from __future__ import annotations import contextlib import functools import itertools import json 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, NoReturn, Optional, TYPE_CHECKING, Union import sympy from sympy import Expr import torch import torch._logging import torch.fx from torch import device, Tensor from torch._decomp import get_decompositions from torch._dynamo.utils import defake, dynamo_timed from torch._library.fake_class_registry import FakeScriptObject from torch._logging import LazyString, trace_structured from torch._prims_common import ( compute_required_storage_length, make_channels_last_strides_for, ) 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 ( free_unbacked_symbols, has_free_symbols, resolve_unbacked_bindings, RuntimeAssert, ShapeEnv, SympyBoolean, SymTypes, ) from torch.fx.node import Node from torch.utils._mode_utils import no_dispatch from torch.utils._ordered_set import OrderedSet from torch.utils._sympy.numbers import int_oo from . import config, ir, metrics from .codegen.common import ( BackendFeature, DeviceOpOverrides, get_backend_features, get_device_op_overrides, get_wrapper_codegen_for_device, init_backend_registration, WorkspaceArg, ) from .exc import ( CppWrapperCodegenError, LoweringException, MissingOperatorWithDecomp, MissingOperatorWithoutDecomp, ) from .ir import ( Constant, DonatedBuffer, FixedLayout, get_device_type, GraphPartitionSignature, InputBuffer, Pointwise, Reduction, StorageBox, TensorBox, TorchBindObject, ) from .lowering import ( constrain_to_fake_tensors, constrain_to_fx_strides, FALLBACK_ALLOW_LIST, fallback_handler, fallback_node_due_to_unsupported_type, lowerings, make_fallback, maybe_layout_constraints, needs_realized_inputs, require_contiguous, unsupported_output_tensor, ) from .runtime import autotune_cache from .runtime.autotune_cache import AutotuneCacheBundler from .sizevars import SizeVarAllocator from .utils import ( convert_shape_to_inductor, gather_origins, get_cloned_parameter_buffer_name, get_donated_idxs, get_sympy_Expr_dtype, is_same_tensor, maybe_get_suppress_shape_guards_ctx, normalize_name, should_assume_input_aligned, ValueWithLineMap, ) from .virtualized import NullHandler, V if TYPE_CHECKING: from collections.abc import Iterable, Iterator, Sequence from types import ModuleType from torch._higher_order_ops.effects import _EffectType from torch.fx import GraphModule from torch.fx.graph import Graph from .codegen.wrapper import PythonWrapperCodegen from .scheduler import BaseSchedulerNode from torch._inductor.codecache import output_code_log log = logging.getLogger(__name__) perf_hint_log = torch._logging.getArtifactLogger(__name__, "perf_hints") aten = torch.ops.aten _post_grad_graph_counter = itertools.count() if config.is_fbcode(): from torch._inductor.fb.utils import log_module_code else: def log_module_code(*args: Any, **kwargs: Any) -> None: pass def may_get_constant_buffer_dtype(constant_buffer: sympy.Expr) -> Optional[torch.dtype]: 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: Any) -> bool: magic_ops = OrderedSet(method_to_operator(m) for m in magic_methods) return op in magic_ops def getattr_recursive( obj: GraphModule, target: str ) -> Union[Tensor, torch._C.ScriptObject, GraphModule]: 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 def get_user_visible_output_strides(g: Graph) -> dict[Node, tuple[int, ...]]: ret: dict[Node, tuple[int, ...]] = {} output_node = g.find_nodes(op="output")[0] if "user_visible_output_idxs" not in output_node.meta: return ret for idx, node in enumerate(output_node.args[0]): if idx in output_node.meta["user_visible_output_idxs"]: ret[node] = output_node.meta["original_output_strides"][idx] return ret def mark_nodes_dislike_padding( g: Graph, user_visible_output_strides: dict[Node, tuple[int, ...]] ) -> None: """ Nodes like convolution/convolution_backward want its input to be dense. If we pad their inputs, we result in extra calls to copy kernels! On the other hand, padding usually helps reduction. The pass finds nodes that dislike padding. These are nodes that can be reached from a convolution/convolution_backward in the backward direction without going thru a reduction. """ if not config.comprehensive_padding: return ops_dislike_padding = OrderedSet( [ aten.convolution, aten.convolution_backward, aten._scaled_mm, ] ) # what's a better way to collect the reduction ops? ops_like_padding = OrderedSet( [ aten.var_mean, aten.sum, aten.mean, aten.prod, aten.any, aten.amin, aten.amax, aten.min, aten.max, aten.argmin, aten.argmax, aten.scatter_reduce, ] ) def _get_overload_packet( node: torch.fx.Node, ) -> Optional[torch._ops.OpOverloadPacket]: return ( node.target._overloadpacket if node.op == "call_function" # hasattr on OpOverloadPacket is slow, do isinstance first and isinstance(node.target, torch._ops.OpOverload) and hasattr(node.target, "_overloadpacket") else None ) for cur in reversed(g.nodes): if isinstance( cur.target, torch._higher_order_ops.triton_kernel_wrap.TritonKernelWrapperMutation, ): cur.meta["dislike_padding"] = True continue op = _get_overload_packet(cur) if not op: continue if op in ops_dislike_padding: cur.meta["dislike_padding"] = True if cur.meta.get("dislike_padding", False): # propagate for prior in cur.all_input_nodes: prior_op = _get_overload_packet(prior) if not prior_op: continue if prior_op not in ops_like_padding: prior.meta["dislike_padding"] = True # We only want to mark output nodes. So, move it after the above prior nodes process. if not config.pad_outputs and cur in user_visible_output_strides: cur.meta["dislike_padding"] = True class GraphLowering(torch.fx.Interpreter): graph_outputs: list[ir.IRNode] def __init__( self, gm: torch.fx.GraphModule, example_inputs: Optional[Sequence[object]] = None, shape_env: Optional[ShapeEnv] = None, graph_id: Optional[int] = None, cpp_wrapper: bool = False, aot_mode: bool = False, layout_opt: Optional[bool] = None, extern_node_serializer: Optional[ Callable[[list[ir.ExternKernelNode]], Any] ] = None, is_inference: bool = False, is_backward: bool = False, is_const_graph: bool = False, const_output_index: Optional[dict[str, int]] = None, const_wrapper_code: Optional[str] = None, const_kernel_code: Optional[str] = None, const_module: Optional[GraphLowering] = None, name: Optional[str] = None, inputs_to_check: Optional[Sequence[int]] = None, ) -> 