def assertUfuncErrorInline(self, yaml_str: str, expect: str) -> None:
# parse a single structured group out of the yaml to g
es = yaml.load(yaml_str, Loader=LineLoader)
parsed_yaml = parse_native_yaml_struct(es)
native_functions, backend_indices = parsed_yaml.native_functions, parsed_yaml.backend_indices
grouped_native_functions = gen.get_grouped_native_functions(
native_functions)
assert len(grouped_native_functions) == 1
g = grouped_native_functions[0]
assert isinstance(g, NativeFunctionsGroup)
assert g.out.ufunc_inner_loop
# this is not ufunc codegen per se, but it does some basic sanity tests for
# ufunc generation
gen.compute_meta_function_declaration(g)
dest.compute_native_function_declaration(
g, backend_indices[DispatchKey.CPU])
dest.compute_native_function_declaration(
g, backend_indices[DispatchKey.CUDA])
try:
# the real kahuna
dest.compute_ufunc_cpu(g)
dest.compute_ufunc_cpu_kernel(g)
dest.compute_ufunc_cuda(g)
except AssertionError as e:
# hack to strip out the context
msg, _ = str(e).split(' in ', 2)
self.assertExpectedInline('\n'.join(textwrap.wrap(msg)),
expect,
skip=1)
return
self.fail(msg="Did not raise when expected to")
def run(source_yaml: str,
output_dir: str,
dry_run: bool,
impl_path: Optional[str] = None) -> None:
# Assumes that this file lives at PYTORCH_ROOT/tools/codegen/gen_backend_stubs.py
pytorch_root = pathlib.Path(__file__).parent.parent.parent.absolute()
template_dir = os.path.join(pytorch_root, "aten/src/ATen/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(
pytorch_root, 'aten/src/ATen/native/native_functions.yaml')
parsed_yaml = parse_native_yaml(native_yaml_path)
native_functions, backend_indices = parsed_yaml.native_functions, parsed_yaml.backend_indices
grouped_native_functions = get_grouped_native_functions(native_functions)
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
class_name = parsed_backend_yaml.class_name
backend_indices = parsed_backend_yaml.backend_indices
selector = SelectiveBuilder.get_nop_selector()
if backend_key is None:
# This could be useful if a backend wants to quickly set up a noop yaml file but doesn't have any kernels ready yet.
return
if class_name is None:
# class_name is an optional argument to backend yaml file.
# if specified it allows an external backend to override
# the name of the class that all generated kernel definitions live under.
# if not specified, its value is given as native_function_class_name.
class_name = backend_indices[backend_key].native_function_class_name()
assert class_name is not None
if impl_path is not None:
error_on_missing_kernels(native_functions, backend_indices,
backend_key, autograd_key, class_name,
impl_path)
gen_dispatchkey_nativefunc_headers(fm, class_name, cpp_namespace,
backend_indices,
grouped_native_functions, backend_key,
autograd_key)
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)
def main() -> None:
parser = argparse.ArgumentParser(description='Generate backend stub files')
parser.add_argument(
'-s',
'--source_yaml',
help='path to source yaml file containing operator external definitions')
parser.add_argument(
'-o', '--output_dir', help='output directory')
parser.add_argument(
'--dry_run', type=bool, default=False, help='output directory')
options = parser.parse_args()
# Assumes that this file lives at PYTORCH_ROOT/tools/codegen/gen_backend_stubs.py
pytorch_root = pathlib.Path(__file__).parent.parent.parent.absolute()
template_dir = os.path.join(pytorch_root, "aten/src/ATen/templates")
def make_file_manager(install_dir: str) -> FileManager:
return FileManager(install_dir=install_dir, template_dir=template_dir, dry_run=options.dry_run)
fm = make_file_manager(options.output_dir)
native_yaml_path = os.path.join(pytorch_root, 'aten/src/ATen/native/native_functions.yaml')
grouped_native_functions = get_grouped_native_functions(native_yaml_path)
cpp_namespace, external_backend_functions = parse_backend_yaml(options.source_yaml, grouped_native_functions)
native_functions = parse_native_yaml(native_yaml_path)
selector = SelectiveBuilder.get_nop_selector()
generated_comment = 'Autogenerated file by gen_backend_stubs.py. Do not edit directly!'
