github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-25 23:08:06 -05:00
Code Issues Packages Projects Releases Wiki Activity
Files
59315ffc7839948a032e366ba8d964c345d835ff
tinygrad/tinygrad/codegen/uopgraph.py
chenyu 59315ffc78 minor cleanup to UOp mod folding [run_process_replay] (#5895) 
some walrus
2024-08-03 21:38:44 -04:00

597 lines
33 KiB
Python
Raw Blame History

from __future__ import annotations
from typing import Iterator, Optional, Tuple, Dict, List, Set, Union, cast, TYPE_CHECKING, Any, DefaultDict, Callable
import functools, itertools, heapq, math, operator
from collections import defaultdict
from tinygrad.dtype import dtypes, PtrDType, ImageDType, DType
from tinygrad.ops import UnaryOps, BinaryOps, ReduceOps, exec_alu
from tinygrad.helpers import DEBUG, getenv, flatten, dedup, TRANSCENDENTAL, prod, CI, all_same
from tinygrad.codegen.uops import UOp, NOp, UOps, UPat, PatternMatcher, END_FOR_UOP, type_verify
from tinygrad.codegen.transcendental import xexp2, xlog2, xsin, TRANSCENDENTAL_SUPPORTED_DTYPES
if TYPE_CHECKING: from tinygrad.renderer import Renderer
# ***** float4/image store handling *****
def fold_expanded(ex, buf):
if buf.dtype != PtrDType(dtypes.float) and buf.dtype != PtrDType(dtypes.half) and not isinstance(buf.dtype, ImageDType): return None
new_srcs = dedup(list(ex.src))
old_new_srcs = new_srcs[:]
is_load, is_image = new_srcs[0].op is UOps.LOAD, isinstance(buf.dtype, ImageDType)
# first, extract all the relevant offsets
offsets_rootsrc: DefaultDict[Any, dict] = defaultdict(dict)
for i,s in enumerate(new_srcs):
if (s.dtype is not None and s.dtype.count != 1) or (is_image and s.src[1].dtype != dtypes.int.vec(3)): continue
idx = s.src[1] if not is_image else s.src[1].src[2] # only id4 for image
if idx.arg is BinaryOps.ADD and idx.src[1].op is UOps.CONST: root_src, arg = idx.src[0], idx.src[1].arg
elif idx.op is UOps.CONST: root_src, arg = "CONST", idx.arg
else: root_src, arg = idx, 0
# add idx and idy for image
if is_image: root_src = (s.src[1].src[0:2], root_src)
# add gates for gated
if len(s.src) >= 4: root_src = (s.src[3], root_src)
assert arg not in offsets_rootsrc[root_src]
offsets_rootsrc[root_src][arg] = i
# then rewrite everything we can
used = set()
for rootsrc, offsets in offsets_rootsrc.items():
for o in offsets:
for fold_length in [4] if is_image else ([8,4,2] if buf.dtype == PtrDType(dtypes.half) and getenv("ALLOW_HALF8") else [4,2]):
if all((rootsrc,o+i) not in used and o+i in offsets for i in range(fold_length)):
load_1 = new_srcs[offsets[o]]
new_src = list(load_1.src)
if not is_image and not new_src[1].divides(fold_length): continue
# for images, we rewrite the index
if is_image: new_src[1] = UOp(UOps.VECTORIZE, dtypes.int.vec(2), (new_src[1].src[0], new_src[1].src[1]))
# vectorize the store/loadconst
if not is_load or len(new_src) >= 4:
new_src[2] = UOp(UOps.VECTORIZE, new_src[2].dtype.vec(fold_length), tuple(new_srcs[offsets[o+i]].src[2] for i in range(fold_length)))
# generate the folded new_srcs
if is_load:
new_load = UOp(UOps.LOAD, load_1.dtype.vec(fold_length), tuple(new_src))
for i in range(fold_length): new_srcs[offsets[o+i]] = UOp(UOps.GEP, load_1.dtype, (new_load,), i)
else:
for i in range(fold_length): new_srcs[offsets[o+i]] = UOp(UOps.STORE, None, tuple(new_src)) if i == 0 else None
for i in range(fold_length): used.add((rootsrc,o+i))
# dedup expand for LOAD
if is_load and len(old_new_srcs) != len(ex.src): new_srcs = [new_srcs[old_new_srcs.index(s)] for s in ex.src]
# remove Nones for STORE
return UOp(ex.op, ex.dtype, tuple(x for x in new_srcs if x is not None), ex.arg) if len(used) else None
def vectorize_reduce(vec:UOp):
if not all_same([(x.src[1:], x.arg) for x in vec.src]): return None
return UOp(UOps.REDUCE, vec.dtype, (UOp(UOps.VECTORIZE, vec.dtype, tuple(x.src[0] for x in vec.src)),) + vec.src[0].src[1:], vec.src[0].arg)
