# mypy: allow-untyped-defs from __future__ import annotations import dataclasses import functools import itertools import logging import os import textwrap from functools import lru_cache from typing import ( Any, Callable, cast, Dict, Iterable, List, Optional, Tuple, TYPE_CHECKING, Union, ) import sympy import torch import torch._logging from torch._dynamo.utils import preserve_rng_state from torch._inductor.runtime.hints import AutotuneHint, DeviceProperties from torch._prims_common import is_integer_dtype from torch.utils._ordered_set import OrderedSet from torch.utils._sympy.functions import CeilDiv, FloorDiv, ModularIndexing from torch.utils._triton import has_triton_package from ...utils._sympy.symbol import free_symbol_is_type, prefix_str, symbol_is_type, SymT from ...utils._sympy.value_ranges import ValueRanges from .. import config, ir from ..codecache import code_hash, get_path, PyCodeCache from ..metrics import is_metric_table_enabled, log_kernel_metadata from ..runtime.benchmarking import benchmarker from ..runtime.hints import ReductionHint, TRITON_MAX_BLOCK from ..runtime.runtime_utils import get_max_y_grid, next_power_of_2 from ..utils import ( cache_on_self, get_bounds_index_expr, get_fused_kernel_name, get_kernel_metadata, is_welford_reduction, Placeholder, sympy_dot, sympy_subs, ) from ..virtualized import _ops as ops, OpsHandler, ReductionType, StoreMode, V from ..wrapper_benchmark import get_kernel_category_by_source_code from .common import ( BackendFeature, CSE, CSEVariable, DeferredLine, IndentedBuffer, OpOverrides, PythonPrinter, SizeArg, TensorArg, WorkspaceArg, ) from .simd import ( constant_repr, IterationRangesEntry, IterationRangesRoot, pexpr, SIMDKernel, SIMDScheduling, ) from .triton_utils import ( config_of, should_unwrap_unspec_arg, signature_of, signature_to_meta, ) if TYPE_CHECKING: from ..ir import IRNode log = logging.getLogger(__name__) perf_hint_log = torch._logging.getArtifactLogger(__name__, "perf_hints") schedule_log = torch._logging.getArtifactLogger(__name__, "schedule") fusion_log = torch._logging.getArtifactLogger(__name__, "fusion") @lru_cache(None) def gen_attr_descriptor_import(): """ import AttrsDescriptor if the triton version is new enough to have this class defined. """ if not has_triton_package(): return "" import triton.compiler.compiler if hasattr(triton.compiler.compiler, "AttrsDescriptor"): return "from triton.compiler.compiler import AttrsDescriptor" else: return "" @lru_cache(None) def gen_common_triton_imports(): imports = IndentedBuffer() imports.splice( """ import triton import triton.language as tl """ ) if attr_desc := gen_attr_descriptor_import(): imports.writeline(attr_desc) imports.splice( """ from torch._inductor.runtime import triton_helpers, triton_heuristics from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties """ ) return imports.getvalue() block_offsets = { symt: sympy.Symbol(f"{prefix_str[symt]}offset", integer=True, nonnegative=True) for symt in [SymT.XBLOCK, SymT.YBLOCK, SymT.RINDEX] } block_sizes = { symt: sympy.Symbol(f"{prefix_str[symt].upper()}BLOCK", integer=True, positive=True) for symt in [SymT.XBLOCK, SymT.YBLOCK, SymT.RINDEX] } @dataclasses.dataclass class IndexingOptions: index_str: str mask_vars: OrderedSet[str] mask_str: str expand_str: Optional[str] _has_rindex: bool index: sympy.Expr def has_mask(self): return bool(self.mask_vars) def has_indirect(self): return free_symbol_is_type(self.index, SymT.TMP) def has_rindex(self): return self._has_rindex def has_tmpmask(self): return "tmp" in self.mask_str def has_rmask(self): return "rmask" in self.mask_str @dataclasses.dataclass class BlockPtrOptions: params: BlockParameters constant_offset: sympy.Expr order: List[int] mask_vars: OrderedSet[str] reshape_suffix: List[str] @property def shape(self) -> List[sympy.Expr]: return self.params.shape @property def block_shape(self) -> List[sympy.Expr]: return self.params.block_shape @property def strides(self) -> List[sympy.Expr]: return self.params.strides @property def offsets(self) -> List[sympy.Expr]: return self.params.offsets @staticmethod def create( *, params: BlockParameters, constant_offset: sympy.Expr, range_trees: List[IterationRangesEntry], mask_vars: OrderedSet[str], ) -> BlockPtrOptions: """Helper to create a BlockPtrOptions instance""" reshape_suffix = [f"{t.prefix.upper()}BLOCK" for t in range_trees] # Only drop broadcast dims if the output has the same # rank as the block. Otherwise, we will get shape errors. drop_broadcasts = len(reshape_suffix) == len(params.strides) broadcasting_dim = [s == 0 for s in params.strides] for i, is_broadcasting in enumerate(broadcasting_dim): if is_broadcasting and drop_broadcasts: # drop any stride==0 dimensions for performance reshape_suffix[i] = "1" if V.kernel.no_x_dim: assert range_trees[0].prefix == "x" reshape_suffix.pop(0) if ( not V.kernel.inside_reduction and len(params.strides) == len(V.kernel.numels) - 1 and V.kernel.numels[-1] != 1 ): # Need to expand rank by 1 to match rank when self.inside_reduction=True reshape_suffix.append("1") def filter(it): """Removes any broadcasting dims from a given sequence""" assert len(it) == len(broadcasting_dim) return [ item for item, is_broadcasting in zip(it, broadcasting_dim) if not is_broadcasting or not drop_broadcasts ] # Drop broadcasting dimensions from the input. params = BlockParameters( **{key: filter(val) for key, val in dataclasses.asdict(params).items()} ) def lookup_size(exprs: Iterable[sympy.Expr]) -> List[sympy.Expr]: return [V.graph.sizevars.lookup_precomputed_size(expr) for expr in exprs] # Look up precomputed sizes params.shape = lookup_size(params.shape) params.strides = lookup_size(params.strides) return BlockPtrOptions( params=params, constant_offset=V.graph.sizevars.lookup_precomputed_size(constant_offset), order=list(reversed(range(len(params.shape)))), mask_vars=mask_vars, reshape_suffix=reshape_suffix, ) def replace_roffset(self, expr: sympy.Expr, replacement: sympy.Expr) -> sympy.Expr: """ Replaces instances of roffset with the new expression. """ roffset = block_offsets[SymT.RINDEX] return sympy_subs(expr, {roffset: replacement}) def format(self, name: str, roffset=True) -> str: """ Codegen a call to tl.make_block_ptr() Args: name: variable name for pointer roffset: should roffset be included in offsets=..., for use with tl.advance() Returns: "tl.make_block_ptr(...)" """ f = V.kernel.index_to_str offsets = [*self.offsets] if not roffset: offsets = [ self.replace_roffset(offset, sympy.Integer(0)) for offset in offsets ] args = [ f"{name} + ({f(self.constant_offset)})" if self.constant_offset != 0 else name, f"shape={f(self.shape)}", f"strides={f(self.strides)}", f"block_shape={f(self.block_shape)}", f"order={f(self.order)}", f"offsets={f(offsets)}", ] return f"tl.make_block_ptr({', '.join(args)})" @cache_on_self def boundary_check(self) -> List[int]: """List of indices to pass to tl.load(boundary_check=...)""" sizevars = V.graph.sizevars # Substitute maximum block sizes in shape expressions. # This works in multiple_of checks because block sizes are powers of 2. block_to_max: Dict[sympy.Expr, Any] = { block_size: TRITON_MAX_BLOCK[prefix_str[symt].upper()] for symt, block_size in block_sizes.items() } return [ idx for idx in range(len(self.shape)) if ( not sizevars.statically_known_equals( self.strides[idx], sympy.Integer(0) ) and not sizevars.statically_known_multiple_of( self.shape[idx], self.block_shape[idx] ) and not sizevars.statically_known_multiple_of( self.shape[idx], sympy_subs(self.block_shape[idx], block_to_max) ) and not ( V.kernel.no_x_dim and self.block_shape[idx] == block_sizes[SymT.XBLOCK] ) ) ] def advance_roffset(self): """ Codegen string to pass to tl.advance(name, ...). Advance is the difference between offsets in each loop iteration. To compute it, we replace roffset with multiples of RBLOCK. Since we expect roffset to vary in range(0, rnumel, RBLOCK), the first iteration has roffset=0, while the second has roffset=RBLOCK. """ rblock = block_sizes[SymT.RINDEX] advance = [ ( self.replace_roffset(offset, rblock) - self.replace_roffset(offset, sympy.Integer(0)) ) for offset in self.offsets ] return V.kernel.index_to_str(advance) def has_indirect(self): return False # block_ptr can't do indirect indexing def has_rindex(self) -> bool: return any(free_symbol_is_type(expr, SymT.RINDEX) for expr in self.block_shape) def has_rmask(self): return self.has_rindex() def has_tmpmask(self): return False # block_ptr can't do indirect indexing def has_mask(self): return bool(self.boundary_check()) def triton_reshape(value: str, old_shape: List[str], new_shape: List[str]): """Workaround https://github.com/openai/triton/issues/2836""" assert isinstance(old_shape, list) and isinstance(new_shape, list) if old_shape == new_shape: return value if [s for s in new_shape if s != "1"] != old_shape: return f"tl.reshape({value}, [{', '.join(new_shape)}])" # rewrite to [:, None] syntax, which is less buggy idx = 0 expand = [] for size in new_shape: if idx < len(old_shape) and size == old_shape[idx]: expand.append(":") idx += 1 else: assert size == "1" expand.append("None") assert idx == len(old_shape) return f"{value}[{', '.join(expand)}]" # NB: Inheriting from PythonPrinter is somewhat dangerous, because there are a # number of operators which Triton "implements", but in a way that is # inconsistent with Python semantics (and consistent with C semantics). We # must override all of these, or it is potential silent correctness problem class TritonPrinter(PythonPrinter): def _print_TruncToInt(self, expr): assert len(expr.args) == 1 return ( f"libdevice.trunc({self._print(expr.args[0])}).to({V.kernel.index_dtype})" ) def _print_ToFloat(self, expr): assert len(expr.args) == 1 return f"{self.paren(self._print(expr.args[0]))}.to(tl.float64)" def _print_PythonMod(self, expr): quot, div = expr.args quot_s = self._print(quot) div_s = self._print(div) if quot.is_nonnegative and div.is_nonnegative: return f"{self.paren(quot_s)} % {self.paren(div_s)}" return f"triton_helpers.remainder_integer({quot_s}, {div_s})" def _print_FloorDiv(self, expr): assert expr.is_integer quot, div = expr.args quot_s = self._print(quot) div_s = self._print(div) if quot.is_nonnegative and div.is_nonnegative: return f"({self.paren(quot_s)} // {self.paren(div_s)})" return f"triton_helpers.div_floor_integer({quot_s}, {div_s})" # TODO: This is wrong, when lhs, rhs > 2**53, Python does a higher # precision algorithm, which we would need to replicate here def _print_IntTrueDiv(self, expr): lhs, rhs = expr.args return f"{self.paren(self._print(lhs))} / {self.paren(self._print(rhs))}" # NB: sympy.floor/ceiling produce integers, so we have to do the # conversion to index dtype def _print_floor(self, expr): assert len(expr.args) == 1 return ( f"libdevice.floor({self._print(expr.args[0])}).to({V.kernel.index_dtype})" ) def _print_FloorToInt(self, expr): assert len(expr.args) == 1 return ( f"libdevice.floor({self._print(expr.args[0])}).to({V.kernel.index_dtype})" ) def _print_ceiling(self, expr): assert len(expr.args) == 1 return f"libdevice.ceil({self._print(expr.args[0])}).to({V.kernel.index_dtype})" def _print_CeilToInt(self, expr): assert len(expr.args) == 1 return f"libdevice.ceil({self._print(expr.args[0])}).to({V.kernel.index_dtype})" def _helper_sqrt(self, expr): return f"libdevice.sqrt({self._print(expr)}.to(tl.float32))" def _print_FloatPow(self, expr): return ( f"libdevice.pow({self._print(expr.args[0])}, {self._print(expr.args[1])})" ) _print_PowByNatural = _print_FloatPow def _print_Where(self, expr): c = self.doprint(expr.args[0]) p = self.doprint(expr.args[1]) q = self.doprint(expr.args[2]) return