Files
tinygrad/tinygrad/schedule/rangeify.py
T
2025-08-15 15:44:31 -07:00

376 lines
16 KiB
Python

from typing import Any
from dataclasses import dataclass, field
from tinygrad.dtype import dtypes, AddrSpace, PtrDType
from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, RewriteNotReady
from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv
from tinygrad.schedule.kernelize import Kernel
from tinygrad.uop.ops import track_rewrites, graph_rewrite_map, graph_rewrite, KernelInfo, identity_element
imported_rewrites = PatternMatcher([
# UOp with size 0 is zero
(UPat(GroupOp.All-{Ops.SINK}, name="root"), lambda root: root.const_like(0) if root.base.st is not None and root.size == 0 else None),
# DETACH and CONTIGUOUS_BACKWARD are NOOPs here
(UPat((Ops.DETACH, Ops.CONTIGUOUS_BACKWARD), name="x"), lambda x: x.src[0]),
# reduce of size 0 is the identity element
(UPat(Ops.REDUCE_AXIS, name="reduce", src=(UPat.var("x"),)),
lambda reduce,x: reduce.const_like(identity_element(reduce.arg[0], reduce.dtype)) if x.size == 0 and reduce.size != 0 else None),
])
earliest_rewrites = imported_rewrites+PatternMatcher([
# RESHAPE on RESHAPE is the second reshape
(UPat(Ops.RESHAPE, src=(UPat(Ops.RESHAPE),), name="x"), lambda x: x.replace(src=(x.src[0].src[0],))),
# non shape changing RESHAPE is NOOP
(UPat(Ops.RESHAPE, name="x"), lambda x: x.src[0] if x.src[0].shape == x.arg else None),
# RESHAPE after COPY
(UPat(Ops.COPY, src=(UPat(Ops.RESHAPE, name="r"),UPat(name="d")), name="c"), lambda c,r,d: c.replace(src=(r.src[0],d)).reshape(r.arg)),
# TODO: this should be BUFFER_VIEW
(UPat(Ops.COPY, src=(UPat(Ops.SHRINK, name="r"),UPat(name="d")), name="c"), lambda c,r,d: c.replace(src=(r.src[0],d)).shrink(r.arg)),
# const hacks
(UPat(Ops.CONST, name="x"), lambda x:
x.replace(src=(x.src[0].src[0],)).reshape((1,)*len(x.shape)).expand(x.shape) if \
len(x.src) and x.src[0].op is Ops.VIEW and not any(s == 0 for s in x.shape) else None),
])
# 1. add contiguous where we have to
ALWAYS_CONTIGUOUS: set[Ops] = {Ops.CONTIGUOUS, Ops.ASSIGN, Ops.COPY, Ops.BUFFER, Ops.BUFFER_VIEW,
Ops.CONST, Ops.BIND, Ops.DEVICE, Ops.MSELECT, Ops.MSTACK, Ops.DEFINE_GLOBAL,
Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.LOAD}
def realize(ctx:dict[UOp, None], tr:UOp) -> None: ctx[tr] = None
def realize_parents(ctx:dict[UOp, None], rb:UOp) -> None:
for s in rb.src:
if s.op not in ALWAYS_CONTIGUOUS: ctx[s] = None
do_realize = PatternMatcher([
# always realize SINK parents
(UPat(Ops.SINK, name="s"), lambda ctx,s: ctx.update((x.base, None) for x in s.src if x.base.op not in ALWAYS_CONTIGUOUS)),
# always realize ASSIGN/CONTIGUOUS/COPY/BUFFER_VIEW
(UPat({Ops.ASSIGN, Ops.CONTIGUOUS, Ops.COPY, Ops.BUFFER_VIEW}, name="tr"), realize),
# realize parents of COPY, MSELECT, MSTACK
(UPat((Ops.COPY, Ops.MSELECT, Ops.MSTACK), name="rb"), realize_parents),
])
add_contiguous = PatternMatcher([(UPat(GroupOp.All-{Ops.CONTIGUOUS, Ops.ASSIGN}, name="x"),
lambda ctx,x: x.replace(tag=1).contiguous() if x in ctx and x.tag is None else None)])
remove_tags = PatternMatcher([(UPat(GroupOp.All, name="x"), lambda x: x.replace(tag=None) if x.tag is not None else None)])