def vectorize_alu(vec:UOp):
if not all_same([x.arg for x in vec.src]): return None
return UOp(vec.src[0].op, vec.dtype, tuple(UOp(UOps.VECTORIZE, cast(DType, vec.src[0].src[i].dtype).vec(cast(DType, vec.dtype).count),
tuple(x.src[i] for x in vec.src)) for i in range(len(vec.src[0].src))), vec.src[0].arg)
float4_folding = PatternMatcher([
(UPat(UOps.EXPAND, src=UPat(UOps.LOAD, src=(UPat(name="buf"), UPat()), allow_any_len=True), name="ex"), fold_expanded),
(UPat({UOps.BARRIER, UOps.SINK}, src=UPat(UOps.STORE, src=(UPat(name="buf"), UPat(), UPat()), allow_any_len=True), name="ex"), fold_expanded),
(UPat(UOps.VECTORIZE, src=UPat(UOps.REDUCE), name="vec"), vectorize_reduce),
(UPat(UOps.VECTORIZE, src=UPat({UOps.ALU, UOps.CAST, UOps.BITCAST}), name="vec"), vectorize_alu),
])
# ***** mod *****
def _get_add_chain(x:UOp):
if x.op is UOps.ALU and x.arg is BinaryOps.ADD:
for s in x.src: yield from _get_add_chain(s)
else: yield x
def mod_folding(x:UOp, c:int) -> Optional[UOp]:
# simplify x in x % c
# None means no change
remainder, something_changed = [], False
for u in _get_add_chain(x):
if u.op is UOps.CONST and (r:=u.arg%c) != u.arg:
if r: remainder.append(u.const(r))
something_changed = True
elif u.op is UOps.ALU and u.arg is BinaryOps.MUL:
if (u0:=u.src[0]).op is UOps.CONST and (r:=u0.arg%c) != u0.arg:
if r: remainder.append(u.src[1] if r==1 else u.const(r)*u.src[1])
something_changed = True
elif (u1:=u.src[1]).op is UOps.CONST and (r:=u1.arg%c) != u1.arg:
if r: remainder.append(u.src[0] if r==1 else u.src[0]*u.const(r))
something_changed = True
else: remainder.append(u)
else: remainder.append(u)
if not something_changed: return None
return functools.reduce(operator.add, remainder) if remainder else x.const(0)
# ***** transcendental *****
transcendental_folding = PatternMatcher([(UPat(UOps.ALU, dtype=TRANSCENDENTAL_SUPPORTED_DTYPES, src=(UPat(name="d"),), arg=k), cast(Callable, v))
for k,v in ((UnaryOps.EXP2, xexp2), (UnaryOps.LOG2, xlog2), (UnaryOps.SIN, xsin))])
# ***** threefry *****
def threefry2x32(x: UOp, seed: UOp):
# split x into two uint32, since x in a uint64
x0, x1 = (x & 0xffffffff).cast(dtypes.uint32), ((x // 2**32) & 0xffffffff).cast(dtypes.uint32)
rotations = [[13, 15, 26, 6], [17, 29, 16, 24]]
ks = [0x0, (seed := seed.cast(dtypes.uint32)) ^ 0x1BD11BDA, seed]
xr = [x0 + ks[-1], x1 + ks[0]]
for i in range(5):
for r in rotations[i % 2]: xr[0], xr[1] = (x0 := xr[0] + xr[1]), x0 ^ ((xr[1] * 2**r) + (xr[1] // 2**(32 - r)))
xr = [(xr[0] + ks[i % 3]), (xr[1] + ks[(i + 1) % 3] + i + 1)]
return xr[1].cast(dtypes.uint64) * 2**32 | xr[0].cast(dtypes.uint64)
# ***** main rewriter *****
def reduce_before_expand(reduce, expand, x):
# if the expand is being reduced, you can't push it through
# NOTE: could do a partial push here in some cases
expands = flatten([x.arg for x in reduce.src[1:] if x.op is UOps.EXPAND])
if any(x in expands for x in expand.arg): return None
red = UOp(UOps.REDUCE, x.dtype, (x,)+reduce.src[1:], reduce.arg)
return UOp(expand.op, expand.dtype, tuple(UOp(UOps.GEP, reduce.dtype, (red,), i) for i in range(x.dtype.count)), expand.arg)
def sum_collapse(phi_input, loop, val1, val2):
for v1,v2 in [(val1, val2), (val2, val1)]:
if loop not in v1.parents: return UOp(UOps.PHI, phi_input.dtype, (phi_input, v2))+v1*(loop.src[1]-loop.src[0]).cast(v1.dtype)
return None
def loop_collapse(loop_start, loop_end, compval, idx, mval, multconst, rng, reduce, idx2=None, idx3=None):
if getenv("DISABLE_LOOP_COLLAPSE") or rng not in reduce.src: return None # must be the right REDUCE
if mval.arg >= 0 or loop_start.arg != 0:
# TODO: support and test this with other mvals and loop_starts
if DEBUG >= 1: print(f"WARNING, NOT FOLDING: mval:{mval.arg} loop_start:{loop_start.arg}")