f"tl.where({c}, {p}, {q})" def _print_min_max_helper(self, expr: sympy.Expr, cmp: str) -> str: """ Helper for max/min code genereration. cmp: > or < """ nargs = len(expr.args) if len(expr.args) == 1: return self._print(expr.args[0]) mid = len(expr.args) // 2 cls = type(expr) a = self._print(cls(*expr.args[:mid])) b = self._print(cls(*expr.args[mid:])) # Use a macro so we can propagate constexprs. # https://github.com/triton-lang/triton/issues/3815 a, b = tuple(f"({x})" for x in (a, b)) assert cmp in (">", "<"), f"Unexpected comparator: '{cmp}'" return f"({a} * ({a} {cmp}= {b}) + {b} * ({b} {cmp} {a}))" def _print_Min(self, expr): return self._print_min_max_helper(expr, "<") def _print_Max(self, expr): return self._print_min_max_helper(expr, ">") def _print_Abs(self, expr): assert len(expr.args) == 1 return f"tl_math.abs({self._print(expr.args[0])})" def _print_OpaqueUnaryFn_cos(self, expr): assert len(expr.args) == 1 return f"libdevice.cos(({self._print(expr.args[0])}).to(tl.float32))" def _print_OpaqueUnaryFn_cosh(self, expr): assert len(expr.args) == 1 return f"libdevice.cosh(({self._print(expr.args[0])}).to(tl.float32))" def _print_OpaqueUnaryFn_acos(self, expr): assert len(expr.args) == 1 return f"libdevice.acos(({self._print(expr.args[0])}).to(tl.float32))" def _print_OpaqueUnaryFn_sin(self, expr): assert len(expr.args) == 1 return f"libdevice.sin(({self._print(expr.args[0])}).to(tl.float32))" def _print_OpaqueUnaryFn_sinh(self, expr): assert len(expr.args) == 1 return f"libdevice.sinh(({self._print(expr.args[0])}).to(tl.float32))" def _print_OpaqueUnaryFn_asin(self, expr): assert len(expr.args) == 1 return f"libdevice.asin(({self._print(expr.args[0])}).to(tl.float32))" def _print_OpaqueUnaryFn_tan(self, expr): assert len(expr.args) == 1 return f"libdevice.tan(({self._print(expr.args[0])}).to(tl.float32))" def _print_OpaqueUnaryFn_tanh(self, expr): assert len(expr.args) == 1 return f"libdevice.tanh(({self._print(expr.args[0])}).to(tl.float32))" def _print_OpaqueUnaryFn_atan(self, expr): assert len(expr.args) == 1 return f"libdevice.atan(({self._print(expr.args[0])}).to(tl.float32))" def _print_RoundToInt(self, expr): assert len(expr.args) == 1 return f"libdevice.llrint({self._print(expr.args[0])})" def _print_RoundDecimal(self, expr): assert len(expr.args) == 2 number, ndigits = expr.args if number.is_integer: # ndigits < 0 should have been filtered by the sympy function assert ndigits < 0 raise ValueError( f"For integer inputs, only non-negative ndigits are currently supported, but got {ndigits}." ) return f"libdevice.nearbyint(1e{ndigits} * {self.paren(self._print(number))}) * 1e{-ndigits}" texpr = TritonPrinter().doprint def triton_compute_type(dtype): triton_type_name = str(dtype).split(".")[-1] if triton_type_name == "bool": triton_type_name = "int1" elif ( triton_type_name in ("float16", "bfloat16") and config.triton.codegen_upcast_to_fp32 ): # float16 math is done in float32 inside the kernel triton_type_name = "float32" elif triton_type_name == "float8_e4m3fn": triton_type_name = "float8e4nv" elif triton_type_name == "float8_e5m2": triton_type_name = "float8e5" elif triton_type_name == "float8_e4m3fnuz": triton_type_name = "float8e4b8" elif triton_type_name == "float8_e5m2fnuz": triton_type_name = "float8e5b16" return f"tl.{triton_type_name}" def _get_primitive_bitwidth(dtype): if hasattr(dtype, "is_floating_point"): if dtype.is_floating_point: # triton_compute_type changes the bitwidth if ( dtype in [torch.bfloat16, torch.float16] and config.triton.codegen_upcast_to_fp32 ): return 32 return torch.finfo(dtype).bits else: return torch.iinfo(dtype).bits else: return -1 def triton_store_type(dtype): triton_type_name = str(dtype).split(".")[-1] if triton_type_name == "bool": triton_type_name = "int8" elif triton_type_name == "float8_e4m3fn": triton_type_name = "float8e4nv" elif triton_type_name == "float8_e5m2": triton_type_name = "float8e5" return f"tl.{triton_type_name}" def triton_acc_type(dtype): if is_integer_dtype(dtype) and dtype.is_signed: nbits = 64 if dtype == torch.int64 else 32 return f"tl.int{nbits}" return triton_compute_type(dtype) class TritonCSEVariable(CSEVariable): def __init__(self, name, bounds: ValueRanges[Any]) -> None: super().__init__(name, bounds) # We'll use this to track which masks the variable needs when used for indirect indexing self.mask_vars: OrderedSet[str] = OrderedSet() def update_on_args(self, name, args, kwargs): for arg in args: if isinstance(arg, TritonCSEVariable): self.mask_vars.update(arg.mask_vars) elif isinstance(arg, sympy.Symbol) and arg.name[0] in "xyr": # most of the time index vars don't need masks associated with them # however, when index vars are used to compute indices for indirect reads # those reads should subsequently be masked, self.mask_vars.update({f"{arg.name[0]}mask"}) class TritonOverrides(OpOverrides): """Map element-wise ops to Triton""" @staticmethod def to_dtype( x, dtype: torch.dtype, src_dtype: Optional[torch.dtype] = None, use_compute_types=True, ): def _get_min_elements_per_thread( src_dtype: torch.dtype, dst_dtype: torch.dtype ) -> int: if src_dtype == dst_dtype: # No data type conversion is needed. No requirements on min_elem_per_thread. return 0 # fp8 data type conversions has min_elem_per_thread requirements. # Refer to Triton implementations here: # https://github.com/openai/triton/blob/10f59d8ce04052521c1bc0cb3a3f8b98918fc7e3/lib/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVM.cpp#L10. fp8_dtypes = ( torch.float8_e4m3fn, torch.float8_e5m2, ) # Triton doesn't support type conversions between fp8_e4m3 and fp8_e5m2. assert not ( src_dtype in fp8_dtypes and dst_dtype in fp8_dtypes and src_dtype != dst_dtype ), "Conversions between float8_e5m2 and float8_e4m3fn is not supported!" if src_dtype == torch.float8_e5m2 or dst_dtype == torch.float8_e5m2: return 4 if src_dtype == torch.float8_e4m3fn or dst_dtype == torch.float8_e4m3fn: return 2 # No requirements on min_elem_per_thread. return 0 if src_dtype is not None: # Both dtype and src_dtype are set. This is used by torch to(dtype=dtype). # It takes the maximum min_elem_per_thread if there are multiple fp8 conversions # in the same kernel. V.kernel.min_elem_per_thread = max( _get_min_elements_per_thread(src_dtype, dtype), V.kernel.min_elem_per_thread, ) if dtype == torch.bool: return f"({x} != 0)" elif dtype == torch.uint8: # to work around llvm uint conversion semantics # that produces 0's for negative values return f"{x}.to(tl.int8).to(tl.uint8)" if use_compute_types: out_dtype = triton_compute_type(dtype) else: out_dtype = triton_store_type(dtype) return f"{x}.to({out_dtype})" @staticmethod def to_dtype_bitcast(x, dtype: torch.dtype, src_dtype: torch.dtype): triton_dtype = triton_compute_type(dtype) # We may promote float16 or bfloat16 to float32 and cause the # bitwidth of dtype to be different from the input tensor (i.e. float32). # In such as case, we will have to convert the input tensor to # its src_type, perform bitcast, and then convert the bit-casted # tensor back to float to ensure we use values with the right precision. if ( src_dtype in (torch.float16, torch.bfloat16) and config.triton.codegen_upcast_to_fp32 ): triton_src_dtype = str(src_dtype).split(".")[-1] cast_x = f"{x}.to(tl.{triton_src_dtype})" if dtype in (torch.float16, torch.bfloat16): triton_type_name = str(dtype).split(".")[-1] triton_dtype = f"tl.{triton_type_name}" cast_x = f"{cast_x}.to({triton_dtype}, bitcast=True)" return f"{cast_x}.to(tl.float32)" else: src_dtype_bitwidth = _get_primitive_bitwidth(src_dtype) target_dtype_bitwidth = _get_primitive_bitwidth(dtype) bitcast = "True" if src_dtype_bitwidth == target_dtype_bitwidth else "False" return f"{x}.to({triton_dtype}, bitcast={bitcast})" @staticmethod def _shaped_constant(value, dtype, shape): type_ = torch._prims_common.dtype_to_type(dtype) triton_val = constant_repr(type_(value)) triton_type = triton_compute_type(dtype) if triton_type == "tl.float32": # Float constants are always f32 in triton return triton_val # NOTE: We use a tensor here in order to get the expected type. # Otherwise, e.g. float64 constants would be trunctated to float32. return f"tl.full({shape}, {triton_val}, {triton_type})" @classmethod def constant(cls, value, dtype): return cls._shaped_constant(value, dtype, shape=[]) @staticmethod def abs(x): return f"tl_math.abs({x})" @staticmethod def libdevice_abs(x): return f"libdevice.abs({x})" @staticmethod def exp(x): return f"tl_math.exp({x})" @staticmethod def libdevice_exp(x): return f"libdevice.exp({x})" @staticmethod def exp2(x): return f"libdevice.exp2({x})" @staticmethod def expm1(x): return f"libdevice.expm1({x})" @staticmethod def sqrt(x): return f"libdevice.sqrt({x})" @staticmethod def libdevice_sqrt(x): return f"libdevice.sqrt({x})" @staticmethod def relu(x): bug = config.triton.inject_relu_bug_TESTING_ONLY if bug == "compile_error": return "compile error!" elif bug == "runtime_error": # NB: this only triggers runtime error as long as input # is not all zero return f'triton_helpers.device_assert_then({x} == 0, "injected assert fail", {x})' elif bug == "accuracy": return f"{x} + 1" elif bug is None: return ops.maximum(ops.constant(0, torch.int32), x) else: raise AssertionError( f"unrecognized config triton.inject_relu_bug_TESTING_ONLY = {bug!r}" ) @staticmethod def minimum(a, b): return f"triton_helpers.minimum({a}, {b})" @staticmethod def maximum(a, b): return f"triton_helpers.maximum({a}, {b})" @staticmethod def where(a, b, c): return f"tl.where({a}, {b}, {c})" @staticmethod def inline_asm_elementwise( *inputs, asm, constraints=None, dtype=torch.float32, is_pure=True, pack=1 ): triton_type = triton_compute_type(dtype) input_refs = ", ".join([str(i) for i in inputs]) if constraints is None: constraints = ", ".join(["=r"] + ["r" for _ in inputs]) return f"tl.inline_asm_elementwise('{asm}', '{constraints}', [{input_refs}], dtype={triton_type}, is_pure={is_pure}, pack={pack})" # noqa: B950 @staticmethod def cos(x): return f"tl_math.cos({x})" @staticmethod def libdevice_cos(x): return f"libdevice.cos({x})" @staticmethod def sin(x): return f"tl_math.sin({x})" @staticmethod def libdevice_sin(x): return f"libdevice.sin({x})" @classmethod def index_expr(cls, expr, dtype): raise NotImplementedError("ops.index_expr not implemented outside a kernel") @staticmethod def masked(mask, body, other): raise NotImplementedError("ops.masked not implemented outside a kernel") @staticmethod def lgamma(x): return f"libdevice.lgamma({x})" @staticmethod def erf(x): return f"libdevice.erf({x})" @staticmethod def cosh(x): return f"libdevice.cosh({x})" @staticmethod def sinh(x): return f"libdevice.sinh({x})" @staticmethod def acos(x): return f"libdevice.acos({x})" @staticmethod def acosh(x): return f"libdevice.acosh({x})" @staticmethod def asin(x): return f"libdevice.asin({x})" @staticmethod def asinh(x): return f"libdevice.asinh({x})" @staticmethod def atan2(x, y): return f"libdevice.atan2({x}, {y})" @staticmethod def atan(x): return f"libdevice.atan({x})" @staticmethod def atanh(x): return f"libdevice.atanh({x})" @staticmethod def copysign(x, y): return f"libdevice.copysign({x}, {y})" @staticmethod def erfc(x): return f"libdevice.erfc({x})" @staticmethod def erfinv(x): return f"libdevice.erfinv({x})" @staticmethod def hypot(x, y): return f"libdevice.hypot({x}, {y})" @staticmethod def log10(x): return f"libdevice.log10({x})" @staticmethod def log2(x): return f"libdevice.log2({x})" @staticmethod def nextafter(x, y): return f"libdevice.nextafter({x}, {y})" @staticmethod def logical_and(a, b): return f"{a} & {b}" @staticmethod def logical_not(a): return f"{a} == 0" @staticmethod def logical_or(a, b): return