early_cleanups = PatternMatcher([(UPat().contiguous(name="c").contiguous(), lambda c: c),])
# 2. mark all children
@dataclass
class ChildrenContext: children: dict[UOp, list[UOp]]|None = None
def extract_children(ctx:ChildrenContext, x:UOp):
if ctx.children is not None: return
# REDUCE_AXIS is fine here, should go to contig only (gate)
ctx.children = {k:list(v.keys()) for k,v in x.get_children_map().items() if len(v) > 1 and any(x.op is Ops.REDUCE_AXIS for x in k.toposort())}
def mark_children(ctx:ChildrenContext, x:UOp):
new_srcs = [(UOp(Ops.CHILD, s.dtype, src=(s,), arg=(ctx.children[s].index(x), len(ctx.children[s]))) if s in ctx.children else s) for s in x.src]
return x.replace(src=tuple(new_srcs))
pm_children = PatternMatcher([
(UPat(Ops.SINK, name="x"), extract_children),
(UPat(GroupOp.All-{Ops.CHILD}, name="x"), mark_children),
])
# 3. rangeify
@dataclass
class RangeifyContext:
idx: int = 0
regs: int = 0
seen_children: dict[UOp, dict[int, UOp]] = field(default_factory=dict)
seen_child: dict[UOp, Any] = field(default_factory=dict)
progress_children: dict[UOp, int] = field(default_factory=dict)
def map_reshape(idx:UOp, r:UOp):
acc = 1
to_sum = []
for s,src in list(zip(idx.shape, idx.src[1:]))[::-1]:
to_sum.append(acc*src)
acc *= s
mish = sum(to_sum)
ret = []
for s in r.src[0].shape[::-1]:
if resolve(s!=1):
# this MOD should limit any ranges outside s
ret.append(mish % s)
mish //= s
else:
ret.append(UOp.const(dtypes.int, 0))
ret = UOp.sink(*ret).simplify().src[::-1] if len(ret) else ()
return r.src[0].index(*ret, dtype=idx.dtype, arg=idx.arg)
def map_pad(idx:UOp, r:UOp):
ret = list(idx.src[1:])
bigwhere = UOp.const(dtypes.bool, True)
for i,(sh,(s,e)) in enumerate(zip(r.shape, r.arg)):
if s == 0 and e == 0: continue
where = UOp.const(dtypes.bool, True)
if e > 0: where = where & (ret[i] < (sh-e))
if s > 0: where = where & (ret[i] >= s)
bigwhere = bigwhere & where
# this is safe but dumb
ret[i] = (ret[i] - s).maximum(0).minimum(r.src[0].shape[i]-1)
# PAD is with 0
return bigwhere.simplify().where(r.src[0].index(*ret, dtype=idx.dtype, arg=idx.arg), UOp.const(r.dtype, 0))
def map_expand(r:UOp, idx:UOp):
new_rngs = []
ending_ranges = []
non_ending_ranges = []
for a,x,y in zip(idx.src[1:], r.src[0].shape, r.shape):
axis_to_range = [u for u in a.toposort() if u.op is Ops.RANGE]
if resolve(x!=y, False):
ending_ranges.extend(axis_to_range)
new_rngs.append(a.const_like(0))
else:
non_ending_ranges.extend(axis_to_range)
new_rngs.append(a)
ending_ranges = [x.arg for x in ending_ranges if x not in non_ending_ranges]
if idx.arg is not None: ending_ranges.append(idx.arg)
return r.src[0].index(*new_rngs, arg=min([x for x in ending_ranges]) if ending_ranges else None)
pm_mops = PatternMatcher([
# this is like the definitions of these
(UPat(Ops.INDEX, src=(UPat(Ops.SHRINK, name="r"),), allow_any_len=True, name="idx"),
lambda r,idx: r.src[0].index(*[a+ss if resolve(ss != 0) else a for a,(ss,_) in zip(idx.src[1:], r.arg)], dtype=idx.dtype, arg=idx.arg)),
(UPat(Ops.INDEX, src=(UPat(Ops.PERMUTE, name="r"),), allow_any_len=True, name="idx"),
lambda r,idx: r.src[0].index(*[idx.src[1+p] for p in argsort(idx.src[0].arg)], dtype=idx.dtype, arg=idx.arg)),
