we don't float anymore

* fix pow(int, float)

* onnx

conftest.py

@@ -0,0 +1,10 @@
+import os, pytest, signal, threading
+
+@pytest.hookimpl(wrapper=True)
+def pytest_runtest_call(item):
+  t = threading.Timer(int(os.getenv("TEST_TIMEOUT", 300)), os.kill, args=(os.getpid(), signal.SIGABRT))
+  t.start()
+  try: yield
+  finally:
+    t.cancel()
+    t.join()

extra/f16_decompress.py

@@ -5,10 +5,10 @@ def bit_extract(x: Tensor, e: int, s: int) -> Tensor:
   return (x >> s) & mask
 
 def u16_to_f16(x: Tensor) -> Tensor:
-  sign = bit_extract(x, 15, 15).float()
+  sign = bit_extract(x, 15, 15).bool()
   exponent = bit_extract(x, 14, 10).float()
   fraction = bit_extract(x, 9, 0).float()
-  return sign.where(-1, 1) * exponent.where((exponent - 15.0).exp2() * (1 + fraction / 1024.0), 6.103515625e-5 * (fraction / 1024.0))
+  return sign.where(-1, 1) * exponent.bool().where((exponent - 15.0).exp2() * (1 + fraction / 1024.0), 6.103515625e-5 * (fraction / 1024.0))
 
 def u32_to_f16(oo: Tensor) -> Tensor:
   f1 = u16_to_f16(oo>>16)

extra/hcq2/hcq2.py

@@ -49,7 +49,7 @@ def make_getaddr(u, device=None):
   return UOp(Ops.GETADDR, dtypes.uint64, src=(u,), arg=device or to_tuple(u.device)[0])
 
 def make_ins(op, *srcs):
-  return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)
+  return UOp(Ops.INS, arg=op, src=tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs))
 
 def make_placeholder(devs, size:int, dtype, name=None, unique=True) -> UOp:
   return UOp.param(next(UOp.unique_num) if unique else 0, dtype, shape=(size,), device=devs).rtag(name or "temp")
@@ -133,7 +133,7 @@ def _build_wait_cmds(dep_lanes:list[tuple[tuple, int, int]], devices:tuple[str,
   for (ddevs, dqueue, dtag), lanes in deps.items():
     sig = make_mstack([make_signal(d if dl is None else ddevs[dl], queue=dqueue, sentinel=dl is None) for dl, d in zip(lanes, devices)])
     val = make_mstack([make_signal_value(d if dl is None else ddevs[dl], queue=dqueue) for dl, d in zip(lanes, devices)])
-    waits.append((sig.index(zero:=UOp.const(dtypes.int, 0)).load() >= val.index(zero) + dtag).wait())
+    waits.append(UOp(Ops.INS, arg="wait", src=(sig, val.index(UOp.const(dtypes.int, 0)) + dtag)))
   return waits, {dtag for _, _, dtag in deps}
 
 def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[tuple[tuple[str, ...], str]],
@@ -154,7 +154,8 @@ def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[t
     waited |= cur_waited
 
     # wait the syncs, store the device epoch; value bumps are a separate call: no lane may bump until every lane has patched its waits
-    submit = make_submit(*waits, make_signal(devs).store((tl:=make_signal_value(devs)).index(zero)), devs=devs, queue="COMPUTE:0")
+    store = UOp(Ops.INS, arg="store", src=(make_signal(devs), (tl:=make_signal_value(devs)).index(zero)))
+    submit = make_submit(*waits, store, devs=devs, queue="COMPUTE:0")
     upd = [(tl, 1)] + [(make_signal_value(devs, queue=qn), n) for qn in dedup([qn for bdevs, qn in batch_info if set(bdevs) & set(devs)])]
     bump = UOp.barrier(*[s.index(zero, dtype=s.dtype).store(s.index(zero) + inc) for s, inc in upd])
     finalizers += [UOp.custom_function("hcq", b.sink()).call(aux=HCQInfo("hcq_finalizer", Estimates(), devs, "COMPUTE:0")) for b in (submit, bump)]
@@ -181,15 +182,15 @@ def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]]) -> list[UOp]:
   for tag, ((call, _), (devices, queue), cmds) in enumerate(zip(batch, batch_info, call_waits)):
     # first queue use, sync prior device work with main signal
     if batch_info.index((devices, queue)) == tag:
-      epoch = (make_signal(devices).index(0).load() >= make_signal_value(devices).index(0) - 1).wait()
-      cmds = [UOp(Ops.BARRIER), epoch] + cmds
-
-    # signal queue timeline if someone waits for us
-    store = make_signal(devices, queue=queue).store(make_signal_value(devices, queue=queue).index(0) + tag) if tag in waited else None
+      epoch = UOp(Ops.INS, arg="wait", src=(make_signal(devices), make_signal_value(devices).index(0) - 1))
+      cmds = [UOp(Ops.INS, arg="barrier", src=()), epoch] + cmds
 
