IQ.Pilot Prebuilt Release @ 27f668a
This commit is contained in:
IQ.Lvbs CI [bot]
140
artifacts/package_runtime/tinygrad/uop/__init__.py
Normal file
140
artifacts/package_runtime/tinygrad/uop/__init__.py
Normal file
@@ -0,0 +1,140 @@
|
||||
# flake8: noqa: E702
|
||||
# allow semicolons to put multiple ops on one line
|
||||
from enum import auto, IntEnum, Enum
|
||||
|
||||
# wrapper around IntEnum that preserves Enum.__str__ and makes auto() unique across all FastEnum subclasses
|
||||
class FastEnum(IntEnum):
|
||||
def __str__(self): return Enum.__str__(self)
|
||||
def __repr__(x): return str(x)
|
||||
@staticmethod
|
||||
def _generate_next_value_(_, __, ___, last_values): return 1 + max([0, *last_values, *[max(c) for c in FastEnum.__subclasses__()]])
|
||||
|
||||
# the order of these Ops controls the order of the toposort
|
||||
class Ops(FastEnum):
|
||||
# ** 1 -- defines/special **
|
||||
|
||||
# BIND pairs a symbolic PARAM with a concrete value
|
||||
BIND = auto()
|
||||
|
||||
# this is a RANGE for GPU dimensions, similar to symbolic shapes but not exactly
|
||||
SPECIAL = auto()
|
||||
|
||||
# BUFFER allocates global/local/register storage depending on its addrspace
|
||||
BUFFER = auto()
|
||||
|
||||
# ** 2 -- non op uops **
|
||||
|
||||
# uops that aren't rendered
|
||||
NOOP = auto(); REWRITE_ERROR = auto()
|
||||
# FUNCTION has a TUPLE body and is gradient-able; CALL is an opaque kernel invocation
|
||||
PARAM = auto(); FUNCTION = auto(); CALL = auto()
|
||||
|
||||
# renderer
|
||||
# LINEAR is a list of UOps, SOURCE has a str arg that's human readable, BINARY has bytes arg that's compiled
|
||||
PROGRAM = auto(); LINEAR = auto(); SOURCE = auto(); BINARY = auto()
|
||||
|
||||
# AFTER passes src[0] through and promises in the toposort that any consumers of the AFTER run after src[1:]
|
||||
# GROUP is a NOOP that just merges things together
|
||||
SINK = auto(); AFTER = auto(); GROUP = auto()
|
||||
|
||||
# vector creation / item selection
|
||||
STACK = auto()
|
||||
|
||||
# tuple/gettuple for function with multiple returns
|
||||
TUPLE = auto(); GETTUPLE = auto()
|
||||
|
||||
# hcq specific
|
||||
GETADDR = auto()
|
||||
|
||||
# ** 3 -- load/store **
|
||||
|
||||
# INDEX is a BinaryOp similar to ADD, but it operates on pointers
|
||||
INDEX = auto(); SHRINK = auto()
|
||||
|
||||
# load/store before math
|
||||
LOAD = auto(); STORE = auto()
|
||||
|
||||
# ** 4 -- math **
|
||||
|
||||
# tensor core math op, not elementwise
|
||||
WMMA = auto()
|
||||
|
||||
# UnaryOps
|
||||
CAST = auto(); BITCAST = auto(); EXP2 = auto(); LOG2 = auto(); SIN = auto()
|
||||
SQRT = auto(); RECIPROCAL = auto(); NEG = auto(); TRUNC = auto()
|
||||
|
||||
# BinaryOps
|
||||
ADD = auto(); MUL = auto(); SHL = auto(); SHR = auto(); CDIV = auto(); MAX = auto(); CMOD = auto()
|
||||
CMPLT = auto(); CMPNE = auto(); CMPEQ = auto()
|
||||
XOR = auto(); OR = auto(); AND = auto()
|
||||
THREEFRY = auto(); SUB = auto(); FDIV = auto(); POW = auto()
|
||||
FLOORDIV = auto(); FLOORMOD = auto()
|
||||
|
||||
# TernaryOps
|
||||
WHERE = auto(); MULACC = auto()
|
||||
|
||||
# ** 5 -- control flow / consts / custom **
|
||||
|
||||
# control flow ops
|
||||
BARRIER = auto(); RANGE = auto(); IF = auto(); END = auto(); ENDIF = auto(); WAIT = auto()
|
||||
|
||||
# const.
|
||||
CONST = auto()
|
||||
|
||||
# CUSTOM/CUSTOMI are used to output strings into codegen. the I makes the string inline
|
||||
CUSTOM = auto(); CUSTOMI = auto()
|
||||
|
||||
# INS is a machine instruction
|
||||
INS = auto()
|
||||
|
||||
# ** 6 -- ops that don't exist in programs **
|
||||
|
||||
# ops that adjust the behavior of the scheduler
|
||||
CONTIGUOUS = auto(); CONTIGUOUS_BACKWARD = auto(); DETACH = auto()
|
||||
|
||||
# buffer ops
|
||||
STAGE = auto(); COPY = auto(); SLICE = auto(); MSELECT = auto(); MSTACK = auto(); CUSTOM_FUNCTION = auto()
|
||||
|
||||
# the core 6 movement ops! these only exist in the tensor graph
|
||||
RESHAPE = auto(); PERMUTE = auto(); EXPAND = auto(); PAD = auto(); FLIP = auto()
|
||||
UNSHARD = auto() # UNSHARD is really a movement op
|
||||
|
||||
# reduce
|
||||
REDUCE = auto(); ALLREDUCE = auto()
|
||||
|
||||
# ** 7 -- pattern compiler IR (used in upat.py) **
|
||||
# PYLITERAL carries a Python literal as an arg for CUSTOM predicates
|
||||
PYLITERAL = auto()
|
||||
|
||||
class GroupOp:
|
||||
Unary = {Ops.EXP2, Ops.LOG2, Ops.SIN, Ops.SQRT, Ops.RECIPROCAL, Ops.NEG, Ops.TRUNC}
|
||||
Binary = {Ops.ADD, Ops.MUL, Ops.CDIV, Ops.MAX, Ops.CMOD, Ops.CMPLT, Ops.CMPNE, Ops.CMPEQ,
|
||||
Ops.XOR, Ops.SHL, Ops.SHR, Ops.OR, Ops.AND, Ops.THREEFRY, Ops.SUB, Ops.FDIV, Ops.POW, Ops.FLOORDIV, Ops.FLOORMOD}
|
||||
Ternary = {Ops.WHERE, Ops.MULACC}
|
||||
ALU = set.union(Unary, Binary, Ternary)
|
||||
Broadcastable = set.union(Binary, Ternary)
|
||||
|
||||
# TODO: is BITCAST always Elementwise if it's shape changing?
|
||||
Elementwise = set.union(ALU, {Ops.CAST, Ops.BITCAST})
|
||||
|
||||
Defines = {Ops.PARAM, Ops.BUFFER}
|
||||
|
||||
Irreducible = {Ops.CONST, Ops.SPECIAL, Ops.RANGE, Ops.PARAM, Ops.GETADDR}
|
||||
Movement = {Ops.RESHAPE, Ops.EXPAND, Ops.PERMUTE, Ops.PAD, Ops.SHRINK, Ops.FLIP}
|
||||
|
||||
# BinaryOps that can be flipped
|
||||
Commutative = {Ops.ADD, Ops.MUL, Ops.MAX, Ops.CMPNE, Ops.CMPEQ, Ops.XOR, Ops.AND, Ops.OR}
|
||||
|
||||
# BinaryOps where f(f(a,b),c) = f(a,f(b,c))
|
||||
Associative = {Ops.ADD, Ops.MUL, Ops.AND, Ops.OR, Ops.MAX}
|
||||
|
||||
# BinaryOps that satisfy f(x,x)=x see https://en.wikipedia.org/wiki/Idempotence
|
||||
Idempotent = {Ops.OR, Ops.AND, Ops.MAX}
|
||||
|
||||
# ALU ops valid as the reduce op in REDUCE/ALLREDUCE arg
|
||||
Reduce = {Ops.ADD, Ops.MUL, Ops.MAX}
|
||||
|
||||
# These can change the dtype to bool
|
||||
Comparison = {Ops.CMPLT, Ops.CMPNE, Ops.CMPEQ}
|
||||
|
||||
All = set(Ops)
|
||||
109
artifacts/package_runtime/tinygrad/uop/divandmod.py
Normal file
109
artifacts/package_runtime/tinygrad/uop/divandmod.py
Normal file
@@ -0,0 +1,109 @@
|
||||
import functools, itertools, math
|
||||
from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp
|
||||
from tinygrad.dtype import dtypes
|
||||
from tinygrad.helpers import unwrap
|
||||
|
||||
# NOTE: this cache is only on index UOps
|
||||
@functools.cache
|
||||
def fold_divmod_general(d: UOp) -> UOp|None:
|
||||
x, y = d.src
|
||||
|
||||
if y.vmin==y.vmax==0: raise ZeroDivisionError(f"{'Division' if d.op is Ops.FLOORDIV else 'Mod'} by zero trying to rewrite {x.alu(d.op, y)}")
|
||||
# x//y is constant
|
||||
if (xdiv:=x//y).vmin == xdiv.vmax: return x - xdiv.vmin*y if d.op is Ops.FLOORMOD else xdiv.const_like(xdiv.vmin)
|
||||
# PARAM // c is irreducible
|
||||
if x.op is Ops.PARAM and y.op is Ops.CONST and x.arg.multiple_of % y.val == 0: return d.const_like(0) if d.op is Ops.FLOORMOD else None
|
||||
|
||||
# split uops for the rest of the processing
|
||||
x_peeled, const = x.pop_const()
|
||||
uops_no_const = list(x_peeled.split_uop(Ops.ADD))
|
||||
|
||||
# ** Constant Denominator Rules **
|
||||
# these rules strictly require y to be a scalar constant > 0
|
||||
if y.op is Ops.CONST and (c := y.val) > 0:
|
||||
# nested_div: (x%(k*c))//c -> (x//c)%k (requires k>0); the mod case is handled by remove_nested_mod below
|
||||
if d.op is Ops.FLOORDIV and x.op is Ops.FLOORMOD and (k := x.src[1].divides(c)) is not None and k > 0: return x.src[0] // y % k
|
||||
|
||||
# remove_nested_mod in sum: (a%4 + b)%2 -> (a+b)%2
|
||||
if d.op is Ops.FLOORMOD:
|
||||
new_xs, changed = [], False
|
||||
for u in uops_no_const:
|
||||
if u.op is Ops.FLOORMOD and u.src[1].divides(c) is not None:
|
||||
u = u.src[0]
|
||||
changed = True
|
||||
new_xs.append(u)
|
||||
if changed: return (UOp.usum(*new_xs) + const) % y
|
||||
|
||||
# Shared decomposition for folding rules
|
||||
decomp = [(u.divides(f:=u.const_factor()),f) for u in uops_no_const]
|
||||
terms, factors = zip(*decomp)
|
||||
|
||||
# fold_divmod_congruence: fold if a is congruent to an expression whose range is between 0 and c
|
||||
# try both signs of the remainder for a lone term (covers a binary numerator that crosses one period)
|
||||
# or on an exact f%c == c//2 tie; otherwise pick the smaller to keep the product over terms small
|
||||
rem_choices = [(r, r-c) if (r:=f%c)*2 == c or len(terms)==1 else (min(r, r-c, key=abs),) for f in factors]
|
||||
for rems in itertools.product(*rem_choices):
|
||||
if (rem:=sum(r*v for r,v in zip(rems,terms))+const%c).vmin//c==rem.vmax//c:
|
||||
if d.op is Ops.FLOORMOD: return rem - rem.vmin//c*c
|
||||
return sum((f-r)//c * v for f,r,v in zip(factors,rems,terms)) + const//c + rem.vmin//c
|
||||
|
||||
# gcd_with_remainder: factor out common gcd from numerator
|
||||
if (g:=math.gcd(*factors, c)) > 1:
|
||||
new_x = unwrap(x_peeled.divides(g)).simplify() + (const//g)%(c//g)
|
||||
if new_x.vmin >= 0:
|
||||
if d.op is Ops.FLOORMOD: return new_x % (c//g) * g + const%g
|
||||
return new_x // (c//g) + const//c
|
||||
|
||||
# nest_by_factor: x//c -> (x//f)//(c//f), x%c -> (x//f%(c//f))*f + b where b=x%f
|
||||
# FLOORDIV identity holds for any sign of x; FLOORMOD reconstruction needs x.vmin>=0
|
||||
results = []
|
||||
for div in {abs(f) for u, f in zip(uops_no_const, factors) if u.op is not Ops.CONST and 1 < abs(f) < c and (c%f)==0}:
|
||||
if (newxs := fold_divmod_general(x//div)) is not None:
|
||||
if d.op is Ops.FLOORDIV:
|
||||
results.append((len(newxs.backward_slice), newxs // (c // div)))
|
||||
elif x.vmin >= 0 and newxs.vmin >= 0:
|
||||
b_parts = [f%div*t for f, t in zip(factors, terms) if f%div]
|
||||
if const % div: b_parts.append(x.const_like(const % div))
|
||||
b = UOp.usum(*b_parts) if b_parts else x.const_like(0)
|
||||
if 0 <= b.vmin and b.vmax < div:
|
||||
results.append((len((r:=(newxs % x.ufix(c//div))*div + b).backward_slice), r))
|
||||
if results: return min(results, key=lambda r: r[0])[1]