return None
if idx2 is not None: idx = idx + idx2
if idx3 is not None: idx = idx + idx3
comprange = UOp.min(loop_end, UOp.max((idx-compval-mval)//mval + (loop_end-loop_start), loop_start))
return UOp(UOps.REDUCE, reduce.dtype, (comprange.cast(multconst.dtype) * multconst,) +
tuple(x for x in reduce.src[1:] if x is not rng), reduce.arg)
def index_collapse(idx,rng,buf,add,mul,ld,reduce):
if rng not in reduce.src: return None
return UOp(reduce.op, reduce.dtype, (UOp(ld.op, ld.dtype, (buf, add+mul*idx)),)+
tuple(x for x in reduce.src[1:] if x is not rng), reduce.arg)
# this is symbolic 2.0
constant_folder = PatternMatcher([
# bigint is rewritten to int32
(UPat({UOps.CONST, UOps.ALU, UOps.SPECIAL, UOps.RANGE, UOps.EXPAND}, dtype=dtypes.bigint, name="x"),
lambda x: UOp(x.op, dtypes.int32, x.src, x.arg)),
# VECTORIZE/GEP
(NOp(UOps.GEP, src=(NOp(UOps.VECTORIZE, name="cast"),), name="gep"), lambda gep, cast: cast.src[gep.arg]),
*[(NOp(UOps.VECTORIZE, dtypes.float.vec(i), tuple(NOp(UOps.GEP, dtypes.float,
src=(NOp.var('x', dtype=dtypes.float.vec(i)),), arg=j) for j in range(i))), lambda x: x) for i in [2, 4, 8, 16]],
*[(NOp(UOps.VECTORIZE, dtypes.half.vec(i), tuple(NOp(UOps.GEP, dtypes.half,
src=(NOp.var('x', dtype=dtypes.half.vec(i)),), arg=j) for j in range(i))), lambda x: x) for i in [2, 4, 8, 16]],
# tensor core with a 0 input is acc
(NOp(UOps.WMMA, src=(NOp.const(None, 0.0), NOp.var(), NOp.var('acc'))), lambda acc: acc),
(NOp(UOps.WMMA, src=(NOp.var(), NOp.const(None, 0.0), NOp.var('acc'))), lambda acc: acc),
# tensor core cleanups
*[(NOp(UOps.REDUCE, src=(NOp(UOps.EXPAND, src=tuple(NOp(UOps.GEP, dtypes.float, src=(NOp.var('x'),), arg=i) for i in range(j)), name="expand"),)
,name="reduce", allow_any_len=True), reduce_before_expand) for j in [2,4,8]],
(NOp.var("add") + NOp(UOps.WMMA, name="wmma"),
lambda add, wmma: UOp(wmma.op, wmma.dtype, (wmma.src[0], wmma.src[1], wmma.src[2]+add), wmma.arg)),
# threefry
(NOp(UOps.ALU, dtype=dtypes.uint64, src=(NOp.var("x"), NOp.var("seed")), arg=BinaryOps.THREEFRY), threefry2x32),
# sum collapse to mul (with possible GEP)
(UPat(UOps.PHI, src=(UPat(UOps.DEFINE_ACC, name="phi_input", src=[UPat(UOps.CONST), UPat(UOps.RANGE, name="loop")]),
UPat(UOps.ALU, BinaryOps.ADD, src=(UPat(name="val1"), UPat(name="val2"))))), sum_collapse),
(UPat(UOps.PHI, src=(UPat(UOps.GEP, name="phi_input", src=(UPat(UOps.DEFINE_ACC, src=[UPat(UOps.CONST), UPat(UOps.RANGE, name="loop")]),)),
UPat(UOps.ALU, BinaryOps.ADD, src=(UPat(name="val1"), UPat(name="val2"))))), sum_collapse),
# extra arange loop folding because we don't fold adds. TODO: fold adds
(NOp(UOps.REDUCE, src=((NOp.var("idx") + NOp.cvar("mval") * NOp(UOps.RANGE, src=(NOp.var("loop_start"), NOp.var("loop_end")), name="rng") +
NOp.var("idx2") + NOp.var("idx3"))
.lt(NOp.cvar("compval")).where(NOp.cvar("multconst"), NOp.const(None, 0)),), arg=ReduceOps.SUM, name="reduce", allow_any_len=True), loop_collapse),
(NOp(UOps.REDUCE, src=((NOp.var("idx") + NOp.cvar("mval") * NOp(UOps.RANGE, src=(NOp.var("loop_start"), NOp.var("loop_end")), name="rng") +
NOp.var("idx2"))
.lt(NOp.cvar("compval")).where(NOp.cvar("multconst"), NOp.const(None, 0)),), arg=ReduceOps.SUM, name="reduce", allow_any_len=True), loop_collapse),
# arange loop folding (reduce)
(NOp(UOps.REDUCE, src=((NOp.var("idx") + NOp.cvar("mval") * NOp(UOps.RANGE, src=(NOp.var("loop_start"), NOp.var("loop_end")), name="rng"))
.lt(NOp.cvar("compval")).where(NOp.cvar("multconst"), NOp.const(None, 0)),), arg=ReduceOps.SUM, name="reduce", allow_any_len=True), loop_collapse),
(NOp(UOps.REDUCE, src=((NOp.var("idx") - NOp(UOps.RANGE, src=(NOp.var("loop_start"), NOp.var("loop_end")), name="rng"))
.lt(NOp.cvar("compval")).where(NOp.cvar("multconst"), NOp.const(None, 0)),), arg=ReduceOps.SUM, name="reduce", allow_any_len=True),
