import itertools import logging import operator import os import re import sys import time from collections import defaultdict from contextlib import contextmanager from typing import Any, Callable, DefaultDict, Dict, List, Optional, Set, Tuple import sympy import torch import torch._logging import torch.fx from torch._decomp import get_decompositions from torch._dynamo.utils import defake, dynamo_timed from torch._logging import LazyString, trace_structured from torch._subclasses.fake_tensor import FakeTensor from torch.fx.experimental._backward_state import BackwardState from torch.fx.experimental.sym_node import magic_methods, method_to_operator from torch.fx.experimental.symbolic_shapes import has_free_symbols, ShapeEnv, SymTypes from torch.utils._mode_utils import no_dispatch from . import config, ir from .codegen.common import ( DeviceOpOverrides, get_device_op_overrides, get_scheduling_for_device, get_wrapper_codegen_for_device, register_backend_for_device, ) from .codegen.cpp_wrapper_cpu import CppWrapperCpu from .codegen.cpp_wrapper_cuda import CppWrapperCuda from .codegen.wrapper import WrapperCodeGen from .exc import ( CppWrapperCodeGenError, LoweringException, MissingOperatorWithDecomp, MissingOperatorWithoutDecomp, ) from .ir import ( Constant, FixedLayout, InputBuffer, Pointwise, Reduction, StorageBox, TensorBox, ) from .lowering import ( constrain_to_fx_strides, FALLBACK_ALLOW_LIST, fallback_handler, fallback_node_due_to_unsupported_type, layout_constraints, lowerings, make_fallback, needs_realized_inputs, unsupported_output_tensor, ) from .sizevars import SizeVarAllocator from .utils import convert_shape_to_inductor, gather_origins, get_sympy_Expr_dtype from .virtualized import V log = logging.getLogger(__name__) perf_hint_log = torch._logging.getArtifactLogger(__name__, "perf_hints") output_code_log = torch._logging.getArtifactLogger(__name__, "output_code") if config.is_fbcode(): from torch._inductor.fb.utils import log_module_code else: def log_module_code(*args, **kwargs): pass def supported_dtype_of_cpp_wrapper(dtype, cuda): supported_dtype = { torch.float32, torch.float64, torch.int64, torch.int32, torch.int16, torch.int8, torch.uint8, torch.bool, torch.bfloat16, torch.complex32, torch.complex64, torch.complex128, torch.float16, } if cuda: supported_dtype.add(torch.float8_e4m3fn) supported_dtype.add(torch.float8_e5m2) supported_dtype.add(torch.float8_e4m3fnuz) supported_dtype.add(torch.float8_e5m2fnuz) return dtype in supported_dtype def may_get_constant_buffer_dtype(constant_buffer): assert isinstance( constant_buffer, (sympy.Symbol, sympy.Expr, sympy.core.numbers.Integer) ), "get_constant_buffer_dtype only supports input of sympy.Symbol, sympy.Expr or sympy.core.numbers.Integer" if isinstance(constant_buffer, sympy.core.numbers.Integer): return torch.int64 if isinstance(constant_buffer, sympy.Expr): return get_sympy_Expr_dtype(constant_buffer) if constant_buffer.is_integer: return torch.int64 elif constant_buffer.is_float: return torch.float32 else: return None def is_magic_method(op): magic_ops = {method_to_operator(m) for m in magic_methods} return op in magic_ops def getattr_recursive(obj, target): target_atoms = target.split(".") attr_itr = obj for i, atom in enumerate(target_atoms): if not hasattr(attr_itr, atom): raise RuntimeError( f"Node referenced nonexistent target {'.'.join(target_atoms[:i])}" ) attr_itr = getattr(attr_itr, atom) return attr_itr class GraphLowering(torch.fx.Interpreter): graph_outputs: List[ir.IRNode] def symbolic_sizes_strides(self, ex: torch.Tensor): """ Support dynamic shapes and dynamic strides by assigning variables to each dimension. We duck-shape tensors, so if two tensors have the same size they get assigned the same symbolic variable. """ if self.reuse_shape_env: return convert_shape_to_inductor(ex.size()), convert_shape_to_inductor( ex.stride() ) else: from torch._dynamo.source import ConstantSource # TODO: this should not be needed once #93059 lands # https://github.com/pytorch/pytorch/pull/94031#discussion_r1096044816 # TODO: make a dedicated UnknownSource for this? # NB: This is using the legacy default behavior from # create_symbolic_sizes_strides_storage_offset but we hope we can # just delete this entirely source = ConstantSource( f"__inductor_unknown_tensor_{len(self._shape_env.var_to_val)}" ) ( size, stride, _, ) = self._shape_env.create_symbolic_sizes_strides_storage_offset( ex, source, ) size = [i.node.expr if isinstance(i, torch.SymInt) else i for i in size] stride = [i.node.expr if isinstance(i, torch.SymInt) else i for i in stride] return size, stride def static_sizes_strides(self, ex: torch.Tensor): """ Primarily used to weights """ size = [sympy.Integer(i) for i in ex.size()] stride = [sympy.Integer(i) for i in ex.stride()] return size, stride def init_backend_registration(self): if get_scheduling_for_device("cpu") is None: from .codegen.cpp import CppScheduling register_backend_for_device("cpu", CppScheduling, WrapperCodeGen) if get_scheduling_for_device("cuda") is None: from .codegen.cuda_combined_scheduling import CUDACombinedScheduling # CUDACombinedScheduling combines Triton and CUDA C++ scheduling for CUDA devices via delegation register_backend_for_device("cuda", CUDACombinedScheduling, WrapperCodeGen) def __init__( self, gm: torch.fx.GraphModule, example_inputs: Optional[List[torch.Tensor]] = None, shape_env=None, num_static_inputs=None, graph_id=None, cpp_wrapper=False, aot_mode=False, user_visible_outputs=frozenset(), layout_opt=None, extern_node_serializer=None, is_inference=False, is_const_graph=False, const_output_index=None, const_code=None, const_module=None, name=None, ): super().