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_backward = is_backward self.is_const_graph = is_const_graph self.const_wrapper_code = const_wrapper_code self.const_kernel_code = const_kernel_code self.const_module = const_module self.inputs_to_check = inputs_to_check 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.reuse_shape_env = True self._shape_env = shape_env # We're going to mutate ras_by_symbol as we finish generating them self.ras_by_symbol: dict[Optional[sympy.Symbol], list[RuntimeAssert]] = ( shape_env.deferred_runtime_asserts.copy() ) self.bound_unbacked_symbols = OrderedSet[sympy.Symbol]() self.sizevars = SizeVarAllocator(shape_env) self.graph_input_names: list[str] = [] self.graph_inputs: dict[str, Union[TensorBox, TorchBindObject, sympy.Expr]] = {} self.graph_inputs_original: dict[str, InputBuffer] = {} self.zero_dim_cpu_tensor_list = OrderedSet[str]() self.device_types: OrderedSet[str] = ( const_module.device_types if const_module else OrderedSet() ) self.device_idxs: OrderedSet[int] = ( const_module.device_idxs if const_module else OrderedSet() ) self.device_type = "cpu" # Inplace padding may require Inductor to allocate slightly larger # tensor for padding. self.buffer_to_padded_size: dict[str, list[int]] = {} self.buffers: list[ir.Buffer] = [] self.operations: list[ir.Operation] = [] self.const_output_index: dict[str, int] = ( const_output_index if const_output_index else {} ) self.folded_constants: OrderedSet[str] = ( OrderedSet(const_output_index.keys()) if const_output_index else OrderedSet() ) self.constants: dict[str, torch.Tensor] = ( const_module.constants if const_module else {} ) self.torchbind_constants: dict[str, torch._C.ScriptObject] = {} self.seen_subgraphs: dict[str, ir.Subgraph] = {} self.constant_reprs: dict[str, str] = {} self.removed_operations = OrderedSet[str]() self.removed_buffers = OrderedSet[str]() self.removed_inplace_buffers = OrderedSet[str]() self.mutated_buffers = OrderedSet[str]() self.never_reuse_buffers = OrderedSet[str]() self.inplaced_to_remove = OrderedSet[str]() self.device_ops: DeviceOpOverrides = None # type: ignore[assignment] self.wrapper_code: PythonWrapperCodegen = None # type: ignore[assignment] # See `ProxyExecutor Design Note` in ir.py for more details self.extern_kernel_nodes: list[ir.ExternKernelNode] = [] from torch._inductor.extern_node_serializer import extern_node_json_serializer self.extern_node_serializer: Callable[[list[ir.ExternKernelNode]], Any] = ( extern_node_serializer if config.is_fbcode() and extern_node_serializer else extern_node_json_serializer ) self.current_node: torch.fx.Node = None # type: ignore[assignment] self.lists: dict[str, list[str]] = {} self.mutated_inputs = OrderedSet[str]() 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.name_to_op: dict[str, ir.Operation] = {} self.creation_time = time.time() self.name = name # type: ignore[assignment] 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 OrderedSet to false for the first pass of codegen. self.record_multi_kernel_choice = cpp_wrapper self.multi_kernel_to_choice: dict[str, str] = {} self.aot_mode = aot_mode self.graph_id = graph_id self.post_grad_graph_id = next(_post_grad_graph_counter) self.scheduler: torch._inductor.scheduler.Scheduler = None # type: ignore[assignment] # current_device is set only during codegen of a device-specific kernel # a graph can have many devices self.current_device: Optional[torch.device] = None self.nodes_prefer_channels_last = ( self.find_nodes_prefer_channels_last() if self.layout_opt else OrderedSet() ) self._warned_fallback = OrderedSet(["aten.convolution_backward"]) self.user_visible_output_strides = get_user_visible_output_strides(gm.graph) mark_nodes_dislike_padding(gm.graph, self.user_visible_output_strides) 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( # type: ignore[operator, union-attr] "dynamo_flat_name_to_original_fqn", {} ) self.allocated_constant_name: dict[str, str] = ( const_module.allocated_constant_name if const_module is not None else {} ) init_backend_registration() self.get_backend_features = functools.lru_cache(None)(get_backend_features) self.effectful_ops: dict[_EffectType, ir.Buffer] = {} self.aligned_inputs: OrderedSet[str] = OrderedSet() self.no_fuse_buffer_names = OrderedSet[str]() self.low_precision_codegen_ops: OrderedSet[str] = OrderedSet() # more aggressive prologue fusion self.invoke_quant_ops: OrderedSet[str] = OrderedSet() # Below field is related to printing debug intermediate tensor values info for debugging self.all_codegen_kernel_names = OrderedSet[str]() # state used by for Kernel.workspace self.workspace_id = itertools.count() # track the current placeholder index that we are processing self.placeholder_idx = -1 self.bw_donated_idxs = get_donated_idxs() def freeze_runtime_asserts(self) -> None: self._shape_env.freeze_runtime_asserts() def symbolic_sizes_strides( self, ex: torch.Tensor ) -> tuple[Sequence[Union[int, Expr]], Sequence[Union[int, Expr]]]: """ 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, ) r_size = [i.node.expr if isinstance(i, torch.SymInt) else i for i in size] r_stride = [i.node.expr if isinstance(i, torch.SymInt) else i for i in stride] return r_size, r_stride def static_sizes_strides( self, ex: torch.Tensor ) -> tuple[list[sympy.Expr], list[sympy.Expr]]: """ 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 get_allocation_size( self, node: Union[ ir.TensorBox, ir.StorageBox, ir.Buffer, WorkspaceArg, ir.TorchBindObject ], ) -> Sequence[Expr]: if isinstance(node, ir.TensorBox): node = node.data # type: ignore[assignment] if isinstance(node, ir.StorageBox): node = node.data # type: ignore[assignment] if ( isinstance(node, ir.ComputedBuffer) and node.name in self.buffer_to_padded_size ): return self.buffer_to_padded_size[node.name] else: return node.get_size() def get_allocation_storage_size( self, node: Union[ir.Buffer, WorkspaceArg, ir.TorchBindObject] ) -> Expr: layout = node.get_layout() size = self.get_allocation_size(node) # consider inplace padding stride = layout.stride offset = layout.offset return compute_required_storage_length(size, stride, offset) # type: ignore[arg-type] def has_feature( self, device: Union[torch._inductor.ir.IRNode, device, None], feature: BackendFeature, ) -> bool: assert isinstance(feature, BackendFeature), feature return feature in self.get_backend_features(get_device_type(device)) def