lambda **kwargs: loop_collapse(mval=UOp.const(dtypes.int, -1), **kwargs)),
# indexing (with a multiply offset)!
(NOp(UOps.REDUCE, src=(NOp.var('idx').eq(NOp(UOps.RANGE, name="rng")).cast()*
NOp(UOps.LOAD, src=(NOp.var("buf"), NOp.var('add')+NOp.var('mul')*NOp(UOps.RANGE, name="rng")), name="ld"),),
arg=ReduceOps.SUM, name="reduce", allow_any_len=True), index_collapse),
(NOp(UOps.REDUCE, src=(NOp.var('idx').eq(NOp(UOps.RANGE, name="rng")).where(
NOp(UOps.LOAD, src=(NOp.var("buf"), NOp.var('add')+NOp.var('mul')*NOp(UOps.RANGE, name="rng")), name="ld"), NOp.const(None, 0.0)),),
arg=ReduceOps.SUM, name="reduce", allow_any_len=True), index_collapse),
# other arange folders
(NOp.cvar("c1") - (NOp.var("x") + NOp.cvar("c2")), lambda c1, c2, x: (c1-c2)-x), # c1 - (x + c2) -> (c1-c2) - x
# max folding
(NOp.max(NOp.var('x'), NOp.var('y')), lambda x,y: x if x.vmin.arg >= y.vmax.arg else y if x.vmax.arg <= y.vmin.arg else None),
# const rules
(NOp(UOps.GEP, src=(NOp.cvar("c"),), name="root"), lambda root, c: root.const(c.arg)),
(UPat(UOps.CAST, name="root", src=UPat(UOps.CONST, name="c")), lambda root, c: root.const(c.arg)),
# a phi on a DEFINE_ACC without loops or a CONST is a noop. this is for correctness, not just speed
(NOp(UOps.PHI, src=(NOp(UOps.DEFINE_ACC, name="acc"), NOp.var("acc"))), lambda acc: UOp.cast(acc.src[0], acc.dtype)),
(NOp(UOps.PHI, src=(NOp(UOps.DEFINE_ACC, src=(NOp.cvar(),)), NOp.var("x"))), lambda x: x),
(NOp(UOps.PHI, src=(NOp.cvar(), NOp.var("x"))), lambda x: x),
# a DEFINE_ACC without inputs is a const + GEP on a const is the const
(NOp(UOps.DEFINE_ACC, src=(NOp.cvar(),), name="root"), lambda root: UOp.cast(root.src[0], root.dtype)),
(NOp(UOps.GEP, src=(NOp.cvar("x"),), name="root"), lambda root,x: root.const(x.arg)),
# a conditional with the same results either way is a noop, also fold const conditionals
(NOp.var().where(NOp.var("val"), NOp.var("val")), lambda val: val),
(NOp.cvar('gate').where(NOp.var('c0'), NOp.var('c1')), lambda gate, c0, c1: c0 if gate.arg else c1),
# ** constant folding **
(UPat(UOps.ALU, name="root", src=UPat(UOps.CONST)), lambda root: root.const(exec_alu(root.arg, root.dtype, [x.arg for x in root.src]))),
# ** self folding **
(-(-NOp.var('x')), lambda x: x), # -(-x) -> x
(NOp.var('x') + 0, lambda x: x), # x+0 -> x
(NOp.var('x') * 1, lambda x: x), # x*1 -> x
(NOp.var('x') * -1, lambda x: -x), # x*-1 -> -x
(NOp.var('x') // NOp.var('x'), lambda x: x.const(1)), # x//x -> 1
(NOp.var('x') // 1, lambda x: x), # x//1 -> x
(NOp.var('x') // -1, lambda x: -x), # x//-1 -> -x
(NOp.var('x') / NOp.var('x'), lambda x: x.const(1)), # x/x -> 1
(NOp.var('x') / NOp.cvar('c'), lambda x,c: x*exec_alu(UnaryOps.RECIP, c.dtype, [c.arg])), # x/c -> x*(1/c)
# ** zero folding **
# 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
(NOp.var('x') * 0, lambda x: x.const(float('nan') if isinstance(x.arg, float) and (math.isnan(x.arg) or math.isinf(x.arg)) else 0)),
# x-x -> 0
(NOp.var('x') - NOp.var('x'), lambda x: x.const(0)),
(UPat(op=UOps.ALU, name='x'), lambda x: x.const(x.vmin.arg) if x.op is not UOps.CONST and x.vmin.arg == x.vmax.arg else None),
# ** load/store folding **
(NOp.store(NOp.var("buf"), NOp.var("idx"), NOp.load(NOp.var("buf"), NOp.var("idx"))), lambda buf,idx:UOp(UOps.NOOP)),
# ** two stage add/mul folding **
((NOp.var('x') + NOp.cvar('c1')) + NOp.cvar('c2'), lambda x,c1,c2: x+x.const(exec_alu(BinaryOps.ADD, x.dtype, [c1.arg, c2.arg]))),
((NOp.var("x") * NOp.cvar("c1")) * NOp.cvar("c2"), lambda x,c1,c2: x*x.const(exec_alu(BinaryOps.MUL, x.dtype, [c1.arg, c2.arg]))),
# *** rules from symbolic ***
# ** lt **
# c0*x<c1 for positive int c0,c1
((NOp.cvar('c0',dtypes.int)*NOp.var('x')).lt(NOp.cvar('c1',dtypes.int)),
lambda x,c0,c1: x.lt(math.ceil(c1.arg/c0.arg)) if c0.arg > 0 and c1.arg > 0 else None),
# mul add lt
(((NOp.cvar('c0')*NOp.var('x'))+NOp.var('x2')).lt(NOp.cvar('c1')),
lambda x,x2,c0,c1: x.lt(c1.arg//c0.arg) if c1.arg % c0.arg == 0 and c0.arg > x2.vmax.arg and x2.vmin.arg >= 0 else None),
