INDEX

toonygrad/engine/realize.py

@@ -8,30 +8,6 @@
 from toonygrad.codegen.linearize import linearize_uop
 from toonygrad.codegen.uopgraph import full_graph_rewrite
 from toonygrad.renderer import Renderer
-
-acc_number = 0
-def do_reduce(root:UOp):
-  global acc_number
-  reduce_parented, reduce_unparented = partition(root.src[1:], lambda x: x in root.src[0].sparents)
-  ret = root.src[0]
-  if len(reduce_parented):
-    acc = UOp(UOps.DEFINE_ACC, root.dtype,
-              (root.const_like(identity_element(root.arg, root.dtype.scalar())),) + tuple(reduce_parented), (acc_number,))
-    acc_number += 1
-    ret = UOp(UOps.ASSIGN, root.dtype, (acc, acc.alu(root.arg, ret)))
-  # for MAX, we can just ignore the unparented
-  if root.arg is BinaryOps.ADD:
-    for r in reduce_unparented:ret = ret * (r.src[1]-r.src[0]).cast(ret.dtype.scalar()).broadcast(ret.dtype.count)
-  return ret
-
-no_pyint = PatternMatcher([(UPat((UOps.CONST, UOps.VCONST, UOps.ALU, UOps.SPECIAL, UOps.RANGE, UOps.EXPAND, UOps.VECTORIZE, UOps.DEFINE_VAR),
-  name="x"), lambda x: UOp(x.op, dtypes.int32.vec(x.dtype.count), x.src, x.arg) if x.dtype.scalar() == dtypes.pyint else None)])
-
-just_reduce = PatternMatcher([
-  # do reduce
-  (UPat(UOps.REDUCE, name="root"), do_reduce),
-])
-
 from toonygrad.codegen.kernel import Kernel
 
 @track_rewrites

toonygrad/engine/schedule.py

@@ -1,6 +1,6 @@
-from typing import List, Dict, Tuple
+from typing import List, Dict, Tuple, Optional
 from dataclasses import dataclass
-from toonygrad.ops import UOp, graph_rewrite, PatternMatcher, UPat, UOps, symbolic, track_rewrites
+from toonygrad.ops import UOp, graph_rewrite, PatternMatcher, UPat, UOps, symbolic, track_rewrites, resolve
 from toonygrad.engine.lazy import LazyBuffer
 from toonygrad.shape.symbolic import Variable
 from toonygrad.shape.shapetracker import ShapeTracker
@@ -44,11 +44,20 @@ def append_kernel(k:List[UOp], base:UOp): k.append(base.sink())
   (UPat(UOps.LOAD, src=(UPat(), UPat(), UPat()), name="ld"), lambda k,ld: UOp.load(ld.src[0], ld.src[1], dtype=ld.dtype)),
 ])
 
-def append_buffer(b:List[Buffer], base:UOp):
-  if base.buffer not in b: b.append(base.buffer)
+def index_buffer(b:List[Buffer], buf:UOp, view:UOp, store:Optional[UOp]=None) -> UOp:
+  if buf.buffer not in b: b.append(buf.buffer)
   # should this be the ptr, or the buffer?
-  return UOp(UOps.DEFINE_GLOBAL, base.dtype.ptr(), (), b.index(base.buffer))
-enumerate_bufs = PatternMatcher([(UPat(UOps.BUFFER, name="base"), append_buffer)])
+  idx, mask = view.st.to_indexed_uops()
+  ret = UOp(UOps.DEFINE_GLOBAL, buf.dtype.ptr(), (), b.index(buf.buffer))
+  if resolve(mask != True): ret = mask.where(ret, ret.const_like(0))
+  ret = UOp(UOps.INDEX, buf.dtype.ptr(), (ret, idx))
+  return UOp.store(ret, store) if store is not None else UOp.load(ret, dtype=buf.dtype)
+
+enumerate_bufs = PatternMatcher([
+  #(UPat(UOps.BUFFER, name="base"), append_buffer),
+  (UPat(UOps.LOAD, src=(UPat(UOps.BUFFER, name="buf"), UPat(UOps.VIEW, name="view"))), index_buffer),
+  (UPat(UOps.STORE, src=(UPat(UOps.BUFFER, name="buf"), UPat(UOps.VIEW, name="view"), UPat.var("store"))), index_buffer),
+])
 
 @track_rewrites
 def _schedule_rewrite(sink:UOp) -> List[ScheduleItem]:

toonygrad/ops.py

@@ -106,6 +106,7 @@ class UOps(FastEnum):
   EMPTY = auto()
   BUFFER_VIEW = auto()
 
+  INDEX = auto()
   EXPAND = auto()
   CONTRACT = auto()
   VIEW = auto()