(UPat(Ops.INDEX, src=(UPat(Ops.FLIP, name="r"),), allow_any_len=True, name="idx"),
lambda r,idx: r.src[0].index(*[((s-1)-a) if f else a for a,s,f in zip(idx.src[1:], r.shape, r.arg)], dtype=idx.dtype, arg=idx.arg)),
# expand needs to end ranges
(UPat(Ops.INDEX, src=(UPat(Ops.EXPAND, name="r"),), allow_any_len=True, name="idx"), map_expand),
# reshape does a lot of symbolic stuff
(UPat(Ops.INDEX, src=(UPat(Ops.RESHAPE, name="r"),), allow_any_len=True, name="idx"), map_reshape),
# pad adds min and max
(UPat(Ops.INDEX, src=(UPat(Ops.PAD, name="r"),), allow_any_len=True, name="idx"), map_pad),
])
def map_contiguous(ctx:RangeifyContext, x:UOp, idx:UOp|None=None):
if x.arg is None and idx is not None: return None
if x.arg is not None and idx is None: return None
ranges = []
new_ranges = []
passthrough_idx = []
for i,s in enumerate(x.shape):
if x.arg is not None and i not in x.arg:
assert idx is not None, "partial contig requires index"
ranges.append(idx.src[1+i])
continue
if idx is not None: passthrough_idx.append(idx.src[1+i])
if resolve(s!=1):
ranges.append(UOp.range(dtypes.int, s, ctx.idx))
new_ranges.append(ranges[-1])
ctx.idx += 1
else:
ranges.append(UOp.const(dtypes.int, 0))
ret = x.src[0].index(*ranges).bufferize(*new_ranges, arg=x.device)
ret = ret.index(*passthrough_idx) if len(passthrough_idx) else ret.reshape(x.shape)
return ret
def map_reduce(ctx:RangeifyContext, idx:UOp, red:UOp):
rngs = list(idx.src[1:])
new_ranges = []
for i,s in enumerate(red.src[0].shape):
if i in red.arg[1]:
rngs[i] = UOp.range(dtypes.int, s, ctx.idx)
ctx.idx += 1
new_ranges.append(rngs[i])
return UOp(Ops.REDUCE, red.dtype, src=(red.src[0].index(*rngs),)+tuple(new_ranges), arg=red.arg[0])
def index_child(ctx:RangeifyContext, c:UOp, x:UOp, idx:UOp):
if c not in ctx.seen_children:
ctx.seen_children[c] = {}
ctx.progress_children[c] = 0
ctx.seen_children[c][x.arg[0]] = idx
# wait here until we have seen all the children
if len(ctx.seen_children[c]) != x.arg[1]:
if ctx.progress_children[c] == len(ctx.seen_children[c]): raise RuntimeError("revisited child before visiting all children")
ctx.progress_children[c] = len(ctx.seen_children[c])
raise RewriteNotReady
if c not in ctx.seen_child:
all_rngs = zip(*[ch.src[1:] for ch in ctx.seen_children[c].values()])
out_rngs = []
end_ranges = []
idx_ranges = []
for i,r in enumerate(all_rngs):
if all_same(r):
out_rngs.append(r[0])
else:
out_rngs.append(UOp.range(dtypes.int, c.shape[i], ctx.idx))
ctx.idx += 1
end_ranges.append(out_rngs[-1])
idx_ranges.append(i)
ctx.seen_child[c] = (idx_ranges, end_ranges)
else:
out_rngs = list(idx.src[1:])
idx_ranges, end_ranges = ctx.seen_child[c]
for i,nr in zip(idx_ranges, end_ranges): out_rngs[i] = nr
if len(idx_ranges) == 0: return c.index(*out_rngs)
return c.index(*out_rngs).bufferize(*end_ranges, arg=x.device).index(*[idx.src[1+i] for i in idx_ranges])
def might_end_axis(idx:UOp):
if idx.arg is None: return None
to_end_axis = []
for i,a in enumerate(idx.src[1:]):
if any(x.arg > idx.arg for x in a.toposort() if x.op is Ops.RANGE):
to_end_axis.append(i)
if to_end_axis: return idx.replace(src=(idx.src[0].contiguous(arg=tuple(to_end_axis)),)+idx.src[1:], arg=None)
return idx.replace(arg=None)