# neg lt -> lt
(NOp.lt(-NOp.var('x'), NOp.cvar('c', dtypes.int)), lambda c,x: UOp.lt(c.const(-c.arg), x)),
# ** div **
# div folding
(NOp.var('x') // NOp.cvar('c'), lambda x,c: x.const(x.vmin.arg//c.arg) if c.arg > 0 and x.vmin.arg//c.arg == x.vmax.arg//c.arg else None),
(NOp.var('x') // NOp.cvar('c'), lambda x,c: d if c.arg > 0 and (d:=x.divides(c.arg)) is not None else None),
# mul div
((NOp.var("x") * NOp.cvar("c0")) // NOp.cvar("c1"),
lambda x,c0,c1: x*(c0.arg//gcd)//(c1.arg//gcd) if c1.arg!=0 and (gcd:=math.gcd(c0.arg,c1.arg))> 1 else None),
# mul add div
(((NOp.cvar('c0')*NOp.var('x'))+NOp.var('x2')) // NOp.cvar('c1'), lambda x,x2,c0,c1:\
x*(c0.arg//g)//(c1.arg//g) if c0.arg > 0 and c1.arg > 0 and (g:=math.gcd(c0.arg,c1.arg)) > 1 and g > x2.vmax.arg and x2.vmin.arg >= 0 else None),
# ** mod **
# apply mod to mod input
(NOp.var('x') % NOp.cvar('c'), lambda x,c: newx%c if 0 < c.arg and (newx:=mod_folding(x,c.arg)) is not None else None),
# remove mod
(NOp.var('x') % NOp.cvar('c'), lambda x,c:\
x-(x.vmin.arg//c.arg)*c.arg if 0 < c.arg and 0 <= x.vmin.arg and x.vmin.arg//c.arg == x.vmax.arg//c.arg else None),
# mul mod
((NOp.cvar('c0')*NOp.var('x')) % NOp.cvar('c1'), lambda x,c0,c1: (x%(c1.arg//c0.arg))*c0 if c1.arg%c0.arg == 0 else None),
# mod mod
((NOp.var('x') % NOp.cvar('c0')) % NOp.cvar('c1'), lambda x,c0,c1: x % c1 if c0.arg % c1.arg == 0 else None),
# ** combine terms **
# -(x+y) -> -x + -y
#(-(NOp.var("x") + NOp.var("y")), lambda x,y: (-x)+(-y)),
# (x*c0)+(x*c1) -> x*(c0+c1)
(NOp.var("x") * NOp.cvar("c0") + NOp.var("x") * NOp.cvar("c1"), lambda x,c0,c1: x*exec_alu(BinaryOps.ADD, x.dtype, [c0.arg, c1.arg])),
# (x*c0)+(y*c0) -> (x+y)*c0
#((NOp.var("x") * NOp.cvar("c0")) + (NOp.var("y") * NOp.cvar("c0")), lambda x,y,c0: c0*(x+y)),
# (x*x2)/x2 -> x
((NOp.var("x") * NOp.var("x2")) / NOp.var("x2"), lambda x,x2: x),
# (x//c0)//c1 -> x//(c0*c1)
((NOp.var("x") // NOp.cvar("c0")) // NOp.cvar("c1"), lambda x,c0,c1: x//x.const(exec_alu(BinaryOps.MUL, x.dtype, [c0.arg, c1.arg]))),
# (x/x1)/x2 -> x/(x1*x2)
((NOp.var("x") / NOp.var("x2")) / NOp.var("x3"), lambda x,x2,x3: x/(x2*x3)),
# c0 + x < c1 -> x < c1 - c0
((NOp.cvar("c0") + NOp.var("x")).lt(NOp.cvar("c1")), lambda x,c0,c1: UOp.lt(x, x.const(exec_alu(BinaryOps.ADD, x.dtype, [c1.arg, -c0.arg])))),
# (x+x*c0)-> x*(c0+1)
(NOp.var("x") + NOp.var("x") * NOp.cvar("c0"), lambda x,c0: x*(c0.arg+1)),
# x!=0 -> (bool)x
(NOp.var("x").ne(0), lambda x: x.cast(dtypes.bool)),
# bool != 1 -> not bool
(NOp.var("x", dtype=dtypes.bool).ne(1), lambda x: -x),
# TODO: can do the invert of this (flip alt/load) when we fix double ops
(NOp.store(NOp.var("buf"), NOp.var("idx"), NOp.var("gate").where(NOp.var("alt"), NOp.load(NOp.var("buf"), NOp.var("idx")))),
lambda buf, idx, gate, alt: UOp.store(buf, idx, alt, gate)),
# VECTORIZE-PHI-GEP -> PHI-VECTORIZE
(NOp(UOps.VECTORIZE, src=tuple(NOp(UOps.PHI, src=(NOp(UOps.GEP, src=(NOp.var("val"),), arg=i), NOp.var(f"v{i}"))) for i in range(4)), name="root"),
lambda root, val, v0, v1, v2, v3: UOp(UOps.PHI, root.dtype, (val, UOp(UOps.VECTORIZE, val.dtype, (v0, v1, v2, v3))))),
(NOp(UOps.VECTORIZE, src=tuple(NOp(UOps.PHI, src=(NOp(UOps.GEP, src=(NOp.var("val"),), arg=i), NOp.var(f"v{i}"))) for i in range(2)), name="root"),
lambda root, val, v0, v1: UOp(UOps.PHI, root.dtype, (val, UOp(UOps.VECTORIZE, val.dtype, (v0, v1))))),
# cast NOOP (NOTE: it's str to deal with PtrDType)
(NOp(UOps.CAST, name="root"), lambda root: root.src[0] if str(root.dtype) == str(root.src[0].dtype) else None),
(NOp(UOps.VECTORIZE, name="root"), lambda root: root.src[0] if str(root.dtype) == str(root.src[0].dtype) else None),
# fold gated LOAD/STORE
(NOp.load(NOp.var("buf"), NOp.var("idx"), NOp.cvar("var"), NOp.const(dtypes.bool, True)), lambda buf,idx,var: UOp.load(buf, idx, dtype=var.dtype)),
(NOp.load(NOp.var("buf"), NOp.var("idx"), NOp.cvar("var"), NOp.const(dtypes.bool, True), NOp.var("barrier")),