fm.write('aten_xla_type.h', lambda: {
'generated_comment': generated_comment,
'cpp_namespace': cpp_namespace,
'dispatch_xla_declarations': list(concatMap(dest.compute_native_function_declaration, external_backend_functions)),
})
fm.write('aten_xla_type_default.h', lambda: {
'generated_comment': generated_comment,
'cpp_namespace': cpp_namespace,
'dispatch_aten_fallback_declarations': list(concatMap(
dest.GenExternalAtenFallback(Target.NAMESPACED_DECLARATION), external_backend_functions
)),
})
fm.write('aten_xla_type_default.cpp', lambda: {
'generated_comment': generated_comment,
'cpp_namespace': cpp_namespace,
# TODO: after cpu fallbacks are moved to a boxed kernel,
# merge registrations / definitions into RegisterDispatchKey
'dispatch_aten_fallback_definitions': list(concatMap(
dest.GenExternalAtenFallback(Target.NAMESPACED_DEFINITION), external_backend_functions
)),
'dispatch_registrations': list(concatMap(
dest.GenExternalAtenFallback(Target.REGISTRATION), [e for e in external_backend_functions if not e.is_autograd_kernel]
)),
'dispatch_autograd_registrations': list(concatMap(
dest.GenExternalAtenFallback(Target.REGISTRATION), [e for e in external_backend_functions if e.is_autograd_kernel]
)),
})
def run(source_yaml: str, output_dir: str, dry_run: bool,
impl_path: Optional[str]) -> None:
# Assumes that this file lives at PYTORCH_ROOT/tools/codegen/gen_backend_stubs.py
pytorch_root = pathlib.Path(__file__).parent.parent.parent.absolute()
template_dir = os.path.join(pytorch_root, "aten/src/ATen/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(
pytorch_root, 'aten/src/ATen/native/native_functions.yaml')
parsed_yaml = parse_native_yaml(native_yaml_path)
native_functions, backend_indices = parsed_yaml.native_functions, parsed_yaml.backend_indices
grouped_native_functions = get_grouped_native_functions(native_functions)
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
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, impl_path)
gen_dispatchkey_nativefunc_headers(fm, class_name, cpp_namespace,
backend_indices,
grouped_native_functions,
backend_key, autograd_key)
for dispatch_key in [backend_key] if autograd_key is None else [
backend_key, autograd_key
]:
gen_dispatcher_registrations(fm, output_dir, cpp_namespace,
backend_indices,
grouped_native_functions, backend_key,
dispatch_key, selector)
def main() -> None:
parser = argparse.ArgumentParser(description='Generate ATen source files')
parser.add_argument(
'-s',
'--source-path',
help='path to source directory for ATen',
default='aten/src/ATen')
parser.add_argument(
'-p',
'--generated-ops-cpp-path',
help='path to directory to generate op dispatcher .cpp file',
default='torch/csrc/jit/runtime/static/generated_ops.cpp')
parser.add_argument(
'-t',
'--generated-ops-test-cpp-path',
help='path to directory to generate op dispatcher .cpp file',
default='benchmarks/static_runtime/test_generated_ops.cc')
options = parser.parse_args()
native_yaml_path = os.path.join(options.source_path, 'native/native_functions.yaml')
parsed_yaml = gen.parse_native_yaml(native_yaml_path)
native_functions, backend_indices = parsed_yaml.native_functions, parsed_yaml.backend_indices
grouped_native_functions = gen.get_grouped_native_functions(native_functions)
structured_native_functions = [g for g in grouped_native_functions
if isinstance(g, NativeFunctionsGroup)]
supported_function_groups = group_functions_by_op_name(structured_native_functions)
gen_out_variant_dispatcher = gen_structured.GenOutVariantDispatcher()
result = [gen_out_variant_dispatcher(groups) for groups in supported_function_groups]
gen_out_variant_dispatcher_test_case = gen_structured.GenOutVariantDispatcherTestCase()
test_result = [gen_out_variant_dispatcher_test_case(groups) for groups in supported_function_groups]
write_cpp(result, options.generated_ops_cpp_path)
write_test_cpp(test_result, options.generated_ops_test_cpp_path)
print("total grouped native ops: %d" % len(grouped_native_functions))