f"{a} | {b}" @staticmethod def logical_xor(a, b): return f"({a} ^ {b})" @staticmethod def bitwise_and(a, b): return f"{a} & {b}" @staticmethod def bitwise_not(a): return f"~{a}" @staticmethod def bitwise_or(a, b): return f"{a} | {b}" @staticmethod def bitwise_xor(a, b): return f"{a} ^ {b}" @staticmethod def bitwise_left_shift(a, b): return f"{a} << {b}" @staticmethod def bitwise_right_shift(a, b): return f"{a} >> {b}" @staticmethod def rand(seed, offset): offset = f"({offset}).to(tl.uint32)" return f"tl.rand({seed}, {offset})" @staticmethod def randn(seed, offset): offset = f"({offset}).to(tl.uint32)" return f"tl.randn({seed}, {offset})" @staticmethod def randint64(seed, offset, low, high): offset = f"({offset}).to(tl.uint32)" return f"triton_helpers.randint64({seed}, {offset}, {low}, {high})" @staticmethod def load_seed(name, offset): raise NotImplementedError("ops.load_seed not implemented outside a kernel") @staticmethod def rsqrt(x): return f"libdevice.rsqrt({x})" @staticmethod def log1p(x): return f"libdevice.log1p({x})" @staticmethod def tan(x): return f"libdevice.tan({x})" @staticmethod def tanh(x): return f"libdevice.tanh({x})" @staticmethod def sigmoid(x): return f"tl.sigmoid({x})" @staticmethod def signbit(x): # XX: This is wrong for the value -0.0 in floating point return f"libdevice.signbit({x}) if ({x}).dtype is tl.float32 else {x} < 0" @staticmethod def fmod(a, b): return f"libdevice.fmod({a}, {b})" @staticmethod def pow(a, b): return f"libdevice.pow({a}, {b})" @staticmethod def log(x): return f"tl_math.log({x})" @staticmethod def libdevice_log(x): return f"libdevice.log({x})" @staticmethod def isinf(x): return f"libdevice.isinf({x}).to(tl.int1)" @staticmethod def isnan(x): return f"libdevice.isnan({x}).to(tl.int1)" @staticmethod def round(x): return f"libdevice.nearbyint({x})" @staticmethod def floor(x): return f"libdevice.floor({x})" @staticmethod def floordiv(a, b): # See the comment in lowering.div_mode. a and b are integer type. # Similar to div_floor_kernel_cuda in pytorch core. # Notice that // in triton behaves as truncdiv instead of floordiv quot = f"{a} // {b}" rem = f"{a} % {b}" return f"tl.where(({a} < 0) != ({b} < 0), tl.where({rem} != 0, {quot} - 1, {quot}), {quot})" @staticmethod def sign(x): z = ops.constant(0, torch.int32) left = ops.to_dtype((ops.lt(z, x)), torch.int8) right = ops.to_dtype((ops.lt(x, z)), torch.int8) sub = ops.sub(left, right) return f"{sub}.to({x}.dtype)" @staticmethod def trunc(x): return f"libdevice.trunc({x})" @staticmethod def truncdiv(a, b): # See the comment in lowering.div_mode. a and b are integer type. # Notice that // in triton behaves as truncdiv instead of floordiv return f"{a} // {b}" @staticmethod def ceil(x): return f"libdevice.ceil({x})" TritonOverrides._initialize_pointwise_overrides("triton") # Use mypy to check protocol implemented correctly def _typecheck_TritonOverrides(h: TritonOverrides) -> OpsHandler[str]: return h class TritonKernelOverrides(TritonOverrides): """Map element-wise ops to Triton within a TritonKernel Unlike TritonOverrides, these assume the code is going to be inserted into the body of the main triton kernel and so it may use indexing and mask variables which are assumed to already be defined in the current scope. """ @classmethod def constant(cls, value, dtype): # NOTE: Cannot use shape=[] as it's not supported by triton-rocm # We could use shape=[1] instead but starting with the correct # ndim avoids extra `tt.expand_dim` ops appearing in the triton IR. ndim = V.kernel.triton_tensor_ndim() shape = [1] * ndim return cls._shaped_constant(value, dtype, shape=shape) @classmethod def index_expr(cls, expr, dtype): indexing = V.kernel.indexing(expr, block_ptr=False) assert isinstance(indexing, IndexingOptions) var = V.kernel.cse.generate( V.kernel.compute, indexing.index_str, bounds=get_bounds_index_expr(expr) ) if dtype not in (torch.int32, torch.int64): var = V.kernel.cse.generate(V.kernel.compute, cls.to_dtype(var, dtype)) var.mask_vars = indexing.mask_vars return var @staticmethod def masked(mask, body, other): if mask is not None and torch.version.hip is not None: mask = V.kernel.cse.generate( V.kernel.compute, f"{mask}.to(tl.int1)", ) nodes = body.graph.find_nodes(op="output") assert nodes, "graph for body does not contain an output" need_where = False for node in nodes: for arg in node.args: if arg.target != "load" or should_unwrap_unspec_arg(arg.args[0]): need_where = True value = None if need_where else other with V.kernel.mask_loads(mask, value=value) as new_mask: result = body() if need_where: # Remove once CSEVariables track the dtype if result.bounds.is_bool: other = bool(other) # Take dtype from result to prevent accidental promotion other = V.kernel.cse.generate( V.kernel.compute, f"tl.full({result}.shape, {constant_repr(other)}, {result}.dtype)", bounds=ValueRanges.wrap(other), ) ret = ops.where(new_mask, result, other) else: ret = result ret.mask_vars.discard(new_mask) return ret @staticmethod def load_seed(name, offset): var = V.kernel.args.input(name) return ( f"tl.load({var} + {V.kernel.args.seed_offset('load_seed_offset', offset)})" ) @staticmethod def frexp(x): cache_key = f"frexp({x})" if cache_key in V.kernel.cse.cache: return V.kernel.cse.cache[cache_key] mantissa = V.kernel.cse.newvar() exponent = V.kernel.cse.newvar() V.kernel.compute.writeline( f"{mantissa}, {exponent} = triton_helpers.frexp({x})" ) V.kernel.cse.cache[cache_key] = (mantissa, exponent) return (mantissa, exponent) # Use mypy to check protocol implemented correctly def _typecheck_TritonKernelOverrides(h: TritonKernelOverrides) -> OpsHandler[str]: return h class HelperFunctions: """An ordered set of helper functions.""" _templates_seen: Dict[str, str] # Template code to function name finalized_helpers: List[str] def __init__(self) -> None: self._templates_seen = {} self.finalized_helpers = [] def add(self, template_code: str, *, base_name="_triton_helper_fn") -> str: """This accepts a function definition with the function name left as a format specifier e.g. @triton.jit def {name}(arg0, arg1): return arg0 + arg1 We add the templated code to the function set and return the name assigned to that function. """ existing_name = self._templates_seen.get(template_code) if existing_name is not None: # Don't duplicate existing helpers return existing_name name = f"{base_name}{len(self.finalized_helpers)}" self._templates_seen[template_code] = name self.finalized_helpers.append(template_code.format(name=name)) return name def __iter__(self): return iter(self.finalized_helpers) def __getitem__(self, idx): return self.finalized_helpers[idx] @dataclasses.dataclass class BlockParameters: """ Class representing ND block dimensions, for block pointer analysis. """ shape: List[sympy.Expr] = dataclasses.field(default_factory=list) block_shape: List[sympy.Expr] = dataclasses.field(default_factory=list) strides: List[sympy.Expr] = dataclasses.field(default_factory=list) offsets: List[sympy.Expr] = dataclasses.field(default_factory=list) def __add__(self, other: BlockParameters) -> BlockParameters: """ Concatenates block parameters. """ cls = type(self) a, b = tuple(dataclasses.asdict(x) for x in (self, other)) return cls(**{key: a[key] + b[key] for key in a}) class TritonKernel(SIMDKernel): overrides = TritonKernelOverrides # type: ignore[assignment] helper_functions: HelperFunctions kexpr: Callable[[sympy.Expr], str] = texpr allow_block_ptr = True def __init__( self, *groups, index_dtype: str, mutations: Optional[OrderedSet[str]] = None, pid_cache=None, reduction_hint=ReductionHint.DEFAULT, min_elem_per_thread=0, override_persistent_reduction=None, optimize_mask=True, ) -> None: self.optimize_mask: bool = optimize_mask super().__init__( *groups, index_dtype=index_dtype, mutations=mutations, reduction_hint=reduction_hint, pid_cache=pid_cache, override_persistent_reduction=override_persistent_reduction, ) self.suffix: IndentedBuffer = IndentedBuffer() # type: ignore[assignment] self.outside_loop_vars: OrderedSet[Any] = OrderedSet() self.min_elem_per_thread = min_elem_per_thread self.block_ptr_id = itertools.count() self.helper_functions = HelperFunctions() # A set of autotuning hints to pass as part of triton_meta self.autotune_hints: OrderedSet[AutotuneHint] = OrderedSet() self.triton_meta: Optional[Dict[str, object]] = None self.codegen_range_tree() def _get_symt(self, tree: IterationRangesEntry) -> SymT: prefix_to_symt = {prefix: symt for symt, prefix in prefix_str.items()} return prefix_to_symt[tree.prefix] def _get_block_size(self, tree: IterationRangesEntry) -> sympy.Symbol: return block_sizes[self._get_symt(tree)] def _get_block_offset(self, tree: IterationRangesEntry) -> sympy.Symbol: return block_offsets[self._get_symt(tree)] def _max_block_size(self, tree: IterationRangesEntry) -> int: return TRITON_MAX_BLOCK[tree.prefix.upper()] def codegen_range_tree(self): for tree in self.range_trees: # reduction indexing goes inside a loop if not tree.is_loop: self.iteration_ranges_codegen_header(tree, self.body) if self.inside_reduction and self.range_trees[-1].is_loop: # workaround for this issue: # https://gist.github.com/jansel/6527126f781559095c5531f98a4235a7 self.body.writeline( f"rbase = {self.iteration_ranges_ranges_code(self.range_trees[-1])}" ) def need_numel_args(self): r""" Indicate whether we need provide numel as arguments for the generated kernel calls in the benchmark. Should be true for pointwise/reduction kernels but false for triton matmul kernels. """ return True def should_use_persistent_reduction(self) -> bool: """ Heuristic to set self.persistent_reduction and add guards if needed. """ if not (self.inside_reduction and config.triton.persistent_reductions): return False threshold = { ReductionHint.INNER: 1024, }.get(self.reduction_hint, 64) # If multi_kernel is enabled, we do more aggressive persistent reduction. # This may result in some persistent reductions slower than the # corresponding non-persistent reductions. MultiKernel will do benchmarking # to pick the faster one. if config.triton.multi_kernel: threshold *= 16 last_numel = self.numels[-1] return V.graph.sizevars.statically_known_leq(last_numel, threshold) # type: ignore[arg-types] def want_no_x_dim(self): return ( self.reduction_hint == ReductionHint.INNER and self.persistent_reduction and len(self.numels) == 2 and V.graph.sizevars.statically_known_geq(self.numels[-1], 256) # type: ignore[arg-types] ) @property def assert_function(self) -> str: return "tl.device_assert" def indexing( self, index: sympy.Expr, *, copy_shape=None, dense_indexing=False, override_mask=None, block_ptr=False, ): """ Compute the index and mask to pass to tl.load() or tl.store() """ index = self.prepare_indexing(index) index_vars = index.free_symbols has_rindex = False mask_vars: OrderedSet[str] = OrderedSet() for var in index_vars: assert isinstance(var, sympy.Symbol) has_rindex = has_rindex or symbol_is_type(var, SymT.RINDEX) if override_mask: pass elif symbol_is_type(var, SymT.TMP): # indirect indexing cse_var = self.cse.varname_map[var.name] mask_vars.update(cse_var.mask_vars) elif symbol_is_type( var, ( SymT.UNBACKED_INT, SymT.SIZE, SymT.PRECOMPUTED_SIZE, SymT.INDEX, SymT.FLOAT, SymT.UNBACKED_FLOAT, ), ): pass else: # var is one of xN, yN or rN assert symbol_is_type( var, (SymT.RINDEX, SymT.XBLOCK, SymT.YBLOCK) ), var.name mask_vars.add(f"{var.name[0]}mask") need_dense = ( config.triton.dense_indexing or dense_indexing or self._load_mask is not None ) and index != 0 have_dense = True have_loop_vars = False dense_mask_vars: OrderedSet[str] = OrderedSet() for tree in self.active_range_trees(): if index_vars.intersection(tree.var_list): have_loop_vars = True else: have_dense = False dense_mask_vars.add(f"{tree.prefix}mask") if ( block_ptr and self.allow_block_ptr and config.triton.use_block_ptr and not override_mask and not self._load_mask and len(mask_vars - dense_mask_vars) == 0 and not self.is_indirect_indexing(index) and have_loop_vars # workaround https://github.com/openai/triton/issues/2821 and self.index_dtype == "tl.int32" ): def match_strided_block( index: sympy.Expr, range_tree: IterationRangesEntry ) -> Optional[BlockParameters]: """ Matches expressions of the form: idx = s * xindex This implies stride (s,), and shape (XBLOCK,). """ symbol = range_tree.symbol() stride = sympy.Wild("stride", exclude=[symbol]) m = index.match(symbol * stride) if m is None: return None return BlockParameters( shape=[range_tree.numel], block_shape=[self._get_block_size(range_tree)], strides=[m[stride]], offsets=[self._get_block_offset(range_tree)], ) def match_mod_div_block( index: sympy.Expr, range_tree: IterationRangesEntry ) -> Optional[BlockParameters]: """ Matches higher-dimensional blocks coming from FloorDiv and ModularIndexing. Example expression to match: sN * ((rindex//(d1 * ... * d(N-1)))) + s1 * ModularIndexing(rindex, 1, d1) + ... + s(N-1) * ModularIndexing(rindex, d1 * ... * d(N-2), d(N-1)) This iterates over a block of shape (dN, ..., d1) and stride (sN, ..., s1). (d1,...,d(N-1)) and (s1,...,sN) are wildcards that we match. Note that dN does not appear in the expression, but we solve for it using range tree numels and the other dims. """ # Bound the possible number of dims. We use the following heuristics: # - At least one dim for each range tree node. # - At least one dim for every FloorDiv or ModularIndexing op. # - At least 2 dims to pattern match. num_dims = max( 2, len(self.range_tree_nodes), (index.count(FloorDiv) + index.count(ModularIndexing)), ) # Pattern match to find the strides and offset. index_var = range_tree.symbol() wild = functools.partial(sympy.Wild, exclude=[index_var]) dims: List[sympy.Expr] = [ wild(f"dim_mod{idx}") for idx in range(num_dims) ] strides: List[sympy.Expr] = [ wild(f"stride_mod{idx}") for idx in range(num_dims) ] def get_slice_numels(dims: List[Any]) -> List[Any]: """ Compute the cumulative size of each dimension's slice. This proceeds from the last dim up to the second. """ numels = [sympy.Integer(1)] for dim in dims[:0:-1]: numel = dim * numels[0] numels.insert(0, numel) return numels # The first dimension's index is computed by division. # The remaining are computed by modulo. slice_numels = get_slice_numels(dims[:num_dims]) block_index_exprs = [FloorDiv(index_var, slice_numels[0])] + [ ModularIndexing(index_var, numel, dim) for dim, numel in zip(dims[1:], slice_numels[1:]) ] # Calculate a linear index from block indices. match_expr = sympy_dot(strides, block_index_exprs) # Pattern match. match = index.match(match_expr) if match is None: return None # Provide default values for unmatched dims and strides. for dim in dims[1:]: if dim not in match: match[dim] = sympy.Integer(1) for stride in strides[1:]: if stride not in match: match[stride] = sympy.Integer(0) sizevars = V.graph.sizevars def get_match(expr: sympy.Expr) -> sympy.Expr: return sizevars.lookup_precomputed_size(match[expr]) # Replace wildcards with matched expressions. dims = [dims[0]] + [get_match(dim) for dim in dims[1:]] strides = [get_match(stride) for stride in strides] slice_numels = get_slice_numels(dims) block_index_exprs = [ sympy_subs(expr, match) for expr in block_index_exprs ] # The leading dimension is not directly matched in our expression. # We solve for it by dividing the range tree numel by the product of # all other dimensions. We quit if they are not known to be divisible. assert ( dims[0] not in match ), "Expected not to match the leading dimension!" if not sizevars.statically_known_multiple_of( range_tree.numel, slice_numels[0] ): return None dims[0] = range_tree.numel / slice_numels[0] # Check for applicable iteration range sizes. # When mapping a 1D block into an ND one, we need to know that # the number of elements is not changed. This means the slice numels of # the ND iteration range must evenly divide the length of the 1D block. # There are two cases where we can guarantee this: # 1. Numels are powers of 2. If numel == 2 ** n, and we know XBLOCK == 2 ** m, # with n and m integers, then either numel is a multiple of XBLOCK, or numel # is less than XBLOCK. (If numel is less than XBLOCK, we round up to 1 below.) # 2. Numels are multiples of the maximum possible block size. max_block = self._max_block_size(range_tree) if any( not sizevars.statically_known_multiple_of(numel, max_block) and not sizevars.statically_known_power_of_2(numel) for numel in slice_numels ): return None def identity(expr: sympy.Expr) -> sympy.Expr: return expr # Compute the ND block shape from the linear block size. # Use CielDiv to round leading dimensions up to 1. # Non-leading dimensions are clamped to the size of the iteration range, # while the leading dimension can exceed this to accomodate a larger # block size. linear_block_size = self._get_block_size(range_tree) block_shape: List[sympy.Expr] = [ CeilDiv(linear_block_size, slice_numels[0]) ] + [ sympy.Min(CeilDiv(linear_block_size, numel), dim) for numel, dim in zip(slice_numels[1:], dims[1:]) ] # Compute block offsets from {xyzr}offset and the matched expressions. block_offsets: List[sympy.Expr] = [ sympy_subs(expr, {index_var: self._get_block_offset(range_tree)}) for expr in block_index_exprs ] return BlockParameters( shape=dims, block_shape=block_shape, strides=strides, offsets=block_offsets, ) def match_block_pointer_subexpr( expr: sympy.Expr, range_tree: IterationRangesEntry ) -> Optional[BlockParameters]: """ Match a block indexing subexpression involving a single range tree. """ for match_func in ( match_strided_block, match_mod_div_block, ): match = match_func(expr, range_tree) if match is not None: return match return None def match_block_pointer() -> Optional[BlockPtrOptions]: index_relative_to_xyr_index = sympy_subs( index, {v: t.expr for v, t in self.range_tree_nodes.items()} ) range_trees = self.active_range_trees(reorder=True) # Match each range tree separately. range_symbols = {tree.symbol() for tree in range_trees} index_terms = sympy.Add.make_args(index_relative_to_xyr_index) block_params = BlockParameters() for tree in range_trees: # Partition the index into subexpressions pertaining to each range tree. # For example xindex * 5 + rindex * 3 is partitioned to # (xindex * 5, rindex * 3). symbol = tree.symbol() subexpr = sympy.Integer(0) + sum( expr for expr in index_terms if symbol in expr.free_symbols ) # Reject mixed terms, e.g. xindex * rindex. # NB: the zero expression is allowed, for broadcasting. if len(range_symbols.intersection(subexpr.free_symbols)) > 1: return None # Match the subexpression for this range tree. params = match_block_pointer_subexpr(subexpr, tree) if params is None: return None block_params += params # Collect leftover terms as a constant offset. offset = sum( expr for expr in index_terms if not range_symbols.intersection(expr.free_symbols) ) # Form the block pointer. self.filter_masks(mask_vars) return BlockPtrOptions.create( params=block_params, constant_offset=offset, range_trees=range_trees, mask_vars=mask_vars, ) # Return a block pointer, if indexing matches the pattern. options = match_block_pointer() if options is not None: return options expand_str = None index_str = self.index_to_str(index) if isinstance(index, sympy.Integer): expand_str = f"{copy_shape}.shape" if copy_shape else self.dense_size_str() index_str = f"tl.full({expand_str}, {index_str}, tl.int32)" return IndexingOptions( index_str, OrderedSet(), "None", expand_str, has_rindex, index ) if need_dense and not have_dense: expand_str = f"{copy_shape}.shape" if copy_shape else self.dense_size_str() index_str = f"tl.broadcast_to({index_str}, {expand_str})" mask_vars = dense_mask_vars elif not have_loop_vars and copy_shape: index_str = f"tl.broadcast_to({index_str}, {copy_shape}.shape)" mask_vars = dense_mask_vars if override_mask: mask_vars = OrderedSet([override_mask]) if self._load_mask: mask_vars.add(self._load_mask) self.filter_masks(mask_vars) mask_str = " & ".join(sorted(map(str, mask_vars))) if mask_vars else "None" return IndexingOptions(index_str, mask_vars, mask_str, expand_str, has_rindex, index) # type: ignore[arg-type] def codegen_block_ptr( self, name: str, var: str, indexing: BlockPtrOptions, other="" ) -> Tuple[str, Optional[DeferredLine], str]: advance_block_ptr = None check = indexing.boundary_check() if not check: # workaround https://github.com/openai/triton/issues/2813 other = "" elif other: assert other == ", other=0.0" other = f", boundary_check={check!r}, padding_option='zero'" else: other = f", boundary_check={check!r}" if ( self.inside_reduction and self.range_trees[-1].is_loop and indexing.has_rindex() ): block_ptr = f"block_ptr{next(self.block_ptr_id)}" self.body.writeline( DeferredLine( name, f"{block_ptr} = {indexing.format(var, roffset=False)}" ) ) advance_block_ptr = DeferredLine( name, f"{block_ptr} = tl.advance({block_ptr}, {indexing.advance_roffset()})", ) else: block_ptr = indexing.format(var) return block_ptr, advance_block_ptr, other def codegen_block_ptr_store_line(self, name, indexing, block_ptr, value, other=""): # broadcasting is not implicit for block_ptrs value = ( f"tl.broadcast_to({value}, {self.index_to_str(indexing.reshape_suffix)})" ) # drop any extra size=1 dimensions block_shape = [V.kernel.index_to_str(expr) for expr in indexing.block_shape] value = triton_reshape(value, indexing.reshape_suffix, block_shape) # workaround https://github.com/openai/triton/issues/2814 value = f"{value}.to({triton_store_type(V.graph.get_dtype(name))})" return f"tl.store({block_ptr}, {value}{other})" def check_bounds( self, expr: sympy.Expr, size: sympy.Expr, lower: bool, upper: bool, ): if not (lower or upper): return assert isinstance(expr, sympy.Expr) indexing = self.indexing(expr, block_ptr=False) assert isinstance(indexing, IndexingOptions) index_str = indexing.index_str mask_str = indexing.mask_str if indexing.has_mask() else None size_str = texpr(self.rename_indexing(size)) if upper else None # expr is already wrapped line = self.indirect_assert( index_str, "0" if lower else None, size_str, mask_str ) indirect = self.is_indirect_indexing(expr) or any( isinstance(m, TritonCSEVariable) for m in indexing.mask_vars ) buffer = self.get_load_buffer(indexing) self.cse.generate(buffer, line, assignment=False) def get_load_buffer(self, indexing): if indexing.has_indirect() or indexing.has_tmpmask(): # Masked loads must come after the mask is computed return self.compute elif ( self.inside_reduction and self.range_trees[-1].is_loop and not indexing.has_rindex() ): # can lift a common load outside of reduction loop # One exception is when this is an indirect_load. return self.body else: return self.loads def load(self, name: str, index: sympy.Expr): var = self.args.input(name) indirect_indexing = self.is_indirect_indexing(index) original_index = index indexing = self.indexing(index, block_ptr=True) has_rindex = indexing.has_rindex() has_tmpmask = indexing.has_tmpmask() # Keep the variable in cache if were going to reuse it. Equiv., if any of the following hold # 1) We are doing broadcasting # 2) It is a non-coalesced load. The intuition is that if it's # non-coalesced, we will likely load each element multiple times in # practice. # 3) It will be used later and it won't be CSE'd. Equiv., if all the following hold # 3.1) We are in a reduction loop # 3.2) Its not its last use # 3.3) This load will not be lifted to the body # is_coalesced = any( i == 1 for i in self.get_strides_of_load(original_index).values() ) if self.is_broadcasted(original_index): ep = ", eviction_policy='evict_last'" elif not is_coalesced: ep = ", eviction_policy='evict_last'" elif self.inside_reduction and self.range_trees[-1].is_loop: if name in self.args.inplace_buffers: names: OrderedSet[str] = OrderedSet( self.args.inplace_buffers[name].other_names ) else: names = OrderedSet([name]) last_use = len(names & self.last_usage) > 0 evict_last = not last_use and (has_rindex or indirect_indexing) if evict_last: ep = ", eviction_policy='evict_last'" else: ep = ", eviction_policy='evict_first'" else: ep = "" if (has_tmpmask or has_rindex) and indexing.has_mask(): if self._load_other: other = f", other={constant_repr(self._load_other)}" else: other = ", other=0.0" else: other = "" advance_block_ptr = None append_broadcast = None if should_unwrap_unspec_arg(name): line = var else: if isinstance(indexing, BlockPtrOptions): block_ptr, advance_block_ptr, other = self.codegen_block_ptr( name, var, indexing, other ) line = f"tl.load({block_ptr}{other}{ep})" # add needed size=1 dimensions block_shape = [str(dim) for dim in indexing.block_shape] line = triton_reshape(line, block_shape, indexing.reshape_suffix) elif isinstance(original_index, sympy.Integer): line = f"tl.load({var} + ({original_index}))" append_broadcast = indexing.expand_str else: line = f"tl.load({var} + ({indexing.index_str}), {indexing.mask_str}{ep}{other})" dtype = V.graph.get_dtype(name) if ( dtype in (torch.float16, torch.bfloat16) and config.triton.codegen_upcast_to_fp32 ): line += ".to(tl.float32)" if dtype == torch.bool and torch.version.hip is None: # Workaround for https://github.com/openai/triton/issues/2151 # tl.load returns int8 when loading from pointer to int1 # NOTE: Currently causes hangs on bool UTs for ROCm line += ".to(tl.int1)" load_buffer = self.get_load_buffer(indexing) result_var = self.cse.generate(load_buffer, line) assert isinstance(result_var, TritonCSEVariable) result_var.mask_vars = indexing.mask_vars # type: ignore[assignment] if append_broadcast: line = f"tl.broadcast_to({result_var}, {append_broadcast})" result_var = self.cse.generate(load_buffer, line) if advance_block_ptr: load_buffer.writeline(advance_block_ptr) if not self.inside_reduction or (not indexing.has_rmask() and not has_rindex): self.outside_loop_vars.add(result_var) return result_var def store( self, name: str, index: sympy.Expr, value: CSEVariable, mode: StoreMode = None ) -> None: var = self.args.output(name) original_index = index indexing = self.indexing(index, dense_indexing=True, block_ptr=mode is None) # Guard against write-after-read corruption in triton. # See # https://github.com/openai/triton/issues/1615 # This triton bug means that a load which is broadcasted over multiple # warps may see the result of a store that happens later in the triton # program. The workaround is to add a barrier before storing, which # enforces that all warps have already read the data. is_inplace = name in self.args.inplace_buffers is_broadcasted = self.is_broadcasted(original_index) if is_inplace and is_broadcasted: self.stores.writeline(DeferredLine(name, "tl.debug_barrier()")) advance_block_ptr = None if isinstance(indexing, BlockPtrOptions): block_ptr, advance_block_ptr, other = self.codegen_block_ptr( name, var, indexing ) # block_ptr stores don't do implicit casting line = self.codegen_block_ptr_store_line( name, indexing, block_ptr, value, other ) elif mode is None: line = f"tl.store({var} + ({indexing.index_str}), {value}, {indexing.mask_str})" elif mode == "atomic_add": line = f"tl.atomic_add({var} + ({indexing.index_str}), {value}, {indexing.mask_str}, sem='relaxed')" else: raise NotImplementedError(f"store mode={mode}") self.stores.writeline(DeferredLine(name, line)) if advance_block_ptr: self.stores.writeline(advance_block_ptr) if not self.inside_reduction: self.outside_loop_vars.add(value) def bucketize( self, values: CSEVariable, offsets_name: str, offsets_size: sympy.Expr, indexing_dtype: torch.dtype, right: bool, ) -> CSEVariable: """ See [Note: Inductor bucketize op] """ # Triton performance for bucketize_binary_search is much better when the number # of threads equals the number of elements. # If we're trying to use a bucketize kernel, we should make sure that an # autotuning config with num_elements_per_warp=32 exists. self.autotune_hints.add(AutotuneHint.ELEMENTS_PER_WARP_32) offsets_ptr = self.args.input(offsets_name) block_size = self.dense_size_str() offsets_size_str = self.index_to_str(offsets_size) if indexing_dtype == torch.int32: triton_dtype = "tl.int32" elif indexing_dtype == torch.int64: triton_dtype = "tl.int64" else: raise NotImplementedError( "Bucketize only supports indexing with int32 and int64" ) result = self.cse.generate( self.compute, f"triton_helpers.bucketize_binary_search({values}, {offsets_ptr}, {triton_dtype}, {right}, {offsets_size_str}, {block_size})", # noqa: B950 line too long ) return result def reduction_resize(self, value): ndims = self.triton_tensor_ndim() if ndims == 1: return f"triton_helpers.promote_to_tensor({value})" sizes = [":"] * ndims sizes[-1] = "None" return f"{value}[{', '.join(sizes)}]" def reduction( self, dtype: torch.dtype, src_dtype: torch.dtype, reduction_type: ReductionType, value: Union[CSEVariable, Tuple[CSEVariable, ...]], ) -> Union[CSEVariable, Tuple[CSEVariable, ...]]: assert self.inside_reduction masks = OrderedSet(f"{tree.prefix}mask" for tree in self.range_trees) self.filter_masks(masks) masks = sorted(masks) if self._load_mask: masks.append(self._load_mask) reduction_range_prefix = self.range_trees[-1].prefix # Say we have # tmp0 = ops.constant(1, torch.int64) # tmp1 = ops.reduction(torch.int64, torch.int64, "sum", tmp0) # tmp0 in the triton code is either a scalar, or single-element tensor # so if we emit tl.sum directly, it will only give 1 instead of RBLOCK * 1 # To avoid this, we broadcast to the expected shape first. dense_size_str = self.dense_size_str() value = self._map_tuple_or_scalar( lambda v: self.cse.generate( self.compute, f"tl.broadcast_to({v}, {dense_size_str})" ), value, ) dim: int root_op: str def final_reduction(value): use_helper = reduction_type in {"any", "max", "min", "prod"} module = "triton_helpers" if use_helper else "tl" if reduction_type in {"max", "min"}: return self.reduction_resize( f"{module}.{reduction_type}2({value}, {dim})" ) return self.reduction_resize(f"{module}.{reduction_type}({value}, {dim})") def final_argreduce(buffer, result_var, value, index): buffer.splice( f"""\ _, {result_var}_tmp = triton_helpers.{root_op}_with_index({value}, {index}, {dim}) {result_var} = {self.reduction_resize(f'{result_var}_tmp')} """ ) cache_key = (src_dtype, reduction_type, value) if cache_key in self.cse.reduction_cache: return self.cse.reduction_cache[cache_key] dim = self.triton_tensor_ndim() - 1 acc_type = triton_acc_type(src_dtype) result_var: Any = self.cse.newvar() result_var.mask_vars = OrderedSet(var for var in masks if var[0] != "r") cond = " & ".join(masks) def where_cond(tval, fval): if not cond: return tval return TritonKernelOverrides.where(cond, tval, fval) if self.persistent_reduction: default = ir.Reduction.default_value(reduction_type, src_dtype) default = self._map_tuple_or_scalar(constant_repr, default) def _mask_value(value, default): return self.cse.generate(self.compute, where_cond(value, default)) if isinstance(value, tuple): masked_value = [_mask_value(v, d) for v, d in zip(value, default)] else: masked_value = _mask_value(value, default) if reduction_type in {"argmax", "argmin"}: accumulator_index = str( self.cse.generate( self.compute, f"tl.broadcast_to({reduction_range_prefix}index, {masked_value}.shape)", ) ) root_op = {"argmax": "max", "argmin": "min"}[reduction_type] final_argreduce( self.compute, result_var, masked_value, accumulator_index ) elif reduction_type == "welford_reduce": # For persistent reductions, don't bother with # welford's algorithm since it uses more registers, and # taking two reductions doesn't increase memory usage. result_var = self.welford_reduce_fallback(dtype, value) elif reduction_type == "welford_combine": mean, m2, weight = masked_value welford = f"triton_helpers.welford({mean}, {m2}, {weight}, {dim})" mean, m2, weight = (self.cse.newvar() for _ in range(3)) self.compute.writeline(f"{mean}, {m2}, {weight} = {welford}") result_var = tuple( self.cse.generate(self.compute, self.reduction_resize(var_name)) for var_name in (mean, m2, weight) ) else: result_var = self.cse.generate( self.compute, final_reduction(masked_value) ) else: accumulator = f"_{result_var}" default = ir.Reduction.default_accumulator(reduction_type, src_dtype) default = self._map_tuple_or_scalar(constant_repr, default) if not isinstance(default, tuple): self.body.writeline( f"{accumulator} = tl.full({self.dense_size_str()}, {default}, {acc_type})" ) if reduction_type in {"argmax", "argmin"}: accumulator_index = f"_{result_var}_index" long_max = torch.iinfo(torch.int64).max self.body.writeline( f"{accumulator_index} = tl.full({self.dense_size_str()}, {long_max}, tl.int64)" ) root_op = {"argmax": "max", "argmin": "min"}[reduction_type] self.compute.splice( f"""\ {accumulator}_next, {accumulator_index}_next = triton_helpers.{root_op}imum_with_index( {accumulator}, {accumulator_index}, {value}, {reduction_range_prefix}index ) {accumulator} = {where_cond(f'{accumulator}_next', accumulator)} {accumulator_index} = {where_cond(f'{accumulator_index}_next', accumulator_index)} """ ) final_argreduce(self.suffix, result_var, accumulator, accumulator_index) elif is_welford_reduction(reduction_type): accumulator = f"{result_var}_mean" accumulator_m2 = f"{result_var}_m2" accumulator_weight = f"{result_var}_weight" self.body.writeline( f"{accumulator} = tl.zeros({self.dense_size_str()}, {acc_type})" ) self.body.writeline( f"{accumulator_m2} = tl.zeros({self.dense_size_str()}, {acc_type})" ) self.body.writeline( f"{accumulator_weight} = tl.zeros({self.dense_size_str()}, {acc_type})" ) if reduction_type == "welford_combine": mean, m2, weight = value self.compute.splice( f"""\ {accumulator}_next, {accumulator_m2}_next, {accumulator_weight}_next = triton_helpers.welford_combine( {accumulator}, {accumulator_m2}, {accumulator_weight}, {mean}, {m2}, {weight} ) """ ) else: assert reduction_type == "welford_reduce" self.compute.splice( f"""\ {accumulator}_next, {accumulator_m2}_next, {accumulator_weight}_next = triton_helpers.welford_reduce( {value}, {accumulator}, {accumulator_m2}, {accumulator_weight}, roffset == 0 ) """ ) self.compute.splice( f"""\ {accumulator} = {where_cond(f'{accumulator}_next', accumulator)} {accumulator_m2} = {where_cond(f'{accumulator_m2}_next', accumulator_m2)} {accumulator_weight} = {where_cond(f'{accumulator_weight}_next', accumulator_weight)} """ ) result_mean = result_var result_m2 = self.cse.newvar() result_weight = self.cse.newvar() self.suffix.splice( f"""\ {result_mean}_tmp, {result_m2}_tmp, {result_weight}_tmp = triton_helpers.welford( {accumulator}, {accumulator_m2}, {accumulator_weight}, {dim} ) {result_mean} = {self.reduction_resize(f'{result_mean}_tmp')} {result_m2} = {self.reduction_resize(f'{result_m2}_tmp')} {result_weight} = {self.reduction_resize(f'{result_weight}_tmp')} """ ) result_var = result_mean, result_m2, result_weight else: combine_fn = ir.get_reduction_combine_fn(reduction_type, src_dtype) updated = combine_fn(accumulator, value) self.compute.writeline( f"{accumulator} = {where_cond(updated, accumulator)}" ) if src_dtype == torch.bool: # This is only really used for aten.any. It changes the # final reduction of a non-persistent reduction from # tmp5 = triton_helpers.max(_tmp5, 1)[:, None] # to # tmp5 = triton_helpers.max(_tmp5.to(tl.int8), 1)[:, None].to(tl.int1) # which is needed because tl.reduce doesn't support tl.int1 accumulator = f"{accumulator}.to(tl.int8)" result_type = triton_compute_type(dtype) self.suffix.writeline( f"{result_var} = {final_reduction(accumulator)}.to({result_type})" ) else: self.suffix.writeline( f"{result_var} = {final_reduction(accumulator)}" ) self.cse.reduction_cache[cache_key] = result_var if isinstance(result_var, tuple): assert all(isinstance(x, TritonCSEVariable) for x in result_var) self.outside_loop_vars |= OrderedSet(result_var) else: assert isinstance(result_var, TritonCSEVariable) self.outside_loop_vars.add(result_var) return result_var def store_reduction(self, name: str, index: sympy.Expr, value: CSEVariable): assert self.inside_reduction self.inside_reduction = False indexing = self.indexing(index, block_ptr=True) self.inside_reduction = True var = self.args.output(name) if isinstance(indexing, BlockPtrOptions): self.suffix.writeline( DeferredLine( name, self.codegen_block_ptr_store_line( name, indexing, indexing.format(var), value, f", boundary_check={indexing.boundary_check()!r}", ), ) ) else: assert isinstance(indexing, IndexingOptions) self.suffix.writeline( DeferredLine( name, f"tl.store({var} + ({indexing.index_str}), {value}, {indexing.mask_str})", ) ) def _lift_helper(self, fn, num_args) -> str: # Lift IR function for scan operations into a triton function # in the global namespace helper = IndentedBuffer() helper.writeline("@triton.jit") args = [tuple(f"arg{i}_{n}" for n in range(num_args)) for i in range(2)] signature = ", ".join(itertools.chain.from_iterable(args)) helper.writeline(f"def {{name}}({signature}):") cse = CSE(prefix="", suffix="") overrides = TritonOverrides(V.MockHandler()) # Build a name that changes depending on fn to workaround a triton bug # where the combine_fn to reduce and scan is not hashed, and so different # scan ops may collide in the triton cache. # This is fixed with the latest triton pin, but not the triton-rocm pin. helper_name = "_triton_helper_fn" class CSEProxy: def __getattr__(self, name: str) -> Callable[..., CSEVariable]: def inner(*args, **kwargs): nonlocal helper_name helper_name += f"_{name}" return cse.generate( helper, getattr(overrides, name)(*args, **kwargs), ) return inner with helper.indent(), V.set_ops_handler(CSEProxy()): outputs = fn(*args) outputs = ", ".join(str(output) for output in outputs) helper.writeline(f"return {outputs}") return self.helper_functions.add(helper.getvalue(), base_name=helper_name) def scan( self, dtypes: Tuple[torch.dtype, ...], combine_fn: Callable[ [Tuple[CSEVariable, ...], Tuple[CSEVariable, ...]], Tuple[CSEVariable, ...] ], values: Tuple[CSEVariable, ...], ) -> Tuple[CSEVariable, ...]: assert self.inside_reduction masks = OrderedSet(f"{tree.prefix}mask" for tree in self.range_trees) self.filter_masks(masks) masks = sorted(masks) assert not self._load_mask, "ops.scan not supported inside ops.masked" reduction_range_prefix = self.range_trees[-1].prefix broadcasted_values = [] accumulators = [] cse_compute = functools.partial(self.cse.generate, self.compute) combine_helper_fn = self._lift_helper(combine_fn, len(values)) dim = self.triton_tensor_ndim() - 1 for value, dtype in zip(values, dtypes): acc_type = triton_acc_type(dtype) cond = " & ".join(masks) value_dtype = self.cse.generate( self.compute, f"{value}.to({triton_compute_type(dtype)})", ) value = self.cse.generate( self.compute, f"tl.broadcast_to({value_dtype}, {self.dense_size_str()})", ) broadcasted_values.append(value) acc_type = triton_acc_type(dtype) cond = " & ".join(masks) if not self.persistent_reduction: accumulator = self.cse.newvar() reduced_size = self.dense_size_list() reduced_size[-1] = "1" reduced_size = f"[{', '.join(reduced_size)}]" default = "float('nan')" if dtype.is_floating_point else "-1" self.body.writeline( f"{accumulator} = tl.full({reduced_size}, {default}, {acc_type})" ) accumulators.append(accumulator) def csv(values): return " ".join(f"{value}," for value in values) def cse_multiple(line, n, masks): cache_keys = [f"{line}, {i}, {masks}" for i in range(n)] if all(cache_key in self.cse.cache for cache_key in cache_keys): return [self.cse.cache[cache_key] for cache_key in cache_keys] result_vars = [self.cse.newvar() for _ in range(n)] self.compute.writeline( f"{csv(result_vars)} = {line}", ) for result_var, cache_key in zip(result_vars, cache_keys): if masks: result_var.mask_vars = masks # type: ignore[attr-defined] self.cse.cache[cache_key] = result_var return tuple(result_vars) partial_scan_vars = cse_multiple( f"tl.associative_scan(({csv(broadcasted_values)}), {dim}, {combine_helper_fn})", len(values), masks, ) if not self.persistent_reduction: # tl.reduce doesn't work for non-commutative operators, so instead # of repeating the scan op as a reduction, we use sum to select the # last scan value partial_reduce_vars = [ cse_compute( f"triton_helpers.select_one(({partial_scan_var}), rbase == (RBLOCK - 1), dim=-1, keep_dims=True)" ) for partial_scan_var in partial_scan_vars ] accs_next = combine_fn(tuple(accumulators), tuple(partial_reduce_vars)) full_scan_vars = combine_fn(tuple(accumulators), partial_scan_vars) result_vars = [ cse_compute(f"tl.where(roffset > 0, {full_scan}, {partial_scan})") for full_scan, partial_scan in zip(full_scan_vars, partial_scan_vars) ] for acc_next, accumulator, partial_reduce in zip( accs_next, accumulators, partial_reduce_vars ): self.compute.writeline( f"{accumulator} = tl.where(roffset > 0, {acc_next}, {partial_reduce})" ) else: result_vars = partial_scan_vars for result_var in result_vars: result_var.mask_vars = masks # type: ignore[attr-defined] return tuple(result_vars) def sort( self, dtypes: Tuple[torch.dtype, ...], values: Tuple[CSEVariable, ...], stable: bool, descending: bool, ) -> Tuple[CSEVariable, ...]: assert self.inside_reduction masks = OrderedSet(f"{tree.prefix}mask" for tree in self.range_trees) self.filter_masks(masks) masks = sorted(masks) assert not self._load_mask, "ops.sort not supported inside ops.masked" assert ( self.persistent_reduction ), "ops.sort is only supported in persistent reductions" reduction_range_prefix = self.range_trees[-1].prefix cse_compute = functools.partial(self.cse.generate, self.compute) dim = self.triton_tensor_ndim() - 1 broadcasted_values = [ cse_compute(f"tl.broadcast_to({value}, {self.dense_size_str()})") for value in values ] def csv(values): return " ".join(f"{value}," for value in values) def cse_multiple(line, n, masks): cache_keys = [f"{line}, {i}, {masks}" for i in range(n)] if all(cache_key in self.cse.cache for cache_key in cache_keys): return [self.cse.cache[cache_key] for cache_key in cache_keys] result_vars = [self.cse.newvar() for _ in range(n)] self.compute.writeline( f"{csv(result_vars)} = {line}", ) for result_var, cache_key in zip(result_vars, cache_keys): if masks: result_var.mask_vars = masks # type: ignore[attr-defined] self.cse.cache[cache_key] = result_var return tuple(result_vars) assert self.range_trees[-1].prefix == "r" rnumel = "None" if self._has_constant_mask(self.range_trees[-1]) else "rnumel" if len(values) == 2: line = ( f"triton_helpers.sort_with_index({broadcasted_values[0]}, {broadcasted_values[1]}," f" {rnumel}, {dim}, stable={stable}, descending={descending})" ) result_vars = cse_multiple(line, len(values), masks) else: raise AssertionError("Unhandled sort") for result_var, input_var in zip(result_vars, values): result_var.mask_vars = masks # type: ignore[attr-defined] result_var.bounds = input_var.bounds return tuple(result_vars) def codegen_body(self): """ Concat output code from index_code, loads, compute, stores, suffix into self.body. For pointwise kernels, this is called just once at the end. For reduction kernels, this generates a loop over the reduction axis. """ if not ( self.indexing_code or self.loads or self.stores or self.compute or self.suffix ): return if self.inside_reduction and self.range_trees[-1].is_loop: self.body.writeline("for roffset in range(0, rnumel, RBLOCK):") with self.body.indent(): # last range tree is always reduction self.iteration_ranges_codegen_header(self.range_trees[-1], self.body) self.body.splice(self.indexing_code) self.body.splice(self.loads) self.body.splice(self.compute) self.body.splice(self.stores) # invalidate any caches that came from inside the reduction loop self.cse.invalidate(self.outside_loop_vars) self.range_trees[-1].cache_clear() else: self.body.splice(self.indexing_code) self.body.splice(self.loads) self.body.splice(self.compute) self.body.splice(self.stores) self.body.splice(self.suffix) self.indexing_code.clear() self.loads.clear() self.compute.clear() self.stores.clear() self.suffix.clear() def codegen_kernel_benchmark(self, num_gb, grid=None): result = IndentedBuffer() argdefs, call_args, signature, _ = self.args.python_argdefs() result.writelines(["", "", "def get_args():"]) with result.indent(): name_cnt = itertools.count() var_names = [] for arg_name, arg_sig in zip(call_args, signature): var_name = f"arg_{next(name_cnt)}" buf = V.graph.try_get_buffer(arg_name) if buf: result.writeline( f"{var_name} = rand_strided({V.graph.sizevars.size_hints(buf.get_size())}, {V.graph.sizevars.size_hints(buf.get_stride())}, device='{buf.get_device()}', dtype={buf.get_dtype()})" # noqa: B950 line too long ) elif arg_name in V.graph.constants: # note that random seed is put in V.graph.constants const_tensor = V.graph.constants[arg_name] result.writeline( f"{var_name} = rand_strided({V.graph.sizevars.size_hints(const_tensor.size())}, {V.graph.sizevars.size_hints(const_tensor.stride())}, device='{const_tensor.device}', dtype={const_tensor.dtype})" # type: ignore[arg-type] # noqa: B950 line too long ) elif isinstance(arg_sig, SizeArg): symval_hint = V.graph.sizevars.size_hint(arg_sig.expr) # Force the seed_offset to be 0 so calls to the same kernel # using different seed offset will have the same benchmark harness. # We can dedup kernel definitions in this case. if "seed_offset" in arg_sig.name: symval_hint = 0 result.writeline(f"{var_name} = {symval_hint}") elif isinstance(arg_sig, WorkspaceArg): device = V.graph.scheduler.get_current_device_or_throw() nbytes = V.graph.sizevars.size_hint(arg_sig.nbytes) result.writeline( f"{var_name} = torch.zeros({nbytes}, device='{device}', dtype=torch.uint8)" ) else: raise KeyError( f"Don't find the buffer or const tensor for {arg_name}" ) var_names.append(var_name) result.writeline(f"return {', '.join(var_names)},") result.writelines(["\n", "\n", "def call(args):"]) if grid is None: grid = [] extra_args = [] extra_args_str = None for tree in self.active_range_trees(): expr = pexpr(V.graph.sizevars.size_hint(tree.numel)) extra_args.append(expr) if tree.prefix != "r": grid.append(expr) if self.need_numel_args(): extra_args_str = ", ".join(map(str, extra_args)) + ", " else: extra_args_str = "" grid_arg = f"{extra_args_str}grid=grid({', '.join(grid)})" else: grid_arg = f"grid={grid}" current_device = V.graph.scheduler.get_current_device_or_throw() index = current_device.index with result.indent(): result.writeline(f"with {V.graph.device_ops.device_guard(index)}:") with result.indent(): result.writeline( V.graph.device_ops.set_device(index) ) # no-op to ensure context stream_name = f"stream{index}" result.writeline(f"{stream_name} = get_raw_stream({index})") result.writeline( f"{str(Placeholder.KERNEL_NAME)}.run(*args, {grid_arg}, stream={stream_name})" ) # benchmark all configs result.writelines(["\n", "\n", "def benchmark_all_configs(args):"]) with result.indent(): result.writeline(f"with {V.graph.device_ops.device_guard(index)}:") with