per_operator_headers: bool = False,
backend_name: str = default_args.backend_name,
gen_forced_fallback_code: bool = False,
# the following arguments are temporary customization points for xla backend migration.
# do not rely on them otherwise, they should be removed once migration is complete
backend_namespace: str = "torch::lazy",
get_tensorlist: str = "GetTensorList",
get_tensor_or_wrap_number: str = "GetLtcTensorOrCreateForWrappedNumber",
try_get_tensor: str = "TryGetLtcTensor",
metrics_counter: str = 'TORCH_LAZY_FN_COUNTER("lazy::")',
create_tensor: str = "LazyTensor::Create",
create_from_first_tensor: bool = False,
create_aten_from_ltc_tensor: str = "torch::lazy::CreateAtenFromLtcTensor",
tuple_aten_from_ltc_tensors: str = "torch::lazy::TupleAtenFromLtcTensors",
lazy_value_class: str = "torch::lazy::Value",
lazy_tensor_ptr: str = "LazyTensorPtr",
get_device_fn: str = "torch::lazy::GetBackendDevice",
) -> None:
lv_tokens = lazy_value_class.split("::")
lv_class = lv_tokens[-1]
lv_ns = "::".join(lv_tokens[:-1])
setValueT(BaseCppType(lv_ns, lv_class))
template_dir = os.path.join(aten_path, "templates")
def make_file_manager(install_dir: str) -> FileManager:
return FileManager(
install_dir=install_dir, template_dir=template_dir, dry_run=dry_run
)
fm = make_file_manager(output_dir)
native_yaml_path = os.path.join(aten_path, "native/native_functions.yaml")
tags_yaml_path = os.path.join(aten_path, "native/tags.yaml")
parsed_yaml = parse_native_yaml(native_yaml_path, tags_yaml_path)
native_functions, backend_indices = (
parsed_yaml.native_functions,
parsed_yaml.backend_indices,
)
grouped_native_functions = get_grouped_native_functions(native_functions)
def sort_native_function(f: Union[NativeFunctionsGroup, NativeFunction]) -> str:
"""
We sort the native function because of the note in concat_map_codegen.
TODO(alanwaketan): Remove this sorting hack once all ops are grouped properly.
"""
func = f.functional.func if isinstance(f, NativeFunctionsGroup) else f.func
return str(func.name.name)
grouped_native_functions = sorted(
grouped_native_functions, key=sort_native_function
)
parsed_backend_yaml = parse_backend_yaml(
source_yaml, grouped_native_functions, backend_indices
)
backend_key = parsed_backend_yaml.backend_key
autograd_key = parsed_backend_yaml.autograd_key
cpp_namespace = parsed_backend_yaml.cpp_namespace
backend_indices = parsed_backend_yaml.backend_indices
full_codegen = parse_full_codegen_ops(source_yaml, grouped_native_functions)
def concat_map_codegen(
func: Callable[[NativeFunction], Sequence[str]],
xs: Iterable[Union[NativeFunctionsGroup, NativeFunction]],
*,
codegenInplaceVariant: bool = False,
) -> Iterator[str]:
"""
We code-gen for the functional variant, which is all we need for IR classes/lowerings/shape inferences, but we
only code-gen additional entries for the inplace variant for the native functions.
Note: If xs is not sorted, there may be an edge case when generating IR classes. Considering relu and relu_, if
we encounter relu_ before relu. we will then generate an IR class with op = at::aten::relu_ for both relu and
relu_ which will cause problems for relu.
TODO(alanwaketan): Once all ops are grouped properly, we should no longer need this hack.
"""
generated = set()
def gen_key(func: FunctionSchema) -> Tuple[str, str]:
# we want to generate unique entries for overloads of functional variants,
# but not for inplace variants unless explicitly told `codegenInplaceVariant`
return (func.name.name.base, func.name.overload_name)
for x in xs:
f = x.functional if isinstance(x, NativeFunctionsGroup) else x
# For the 'or'd terms:
# 1. codegenInplaceVariant means we can generate the in-place variant corresponding items.
# 2. not f.func.name.name.inplace means the op is not a in-place variant, so we can generate the item.
# 3. f.func.name.name.base not in generated means even for in-place ops we still need to generate the item
# as if they were the functional variants for one time.
if f.func.name in full_codegen and (
codegenInplaceVariant
or not f.func.name.name.inplace
or gen_key(f.func) not in generated
):
generated.add(gen_key(f.func))
for r in func(f):
yield r
selector = SelectiveBuilder.get_nop_selector()
assert backend_key is not None
class_name = backend_indices[backend_key].native_function_class_name()
if impl_path is not None:
error_on_missing_kernels(
native_functions,
backend_indices,
backend_key,
autograd_key,
class_name,
impl_path,
full_codegen,
)
""" Validate Shape Inference Definitions
Generated lazy native functions all perform shape inference, by first using a meta:: kernel
if available for that op, and otherwise using a 'compute_shape_{op}' function instead. The generator
knows the call signature for compute_shape_{op} becuase it matches the nativefunction (and meta::) signature,
so it just has to check whether the op is structured and generate a call for one or the other. It's up to the dev
to supply the missing compute_shape_{op} function, but the codegen at least warns you about this and provides
the expected signature which can be copy-pasted into shape_inference.h.
compute_shape_{op} functions are handwritten and should be replaced over time as ops get ported
to structured kernels.