|
||||
|
||||
# ** Variable Denominator / Fallback Rules **
|
||||
# These rules apply to variables OR constants that failed the checks above.
|
||||
# Reconstruct all uops including const for these checks.
|
||||
all_uops = list(x.split_uop(Ops.ADD))
|
||||
|
||||
# divide_by_gcd: x//y -> (x//gcd)//(y//gcd)
|
||||
gcd = UOp.gcd(*all_uops, y).simplify()
|
||||
if not (gcd.op is Ops.CONST and gcd.val==1):
|
||||
ret = unwrap(x.divide_exact(gcd)).alu(d.op, unwrap(y.divide_exact(gcd)))
|
||||
return ret*gcd if d.op is Ops.FLOORMOD else ret
|
||||
|
||||
# factor_remainder: (d*x+y)//d -> x+y//d
|
||||
if y.vmin<0 or x.vmin<0: return None
|
||||
quo, rem = [], []
|
||||
for u in all_uops:
|
||||
if (q:=u.divide_exact(y)) is not None: quo.append(q)
|
||||
elif y.op is Ops.CONST and (c:=u.const_factor())%y.val!=c:
|
||||
rem.append(u.divides(c)*(c%y.val))
|
||||
quo.append(u.divides(c)*(c//y.val) if d.op is Ops.FLOORDIV else u.const_like(0))
|
||||
else: rem.append(u)
|
||||
|
||||
if not quo: return None
|
||||
new_x = sum(rem)+x.const_like(0)
|
||||
if new_x.vmin<0: return None
|
||||
return new_x%y if d.op is Ops.FLOORMOD else new_x//y+sum(quo)
|
||||
|
||||
div_and_mod_symbolic = PatternMatcher([
|
||||
# ** 1. Fast Inline Rules **
|
||||
# (x//c+a)//d -> (x+a*c)//(c*d) for c>0, d>0
|
||||
((UPat.var("x")//UPat.cvar("c") + UPat.cvar("a"))//UPat.cvar("d"), lambda x,c,a,d: (x+a*c)//(c*d) if d.vmin>0 else None),
|
||||
# (x+c)//d -> (x+c%d)//d + c//d ; (x+c)%d -> (x+c%d)%d (split the multiple of d out of the const, holds for any d!=0)
|
||||
(UPat((Ops.FLOORDIV, Ops.FLOORMOD), src=(UPat.var("x", dtypes.weakint)+UPat.cvar("c"), UPat.cvar("d")), name="n"),
|
||||
lambda n,x,c,d: None if d.val==0 or c.val%d.val==c.val else
|
||||
(x+c.val%d.val)//d + c.val//d.val if n.op is Ops.FLOORDIV else (x+c.val%d.val)%d),
|
||||
|
||||
# ** 2. Slow Rules **
|
||||
(UPat((Ops.FLOORDIV, Ops.FLOORMOD), dtypes.weakint, name="d"), lambda d: fold_divmod_general(d)),
|
||||
])
|
||||
26
artifacts/package_runtime/tinygrad/uop/movement.py
Normal file
26
artifacts/package_runtime/tinygrad/uop/movement.py
Normal file
@@ -0,0 +1,26 @@
|
||||
from tinygrad.uop.ops import PatternMatcher, UPat, Ops
|
||||
|
||||
# TODO: pm_mops from rangeify belongs here. this is all pattern matchers that strictly clean up movement ops
|
||||
|
||||
mop_cleanup = PatternMatcher([
|
||||
# merge adjacent RESHAPES
|
||||
(UPat(Ops.RESHAPE, src=(UPat(Ops.RESHAPE, name="x2"), UPat()), name="x"), lambda x,x2: x.replace(src=(x2.src[0], x.src[1]))),
|
||||
# remove noop RESHAPEs
|
||||
(UPat(Ops.RESHAPE, src=(UPat(name="x2"), UPat()), name="x"), lambda x,x2: x2 if x2._shape is not None and x2.shape == x.shape else None),
|
||||
# merge PERMUTEs
|
||||
(UPat(Ops.PERMUTE, src=(UPat(Ops.PERMUTE, name="x2"),), name="x"), lambda x,x2: x2.replace(arg=tuple(x2.arg[i] for i in x.arg))),
|
||||
# remove noop PERMUTEs
|
||||
(UPat(Ops.PERMUTE, name="x"), lambda x: x.src[0] if list(x.arg) == list(range(len(x.arg))) else None),
|
||||
# STACK on INDEX CONST
|
||||
(UPat(Ops.STACK, src=UPat(Ops.INDEX, src=(UPat.var("src"), UPat(Ops.CONST))), name="stk"),
|
||||
lambda src,stk: src if stk.shape == src.shape and list(range(len(stk.src))) == [x.src[1].val for x in stk.src] else None),
|
||||
# const INDEX into STACK is src
|
||||
(UPat(Ops.INDEX, src=(UPat(Ops.STACK, name="a"), UPat.cvar("i")), name="idx", allow_any_len=True),
|
||||
lambda a,i,idx: a.src[i.val] if len(idx.src) <= 2 else a.src[i.val].index(*idx.src[2:])),
|
||||
# INDEX on INDEX is INDEX
|
||||
(UPat(Ops.INDEX, src=(UPat(Ops.INDEX, name="idx1", allow_any_len=True),), allow_any_len=True, name="idx2"),
|
||||
lambda idx1,idx2: idx1.src[0].index(*idx1.src[1:], *idx2.src[1:]) if all(x.shape == () for x in idx1.src[1:]+idx2.src[1:]) else None),
|
||||
# INDEX on shaped INDEX (TODO: this can be more generic)
|
||||
(UPat(Ops.INDEX, src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx1_arg"))),), allow_any_len=True, name="idx2"),
|
||||
lambda buf,idx1_arg,idx2: buf.index(idx1_arg.index(*idx2.src[1:])) if len(idx1_arg.shape) == len(idx2.src[1:]) else None),
|
||||
])
|
||||
1782
artifacts/package_runtime/tinygrad/uop/ops.py
Normal file
1782
artifacts/package_runtime/tinygrad/uop/ops.py
Normal file
File diff suppressed because it is too large
Load Diff
160
artifacts/package_runtime/tinygrad/uop/render.py
Normal file
160
artifacts/package_runtime/tinygrad/uop/render.py
Normal file
@@ -0,0 +1,160 @@
|
||||
from tinygrad.dtype import AddrSpace, dtypes
|
||||
from tinygrad.uop import Ops, GroupOp
|
||||
from tinygrad.uop.ops import ParamArg, UOp, PatternMatcher, UPat, multirange_str, range_str, consumer_map_from_toposort
|
||||
from tinygrad.helpers import strip_parens
|
||||
|
||||
def pretty_print(x:UOp, cache=None, d=0)->str:
|
||||
def dfs(x:UOp, cache:dict):
|
||||
for s in x.src:
|
||||
cache.setdefault(s, [len(cache), 0, False])[1] += 1
|
||||
if cache[s][1] == 1: dfs(s, cache)
|
||||
if cache is None: dfs(x, cache:={})
|
||||
if (cx:=cache.setdefault(x, [0,0,False]))[2]: return f"{' '*d}x{cx[0]}"
|
||||
cx[2], srcs = True, (''.join(f'\n{pretty_print(s, cache, d+2)},' for s in x.src))
|
||||
return f"{' '*d}{f'x{cx[0]}:=' * (cx[1]>1)}{type(x).__name__}({x.op}, {x.dtype}, arg={x.argstr()}{x.tagstr()}, src=({srcs}))"
|
||||
|
||||
# ***** uop helpers *****
|
||||
|
||||
def print_uops(uops:list[UOp]):
|
||||
uops_index = {u:i for i,u in enumerate(uops)}
|
||||
for i,u in enumerate(uops):
|
||||
formatted_srcs = [(uops_index[x] if x.op is not Ops.CONST else f"{x.val}") if x in uops else "--" for x in u.src]
|
||||
print(f"{i:4d} {str(u.op):20s}: {multirange_str(u.ranges, color=True, pad=10)} {str(u.dtype):40s} " f"{str(formatted_srcs):32s} {u.arg}")
|
||||
|
||||
# for debug
|
||||
syms = { Ops.ADD: "+", Ops.SUB: "-", Ops.FLOORDIV: "//", Ops.FLOORMOD: "%", Ops.SHL: "<<", Ops.SHR: ">>",
|
||||
Ops.MUL: "*", Ops.CMPLT: "<", Ops.CMPNE: "!=", Ops.AND: "&", Ops.OR: "|", Ops.XOR: "^"}
|
||||
# comparison operators are not in here because they are chained in python, not left-associative
|
||||
precedence = {Ops.MUL:1, Ops.FLOORDIV:1, Ops.FLOORMOD:1, Ops.ADD:2, Ops.SUB:2, Ops.SHL:3, Ops.SHR:3, Ops.AND:4, Ops.XOR:5, Ops.OR:6}
|
||||
def strip_binary_parens(x:UOp, left:str, right:str, code_for_op) -> str:
|
||||
if x.op not in precedence: return code_for_op(left, right)
|
||||
return code_for_op(strip_parens(left) if precedence.get(x.src[0].op,99)<=precedence[x.op] else left, strip_parens(right) if
|
||||
precedence.get(x.src[1].op,99)<precedence[x.op] else right)
|
||||
|
||||
renderer = PatternMatcher([
|
||||
(UPat(Ops.PARAM, name="x"), lambda x: x.arg.name if x.arg.name is not None else f"p{x.arg.slot}"),
|
||||
(UPat((Ops.SPECIAL), name="x"), lambda x: x.arg),
|
||||
(UPat(Ops.RANGE, dtypes.void, name="x"), lambda x: f"loop{x.arg[0]}"),
|
||||
(UPat(Ops.RANGE, name="x"), lambda x: f"r{range_str(x)}"),
|
||||
(UPat(Ops.CONST, name="x"), lambda x: str(x.val)),
|
||||
(UPat(Ops.CAST, name="x"), lambda ctx,x: f"({str(x.dtype)[7:]})({ctx[x.src[0]]})"),
|
||||
(UPat(Ops.BIND, name="x"), lambda ctx,x: ctx[x.src[0]]),
|
||||
(UPat(Ops.NEG, name="x"), lambda ctx,x: f"(-{ctx[x.src[0]]})"),
|
||||
(UPat(Ops.RECIPROCAL, name="x"), lambda ctx,x: f"(1/{ctx[x.src[0]]})"),
|
||||
(UPat(Ops.MAX, name="x"), lambda ctx,x: f"max({ctx[x.src[0]]}, {ctx[x.src[1]]})"),
|
||||
(UPat(Ops.MULACC, name="x"), lambda ctx,x: f"({ctx[x.src[0]]}*{ctx[x.src[1]]}+{ctx[x.src[2]]})"),
|
||||
(UPat(Ops.WHERE, name="x"), lambda ctx,x: f"({ctx[x.src[1]]} if {ctx[x.src[0]]} else {ctx[x.src[2]]})"),
|
||||
(UPat(Ops.CDIV, name="x"), lambda ctx,x: f"cdiv({ctx[x.src[0]]}, {ctx[x.src[1]]})"),
|
||||
(UPat(Ops.CMOD, name="x"), lambda ctx,x: f"cmod({ctx[x.src[0]]}, {ctx[x.src[1]]})"),
|
||||
(UPat(GroupOp.Movement, name="x"), lambda ctx,x: f"{ctx[x.src[0]]}.{x.op.name.lower()}({render_marg(ctx, x)})"),
|
||||
(UPat(set(syms.keys()), name="x"), lambda ctx,x: strip_binary_parens(x, ctx[x.src[0]], ctx[x.src[1]], lambda a,b: f"({a}{syms[x.op]}{b})")),
|
||||
(UPat((Ops.INDEX, Ops.STAGE), name="x"), lambda x, ctx: ''.join([f"[{strip_parens(ctx[y])}]" for y in x.src[1:]])),
|
||||
(UPat(Ops.STACK, name="x"), lambda ctx,x: f"{{{','.join([ctx[y] for y in x.src])}}}"),
|
||||
(UPat(GroupOp.All, name="x"), lambda x: str(x)),
|
||||
])
|
||||
|
||||
renderer_infer = PatternMatcher([
|
||||
(UPat(Ops.CMOD, name="x"), lambda ctx,x: f"cmod({ctx[x.src[0]]}, {ctx[x.src[1]]})"),
|
||||
(UPat(Ops.CDIV, name="x"), lambda ctx,x: f"cdiv({ctx[x.src[0]]}, {ctx[x.src[1]]})"),
|
||||
(UPat(Ops.FLOORMOD, name="x"), lambda ctx,x: f"floormod({ctx[x.src[0]]}, {ctx[x.src[1]]})"),
|
||||
(UPat(Ops.FLOORDIV, name="x"), lambda ctx,x: f"floordiv({ctx[x.src[0]]}, {ctx[x.src[1]]})"),
|
||||
(UPat(Ops.CAST, name="x"),
|
||||
lambda ctx,x: f"{'float' if dtypes.is_float(x.dtype) else 'bool' if x.dtype is dtypes.bool else 'int'}({ctx[x.src[0]]})"),
|
||||
(UPat(Ops.BITCAST, name="x"), lambda ctx,x: f"bitcast({ctx[x.src[0]]}, {x.src[0].dtype!r}, {x.dtype!r})"),
|
||||
]) + renderer
|
||||
|
||||
# *** pyrender ***
|
||||
|
||||
def srcs(ctx, src): return f"({ctx[src[0]]},)" if len(src) == 1 else f"({', '.join([ctx[x] for x in src])})"
|
||||
def render_marg(ctx,x:UOp):
|
||||
if x.op is Ops.PERMUTE: return str(x.marg)
|
||||
if x.op is Ops.FLIP: return str(tuple([i for i,x in enumerate(x.marg) if x]))
|
||||
pieces = []
|
||||
if x.op in {Ops.RESHAPE, Ops.EXPAND}:
|
||||
pieces = [f"{ctx[a] if isinstance(a, UOp) else str(a)}" for a in x.marg]
|
||||
if x.op in {Ops.PAD, Ops.SHRINK}:
|
||||
pieces = [f"({ctx[a[0]] if isinstance(a[0], UOp) else str(a[0])}, {ctx[a[1]] if isinstance(a[1], UOp) else str(a[1])})" for a in x.marg]
|
||||
return f"({','.join(pieces)})" if len(pieces) != 1 else f"({pieces[0]},)"
|
||||
|
||||
sugar = {Ops.SINK, Ops.END, Ops.STORE, Ops.LOAD, Ops.SQRT, Ops.INDEX, Ops.REDUCE, Ops.AFTER, Ops.THREEFRY,
|
||||
Ops.RECIPROCAL, Ops.EXP2, Ops.LOG2, Ops.SIN, Ops.CONTIGUOUS, Ops.BARRIER, Ops.DETACH}
|
||||
pm_pyrender_extra = PatternMatcher([
|
||||
(UPat(Ops.CONST, src=(), name="x"), lambda x: f"UOp.const({x.val}, {x.dtype})"),
|
||||
(UPat((Ops.CAST, Ops.BITCAST), name="x"), lambda ctx,x: f"{ctx[x.src[0]]}.{x.op.name.lower()}({x.dtype})"),
|
||||
(UPat(Ops.SPECIAL, src=(UPat(Ops.CONST),), name="x"), lambda x: f"UOp.special({x.src[0].val}, {repr(x.arg)}, dtype={x.dtype})"),
|
||||
(UPat(Ops.BUFFER, src=(UPat(),), name="x"), lambda x:
|
||||
f"UOp.new_buffer({repr(x.arg.device)}, {x.max_numel()}, {x.dtype}, {x.arg.slot})"
|
||||
if isinstance(x.arg, ParamArg) and x.addrspace is AddrSpace.GLOBAL else None),
|
||||
(UPat(Ops.COPY, src=(UPat(name="x"),), name="copy"), lambda ctx,x,copy: f"{ctx[x]}.copy_to_device({repr(copy.arg)})"),
|
||||
(UPat(Ops.CUSTOM_FUNCTION, name="x"), lambda ctx,x: f"UOp(Ops.CUSTOM_FUNCTION, src={srcs(ctx, x.src)}, arg={x.arg!r})"),
|
||||
(UPat(Ops.REDUCE, name="r"), lambda ctx,r: f"{ctx[r.src[0]]}._rop({r.arg[0]}, {tuple(range(r.arg[1]))})" if r.arg[1] else None),
|
||||
# NOTE: range has srcs sometimes after control flow
|
||||
(UPat(Ops.RANGE, src=(UPat(Ops.CONST, name="c"),), allow_any_len=True, name="x"), lambda ctx,x,c:
|
||||
"UOp.range("+', '.join([str(c.val)] + [repr(y) for y in x.arg])+
|
||||
(f', src={srcs(ctx, x.src[1:])}' if len(x.src) > 1 else '')+(', dtype='+str(x.dtype) if x.dtype is not dtypes.weakint else '')+")"),
|
||||
# TODO: index shouldn't mismatch dtype
|
||||
(UPat(Ops.INDEX, src=(UPat(), UPat()), allow_any_len=True, name="x"), lambda ctx,x:
|
||||
f"{ctx[x.src[0]]}.index({ctx[x.src[1]]}, "+''.join([f"{ctx[xx]}, " for xx in x.src[2:]])+
|
||||
f"dtype={x.dtype})" if x.src[0].dtype != x.dtype else None),
|
||||
# TODO: movement ops simplify stuff, this can break SPEC=2
|
||||
#(UPat(GroupOp.Movement, name="x"), lambda ctx,x: f"{ctx[x.src[0]]}.{x.op.name.lower()}({render_marg(ctx,x)})"),
|
||||
# NOTE: CMPNE doesn't work cause there's no __rne__
|
||||
# explicit trunc ops: `//` and `%` parse as FLOORDIV/FLOORMOD, so render CDIV/CMOD via .alu()
|
||||
(UPat(Ops.CDIV, name="x"), lambda ctx,x: f"{ctx[x.src[0]]}.alu(Ops.CDIV, {ctx[x.src[1]]})"),
|
||||
(UPat(Ops.CMOD, name="x"), lambda ctx,x: f"{ctx[x.src[0]]}.alu(Ops.CMOD, {ctx[x.src[1]]})"),
|
||||
# `.where` re-promotes its operands, so render WHERE via .alu() too
|
||||
(UPat(Ops.WHERE, name="x"), lambda ctx,x: f"{ctx[x.src[0]]}.alu(Ops.WHERE, {ctx[x.src[1]]}, {ctx[x.src[2]]})"),
|
||||
# the binary operators re-promote their operands (a weak src meeting a strong one gets a cast), render those via .alu() too
|
||||
(UPat(set(syms.keys())-{Ops.SUB, Ops.CDIV, Ops.CMOD}, name="x"), lambda ctx,x:
|
||||
strip_binary_parens(x, ctx[x.src[0]], ctx[x.src[1]], lambda a,b: f"({a}{syms[x.op]}{b})")
|
||||
if x.src[0]._broadcasted(x.src[1]) == x.src else f"{ctx[x.src[0]]}.alu({x.op}, {ctx[x.src[1]]})"),
|
||||
(UPat(sugar, src=(), name="x"), lambda x: f"UOp.{x.op.name.lower()}("+', '.join(([f'arg={repr(x.arg)}'] if x.arg is not None else []))+")"),
|
||||
(UPat(sugar, name="x"), lambda ctx,x: f"{ctx[x.src[0]]}.{x.op.name.lower()}("+', '.join([ctx[y] for y in x.src[1:]] + \
|
||||
([f'arg={repr(x.arg)}'] if x.arg is not None else []))+")"),
|
||||
])
|
||||
|
||||
# NOTE: you can remove pm_pyrender_extra and it'll still be correct
|
||||
pm_pyrender = pm_pyrender_extra+PatternMatcher([
|
||||
(UPat(GroupOp.All, name="u"), lambda ctx,u: f"UOp({u.op}, {u.dtype}, {srcs(ctx,u.src)}"+(f", {repr(u.arg)})" if u.arg is not None else ")")),
|
||||
])
|
||||
|
||||
def _render_with_splits(lst:list[UOp], pm:PatternMatcher, to_render:set[UOp], split_depth:int=100) -> dict[str, str]:
|
||||
r: dict[UOp, str] = {}
|
||||