get_current_device_or_throw(self) -> torch.device: if device := self.current_device: return device else: raise RuntimeError("No current device") @contextlib.contextmanager def set_current_device(self, device: torch.device) -> Iterator[None]: prior = self.current_device self.current_device = device try: yield finally: self.current_device = prior def get_training_phase(self) -> str: if self.is_inference: return "inference" if self.is_backward: return "backward" return "forward" @staticmethod def decide_layout_opt(gm: GraphModule, *, is_inference: bool) -> 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: Any) -> bool: meta_val = n.args[1].meta["val"] # type: ignore[union-attr, operator] assert isinstance(meta_val, torch.Tensor) return n.args[-1] > 1 and meta_val.size(1) > 1 # type: ignore[union-attr, operator] def is_in_out_channel(n: torch.fx.Node) -> bool: return ( n.args[1].meta["val"].size(0) * 2 <= n.args[1].meta["val"].size(1) # type: ignore[union-attr, operator] and n.args[1].meta["val"].size(2) > 1 # type: ignore[union-attr, operator] ) def is_small_channel(n: torch.fx.Node) -> bool: return ( n.args[1].meta["val"].size(0) <= 64 # type: ignore[union-attr, operator] and n.args[1].meta["val"].size(1) <= 64 # type: ignore[union-attr, operator] ) # 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, ) -> SubgraphLowering: """ Make a subgraph of the current graph with all inherited parts, except the graph module (`gm`) and `example_inputs`. The subgraphs are lowered separately and lifted into a separate function in the parent output wrapper code. The subgraph name is qualified by the parent graph's name. Note that the lifting of subgraph is supported for python wrapper only. For cpp wrapper, we inline the subgraphs in the parent wrapper. """ return SubgraphLowering( parent=self, 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, is_backward=self.is_backward, name=self.qualify_name(subgraph_name), ) def find_nodes_prefer_channels_last(self) -> OrderedSet[Node]: """ 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 = OrderedSet[Node]() for n in reversed(self.module.graph.nodes): # type: ignore[arg-type, union-attr] 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: # type: ignore[union-attr] if n in output_set: output_set.update(n.users) return output_set def warn_fallback(self, name: str) -> None: 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) -> None: 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) -> torch._subclasses.fake_tensor.FakeTensorMode: return V.fake_mode def try_get_buffer( self, buffer_name: str ) -> Optional[Union[ir.TensorBox, ir.Buffer, ir.TorchBindObject]]: 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] if buffer_name in self.constants: data = V.graph.constants[buffer_name] return ir.ConstantBuffer( name=buffer_name, layout=ir.FixedLayout( data.device, data.dtype, *V.graph.static_sizes_strides(data) ), ) return None def add_symbol_graph_input(self, symbol: sympy.Expr) -> None: raise RuntimeError("Should not be called for the main graph") def get_buffer( self, buffer_name: str ) -> Union[ir.TensorBox, ir.Buffer, ir.TorchBindObject]: buf = self.try_get_buffer(buffer_name) if buf is not None: return buf raise RuntimeError(f"Failed to find buffer matching name {buffer_name}") def get_dtype(self, buffer_name: str) -> torch.dtype: if buffer_name in self.constants: return self.constants[buffer_name].dtype # For a mutation op we should return the dtype of the buffer being mutated if ( hasattr(self.scheduler, "mutation_real_name") and buffer_name in self.scheduler.mutation_real_name ): mutated_buf = self.scheduler.mutation_real_name[buffer_name] if mutated_buf in self.name_to_buffer: return self.name_to_buffer[mutated_buf].get_dtype() if mutated_buf in self.graph_inputs: return self.graph_inputs[mutated_buf].get_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) -> Union[int, Expr]: 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 not buf.has_tensor_output(): 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}") def run(self, *args: Any) -> Any: # type: ignore[override] with dynamo_timed("GraphLowering.run"): return super().run(*args) def register_operation(self, op: ir.Operation) -> str: assert op.operation_name is None, f"Operation registered twice: {op}" assert isinstance(op, ir.Operation) name = self.qualify_name(f"op{len(self.operations)}") self.operations.append(op) self.name_to_op[name] = op op.operation_name = name return name def register_buffer(self, buffer: ir.Buffer, *, set_name: bool = False) -> str: name = self.qualify_name(f"buf{len(self.buffers)}") self.buffers.append(buffer) self.name_to_buffer[name] = buffer device = buffer.get_device() if ( # Skip empty CPU tensor so that CUDA graphs can succeed, see https://github.com/pytorch/pytorch/pull/114144 device is not None and not ( isinstance(buffer, ir.ComputedBuffer) and buffer.is_zero_elements() and device == torch.device("cpu") ) ): self.add_device_info(device) if set_name: buffer.name = name return name def register_operation_list(self, operation_names: list[str]) -> str: name = self.qualify_name("list_" + "_".join(operation_names)) self.lists[name] = operation_names return name def register_users_of( self, node_output: Union[Iterable[ir.IRNode], ir.IRNode] ) -> None: def register(value: Union[Iterable[ir.IRNode], ir.IRNode]) -> None: if isinstance(value, (list, tuple)): for x in value: register(x) if isinstance(value, ir.TensorBox): 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) -> None: """ 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 get_original_value_of_constant(self, name: str) -> torch.Tensor: """ In AOTI, module buffers may have been mutated during the tracing and compilation. Thus we need to read from previously stored original buffers, to make sure the generated model.so uses correct initial values. """ assert name in self.allocated_constant_name and name in self.constants, ( "Can not find the original value for " + name ) orig_name = get_cloned_parameter_buffer_name(self.allocated_constant_name[name]) return ( self.module.meta[orig_name] # type: ignore[index] if orig_name in self.module.meta # type: ignore[operator] else self.constants[name] ) def allocate_non_dup_const_name( self, name: Optional[str], data: Union[Tensor] ) -> str: if not config.aot_inductor.use_runtime_constant_folding: for