See torch/csrc/lazy/core/shape_inference.cpp #READ THIS! for more information.
"""
if shape_inference_hdr is not None:
expected_shape_infr_decls = list(
concat_map_codegen(
dest.GenLazyShapeInferenceDefinition(
backend_indices[backend_key], tensor_class
),
grouped_native_functions,
codegenInplaceVariant=True,
)
)
validate_shape_inference_header(shape_inference_hdr, expected_shape_infr_decls)
assert class_name is not None
# Generate nativefunction declarations
# Note, eager registrations is set to False for the lazy TS backend as another LTC backend
# may want to register their own lazy kernels instead of registering the TS ones.
# The registration will lazily happen when init_ts_backend is called.
gen_dispatchkey_nativefunc_headers(
fm,
class_name,
cpp_namespace,
backend_indices,
grouped_native_functions,
backend_key,
autograd_key,
backend_name,
)
# Generate Dispatcher registrations which hook up the nativefunctions
for dispatch_key in (
[backend_key] if autograd_key is None else [backend_key, autograd_key]
):
gen_dispatcher_registrations(
fm,
output_dir,
class_name,
cpp_namespace,
backend_indices,
grouped_native_functions,
backend_key,
dispatch_key,
selector,
build_in_tree=build_in_tree,
per_operator_headers=per_operator_headers,
backend_name=backend_name,
eager_registration=False,
)
# Generate native function impls that build IR nodes
ns_helper = NamespaceHelper(cpp_namespace)
fm.write_with_template(
f"{backend_key}NativeFunctions.cpp",
"DispatchKeyNativeFunctions.cpp",
lambda: {
"includes": [
f"#include <{path}>"
for path in [
tensor_class_hdr,
shape_inference_hdr,
"ATen/Functions.h",
"ATen/MetaFunctions.h",
"ATen/Operators.h",
"ATen/native/CPUFallback.h",
"torch/csrc/lazy/core/ir_builder.h",
"torch/csrc/lazy/core/lazy_graph_executor.h",
"torch/csrc/lazy/core/metrics.h",
"torch/csrc/lazy/core/shape.h",
f"{output_dir}/{backend_key}NativeFunctions.h",
f"{output_dir}/LazyIr.h",
]
+ (
["torch/csrc/lazy/ts_backend/ts_eager_fallback.h"]
if gen_forced_fallback_code
else []
)
],
"native_functions_include": "",
"namespace_prologue": ns_helper.prologue,
"namespace_epilogue": ns_helper.epilogue,
"native_function_definitions": list(
concat_map_codegen(
dest.GenLazyNativeFuncDefinition(
f"{backend_key}NativeFunctions",
backend_indices[backend_key],
tensor_class,
gen_forced_fallback_code,
backend_namespace,
get_tensorlist,
get_tensor_or_wrap_number,
try_get_tensor,
metrics_counter,
create_tensor,
create_from_first_tensor,
create_aten_from_ltc_tensor,
tuple_aten_from_ltc_tensors,
lazy_tensor_ptr,
get_device_fn,
),
grouped_native_functions,
codegenInplaceVariant=True,
)
),
},
)
# Generate IR node classes
fm.write_with_template(
"LazyIr.h",
"LazyIr.h",
lambda: {
"lazy_ir_sysinc": [
f"#include <{path}>"
for path in [
"ATen/core/Formatting.h",
"c10/core/ScalarType.h",
"c10/util/Optional.h",
"torch/csrc/lazy/core/hash.h",
"torch/csrc/lazy/core/ir.h",
"torch/csrc/lazy/core/shape.h",
"vector",
]
],
"lazy_ir_inc": [
f'#include "{path}"'
for path in [node_base_hdr if node_base_hdr is not None else None]
if path is not None
],
"ir_declarations": list(
concat_map_codegen(
lazy_ir_generator(backend_indices[backend_key], node_base),
grouped_native_functions,
)
),
"namespace_prologue": ns_helper.prologue,
"namespace_epilogue": ns_helper.epilogue,
},
)
if not _valueT:
raise NotImplementedError(
"The value type needs to be set with setValueT() in run_gen_lazy_tensor()"
)
return _valueT
def setValueT(val: BaseCppType) -> None:
global _valueT
_valueT = val
# this is a bad hack. I need to refactor the data model to represent each arg in the schema as an object,
# making it easier to represent special properties of an arg.
tensorListValueT = BaseCppType("torch::lazy", "Value")
def process_ir_type(
typ: Type, properties: "LazyIrProperties"
) -> Union[BaseCType, VectorCType, OptionalCType, ListCType]:
"""
This function takes a type from NativeFunctions and converts it for use with
lazy tensor codegen.
Type conversion for lazy currently consists of
(1) changing at::Tensors into lazy::Values
(2) wrapping everything in a BaseCType
(3) making cpp-reference types into cpp-value types (e.g. vector instead of IntArrayRef)
(1) converts at::Tensors to lazy::Values (which wrap lazy::Nodes, with which Lazy IR represents tensors.)