__init__(gm) self.example_inputs = example_inputs self.layout_opt = ( layout_opt if layout_opt is not None else self.decide_layout_opt(gm, is_inference=is_inference) ) self.num_channels_last_conv = 0 self.is_inference = is_inference self.is_const_graph = is_const_graph self.const_code = const_code self.const_module = const_module self.extra_traceback = False # we do our own error wrapping if shape_env is None: shape_env = ShapeEnv() self.reuse_shape_env = False else: self._shape_env = shape_env self.reuse_shape_env = True self._shape_env = shape_env self.sizevars = SizeVarAllocator(shape_env) self.graph_input_names: List[str] = [] self.graph_inputs: Dict[str, TensorBox] = {} self.graph_inputs_original: Dict[str, InputBuffer] = {} self.device_types: Set[str] = ( const_module.device_types if const_module else set() ) self.device_idxs: Set[int] = const_module.device_idxs if const_module else set() self.cuda = False self.buffers: List[ir.Buffer] = [] self.const_output_index: Dict[str, int] = ( const_output_index if const_output_index else {} ) self.folded_constants: Set[str] = ( set(const_output_index.keys()) if const_output_index else set() ) self.constants: Dict[str, torch.Tensor] = ( const_module.constants if const_module else {} ) self.constant_reprs: Dict[str, str] = {} self.removed_buffers: Set[str] = set() self.removed_inplace_buffers: Set[str] = set() self.mutated_buffers: Set[str] = set() self.never_reuse_buffers: Set[str] = set() self.inplaced_to_remove: Set[str] = set() self.device_ops: DeviceOpOverrides = None # type: ignore[assignment] self.wrapper_code: WrapperCodeGen = None # type: ignore[assignment] # See `ProxyExecutor Design Note` in ir.py for more details self.extern_kernel_nodes: List[ir.ExternKernelNode] = [] self.extern_node_serializer: Optional[ Callable[[List[ir.ExternKernelNode]], Any] ] = extern_node_serializer self.current_node: torch.fx.Node = None # type: ignore[assignment] self.num_static_inputs = num_static_inputs self.lists: Dict[str, List[str]] = {} self.mutated_inputs: Set[str] = set() self.mutated_input_idxs: List[int] = [] self.name_to_buffer: Dict[str, ir.Buffer] = {} self.name_to_users: DefaultDict[str, List[ir.IRNode]] = defaultdict(list) self.creation_time = time.time() self.name = name self.cpp_wrapper = cpp_wrapper # record multi_kernel choice for cpp_wrapper so the second pass knows # which sub-kernel is picked. Copy cpp_wrapper to another variable # since cpp_wrapper flag is set to false for the first pass of codegen. self.record_multi_kernel_choice = cpp_wrapper self.multi_kernel_to_choice: Dict[str, int] = {} self.aot_mode = aot_mode self.graph_id = graph_id self.scheduler: "torch._inductor.scheduler.Scheduler" = None # type: ignore[assignment] self.nodes_prefer_channels_last = ( self.find_nodes_prefer_channels_last() if self.layout_opt else set() ) self._warned_fallback = {"aten.convolution_backward"} self.user_visible_outputs = user_visible_outputs self.cache_key: str = "" # This is the cache key for the compiled artifact self.cache_path: str = "" # This is the path in the filesystem where the compiled artifact is stored self.cache_linemap: List[ Tuple[int, str] ] = ( [] ) # This is the linemap used by the profiler to mark custom compiled kernels getting run # Used if lowering encounters cases where cudagraphs are not supported self.disable_cudagraphs_reason: Optional[str] = None # only keeping one node per device for stack trace purposes self.device_node_mapping: Dict[torch.device, torch.fx.Node] = {} self.orig_gm: torch.fx.GraphModule = gm.__copy__() self.dynamo_flat_name_to_original_fqn = self.module.meta.get( "dynamo_flat_name_to_original_fqn", {} ) self.allocated_constant_name = ( const_module.allocated_constant_name if const_module is not None else {} ) self.init_backend_registration() @staticmethod def decide_layout_opt(gm, *, is_inference) -> bool: """ Decide if we should enable layout optimization for this graph based on heuristics. """ if not config.layout_optimization: return False if config.force_layout_optimization: return True conv_nodes = [ n for n in gm.graph.nodes if n.target == torch.ops.aten.convolution.default ] nconv = len(conv_nodes) if nconv == 0: return False # For cpu backend and mkldnn enabled, we always use channels_last for better performance. if ( torch.backends.mkldnn.enabled and torch.backends.mkldnn.is_available() and all( n.args[idx].meta["val"].device == torch.device("cpu") for n in conv_nodes for idx in [0, 1] ) ): return True # Following models are skipped due to this: # jx_nest_base # volo_d1_224 if