pm_rangeify = pm_mops+PatternMatcher([
# if there are new ended children, tag the SINK
(UPat(Ops.INDEX, src=(UPat(Ops.CHILD, src=(UPat(name="c"), ), name="x"),), allow_any_len=True, name="idx"), index_child),
# if there's an INDEX it can support partial contig
(UPat(Ops.INDEX, src=(UPat(Ops.CONTIGUOUS, src=(UPat(),), name="x"),), allow_any_len=True, name="idx"), map_contiguous),
# CONST can't have axes. remove srcs when we idx
(UPat(Ops.INDEX, src=(UPat(Ops.CONST, name="c"),)), lambda c: c.replace(src=())),
# sink contigs to kick it off
(UPat(Ops.CONTIGUOUS, src=(UPat(),), name="x"), lambda ctx,x: map_contiguous(ctx, x)),
# handle arg on any op with weight. old endrange stuff
(UPat(Ops.INDEX,src=(UPat(GroupOp.Elementwise.union({Ops.REDUCE_AXIS})),), allow_any_len=True, name="idx"), might_end_axis),
# move MAP through elementwise ALU / reduce. these are the items with cost
(UPat(Ops.INDEX, src=(UPat(GroupOp.Elementwise.union({Ops.STORE, Ops.ASSIGN, Ops.COPY, Ops.DEVICE})),), allow_any_len=True, name="x"),
lambda x: x.src[0].replace(src=tuple([s.index(*x.src[1:]) for s in x.src[0].src]))),
(UPat(Ops.INDEX, src=(UPat(Ops.REDUCE_AXIS, name="red"),), allow_any_len=True, name="idx"), map_reduce),
# CONTIGUOUS on ASSIGN is STORE
# TODO: tag in UPat?
(UPat(Ops.CONTIGUOUS, src=(UPat(Ops.ASSIGN, name="a"),), name="c", allow_any_len=True),
lambda c,a: UOp(Ops.STORE, src=a.src+c.src[1:]) if c.tag == 1 else None),
])
# 4. remove bufferize
def bufferize_to_store(x:UOp):
rngs = x.src[1:]
shape = tuple([r.vmax+1 for r in rngs])
assert prod(shape) > 0, f"no zero sized buffers {shape}"
buf = UOp.new_buffer(x.arg, prod(shape), x.dtype)
return buf.reshape(shape).index(*rngs, dtype=x.dtype.ptr(size=prod(shape))).store(x.src[0], *rngs)
def add_load_on_buffer(x:UOp, b:UOp):
if isinstance(x.dtype, PtrDType): return None
return x.replace(dtype=x.dtype.ptr(b.size)).load()
def add_load_on_store(x:UOp, st:UOp):
if isinstance(x.dtype, PtrDType): return None
rngs = x.src[1:]
shape = tuple([r.vmax+1 for r in rngs])
b = st.src[0].src[0]
assert b.op is Ops.BUFFER
return b.shrink(((0,prod(shape)),)).reshape(shape).index(*rngs, dtype=x.dtype.ptr(size=b.size)).load(st)
pm_add_buffers = pm_mops+PatternMatcher([
(UPat(Ops.BUFFERIZE, name="x"), bufferize_to_store),
(UPat(Ops.INDEX, src=(UPat(Ops.BUFFER, name="b"), UPat()), name="x"), add_load_on_buffer),
(UPat(Ops.INDEX, src=(UPat(Ops.STORE, name="st"),), allow_any_len=True, name="x"), add_load_on_store),
])
# 5 (alt). create pointers
def debuf(ctx, b:UOp):
ret = UOp(Ops.DEFINE_GLOBAL, b.dtype.ptr(b.arg), arg=ctx[0])
ctx[0] += 1
return ret
pm_debuf = PatternMatcher([
(UPat(Ops.BUFFER, name="b"), debuf),
# HACK: consts shouldn't have srcs by here
(UPat(Ops.CONST, name="x"), lambda x: x.replace(src=()) if len(x.src) else None),
# no movement ops
(UPat(GroupOp.Movement, name="x"), lambda x: x.src[0]),
# HACK: no copy
(UPat(Ops.COPY, name="x"), lambda x: x.src[0]),
])
# 5. split into kernels
@dataclass
class LocalAddBufferContext:
dg:int = 0
map:dict = field(default_factory=dict)
def debuf(ctx:LocalAddBufferContext, b:UOp): return UOp(Ops.DEFINE_GLOBAL, b.dtype.ptr(b.arg), arg=ctx.map[b][1])