print("structured grouped native ops: %d" % len(structured_native_functions))
supported_grouped_functions = sum([len(groups) for groups in supported_function_groups])
print("generated grouped native ops: %d" % supported_grouped_functions)
def run(source_yaml: str, output_dir: str, dry_run: bool,
impl_path: Optional[str], gen_ts_lowerings: bool, node_base: str,
node_base_hdr: Optional[str], tensor_class: str,
tensor_class_hdr: str) -> None:
# Assumes that this file lives at PYTORCH_ROOT/tools/codegen/gen_backend_stubs.py
pytorch_root = pathlib.Path(__file__).parent.parent.parent.absolute()
template_dir = os.path.join(pytorch_root, "aten/src/ATen/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(
pytorch_root, 'aten/src/ATen/native/native_functions.yaml')
parsed_yaml = parse_native_yaml(native_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, impl_path,
full_codegen)
assert class_name is not None
# Generate nativefunction declarations
gen_dispatchkey_nativefunc_headers(fm, class_name, cpp_namespace,
backend_indices,
grouped_native_functions, backend_key,
autograd_key)
# 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, cpp_namespace,
backend_indices, grouped_native_functions,
backend_key, dispatch_key, selector)
# Generate native function impls that build IR nodes
fm.write_with_template(
f'{backend_key}NativeFunctions.cpp',
'DispatchKeyNativeFunctions.cpp',
lambda: {
'includes': [
f'#include <{path}>' for path in [
tensor_class_hdr,
"ATen/MetaFunctions.h",
"torch/csrc/lazy/core/metrics.h",
"torch/csrc/lazy/core/shape.h",
"lazy_tensor_core/csrc/aten_ltc_bridge.h",
"lazy_tensor_core/csrc/lazy_graph_executor.h",
f"{output_dir}/{backend_key}NativeFunctions.h",
f"{output_dir}/{backend_key}LazyIr.h",
f"{output_dir}/{backend_key}ShapeInference.h",
]
],
'native_functions_include':
'',
'backend_namespace':
'torch_lazy_tensors', # this is wrong
'native_function_definitions':
list(
concat_map_codegen(dest.GenLazyNativeFuncDefinition(
f'{backend_key}NativeFunctions', backend_indices[
backend_key], tensor_class),
grouped_native_functions,
codegenInplaceVariant=True)),
})
# Generate headers for shape/dtype funcs for non-meta kernels
fm.write_with_template(
f'{backend_key}ShapeInference.h', 'ShapeInference.h', lambda: {
'lazy_ir_sysinc': [
f'#include <{path}>' for path in [
"ATen/Tensor.h",
"c10/core/ScalarType.h",
"c10/util/Optional.h",
"torch/csrc/lazy/core/ir.h",
"torch/csrc/lazy/core/shape.h",
"vector",
]
],
'lazy_ir_inc': [],
'DispatchKey':
backend_key,
'dispatch_namespace':
backend_key.lower(),
'func_declarations':
list(
concat_map_codegen(
dest.GenLazyShapeInferenceDefinition(
backend_indices[backend_key], tensor_class),
grouped_native_functions)),
})
# Generate IR node classes
fm.write_with_template(
f'{backend_key}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",
"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
],
'external_backend_headers':
f'#include "{output_dir}/{backend_key}NativeFunctions.h"',
'namespaced_headers':
'',
'DispatchKey':
backend_key,
'dispatch_namespace':
backend_key.lower(),
'ir_declarations':
list(
concat_map_codegen(
dest.LazyIR(backend_indices[backend_key], node_base),
grouped_native_functions)),
})
def run(source_yaml: str, output_dir: str, dry_run: bool,
impl_path: Optional[str]) -> None:
# Assumes that this file lives at PYTORCH_ROOT/tools/codegen/gen_backend_stubs.py
pytorch_root = pathlib.Path(__file__).parent.parent.parent.absolute()
template_dir = os.path.join(pytorch_root, "aten/src/ATen/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(
pytorch_root, 'aten/src/ATen/native/native_functions.yaml')
parsed_yaml = parse_native_yaml(native_yaml_path)
native_functions, backend_indices = parsed_yaml.native_functions, parsed_yaml.backend_indices
grouped_native_functions = get_grouped_native_functions(native_functions)
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
selector = SelectiveBuilder.get_nop_selector()