constant_name, value in self.constants.items(): if is_same_tensor(data, value): return constant_name if name is None: name = f"constant{len(self.constants)}" orig_name = name 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 = normalize_name(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}" ) self.allocated_constant_name[name] = orig_name # type: ignore[assignment] return name def add_tensor_constant( self, data: Tensor, name: Optional[str] = None ) -> TensorBox: new_name = self.allocate_non_dup_const_name(name, data) return TensorBox.create( ir.ConstantBuffer( name=new_name, layout=FixedLayout( data.device, data.dtype, *self.static_sizes_strides(data) ), ) ) def constant_name(self, name: str, device_override: Optional[torch.device]) -> str: """ 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 with torch.utils._python_dispatch._disable_current_modes(): # caller might have OrderedSet fake tensor mode which will create a fake tensor # when calling .to, so unset modes here return self.allocate_non_dup_const_name( f"{name}_{device_override.type}{device_override.index or 0}", self.constants[name].to(device_override), ) def placeholder( self, target: str, # type: ignore[override] args: tuple[object], # type: ignore[override] kwargs: dict[str, object], ) -> Union[Expr, TensorBox, None]: self.placeholder_idx += 1 example = super().placeholder(target, args, kwargs) # type: ignore[arg-type] target = self.qualify_name(target) if isinstance(example, SymTypes): expr = example.node.expr self.graph_inputs[target] = expr self.graph_input_names.append(target) return expr elif isinstance(example, (int, bool, float)): expr = sympy.sympify(example) self.graph_inputs[target] = expr self.graph_input_names.append(target) return expr elif isinstance(example, FakeScriptObject): obj = TorchBindObject(name=target, value=example) self.graph_inputs[target] = obj self.graph_input_names.append(target) return obj elif example is None: self.graph_input_names.append(target) return None if isinstance(example, BackwardState): # Ignored arg, must be unused # Alternately we could filter this out in AotAutograd self.graph_input_names.append(target) return None # See note: Note: [Generator arguments in AOTDispatcher] elif isinstance(example, torch.Generator): assert ( len(V.graph.current_node.users) == 1 and next(iter(V.graph.current_node.users)).target is torch._prims.rng_prims.graphsafe_run_with_rng_state ) gen = ir.GeneratorState(name=target, device=example.device) self.graph_inputs[target] = gen # type: ignore[assignment] self.graph_input_names.append(target) return gen 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) # type: ignore[assignment] if ( self.is_backward and self.bw_donated_idxs and self.placeholder_idx in self.bw_donated_idxs ): tensor = TensorBox.create( DonatedBuffer( name=target, layout=FixedLayout(example.device, example.dtype, sizes, strides), ) ) else: # TODO(jansel): handle input aliasing tensor = TensorBox.create( InputBuffer( name=target, layout=FixedLayout(example.device, example.dtype, sizes, strides), ) ) self.graph_inputs[target] = tensor self.graph_input_names.append(target) self.graph_inputs_original[target] = tensor.data.data if self.current_node.users: # cudagraphs should work with an unused CPU input self.add_device_info(example.device) # Note: [Input Alignment handling in Inductor] # Alignment matters for generating efficient code. Some operations, # e.g. vectorized loads, can only be performed on aligned inputs. # # But if we codegen assuming aligned inputs and then get unaligned # inputs at runtime, then we are forced to clone - which is bad for # both perf and memory usage. # # One option would be to guard on storage_offset%ALIGNMENT, and then # codegen based on this. But storage_offset guards turned out to be # expensive and cause recompiles; Instead, we're generating code # based on the alignment of the example input without guarding. with maybe_get_suppress_shape_guards_ctx(): if should_assume_input_aligned(example): self.aligned_inputs.add(target) return tensor def call_function(self, target: Callable, args: Any, kwargs: dict[str, Any]) -> Any: # type: ignore[type-arg, override] if target is operator.getitem and isinstance(args[0], (list, tuple, dict)): return super().call_function(target, args, kwargs) # hasattr on OpOverloadPacket is slow, check isinstance first if not isinstance(target, torch._ops.OpOverloadPacket) and hasattr( target, "_inductor_lowering_function" ): # passthrough lowerings from .pattern_matcher return target(*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, warn=False, override_decomp=True) elif config.implicit_fallbacks: error = ( MissingOperatorWithDecomp if get_decompositions([target]) else MissingOperatorWithoutDecomp ) log.info( "Creating implicit fallback for:\n%s", error.operator_str(target, args, kwargs), ) # use contiguous unless the (custom) op asks something else # explicitly if torch._C.Tag.needs_fixed_stride_order in target.tags: decided_constraint = constrain_to_fx_strides # type: ignore[assignment] elif torch._C.Tag.flexible_layout in target.tags: decided_constraint = None # type: ignore[assignment] else: # If there are no tags, we do different things depending on # if it's a builtin ATen/prim ops or custom ops. # For ATen ops, we require_contiguous to fix https://github.com/pytorch/pytorch/issues/140452 # For custom ops, we constrain_to_fx_strides to maintain the # behavior of PyTorch 2.5: https://github.com/pytorch/pytorch/issues/148356 # # For ATen ops, only apply the constraint for backward # ops since fwd ops should work for any strides. if torch._library.utils.is_builtin(target) and self.is_backward: decided_constraint = require_contiguous # type: ignore[assignment] else: # maybe_layout_constraints will decide the layout constraint for the custom op # lazily decided_constraint = None # type: ignore[assignment] # for implicitly fallback ops, we conservatively requires # contiguous input since some eager kernels does not # support non-contiguous inputs. They may silently cause # accuracy problems. Check https://github.com/pytorch/pytorch/issues/140452 make_fallback(target, layout_constraint=decided_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]) # type: ignore[index] n = self.current_node layout_constraints = maybe_layout_constraints(target) if layout_constraints: old_args, old_kwargs = args, kwargs args, kwargs = layout_constraints(n, *args, **kwargs) out = lowerings[target](*args, **kwargs) # type: ignore[index] if layout_constraints: # layout_constraints are allowed to make new copies of the inputs. # if they do, and if the target is mutable, then we need to # write the new values back into the original inputs. self.propagate_mutation(n, old_args, old_kwargs, args, kwargs) # type: ignore[possibly-undefined] 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: str, # type: ignore[override] args: tuple[()], # type: ignore[override] kwargs: dict[str, object], ) -> Union[Constant, TensorBox, ir.Subgraph, TorchBindObject]: # this is a constant value = getattr_recursive(self.module, target) # type: ignore[arg-type] if isinstance(value, torch.fx.GraphModule): # Reuse the existing subgraph if we have seen it before already. if target in self.seen_subgraphs: return self.seen_subgraphs[target] out = ir.Subgraph(name=target, graph_module=value) self.seen_subgraphs[target] = out return out if isinstance(value, torch._C.ScriptObject): self.torchbind_constants[target] = value self.constant_reprs[target] = "" return TorchBindObject(name=target, value=value) elif isinstance(value, FakeScriptObject): self.torchbind_constants[target] = value.real_obj self.constant_reprs[target] = "" return TorchBindObject(name=target, value=value.real_obj) assert isinstance(value, torch.Tensor) 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=value.item(), dtype=value.dtype, device=value.device ) if self.can_inline_constant(value): log.debug("Inlining constant: %s ", str(target)) # 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: Any, args: Any, kwargs: Any) -> NoReturn: raise AssertionError def call_method(self, target: Any, args: Any, kwargs: Any) -> NoReturn: raise AssertionError def output( self, target: str, # type: ignore[override] args: tuple[object], # type: ignore[override] kwargs: dict[str, object], ) -> None: result = super().output(target, args, kwargs) # type: ignore[arg-type] if not isinstance(result, (tuple, list)): # nested subgraphs can have singleton outputs result = (result,) 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, ir.EffectfulKernel, ir.ShapeAsConstantBuffer, ), ) for x in result ), result fx_node_args = V.graph.current_node.args[0] # type: ignore[arg-type] if not isinstance(fx_node_args, (tuple, list)): # nested subgraphs can have singleton outputs fx_node_args = (fx_node_args,) result = [ir.ExternKernel.realize_input(x) for x in result] result_correct_strides = [] assert len(fx_node_args) == len(result) for r, fx_node in zip(result, fx_node_args): if not isinstance(r, (ir.TensorBox, ir.BaseView)): result_correct_strides.append(r) elif isinstance(r.get_output_spec(), ir.CommBufferLayout): # Active references to persistent comm buffers are not allowed # outside of graphs result_correct_strides.append(ir.ExternKernel.copy_input(r)) else: # AOT Autograd tries to detect stride divergence of inductor from output metadata. # Here, we try to avoid spurious divergence by matching insignificant strides such as # should have already been realized assert torch._inductor.ir.is_storage_and_layout(r) meta_strides = [ s.node.expr if isinstance(s, torch.SymInt) else s for s in fx_node.meta["val"].stride() ] result_correct_strides.append( ir.try_match_insignificant_strides(r, meta_strides) ) self.graph_outputs = result_correct_strides value: ir.IRNode for name, value in self.graph_inputs.items(): if isinstance(value, TorchBindObject): continue assert isinstance( value, (TensorBox, sympy.Expr, torch._inductor.ir.GeneratorState) ), 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.MutationLayoutSHOULDREMOVE.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) -> None: for buf in self.buffers: buf.decide_layout() @contextmanager def set_current_node(self, node: torch.fx.Node): # type: ignore[no-untyped-def] old = self.current_node try: self.current_node = node yield finally: self.current_node = old @contextmanager def set_current_wrapper_code(self) -> Iterator[None]: old = self.wrapper_code try: yield finally: self.wrapper_code = old def propagate_mutation( self, fx_node: torch.fx.Node, old_args: tuple[Any], old_kwargs: dict[str, Any], new_args: tuple[Any], new_kwargs: dict[str, Any], ) -> None: """Propagate mutations on new_args/new_kwargs back to old_args/old_kwargs. Assumes we may have cloned old_args/old_kwargs into new_args/new_kwargs and then called fx_node(*new_args, **new_kwargs). If fx_node mutates any of new_args/new_kwargs, and they are different from old_args/old_kwargs, then we need to update the original tensor. """ assert len(old_args) == len(new_args) assert len(old_kwargs) == len(new_kwargs) if fx_node.target is torch.ops.higher_order.triton_kernel_wrapper_mutation: kwargs = fx_node.kwargs["kwargs"] assert isinstance(kwargs, dict) mutated = torch._higher_order_ops.triton_kernel_wrap.get_mutated_tensors( old_kwargs["kernel_idx"], old_kwargs["constant_args_idx"], { k: v.meta["val"] if isinstance(v, torch.fx.Node) else v for k, v in kwargs.items() }, ) for name in mutated: old_arg = old_kwargs["kwargs"][name] new_arg = new_kwargs["kwargs"][name] if old_arg is new_arg: continue self.call_function(torch.ops.aten.copy_.default, (old_arg, new_arg), {}) return assert isinstance(fx_node.target, torch._ops.OpOverload) def maybe_propagate( schema_arg: torch._C.Argument, old_arg: ir.IRNode, new_arg: ir.IRNode ) -> None: if old_arg is new_arg: return if schema_arg.alias_info is not None and schema_arg.alias_info.is_write: # The lowering for copy_ is smart enough to "replace" old_arg with # new_arg in all future uses so a copy_ kernel never gets emitted. # old_arg, new_arg may be immutable_list if isinstance(old_arg, ir.IRNode): old_arg = (old_arg,) # type: ignore[assignment] new_arg = (new_arg,) # type: ignore[assignment] for old_arg_item, new_arg_item in zip(old_arg, new_arg): # type: ignore[call-overload] if old_arg_item is new_arg_item: continue self.call_function( torch.ops.aten.copy_.default, (old_arg_item, new_arg_item), {} ) schema = fx_node.target._schema for idx, (old_arg, new_arg) in enumerate(zip(old_args, new_args)): schema_arg = schema.arguments[idx] maybe_propagate(schema_arg, old_arg, new_arg) schema_kwargs = {arg.name: arg for arg in schema.arguments} for key in old_kwargs.keys(): old_arg = old_kwargs[key] new_arg = new_kwargs[key] schema_arg = schema_kwargs[key] maybe_propagate(schema_arg, old_arg, new_arg) def run_node(self, n: torch.fx.Node) -> object: def debug(msg: str) -> None: log.debug("lowering %s %s", LazyString(n.format_node), msg) from torch._inductor.compiler_bisector import CompilerBisector buffer_watermark = len(self.buffers) operation_watermark = len(self.operations) # origins: OrderedSet[Union[Node, ir.IRNode]] = OrderedSet([n]) origins: OrderedSet[Any] = OrderedSet([n]) is_call_function = n.op == "call_function" if is_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) or CompilerBisector.disable_subsystem( "inductor", "lowerings", lambda: repr(n) ) ) ): debug("fallback_handler") result = fallback_handler(n.target, add_to_fallback_set=False)( *args, # type: ignore[possibly-undefined] **kwargs, # type: ignore[possibly-undefined] ) elif ( n.op == "call_function" and n.target is torch.ops.higher_order.triton_kernel_wrapper_mutation and config.triton_kernel_default_layout_constraint != "flexible_layout" ): debug("user_defined_triton_kernel_layout_constraints") if ( config.triton_kernel_default_layout_constraint == "needs_fixed_stride_order" ): old_args = args # type: ignore[possibly-undefined] old_kwargs = kwargs # type: ignore[possibly-undefined] if arg_kwarg_vals := n.meta.get("arg_kwarg_vals"): inp_args = arg_kwarg_vals[0] inp_kwargs = arg_kwarg_vals[1] args, kwargs = constrain_to_fake_tensors( args, kwargs, inp_args, inp_kwargs ) else: args, kwargs = constrain_to_fx_strides(n, *args, **kwargs) # type: ignore[index] result = self.call_function(n.target, args, kwargs) # type: ignore[arg-type] self.propagate_mutation(n, old_args, old_kwargs, args, kwargs) # type: ignore[possibly-undefined] else: raise RuntimeError( f"Unknown triton_kernel_default_layout_constraint: {config.triton_kernel_default_layout_constraint}" ) 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 # https://github.com/pytorch/pytorch/issues/127789 debug("is_magic_method") if isinstance( n.meta["val"], (torch.SymInt, torch.SymFloat, torch.SymBool) ): 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, torch.ops.aten.resize.default, torch.ops.aten.resize_as.default, ] is_output = any(user.op == "output" for user in n.users) is_user_visible = n in self.user_visible_output_strides is_input_for_as_strided = any( user.target in as_strided_ops for user in n.users ) if n.meta.get("inductor_realize_to_strides", False) and isinstance( result, TensorBox ): result.realize() strides = n.meta["val"].stride() sym_strides = torch._inductor.utils.any_is_symbolic(*strides) if result.maybe_get_stride() != strides and not sym_strides: stride_order = ir.get_stride_order(strides) result = ir.ExternKernel.require_stride_order(result, stride_order) 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 ): if is_user_visible: strides = self.user_visible_output_strides.get(n) else: strides = n.meta["val"].stride() if strides is not None and len(strides) > 0: allow_padding = ( config.pad_outputs or not is_user_visible ) and not is_input_for_as_strided dense = torch._prims_common.is_non_overlapping_and_dense( n.meta["val"] ) unbacked_symbols_in_strides = ( len(free_unbacked_symbols(strides)) > 0 ) if ( not unbacked_symbols_in_strides and dense and len(result.get_size()) == 4 and n in self.nodes_prefer_channels_last and not is_user_visible and not is_input_for_as_strided ): strides = ir.FlexibleLayout.stride_ordered_for_memory_format( result.get_size(), torch.channels_last ) if not unbacked_symbols_in_strides and len(strides): # To avoid converting possible view ops to a copy kernel, we use the previous # require_exact_strides to handle views. But ultimately it's better to require # the right strides at the tensor definition. if n.meta["val"]._is_view() or isinstance( result.data, ir.BaseView ): result = ir.ExternKernel.require_stride_order( result, ir.get_stride_order(strides), allow_padding=allow_padding, ) else: strides = [ s.node.expr if isinstance(s, torch.SymInt) else s for s in strides ] result = ir.ExternKernel.require_exact_strides( result, strides, allow_padding=allow_padding ) # Realize if (1) any user need inputs realized, or (2) there is # already too many reads and rematerializing can be bad. num_users = len(OrderedSet(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, ] need_fixed_channels_last_layout = [] 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._linear_pointwise.default, torch.ops.mkldnn._linear_pointwise.binary, torch.ops.aten.mkldnn_rnn_layer.default, torch.ops.onednn.qlinear_pointwise.default, torch.ops.onednn.qlinear_pointwise.tensor, torch.ops.onednn.qlinear_pointwise.binary, torch.ops.onednn.qlinear_pointwise.binary_tensor, ] need_fixed_channels_last_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.onednn.qconv2d_pointwise.default, torch.ops.onednn.qconv2d_pointwise.binary, ] 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()), allow_padding=True, ) if ( user.target in need_fixed_channels_last_layout and n is user.args[0] ): result = ir.ExternKernel.require_stride_order( result, ir.get_stride_order( make_channels_last_strides_for(n.meta["val"].shape) ), ) 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(threshold=100): 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._post_init_setattr("origin_node", n) elif isinstance(result.data.data, ir.Buffer): result.data.data._post_init_setattr("origin_node", n) if isinstance(result.data.data, ir.ComputedBuffer) and isinstance( result.data.data.data, ir.Loops ): result.data.data.data._post_init_setattr("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]._post_init_setattr("origin_node", n) self.register_users_of(result) new_unbacked_defs = OrderedSet[sympy.Symbol]() for buf in self.buffers[buffer_watermark:]: new_unbacked_defs |= buf.get_unbacked_symbol_defs() for op in self.operations[operation_watermark:]: new_unbacked_defs |= op.get_unbacked_symbol_defs() def format_new_defs() -> str: r = [ f"unbacked_symbol_defs={buf.get_unbacked_symbol_defs()} in:\n{buf}\n" for buf in self.buffers[buffer_watermark:] ] r.extend( f"unbacked_symbol_defs={op.get_unbacked_symbol_defs()} in:\n{op}\n" for op in self.operations[operation_watermark:] ) return "***\n".join(r) if n.op != "placeholder": # Note [Backwards runtime asserts] # Backwards poses an interesting problem for deferred runtime # asserts. In the easy case, we may solely