nimlgen/tinygrad
1
0
Fork 0
forked from tinygrad/tinygrad
Code Pull Requests Activity

move NOOPs to pre_isel_matcher and rm NOOP from spec

Browse Source
This commit is contained in:
ttomsa
2026-02-10 17:22:21 +00:00
parent 86b5441781
commit ce31a4fbec
2 changed files with 21 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tinygrad/renderer/isa/x86.py
+19 -18
View File
@@ -16,17 +16,8 @@ extra_matcher = PatternMatcher([
(UPat.var('x', dtypes.bool).alu(Ops.CMPEQ, UPat.var('y')), lambda x,y: (x^y)^True),
(UPat.var('x', dtypes.bool)<UPat.var('y'), lambda x,y: (x^True)&y),
# *** NOOP ***
# cast to/from pointer is a noop
# cast to pointer is a noop
(UPat.var("y").cast(name="x"), lambda y,x: y if isinstance(x.dtype, PtrDType) or y.dtype == dtypes.void else None),
(UPat.var("y").cast(name="x"), lambda y,x: x.replace(op=Ops.NOOP) if isinstance(y.dtype, PtrDType) else None),
# zero extending scalar 32bit int is a noop
(UPat.var("y", dtypes.uint32).cast(dtypes.int64s, name="x"), lambda y,x: x.replace(op=Ops.NOOP) if y.dtype.count == 1 else None),
# cast between signed and unsigned int is a noop
(UPat.var("y", dtypes.ints+(dtypes.bool,)).cast(dtypes.ints, name="x"),
lambda y,x: x.replace(op=Ops.NOOP) if x.dtype.itemsize == y.dtype.itemsize else None),
# bitcasts between scalar float and scalar int are real, rest are noops
(UPat.var("y").bitcast().named("x"), lambda y,x: None if (y.dtype in dtypes.floats and x.dtype in dtypes.ints) or \
(y.dtype in dtypes.ints and x.dtype in dtypes.floats) else x.replace(op=Ops.NOOP)),
# TODO: this should be removed when bool cast is canonicalized
# rewrite cast to bool to CMPNE 0
(UPat.var("y").cast(dtypes.bool), lambda y: y != y.const_like(0)),
@@ -59,26 +50,36 @@ extra_matcher = PatternMatcher([
# no cmpne for packed ints, y != x => !(y==x)
(UPat(Ops.CMPNE, src=(UPat.var("y", dtypes.ints), UPat.var("x")), name="cmp"),
lambda y,x,cmp: UOp(Ops.CMPEQ, cmp.dtype, (y,x))^True if y.dtype.count > 1 else None),
# noop of a noop is removed
(UPat(Ops.NOOP, src=(UPat(Ops.NOOP),), name="x"), lambda x: x.replace(src=x.src[0].src)),
# cast to < scalar int is a noop
(UPat.var("y", dtypes.ints).cast(dtypes.ints, name="x"),
lambda y,x: x.replace(op=Ops.NOOP) if x.dtype.itemsize < y.dtype.itemsize and y.dtype.count == 1 else None),
# float where expects a mask TODO: handle float64 cmp to float32 where
(UPat.var("m", dtypes.bool).where(UPat.var("a", dtypes.floats), UPat.var("b")),
lambda m,a,b: m.cast(a.dtype).ne(0).where(a, b) if m.src[0].dtype not in dtypes.floats else None),
# TODO: do we want this? Kinda not needed if DEVECTORIZE=0. If yes make it general
(UPat(Ops.VECTORIZE, dtypes.float16, name="x"), lambda x: x.replace(dtype=dtypes.float32.vec(x.dtype.count),
src=tuple(s.src[0] for s in x.src)).cast(x.dtype) if all(s.op is Ops.CAST for s in x.src) else None),
# moving elements of a single register to another without shuffling is a noop
(UPat(Ops.VECTORIZE, src=(UPat.var("y"),), allow_any_len=True, name="x"),
lambda y,x: UOp(Ops.NOOP, x.dtype, y.src) if all(s.op is Ops.GEP and s.src == y.src and s.arg[0] == i for i,s in enumerate(x.src)) else None),
])