# TODO: handle cases when yaml contains zero ops properly in a later PR.
if backend_key is not None and autograd_key is not None:
backend_dispatch_key: DispatchKey = backend_key
autograd_dispatch_key: DispatchKey = autograd_key
class_name = backend_indices[
backend_dispatch_key].native_function_class_name()
if impl_path is not None:
error_on_missing_kernels(native_functions, backend_indices,
backend_key, autograd_key, impl_path)
assert class_name is not None
generated_comment = 'Autogenerated file by gen_backend_stubs.py. Do not edit directly!'
fm.write_with_template(
f'{backend_dispatch_key}NativeFunctions.h',
'DispatchKeyNativeFunctions.h',
lambda: {
'generated_comment':
generated_comment,
'cpp_namespace':
cpp_namespace,
'class_name':
class_name,
# Convert to a set first to remove duplicate kernel names.
# Backends are allowed to repeat kernel names; only generate the declaration once!
'dispatch_declarations':
list(
set(
concatMap(
lambda f: dest.compute_native_function_declaration(
f, backend_indices[backend_dispatch_key]),
grouped_native_functions))) +
list(
set(
concatMap(
lambda f: dest.compute_native_function_declaration(
f, backend_indices[autograd_dispatch_key]),
grouped_native_functions))),
})
for dispatch_key in [backend_dispatch_key, autograd_dispatch_key]:
fm.write_with_template(
f'Register{dispatch_key}.cpp', 'RegisterDispatchKey.cpp',
lambda: {
'extra_cuda_headers':
'',
'external_backend_headers':
f'#include "{output_dir}/{backend_key}NativeFunctions.h"',
'namespaced_headers':
'',
'DispatchKey':
dispatch_key,
'dispatch_namespace':
dispatch_key.lower(),
'dispatch_helpers':
dest.gen_registration_helpers(backend_indices[dispatch_key]
),
'dispatch_namespaced_definitions':
list(
concatMap(
dest.RegisterDispatchKey(
backend_indices[dispatch_key],
Target.NAMESPACED_DEFINITION,
selector,
rocm=False,
cpp_namespace=cpp_namespace,
class_method_name=
f'{backend_dispatch_key}NativeFunctions'),
grouped_native_functions)),
'dispatch_anonymous_definitions':
list(
concatMap(
dest.RegisterDispatchKey(
backend_indices[dispatch_key],
Target.ANONYMOUS_DEFINITION,
selector,
rocm=False,
cpp_namespace=cpp_namespace,
class_method_name=
f'{backend_dispatch_key}NativeFunctions'),
grouped_native_functions)),
'dispatch_registrations':
list(
concatMap(
dest.RegisterDispatchKey(
backend_indices[dispatch_key],
Target.REGISTRATION,
selector,
rocm=False,
cpp_namespace=cpp_namespace,
class_method_name=
f'{backend_dispatch_key}NativeFunctions'),
grouped_native_functions)),
})
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_cls: Type[LazyIR] = default_args.lazy_ir_cls,
# 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) -> None:
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')
parsed_yaml = parse_native_yaml(native_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/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),
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_cls(backend_indices[backend_key], node_base),
grouped_native_functions)),
'namespace_prologue':
ns_helper.prologue,
'namespace_epilogue':
ns_helper.epilogue,
})
def run(source_yaml: str, output_dir: str, dry_run: bool) -> None:
# Assumes that this file lives at PYTORCH_ROOT/tools/codegen/gen_backend_stubs.py
pytorch_root = pathlib.Path(__file__).parent.parent.parent.absolute()
template_dir = os.path.join(pytorch_root, "aten/src/ATen/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(
pytorch_root, 'aten/src/ATen/native/native_functions.yaml')
parsed_yaml = parse_native_yaml(native_yaml_path)
native_functions, backend_indices = parsed_yaml.native_functions, parsed_yaml.backend_indices
grouped_native_functions = get_grouped_native_functions(native_functions)
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
selector = SelectiveBuilder.get_nop_selector()