result.indent(): result.writeline( V.graph.device_ops.set_device(index) ) # no-op to ensure context result.writeline( f"return {str(Placeholder.KERNEL_NAME)}.benchmark_all_configs(*args, {grid_arg})" ) result.writelines(["\n", "\n", "if __name__ == '__main__':"]) with result.indent(): result.writeline( "from torch._inductor.runtime.benchmarking import benchmarker" ) result.writeline("") result.writeline("args = get_args()") result.writeline( "ms = benchmarker.benchmark_gpu(lambda: call(args), rep=40, fast_flush=True)" ) result.writeline(f"num_gb = {num_gb}") result.writeline("gb_per_s = num_gb / (ms / 1e3)") result.writeline( 'print(f"{ms:.3f}ms {num_gb:.3f}GB {gb_per_s:.2f}GB/s")' ) return result def imports_for_benchmark_kernel(self): return textwrap.dedent( """ from torch._dynamo.testing import rand_strided {} import torch from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid """.format( V.graph.device_ops.import_get_raw_stream_as("get_raw_stream") ) ) def _get_heuristic(self): if self.persistent_reduction: assert self.inside_reduction return "persistent_reduction" elif self.inside_reduction: return "reduction" return "pointwise" @staticmethod def inductor_meta_common(): inductor_meta = { "backend_hash": torch.utils._triton.triton_hash_with_backend(), "are_deterministic_algorithms_enabled": torch.are_deterministic_algorithms_enabled(), "assert_indirect_indexing": config.assert_indirect_indexing, "autotune_local_cache": config.autotune_local_cache, "autotune_pointwise": config.triton.autotune_pointwise, "autotune_remote_cache": config.autotune_remote_cache, "force_disable_caches": config.force_disable_caches, "dynamic_scale_rblock": config.dynamic_scale_rblock, "max_autotune": config.max_autotune, "max_autotune_pointwise": config.max_autotune_pointwise, "min_split_scan_rblock": config.triton.min_split_scan_rblock, "spill_threshold": config.triton.spill_threshold, "store_cubin": config.triton.store_cubin, } if torch.version.hip is not None: inductor_meta["is_hip"] = True if config.is_fbcode(): inductor_meta["is_fbcode"] = True if config.profile_bandwidth: inductor_meta["profile_bandwidth"] = config.profile_bandwidth inductor_meta["profile_bandwidth_regex"] = config.profile_bandwidth_regex inductor_meta["profile_bandwidth_output"] = config.profile_bandwidth_output inductor_meta[ "profile_bandwidth_with_do_bench_using_profiling" ] = config.profile_bandwidth_with_do_bench_using_profiling if config.coordinate_descent_tuning: inductor_meta[ "coordinate_descent_tuning" ] = config.coordinate_descent_tuning inductor_meta[ "coordinate_descent_search_radius" ] = config.coordinate_descent_search_radius inductor_meta[ "coordinate_descent_check_all_directions" ] = config.coordinate_descent_check_all_directions return inductor_meta def codegen_kernel(self, name=None): code = IndentedBuffer() size_hints = [] for numel in self.numels: numel_hint = V.graph.sizevars.symbolic_hint(numel) if not isinstance(numel_hint, (int, sympy.Integer)): # This default heuristic hint was picked carefully: it is # large, to ensure that we don't shrink the block size (since # if you don't have many elements, it'd be wasteful to pick a # large block size). Since we don't know how many elements we # might have, we should be OK with some inefficiency to make # sure we handle the large case well. 8192 is the largest # block size we support, so we pick that. # # If we have a better hint for unbacked SymInts (e.g., because # a user told us, or we are tracking upper bounds) we could # use that here. size_hint = 8192 else: size_hint = next_power_of_2(int(numel_hint)) size_hints.append(size_hint) if not self.inside_reduction: size_hints.pop() heuristics = self._get_heuristic() if name is None: code.splice(gen_common_triton_imports()) if config.benchmark_kernel: code.splice(self.imports_for_benchmark_kernel()) argdefs, _, signature, _ = self.args.python_argdefs() # maps actual expression to SizeArg if it is in sizevars replacements for i, arg in enumerate(signature): if isinstance(arg, SizeArg): # mypy is unhappy about the sympy.Expr # type for the key of the dict below symbol = cast(sympy.Symbol, arg.expr) if symbol in V.graph.sizevars.inv_precomputed_replacements: signature[i] = SizeArg( arg.name, V.graph.sizevars.inv_precomputed_replacements[symbol] ) mutated_args: OrderedSet[str] = OrderedSet() for mutation in self.mutations: if mutation in self.args.input_buffers: mutated_args.add(self.args.input_buffers[mutation]) if ( mutation in self.args.inplace_buffers and mutation not in V.graph.removed_buffers and mutation not in self.removed_buffers ): mutated_args.add(self.args.inplace_buffers[mutation].inner_name) if mutation in self.args.output_buffers: mutated_args.add(self.args.output_buffers[mutation]) # workspace arguments are mutated, but are not marked as mutations in self.mutations # because their buffers are added during codegen, and aren't tracked during # lowering/scheduling. So we add them as mutated_args explicitly below. # # In the logic below, we only mark the workspaces a mutated if they are marked with # zero_fill: that's because, if we don't expect the buffer to be pre-filled with # zeros, then, although we still mutate the data, we don't care about those # mutations because we don't make any assumptions about the contents of the # workspace buffer. for argname, arg in zip(argdefs, signature): if isinstance(arg, WorkspaceArg) and arg.zero_fill: mutated_args.add(argname) mutated_args = sorted(mutated_args) triton_meta_signature = signature_to_meta( signature, size_dtype=self.index_dtype ) triton_meta = { "signature": triton_meta_signature, "device": DeviceProperties.create( V.graph.scheduler.get_current_device_or_throw() ), "constants": {}, } inductor_meta = { "autotune_hints": set(self.autotune_hints), "kernel_name": str(Placeholder.DESCRIPTIVE_NAME), "mutated_arg_names": mutated_args, "no_x_dim": self.no_x_dim, "num_load": self.num_load, "num_reduction": self.num_reduction, **self.inductor_meta_common(), } num_gb = None if config.benchmark_kernel or config.profile_bandwidth: num_gb = self.estimate_kernel_num_bytes() / 1e9 inductor_meta["kernel_num_gb"] = num_gb for tree in self.active_range_trees(): sizearg = SizeArg(f"{tree.prefix}numel", tree.numel) signature.append(sizearg) triton_meta_signature[len(argdefs)] = signature_of( sizearg, size_dtype=self.index_dtype ) argdefs.append(f"{tree.prefix}numel") # constexpr version causes issues, see # https://github.com/pytorch/torchdynamo/pull/1362 # triton_meta["constants"][len(argdefs)] = V.graph.sizevars.size_hint( # tree.numel # ) # argdefs.append(f"{tree.prefix}numel: tl.constexpr") triton_meta["configs"] = [config_of(signature)] # Triton compiler includes equal_to_1 args into constants even # when they are not constexpr. otherwise there may be a segfault # during launching the Inductor-compiled Triton kernel. # https://github.com/pytorch/pytorch/issues/120478#issuecomment-1962822307 # https://github.com/openai/triton/blob/231efe9ed2d200be0f69a07c298e4342b08efe3d/python/triton/runtime/jit.py#L384 for arg_num in triton_meta["configs"][0].equal_to_1: # type: ignore[index] triton_meta["constants"][arg_num] = 1 # type: ignore[index] self.triton_meta = triton_meta for tree in self.range_trees: if tree.prefix == "r" and self.persistent_reduction: # RBLOCK for persistent_reduction is defined in codegen_static_numels continue if tree.tensor_dim is None: continue argdefs.append(f"{tree.prefix.upper()}BLOCK : tl.constexpr") self.codegen_body() for helper in self.helper_functions: code.writeline("") code.splice(helper) if self.inside_reduction: reduction_hint = self.reduction_hint heuristics_line = f""" @triton_heuristics.{heuristics}( size_hints={size_hints!r}, reduction_hint={reduction_hint}, filename=__file__, triton_meta={triton_meta!r}, inductor_meta={inductor_meta!r} ) @triton.jit """ else: tile_hint = "" if len(size_hints) == 2: if len(signature) == 4: # input, output and 2 args tile_hint = "tile_hint=TileHint.SQUARE," else: tile_hint = "tile_hint=TileHint.DEFAULT," heuristics_line = f""" @triton_heuristics.{heuristics}( size_hints={size_hints!r}, {tile_hint} filename=__file__, triton_meta={triton_meta!r}, inductor_meta={inductor_meta!r}, min_elem_per_thread={self.min_elem_per_thread} ) @triton.jit """ code.splice(heuristics_line) code.writeline( f"def {name or str(Placeholder.KERNEL_NAME)}({', '.join(argdefs)}):" ) with code.indent(): self.codegen_static_numels(code) for old, new in self.args.aliases(): code.writeline(f"{old} = {new}") code.splice(self.body) if config.benchmark_kernel: code.splice(self.codegen_kernel_benchmark(num_gb)) return code.getvalue() def _get_persistent_RBLOCK(self, rnumel): rnumel = V.graph.sizevars.simplify(rnumel) if isinstance(rnumel, (sympy.Integer, int)): val = int(rnumel) val = next_power_of_2(val) else: val = 128 while not V.graph.sizevars.statically_known_leq(rnumel, val): assert val <= 16 * 1024, f"Failed to find static RBLOCK for {rnumel}" val *= 2 return val def codegen_static_numels(self, code): """ We get a small speedup from hard coding numels if they are static. This code stomps on the passed-in values by writing an constant to the top of the kernel. In a kernel like: def KERNEL_NAME(in_ptr0, in_ptr1, out_ptr2, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr): We would add xnumel = 4096 rnumel = 768 After the signature, before the kernel code, if we decided to make these static. As its hardcoded, it becomes a better signal to triton on how to unroll and do some static indexing. So, it's not so much that downstream knows that its a static numel, as that you just plop a constant into the kernel. """ for tree in self.range_trees: if tree.prefix != "r" or self.inside_reduction: simplified_tree_numel = V.graph.sizevars.simplify(tree.numel) if isinstance(simplified_tree_numel, (sympy.Integer, int)): code.writeline(f"{tree.prefix}numel = {int(simplified_tree_numel)}") if tree.prefix == "r" and self.persistent_reduction: val = self._get_persistent_RBLOCK(tree.numel) code.writeline(f"RBLOCK: tl.constexpr = {val}") if tree.prefix == "x" and self.no_x_dim: code.writeline("XBLOCK: tl.constexpr = 1") def _get_grid_fn(self): return "grid" def add_numel_to_call_args_and_grid(self, name, call_args, arg_types, grid): # TODO(jansel): if there are constants, we shouldn't bother passing them as args for tree in self.range_trees: if isinstance(tree.numel, (sympy.Integer, sympy.Symbol)): expr = tree.numel else: expr = V.graph.wrapper_code.generate_numel_expr(name, tree) if tree.prefix != "r" or self.inside_reduction: call_args.append(expr) arg_types.append(type(expr)) if tree.grid_dim is not None: grid.append(expr) def call_kernel(self, name: str, node: Optional[IRNode] = None): wrapper = V.graph.wrapper_code wrapper.write_triton_header_once() _, call_args, _, arg_types = self.args.python_argdefs() grid: List[Any] = [] self.add_numel_to_call_args_and_grid(name, call_args, arg_types, grid) current_device = V.graph.scheduler.get_current_device_or_throw() if self.args.workspace_arg is not None: ws = self.args.workspace_arg wrapper.generate_workspace_allocation( ws.nbytes, current_device, ws.zero_fill ) grid = wrapper.generate_default_grid(name, grid) wrapper.generate_kernel_call( name, call_args, grid, current_device.index, cuda=True, triton=True, arg_types=arg_types, grid_fn=self._get_grid_fn(), triton_meta=self.triton_meta, ) if self.args.workspace_arg is not None: wrapper.writeline(wrapper.make_free_by_names(["workspace"])) def codegen_nan_check(self): wrapper = V.graph.wrapper_code _, call_args, arg_signatures, _ = self.args.python_argdefs() for arg, arg_signature in zip(call_args, arg_signatures): if