close over data # dependent sized tensors, and there are no binding sites for # unbacked SymInts. In this case, we can just drop all the # runtime asserts on the floor: no non-placeholder bindings, no # problem. # # However, it is *possible* for a fresh runtime assert to show up # between forwards and backwards. Right now, the freezing process # that happens when we lower forwards means that we will freeze # runtime asserts, and then the moment the backwards lowering # process attempts to add a new deferred runtime assert, we will # fail. Let's say you remove that assert. Now when we get here, # we need to make sure we actually emit these asserts (because we # can't emit them in forwards, we already compiled it). So we # have to do something here. But we don't want to reemit ALL # deferred runtime asserts, we only want to emit the NEW ones. # Therefore needing some sort of stratification in the ShapeEnv. # This is all doable, it just hasn't been done yet. shape_env = V.graph.sizevars.shape_env def make_assert(expr: SympyBoolean, msg: str) -> None: assert_op = ir.AssertScalar(expr, msg) self.register_buffer(assert_op, set_name=True) self.register_operation(assert_op) for i0 in new_unbacked_defs: ras = self.ras_by_symbol.pop(i0, []) # NB: size-like not needed, we won't retrace vr = shape_env.var_to_range[i0] if not shape_env._default_unspecified_value_range().issubset(vr): def is_convertible(s: Expr) -> bool: if s in (int_oo, -int_oo): return False try: int(s) return True except TypeError: return False if is_convertible(vr.lower): make_assert(i0 >= vr.lower, f"{i0} >= {vr.lower}") if is_convertible(vr.upper): make_assert(i0 <= vr.upper, f"{i0} <= {vr.upper}") for ra in ras: fvs = free_unbacked_symbols(ra.expr) missing = fvs - self.bound_unbacked_symbols if missing: i1 = min(missing, key=str) self.ras_by_symbol.setdefault(i1, []).append(ra) else: make_assert(ra.expr, f"{ra.expr}") self.bound_unbacked_symbols |= new_unbacked_defs unbacked_bindings = resolve_unbacked_bindings( V.graph.sizevars.shape_env, n.meta.get("unbacked_bindings", {}) ) assert unbacked_bindings is not None # When we do lowering, it is possible we reallocate unbacked SymInts. # So we need to line up the unbacked SymInts when performing the test # here # # In principle, we could permit lowering to introduce MORE unbacked # SymInts: as long as all the old unbacked ones are accounted for, # it's fine for inductor to introduce extra calls to item()/unbacked() # whatever. This actually happens in practice when an unbacked SymInt # gets memoized away; naively, when Inductor reprocesses a kernel, it # doesn't know that the memo still applies, and ends up allocating a # new symbol. However, this is generally a bad thing: we may still # end up needing to test equalities on the symbols, and a fresh # symbol is likely to hit lots of GuardOnDataDependent errors that # we already know facts for. renamed_unbacked_bindings = OrderedSet( V.fake_mode.shape_env.unbacked_renamings.get(s, s) for s in unbacked_bindings.keys() ) assert new_unbacked_defs >= renamed_unbacked_bindings, ( f"failed {new_unbacked_defs} >= {renamed_unbacked_bindings} (inductor >= fx)\n" f"fx node is: {n.format_node()}\n" f"new operations are:\n\n{format_new_defs()}" ) return result def validate_can_generate_cpp_wrapper(self) -> None: if config.disable_cpp_codegen: raise CppWrapperCodegenError("C++ codegen is disabled") if sys.platform not in ("linux", "darwin", "win32"): raise CppWrapperCodegenError(f"Unsupported platform {sys.platform}") def init_wrapper_code( self, is_subgraph: bool = False, subgraph_name: Optional[str] = None, parent_wrapper_code: Optional[PythonWrapperCodegen] = None, partition_signatures: Optional[GraphPartitionSignature] = None, ) -> None: device_types = self.device_types.copy() device_types.discard("cpu") device_types.discard("meta") # 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 self.device_type = "cpu" if only_cpu else device_types.pop() if self.cpp_wrapper: self.validate_can_generate_cpp_wrapper() self.device_ops = get_device_op_overrides(self.device_type) wrapper_code_gen_cls = get_wrapper_codegen_for_device( self.device_type, self.cpp_wrapper ) assert wrapper_code_gen_cls is not None, ( f"Device {self.device_type} not supported" ) self.wrapper_code = wrapper_code_gen_cls.create( is_subgraph, subgraph_name, parent_wrapper_code, partition_signatures, ) 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, ) -> tuple[ValueWithLineMap, ValueWithLineMap]: """ For GPU, Triton kernels are autotuned and stored as cubin files """ if any(device in self.device_types for device in ["cuda", "xpu"]): if config.triton.autotune_at_compile_time: # If autotune_at_compile_time is True, we can do the codegen in one-pass # TODO: once autotune_at_compile_time is stable, we should delete the else branch return self.codegen() else: # first pass self.cpp_wrapper = False compiled = self.compile_to_module().call def materialize( x: Union[torch.SymInt, torch.SymFloat, torch.Tensor], ) -> Union[int, float, torch.Tensor]: if x is None: return None elif 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 tracing_context = torch._guards.TracingContext.try_get() if tracing_context is not None and not isinstance( V.real_inputs, NullHandler ): 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: # In the backward pass, V.real_inputs is not OrderedSet. # Generating random inputs based on self.example_inputs sometimes can be problematic, # e.g. illegal memory access. A comprehensive fix is to autotune in a separate process. real_inputs = [ materialize(x) # type:ignore[arg-type] for x in ( self.example_inputs # type:ignore[union-attr] if isinstance(V.real_inputs, NullHandler) else V.real_inputs ) ] if self.mutated_inputs: from .compile_fx import clone_preserve_strides mutated_input_idxs = [ idx for idx, name in enumerate(self.graph_inputs) if name in self.mutated_inputs and isinstance(real_inputs[idx], torch.Tensor) ] for idx in mutated_input_idxs: # clone mutated Tensor inputs to avoid mutating them in # the first pass of the CPP wrapper-based compilation, as # this will lead to a side effect on the example inputs: # e.g. if torch.compile(f)(x) if called on input-mutating # f, the inputs x will be mutated twice in the process: # once here, and again when running the compiled model; # this will also lead to a numerically incorrect output