def split_load(ctx:LocalAddBufferContext, s:UOp):
b = s.src[0].src[0]
if b.op is not Ops.BUFFER: return None
if len(s.src) == 2 and s.src[1].op is Ops.ASSIGN:
assert len(s.src) == 2
lb = s.src[1]
assert b not in ctx.map or ctx.map[b][0] == lb
else:
lb = b
if b not in ctx.map:
ctx.map[b] = (lb, ctx.dg)
ctx.dg += 1
return s.replace(src=s.src[0:1]) if b is not lb else None
def handle_store(ctx:LocalAddBufferContext, s:UOp):
b = s.src[0].src[0]
if b.op is not Ops.BUFFER: return None
if b not in ctx.map:
ctx.map[b] = (b, ctx.dg)
ctx.dg += 1
if s.src[1].op is Ops.COPY: return s.src[1]
return None
to_define_global = PatternMatcher([
(UPat(Ops.BUFFER, name="b"), debuf),
(UPat(Ops.LOAD, name="s"), split_load),
(UPat(Ops.STORE, name="s"), handle_store),
])
def split_store(x:UOp):
if len(x.ranges): return None
shape = tuple([r.vmax+1 for r in x.src[2:]])
name = "k_"+'_'.join([str(s) for s in shape])
ctx = LocalAddBufferContext()
ret = graph_rewrite(x, to_define_global, ctx=ctx, name="kernel split", bottom_up=True)
ret = ret.sink(arg=KernelInfo(name=name)) if ret.op is Ops.STORE else ret
kernel = UOp(Ops.KERNEL, src=tuple([x[0] for x in ctx.map.values()]), arg=Kernel(ret, ()))
return kernel.src[0].assign(kernel)
split_kernels = PatternMatcher([
(UPat(Ops.STORE, name="x"), split_store),
])
@track_rewrites(name=lambda sink,ret: f"Schedule {pluralize('Kernel',len([u for u in ret[sink].toposort() if u.op is Ops.KERNEL]))}", replay=True)
def get_kernelize_map(sink:UOp) -> dict[UOp, UOp]:
tensor_map = graph_rewrite_map(sink, earliest_rewrites, name="earliest")
realize_map = {}
graph_rewrite(tensor_map[sink], do_realize, ctx=realize_map, name="Input Graph")
tensor_map = graph_rewrite_map(tensor_map[sink], add_contiguous, ctx=realize_map, bottom_up=True, input_map=tensor_map, name="add contiguous")
tensor_map = graph_rewrite_map(tensor_map[sink], early_cleanups+remove_tags, input_map=tensor_map, name="cleanup")
tensor_map = graph_rewrite_map(tensor_map[sink], pm_children, ctx=ChildrenContext(), bottom_up=True, input_map=tensor_map, name="children")
tensor_map = graph_rewrite_map(tensor_map[sink], pm_rangeify, ctx=RangeifyContext(), bottom_up=True, input_map=tensor_map, name="rangeify")
tensor_map = graph_rewrite_map(tensor_map[sink], pm_add_buffers, bottom_up=True, input_map=tensor_map, name="add buffers")
if getenv("VIZ"): graph_rewrite(tensor_map[sink], PatternMatcher([]), name="View Rangeify Graph")
# render
if getenv("SRC"):
rsink = tensor_map[sink]
from tinygrad.codegen.devectorizer import pm_reduce, ReduceContext
rsink = graph_rewrite(rsink, pm_reduce, ctx=ReduceContext(), name="remove reduce")
rsink = graph_rewrite(rsink, pm_debuf, ctx=[0], name="debuf", bottom_up=True)
from tinygrad.codegen import rewrites_for_linearizer, apply_rewrites
rsink = apply_rewrites(rsink, rewrites_for_linearizer)
from tinygrad.renderer.cstyle import CStyleLanguage
src = CStyleLanguage().render(rsink.arg.lst)
print(src)
return {sink:sink}
tensor_map = graph_rewrite_map(tensor_map[sink], split_kernels, input_map=tensor_map, name="split kernels")
if getenv("VIZ"): graph_rewrite(tensor_map[sink], PatternMatcher([]), name="View Kernel Graph")
return tensor_map