len(list(gm.graph.nodes)) >= 300 * nconv: log.debug("Skipped layout opt because only a few conv") return False if any( has_free_symbols(n.args[idx].meta["val"]) for n in conv_nodes for idx in [0, 1] ): log.debug( "See perf regression with dynamic shape. Follow up in https://github.com/pytorch/pytorch/issues/102670" ) return False def is_grouped(n): return n.args[-1] > 1 and n.args[1].meta["val"].size(1) > 1 def is_in_out_channel(n): return ( n.args[1].meta["val"].size(0) * 2 <= n.args[1].meta["val"].size(1) and n.args[1].meta["val"].size(2) > 1 ) def is_small_channel(n): return ( n.args[1].meta["val"].size(0) <= 64 and n.args[1].meta["val"].size(1) <= 64 ) # only grouped convolutions benchmarked as slower in conv samples for inference only if is_inference: from torch.utils.flop_counter import FlopCounterMode flop_counts: Dict[str, float] = defaultdict(float) for node in conv_nodes: success, args, kwargs = torch._inductor.fx_utils.get_fake_args_kwargs( node ) if success: with FlopCounterMode(display=False) as flop_counter_mode: with V.fake_mode: node.target(*args, **kwargs) counted_flops = flop_counter_mode.get_total_flops() if is_grouped(node): node_type = "grouped" elif is_small_channel(node): node_type = "small" elif is_in_out_channel(node): node_type = "in_out" else: node_type = "default" flop_counts[node_type] += counted_flops else: log.debug("Conv inputs meta not found") # average benchmarked channels last speedup / slowdown, < 1 is speedup. # taken from the set of convolution inputs in benchmarks/dynamo/microbenchmarks/operator_inp_logs/torchbench_train/ # To regenerate these numbers follow https://gist.github.com/eellison/55d7a6ed6f39829d68ac56f95f4df5bb GROUPED_MULTIPLIER = 1.358 DEFAULT_MULTIPLIER = 0.823 IN_OUT_MULTIPLIER = 0.725 SMALL_MULTIPLIER = 0.783 total_flops = sum(flop_counts.values()) # TODO - get different values per hardware weighted_flops = ( flop_counts["grouped"] * GROUPED_MULTIPLIER + flop_counts["small"] * SMALL_MULTIPLIER + flop_counts["in_out"] * IN_OUT_MULTIPLIER + flop_counts["default"] * DEFAULT_MULTIPLIER ) do_layout_opt = weighted_flops <= total_flops if not do_layout_opt: log.debug( "Skipped layout opt in inference because weighted flops indicate slowdown, default: %d, channels last: %d", total_flops, weighted_flops, ) return do_layout_opt # Channels last layout can dramatically hurt grouped conv perf. E.g. # Conv with arguments like # {"input_shape": [32, 224, 112, 112], "weight_shape": [224, 112, 3, 3], # "stride": [2, 2], "padding": [1, 1], "groups": 2} # slows down 31x using channels last.. # But a lot of timm models use depthwise separable convolution which will # result in grouped convolution with in-channel size == 1. # For those grouped convolution, channels last still helps a lot. # E.g. # Conv with arguments # {"input_shape": [128, 58, 56, 56], "weight_shape": [58, 1, 3, 3], # "stride": [2, 2], "padding": [1, 1], "groups": 58} # get 1.86x speedup with channels last layout. # # The following heuristics skip using channels-last if the model contains # grouped convolution with in-channels > 1. if any(map(is_grouped, conv_nodes)): log.debug( "Skip layout opt because found grouped convolution with >1 in_channels!" ) return False # For some models that contain convolution with larger in-channel than out-channel, applying # channels last hurts performance. # Following models are skipped due to this: # - pytorch_unet # - phlippe_densenet (slightly worse) # - Background_Matting (1.22x -> 0.821x) # - pytorch_CycleGAN_and_pix2pix (1.597x -> 1.294x) if any(map(is_in_out_channel, conv_nodes)): log.debug( "Skip layout opt because some convolutions have smaller out_channel" ) return False # Following models are skipped due to this: # - functorch_maml_omniglot if all(map(is_small_channel, conv_nodes)): log.debug("Skip layout opt because all convolution channels are too small") return False return True def qualify_name(self, name: str) -> str: """Prepend the given name with the graph name if any.""" if self.name is not None: return f"{self.name}_{name}" return name def make_subgraph( self, gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor], subgraph_name: str, ) -> "GraphLowering": """ Make a subgraph of the current graph with all inherited parts, except the graph module (`gm`) and `example_inputs`. The subgraphs are lowered separately, but intended to be inlined in the parent graph's codegening. Hence the need for maintaining the same `shape_env` and other properties. The subgraph name is qualified by the parent graph's name. """ return GraphLowering( gm=gm, example_inputs=example_inputs, shape_env=self._shape_env, cpp_wrapper=self.cpp_wrapper, aot_mode=self.aot_mode, extern_node_serializer=self.extern_node_serializer, is_inference=self.is_inference, name=self.qualify_name(subgraph_name), ) def find_nodes_prefer_channels_last(self): """ The rule to decide if an node prefer channels last is simple. 