     # and make hcq call
     info = HCQInfo(get_call_name(call, get_call_arg_uops(call)), estimate_uop(call), devices, queue)
-    cmds = [*cmds, call.replace(arg=replace(call.arg, aux=info))] + ([store] if store is not None else [])
+    cmds = [*cmds, call.replace(arg=replace(call.arg, aux=info))]
+
+    # signal queue timeline if someone waits for us
+    if tag in waited: cmds += [UOp(Ops.INS, arg="store", src=(make_signal(devices, queue), make_signal_value(devices, queue).index(0) + tag))]
     src.append(UOp.custom_function("hcq", make_submit(*cmds, devs=devices, queue=queue).sink()).call(name="hcq", aux=info))
   return src + finalizers
 

extra/hcq2/ops_amd2.py

@@ -90,7 +90,7 @@ def memory_barrier(ctx):
                  reg_done=getattr(ctx.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_DONE').addr[0], value=0xffffffff),
     acquire_mem(ctx)))
 
-def pm4_wait(ctx, x, y): return wait_reg_mem(ctx, y, mem=make_getaddr(x.buf_uop, ctx.devs))
+def pm4_wait(ctx, dst, val): return wait_reg_mem(ctx, val, mem=make_getaddr(dst, ctx.devs))
 
 def pm4_barrier(ctx): return memory_barrier(ctx)
 
@@ -138,10 +138,10 @@ def pm4_program(ctx, call, prg):
 pm_pm4_opsel = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="prg"),), name="call", allow_any_len=True), pm4_program),
 
-  (UPat(Ops.WAIT, src=(UPat.var("x") >= UPat.var("y"),)), pm4_wait),
-  (UPat(Ops.BARRIER), pm4_barrier),
-  (UPat(Ops.CUSTOM_FUNCTION, arg="timestamp", src=(UPat(name="dst"),)), pm4_timestamp),
-  (UPat(Ops.STORE, src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), pm4_store),
+  (UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val"))), pm4_wait),
+  (UPat(Ops.INS, arg="barrier"), pm4_barrier),
+  (UPat(Ops.INS, arg="timestamp", src=(UPat(name="dst"),)), pm4_timestamp),
+  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), pm4_store),
 ])
 
 def pm4_submit(cmdbuf, devs):
@@ -184,10 +184,10 @@ def sdma_copy(ctx, call):
      ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz - off, ctx.max_copy_size) - 1), 0,
      *data64_le(src_addr + off), *data64_le(dst_addr + off)) for off in range(0, sz, ctx.max_copy_size)]))
 
-def sdma_wait(ctx, x, y):
+def sdma_wait(ctx, dst, val):
   op = ctx.sdma.SDMA_OP_POLL_REGMEM | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_FUNC(WAIT_REG_MEM_FUNCTION_GEQ) \
      | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)
-  return make_ins(SDMAOps.POLL_REGMEM, op, *data64_le(make_getaddr(x.buf_uop, ctx.devs)), y, 0xffffffff,
+  return make_ins(SDMAOps.POLL_REGMEM, op, *data64_le(make_getaddr(dst, ctx.devs)), val, 0xffffffff,
     ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(0x04) | ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff))
 
 def sdma_store(ctx, dst, val):
@@ -202,10 +202,10 @@ def sdma_timestamp(ctx, dst):
 pm_sdma_opsel = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.COPY),), name="call", allow_any_len=True), sdma_copy),
 
-  (UPat(Ops.BARRIER), lambda: UOp(Ops.NOOP, dtypes.void, ())),
-  (UPat(Ops.WAIT, src=(UPat.var("x") >= UPat.var("y"),)), sdma_wait),
-  (UPat(Ops.CUSTOM_FUNCTION, arg="timestamp", src=(UPat(name="dst"),)), sdma_timestamp),
-  (UPat(Ops.STORE, src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), sdma_store),
+  (UPat(Ops.INS, arg="barrier"), lambda: UOp(Ops.NOOP, dtypes.void, ())),
+  (UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val"))), sdma_wait),
+  (UPat(Ops.INS, arg="timestamp", src=(UPat(name="dst"),)), sdma_timestamp),
+  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), sdma_store),
 ])
 
 def sdma_submit(cmdbuf, devs):