lambda buf,idx,var,barrier: UOp.load(buf, idx, barrier, dtype=var.dtype)),
(NOp.load(NOp.var(), NOp.var(), NOp.cvar("var"), NOp.const(dtypes.bool, False)), lambda var: var),
(NOp.load(NOp.var(), NOp.var(), NOp.cvar("var"), NOp.const(dtypes.bool, False), NOp.var()), lambda var: var),
(NOp.store(NOp.var("buf"), NOp.var("idx"), NOp.var("val"), NOp.const(dtypes.bool, True)), UOp.store),
(NOp.store(NOp.var(), NOp.var(), NOp.var(), NOp.const(dtypes.bool, False)), lambda: UOp(UOps.NOOP)),
# remove NOOPs from SINK
(NOp(UOps.SINK, name="root"),
lambda root: UOp(UOps.SINK, root.dtype, a, root.arg) if len(a:=tuple(x for x in root.src if x.op is not UOps.NOOP)) != len(root.src) else None),
# ** move add consts to end (NOTE: this is still happening before constant folding) **
(UPat(UOps.ALU, BinaryOps.ADD, src=(UPat(UOps.CONST, name='c1'), UPat(name='x'))), lambda c1,x: x+c1 if x.op is not UOps.CONST else None),
(UPat(UOps.ALU, BinaryOps.ADD, src=(UPat(UOps.ALU, BinaryOps.ADD, src=(UPat(name='x'), UPat(UOps.CONST, name='c1'))), UPat(name='y'))),
lambda x,c1,y: (x+y)+c1),
])
# *** uop expander ***
def _expand_arg_to_idx(args:Tuple[Tuple[int, int], ...], rpk:Dict[int, int]) -> int:
idx, mul = 0, 1
for axis,m in args[::-1]:
idx += rpk[axis] * mul
mul *= m
return idx
def _choices_from_args(args:Tuple[Tuple[int, int], ...]) -> List[Dict[int, int]]:
return [dict(x) for x in itertools.product(*[zip(itertools.repeat(axis), range(m)) for axis,m in args])]
def do_expand(root:UOp):
if root.op is UOps.REDUCE:
if root.src[0].op is not UOps.EXPAND: return None
reduce_expand_args = flatten([x.arg for x in root.src[1:] if x.op is UOps.EXPAND])
expand_args = tuple(x for x in root.src[0].arg if x not in reduce_expand_args)
if len(expand_args) == 0: return None
dont_expand_args = tuple(x for x in root.src[0].arg if x in reduce_expand_args)
else:
expands = [x for x in root.src if x.op is UOps.EXPAND]
if len(expands) == 0: return None
expand_args = tuple(sorted(dedup(flatten([x.arg for x in expands]))))
if root.op is UOps.WMMA:
# both the reduce and upcast args are not expanded here
dont_expand_args = tuple(x for x in expand_args if x[0] in root.arg[-1] or x[0] in [y[0] for y in flatten(root.arg[-2])])
expand_args = tuple(x for x in expand_args if x not in dont_expand_args)
else:
dont_expand_args = ()
new_srcs: List[UOp] = []
lrpks = _choices_from_args(dont_expand_args)
for rpk in _choices_from_args(expand_args):
new_src: List[UOp] = []
for src in root.src:
if src.op is UOps.EXPAND:
lnew_src = tuple(src.src[_expand_arg_to_idx(src.arg, {**rpk, **lrpk})] for lrpk in lrpks)
# TODO: is this right for UOps.WMMA? when there's more than one, all lnew_src should be the same
new_src.append(lnew_src[0] if len(lnew_src) == 1 or root.op is UOps.WMMA else UOp(UOps.EXPAND, root.dtype, lnew_src, dont_expand_args))
else:
new_src.append(src)
new_srcs.append(UOp(root.op, root.dtype, tuple(new_src), root.arg))
if root.op is UOps.EXPAND:
# merge two expands
expand_args, old_args = tuple(sorted(root.arg+expand_args)), expand_args
assert len(expand_args) == (len(old_args) + len(root.arg))
new_srcs = [new_srcs[_expand_arg_to_idx(old_args, rpk)].src[_expand_arg_to_idx(root.arg, rpk)] for rpk in _choices_from_args(expand_args)]
if root.op is UOps.IF:
# merge ifs into an or
conditions = functools.reduce(lambda x,y: x|y, dedup(x.src[0] for x in new_srcs if x.src[0].op is not UOps.CONST))
barriers = tuple(set(x.src[1] for x in new_srcs))
new_srcs = [UOp(UOps.IF, src=(conditions,)+barriers) for _ in new_srcs]
assert prod([x[1] for x in expand_args]) == len(new_srcs)
return UOp(UOps.EXPAND, root.dtype, tuple(new_srcs), expand_args)
acc_number = 0
def do_reduce_with_expand(root):
global acc_number
expands = [x for x in root.src[1:] if x.op is UOps.EXPAND]
expands_reduce = [x for x in expands if root.src[0].op is UOps.EXPAND and all(y in root.src[0].arg for y in x.arg)]
expands_non_reduce = [x for x in expands if x not in expands_reduce]