1. if it's input/output of a convolution 2. if one of its user prefers channels last We have rule 1 because cudnn runs a faster convolution kernel for channels last inputs; Rule 2 is also important. It makes sure that indirect inputs to convolution also prefers channels last. Consider the scenario: conv -> batch-norm -> relu -> conv Without rule 2, batch-norm output may use a contiguous layout. That will cause 2 extra copies: 1. the output of batch-norm should be channels last initially since its input is a conv's output. Forcing the batch-norm's output to be contiguous results in the first copy 2. The second conv's input is initially contiguous. This layout is propagated from the batch-norm's output. We need convert it to channels last layout which results in the second copy. With rule 2, we makes sure all the tensors in the chain uses channels last layout. So both copies can be saved. """ output_set = set() for n in reversed(self.module.graph.nodes): if n.target == torch.ops.aten.convolution.default: output_set.add(n) continue for user in n.users: if user in output_set: output_set.add(n) break # need a second pass to add downstream nodes of those channel last nodes to the sets. # This pass is especially needed to avoid mix-layout kernel inputs in backward pass. # # Let's say a conv-batchnorm 's output is passed to relu whose output is in turn returned # from the fwd graph. Without this second pass, we will force relu's output to be contiguous. # Then in the kernel in backward pass, the contiguous output of relu may be mix with other channels last # tensors and passed to a kernel. # # This pass improve yolov3 training speedup from 1.116x (worse than disabling layout optimization speedup 1.196x) to 1.457x. # It also improves dla102 training speedup from 1.240x (worse than disabling layout optimization speedup 1.523x) to 1.835x . # This also helps the following models: # - res2net101_26w_4s # - res2net50_14w_8s # - sebotnet33ts_256 for n in self.module.graph.nodes: if n in output_set: for child in n.users: output_set.add(child) return output_set def warn_fallback(self, name): if name not in self._warned_fallback: self._warned_fallback.add(name) perf_hint_log.info("Using FallbackKernel: %s", name) def add_device_info(self, device: torch.device): self.device_types.add(device.type) if device.index is not None: self.device_idxs.add(device.index) if V.graph.current_node and device not in self.device_node_mapping: self.device_node_mapping[device] = V.graph.current_node @property def fake_mode(self): return V.fake_mode def get_buffer(self, buffer_name: str): if buffer_name in self.name_to_buffer: return self.name_to_buffer[buffer_name] if buffer_name in self.graph_inputs: return self.graph_inputs[buffer_name] return None def get_dtype(self, buffer_name: str): if buffer_name in self.constants: return self.constants[buffer_name].dtype if buffer_name in self.name_to_buffer: return self.name_to_buffer[buffer_name].get_dtype() if buffer_name in self.graph_inputs: return self.graph_inputs[buffer_name].get_dtype() m = re.match(r"(as_strided|reinterpret_tensor)\(([a-zA-Z0-9_]+),", buffer_name) if m: return self.get_dtype(m.group(1)) raise KeyError(f"could not find {buffer_name}") def get_numel(self, buffer_name: str): from .ir import MultiOutputLayout if buffer_name in self.constants: return self.constants[buffer_name].numel() if buffer_name in self.name_to_buffer: buf = self.name_to_buffer[buffer_name] if isinstance(getattr(buf, "layout", None), MultiOutputLayout): return 1 return buf.get_numel() if buffer_name in self.graph_inputs: return self.graph_inputs[buffer_name].get_numel() raise KeyError(f"could not find {buffer_name}") @dynamo_timed def run(self, *args): return super().run(*args) def register_buffer(self, buffer: ir.Buffer): name = self.qualify_name(f"buf{len(self.buffers)}") self.buffers.append(buffer) self.name_to_buffer[name] = buffer # Skip empty CPU tensor so that CUDA graphs can succeed, see https://github.com/pytorch/pytorch/pull/114144 if not isinstance(buffer, ir.ComputedBuffer) or not buffer.is_zero_elements(): self.add_device_info(buffer.get_device()) return name def register_list(self, buffer_names: List[str]): name = self.qualify_name("list_" + "_".join(buffer_names)) self.lists[name] = buffer_names return name def register_users_of(self, node_output): def register(value): if isinstance(value, (list, tuple)): for x in value: register(x) if isinstance(value, ir.IRNode): if ( not hasattr(value, "data") or not isinstance(value.data, ir.IRNode) or not ( hasattr(value.data, "data") and isinstance(value.data.data, ir.IRNode) ) ): return for read_name in value.get_read_names(): self.name_to_users[read_name].append(value) register(node_output) def mark_buffer_mutated(self, name: str): """ When a buffer is mutated we need to make sure all the reads to the old version are realized before the mutation happens. """ assert isinstance(name, str) self.mutated_buffers.add(name) if name not in self.name_to_users: return for user in self.name_to_users[name]: user.realize() def add_tensor_constant(self, data, name=None): def allocate(name): if not config.aot_inductor.use_runtime_constant_folding: for constant_name, value in self.constants.items(): if ( not data.is_mkldnn and data.size() == value.size() and data.stride() == value.stride() and