# ***** X86 pre instruction selection *****
# these must be done in a separate matcher because they violate the spec
pre_isel_matcher = PatternMatcher([
# cast from pointer is a noop
(UPat.var("y").cast(name="x"), lambda y,x: x.replace(op=Ops.NOOP) if isinstance(y.dtype, PtrDType) else None),
# zero extending scalar 32bit int is a noop
(UPat.var("y", dtypes.uint32).cast(dtypes.int64s, name="x"), lambda y,x: x.replace(op=Ops.NOOP) if y.dtype.count == 1 else None),
# cast between signed and unsigned int is a noop
(UPat.var("y", dtypes.ints+(dtypes.bool,)).cast(dtypes.ints, name="x"),
lambda y,x: x.replace(op=Ops.NOOP) if x.dtype.itemsize == y.dtype.itemsize else None),
# cast to < scalar int is a noop
(UPat.var("y", dtypes.ints).cast(dtypes.ints, name="x"),
lambda y,x: x.replace(op=Ops.NOOP) if x.dtype.itemsize < y.dtype.itemsize and y.dtype.count == 1 else None),
# bitcasts between scalar floats and ints are real, rest are noops
(UPat.var("y").bitcast().named("x"), lambda y,x: None if y.dtype in dtypes.floats and x.dtype in dtypes.ints or \
y.dtype in dtypes.ints and x.dtype in dtypes.floats else x.replace(op=Ops.NOOP)),
# noop of a noop is removed
(UPat(Ops.NOOP, src=(UPat(Ops.NOOP),), name="x"), lambda x: x.replace(src=x.src[0].src)),
# moving elements of a single register to another without shuffling is a noop
(UPat(Ops.VECTORIZE, src=(UPat.var("y"),), allow_any_len=True, name="x"),
lambda y,x: UOp(Ops.NOOP, x.dtype, y.src) if all(s.op is Ops.GEP and s.src == y.src and s.arg[0] == i for i,s in enumerate(x.src)) else None),
# gated index becomes a conditional move on the index, the load/store are unconditional
(UPat.var("base").index(UPat.var("idx"), UPat.var("gate")).load(UPat.var("alt"), name="x"), lambda base,idx,gate,alt,x:
gate.where(base.index(idx, ptr=True), (l:=UOp(Ops.DEFINE_LOCAL, base.dtype.base.ptr(x.dtype.count), arg=0)).after(l.store(alt))
tinygrad/uop/spec.py
+2 -6
View File
@@ -14,11 +14,10 @@ def validate_index(buf:UOp, idx:UOp, gate:UOp|None=None):
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, X86 turns certain casts into NOOPs
# WEBGPU has a BITCAST in the index, PTX casts pointer to long
# VECTORIZE/GEP can't be properly modeled in z3 since it doesn't support vectors
for x in idx.toposort() | gate.toposort():
if x.op in {Ops.BITCAST, Ops.VECTORIZE, Ops.GEP} or (x.op is Ops.CAST and isinstance(x.src[0].dtype, PtrDType)) \
or x.op is Ops.NOOP: return True
if x.op in {Ops.BITCAST, Ops.VECTORIZE, Ops.GEP} or (x.op is Ops.CAST and isinstance(x.src[0].dtype, PtrDType)): return True
# if all is good and CHECK_OOB=1, validate with z3
from tinygrad.uop.validate import validate_index_with_z3
@@ -36,9 +35,6 @@ def validate_index(buf:UOp, idx:UOp, gate:UOp|None=None):
shared_spec = PatternMatcher([
(UPat(Ops.SINK, dtypes.void), lambda: True), # NOTE: for testing, we let sinks be anything
# NOOP
(UPat(Ops.NOOP), lambda: True),
# CONST/DEFINE_VAR are everywhere
(UPat(Ops.CONST, src=(), name="x"), lambda x: type(x.arg) is type(dtypes.as_const(x.arg, x.dtype))),
(UPat(Ops.DEFINE_VAR, name="x"), lambda x: isinstance(x.arg[1], int) and isinstance(x.arg[2], int)),