ret: dict[str, str] = {}
|
||||
depth: dict[UOp, int] = {}
|
||||
for i,u in enumerate(lst):
|
||||
# limit inline depth to avoid "too many nested parentheses" in Python parser
|
||||
op_depth = 1 + max([depth.get(s, 0) for s in u.src], default=0)
|
||||
if op_depth > split_depth: to_render.add(u)
|
||||
depth[u] = 0 if u in to_render else op_depth
|
||||
ren = pm.rewrite(u, ctx=r)
|
||||
assert isinstance(ren, str)
|
||||
if u.tag is not None: ren += f".rtag({repr(u.tag)})"
|
||||
if u not in to_render: r[u] = ren
|
||||
else:
|
||||
r[u] = f"c{i}" if u is not lst[-1] else "ast"
|
||||
ret[r[u]] = ren
|
||||
return ret
|
||||
|
||||
def pyrender(ast:UOp) -> str:
|
||||
lst = list(ast.toposort())
|
||||
|
||||
cmap = consumer_map_from_toposort(lst)
|
||||
not_rendered = {Ops.CONST}
|
||||
always_rendered = {Ops.PARAM, Ops.LOAD, Ops.SPECIAL, Ops.RANGE, Ops.CONTIGUOUS, Ops.STACK,
|
||||
Ops.BUFFER, Ops.COPY, Ops.CALL, Ops.FUNCTION, Ops.WHERE, Ops.END}
|
||||
|
||||
to_render: set[UOp] = {ast}
|
||||
for u in lst:
|
||||
if u.op in {Ops.SINK}:
|
||||
for s in u.src: to_render.add(s)
|
||||
if u.op is Ops.STORE: to_render.add(u.src[1])
|
||||
if u.op is Ops.REDUCE: to_render.add(u.src[0])
|
||||
if u.op is Ops.FUNCTION or (u.op is Ops.CALL and u.src[0].dtype is dtypes.void): raise NotImplementedError("call can't be pyrendered")
|
||||
if u.op in not_rendered: continue
|
||||
# checking the consumers is not enough, you have to make sure it's not used twice by the one consumer
|
||||
if len(cmap[u]) == 1 and len([x for x in list(cmap[u].keys())[0].src if x is u]) == 1 and u.op not in always_rendered: continue
|
||||
to_render.add(u)
|
||||
|
||||
ret = _render_with_splits(lst, pm_pyrender, to_render)
|
||||
return '\n'.join([f"{k} = {strip_parens(v)}" for k,v in ret.items()])
|
||||
278
artifacts/package_runtime/tinygrad/uop/spec.py
Normal file
278
artifacts/package_runtime/tinygrad/uop/spec.py
Normal file
@@ -0,0 +1,278 @@
|
||||
import math
|
||||
from typing import Any
|
||||
from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, AxisType, KernelInfo, ParamArg
|
||||
from tinygrad.uop.render import print_uops, pyrender
|
||||
from tinygrad.dtype import DType, dtypes, AddrSpace, Invalid, ConstFloat
|
||||
from tinygrad.helpers import DEBUG, Context, SPEC, Metadata, panic, CHECK_OOB, all_same, is_image_shape
|
||||
|
||||
# ***** uop helpers *****
|
||||
|
||||
def validate_index(uidx:UOp, gate:UOp|None=None):
|
||||
if len(uidx.src) != 2: return True # skip for non final index. TODO: check more complex index with shape
|
||||
buf,idx = uidx.src
|
||||
if idx.is_invalid: return True
|
||||
if gate is None: gate = UOp.const(True)
|
||||
# TODO: check for overflow
|
||||
if not CHECK_OOB or is_image_shape(buf._shape): return True
|
||||
|
||||
# buffer size
|
||||
sz = buf.max_numel()
|
||||
|
||||
# We can use UOp min/max to do a faster check, but it can give false positive since its not an exact bound and doesn't consider the mask
|
||||
if 0<=idx.vmin and idx.vmax<sz: return True
|
||||
|
||||
# TODO: validate these
|
||||
# WEBGPU has a BITCAST in the index, PTX casts pointer to long
|
||||
# VECTORIZE can't be properly modeled in z3 since it doesn't support vectors
|
||||
# don't descend into PARAM shape metadata; only the PARAM value participates in index arithmetic
|
||||
for x in idx.toposort(gate=lambda x: x.op is not Ops.PARAM) | gate.toposort(gate=lambda x: x.op is not Ops.PARAM):
|
||||
if x.op in {Ops.BITCAST, Ops.STACK}: return True
|
||||
|
||||
# if all is good and CHECK_OOB=1, validate with z3
|
||||
from tinygrad.uop.validate import validate_index_with_z3
|
||||
return validate_index_with_z3(sz, idx, gate)
|
||||
|
||||
def type_verify(ast:UOp|list[UOp], check_spec:PatternMatcher):
|
||||
lst = list(ast.toposort()) if isinstance(ast, UOp) else ast
|
||||
if SPEC > 1: test_pyrender(lst[-1]) # assume this is the sink
|
||||
|
||||
with Context(TRACK_MATCH_STATS=0):
|
||||
for i,u in enumerate(lst):
|
||||
ret: bool|None = check_spec.rewrite(u)
|
||||
if ret is not True:
|
||||
if DEBUG >= 3: print_uops(lst)
|
||||
raise RuntimeError(f"UOp verification failed at {i} on {u.op} {u.dtype} {len(u.src)} {[(x.op, x.dtype, x.arg) for x in u.src]} {u.arg}")
|
||||
|
||||
# ***** new specs *****
|
||||
def matches_dtype(x:UOp, dtype:DType) -> bool: return x.dtype == dtype or x.base.is_invalid # Invalid matches any dtype
|
||||
# these ops can be used in the tensor graph and programs
|
||||
spec_shared = PatternMatcher([
|
||||
# NOTE: for testing, we let sinks be anything
|
||||
(UPat(Ops.SINK, dtypes.void), lambda: True),
|
||||
|
||||
# NOOP. TODO: remove this
|
||||
(UPat(Ops.NOOP), lambda: True),
|
||||
|
||||
# CONST is everywhere; Invalid is a bool const
|
||||
(UPat(Ops.CONST, src=(), name="x"), lambda x: x.dtype is dtypes.bool if x.is_invalid else type(x.val) is type(x.dtype.const(x.val))),
|
||||
|
||||
# STACK is everywhere too
|
||||
(UPat(Ops.STACK, dtype=dtypes.void, src=()), lambda: True),
|
||||
(UPat(Ops.STACK, src=(UPat(),), allow_any_len=True, name="s"),
|
||||
lambda s: all_same([x.shape for x in s.src]) and all(matches_dtype(x, s.dtype) or x.dtype in dtypes.weaks for x in s.src)),
|
||||
|
||||
# ALUs: operands match the result dtype, except comparisons/WHERE; renderer-lowered shifts may use a uint32 count
|
||||
# a weak dtype matches any dtype until lowering commits its operand
|
||||
(UPat(Ops.WHERE, name="w", src=(UPat(dtype=dtypes.bool), UPat(), UPat())),
|
||||
lambda w: all(matches_dtype(s, w.dtype) or s.dtype in dtypes.weaks for s in w.src[1:])),
|
||||
(UPat(GroupOp.Comparison, dtype=dtypes.bool, src=(UPat.var("x"), UPat.var("y"))),
|
||||
lambda x,y: matches_dtype(x, y.dtype) or matches_dtype(y, x.dtype) or x.dtype in dtypes.weaks or y.dtype in dtypes.weaks),
|
||||
(UPat((Ops.AND, Ops.OR, Ops.XOR, Ops.SHL, Ops.SHR), name="x"), lambda x: False if any(dtypes.is_float(s.dtype) for s in x.src) else None),
|
||||
(UPat((Ops.SHL, Ops.SHR), src=(UPat.var("x"), UPat(dtype=dtypes.uint)), name="a"), lambda a,x: matches_dtype(x, a.dtype) or None),
|
||||
(UPat((Ops.CDIV, Ops.CMOD, Ops.FLOORDIV, Ops.FLOORMOD), name="x"), lambda x: None if dtypes.is_int(x.dtype) else False),
|
||||
(UPat(GroupOp.ALU, name="x"), lambda x: all(matches_dtype(y, x.dtype) or y.dtype in dtypes.weaks for y in x.src)),
|
||||
|
||||
# CAST
|
||||
(UPat((Ops.BITCAST, Ops.CAST), src=(UPat(),), name="x"), lambda x: isinstance(x.arg, DType)),
|
||||
|
||||
# RANGE can be in the big graph now. a void RANGE is a bound-less loop header, the arg is an axis id like RANGE
|
||||
(UPat(Ops.RANGE, src=(UPat.var("x"),), allow_any_len=True, name="rng"), lambda rng,x:
|
||||
matches_dtype(x, rng.dtype) and isinstance(rng.arg, tuple) and len(rng.arg) >= 2 and \
|
||||
all(isinstance(ra, int) for ra in rng.arg[0:-1]) and isinstance(rng.arg[-1], AxisType)),
|
||||
(UPat(Ops.INDEX, name="x"), lambda x: len(x.src)>0 and all(dtypes.is_int(y.dtype) or y.base.is_invalid for y in x.src[1:]) or None),
|
||||
# END closes RANGEs
|
||||
(UPat(Ops.END, src=(UPat(),), allow_any_len=True, name="x"), lambda x: all(u.op is Ops.RANGE for u in x.src[1:]) or None),
|
||||
# a loop-ended END requires a trailing bool condition for the backedge (loop again while true)
|
||||
(UPat(Ops.END, src=(UPat(), UPat(Ops.RANGE, dtypes.void), UPat(dtype=dtypes.bool))), lambda: True),
|
||||
|
||||
# PARAM
|
||||
(UPat(Ops.PARAM, name="x"), lambda x: isinstance(x.arg, ParamArg)),
|
||||
(UPat(Ops.BUFFER, src=(UPat(),), name="x"), lambda x:
|
||||
isinstance(x.arg, ParamArg) and x.addrspace in (AddrSpace.REG, AddrSpace.LOCAL)),
|
||||
|
||||
# GROUP of stores (or groups, or NOOPs)
|
||||
(UPat(Ops.GROUP, dtypes.void, src=UPat((Ops.GROUP, Ops.STORE, Ops.NOOP, Ops.INS, Ops.END))), lambda: True),
|
||||
|
||||
# AFTER on Movement Op, PARAM, BUFFER, CONTIGUOUS, or another AFTER
|
||||
(UPat(Ops.AFTER, src=(UPat(GroupOp.Movement.union({Ops.PARAM, Ops.BUFFER, Ops.CONTIGUOUS, Ops.INDEX,
|
||||
Ops.AFTER, Ops.UNSHARD, Ops.BITCAST, Ops.INS})),),
|
||||
allow_any_len=True, name="x"), lambda x: matches_dtype(x.src[0], x.dtype)),
|
||||
|
||||
# CUSTOM (inline and non inline)
|
||||
(UPat((Ops.CUSTOMI, Ops.CUSTOM)), lambda: True),
|
||||
|
||||
# CALL of an external function
|
||||
(UPat(Ops.CALL, src=(UPat(),), allow_any_len=True, name="x"),
|
||||
lambda x: matches_dtype(x.src[0], dtypes.uint64) if x.src[0].dtype is not dtypes.void else None),
|
||||
|
||||
# pattern compiler IR ops (not in tensor/program graphs, but spec-compliant)
|
||||
(UPat(Ops.PYLITERAL), lambda: True),
|
||||
|
||||
# BARRIER (on any length). TODO: this should only be in spec_program
|
||||
(UPat(Ops.BARRIER, dtypes.void), lambda: True),
|
||||
|
||||
# assembly instruction
|
||||
(UPat(Ops.INS), lambda: True),
|
||||
|
||||
# LOAD(idx) / STORE(idx, val) with gates on the LOAD/STORE
|
||||
(UPat((Ops.INDEX, Ops.SHRINK), name="uidx").or_casted().load(), validate_index),
|
||||
(UPat((Ops.INDEX, Ops.SHRINK), name="uidx").or_casted().load(UPat.var("alt"), UPat.var("gate", dtype=dtypes.bool), name="load"),
|
||||
lambda uidx,gate,alt,load: validate_index(uidx, gate) if matches_dtype(alt, load.dtype) else False),
|
||||
(UPat((Ops.INDEX, Ops.SHRINK), name="uidx").or_casted().store(UPat()), validate_index),
|
||||
(UPat((Ops.INDEX, Ops.SHRINK), name="uidx").or_casted().store(UPat(), UPat.var("gate", dtype=dtypes.bool)), validate_index),
|
||||
|
||||
# STORE in tensor graph: store a value into a target
|
||||
(UPat(Ops.STORE, dtypes.void, (UPat(name="x"), UPat())), lambda x: True),
|
||||
|
||||
# WMMA has a <a, b, acc>
|
||||
(UPat(Ops.WMMA, src=(UPat(), UPat(), UPat()), name="x"), lambda x: isinstance(x.arg, tuple) and len(x.arg) == 5),
|
||||
])
|
||||
|
||||
def is_device(d): return isinstance(d, str) or (isinstance(d, tuple) and all(isinstance(s, str) for s in d))
|
||||
|
||||
def valid_gettuple(g:UOp, t:UOp): return isinstance(g.arg, int) and 0 <= g.arg < len(t.src) and matches_dtype(t.src[g.arg], g.dtype)
|
||||
|
||||
# these ops can exist in tensor but not programs. example: movement
|
||||
spec_tensor = PatternMatcher([
|
||||
(UPat((Ops.SIN, Ops.LOG2, Ops.EXP2, Ops.SQRT, Ops.RECIPROCAL), src=(UPat(),), name="u"), lambda u: dtypes.is_float(u.dtype)),
|
||||
|
||||
# BUFFER
|
||||
(UPat(Ops.BUFFER, src=(UPat(),), name="buf"), lambda buf:
|
||||
(isinstance(buf.dtype, DType) and matches_dtype(buf.src[0], dtypes.weakint) and is_device(buf.arg.device))
|
||||
if isinstance(buf.arg, ParamArg) and buf.addrspace is AddrSpace.GLOBAL else None),
|
||||
|
||||
# Tensor variable bindings
|
||||
(UPat(Ops.BIND, (dtypes.int, dtypes.long, dtypes.weakint,), (UPat(Ops.PARAM), UPat.cvar(dtype=(dtypes.int,dtypes.long,dtypes.weakint,))), arg=None),
|
||||
lambda: True),
|
||||
|
||||
# custom function
|
||||
(UPat(Ops.CUSTOM_FUNCTION, name="x"), lambda x: isinstance(x.arg, str)),
|
||||
|
||||
# CALL
|
||||
(UPat(Ops.CALL, dtypes.void, src=(UPat((Ops.SINK, Ops.LINEAR, Ops.PROGRAM, Ops.COPY, Ops.CUSTOM_FUNCTION)),), allow_any_len=True), lambda: True),
|
||||
|
||||
# FUNCTION + TUPLE must have void dtype, GETTUPLE can only appear on FUNCTION or TUPLE
|
||||
(UPat(Ops.FUNCTION, dtypes.void, src=(UPat(Ops.TUPLE),), allow_any_len=True), lambda: True),
|
||||
(UPat(Ops.TUPLE, dtypes.void), lambda: True),
|
||||
(UPat(Ops.GETTUPLE, src=(UPat(Ops.FUNCTION, src=(UPat(Ops.TUPLE, name="t"),), allow_any_len=True),), name="g"), valid_gettuple),
|
||||
(UPat(Ops.GETTUPLE, src=(UPat(Ops.TUPLE, name="t"),), name="g"), valid_gettuple),
|
||||
|
||||
# SPECIAL is index before index lowering. custom_kernel currently has this
|
||||
(UPat(Ops.SPECIAL, src=(UPat.var("x", dtypes.weakint),), name="s"), lambda s,x: matches_dtype(x, s.dtype) and isinstance(s.arg, str)),
|
||||
|
||||
# movement ops
|
||||
(UPat((Ops.RESHAPE, Ops.EXPAND), src=(UPat(), UPat())), lambda: True),
|
||||
(UPat((Ops.PAD, Ops.SHRINK), src=(UPat(), UPat(), UPat()), name="x"), lambda x: x.src[1].shape == x.src[2].shape),
|
||||
(UPat((Ops.PERMUTE, Ops.FLIP), name="mv", src=(UPat(),)), lambda mv: isinstance(mv.arg, tuple)),
|
||||
|
||||
# REDUCE has arg=(op, num_axes), src[1:] are ranges after lowering
|
||||
(UPat(Ops.REDUCE, src=(UPat(),), allow_any_len=True, name="x"),
|
||||
lambda x: isinstance(x.arg, tuple) and len(x.arg) == 2 and x.arg[0] in GroupOp.Reduce
|
||||
and isinstance(x.arg[1], int) and all(y.dtype in (dtypes.weakint, dtypes.int) for y in x.src[1:])),
|
||||
|
||||
# COPY
|
||||
(UPat(Ops.COPY, name="copy", src=(UPat.var("x"),)), lambda copy,x: matches_dtype(x, copy.dtype) and is_device(copy.arg)),
|
||||
(UPat(Ops.ALLREDUCE, name="red", src=(UPat.var("x"),)), lambda red,x: matches_dtype(x, red.dtype) and isinstance(red.arg, tuple) and
|
||||
len(red.arg) == 2 and red.arg[0] in GroupOp.Reduce and is_device(red.arg[1])),
|
||||
|
||||
# UNSHARD/MSELECT/MSTACK
|
||||
# an UNSHARD carries the value and one sharding range per sharded axis (usually a DEVICE RANGE, but can be a derived expression)
|
||||
(UPat(Ops.UNSHARD, name="multi"), lambda multi: len(multi.src) == 1+len(multi.arg) and matches_dtype(multi.src[0], multi.dtype)
|
||||
and all(isinstance(a, int) for a in multi.arg) and all(r.dtype in dtypes.weaks for r in multi.src[1:])),
|
||||
(UPat(Ops.MSELECT, name="x"), lambda x: isinstance(x.src[0].device, tuple) and x.arg < len(x.src[0].device)),
|
||||
(UPat(Ops.MSTACK, name="x"), lambda x: all(isinstance(s.device, str) for s in x.src) or (all_same(x.src) and x.src[0].device is None)),
|
||||
|
||||
# CONTIGUOUS ensures the source UOp realizes
|
||||
(UPat((Ops.DETACH, Ops.CONTIGUOUS, Ops.CONTIGUOUS_BACKWARD), name="root", src=(UPat.var("x"),), arg=None),