data.dtype == value.dtype and data.device == value.device and torch.eq(data, value).all() ): return constant_name if name is None: name = f"constant{len(self.constants)}" if name[0].isdigit(): name = f"constant_{name}" name = self.qualify_name(name) # We may generate a var name for each constant in the codegen. # Let's only keep sane characters. prefix = re.sub(r"[^a-zA-Z0-9_]", "_", name) name = prefix cnt = 0 while name in self.constants: name = f"{prefix}_{cnt}" cnt += 1 self.constants[name] = data self.constant_reprs[name] = ( f"{data.device!r} {data.dtype!r} " f"{tuple(data.size())!r} {tuple(data.stride())!r} " f"{hash(data):x}" ) return name new_name = allocate(name) self.allocated_constant_name[new_name] = name return TensorBox.create( ir.ConstantBuffer( new_name, FixedLayout(data.device, data.dtype, *self.static_sizes_strides(data)), ) ) def constant_name(self, name: str, device_override: Optional[torch.device]): """ We AOT copy constants to the devices they are needed on. If device_override doesn't match the constant's device, then copy it and return a different name. """ if self.constants[name].device == device_override or device_override is None: return name alt_name = f"{name}_{device_override.type}{device_override.index or 0}" if alt_name not in self.constants: self.constants[alt_name] = self.constants[name].to(device_override) return alt_name def placeholder(self, target: str, args, kwargs): example = super().placeholder(target, args, kwargs) self.graph_input_names.append(target) if isinstance(example, SymTypes): expr = example.node.expr self.graph_inputs[target] = expr return expr elif isinstance(example, (int, bool, float)): expr = sympy.sympify(example) self.graph_inputs[target] = expr return expr if isinstance(example, BackwardState): # Ignored arg, must be unused # Alternately we could filter this out in AotAutograd return None assert isinstance(example, torch.Tensor), example # todo(chilli): We can remove the last check once we turn buffers into # static shape tensors. That's a hack to workaround Inductor believing # the buffer should be static but us passing in a fake tensor with # symbolic shapes. if not example._has_symbolic_sizes_strides: # the first N inputs are weights sizes, strides = self.static_sizes_strides(example) else: sizes, strides = self.symbolic_sizes_strides(example) # TODO(jansel): handle input aliasing target = self.qualify_name(target) tensor = TensorBox.create( InputBuffer( target, FixedLayout(example.device, example.dtype, sizes, strides), ) ) self.graph_inputs[target] = tensor self.graph_inputs_original[target] = tensor.data.data self.add_device_info(example.device) return tensor def call_function(self, target, args, kwargs): if target is operator.getitem and isinstance(args[0], (list, tuple, dict)): return super().call_function(target, args, kwargs) if hasattr(target, "_inductor_lowering_function"): # passthrough lowerings from .pattern_matcher return target(*args, **kwargs) def get_custom_op_layout_constraints(target, args, kwargs): # Custom operations that require preserving stride order # which run through implicit fallback must constrain their # arguments' fx strides layout_constraint = None if torch._C.Tag.needs_fixed_stride_order in target.tags: # We have to set the current args because call_function will immediately # evaluate this lowering after creating the fallback, without evaluating # the layout constraint args, kwargs = constrain_to_fx_strides( self.current_node, *args, **kwargs ) # Also register the layout constraint so when the fallback # is used again, we can constrain the args to the same layout layout_constraint = constrain_to_fx_strides return layout_constraint, args, kwargs if target not in lowerings: assert isinstance( target, torch._ops.OpOverload ), f"{target} is not an OpOverload" base_name = target.name().split(".")[0] if base_name in FALLBACK_ALLOW_LIST: make_fallback(target) elif config.implicit_fallbacks: layout_constraint, args, kwargs = get_custom_op_layout_constraints( target, args, kwargs ) error = ( MissingOperatorWithDecomp if get_decompositions([target]) else MissingOperatorWithoutDecomp ) log.info( "Creating implicit fallback for:\n%s", error.operator_str(target, args, kwargs), ) make_fallback(target, layout_constraint) elif get_decompositions([target]): # There isn't a good way to dynamically patch this in # since AOT Autograd already ran. The error message tells # the user how to fix it. raise MissingOperatorWithDecomp(target, args, kwargs) else: raise MissingOperatorWithoutDecomp(target, args, kwargs) try: log.debug(" via %s", lowerings[target]) out = lowerings[target](*args, **kwargs) return out except Exception as e: raise LoweringException(e, target, args, kwargs).with_traceback( e.