isinstance(arg_signature, TensorArg): if V.graph.cpp_wrapper: if config.abi_compatible: wrapper.writeline( f'AOTI_TORCH_ERROR_CODE_CHECK(aoti_torch_check_inf_and_nan("{arg}", {arg}));' ) else: wrapper.writeline(f'assert_inf_and_nan("{arg}", {arg});') else: line = f"assert not {arg}.isnan().any().item()" wrapper.writeline(line) line = f"assert not {arg}.isinf().any().item()" wrapper.writeline(line) def create_cse_var(self, *args, **kwargs): return TritonCSEVariable(*args, **kwargs) def codegen_iteration_ranges_entry(self, entry: IterationRangesEntry): line = f"{entry.name} = {self.kexpr(self.rename_indexing(entry.expr))}" if entry.root.is_loop: self.indexing_code.writeline(line) else: # lift non-reduction stores outside loop self.body.writeline(line) def iteration_ranges_ranges_code(self, entry): assert entry.tensor_dim is not None size = self.indexing_size_str(entry.tensor_dim) index_dtype = self.index_dtype convert = f".to({index_dtype})" if index_dtype != "tl.int32" else "" return f"tl.arange(0, {entry.prefix.upper()}BLOCK){size}{convert}" def iteration_ranges_scalar_code(self, entry, value): index_dtype = self.index_dtype ndim = self.triton_tensor_ndim() size = [1] * ndim return f"tl.full({size}, {value}, {index_dtype})" def iteration_ranges_get_pid(self, entry): assert entry.grid_dim is not None key = f"tl.program_id({entry.grid_dim})" # y_grid has a limit, so express it in terms of y and z in case of overflow. # z grid is only exercised when max_tiles == 3 (off by default). if ( entry.grid_dim == 1 and not entry.has_zdim and not V.graph.sizevars.statically_known_leq(entry.numel, get_max_y_grid()) ): # For ynumel larger than max_ygrid, we need to use zdim. # For each z dimension, there are tl.num_programs(1) yblocks which is passed by grad(x,y,z). # So, we need to add tl.program_id(z) * tl.num_programs(y) *YBLOCK to get the correct yoffset. key = f"({key} + tl.program_id({entry.grid_dim + 1}) * tl.num_programs({entry.grid_dim}))" pid = entry.pid_cache.get(key, key) if self.index_dtype != "tl.int32": return f"{pid}.to({self.index_dtype})" return pid def _has_constant_mask(self, tree: IterationRangesRoot): if not self.optimize_mask: return False if V.graph.sizevars.statically_known_equals(tree.numel, 1): # type: ignore[arg-type] return True # Masks are superfluous if numel is a multiple of BLOCK # (We use the fact that BLOCK is required by triton to be a power of 2) if tree.prefix == "r" and self.persistent_reduction: max_block = self._get_persistent_RBLOCK(tree.numel) elif tree.prefix == "x" and self.no_x_dim: max_block = 1 else: if tree.prefix.upper() not in TRITON_MAX_BLOCK: return False max_block = TRITON_MAX_BLOCK[tree.prefix.upper()] # Optional optimization: if block divides numel exactly, we will # never need to do a masked load to handle stragglers at the end. # It's faster to avoid masking at all. But it is sound to always # mask. return V.graph.sizevars.statically_known_multiple_of(tree.numel, max_block) def filter_masks(self, mask_vars): for tree in self.range_trees: if self._has_constant_mask(tree): mask_vars.discard(f"{tree.prefix}mask") def iteration_ranges_codegen_header(self, entry, code): x = entry.prefix if entry.is_loop: code.writeline(f"{entry.name} = {x}offset + {x}base") elif entry.grid_dim is None: # no need to "{x}offset = " code.writeline(f"{entry.name} = {self.iteration_ranges_ranges_code(entry)}") code.writeline(f"{x}offset = 0") else: if entry.tensor_dim is not None: line = f"{x}offset + {self.iteration_ranges_ranges_code(entry)}" else: line = self.iteration_ranges_scalar_code(entry, f"{x}offset") code.writelines( [ f"{x}offset = {self.iteration_ranges_get_pid(entry)} * {x.upper()}BLOCK", f"{entry.name} = {line}", ] ) if self._has_constant_mask(entry): sizes = self.dense_size_str() code.writeline(f"{x}mask = tl.full({sizes}, True, tl.int1)") else: code.writeline(f"{x}mask = {entry.name} < {x}numel") class TritonScheduling(SIMDScheduling): int32_type = "tl.int32" int64_type = "tl.int64" kernel_type = TritonKernel backend_features = dict.fromkeys( # dict for deterministic order [ BackendFeature.FOREACH, BackendFeature.BUCKETIZE, BackendFeature.INPLACE_BUFFERS, BackendFeature.MASKED_SCATTER_WITH_INDEX, BackendFeature.SCAN, BackendFeature.TRITON_TEMPLATES, ] ) if torch.version.hip is None: backend_features.update( dict.fromkeys( [ # TODO: Move this above when ROCm triton adds support for multiple inputs BackendFeature.TUPLE_REDUCTION, BackendFeature.SORT, ] ) ) @classmethod def get_backend_features(cls, device: torch.device): return cls.backend_features def codegen_comment(self, node_schedule): wrapper = V.graph.wrapper_code origins, detailed_origins = get_kernel_metadata(node_schedule, wrapper) if origins: wrapper.writeline(origins) if config.debug_fusion: from torch._inductor.scheduler import ( BaseSchedulerNode, ForeachKernelSchedulerNode, ) if not any( isinstance(n, ForeachKernelSchedulerNode) for n in node_schedule ): # We probably should look what are the nodes inside a foreach # schedule node node_names = [ n.get_name() for n in node_schedule if isinstance(n, BaseSchedulerNode) ] wrapper.writeline( f"{wrapper.comment} Fused node name list: {', '.join(node_names)}" ) def define_kernel(self, src_code, node_schedule, kernel): wrapper = V.graph.wrapper_code if src_code in wrapper.src_to_kernel: kernel_name = wrapper.src_to_kernel[src_code] else: fused_name = ( get_fused_kernel_name(node_schedule, config.triton.descriptive_names) if config.triton.descriptive_names else "" ) kernel_category = get_kernel_category_by_source_code(src_code)[:3] kernel_name = "_".join( ["triton", kernel_category, fused_name, wrapper.next_kernel_suffix()] ) # use the original src_code as the key wrapper.src_to_kernel[src_code] = kernel_name subs_name = kernel_name if config.triton.unique_kernel_names else "triton_" # DESCRIPTIVE_NAME is used for profiling purposes; it shows the full kernel name # even when unique_kernel_names is turned off. Meanwhile, KERNEL_NAME is sometimes set # to "triton_" to maximize caching opportunities (when unique_kernel_names = False). src_code = src_code.replace(str(Placeholder.DESCRIPTIVE_NAME), kernel_name) src_code = src_code.replace(str(Placeholder.KERNEL_NAME), subs_name) # TODO(voz): Ostensibly, we should not need this. But there are cases where C++ codegen does # not use BracesBuffer, so we have no good indicator of a C++ buffer atm. src_code = src_code.replace("#pragma CMT", "#") basename, _, kernel_path = get_path(code_hash(src_code.strip()), "py") compile_wrapper = IndentedBuffer() compile_wrapper.writeline(f"async_compile.triton({subs_name!r}, '''") compile_wrapper.splice(src_code, strip=True) current_device = V.graph.scheduler.get_current_device_or_throw() compile_wrapper.writeline(f"''', device_str='{current_device.type}')") metadata_comment = f"# kernel path: {kernel_path}" origins, detailed_origins = get_kernel_metadata(node_schedule, wrapper) metadata_comment += "\n" + origins + "\n" + detailed_origins wrapper.define_kernel( kernel_name, compile_wrapper.getvalue(), metadata_comment ) # log kernel metadata for offline analysis. # E.g. one can find all unaligned inner reduction and check if # padding helps with the perf kernel by kernel. if is_metric_table_enabled("kernel_metadata"): log_kernel_metadata(kernel_name, kernel_path, src_code) return kernel_name def benchmark_fused_nodes(self, nodes): with preserve_rng_state(): src_code = self.generate_kernel_code_from_nodes( nodes, benchmark_kernel=True ) mod = PyCodeCache.load(src_code) def cache_file_path(): assert mod.__file__ is not None return os.path.splitext(mod.__file__)[0] + ".kernel_perf" def load_cache(): path = cache_file_path() if os.path.exists(path): with open(path) as fd: return float(fd.read()) return None def store_cache(): path = cache_file_path() with open(path, "w") as fd: fd.write(str(ms)) log.debug( "kernel src code for %s written to: %s", {n.get_name() for n in nodes}, mod.__file__, ) ms = load_cache() if ms is not None: return ms, mod.__file__ args = mod.get_args() call = mod.call wrapped_jit_function = mod.triton_ # call once to trigger the compilation try: call(wrapped_jit_function.clone_args(*args)[0]) except Exception as e: log.debug( "Exception (%s) in compiling fused nodes %s", e, {n.get_name() for n in nodes}, ) ms = float("inf") store_cache() return ms, mod.__file__ launchers = wrapped_jit_function.launchers assert len(launchers) == 1 if launchers[0].n_spills > 0: # skip benchmarking the kernel if there are register spills ms = float("inf") else: # We have to clone the inplace updated arguments to avoid earlier calls # generating out of range indices for later calls. ms = benchmarker.benchmark_gpu( lambda: call(wrapped_jit_function.clone_args(*args)[0]) ) # overhead of cloning args gives bias for fusing the kernel # in the case of mutating/in-placeable second fusion # TODO - would be better as a hook in triton do_bench that reset # the input values between benchmarking ms = ms - benchmarker.benchmark_gpu( lambda: wrapped_jit_function.clone_args(*args) ) log.debug( "The fused kernel for %s took %.3f ms to run", {n.get_name() for n in nodes}, ms, ) store_cache() return ms, mod.__file__ def benchmark_combo_kernel(self, node_list): def cache_file_path(): assert mod.__file__ is not None return os.path.splitext(mod.__file__)[0] + ".kernel_perf" def load_cache(): path = cache_file_path() if os.path.exists(path): with open(path) as fd: return tuple(float(e) for e in fd.read().split()) return (None, None) def store_cache(): path = cache_file_path() with open(path, "w") as fd: fd.write(str(ms) + " " + str(ms_clone)) total_ms, file_list = 0, [] total_clone_ms = 0 removed_buffers_orig = V.graph.removed_buffers V.graph.removed_buffers = OrderedSet(removed_buffers_orig) inplaced_to_remove_orig = V.graph.inplaced_to_remove V.graph.inplaced_to_remove = OrderedSet(inplaced_to_remove_orig) enable_autotune = config.combo_kernels_autotune > 0 mixed_sizes = config.combo_kernel_allow_mixed_sizes > 0 kernel_code_list = self.generate_combo_kernel_code( subkernel_nodes=node_list, custom_part_algorithm=True, enable_autotune=enable_autotune, mixed_sizes=mixed_sizes, only_gen_src_code=True, ) for src_code, _, node_group in kernel_code_list: fused_node_lists = [node.get_nodes() for node in node_group] names = [n.get_name() for nodes in fused_node_lists for n in nodes] src_code = src_code.replace(str(Placeholder.KERNEL_NAME), "triton_") mod = PyCodeCache.load(src_code) log.debug( "kernel src code for %s written to: %s", names, mod.__file__, ) ms, ms_clone = load_cache() if ms is not None: total_ms += ms total_clone_ms += ms_clone file_list.append(mod.__file__) continue args = mod.get_args() call = mod.call wrapped_jit_function = mod.triton_ # call once to trigger the compilation call(wrapped_jit_function.clone_args(*args)[0]) launchers = wrapped_jit_function.launchers assert len(launchers) == 1 if launchers[0].n_spills > 0: # skip benchmarking the kernel if there are register spills ms = ms_clone = float("inf") else: # We have to clone the inplace updated arguments to avoid earlier calls # generating out of range indices for later calls. ms = benchmarker.benchmark_gpu( lambda: call(wrapped_jit_function.clone_args(*args)[0]) ) ms_clone = benchmarker.benchmark_gpu( lambda: wrapped_jit_function.clone_args(*args)[0] ) log.debug( "The fused kernel for %s took %.3f ms to run, %.3f ms to clone inputs", {n.get_name() for n in node_group}, ms, ms_clone, ) store_cache() total_ms += ms total_clone_ms += ms_clone file_list.append(mod.__file__) V.graph.removed_buffers = removed_buffers_orig V.graph.inplaced_to_remove = inplaced_to_remove_orig return total_ms, total_clone_ms, file_list