tinygrad/extra/ops.py

136 lines
6.7 KiB
Python

from __future__ import annotations
from typing import Dict, Union, Tuple, Any, List, cast
import functools, hashlib
from enum import Enum, auto
from dataclasses import dataclass
from tinygrad.helpers import dedup, pretty_print, prod
from tinygrad.ops import ReduceOps, UnaryOps, BinaryOps, TernaryOps, UOp, UOps
from tinygrad.dtype import ImageDType, PtrDType, dtypes, DType, ConstType
from tinygrad.shape.symbolic import Variable, sint
from tinygrad.shape.shapetracker import ShapeTracker
# these ops are deleted after AST is UOp
class BufferOps(Enum): LOAD = auto(); CONST = auto(); STORE = auto() # noqa: E702
class MetaOps(Enum): KERNEL = auto();
Op = Union[UnaryOps, BinaryOps, ReduceOps, MetaOps, TernaryOps, BufferOps]
@dataclass(frozen=True)
class MemBuffer:
idx: int
dtype: DType
st: ShapeTracker
@dataclass(frozen=True)
class ConstBuffer:
val: ConstType | Variable
dtype: DType
st: ShapeTracker
@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:LazyOp): return pretty_print(self, lambda x: f'LazyOp({x.op}, arg={x.arg}, src=(%s))')
@functools.cached_property
def dtype(self) -> DType:
if self.op in BufferOps: return self.arg.dtype
if self.op in [UnaryOps.CAST, UnaryOps.BITCAST]: return self.arg
return dtypes.bool if self.op in {BinaryOps.CMPLT, BinaryOps.CMPNE} else self.src[-1].dtype
@functools.cached_property
def full_shape(self) -> Tuple[sint, ...]:
if len(self.src) == 0 and self.op in BufferOps: return self.arg.st.shape
return tuple(max(x) for x in zip(*[x.full_shape for x in self.src]))
@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]:
extract_vars = [x.arg.st.vars() for x in self.lazyops if x.op in BufferOps]
const_vars = [x.arg.val for x in self.lazyops if x.op is BufferOps.CONST and isinstance(x.arg.val, Variable)]
return sorted(set.union(*extract_vars, set(const_vars)), key=lambda v: v.expr)
def __add__(self, x:LazyOp): return LazyOp(BinaryOps.ADD, (self, x))
def __sub__(self, x:LazyOp): return LazyOp(BinaryOps.ADD, (self, -x))
def __mul__(self, x:LazyOp): return LazyOp(BinaryOps.MUL, (self, x))
def ne(self, x:LazyOp): return LazyOp(BinaryOps.CMPNE, (self, x))
def eq(self, x:LazyOp): return -self.ne(x)
def __neg__(self): return LazyOp(UnaryOps.NEG, (self,))
@staticmethod
def const(val, dtype:DType, shape:Tuple[sint, ...]):
return LazyOp(BufferOps.CONST, (), ConstBuffer(val, dtype, ShapeTracker.from_shape(()).reshape((1,)*len(shape)).expand(shape)))
# the living definition of LazyOps
def verify_lazyop(ast:LazyOp) -> Dict[LazyOp, ShapeTracker]:
assert ast.op is MetaOps.KERNEL, "must be SINK"
sts: Dict[LazyOp, ShapeTracker] = {}
def assert_valid(op:LazyOp, st:ShapeTracker):
if op in sts: return
# restore globals from the two stage reduce
if op.op is BufferOps.LOAD and op.arg.idx < 0:
assert_valid(local_reduce:=op.src[0].src[0], op.arg.st)
return sts.setdefault(op, sts[local_reduce])
for x in op.src: assert_valid(x, st)
# only reduceop is allowed to change shape, limited to turning n to 1
if op.op in ReduceOps:
axis = op.arg
assert isinstance(axis, tuple) and all(isinstance(i, int) for i in axis), f"reduceop must have axis {op.arg}"
st = ShapeTracker.from_shape(sts[op.src[0]].reduce(axis))
else:
# movementops are pushed to the edges with LOAD
# elementwise inherits shape
st = op.arg.st if op.op in BufferOps else sts[op.src[0]]
for x in op.src:
if sts[x].shape != st.shape:
if prod(sts[x].shape) == prod(st.shape): raise AssertionError(f"found implicit reshape {x.op} {op.op} {sts[x].shape} != {st.shape}")
raise AssertionError(f"found implicit expand {x.op} {sts[x].shape} != {op.op} {st.shape} {prod(sts[x].shape)} != {prod(st.shape)}")
sts[op] = st
for i, out in enumerate(ast.src):
assert out.arg.idx == i, f"unexpected output buffer idx {out.arg.idx} != {i}"
assert out.op is BufferOps.STORE, f"kernels must have stores as the output, got {out.op}"
assert out.arg.st.size == ast.src[-1].arg.st.size, f"outputs must have the same size, got {out.arg.st.size}"
assert_valid(out, out.arg.st)
shape_dims = [sorted(dedup(dims)) for dims in zip(*[x.shape for x in sts.values()])]
assert all(len(x) == 1 or (len(x) == 2 and x[0] == 1) for x in shape_dims), f"shapes must have either 1 or n in each dimension, {shape_dims}"
return sts
def to_uop(*a) -> UOp:
assert isinstance(a[0], LazyOp), f"{a} must be a LazyOp ast"
if a[0].op is BufferOps.STORE: ast = LazyOp(MetaOps.KERNEL, a)
else:
assert a[0].op is MetaOps.KERNEL
ast = a[0]
verify_lazyop(ast)
@functools.lru_cache(None)
def create_uop(lop:LazyOp) -> UOp:
if lop.op in BufferOps:
st_uop = lop.arg.st.to_uop()
membuf_dtype: DType = lop.arg.dtype
dtype = membuf_dtype.base if isinstance(membuf_dtype, ImageDType) else membuf_dtype
if lop.op is BufferOps.CONST:
return UOp(UOps.CONST, dtype, (st_uop,), lop.arg.val)
buf = UOp(UOps.DEFINE_GLOBAL, membuf_dtype if isinstance(membuf_dtype, ImageDType) else PtrDType(membuf_dtype), (), lop.arg.idx)
if lop.op is BufferOps.LOAD: return UOp(UOps.LOAD, dtype, (buf, st_uop))
return UOp(UOps.STORE, dtypes.void, (buf, st_uop, create_uop(lop.src[0])))
src = tuple(create_uop(x) for x in lop.src)
if lop.op is MetaOps.KERNEL: return UOp(UOps.SINK, dtypes.void, src)
if lop.op in ReduceOps:
alu_op = {ReduceOps.SUM:BinaryOps.ADD, ReduceOps.PROD:BinaryOps.MUL, ReduceOps.MAX:BinaryOps.MAX}[cast(ReduceOps, lop.op)]
return UOp(UOps.REDUCE_AXIS, src[0].dtype, src, (alu_op, lop.arg))
if lop.op is UnaryOps.CAST: return UOp(UOps.CAST, lop.arg.scalar(), src)
if lop.op is UnaryOps.BITCAST: return UOp(UOps.BITCAST, lop.arg.scalar(), src)
return src[0].alu(lop.op, *src[1:])
ret = create_uop(ast)
#with open("/tmp/ast", "w") as f: f.write(str(ret))
return ret