__traceback__ ) from None @staticmethod def can_inline_constant(t: torch.Tensor) -> bool: """ True if this is a small constant attr that will be inlined. """ return len(t.shape) == 1 and t.shape[0] <= 8 def get_attr(self, target, args, kwargs): # this is a constant value = getattr_recursive(self.module, target) if isinstance(value, torch.fx.GraphModule): return ir.Subgraph(name=target, graph_module=value) if ( config.aot_inductor.use_runtime_constant_folding or config.always_keep_tensor_constants or unsupported_output_tensor(value) ): return self.add_tensor_constant(value, target) with no_dispatch(): if value.shape == (): return Constant(value.item(), value.dtype, value.device) if self.can_inline_constant(value): # tensor lowering has constant inlining logic from .lowering import tensor return tensor(value.tolist(), dtype=value.dtype, device=value.device) return self.add_tensor_constant(value, target) def call_module(self, target, args, kwargs): raise AssertionError() def call_method(self, target, args, kwargs): raise AssertionError() def output(self, target, args, kwargs): result = super().output(target, args, kwargs) assert isinstance(result, (tuple, list)), type(result) assert all( isinstance( x, ( TensorBox, ir.Constant, type(None), ir.ConstantBuffer, sympy.Expr, sympy.logic.boolalg.Boolean, int, ), ) for x in result ), result self.graph_outputs = [ir.ExternKernel.realize_input(x) for x in result] value: ir.IRNode for name, value in self.graph_inputs.items(): assert isinstance( value, (TensorBox, sympy.Expr) ), f"Unsupported inductor graph input type: {type(value)}" if not isinstance(value, TensorBox): continue value.realize() assert isinstance(value, TensorBox) value = value.data assert isinstance(value, ir.StorageBox) value_storage_box = value value = value.data if not isinstance(value, InputBuffer) or value.get_name() != name: # one of our inputs was mutated, need to turn that into a copy ir.MutationLayout.realize_into(value, self.graph_inputs_original[name]) # replace output with mutated input try: ind = self.graph_outputs.index(value_storage_box) self.graph_outputs[ind] = self.graph_inputs_original[name] except ValueError: pass self.finalize() log.debug( "Force channels last inputs for %d conv for the current graph with id %d", self.num_channels_last_conv, self.graph_id if self.graph_id is not None else -1, ) def finalize(self): for buf in self.buffers: buf.decide_layout() @contextmanager def set_current_node(self, node: torch.fx.Node): old = self.current_node try: self.current_node = node yield finally: self.current_node = old def run_node(self, n: torch.fx.Node): def debug(msg): log.debug("lowering %s %s", LazyString(n.format_node), msg) origins = {n} if n.op == "call_function": args, kwargs = self.fetch_args_kwargs_from_env(n) origins |= gather_origins(args, kwargs) with ir.IRNode.current_origins(origins), self.set_current_node( n ), V.set_current_node(n): if ( n.op == "call_function" and n.target is not operator.getitem and fallback_node_due_to_unsupported_type(n) ): debug("fallback_handler") result = fallback_handler(n.target, add_to_fallback_set=False)( *args, **kwargs # type: ignore[possibly-undefined] ) elif n.op == "call_function" and n.target in layout_constraints: debug("layout_constraints") args, kwargs = layout_constraints[n.target](n, *args, **kwargs) # type: ignore[index] result = self.call_function(n.target, args, kwargs) elif is_magic_method(n.target): # TODO: this is sus, it probably should be handled in the # lowerings themselves similarly to sym_size/sym-stride debug("is_magic_method") if isinstance(n.meta["val"], torch.SymInt): result = n.meta["val"].node.expr else: result = super().run_node(n) else: debug("") result = super().run_node(n) # require the same stride order for dense outputs, # 1. user-land view() will not throw because inductor # output different strides than eager # long term the solution is to make view() always succeed # with infallible strides. # 2: as_strided ops, we need make sure its input has same size/stride with # eager model to align with eager behavior. as_strided_ops = [ torch.ops.aten.as_strided.default, torch.ops.aten.as_strided_.default, torch.ops.aten.as_strided_scatter.default, ] is_output = any(user.op == "output" for user in n.users) is_input_for_as_strided = any( user.target in as_strided_ops for user in n.users ) if ( is_output and isinstance(result, TensorBox) and isinstance(result.data, ir.BaseView) ): # Realize so that outputs are correctly aliased result.realize() if (is_output or is_input_for_as_strided) and isinstance( n.meta["val"], torch.Tensor ): strides = n.meta["val"].stride() dense = torch._prims_common.is_non_overlapping_and_dense(n.meta["val"]) # requiring a stride order for a non-dense output wouldn't # recreate the same strides, and would fail with view, defer for now. if dense and len(strides): stride_order = ir.get_stride_order(strides) if ( len(result.get_size()) == 4 and n in self.nodes_prefer_channels_last and n.name not in self.user_visible_outputs and not is_input_for_as_strided ): stride_order = ir.NHWC_STRIDE_ORDER result = ir.ExternKernel.require_stride_order(result, stride_order) # Realize if (1) any user need inputs realized, or (2) there is # already too many reads and rematerializing can be bad. num_users = len(set(n.users)) if num_users > 1 and isinstance(result, TensorBox): for user in n.users: if user.target in needs_realized_inputs: result.realize_hint() # This inclusion is