forked from tinygrad/tinygrad
99 lines
5.0 KiB
Python
99 lines
5.0 KiB
Python
from __future__ import annotations
|
|
from typing import Union, Type, Tuple, Any, List, Dict, Callable
|
|
import functools, hashlib
|
|
from enum import Enum, auto
|
|
from dataclasses import dataclass
|
|
from tinygrad.helpers import prod, dedup
|
|
from tinygrad.dtype import dtypes, DType
|
|
from tinygrad.shape.symbolic import Variable, sint
|
|
from tinygrad.shape.shapetracker import ShapeTracker
|
|
from tinygrad.buffer import Buffer
|
|
|
|
# these are the llops your accelerator must implement, along with toCpu
|
|
# the Enum class doesn't work with mypy, this is static. sorry it's ugly
|
|
# NOTE: MOD, CMPLT don't have to be implemented on vectors, just scalars
|
|
# NOTE: many GPUs don't have DIV, but UnaryOps.RECIP doesn't work for integer division
|
|
class UnaryOps(Enum): EXP2 = auto(); LOG2 = auto(); CAST = auto(); SIN = auto(); SQRT = auto(); NEG = auto() # noqa: E702
|
|
class BinaryOps(Enum):
|
|
ADD = auto(); SUB = auto(); MUL = auto(); DIV = auto(); MAX = auto(); MOD = auto(); CMPLT = auto(); CMPEQ = auto(); XOR = auto() # noqa: E702
|
|
class TernaryOps(Enum): WHERE = auto() # noqa: E702
|
|
class ReduceOps(Enum): SUM = auto(); MAX = auto() # noqa: E702
|
|
class BufferOps(Enum): LOAD = auto(); CONST = auto(); STORE = auto() # noqa: E702
|
|
class LoadOps(Enum): EMPTY = auto(); CONST = auto(); COPY = auto(); CONTIGUOUS = auto(); CUSTOM = auto(); SYNC = auto(); ASSIGN = auto() # noqa: E702
|
|
|
|
Op = Union[UnaryOps, BinaryOps, ReduceOps, LoadOps, TernaryOps, BufferOps]
|
|
OpType = Union[Type[UnaryOps], Type[BinaryOps], Type[ReduceOps], Type[LoadOps], Type[TernaryOps], Type[BufferOps]]
|
|
|
|
@dataclass(frozen=True)
|
|
class MemBuffer:
|
|
idx: int
|
|
dtype: DType
|
|
st: ShapeTracker
|
|
|
|
@dataclass(frozen=True)
|
|
class ConstBuffer:
|
|
val: Union[int, float]
|
|
dtype: DType
|
|
st: ShapeTracker
|
|
|
|
@dataclass(frozen=True)
|
|
class ScheduleItem:
|
|
ast: Tuple[LazyOp, ...]
|
|
outputs: Tuple[Buffer, ...]
|
|
inputs: Tuple[Buffer, ...]
|
|
var_vals: Dict[Variable, int]
|
|
|
|
@dataclass(frozen=True, eq=False)
|
|
class LazyOp:
|
|
op: Op
|
|
src: Tuple[LazyOp, ...] = ()
|
|
arg: Any = None
|
|
def cached_compare(self, x, context):
|
|
if id(self) == id(x): return True
|
|
if self.op != x.op or self.arg != x.arg or len(self.src) != len(x.src): return False
|
|
if (key := (id(self), id(x))) in context: return context[key]
|
|
ret = context[key] = all(a.cached_compare(b, context) for a,b in zip(self.src, x.src))
|
|
return ret
|
|
def __eq__(self, x): return self.cached_compare(x, context={})
|
|
def __repr__(self): return f"LazyOp(op={self.op}, src={self.src}, arg={self.arg})"
|
|
@functools.cached_property
|
|
def key(self) -> bytes:
|
|
return hashlib.sha256(functools.reduce(lambda x,y: x+y, [s.key for s in self.src], str((self.op, self.arg)).encode())).digest()
|
|
@functools.cached_property
|
|
def hash(self): return hash((self.op, self.src, self.arg))
|
|
def __hash__(self): return self.hash
|
|
@functools.cached_property
|
|
def lazyops(self) -> List[LazyOp]: return dedup([self] + [item for x in self.src for item in x.lazyops])
|
|
def vars(self) -> List[Variable]:
|
|
return sorted(set.union(*[x.arg.st.vars() for x in self.lazyops if x.op in BufferOps], set()), key=lambda x: str(x.expr))
|
|
|
|
# **************** independent FlopCounter ****************
|
|
|
|
@dataclass
|
|
class FlopCounter:
|
|
shape: Tuple[int, ...]
|
|
dtype: DType
|
|
flops: sint
|
|
mem: Dict[int, int]
|
|
@property
|
|
def mem_estimate(self): return sum(self.mem.values())
|
|
def consume_flops(self):
|
|
self.flops, ret = 0, self.flops
|
|
return ret
|
|
|
|
InterpretedFlopCounter: Dict[Op, Callable] = {
|
|
BufferOps.LOAD: lambda arg: FlopCounter(arg.st.shape, arg.dtype, 0, {arg.idx: arg.dtype.itemsize*arg.st.real_size()}),
|
|
BufferOps.CONST: lambda arg: FlopCounter(arg.st.shape, arg.dtype, 0, {}),
|
|
BufferOps.STORE: lambda self,arg: FlopCounter(arg.st.shape, arg.dtype, self.consume_flops(), {**self.mem, arg.idx: arg.dtype.itemsize*arg.st.real_size()}), # noqa: E501
|
|
UnaryOps.CAST: lambda self,arg: FlopCounter(self.shape, arg[0], self.consume_flops(), self.mem), # cast uses no flops
|
|
**{op:lambda self: FlopCounter(self.shape, self.dtype, self.consume_flops() + prod(self.shape), self.mem) for op in UnaryOps if op is not UnaryOps.CAST}, # noqa: E501
|
|
**{op:lambda self,y,op=op: FlopCounter(self.shape, dtypes.bool if op in (BinaryOps.CMPLT, BinaryOps.CMPEQ) else self.dtype, self.consume_flops() + y.consume_flops() + prod(self.shape), {**self.mem, **y.mem}) for op in BinaryOps}, # noqa: E501
|
|
**{op:lambda self,axis: FlopCounter(tuple(1 if i in axis else s for i,s in enumerate(self.shape)), self.dtype, self.consume_flops() + prod(self.shape), self.mem) for op in ReduceOps}, # noqa: E501
|
|
TernaryOps.WHERE: lambda self,y,z: FlopCounter(self.shape, y.dtype, self.consume_flops() + y.consume_flops() + z.consume_flops() + prod(self.shape), {**self.mem, **y.mem, **z.mem})} # noqa: E501
|
|
|
|
@functools.lru_cache(None)
|
|
def get_lazyop_info(ast:LazyOp) -> FlopCounter:
|
|
@functools.lru_cache(None) # NOTE: this cache needs to be recreated for new ASTs
|
|
def run_ast(ast): return InterpretedFlopCounter[ast.op](*([run_ast(x) for x in ast.src]+([ast.arg] if ast.arg is not None else [])))
|
|
return run_ast(ast)
|