const = UOp.const(root.dtype.scalar(), 0 if root.arg is ReduceOps.SUM else dtypes.min(root.dtype))
ret = acc = UOp(UOps.DEFINE_ACC, root.dtype, (const,) + tuple(x for x in root.src[1:] if x.op is not UOps.EXPAND), (acc_number,))
acc_number += 1
alu_op = {ReduceOps.SUM:BinaryOps.ADD, ReduceOps.MAX:BinaryOps.MAX}[cast(ReduceOps, root.arg)]
if len(expands_reduce):
assert root.src[0].op is UOps.EXPAND
expand_reduce_args = dedup(flatten([x.arg for x in expands_reduce]))
assert prod([y[1] for y in expand_reduce_args]) == len(root.src[0].src)
ret = functools.reduce(lambda x,y: x.alu(alu_op, y), (ret,)+root.src[0].src)
else:
ret = ret.alu(alu_op, root.src[0])
ret = UOp(UOps.PHI, ret.dtype, (acc, ret))
if len(expands_non_reduce): ret = ret * prod([sz for _,sz in flatten([x.arg for x in expands_non_reduce])])
return ret
def do_contract(con:UOp):
ex = con.src[0]
assert con.dtype is not None
# CONTRACT without EXPAND repeats the element VECTORIZED
if ex.op is not UOps.EXPAND or not all(x in ex.arg for x in con.arg):
assert ex.op is not UOps.EXPAND or not any(x in ex.arg for x in con.arg), "partial contract not supported"
return UOp(UOps.VECTORIZE, con.dtype, con.src*con.dtype.count)
# CONTRACT may remove several axes from EXPAND
assert con.dtype.count == prod([x[1] for x in con.arg]), "dtype is wrong"
srcs = []
for rpk in _choices_from_args(new_ex_args:=tuple(x for x in ex.arg if x not in con.arg)):
lsrcs = [ex.src[_expand_arg_to_idx(ex.arg, {**rpk, **lrpk})] for lrpk in _choices_from_args(con.arg)]
srcs.append(UOp(UOps.VECTORIZE, con.dtype, tuple(lsrcs)))
return srcs[0] if len(srcs) == 1 else UOp(UOps.EXPAND, con.dtype, tuple(srcs), new_ex_args)
def no_vectorized_alu(alu):
if alu.dtype.count == 1: return None
alus = tuple(UOp(alu.op, alu.dtype.scalar(),
tuple(UOp(UOps.GEP, s.dtype.scalar(), (s,), i) for s in alu.src), alu.arg) for i in range(alu.dtype.count))
return UOp(UOps.VECTORIZE, alu.dtype, alus)
def create_gate(root:UOp) -> Optional[UOp]:
@functools.lru_cache(None)
def _gate_srcs(u:UOp, gate:UOp) -> UOp:
if u.op is UOps.LOAD and u.src[-1].op is UOps.BARRIER: return UOp(u.op, u.dtype, u.src[:-1]+(UOp(UOps.IF, None, (gate, u.src[-1])),), u.arg)
return u if (replace_source:=tuple(_gate_srcs(x, gate) for x in u.src)) == u.src else UOp(u.op, u.dtype, replace_source, u.arg)
return None if len(root.src) == 3 or (ret:=_gate_srcs(root, root.src[3])) is root else ret
expander = PatternMatcher([
# create gate MUST BE BEFORE expander
(NOp(UOps.STORE, name="root"), create_gate),
# do expansion
(UPat({UOps.ALU, UOps.CAST, UOps.BITCAST, UOps.GEP, UOps.WMMA, UOps.LOAD, UOps.STORE,
UOps.VECTORIZE, UOps.REDUCE, UOps.EXPAND, UOps.IF}, name="root"), do_expand),
(NOp(UOps.CONTRACT, name="con"), do_contract),
# remove EXPANDs from SINK
(NOp(UOps.SINK, name="root"),
lambda root: UOp(UOps.SINK, root.dtype, a, root.arg)
if len(a:=tuple(flatten(x.src if x.op is UOps.EXPAND else (x,) for x in root.src))) != len(root.src) else None),
# BARRIERs aren't actually expanded
(NOp(UOps.BARRIER, src=(NOp(UOps.EXPAND, name="ex"),)), lambda ex: UOp(UOps.EXPAND, None, (UOp(UOps.BARRIER, None, ex.src),)*len(ex.src), ex.arg)),
# empty EXPAND is NOOP
(NOp(UOps.EXPAND, src=(NOp.var('x'),), arg=()), lambda x: x),
])
def delete_redundant_gates(root:UOp) -> Optional[UOp]:
@functools.lru_cache(None)
def find_gate(x:UOp) -> Optional[UOp]:
if x.op is UOps.IF: return x
return next((ret for s in x.src if (ret:=find_gate(s)) is not None), None)
if len(root.src) == 3 or (gate:=find_gate(root)) is None or gate.src[0] is not root.src[3]: return None
return UOp(UOps.STORE, root.dtype, root.src[:3], root.arg)
reducer = PatternMatcher([
(NOp(UOps.REDUCE, name="root"), do_reduce_with_expand),
# no ALU on vectorized dtypes
(UPat({UOps.ALU, UOps.CAST, UOps.BITCAST}, name="alu"), no_vectorized_alu),
# VECTORIZE a CONST is a CONST (eventually remove this rule)
(UPat(UOps.VECTORIZE, name="root", src=UPat(UOps.CONST, name="c")), lambda root, c: root.const(c.arg)),