|
||||
lambda root,x: matches_dtype(x, root.dtype)),
|
||||
|
||||
# TODO: this should not be here. STAGE is transformed to BUFFER later
|
||||
(UPat(Ops.STAGE, src=(UPat(),), allow_any_len=True), lambda: True),
|
||||
|
||||
# codegen: PROGRAM with progressive sources through the pipeline (SINK, LINEAR?, SOURCE?, BINARY?)
|
||||
(UPat(Ops.LINEAR, dtypes.void), lambda: True),
|
||||
(UPat(Ops.SOURCE, dtypes.void, src=()), lambda: True),
|
||||
(UPat(Ops.BINARY, dtypes.uint8, src=(), name="x"), lambda x: isinstance(x.arg, bytes)),
|
||||
(UPat(Ops.PROGRAM, dtypes.void, src=(UPat(Ops.SINK),)), lambda: True),
|
||||
(UPat(Ops.PROGRAM, dtypes.void, src=(UPat(Ops.SINK), UPat(Ops.LINEAR))), lambda: True),
|
||||
(UPat(Ops.PROGRAM, dtypes.void, src=(UPat(Ops.SINK), UPat(Ops.LINEAR), UPat(Ops.SOURCE))), lambda: True),
|
||||
(UPat(Ops.PROGRAM, dtypes.void, src=(UPat(Ops.SINK), UPat(Ops.LINEAR), UPat(Ops.SOURCE), UPat(Ops.BINARY))), lambda: True),
|
||||
])+spec_shared
|
||||
|
||||
# these ops can exist in programs but not the tensor spec. example: LOAD
|
||||
spec_program = PatternMatcher([
|
||||
# index and weak dtypes are not allowed in programs
|
||||
(UPat(GroupOp.All, (dtypes.weakint, dtypes.weakfloat)), lambda: False),
|
||||
|
||||
# allow special SHRINK
|
||||
(UPat(Ops.SHRINK, src=(UPat((Ops.PARAM, Ops.BUFFER, Ops.AFTER)), UPat(), UPat(Ops.CONST))), lambda: True),
|
||||
|
||||
# movement ops are not allowed in programs
|
||||
(UPat(GroupOp.Movement), lambda: False),
|
||||
|
||||
# REG/LOCAL buffer
|
||||
(UPat(Ops.BUFFER, name="x"), lambda x: isinstance(x.arg, ParamArg) and x.addrspace in (AddrSpace.REG, AddrSpace.LOCAL)),
|
||||
|
||||
# Invalid is not allowed in program
|
||||
(UPat(Ops.CONST, arg=Invalid), lambda: False),
|
||||
|
||||
# if has a <gate, index_for_dedup>
|
||||
(UPat(Ops.IF, dtype=dtypes.void, src=(UPat(dtype=dtypes.bool), UPat((Ops.CAST, Ops.INDEX, Ops.SHRINK)))), lambda: True),
|
||||
(UPat(Ops.ENDIF, dtype=dtypes.void, src=(UPat(Ops.IF),)), lambda: True),
|
||||
|
||||
# SPECIAL is int32 after index lowering
|
||||
(UPat(Ops.SPECIAL, src=(UPat.var("x", dtypes.int32),), name="s"), lambda s,x: matches_dtype(x, s.dtype) and isinstance(s.arg, str)),
|
||||
])+spec_shared
|
||||
|
||||
spec_hcq = PatternMatcher([
|
||||
(UPat(Ops.GETADDR, dtypes.uint64, src=(UPat((Ops.BUFFER, Ops.PARAM)).or_after(),), name="x"), lambda x: is_device(x.arg)),
|
||||
(UPat(Ops.PROGRAM, dtypes.void, src=(UPat((Ops.BUFFER, Ops.PARAM)).or_after(),)), lambda: True),
|
||||
])+spec_shared
|
||||
|
||||
# these are intermediate ops. everything should be deleted from here
|
||||
spec_full = PatternMatcher([
|
||||
(UPat(Ops.REWRITE_ERROR, dtypes.void, name="x"), lambda x: isinstance(x.arg, str)),
|
||||
|
||||
# SLICE on BUFFER is allowed if BUFFER is
|
||||
(UPat(Ops.SLICE, src=(UPat(GroupOp.Movement.union({Ops.BUFFER, Ops.PARAM, Ops.STAGE, Ops.AFTER})),
|
||||
UPat(Ops.CONST, dtype=dtypes.weakint)), allow_any_len=True, name="bv"),
|
||||
lambda bv: isinstance(bv.arg, int)),
|
||||
|
||||
(UPat(Ops.CALL, dtypes.void, src=(UPat((Ops.SLICE,)),), allow_any_len=True), lambda: True),
|
||||
|
||||
# codegen may end ranges after gpudims has replaced RANGE with SPECIAL.
|
||||
(UPat(Ops.END, src=(UPat(), UPat()), allow_any_len=True), lambda: True),
|
||||
|
||||
# allow any AFTER
|
||||
(UPat(Ops.AFTER, src=(UPat(),), allow_any_len=True), lambda: True),
|
||||
|
||||
# all loads/stores
|
||||
(UPat((Ops.LOAD, Ops.STORE)), lambda: True),
|
||||
|
||||
# while BIND is being casted
|
||||
(UPat(Ops.BIND, (dtypes.int, dtypes.weakint), (UPat(), UPat()), arg=None), lambda: True),
|
||||
])+spec_tensor+spec_program+spec_hcq
|
||||
|
||||
# **** pyrender (move this) ****
|
||||
|
||||
# late imports to avoid circular import
|
||||
from tinygrad.codegen.opt import Opt, OptOps
|
||||
from tinygrad.schedule.rangeify import BufferizeOpts
|
||||
glbls:dict[str, Any] = {"inf": math.inf, "nan": math.nan, "KernelInfo": KernelInfo, "Metadata": Metadata,
|
||||
"UOp": UOp, "dtypes": dtypes, "Ops": Ops, "AxisType": AxisType, "Invalid": Invalid,
|
||||
"Opt": Opt, "OptOps": OptOps, "BufferizeOpts": BufferizeOpts, "AddrSpace": AddrSpace, "panic": panic,
|
||||
"ConstFloat": ConstFloat, "ParamArg": ParamArg}
|
||||
def eval_pyrender(code:str) -> UOp:
|
||||
lcls:dict[str, Any] = {}
|
||||
exec(code, glbls, lcls)
|
||||
return lcls['ast']
|
||||
|
||||
def test_pyrender(test_ast:UOp, assert_parents=True):
|
||||
try: code = pyrender(test_ast)
|
||||
except NotImplementedError: return None # this is okay, not all ops can be pyrendered
|
||||
ast:UOp = eval_pyrender(code)
|
||||
if ast is not test_ast:
|
||||
if assert_parents:
|
||||
for u in test_ast.toposort(): test_pyrender(u, assert_parents=False)
|
||||
raise RuntimeError(f"PYRENDER ISSUE:\nSTR MATCH: {str(test_ast) == str(ast)}\nUOP:\n{test_ast}\nPRODUCED:\n{ast}\nCODE:\n{code}")
|
||||
return code
|
||||
461
artifacts/package_runtime/tinygrad/uop/symbolic.py
Normal file
461
artifacts/package_runtime/tinygrad/uop/symbolic.py
Normal file
@@ -0,0 +1,461 @@
|
||||
# all of symbolic lives here now
|
||||
import math
|
||||
from collections import defaultdict
|
||||
from tinygrad.uop.ops import Ops, PatternMatcher, UPat, UOp, GroupOp, exec_alu
|
||||
from tinygrad.dtype import PyConst, ConstType, dtypes, can_lossless_cast, Invalid, bitcast
|
||||
from tinygrad.helpers import partition, all_same, prod, flatten, unwrap, IMAGE, dedup
|
||||
from tinygrad.uop.divandmod import div_and_mod_symbolic
|
||||
from tinygrad.uop.movement import mop_cleanup
|
||||
|
||||
# TODO: symbolic shouldn't be importing from codegen
|
||||
from tinygrad.codegen.decomp.transcendental import xpow
|
||||
|
||||
# ******** phase 1 of symbolic used to live in ops, it's the most generic folding rules ********
|
||||
|
||||
def simplify_pow(x:UOp, c:UOp) -> UOp|None:
|
||||
if c.val < 0: return x.reciprocal().pow(-c.val)
|
||||
if c.val == 0: return x.const_like(1)
|
||||
if int(c.val-0.5)+0.5 == c.val: return x.pow(c.val-0.5) * x.sqrt()
|
||||
if int(c.val) == c.val: return (y := x.pow(c.val//2)) * y * (x if c.val%2 == 1 else 1)
|
||||
return None
|
||||
|
||||
def fold_bitcast(root:UOp, c:UOp) -> UOp|None:
|
||||
if c.dtype.fmt is None or root.dtype.fmt is None or c.dtype.itemsize != root.dtype.itemsize: return None
|
||||
return root.const_like(bitcast(c.val, c.dtype, root.dtype))
|
||||
|
||||
def const_arg(u:UOp) -> ConstType|tuple[ConstType, ...]|None:
|
||||
if u.op is Ops.CONST: return u.val
|
||||
if u.op is Ops.STACK and all(s.op is Ops.CONST for s in u.src): return tuple(s.val for s in u.src)
|
||||
return None
|
||||
|
||||
def fold_const_alu(a:UOp) -> UOp|None:
|
||||
vals = [const_arg(s) for s in a.src]
|
||||
return None if any(v is None for v in vals) else a.const_like(exec_alu(a.op, a.dtype, vals, False))
|
||||
|
||||
def _quotient_base(q:UOp, base:UOp, div:int) -> UOp|None:
|
||||
# the B with q == B//div and B%div == base%div, or None. only such congruence is needed to recombine, and canonicalization
|
||||
# moves consts freely: the quotient may be merged ((x//c + a)//div -> (x + a*c)//(c*div) for div>0) and shifted ((y + k*D)//D == y//D + k)
|
||||
(q, s), (num, a) = q.pop_const(), base.pop_const()
|
||||
if q.op is not Ops.FLOORDIV or q.src[1].op is not Ops.CONST: return None
|
||||
if div > 0 and num.op is Ops.FLOORDIV and num.src[1].op is Ops.CONST and q.src[1].val == (c:=num.src[1].val)*div: num, a, D = num.src[0], a*c, c*div
|
||||
elif q.src[1].val == div: D = div
|
||||
else: return None
|
||||
(x, xa), (p, pa) = num.pop_const(), q.src[0].pop_const()
|
||||
if p is not x or (t:=xa + a - pa) % D: return None
|
||||
return base - k*div if (k:=t//D - s) else base
|
||||
|
||||
def fold_add_divmod_recombine(x:UOp) -> UOp|None:
|
||||
# a scaled mod (base%div)*mul recombines with a partner q*(div*mul) carrying the quotient of a b == base (mod div):
|
||||
# q == b//div -> b*mul (full recombine)
|
||||
# q == (b//div)%d -> (b%(div*d))*mul (partial recombine into a wider mod, needs d>0)
|
||||
terms = list(x.split_uop(Ops.ADD))
|
||||
for i,u in enumerate(terms):
|
||||
mod, mul = u.pop_const(Ops.MUL)
|
||||
if mod.op is not Ops.FLOORMOD or mod.src[1].op is not Ops.CONST: continue
|
||||
base, div = mod.src[0], mod.src[1].val
|
||||
for j,v in enumerate(terms):
|
||||
q, scale = v.pop_const(Ops.MUL)
|
||||
if i == j or scale != div*mul: continue
|
||||
rest = [t for k,t in enumerate(terms) if k not in (i,j)]
|
||||
if (b:=_quotient_base(q, base, div)) is not None: return (b*mul).usum(*rest)
|
||||
if q.op is Ops.FLOORMOD and q.src[1].op is Ops.CONST and (d:=q.src[1].val) > 0 and (b:=_quotient_base(q.src[0], base, div)) is not None:
|
||||
return ((b % (div*d))*mul).usum(*rest)