# TODO: handle cases when yaml contains zero ops properly in a later PR.
if backend_key is not None and autograd_key is not None:
backend_dispatch_key: DispatchKey = backend_key
autograd_dispatch_key: DispatchKey = autograd_key
generated_comment = 'Autogenerated file by gen_backend_stubs.py. Do not edit directly!'
fm.write(
'aten_xla_type.h',
lambda: {
'generated_comment':
generated_comment,
'cpp_namespace':
cpp_namespace,
# Convert to a set first to remove duplicate kernel names.
# Backends are allowed to repeat kernel names; only generate the declaration once!
'dispatch_xla_declarations':
list(
set(
concatMap(
lambda f: dest.compute_native_function_declaration(
f, backend_indices[backend_dispatch_key]),
grouped_native_functions))) +
list(
set(
concatMap(
lambda f: dest.compute_native_function_declaration(
f, backend_indices[autograd_dispatch_key]),
grouped_native_functions))),
})
external_backend_headers = '''\
#include <tensorflow/compiler/xla/xla_client/debug_macros.h>
#include <tensorflow/compiler/xla/xla_client/metrics.h>
#include <tensorflow/compiler/xla/xla_client/tf_logging.h>
#include <torch_xla/csrc/function_call_tracker.h>
#include <torch_xla/csrc/aten_xla_type.h>
#include <torch_xla/csrc/aten_xla_type_default.h>'''
for dispatch_key in [backend_dispatch_key, autograd_dispatch_key]:
fm.write_with_template(
f'Register{dispatch_key}.cpp', 'RegisterDispatchKey.cpp',
lambda: {
'extra_cuda_headers':
'',
'legacy_th_headers':
'',
'external_backend_headers':
external_backend_headers,
'DispatchKey':
dispatch_key,
'dispatch_namespace':
dispatch_key.lower(),
'dispatch_namespaced_definitions':
list(
concatMap(
dest.RegisterDispatchKey(
backend_indices[dispatch_key],
Target.NAMESPACED_DEFINITION,
selector,
rocm=False,
cpp_namespace=cpp_namespace),
grouped_native_functions)),
'dispatch_anonymous_definitions':
list(
concatMap(
dest.RegisterDispatchKey(
backend_indices[dispatch_key],
Target.ANONYMOUS_DEFINITION,
selector,
rocm=False,
cpp_namespace=cpp_namespace),
grouped_native_functions)),
'dispatch_registrations':
list(
concatMap(
dest.RegisterDispatchKey(
backend_indices[dispatch_key],
Target.REGISTRATION,
selector,
rocm=False,
cpp_namespace=cpp_namespace),
grouped_native_functions)),
})
fm.write(
'aten_xla_type_default.h', lambda: {
'generated_comment':
generated_comment,
'cpp_namespace':
cpp_namespace,
'dispatch_aten_fallback_declarations':
list(
concatMap(
dest.GenExternalAtenFallback(
Target.NAMESPACED_DECLARATION, backend_indices[
backend_dispatch_key]),
grouped_native_functions)),
})
fm.write(
'aten_xla_type_default.cpp',
lambda: {
'generated_comment':
generated_comment,
'cpp_namespace':
cpp_namespace,
# TODO: after cpu fallbacks are moved to a boxed kernel,
# merge registrations / definitions into RegisterDispatchKey
'dispatch_aten_fallback_definitions':
list(
concatMap(
dest.GenExternalAtenFallback(
Target.NAMESPACED_DEFINITION, backend_indices[
backend_dispatch_key]),
grouped_native_functions)),
'dispatch_registrations':
list(
concatMap(
dest.GenExternalAtenFallback(
Target.REGISTRATION, backend_indices[
backend_dispatch_key]),
grouped_native_functions)),
})