mutated_inp = real_inputs[idx] assert isinstance(mutated_inp, torch.Tensor) real_inputs[idx] = clone_preserve_strides(mutated_inp) del mutated_inp with torch.utils._python_dispatch._disable_current_modes(): compiled(real_inputs) del real_inputs # second pass self.cpp_wrapper = True self.removed_buffers.clear() self.removed_operations.clear() self.inplaced_to_remove.clear() V.graph.sizevars.precomputed_replacements.clear() V.graph.sizevars.inv_precomputed_replacements.clear() metrics.reset() with config.patch({"triton.autotune_at_compile_time": False}): return self.codegen() else: # cpu return self.codegen() def _update_scheduler(self) -> None: """ (Re)initializes the scheduler member. When initializing the scheduler, no CUBIN files should be generated (to avoid biasing any benchmarks and pessimizing fusion decisions). """ from .scheduler import Scheduler with config.patch("triton.store_cubin", False): self.scheduler = Scheduler(self.operations) def codegen(self) -> tuple[ValueWithLineMap, ValueWithLineMap]: with dynamo_timed("GraphLowering.codegen", log_pt2_compile_event=True): self.init_wrapper_code() self._update_scheduler() V.debug.draw_orig_fx_graph(self.orig_gm, self.scheduler.nodes) self.wrapper_code.push_codegened_graph(self) self.scheduler.codegen() log.debug( "Finished codegen for all nodes. The list of kernel names available: %s", V.graph.all_codegen_kernel_names, ) # Dump provenance artifacts for debugging trace provenance_info = ( V.debug.log_inductor_triton_kernel_to_post_grad_node_info() ) # provenance_info might be None if config.trace.enabled is not set if provenance_info: ( debug_info, node_mappings, ) = provenance_info trace_structured( "artifact", metadata_fn=lambda: { "name": "inductor_triton_kernel_to_post_grad_nodes", "encoding": "json", }, payload_fn=lambda: json.dumps(debug_info), ) trace_structured( "artifact", metadata_fn=lambda: { "name": "inductor_provenance_tracking_node_mappings", "encoding": "json", }, payload_fn=lambda: json.dumps(node_mappings), ) result = self.wrapper_code.generate(self.is_inference) self.wrapper_code.pop_codegened_graph() return result def codegen_subgraph(self, parent_graph: GraphLowering) -> None: """ 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()`. """ with dynamo_timed("GraphLowering.codegen_subgraph", log_pt2_compile_event=True): self.wrapper_code = parent_graph.wrapper_code self.device_ops = parent_graph.device_ops self.cpp_wrapper = parent_graph.cpp_wrapper self._update_scheduler() self.scheduler.codegen() def count_bytes( self, ) -> tuple[ int, list[tuple[BaseSchedulerNode, int]], list[tuple[BaseSchedulerNode, float]] ]: total_bytes = 0 node_counts = [] node_runtimes = [] for node in self.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 # No-op to be patched for unit tests save_output_code: Optional[Callable[[str], None]] = None def compile_to_module(self) -> ModuleType: with dynamo_timed( "GraphLowering.compile_to_module", phase_name="code_gen", log_pt2_compile_event=True, dynamo_compile_column_us="inductor_code_gen_cumulative_compile_time_us", ): return self._compile_to_module() def _compile_to_module(self) -> ModuleType: from .codecache import PyCodeCache # Currently, if we're here, we don't have to worry about the kernel code, which # is only available in AOTInductor mode. wrapper_code, _ = ( self.codegen_with_cpp_wrapper() if self.cpp_wrapper else self.codegen() ) if config.triton.autotune_at_compile_time: tuning_code = ( '"""\n' + "Compile-time auto-tuning block: \n" + self.wrapper_code.kernel_autotune_defs.getvalue() + self.wrapper_code.kernel_autotune_calls.getvalue() + '"""\n' ) wrapper_code.value = tuning_code + wrapper_code.value if GraphLowering.save_output_code is not None: GraphLowering.save_output_code(wrapper_code.value) output_code_log.debug("Output code: \n%s", wrapper_code.value) inductor_meta = autotune_cache.inductor_meta_from_config() AutotuneCacheBundler.begin_compile(inductor_meta, code=wrapper_code.value) try: linemap = [ (line_no, node.stack_trace) # type: ignore[attr-defined] for line_no, node in wrapper_code.line_map ] key, path = PyCodeCache.write(wrapper_code.value) output_code_log.debug("Output code written to: %s", path) except Exception: trace_structured( "inductor_output_code", # Just omit the filename, I still want the code though! payload_fn=lambda: wrapper_code.value, ) raise else: trace_structured( "inductor_output_code", lambda: {"filename": path}, payload_fn=lambda: wrapper_code.value, ) with dynamo_timed("PyCodeCache.load_by_key_path", log_pt2_compile_event=True): mod = PyCodeCache.load_by_key_path( key, path, linemap=linemap, # type: ignore[arg-type] attrs={**self.constants, **self.torchbind_constants}, ) self.cache_key = key self.cache_path = path self.cache_linemap = linemap # type: ignore[assignment] if config.benchmark_harness and config.profile_bandwidth_output: # run the inputs code gen to get the bandwidth info mod.benchmark_compiled_module(times=1, repeat=1) # 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.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 get_output_names(self) -> list[str]: names = [] shape_counter = itertools.count(0) none_counter = itertools.count(0) for node in self.graph_outputs: if isinstance(node, ir.NoneAsConstantBuffer): names.append(f"{self.name}_none{next(none_counter)}") elif isinstance(node, ir.ShapeAsConstantBuffer): names.append(f"{self.name}_shape{next(shape_counter)}") else: names.append(node.get_name()) return names def is_unspec_arg(self, name: str) -> bool: # 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 len(self.graph_inputs[name].get_size()) == 0 and get_device_type(self.graph_inputs[name]) == "cpu" ) or name in self.zero_dim_cpu_tensor_list class SubgraphLowering(GraphLowering): """ Mostly a helper class for the subgraph lowering. The main goal is to call init_wrapper_code with the subgraph related arguments. """ def __init__(self, parent: GraphLowering, *args: Any, **kwargs: Any) -> None: self.parent = parent super().__init__(*args, **kwargs) def init_wrapper_code( self, is_subgraph: bool = False, subgraph_name: Optional[str] = None, parent_wrapper_code: Optional[PythonWrapperCodegen] = None, partition_signatures: Optional[GraphPartitionSignature] = None, ) -> None: super().init_wrapper_code( is_subgraph=True, subgraph_name=self.name, parent_wrapper_code=self.parent.wrapper_code, )