|
||||
return None
|
||||
|
||||
# Invalid poisons the value: ops move inside the gate so the Invalid reaches the LOAD/STORE and folds there.
|
||||
# this needs to be before symbolic so that 0*something_that_might_be_invalid doesnt become 0
|
||||
invalid_pat = UPat(Ops.CONST, arg=Invalid, name="i")
|
||||
invalid_gate = UPat.var("cond").where(UPat.var("x"), invalid_pat)
|
||||
pm_data_invalid = PatternMatcher([
|
||||
(invalid_pat.broadcast(), lambda i: i),
|
||||
(UPat(GroupOp.Unary|{Ops.CAST, Ops.BITCAST}, src=(invalid_pat,)), lambda i: i),
|
||||
(UPat(GroupOp.Unary|{Ops.CAST, Ops.BITCAST}, src=(invalid_gate,), name="op"),
|
||||
lambda cond,x,op,i: cond.where(op.replace(src=(x,)), i)),
|
||||
# binary ops move inside the gate, with Invalid in the false branch
|
||||
(UPat(GroupOp.Binary, src=(invalid_gate, UPat.var("y")), name="alu"), lambda cond,x,y,alu,i: cond.where(x.alu(alu.op,y), i)),
|
||||
(UPat(GroupOp.Binary, src=(UPat.var("y"), invalid_gate), name="alu"), lambda cond,x,y,alu,i: cond.where(y.alu(alu.op,x), i)),
|
||||
(UPat(GroupOp.Binary-GroupOp.Comparison, src=[invalid_pat, UPat()]), lambda i: i),
|
||||
# an Invalid condition poisons the whole where; a gated Invalid condition lifts the gate out
|
||||
(invalid_pat.where(UPat(), UPat()), lambda i: i),
|
||||
(invalid_gate.where(UPat.var("a"), UPat.var("b")), lambda cond,x,i,a,b: cond.where(x.where(a,b), i)),
|
||||
# normalize where(cond, Invalid, val) -> where(~cond, val, Invalid)
|
||||
(UPat.var("cond").where(invalid_pat, UPat.var("val")), lambda cond, i, val: cond.logical_not().where(val, i) if not val.is_invalid else i),
|
||||
# lift Invalid out: a.where(cond.where(x, Invalid), c) -> (~a|cond).where(a.where(x, c), Invalid)
|
||||
(UPat.var("a").where(invalid_gate, UPat.var("c")), lambda cond,i,x,a,c:
|
||||
(a.logical_not()|cond).where(a.where(x,c), i) if not c.is_invalid else None),
|
||||
(UPat.var("a").where(UPat.var("b"), invalid_gate), lambda cond,i,x,a,b: (a|cond).where(a.where(b, x), i) if not b.is_invalid else None),
|
||||
# fold gated LOAD/STORE
|
||||
(UPat(Ops.STORE, src=(UPat(Ops.INDEX, src=(UPat(), invalid_pat), allow_any_len=True).or_casted(), UPat())), lambda i: UOp(Ops.NOOP)),
|
||||
(UPat(Ops.LOAD, src=(UPat(Ops.INDEX, src=(UPat(), invalid_pat), allow_any_len=True).or_casted(),), allow_any_len=True, name="x"),
|
||||
lambda x,i: x.src[1] if len(x.src) > 1 else x.const_like(0)),
|
||||
])
|
||||
|
||||
pm_remove_invalid = PatternMatcher([
|
||||
(invalid_gate.named("w"), lambda cond,x,i,w: w.replace(src=(cond,x,w.const_like(0)))),
|
||||
(UPat(Ops.STACK, name="s"), lambda s: s.replace(src=tuple(UOp.const(0, s.dtype) if x.is_invalid else x for x in s.src))
|
||||
if any(x.is_invalid for x in s.src) else None),
|
||||
])
|
||||
|
||||
# the one rule that collapses the pair CAST(dt, CONST(v)) into a typed CONST
|
||||
# TODO: delete this once CONST has no dtype
|
||||
pm_fold_cast_const = PatternMatcher([(UPat(Ops.CAST, name="root", src=(UPat.cvar("c"),)), lambda root, c: root.const_like(c.val))])
|
||||
|
||||
symbolic_simple = pm_data_invalid + PatternMatcher([
|
||||
# ** self folding **
|
||||
(UPat.var("x") + 0, lambda x: x), # x+0 -> x
|
||||
(UPat.var("x") * 1, lambda x: x), # x*1 -> x
|
||||
(UPat.var("x", dtype=dtypes.ints+(dtypes.bool, dtypes.weakint)) ^ 0, lambda x: x), # x^0 -> x
|
||||
(UPat.var("x") // UPat.var("x"), lambda x: x.const_like(1)), # x//x -> 1
|
||||
(UPat.var("x") // 1, lambda x: x), # x//1 -> x
|
||||
(UPat.var("x") // -1, lambda x: -x), # x//-1 -> -x
|
||||
((UPat.var("x") ^ UPat.var("y")) ^ UPat.var("y"), lambda x,y: x), # (x^y)^y -> x
|
||||
((UPat.var() % UPat.var("y")).named("base") % UPat.var("y"), lambda base,y: base), # (x%y)%y = -> x%y (rewritten with base for speed)
|
||||
# variations of (x%c)+(x//c)*c = x
|
||||
(UPat(Ops.ADD, dtype=dtypes.weakint, name="x"), fold_add_divmod_recombine),
|
||||
(UPat.var("x", dtype=dtypes.bool) & UPat.cvar("c"), lambda x,c: x if c.val else c),
|
||||
(UPat.var("x", dtype=dtypes.bool) | UPat.cvar("c"), lambda x,c: c if c.val else x),
|
||||
(UPat.var("x", dtype=dtypes.bool) != UPat.const(False, dtypes.bool), lambda x: x), # x != False -> x
|
||||
(UPat(GroupOp.Idempotent, src=(UPat.var("x"), UPat.var("x"))), lambda x: x),
|
||||
(UPat.var("x", dtype=dtypes.bool).logical_not().logical_not(), lambda x: x),
|
||||
(UPat.var("x", dtype=dtypes.bool).where(UPat.const(True, dtypes.bool), UPat.const(False, dtypes.bool)), lambda x: x),
|
||||
(UPat.var("x", dtype=dtypes.bool).where(UPat.const(False, dtypes.bool), UPat.const(True, dtypes.bool)), lambda x: x.logical_not()),
|
||||
# CAST(bool -> int) != const — CAST(True)=1, CAST(False)=0, so fold based on const value
|
||||
(UPat.var("x", dtype=dtypes.bool).cast(dtypes.ints+(dtypes.weakint,)) != UPat.cvar("c"),
|
||||
lambda x,c: x if c.val == 0 else x.logical_not() if c.val == 1 else x.const_like(True)),
|
||||
(UPat.var("x", dtype=dtypes.ints+(dtypes.bool, dtypes.weakint)).trunc(), lambda x: x),
|
||||
# ** zero folding **
|
||||
(UPat.var("x") < UPat.var("x"), lambda x: x.const_like(False, dtypes.bool)), # x < x -> False
|
||||
(UPat.var("x") % UPat.var("x"), lambda x: x.const_like(0)), # x%x -> 0
|
||||
(UPat.var("x") ^ UPat.var("x"), lambda x: x.const_like(0)), # x^x -> 0
|
||||
(UPat.var("x") & 0, lambda x: x.const_like(0)), # x&0 -> 0
|
||||
# (x&mask)>>k -> x>>k when mask only clears bits below k
|
||||
((UPat.var("x") & UPat.cvar("mask")) >> UPat.cvar("k"),
|
||||
lambda x,mask,k: x >> k.val if mask.val | ((1 << k.val) - 1) == -1 else None),
|
||||
((UPat.var("x") & UPat.cvar("mask")) // UPat.cvar("c"),
|
||||
lambda x,mask,c: x // c.val if c.val > 0 and c.val & (c.val-1) == 0 and mask.val | (c.val-1) == -1 else None),
|
||||
(UPat.var("x", dtype=dtypes.ints+(dtypes.bool, dtypes.weakint)) != UPat.var("x"),
|
||||
lambda x: x.const_like(False, dtypes.bool)), # x != x -> False (only ints)
|
||||
# ** constant folding **
|
||||
(UPat(GroupOp.Unary, src=(UPat((Ops.CONST, Ops.STACK)),), name="a"), fold_const_alu),
|
||||
# NOTE: THREEFRY(const,const) folds via its decomposition
|
||||
(UPat(GroupOp.Binary-{Ops.THREEFRY}, src=(UPat((Ops.CONST, Ops.STACK)),)*2, name="a"), fold_const_alu),
|
||||
(UPat(GroupOp.Ternary, src=(UPat((Ops.CONST, Ops.STACK)),)*3, name="a"), fold_const_alu),
|
||||
# bool MUL is AND, ADD/MAX is OR. prevents other rules to rewrite bool ADD/MUL incorrectly
|
||||
(UPat.var('x', dtype=dtypes.bool) * UPat.var('y', dtype=dtypes.bool), lambda x,y: x&y),
|
||||
(UPat.var('x', dtype=dtypes.bool) + UPat.var('y', dtype=dtypes.bool), lambda x,y: x|y),
|
||||
(UPat.var('x', dtype=dtypes.bool).maximum(UPat.var('y', dtype=dtypes.bool)), lambda x,y: x|y),
|
||||
# *** div rules ***
|
||||
(UPat.cvar('x', arg=0) / 0, lambda x: x.const_like(float('nan'))), # 0/0 -> nan
|
||||
((UPat.var("x") * 0) / 0, lambda x: x.const_like(float('nan'))), # (x*0)/0 -> nan
|
||||
# can be wrong if x or x2 is 0
|
||||
(UPat.var("x") / UPat.var("x"), lambda x: x.const_like(1)), # x/x -> 1
|
||||
((UPat.var("x") * UPat.var("x2")) / UPat.var("x2"), lambda x,x2: x), # (x*x2)/x2 -> x
|
||||
# x*0 -> 0 or 0*x -> 0
|
||||
# if x is nan or inf it should render the nan value.
|
||||
# NOTE: this can be wrong for loaded NaN
|
||||
(UPat.var("x") * 0, lambda x: x.const_like(float("nan") if x.op is Ops.CONST
|
||||
and isinstance(x.val, float) and (math.isnan(x.val) or math.isinf(x.val)) else 0)),
|
||||
# *** cast/bitcast ***
|
||||
(UPat((Ops.CAST, Ops.BITCAST), name="root"), lambda root: root.src[0] if root.dtype == root.src[0].dtype else None),
|
||||
(UPat(Ops.BITCAST, name="root", src=(UPat.cvar("c"),)), fold_bitcast),
|
||||
# b.cast(a).cast(b) -> b if a preserves all values in b
|
||||
(UPat.var('x').cast(name="a").cast(name="b"), lambda x,a,b: x if x.dtype == b.dtype and can_lossless_cast(b.dtype, a.dtype) else None),
|
||||
# bitcast twice
|
||||
(UPat(Ops.BITCAST, name="b", src=(UPat.var('x').bitcast(),)), lambda x,b: x.bitcast(b.dtype)),
|
||||
(UPat.var("x").cast(dtypes.bool), lambda x: x != 0),
|
||||
# ** pow **
|
||||
(UPat.var("x").alu(Ops.POW, UPat.cvar("c")), simplify_pow),
|
||||
# positive const ** x
|
||||
(UPat.cvar("c").alu(Ops.POW, UPat.var("x")), lambda c,x: c if c.val == 1 else (x*math.log2(c.val)).exp2() if c.val > 0 else None),
|
||||
# unpack a uint64 packed from two uint32 (threefry)
|
||||
(((UPat.var(None, dtypes.uint64)<<32) | UPat.var('y', dtypes.uint32).cast(dtypes.uint64)).cast(dtypes.uint32), lambda y: y),
|
||||
(((UPat.var('x', dtypes.uint32).cast(dtypes.uint64)<<32) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))>>32,
|
||||
lambda x: x.cast(dtypes.uint64)),
|
||||
# ** simple where folding **
|
||||
# a conditional with the same results either way is a noop, also fold const conditionals
|
||||
(UPat.var().where(UPat.var("val"), UPat.var("val")), lambda val: val),
|
||||
(UPat.cvar("gate").where(UPat.var("c0"), UPat.var("c1")), lambda gate, c0, c1: c0 if gate.val else c1),
|
||||
# a.where(b.where(c, d), d) -> (a & b).where(c, d)
|
||||
(UPat.var("a").where(UPat.var("b").where(UPat.var("c"), UPat.var("d")), UPat.var("d")), lambda a,b,c,d: (a&b).where(c,d)),
|
||||
# a.where(c, b.where(c, d)) -> (a | b).where(c, d)
|
||||
(UPat.var("a").where(UPat.var("c"), UPat.var("b").where(UPat.var("c"), UPat.var("d"))), lambda a,b,c,d: (a|b).where(c,d)),
|
||||
])+mop_cleanup
|
||||
|
||||
# ******** phase 2 builds on phase 1, it includes the old "symbolic", rules that match deeper ********
|
||||
|
||||
def lt_folding(x:UOp, c:int) -> UOp|None:
|
||||
p, np = partition(x.split_uop(Ops.ADD), lambda u: u.const_factor() == 1)
|
||||
if np and (d:=math.gcd(*[u.const_factor() for u in np], c)) > 1 and 0 <= sum(u.vmin for u in p) and sum(u.vmax for u in p) < d:
|
||||
return unwrap(UOp.usum(*np).divides(d))<(c//d)
|
||||
return None
|
||||
|
||||
def canonicalize_simplex(X:UOp) -> UOp|None:
|
||||
# (X := a0*x0 + a1*x1 + ...) > 0 is equivalent to x0 + x1 + ... > 0 if xi >= 0 and ai > 0 for ints.
|
||||
# returns x0 + x1 + ... in such case, or None if not
|
||||
changed, ret = False, []
|
||||
for u in X.split_uop(Ops.ADD):
|
||||
# assumed the const is the last src of MUL
|
||||
if u.op is Ops.MUL and u.src[1].op is Ops.CONST and u.src[1].val > 0:
|
||||
changed = True
|
||||
u = u.src[0]
|
||||
if not (u.op in GroupOp.Irreducible and u.vmin >= 0): return None
|
||||
ret.append(u)
|
||||
return UOp.usum(*ret) if changed else None
|
||||
|
||||
commutative = PatternMatcher([
|
||||
# ** COMMUTATIVE flipping (only for index) **
|
||||
# NOTE: this can break merging vector math by only flipping some of them
|
||||
(UPat(GroupOp.Commutative, dtype=dtypes.weakint, name='x'), lambda x:
|
||||
x.replace(src=x.src[::-1]) if x.src[1].tuplize < x.src[0].tuplize and not x.src[0].tuplize < x.src[1].tuplize else None),
|
||||
])
|
||||
|
||||
def fold_where_closure(cond:UOp, t:UOp, f:UOp) -> UOp|None:
|
||||
"""in cond.where(t, f), cond is True within t and False within f"""
|
||||
if cond not in t.bool_slice and cond not in f.bool_slice: return None
|
||||
# INDEX gates are owned by the valid/store-coalescing machinery, leave them alone
|
||||
if any(u.op_in_backward_slice_with_self(Ops.INDEX) for u in (cond, t, f)): return None
|
||||
return cond.where(t.substitute({cond: cond.const_like(True)}), f.substitute({cond: cond.const_like(False)}))
|
||||
|
||||
symbolic = symbolic_simple+commutative+PatternMatcher([
|
||||
# ** boolean algebra **
|
||||
# TODO: make a more general or folder like simplify_valid
|
||||
(UPat.var("x", dtype=dtypes.bool) | UPat.var("x", dtype=dtypes.bool).logical_not(), lambda x: x.const_like(True)), # x|!x -> True
|
||||
# ** combine terms **
|
||||
(UPat.var("x") * UPat.cvar("c0") + UPat.var("x") * UPat.cvar("c1"), lambda x,c0,c1: x*(c0+c1)), # (x*c0)+(x*c1) -> x*(c0+c1)
|
||||
((UPat.var("y") + UPat.var("x") * UPat.cvar("c0")) + UPat.var("x") * UPat.cvar("c1"), lambda x,y,c0,c1: y+x*(c0+c1)),
|
||||
(UPat.var("x") + UPat.var("x") * UPat.cvar("c"), lambda x,c: x*(c+1)), # (x+x*c)-> x*(c+1)
|
||||
((UPat.var("y") + UPat.var("x")) + UPat.var("x") * UPat.cvar("c"), lambda x,y,c: y+x*(c+1)),
|
||||
((UPat.var("y") + UPat.var("x") * UPat.cvar("c")) + UPat.var("x"), lambda x,y,c: y+x*(c+1)),
|
||||
(UPat.var("x") + UPat.var("x"), lambda x: x*2), # (x+x)-> x*2
|
||||
((UPat.var("y") + UPat.var("x")) + UPat.var("x"), lambda y,x: y+x*2),
|
||||
((UPat.var("x") / UPat.var("x2")) / UPat.var("x3"), lambda x,x2,x3: x/(x2*x3) if x2 is not x3 else None), # (x/x2)/x3 -> x/(x2*x3)
|
||||
(-1 * (UPat.var("x") + UPat.cvar("c")), lambda x,c: (-x)+(-c)), # -(x+c) -> -x + -c
|
||||
(UPat.cvar("y") * (UPat.var("x", dtype=dtypes.weakint) + UPat.cvar("c")), lambda x,y,c: (y*x)+(y*c)), # y*(x+c) -> y*x + y*c
|
||||
# ** where folding **
|
||||
(UPat.var("cond", dtype=dtypes.bool).logical_not().where(UPat.var("t"), UPat.var("f")),
|
||||
lambda cond, t, f: cond.where(f,t) if not f.is_invalid else None),
|
||||
# in cond.where(t, f), uses of cond fold to True within t and False within f
|
||||
(UPat.var("cond", dtype=dtypes.bool).where(UPat.var("t"), UPat.var("f")), fold_where_closure),
|
||||
# alu of two where with same conds can combine, only do if true branch or false branch is const
|
||||
(UPat(GroupOp.Binary, name="alu", src=(UPat.var("c").where(UPat.var("t"), UPat.var("f")), UPat.var("c").where(UPat.var("tt"), UPat.var("ff")))), \
|
||||
lambda alu,c,t,tt,f,ff: c.where(t.alu(alu.op, tt), f.alu(alu.op, ff)) if t.op == tt.op == Ops.CONST or f.op == ff.op == Ops.CONST else None),
|
||||
# if its a plus we add the associative variation too
|
||||
((UPat.var("y")+UPat.var("c").where(UPat.var("t"), UPat.var("f"))) + UPat.var("c").where(UPat.var("tt"), UPat.var("ff")), \
|
||||
lambda y,c,t,tt,f,ff: y+c.where(t+tt, f+ff) if t.op == tt.op == Ops.CONST or f.op == ff.op == Ops.CONST else None),
|
||||
# complementary zero branches under the same condition select directly
|
||||
(UPat.var("c").where(UPat.var("t"), 0) + UPat.var("c").where(0, UPat.var("f")), lambda c,t,f: c.where(t, f)),
|
||||
# ALU/variable min==max -> CONST
|
||||
(UPat({Ops.CMPLT, Ops.CMPNE, Ops.FLOORDIV, Ops.FLOORMOD, Ops.PARAM, Ops.BIND, Ops.SPECIAL}, name="x"),
|
||||
lambda x: x.const_like(x.vmin) if x.vmin == x.vmax else None),
|
||||
(UPat(Ops.RANGE, src=(UPat(Ops.CONST,)), name="x"), lambda x: x.const_like(x.vmin) if x.vmin == x.vmax else None),
|
||||
# max folding
|
||||
(UPat.maximum(UPat.var("x"), UPat.var("y")), lambda x,y: x if x.vmin >= y.vmax else y if x.vmax <= y.vmin else None),