pyproject.toml

@@ -74,7 +74,6 @@ testing_minimal = [
   "torch==2.9.1",
   "pytest",
   "pytest-xdist",
-  "pytest-timeout",
   "pytest-split",
   "hypothesis>=6.148.9",
   "z3-solver<4.15.4",  # 4.15.4 has a segfault when creating many z3.Context()
@@ -160,8 +159,6 @@ norecursedirs = [
   ".hypothesis",
   ".git",
 ]
-timeout = 300
-timeout_func_only = true
 testpaths = ["test"]
 filterwarnings = [
   # Ignore SWIG warnings from importlib

test/backend/test_linearizer.py

@@ -290,7 +290,7 @@ class TestLinearizer(unittest.TestCase):
   @unittest.skipIf(MOCKGPU and isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, CUDARenderer)), "PTX indexes differently. might be ok?")
   def test_where_fold(self):
     a = Tensor.ones(4, 4).contiguous().realize()
-    b = a.shrink(((1, 2), None)).pad(((1, 2), None))
+    b = a.shrink(((1, 2), None)).pad(((1, 2), None)).bool()
     a.assign(b.where(2, a))
     linear, var_vals = a.linear_with_vars()
     assert len(linear.src) == 1

test/backend/test_ops.py

@@ -5,7 +5,6 @@ import torch
 from tinygrad.helpers import getenv, DEBUG, DEV, IMAGE, Context
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.renderer.cstyle import QCOMCLRenderer
 from tinygrad.renderer.nir import NIRRenderer
 
 TINY_BACKEND = getenv("TINY_BACKEND")
@@ -450,7 +449,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(45,35), (45,35), (45,35)], lambda x,y,z: x.lerp(y,z))
     helper_test_op(None, lambda x,y,z: x.lerp(y,z), vals=[[1.,2.,3.], [4.,5.,6.], 0.5])
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_tril(self):
     helper_test_op([(3,3)], lambda x: x.tril())
     helper_test_op([(3,3)], lambda x: x.tril(1))
@@ -468,7 +466,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(5,3,3)], lambda x: x.tril(1))
     helper_test_op(None, lambda x: x.tril(), vals=[[[True] * 3] * 3], forward_only=True)
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_triu(self):
     helper_test_op([(3,3)], lambda x: x.triu())
     helper_test_op([(3,3)], lambda x: x.triu(1))
@@ -771,6 +768,11 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: torch.tensor([2], dtype=torch.int) ** torch.tensor(-2, dtype=torch.int),
                        lambda: Tensor([2]) ** Tensor(-2), forward_only=True)
 
+  def test_pow_int_base_float_exponent(self):
+    for exponent in (0.5, 1.5, 2.0, -1.0, 0.0):
+      helper_test_op([], lambda: torch.tensor([1, 2, 3, 4], dtype=torch.int) ** exponent,
+                         lambda: Tensor([1, 2, 3, 4], dtype=dtypes.int32) ** exponent, forward_only=True)
+
   def test_sqrt(self):
     helper_test_op([(45,65)], lambda x: x.sqrt())
     helper_test_op(None, lambda x: x.sqrt(), vals=[[0.0]])
@@ -791,7 +793,6 @@ class TestOps(unittest.TestCase):
 
     self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_xor(y), expected=RuntimeError)
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_and(self):
     data = [[1,-8,1],[32,1,6]]
     tor = torch.tensor(data, dtype=torch.int)
@@ -809,7 +810,6 @@ class TestOps(unittest.TestCase):
 
     self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_and(y), expected=RuntimeError)
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_or(self):
     data = [[1,-8,1],[32,1,6]]
     tor = torch.tensor(data, dtype=torch.int)
@@ -1229,7 +1229,6 @@ class TestOps(unittest.TestCase):
     helper_test_op(None, lambda x: x.type(torch.int32).argmax().type(torch.int32), lambda x: x.argmax(), forward_only=True, vals=[[False, True]])
     helper_test_op(None, lambda x: x.type(torch.int32).argmax().type(torch.int32), lambda x: x.argmax(), forward_only=True, vals=[[True, False]])
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_argmin(self):
     # check if it returns the first index for multiple occurrences
     helper_test_op(None, lambda x: x.argmin().type(torch.int32), lambda x: x.argmin(), forward_only=True, vals=[[2, 2]])
@@ -1535,7 +1534,6 @@ class TestOps(unittest.TestCase):
   def test_prod_dtype_arg(self):
     with self.assertRaises(AttributeError): Tensor([1.0, 2.0]).prod(dtype="")
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_min(self):
     helper_test_op([(3,3)], lambda x: x.min())
     helper_test_op([(45,3)], lambda x: x.min())
@@ -1575,7 +1573,6 @@ class TestOps(unittest.TestCase):
   def test_any_zero_axis(self):
     helper_test_op([(1,0,3,0,5)], lambda x: x.any(axis=(1,3)), forward_only=True)
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_all(self):
     helper_test_op([(3,4,5,6)], lambda x: x.all(), forward_only=True)
     helper_test_op(None, lambda x: x.all(), vals=[[True, True]], forward_only=True)
@@ -2953,7 +2950,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[...,c,:,e], lambda x: x[...,k,:,p])
 