# delete_redundant_gates (after expand, is this still needed?)
(NOp(UOps.STORE, name="root"), delete_redundant_gates),
])
# *** uop graph ***
def get_children_dfs(u:UOp, children:Dict[UOp, List[UOp]], in_degree:Dict[UOp, int]):
if u in children: return
children[u] = []
for x in u.src:
get_children_dfs(x, children, in_degree)
children[x].append(u)
in_degree[u] = len(u.src)
def graph_rewrite(sink:UOp, pm:PatternMatcher) -> UOp:
nodes: Dict[Tuple, UOp] = {}
replace: Dict[UOp, UOp] = {}
def __inner_rewrite(n:UOp) -> UOp:
if n in replace: return replace[n]
replace_source = (n.op, n.dtype, tuple(__inner_rewrite(y) for y in n.src), n.arg)
if found := nodes.get(replace_source): replace[n] = found
else: nodes[replace_source] = replace[n] = found = __inner_rewrite(new_x) if (new_x := pm.rewrite(x:=UOp(*replace_source))) else x
return found
return __inner_rewrite(sink)
class UOpGraph:
def __init__(self, sink:Union[UOp, List[UOp]], opts:Optional[Renderer]=None):
self.sink: UOp = sink if isinstance(sink, UOp) else UOp(UOps.SINK, None, tuple(sink))
assert self.sink.op is UOps.SINK, f"sink isn't sink, it's {self.sink.op}"
# used by linearizer
self._uops: Optional[List[UOp]] = None
self.opts = opts
self.folder = constant_folder
if TRANSCENDENTAL >= 2 or (opts is not None and TRANSCENDENTAL >= 1 and opts.device in {"CLANG", "LLVM"}):
self.folder = self.folder + transcendental_folding
def __reduce__(self): return self.__class__, (self.sink, self.opts)
def __iter__(self) -> Iterator[UOp]: return iter(self.uops)
def __getitem__(self, index) -> UOp: return self.uops[index]
@property
def uops(self) -> List[UOp]:
if self._uops is None: self.linearize()
return cast(List[UOp], self._uops)
def graph(self):
from tinygrad.engine.graph import graph_uops
graph_uops(self.uops)
def print(self):
for i,u in enumerate(self):
formatted_parents = [self.uops.index(x) if x.op is not UOps.CONST else f"{x.arg}" for x in u.src]
print(f"{i:4d} {str(u.op):20s}: {str(u.dtype) if u.dtype is not None else '':25s} " f"{str(formatted_parents):32s} {u.arg}")
cnt = 0
def linearize(self, extra_pm:Optional[PatternMatcher]=None, skip_check=False) -> UOpGraph:
global acc_number
acc_number = 0
# NOTE: relinearizering should be okay
#assert self._uops is None, "already linearized"
# do graph rewrite
sink = graph_rewrite(self.sink, self.folder)
# expand
UOpGraph.cnt += 1
if UOpGraph.cnt != getenv("DEBUG_EXPAND", 0):
sink = graph_rewrite(sink, self.folder+expander+float4_folding if self.opts is not None and self.opts.supports_float4 else self.folder+expander)
sink = graph_rewrite(sink, self.folder+expander+reducer)
# for PTX only
if extra_pm: sink = graph_rewrite(sink, self.folder+extra_pm)
# filter nodes that don't link to a sink
# BFS toposort
children: Dict[UOp, List[UOp]] = {}
in_degree: Dict[UOp, int] = {}
get_children_dfs(sink, children, in_degree)
@functools.lru_cache(None)
def get_recursive_children(x:UOp, end:UOps, include_self=False) -> Set[UOp]:
if x.op is UOps.SINK: return set()
return set.union({x} if include_self else set(), *([get_recursive_children(u, end, True) for u in children[x] if x.op is not end]))
# scope children impact the toposort and END* insertion
scope_children = {p:get_recursive_children(p, END_FOR_UOP[p.op][0]) for p in reversed(in_degree) if p.op in END_FOR_UOP}
queue:List[Tuple[int, UOp]] = []
def push(u:UOp):
priority = 0
# prefer uops that are loop children
for l, ss in scope_children.items():
if l.op is UOps.RANGE and u in ss: priority -= l.arg[0]*1000 + l.arg[1]
heapq.heappush(queue, (priority, u))
for u in children:
if in_degree[u] == 0: push(u)
scope_end: Dict[UOp, UOp] = {}
self._uops = []
while queue:
p,x = heapq.heappop(queue)
if DEBUG >= 7: print(p,x)
if x in scope_children: scope_end[x] = x
if x.op is UOps.DEFINE_ACC:
idx = min([self._uops.index(l) for l in x.src if l.op is UOps.RANGE])
self._uops.insert(idx, x)
else: self._uops.append(x)
for u, ss in scope_children.items():
if x in ss:
ss.remove(x)
if len(ss) == 0: scope_end[u] = x
for u in children[x]:
in_degree[u] -= 1
if in_degree[u] == 0: push(u)
# end scopes in toposort order
for u, x in scope_end.items(): self._uops.insert(self._uops.index(x)+1, UOp(END_FOR_UOP[u.op][1], None, (u,)))
# sanity checks (NOTE: these can cause things to be skipped in BEAM)
if not skip_check:
bad_ops = dedup([x.op for x in self._uops if x.op in {UOps.EXPAND, UOps.CONTRACT, UOps.REDUCE}])
try:
type_verify(self.uops)
assert self._uops[-1].op is UOps.SINK, f"didn't end with SINK, ended with {self._uops[-1]}"
assert len(bad_ops) == 0, f"bad UOps left in list: {bad_ops}"
# TODO: this should be enabled, and the valid clause should be removed
# NOTE: multiple identical stores to DEFINE_LOCAL is okay
assert len(all_stores := [x.src[0:2]+x.src[3:] for x in self._uops if x.op is UOps.STORE and x.src[0].op is not UOps.DEFINE_LOCAL]) \
== len(dedup(all_stores)), "repeated stores in uops"
except AssertionError as e:
self.print()
if not CI: self.graph()
raise e
# strip the SINK
self._uops = self._uops[:-1]
return self
Reference in New Issue View Git Blame Copy Permalink