|
||||
# TODO: why does this rule break beautiful_mnist?
|
||||
#((UPat.var("x")+UPat.var("z")).maximum(UPat.var("y")+UPat.var("z")), lambda x,y,z: x.maximum(y) + z),
|
||||
# ** two stage ALU folding **
|
||||
*((UPat.var("x").alu(op, UPat.cvar("c1")).alu(op, UPat.cvar("c2")).named("f"),
|
||||
lambda f,x,c1,c2: x.alu(f.op,c1.alu(f.op,c2))) for op in GroupOp.Associative),
|
||||
((UPat.cvar("c0") + UPat.var("x")) < UPat.cvar("c1"), lambda x,c0,c1: x<(c1-c0)), # c0 + x < c1 -> x < c1 - c0
|
||||
# (x//c1)//c2 -> x//(c1*c2) for c2>0
|
||||
((UPat.var("x") // UPat.cvar("c1")) // UPat.cvar("c2"), lambda x,c1,c2: x//(c1*c2) if c2.vmin>0 else None),
|
||||
# ** lt **
|
||||
# c0*x<c1 -> sign(c0)*x < ceil(c1/abs(c0))
|
||||
((UPat.cvar("c0")*UPat.var("x", dtype=dtypes.weakint))<UPat.cvar("c1"),
|
||||
lambda x,c0,c1: (x if c0.val > 0 else -x)<-(-c1.val//abs(c0.val)) if abs(c0.val) > 1 else None),
|
||||
# x//d<c -> x<c*d for d>0, and -> c*d<x for d<0
|
||||
((UPat.var("x", dtype=dtypes.weakint)//UPat.cvar("d"))<UPat.cvar("c"),
|
||||
lambda x,d,c: (x<c.val*d.val) if d.val > 0 else (x>c.val*d.val) if d.val < 0 else None),
|
||||
# ** move add/mul consts to end (NOTE: this is still happening before constant folding) **
|
||||
((UPat.var("x") + UPat.cvar("c1")) + UPat.var("y"), lambda x,c1,y: (x+y)+c1 if y.op is not Ops.CONST else None),
|
||||
((UPat.var("x") * UPat.cvar("c1")) * UPat.var("y"), lambda x,c1,y: (x*y)*c1 if y.op is not Ops.CONST else None),
|
||||
# *** rules from symbolic ***
|
||||
# generic lt folding
|
||||
(UPat.var("x", dtypes.weakint)<UPat.cvar("c"), lambda x,c: lt_folding(x, c.val) if 0 < c.val else None),
|
||||
(UPat.var("x", dtypes.weakint)*-1 < UPat.var("y")*-1, lambda x,y: y<x),
|
||||
# canonicalize a simplex with positive coefficients > 0. NOTE: not x < 1 means x > 0
|
||||
((UPat.var("x", dtypes.weakint)<1).ne(True), lambda x: (newx<1).ne(True) if (newx:=canonicalize_simplex(x)) is not None else None),
|
||||
# a range mod its own upper bound is just the range
|
||||
(UPat(Ops.RANGE, src=UPat.var("end"), name="r")%UPat.var("end"), lambda r,end: r),
|
||||
(UPat(Ops.RANGE, src=UPat.var("end"), name="r")//UPat.var("end"), lambda r,end: r.const_like(0)),
|
||||
# cast/long folding
|
||||
# if the intermediate cast doesnt narrow we can do it in one cast
|
||||
(UPat.var('x').cast(name="a").cast(name="b"), lambda x,a,b: x.cast(b.dtype) if can_lossless_cast(x.dtype, a.dtype) else None),
|
||||
(UPat.var('x', dtypes.ints+(dtypes.weakint,)).cast(dtypes.ints+(dtypes.weakint,), name="a").cast(name="b"),
|
||||
lambda x,a,b: x.cast(b.dtype) if a.dtype.min<=x.vmin and x.vmax<=a.dtype.max else None),
|
||||
# try to do math in int instead of long, keep weak const weak
|
||||
(UPat(GroupOp.Binary, src=(UPat.var("x", dtypes.long), UPat.var("y", dtypes.long)), name="u"), lambda u,x,y:
|
||||
(UOp.const(x.val) if x.op is Ops.CONST else x.cast(dtypes.int)).alu(u.op,
|
||||
UOp.const(y.val) if y.op is Ops.CONST else y.cast(dtypes.int)).cast(u.dtype)
|
||||
if not any(v.overflows(dtypes.int) for v in (u,x,y)) else None),
|
||||
((UPat.var("x", dtypes.weakint) + UPat.cvar("c")).cast(dtypes.sints, name="cast"), lambda x,c,cast:x.cast(cast.dtype)+cast.const_like(c.val)),
|
||||
# only RANGE/IF/STORE/KERNEL have side effects
|
||||
(UPat(Ops.AFTER, name="x"), lambda x: x.replace(src=(x.src[0],)+
|
||||
tuple(dedup(flatten([(y,) if y.op in {Ops.RANGE, Ops.STORE, Ops.CALL, Ops.FUNCTION, Ops.BARRIER, Ops.END, Ops.LINEAR, Ops.STAGE}
|
||||
else y.src for y in x.src[1:]]))))),
|
||||
# after with 1 src is just src[0]
|
||||
(UPat(Ops.AFTER, src=(UPat.var("s"),)), lambda s: s),
|
||||
])+div_and_mod_symbolic
|
||||
|
||||
# ******** we take a small aside to "simplify_valid" to rewrite valids ********
|
||||
|
||||
def parse_valid(v:UOp) -> tuple[UOp, bool, int]|None:
|
||||
# if it's X <= c, returns X, True, c
|
||||
# if it's X >= c, returns X, False, c
|
||||
|
||||
if v.op is Ops.CMPNE and v.src[1].op is Ops.CONST and v.src[1].val == 1 and (s0:=v.src[0]).op is Ops.CMPLT and dtypes.is_int(s0.src[0].dtype):
|
||||
# (X < c).ne(True) -> X >= c
|
||||
return s0.src[0], False, int(s0.src[1].vmin)
|
||||
if v.op is Ops.CMPLT and dtypes.is_int(v.src[0].dtype):
|
||||
# c < X -> X >= c+1 (a const on the left is a lower bound on the right)
|
||||
if v.src[0].op is Ops.CONST: return v.src[1], False, int(v.src[0].val)+1
|
||||
# X < c -> X <= c-1
|
||||
return v.src[0], True, int((v.src[1]).vmax)-1
|
||||
return None
|
||||
|
||||
def uop_given_valid(valid:UOp, uop:UOp, try_simplex=True) -> UOp:
|
||||
# return simplified uop (might be the same as input)
|
||||
|
||||
# first, parse valid into {expr: (lower_bound, upper_bound)}
|
||||
bounds:defaultdict[UOp, list[PyConst|None]] = defaultdict(lambda: [None, None])
|
||||
for stmt in valid.split_uop(Ops.AND):
|
||||
if (res:=parse_valid(stmt)) is None: continue
|
||||
expr, is_upper, c = res
|
||||
bounds[expr][int(is_upper)] = c
|
||||
|
||||
# simplify uop given that valid is True
|
||||
all_candidates = []
|
||||
for i,(expr,v) in enumerate(bounds.items()):
|
||||
v0, v1 = (expr.vmin if v[0] is None else v[0], expr.vmax if v[1] is None else v[1])
|
||||
# try checking the whole clause
|
||||
all_candidates.append((expr, UOp.variable(f"fake{i}", v0, v1, expr.dtype)))
|
||||
|
||||
if try_simplex:
|
||||
# every candidate is a set of constrained UOp based on valid, and if every item in a set simplifies the uop into a same output, we rewrite uop
|
||||
candidates = [[all_candidates[-1]]]
|
||||
if expr.op is Ops.ADD and v0 == 1 and all(u.op in GroupOp.Irreducible for u in expr.split_uop(Ops.ADD)):
|
||||
# if the constraint is a simplex: X0 + X1 + ... > 0, we can check if all Xi > 0 simplify into the same output
|
||||
candidates.append([(Xi, UOp.variable(f"fake{i}", 1, Xi.vmax, Xi.dtype)) for Xi in expr.split_uop(Ops.ADD)])
|
||||
|
||||
for candidate in candidates:
|
||||
# if every branch in candidate gives the same simplified uop, we can rewrite the uop
|
||||
if any(X not in uop.backward_slice_with_self for X,_ in candidate): continue # skip if a branch var isn't in uop
|
||||
newuops = [uop.substitute({X:newX}).simplify().substitute({newX:X}).simplify() for X,newX in candidate]
|
||||
if all_same(newuops): uop = newuops[0]
|
||||
elif uop.op is Ops.STACK and len(uop.src) == 2:
|
||||
if all_same([uops.src[0] for uops in newuops]): uop = uop.replace(src=(newuops[0].src[0], uop.src[1]))
|
||||
if all_same([uops.src[1] for uops in newuops]): uop = uop.replace(src=(uop.src[0], newuops[0].src[1]))
|
||||
|
||||
# try all the valids together (but only the whole expressions)
|
||||
if (s_uop:=uop.substitute(sub_dict:=dict(all_candidates))) is not uop:
|
||||
uop = s_uop.simplify().substitute({newX:X for X,newX in sub_dict.items()}).simplify()
|
||||
return uop
|
||||
|
||||
def _valid_priority(v: UOp, valids:list[UOp]) -> int:
|
||||
# we want valid that's in other valids' parents to be first, so it's more likely the other valids get simplified
|
||||
return 0 if (res:=parse_valid(v)) is None else sum(-1 for other in valids if res[0] in other.backward_slice_with_self)
|
||||
|
||||
def simplify_valid(valid:UOp) -> UOp|None:
|
||||
if valid.op_in_backward_slice_with_self(Ops.INDEX): return None # this should only be for indexing, skip if there's a INDEX
|
||||
ret:list[UOp] = []
|
||||
valids = list(valid.split_uop(Ops.AND))
|
||||
valids = sorted(valids, key=lambda v: _valid_priority(v, valids))
|
||||
for stmt in dedup(valids):
|
||||
if ret: stmt = uop_given_valid(UOp.uprod(*ret), stmt)
|
||||
ret.append(stmt)
|
||||
return UOp.uprod(*ret) if ret != valids else None
|
||||
|
||||
# ******** phase 3 is the complete symbolic ********
|
||||
|
||||
def reduce_mul_chain(r:UOp) -> UOp|None:
|
||||
if r.arg[0] not in {Ops.ADD, Ops.MAX}: return None
|
||||
if r.dtype != r.src[0].dtype: return None
|
||||
inside, outside = [], []
|
||||
for m in r.src[0].split_uop(Ops.MUL):
|
||||
m_parents = m.backward_slice
|
||||
if m not in r.src[1:] and all(r not in m_parents for r in r.src[1:]) and (r.arg[0] != Ops.MAX or m.vmin >= 0): outside.append(m)
|
||||
else: inside.append(m)
|
||||
if len(outside) == 0: return None
|
||||
return r.replace(src=(prod(inside) if len(inside) else r.src[0].const_like(1),)+r.src[1:])*prod(outside)
|
||||
|
||||
def drop_and_clauses(cond:UOp, x:UOp, i:UOp) -> UOp|None:
|
||||
keep, drop = partition(cond.split_uop(Ops.AND), lambda c: any(r in x.ranges for r in c.ranges))
|
||||
return UOp.const(True).uprod(*keep).where(x, i) if drop else None
|
||||
pm_drop_and_clauses = PatternMatcher([(invalid_gate, drop_and_clauses)])
|
||||
|
||||
# move conditions from where to load's valid, drop clauses already in load
|
||||
def where_on_load(cond:UOp, buf:UOp, idx:UOp, or_cast:UOp) -> UOp|None:
|
||||
where_clauses, load_valid = list(cond.split_uop(Ops.AND)), idx.get_valid()
|
||||
in_load = set(load_valid.split_uop(Ops.AND))
|
||||
idx_index = {u for u in idx.backward_slice_with_self if u.op is Ops.INDEX}
|
||||
# can move if: condition's ranges are subset of idx's ranges, and no data dependent INDEX (only idx's INDEX allowed)
|
||||
def can_move(c:UOp) -> bool:
|
||||
return c.ranges.keys() <= idx.ranges.keys() and all(u in idx_index for u in c.backward_slice_with_self if u.op is Ops.INDEX)
|
||||
moved, keep = partition([c for c in where_clauses if c not in in_load], can_move)
|
||||
if len(keep) == len(where_clauses): return None
|
||||
idx = buf.index(idx.get_idx().valid(load_valid.uprod(*moved)))
|
||||
ret_idx = idx.cast(or_cast.dtype) if or_cast.op is Ops.CAST else idx
|
||||
return UOp.const(True).uprod(*keep).where(ret_idx, ret_idx.const_like(0))
|
||||
|
||||
# where after gated load becomes alt value, TODO: this is sort of duplicated with rules in devectorizer
|
||||
pm_move_where_on_load = PatternMatcher([
|
||||
(UPat.var("cond").where(UPat.var("buf").index(UPat.var("idx")).or_casted("or_cast"), 0), where_on_load),
|
||||
(UPat.var("cond").where(0, UPat.var("buf").index(UPat.var("idx")).or_casted("or_cast")),
|
||||
lambda cond,buf,idx,or_cast: where_on_load(cond.logical_not(),buf,idx,or_cast)),
|
||||
])
|
||||
|
||||
def gated_given_valid(cond:UOp, x:UOp, i:UOp) -> UOp|None:
|
||||
if x.dtype is not dtypes.weakint: return None
|
||||
# Skip if x contains DIV/MOD AND IMAGE mode is enabled -> image index e.g. openpilot
|
||||
if IMAGE.value > 0 and x.op_in_backward_slice_with_self(Ops.CDIV, Ops.CMOD, Ops.FLOORDIV, Ops.FLOORMOD): return None
|
||||
return cond.where(uop_given_valid(cond, x, try_simplex=False), i)
|
||||
|
||||
pm_simplify_valid = PatternMatcher([
|
||||
# simplify valid
|
||||
(UPat(Ops.AND, name="valid"), simplify_valid),
|
||||
(invalid_gate, gated_given_valid),
|
||||
])
|
||||
|
||||
# this is symbolic 2.0
|
||||
REMOVE_FROM_SINK_LIKE = {Ops.NOOP, Ops.STACK, Ops.SINK, Ops.GROUP}
|
||||
pm_clean_up_group_sink = PatternMatcher([
|
||||
# clean up GROUP/SINK
|
||||
(UPat(Ops.GROUP, src=(UPat.var("x"),)), lambda x: x),
|
||||
(UPat((Ops.SINK, Ops.GROUP), name="root"),
|
||||
lambda root: UOp(root.op, src=tuple(flatten(x.src if x.op in REMOVE_FROM_SINK_LIKE else (x,) for x in root.src)), arg=root.arg)
|
||||
if any(x.op in REMOVE_FROM_SINK_LIKE for x in root.src) else None),
|
||||
])
|
||||
|
||||
sym = symbolic+pm_simplify_valid+PatternMatcher([
|
||||
# reorder ALU/VECTORIZE
|
||||
(UPat(GroupOp.ALU, src=(UPat(Ops.STACK, src=UPat(name='x')), UPat(Ops.STACK, src=UPat(name='y'))), name='alu'),
|
||||
lambda x,y,alu: UOp(Ops.STACK, src=(UOp(alu.op, src=(x,y)),))),
|
||||
# ** where **
|
||||
# push cast to branches
|
||||