somewhat controversial (from # discussion between Horace, Natalia, and Elias). # Currently, it's not very clear why this is helpful. # The general idea here is that even though a node may # have FlexibleLayout, we still often *treat* it as if # it was contiguous. This appears to sometimes result in # suboptimal behavior. # # When we do a better job selecting layout, we should # revisit this. need_fixed_layout = [ torch.ops.aten.convolution_backward.default, torch.ops.aten.mm.default, torch.ops.aten._int_mm.default, ] if not self.layout_opt: need_fixed_layout.append(torch.ops.aten.convolution.default) if torch._C._has_mkldnn: need_fixed_layout += [ torch.ops.mkldnn._convolution_pointwise.default, torch.ops.mkldnn._convolution_pointwise.binary, torch.ops.mkldnn._convolution_pointwise_.binary, torch.ops.mkldnn._convolution_transpose_pointwise.default, torch.ops.mkldnn._linear_pointwise.default, torch.ops.mkldnn._linear_pointwise.binary, torch.ops.aten.mkldnn_rnn_layer.default, torch.ops.onednn.qconv2d_pointwise.default, torch.ops.onednn.qconv2d_pointwise.binary, torch.ops.onednn.qlinear_pointwise.default, torch.ops.onednn.qlinear_pointwise.tensor, ] if torch._C.has_mkl: need_fixed_layout += [torch.ops.mkl._mkl_linear.default] if user.target in need_fixed_layout: result = ir.ExternKernel.require_stride_order( result, ir.get_stride_order(n.meta["val"].stride()) ) if user.op == "output": if isinstance(result.data.data, (Pointwise, Reduction)): result.realize() # TODO(jansel): introduce a store vs inline choice result.mark_reuse(len(n.users)) # Realize if the IRNode already has accumulated lots of reads if isinstance(result, TensorBox) and result.has_exceeded_max_reads(): # Prevent excessive accumulation in a computed buffer, when # there are multiple branches each with small number of memory # reads, but they converge to a user. result.realize_hint() # Realize if a Pointwise has too much stuff to be inlined. # As this may cause RecursionError during Inductor's evaluation. if isinstance(result, TensorBox) and isinstance(result.data, StorageBox): curr = result.data.data if isinstance(curr, Pointwise): # Use inner fn as a rough proxy. Good enough. if curr.has_large_inner_fn(): result.realize() # This is not complete, but it doesn't have to be: origin_node # tracking is best effort. The logic here critically relies on direct # TensorBox -> StorageBox denoting a non-view; we don't bother trying # to get views to work. Feel free to add any extra cases as needed. # # Note: we can't YOLO tree_map over this result, because if there are # buffers or a view involved, we might not be able to validly assign # the origin_node here. if isinstance(result, TensorBox) and isinstance(result.data, ir.StorageBox): if isinstance(result.data.data, ir.Loops): result.data.data.origin_node = n elif isinstance(result.data.data, ir.Buffer): result.data.data.origin_node = n if isinstance(result.data.data, ir.ComputedBuffer) and isinstance( result.data.data.data, ir.Loops ): result.data.data.data.origin_node = n # Not really multi-output, can straightforwardly recurse in elif ( isinstance(result.data.data, ir.MultiOutput) and not result.data.data.indices ): if isinstance(result.data.data.inputs[0], ir.Buffer): result.data.data.inputs[0].origin_node = n self.register_users_of(result) return result def validate_can_generate_cpp_wrapper(self): if config.disable_cpp_codegen: raise CppWrapperCodeGenError("C++ codegen is disabled") if sys.platform not in ["linux", "darwin"]: raise CppWrapperCodeGenError(f"Unsupported platform {sys.platform}") for value in self.graph_inputs.values(): dtype = None if isinstance(value, TensorBox): dtype = value.get_dtype() elif isinstance( value, (sympy.Symbol, sympy.Expr, sympy.core.numbers.Integer) ): dtype = may_get_constant_buffer_dtype(value) if not supported_dtype_of_cpp_wrapper(dtype, self.cuda): raise CppWrapperCodeGenError(f"Unsupported input dtype {dtype}") def init_wrapper_code(self): self.cuda = "cuda" in self.device_types if self.cpp_wrapper: self.validate_can_generate_cpp_wrapper() self.wrapper_code = CppWrapperCuda() if self.cuda else CppWrapperCpu() else: device_types = self.device_types.copy() device_types.discard("cpu") # TODO(Eikan): Only support mixing cpu and other device now. assert len(device_types) <= 1, "Does not support mixing {}".format( "+".join(device_types) ) only_cpu = len(device_types) == 0 device_type = "cpu" if only_cpu else device_types.pop() self.device_ops = get_device_op_overrides(device_type) wrapper_code_gen_cls = get_wrapper_codegen_for_device(device_type) assert ( wrapper_code_gen_cls is not None ), f"Device {device_type} not supported" self.wrapper_code = wrapper_code_gen_cls() if self.const_module: # If we have const module, we could reuse the kernels # This could avoid duplication and save time on doing recompilation (if Triton.) self.wrapper_code._names_iter = self.const_module.wrapper_code._names_iter self.wrapper_code.src_to_kernel = ( self.const_module.wrapper_code.src_to_kernel ) def codegen_with_cpp_wrapper(self): """ For CPU, the cpp wrapper codegen is done in one pass. For GPU, the cpp