(UPat.var("s").where(UPat.var("a"), UPat.var("b")).cast().named("cast"), lambda s,a,b,cast: s.where(a.cast(cast.dtype), b.cast(cast.dtype))),
|
||||
# ** pow **
|
||||
((UPat(Ops.POW, name="p"), lambda p: xpow(*p.src))),
|
||||
# ** load/store folding **
|
||||
(UPat.store(UPat(Ops.INDEX, name="index"), UPat.load(UPat(Ops.INDEX, name="index"))), lambda index: UOp(Ops.NOOP)),
|
||||
(UPat.store(UPat(Ops.INDEX, name="index"), UPat.var("gate").where(UPat.var("alt"),
|
||||
UPat.load(UPat(Ops.INDEX, name="index")))),
|
||||
lambda index, gate, alt: UOp.store(index.src[0].index(index.src[1].valid(gate)), alt)),
|
||||
# fold gated LOAD/STORE
|
||||
(UPat(Ops.STORE, src=(UPat(), invalid_pat)), lambda i: UOp(Ops.NOOP)),
|
||||
# store of where with invalid -> gated store
|
||||
(UPat(Ops.STORE, src=(UPat(Ops.INDEX, name="index"), UPat.var("cond").where(UPat.var("val"), invalid_pat))),
|
||||
lambda index, cond, val, i: UOp.store(index.src[0].index(index.src[1].valid(cond)), val)),
|
||||
((UPat.var("x") * UPat.var("x")).reciprocal(), lambda x: x.reciprocal()*x.reciprocal()), # 1/(x^c) -> (1/x)^c
|
||||
((UPat.var("x") * UPat.var("x") * UPat.var("x")).reciprocal(), lambda x: x.reciprocal()*x.reciprocal()*x.reciprocal()),
|
||||
((UPat.var("x") * UPat.cvar("c")).reciprocal(), lambda x,c: x.reciprocal()*c.reciprocal()), # 1/(x*c) -> (1/c)*(1/x)
|
||||
(UPat.var("x") * ((1+UPat.var("x")).reciprocal().named("d")), lambda x,d: 1-d), # x*/(1+x) -> 1-1/(1+x)
|
||||
(UPat.var("x") * ((1+UPat.var("x")).reciprocal().named("d")*UPat.var("y")), lambda x,y,d: y*(1-d)),
|
||||
(UPat.var("x") * ((1+UPat.var("x")).reciprocal().named("d")+UPat.var("y")), lambda x,y,d: (1-d)+x*y),
|
||||
# move const multiply after REDUCE (NOTE: the mul chain can do this, but only if it's a same dtype reduce)
|
||||
((UPat.var("x")*UPat.cvar("c")).reduce(arg=Ops.ADD, name="r", allow_any_len=True), lambda x,c,r: r.replace(src=(x,)+r.src[1:])*c.val),
|
||||
# reduce mul chain, move muls after the reduce
|
||||
(UPat(Ops.MUL).reduce(name="r", allow_any_len=True), reduce_mul_chain),
|
||||
# ** combine terms (opinionated) **
|
||||
(-1 * (UPat.var("x") + UPat.var("y")), lambda x,y: (-x)+(-y)), # -(x+y) -> -x + -y
|
||||
# (x+y)*c -> x*c+y*c. only for int, float has inf*0=nan issue
|
||||
((UPat.var("x", dtypes.weakint) + UPat.var("y")) * UPat.cvar("c"), lambda x,y,c: x*c+y*c),
|
||||
])+pm_clean_up_group_sink
|
||||
171
artifacts/package_runtime/tinygrad/uop/upat.py
Normal file
171
artifacts/package_runtime/tinygrad/uop/upat.py
Normal file
@@ -0,0 +1,171 @@
|
||||
from typing import Any, Callable
|
||||
import itertools, inspect, functools, types
|
||||
from tinygrad.helpers import partition, dedup, Context
|
||||
from tinygrad.uop.ops import UPat, UOp, Ops, PatternMatcher, graph_rewrite, deconstruct_function
|
||||
|
||||
class UPatCompileError(Exception): pass
|
||||
|
||||
# **** UPat compiled ****
|
||||
# This file builds an IR of match predicates and compiles them to Python source.
|
||||
# Ops used: CUSTOM (format-string predicate over operands), CUSTOMI (inline string fragment),
|
||||
# STORE (bind a matched UOp to a name), PYLITERAL (Python literal for CUSTOM operands),
|
||||
# AND/OR (clause combininers).
|
||||
|
||||
def _get_clause(self:UPat, base:UOp, depth=0) -> UOp:
|
||||
if self.is_any:
|
||||
assert len(self.src) == 1
|
||||
return UOp(Ops.AND, src=(UOp(Ops.OR, src=tuple(_get_clause(s, base, depth) for s in self.src[0])),))
|
||||
# build the and_clause for acceptance
|
||||
and_clause:list[UOp] = []
|
||||
if self.op is not None:
|
||||
if len(self.op) > 1: and_clause.append(UOp(Ops.CUSTOM, src=(base, UOp(Ops.PYLITERAL, arg=tuple(int(x) for x in self.op))), arg="{0}.op in {1}"))
|
||||
else: and_clause.append(UOp(Ops.CUSTOM, src=(base,), arg="{0}.op == "+str(self.op[0].value)))
|
||||
if self.arg is not None:
|
||||
if isinstance(self.arg, int): and_clause.append(UOp(Ops.CUSTOM, src=(base,), arg="{0}.arg == "+str(int(self.arg))))
|
||||
else: and_clause.append(UOp(Ops.CUSTOM, src=(base, UOp(Ops.PYLITERAL, arg=self.arg)), arg="{0}.arg == {1}"))
|
||||
if self.strict_length or self.required_len > 0:
|
||||
and_clause.append(UOp(Ops.CUSTOM, src=(base,), arg=("len({0}.src)"+(" == " if self.strict_length else " >= ")+str(self.required_len))))
|
||||
if self.name is not None: and_clause.append(UOp(Ops.STORE, src=(UOp(Ops.CUSTOMI, arg=self.name), base)))
|
||||
if self.match_dtype is not None:
|
||||
if len(self.match_dtype) > 1:
|
||||
and_clause.append(UOp(Ops.CUSTOM, src=(base, UOp(Ops.PYLITERAL, arg=tuple(self.match_dtype))),
|
||||
arg="{0}.dtype in {1}"))
|
||||
else:
|
||||
and_clause.append(UOp(Ops.CUSTOM, src=(base, UOp(Ops.PYLITERAL, arg=self.match_dtype[0])),
|
||||
arg="{0}.dtype == {1}"))
|
||||
if self.match_tag is not None:
|
||||
if len(self.match_tag) > 1:
|
||||
and_clause.append(UOp(Ops.CUSTOM, src=(base, UOp(Ops.PYLITERAL, arg=tuple(self.match_tag))), arg="{0}.tag in {1}"))
|
||||
else: and_clause.append(UOp(Ops.CUSTOM, src=(base, UOp(Ops.PYLITERAL, arg=self.match_tag[0])), arg="{0}.tag == {1}"))
|
||||
if self.src is not None:
|
||||
# single match
|
||||
if len(self.src) == 1 and isinstance(self.src[0], tuple):
|
||||
and_clause += [_get_clause(s, base.index(i), depth) for i,s in enumerate(self.src[0])]
|
||||
# repeat match
|
||||
elif len(self.src) == 1 and isinstance(self.src[0], itertools.repeat):
|
||||
it = UOp(Ops.CUSTOMI, arg=f"ituop{depth}")
|
||||
match = _get_clause(next(self.src[0]), it, depth+1)
|
||||
and_clause.append(UOp(Ops.CUSTOM, src=(match, it, base), arg="all([{0} for {1} in {2}.src])"))
|
||||
# multi match (fork)
|
||||
elif len(self.src) > 1 and all(isinstance(x, tuple) for x in self.src):
|
||||
fork_cond = [UOp(Ops.AND, src=tuple([_get_clause(s, base.index(i), depth) for i,s in enumerate(ss)])) for ss in self.src]
|
||||
and_clause.append(UOp(Ops.OR, src=tuple(fork_cond)))
|
||||
else: raise RuntimeError("broken")
|
||||
return UOp(Ops.AND, src=tuple(and_clause)) if and_clause else UOp(Ops.CUSTOMI, arg="True")
|
||||
|
||||
# *** pattern matcher ***
|
||||
|
||||
def do_process_and(a:UOp) -> UOp|None:
|
||||
found = False
|
||||
new_src:list[UOp] = []
|
||||
or_clause:list[UOp] = []
|
||||
|
||||
# remove any nested ANDs, extract or clauses
|
||||
for x in a.src:
|
||||
if x.op is Ops.AND:
|
||||
new_src.extend(x.src)
|
||||
found = True
|
||||
elif x.op is Ops.OR: or_clause.append(x)
|
||||
else: new_src.append(x)
|
||||
|
||||
# too big to compile
|
||||
if len(or_clause) >= 4: raise UPatCompileError("too big to compile")
|
||||
|
||||
# one or clause max
|
||||
if len(or_clause) > 1:
|
||||
# need the product of the or clauses
|
||||
or_clause = [UOp(Ops.OR, src=tuple([UOp(Ops.AND, src=x) for x in itertools.product(*[x.src for x in or_clause])]))]
|
||||
found = True
|
||||
|
||||
# handle stores
|
||||
stores, new_src = partition(new_src, lambda x: x.op is Ops.STORE)
|
||||
if len(stores):
|
||||
if len(or_clause):
|
||||
# push stores to the top if we have an or_clause
|
||||
assert len(or_clause) == 1 and all(x.op is Ops.AND for x in or_clause[0].src)
|
||||
or_clause = [UOp(Ops.OR, src=tuple([x.replace(src=x.src+tuple(stores)) for x in or_clause[0].src]))]
|
||||
found = True
|
||||
else:
|
||||
# check for duplicate stores
|
||||
dict_stores: dict[UOp, UOp] = {}
|
||||
for store in stores:
|
||||
if store.src[0] in dict_stores:
|
||||
# duplicate store is an identity compare
|
||||
new_src.append(UOp(Ops.CUSTOM, src=(dict_stores[store.src[0]], store.src[1]), arg="{0} is {1}"))
|
||||
found = True
|
||||
else:
|
||||
dict_stores[store.src[0]] = store.src[1]
|
||||
# put the stores back
|
||||
for k,v in dict_stores.items(): new_src.append(UOp(Ops.STORE, src=(k,v)))
|
||||
|
||||
# reassemble, if there's any deduping to do, do it
|
||||
if len(dretand:=dedup(new_src+or_clause)) != len(new_src)+len(or_clause): found = True
|
||||
return UOp(Ops.AND, src=tuple(dretand)) if found else None
|
||||
|
||||
# processor
|
||||
pm_proc = PatternMatcher([(UPat(Ops.AND, name="a"), do_process_and)], compiled=False)
|
||||
|
||||
# renderer
|
||||
def wrap(ctx, x) -> UOp:
|
||||
ctx[ret:=f"a{len(ctx)}"] = x.arg
|
||||
return UOp(Ops.CUSTOMI, arg=ret)
|
||||
|
||||
pm_renderer = PatternMatcher([
|
||||
(UPat(Ops.PYLITERAL, name="x"), wrap),
|
||||
|
||||
# AND of CUSTOMI fragments inside a CUSTOM becomes a single CUSTOMI (joined with " and ")
|
||||
(UPat(Ops.CUSTOM, src=(UPat(Ops.AND, src=UPat(Ops.CUSTOMI), name="x"), UPat(), UPat()), name="r"),
|
||||
lambda r,x: r.replace(src=(UOp(Ops.CUSTOMI, arg="(" + ' and '.join(y.arg for y in x.src) + ")"),)+r.src[1:])),
|
||||
|
||||
(UPat(Ops.CUSTOM, src=UPat(Ops.CUSTOMI), name="x"), lambda x: UOp(Ops.CUSTOMI, arg=x.arg.format(*[y.arg for y in x.src]))),
|
||||
(UPat(Ops.INDEX, src=(UPat(Ops.CUSTOMI, name="x"), UPat(Ops.CONST, name="c")), name="g"), lambda x,c,g: x.replace(arg=x.arg+f".src[{c.val}]"))
|
||||
], compiled=False)
|
||||
|
||||
def _final_render(x:UOp, has_ctx:bool, depth=1) -> list[str]:
|
||||
assert x.op is Ops.AND
|
||||
and_pieces, store_pieces = [], []
|
||||
or_pieces: list[str] = []
|
||||
for s in x.src:
|
||||
if s.op is Ops.OR:
|
||||
assert len(or_pieces) == 0 and len(s.src) >= 1
|
||||
for ss in s.src: or_pieces.extend(_final_render(ss, has_ctx, depth+1))
|
||||
elif s.op is Ops.STORE:
|
||||
assert s.src[0].op is Ops.CUSTOMI and s.src[1].op is Ops.CUSTOMI
|
||||
store_pieces.append(f"{s.src[0].arg}={s.src[1].arg}")
|
||||
elif s.op is Ops.CUSTOMI: and_pieces.append(s.arg)
|
||||
else: raise UPatCompileError(f"can't compile this {s}")
|
||||
# if we have an or, render it
|
||||
if len(or_pieces):
|
||||
assert len(store_pieces) == 0
|
||||
and_clause = ' and '.join(and_pieces)
|
||||
return [f"{' '*depth}if {and_clause if len(and_clause) else 'True'}:"] + or_pieces
|
||||
# if we don't, this is a final return
|
||||
store_clause = ', '.join((["ctx=ctx"] if has_ctx else [])+store_pieces)
|
||||
and_clause = ' and '.join(and_pieces + [f"(_ret:=_fxn({store_clause})) is not None"])
|
||||
return [f"{' '*depth}if {and_clause}: return _ret"]
|
||||
|
||||
def _get_code(self:UPat, has_ctx:bool):
|
||||
ret = _get_clause(self, UOp(Ops.CUSTOMI, arg="uop"))
|
||||
try:
|
||||
# TODO: this should be tracked in a "system" rewrite, not untracked or tracked with kernel
|
||||
with Context(TRACK_MATCH_STATS=0):
|
||||
ret = graph_rewrite(ret, pm_proc, name="process UPat")
|
||||
dyn_lookup: dict[str, Any] = {}
|
||||
out = graph_rewrite(ret, pm_renderer, ctx=dyn_lookup, name="compile UPat")
|
||||
rendered = _final_render(out, has_ctx)
|
||||
except UPatCompileError:
|
||||
#print("FAILED", self, self.location)
|
||||
return None
|
||||
return '\n'.join([f"# match for {self.location}", "def compiled_match(uop, ctx):"] + rendered + [" return None"]), dyn_lookup
|
||||
|
||||
@functools.cache
|
||||
def upat_compile(self:UPat, fxn) -> Callable|None:
|
||||
real_fxn = types.FunctionType(*deconstruct_function(fxn))
|
||||
code = _get_code(self, 'ctx' in inspect.signature(real_fxn).parameters)
|
||||
if code is None: return None
|
||||
code_str, dyn_lookup = code
|
||||
globs = dyn_lookup.copy()
|
||||
globs["_fxn"] = real_fxn
|
||||
namespace: dict = {}
|
||||
exec(code_str, globs, namespace) # pylint: disable=W0122
|
||||
return namespace["compiled_match"]
|
||||
88
artifacts/package_runtime/tinygrad/uop/validate.py
Normal file
88
artifacts/package_runtime/tinygrad/uop/validate.py
Normal file
@@ -0,0 +1,88 @@
|
||||
from typing import Callable
|
||||
from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, python_alu
|
||||
from tinygrad.dtype import dtypes, Invalid
|
||||
from tinygrad.helpers import cpu_profile
|
||||
import z3
|
||||
|
||||
# older versions of z3 dont have some operators like & overloaded
|
||||
if z3.get_version() < (4, 12, 4, 0):
|
||||
raise ImportError("bounds checking requires z3 >= 4.12.4, use CHECK_OOB=0 to disable, or \"pip install 'z3-solver>=4.12.4\"")
|
||||
|
||||
# IDIV is truncated division but z3 does euclidian division (floor if b>0 ceil otherwise); mod by power of two sometimes uses Ops.AND
|
||||
def z3_cdiv(a:z3.ArithRef, b:z3.ArithRef) -> z3.ArithRef:return z3.If((a<0), z3.If(0<b, (a+(b-1))/b, (a-(b+1))/b), a/b)
|
||||
def z3_floordiv(a:z3.ArithRef, b:z3.ArithRef) -> z3.ArithRef: return z3.If(b > 0, a/b, (-a)/(-b))
|
||||
def z3_xor(a:z3.ExprRef, b:z3.ExprRef) -> z3.ExprRef:
|
||||
if isinstance(a, z3.BoolRef): return a^b
|
||||
# x ^ -1 = -(x+1), i.e. bitwise NOT
|
||||
if isinstance(b, z3.IntNumRef) and b.as_long() == -1: return -(a+1)
|
||||
if isinstance(a, z3.IntNumRef) and a.as_long() == -1: return -(b+1)
|
||||
raise RuntimeError(f"z3 int XOR only supports XOR with -1, got {a=} {b=}")
|
||||
def z3_and(a:z3.ExprRef, b:z3.ExprRef) -> z3.ExprRef:
|
||||
if isinstance(a, z3.BoolRef): return a&b
|
||||
if isinstance(a, z3.IntNumRef): a, b = b, a
|
||||
if isinstance(b, z3.IntNumRef):
|
||||
# x & (2^k-1) = x % 2^k and x & -(2^k) = x - x % 2^k for any x in two's complement
|
||||
if (m:=b.as_long()+1) > 0 and m&(m-1) == 0: return a%m
|
||||
if (m:=-b.as_long()) > 0 and m&(m-1) == 0: return a - a%m
|
||||
raise RuntimeError(f"z3 int AND only supports 2**k-1 and -2**k masks, got {a=} {b=}")
|
||||
z3_alu: dict[Ops, Callable[..., z3.ExprRef]] = python_alu | {Ops.CMOD: lambda a,b: a-z3_cdiv(a,b)*b, Ops.CDIV: z3_cdiv, Ops.FLOORDIV: z3_floordiv,
|
||||
Ops.FLOORMOD: lambda a,b: a-z3_floordiv(a,b)*b,
|
||||
Ops.SHR: lambda a,b: a/(2**b.as_long()), Ops.SHL: lambda a,b: a*(2**b.as_long()),
|
||||
Ops.AND: z3_and, Ops.WHERE: z3.If, Ops.XOR: z3_xor, Ops.MAX: lambda a,b: z3.If(a<b, b, a),}
|
||||
def create_bounded(name:str, vmin:int, vmax:int, z3ctx:z3.Context) -> tuple[z3.ArithRef, z3.BoolRef]:
|
||||
return (s:=z3.Int(name, ctx=z3ctx)), (vmin <= s)&(s <= vmax)
|
||||
|
||||
z3_renderer = PatternMatcher([
|
||||
(UPat.var("cond").where(UPat.var("x"), UPat(Ops.CONST, arg=Invalid)), lambda x,cond,ctx: (ctx[1][x], ctx[1][cond])),
|
||||
# variables
|
||||
(UPat(Ops.SPECIAL, name="x"), lambda x,ctx: create_bounded(x.arg, 0, ctx[1][x.src[0]]-1, ctx[0])),
|
||||
(UPat(Ops.PARAM, name="x"), lambda x,ctx: create_bounded(x.arg.name, x.vmin, x.vmax, ctx[0])),
|
||||
(UPat(Ops.RANGE, name="x"), lambda x,ctx: create_bounded(x.render(simplify=False), 0, ctx[1][x.src[0]]-1, ctx[0])),
|
||||
# loads are variables bounded by the min/max of the dtype. non-pointer INDEX is also a LOAD
|
||||
(UPat((Ops.LOAD, Ops.INDEX), dtypes.ints+(dtypes.weakint,), name="x"), lambda x,ctx:
|
||||
create_bounded(f"load{len(ctx[1])}", x.dtype.min, x.dtype.max, ctx[0])),
|
||||
(UPat((Ops.LOAD, Ops.INDEX), dtypes.bool), lambda ctx: (z3.Bool(f"load{len(ctx[1])}", ctx=ctx[0]), None)),
|
||||
# constants
|
||||
(UPat(Ops.CONST, arg=Invalid), lambda ctx: (z3.Int("Invalid", ctx=ctx[0]), None)),
|
||||
(UPat(Ops.CONST, dtypes.ints+(dtypes.weakint,), name="x"), lambda x,ctx: (z3.IntVal(x.val, ctx=ctx[0]), None)),
|
||||
(UPat(Ops.CONST, dtypes.bool, name="x"), lambda x,ctx: (z3.BoolVal(x.val, ctx=ctx[0]), None)),
|
||||
# casts from floats create new variables
|
||||
(UPat(Ops.CAST, dtypes.ints+(dtypes.weakint,), src=(UPat(dtype=dtypes.floats),), name="x"), lambda x,ctx:
|
||||
create_bounded(f"cast{len(ctx[1])}", x.dtype.min, x.dtype.max, ctx[0])),
|
||||
# A comparison between floats introduces a new bool variable
|
||||
(UPat(GroupOp.Comparison, src=UPat(dtype=dtypes.floats)), lambda ctx: (z3.Bool(f"float_cmp{len(ctx[1])}", ctx=ctx[0]), None)),
|
||||
# casts from bool/int to int/bool
|
||||
(UPat(Ops.CAST, dtypes.ints+(dtypes.weakint,),src=(UPat.var("x", dtypes.bool),)), lambda x,ctx: (z3.If(ctx[1][x], 1, 0), None)),
|
||||
(UPat(Ops.CAST, dtypes.ints+(dtypes.weakint,), src=(UPat.var("x", dtypes.ints+(dtypes.weakint,)),)), lambda x,ctx: (ctx[1][x], None)),
|
||||
(UPat(Ops.CAST, dtypes.bool, name="x"), lambda x,ctx: (ctx[1][x.src[0]]!=0, None)),
|
||||
(UPat(GroupOp.ALU, name="x"), lambda x,ctx: (z3_alu[x.op](*(ctx[1][s] for s in x.src)), None)),
|
||||
])
|
||||
|
||||
def uops_to_z3(solver:z3.Solver, *uops: UOp) -> list[z3.ExprRef]:
|
||||
# gate on upstream memory addressing, but keep INDEX as an unknown LOAD
|
||||
lst = list(UOp.sink(*uops).toposort(gate=lambda x: x.op not in {Ops.AFTER, Ops.BUFFER, Ops.SHRINK} and \
|
||||
(x.dtype in dtypes.ints+(dtypes.bool, dtypes.weakint) or x.op is Ops.SINK)))[:-1]
|
||||
z3map: dict[UOp, z3.ExprRef] = {}
|
||||
for u in lst:
|
||||
# NOTE: we skip STACK here, it can't actually be accessed
|
||||
if u.op is Ops.STACK: continue
|
||||
z3_rewritten: tuple[z3.ExprRef, z3.BoolRef|None]|None = z3_renderer.rewrite(u, ctx=(solver.ctx, z3map))
|
||||
if z3_rewritten is None: raise NotImplementedError(f"{u.op} is not supported by z3")
|
||||
new_u, constraint = z3_rewritten
|
||||
if constraint is not None: solver.add(constraint)
|
||||
z3map[u] = new_u
|
||||
assert all(u in z3map for u in uops), "UOp failed to rewrite to z3!"
|
||||
return [z3map[u] for u in uops]
|
||||
|
||||
def validate_index_with_z3(sz:int, idx:UOp, gate:UOp) -> bool:
|
||||
solver = z3.Solver(ctx=z3.Context())
|
||||
z3_idx, z3_mask = uops_to_z3(solver, idx, gate)
|
||||
solver.add(z3_mask)
|
||||
with cpu_profile("validate index with z3", "TINY"):
|
||||
match solver.check((z3_idx<0)|(sz<=z3_idx)):
|
||||
case z3.unsat: return True
|
||||
case z3.sat: print(f"# OUT OF BOUNDS ACCESS: at {solver.model()} INDEX not in 0 - {sz}\nconstraints = {solver}")
|
||||
case z3.unknown: print(f"# UNKNOWN RESULT FROM Z3: {solver.reason_unknown()}\nconstraints = {solver}")
|
||||
print(f"idx={idx.render(simplify=False)}")
|
||||
print(f"mask={gate.render(simplify=False)}")
|
||||
return False
|
||||
78
artifacts/package_runtime/tinygrad/uop/weak.py
Normal file
78
artifacts/package_runtime/tinygrad/uop/weak.py
Normal file
@@ -0,0 +1,78 @@
|
||||
from dataclasses import replace
|
||||
from tinygrad.dtype import dtypes, DType, AddrSpace, Invalid, least_upper_dtype, strong_dtype, weak_dtype
|
||||
from tinygrad.helpers import unwrap
|
||||
from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, GroupOp, graph_rewrite, dtype_from_uop
|
||||
|
||||
def select_dtype(u:UOp):
|
||||
if u.dtype is dtypes.weakfloat: return dtypes.default_float
|
||||
return dtypes.long if u.overflows(dtypes.int32) else dtypes.int
|
||||
|
||||
def lower_weak_node(u:UOp) -> UOp|None:
|
||||
start, src = (1 if u.op is Ops.WHERE else 0), tuple(s.src[0] if s.op is Ops.CAST and s.dtype in dtypes.weaks else s for s in u.src)
|
||||
if src == u.src or any(s.dtype in dtypes.weaks for s in src[start:]): return None
|
||||
dt = strong_dtype(least_upper_dtype(select_dtype(u), *(s.dtype for s in src)) if u.op in GroupOp.Binary
|
||||
else unwrap(dtype_from_uop(u.op, src, u.arg)))
|
||||
return u.replace(dtype=None, src=src[:start]+tuple(s if s.base.is_invalid else s.cast(dt) for s in src[start:])).cast(u.dtype)
|
||||
|
||||
pm_lower_weak = PatternMatcher([
|
||||
(UPat(Ops.CONST, dtype=dtypes.weaks, name="u"), lambda u: UOp.const(u.val, select_dtype(u)).cast(u.dtype)),
|
||||
# two stacked weak casts are a weakint value used as weakfloat (or vice versa): resolve the inner one at the outer kind's default.
|
||||
# a SINGLE weak cast is never rewritten here, each consumer absorbs it on its own edge (see lower_weak_srcs)
|
||||
(UPat(Ops.CAST, dtype=dtypes.weaks, src=(UPat(Ops.CAST, dtype=dtypes.weaks, src=(UPat.var("x"),)),), name="u"),
|
||||
lambda u,x: x.cast(select_dtype(u)).cast(u.dtype) if x.dtype not in dtypes.weaks else None),
|
||||
# Binary can widen from the bounds, all other nodes derive from the lowered sources.
|
||||
# a weakfloat Unary (sin/exp2/...) must resolve here, before the transcendental decomposition
|
||||
(UPat(GroupOp.Binary|GroupOp.Unary|{Ops.WHERE, Ops.RANGE, Ops.STACK, Ops.SPECIAL}, name="u"), lower_weak_node),
|
||||
(UPat(Ops.PARAM, dtype=dtypes.weakint, name="u"),
|
||||
lambda u: u.replace(dtype=None, arg=replace(u.arg, dtype=select_dtype(u))).cast(dtypes.weakint) if u.addrspace == AddrSpace.ALU else None),
|
||||
])
|
||||
|
||||
def lower_weak_srcs(ctx:dict[UOp, UOp]|None, u:UOp) -> UOp|None:
|
||||
if ctx is None: ctx = {}
|
||||
def lower(s:UOp) -> UOp:
|
||||
if (r:=ctx.get(s)) is None:
|
||||
r = graph_rewrite(s, pm_lower_weak)
|
||||
# the consumer absorbs the cast on its own edge
|
||||
ctx[s] = r = r.src[0] if r.op is Ops.CAST and r.dtype in dtypes.weaks else r
|
||||
return r
|
||||
# a comparison demands a common operand width: lower it whole so the Binary rule unifies its operands
|
||||
ret = lower(u) if u.op in GroupOp.Comparison else u.replace(src=tuple(lower(s) if s.dtype in dtypes.weaks else s for s in u.src))
|
||||
return None if ret is u else ret
|
||||
|
||||
def commit_weak(s:UOp, dt:DType) -> UOp:
|
||||
# a bare weak CONST commits directly (the value stays mathematical, emission truncates), a weak non-const src takes the demand cast
|
||||
return UOp.const(s.val, dt) if s.op is Ops.CONST else s.cast(dt)
|
||||
|
||||
def commit_weak_srcs(u:UOp) -> UOp|None:
|
||||
if not any(s.dtype in dtypes.weaks for s in u.src): return None
|
||||
if (dt:=least_upper_dtype(*(s.dtype for s in u.src))) in dtypes.weaks: return None
|
||||
# the root re-derives: a shift's dtype is its lhs's, so committing the lhs commits the node too
|
||||
return u.replace(dtype=None, src=tuple(commit_weak(s, dt) if s.dtype in dtypes.weaks else s for s in u.src))
|
||||
|
||||
# runs in index lowering and in the decomps: a rule that mints a weak const commits it in the same rewrite, so none reaches the renderer
|
||||
pm_commit_weak = PatternMatcher([
|
||||
(UPat(GroupOp.Broadcastable, name="u"), commit_weak_srcs),
|
||||
# demand from the destination: a STORE's weak value commits at the destination's dtype
|
||||
(UPat(Ops.STORE, src=(UPat(), UPat(dtype=dtypes.weaks)), allow_any_len=True, name="u"),
|
||||
lambda u: u.replace(src=(u.src[0], commit_weak(u.src[1], u.src[0].dtype), *u.src[2:]))),
|
||||
])
|
||||
|
||||
# a concrete CAST over a weak node states the width the value will live at. that width is a floor, never a narrowing
|
||||
def cast_weak_srcs(c:UOp, u:UOp) -> UOp|None:
|
||||
if c.dtype in dtypes.weaks or weak_dtype(c.dtype) is not u.dtype: return None
|
||||
dt = least_upper_dtype(c.dtype, select_dtype(u))
|
||||
return u.replace(dtype=None, src=tuple(commit_weak(s, dt) if s.dtype in dtypes.weaks else s for s in u.src)).cast(c.dtype)
|
||||
|
||||
pm_cast_weak = PatternMatcher([
|
||||
(UPat(Ops.CAST, name="c", src=(UPat(GroupOp.ALU, dtype=dtypes.weaks, name="u"),)), cast_weak_srcs),
|
||||
])
|
||||
|
||||
pm_lower_index_dtype = pm_commit_weak+pm_cast_weak+PatternMatcher([
|
||||
(UPat(GroupOp.All, name="u"),
|
||||
lambda ctx,u: lower_weak_srcs(ctx, u) if u.dtype not in dtypes.weaks and any(s.dtype in dtypes.weaks for s in u.src) else None),
|
||||
# a valid index into an n-element buffer lives in [0,n): a gated long index narrows when n-1 fits int32 (out-of-gate wraps, discarded)
|
||||
# TODO: more generic
|
||||
(UPat((Ops.INDEX, Ops.SHRINK), src=(UPat.var("buf"), UPat.var("gate").where(UPat.var("idx", dtypes.long), UPat(Ops.CONST, arg=Invalid))),
|
||||
allow_any_len=True, name="u"),
|
||||
lambda u,buf,gate,idx: u.replace(src=(buf, idx.cast(dtypes.int).valid(gate))+u.src[2:]) if buf.max_numel()-1 <= dtypes.int32.max else None),
|
||||
])
|
||||
Reference in New Issue
Block a user