wrapper codegen is done in two steps: JIT-compile the model with python wrapper code and run it to generate autotuned kernel binaries in the first pass; and then generate cpp wrapper code and compile it to a dynamic library in the second pass. """ if "cuda" in self.device_types: # first pass self.cpp_wrapper = False compiled = self.compile_to_module().call def materialize(x): if isinstance(x, (torch.SymInt, torch.SymFloat)): # Need concrete value to run dynamic shapes and tune the result return x.node.hint elif isinstance(x, FakeTensor): return defake(x) else: assert isinstance( x, torch.Tensor ), "Unknown type when creating real inputs" + str(type(x)) return x if tracing_context := torch._guards.TracingContext.try_get(): if tracing_context.output_strides: tracing_context.output_strides.clear() params_flat = [ param for param in tracing_context.params_flat # type: ignore[union-attr] if param is not None ] real_inputs = [ materialize(x) for x in itertools.chain(params_flat, V.real_inputs) ] else: real_inputs = [materialize(x) for x in V.real_inputs] with torch.utils._python_dispatch._disable_current_modes(): assert self.example_inputs is not None compiled(real_inputs) del real_inputs # second pass # TODO: reuse self.scheduler from the first pass to speed up the second pass self.cpp_wrapper = True self.removed_buffers.clear() self.inplaced_to_remove.clear() return self.codegen() else: # cpu return self.codegen() def codegen(self): from .scheduler import Scheduler self.init_wrapper_code() self.scheduler = Scheduler(self.buffers) V.debug.draw_orig_fx_graph(self.orig_gm, self.scheduler.nodes) self.scheduler.codegen() return self.wrapper_code.generate(self.is_inference) def codegen_subgraph(self, parent_graph): """ This is a more compact version of the `codegen()` above where we codegen this graph as a subgraph of some parent graph. The parent graph is passed as an argument: the intention is to inline codegening of the subgraph in the parent graph's wrapper code (including the generated kerenls). The wrapper code is not finalized (via `.generate()` call), as this will be done in the parent graph's `codegen()`. """ from .scheduler import Scheduler self.wrapper_code = parent_graph.wrapper_code self.device_ops = parent_graph.device_ops self.cpp_wrapper = parent_graph.cpp_wrapper self.scheduler = Scheduler(self.buffers) self.scheduler.codegen() def count_bytes(self): from .scheduler import Scheduler scheduler = Scheduler(self.buffers) total_bytes = 0 node_counts = [] node_runtimes = [] for node in scheduler.nodes: num_bytes = node.get_read_write_buffers_sizes() total_bytes += num_bytes node_counts.append((node, num_bytes // 4)) node_runtimes.append((node, node.get_estimated_runtime())) return total_bytes, node_counts, node_runtimes @dynamo_timed(phase_name="code_gen") def compile_to_module(self): from .codecache import PyCodeCache code, linemap = ( self.codegen_with_cpp_wrapper() if self.cpp_wrapper else self.codegen() ) linemap = [(line_no, node.stack_trace) for line_no, node in linemap] key, path = PyCodeCache.write(code) mod = PyCodeCache.load_by_key_path( key, path, linemap=linemap, attrs=self.constants ) self.cache_key = key self.cache_path = path self.cache_linemap = linemap # Logged twice as per https://github.com/pytorch/pytorch/pull/99038#discussion_r1167826029 # TODO. Revisit this once the logging API is more mature assert mod.__file__ is not None log_module_code(mod.__file__) log.debug("Output code written to: %s", mod.__file__) output_code_log.debug("Output code: \n%s", code) trace_structured( "inductor_output_code", lambda: {"filename": mod.__file__}, payload_fn=lambda: code, ) output_code_log.info("Output code written to: %s", mod.__file__) if config.benchmark_kernel: print(f"Compiled module path: {mod.__file__}", file=sys.stderr) V.debug.output_code(mod.__file__) V.debug.copy(os.path.splitext(mod.__file__)[0] + ".debug") return mod def compile_to_fn(self): if self.aot_mode: from .codecache import AotCodeCompiler assert self.cpp_wrapper, "AOT mode only supports C++ wrapper" code, linemap = self.codegen_with_cpp_wrapper() output_code_log.debug("Output code: \n%s", code) serialized_extern_kernel_nodes = None if ( config.is_fbcode() and self.extern_kernel_nodes and self.extern_node_serializer ): serialized_extern_kernel_nodes = self.extern_node_serializer( self.extern_kernel_nodes ) output_code_log.debug( "Serialized Extern Kernel Nodes: \n%s", serialized_extern_kernel_nodes, ) # Directly return the file path with the compiled code return AotCodeCompiler.compile( self, code, serialized_extern_kernel_nodes, cuda=self.cuda ) else: return self.compile_to_module().call def get_output_names(self): return [ node.get_name() for node in self.graph_outputs if not isinstance(node, ir.NoneAsConstantBuffer) and not isinstance(node, ir.ShapeAsConstantBuffer) ] def is_unspec_arg(self, name: str): # dynamo wraps unspec variable as 0d CPU tensor, # need to convert to scalar during codegen (triton only) return ( name in self.graph_inputs.keys() and self.graph_inputs[name].get_numel() == 1 and self.graph_inputs[name].get_device().type == "cpu" )