   @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_dim_collapse_int(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     # dim collapse from int
@@ -2964,7 +2960,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[1,:,3:11:2,d,0:2], lambda x: x[1,:,3:11:2,o,0:2])
 
   @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_dim_inject_none(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     # dim injection from None
@@ -2999,7 +2994,6 @@ class TestOps(unittest.TestCase):
                             lambda x: x[Tensor([[0,1,-1],[-1,-2,0]]), Tensor([2,1,-1])])
 
   @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_list_indices(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[((0,),)])
@@ -3011,7 +3005,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,(2,1,0),c,(-2,1,0),e], lambda x: x[i,(2,1,0),k,(-2,1,0),p])
 
   @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_tuple_indices(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[(((0,),),)], lambda x: x[(((0,),),)])

test/backend/test_pickle.py

@@ -1,7 +1,7 @@
-import unittest, pickle, types
+import unittest, pickle, types, tracemalloc
 import numpy as np
-from tinygrad import Tensor, TinyJit, Variable, dtypes
-from tinygrad.helpers import GlobalCounters, ContextVar, Context
+from tinygrad import Tensor, Device, TinyJit, Variable, dtypes
+from tinygrad.helpers import GlobalCounters, ContextVar, Context, DEV
 from tinygrad.uop.ops import PatternMatcher, UPat, UOp
 
 class TestPickle(unittest.TestCase):
@@ -78,6 +78,22 @@ class TestPickle(unittest.TestCase):
     a2:UOp = pickle.loads(s)
     self.assertListEqual(a2.base.realized.as_memoryview().cast("I").tolist(), [0, 1, 2, 3])
 
+  @unittest.skipIf(DEV.interface.startswith("MOCK"), "mock device buffers live in host RAM, not VRAM")
+  def test_pickle_oob_ram(self):
+    N, M = 8, 10**6
+    ts = [Tensor.rand(M, dtype='float32').realize() for _ in range(N)]
+    tracemalloc.start()
+    st = pickle.dumps(ts, protocol=5, buffer_callback=lambda pb: pb.release())
+    self.assertLess(tracemalloc.get_traced_memory()[1], N*M*4)
+    tracemalloc.reset_peak()
+    def make_fake_buffers():
+      for _ in range(N):
+        Device[Device.DEFAULT].synchronize()
+        yield pickle.PickleBuffer(bytearray(M*4))
+    pickle.loads(st, buffers=make_fake_buffers())
+    self.assertLess(tracemalloc.get_traced_memory()[1], N*M*4)
+    tracemalloc.stop()
+
   def test_pickle_unrealized_tensor(self):
     t = Tensor.ones(10, 10)
     st = pickle.dumps(t)

test/backend/test_wait_loop.py

@@ -0,0 +1,104 @@
+import unittest
+from tinygrad import Tensor, UOp
+from tinygrad.device import Device
+from tinygrad.dtype import AddrSpace, dtypes
+from tinygrad.renderer.nir import NIRRenderer
+from tinygrad.renderer.isa.x86 import X86Renderer
+from tinygrad.uop.ops import KernelInfo
+
+def wait_loop_kernel(C:UOp) -> UOp:
+  N = 10
+
+  # a RANGE with no src is a bound-less loop header: a jump target with no induction variable.
+  # the compare and conditional backedge are expanded by the renderers from the loop RANGE/END
+  l = UOp.loop(0)
+
+  i = UOp.placeholder((1,), dtypes.int, 0, addrspace=AddrSpace.REG)
+
+  # i = 0
+  i = i.after(i[0].store(0))
+
+  # i + 1, read loop-carried through after(l)
+  inc = i.after(l)[0].load() + 1
+
+  # i = inc; END(store, l, cond): conditional backedge, loop again while inc < N (do-while)
+  # NOTE: the cond uses the computed value, not a reload of the register
+  st = i[0].store(inc)
+  i = i.after(st.end(l, inc < N))
+
+  return C[0].store(i[0].load()).sink(arg=KernelInfo(name="wait_loop"))
+
+def nested_loop_kernel(C:UOp) -> UOp:
+  r = UOp.range(4, 0)
+  l = UOp.loop(1)
+
+  i = UOp.placeholder((1,), dtypes.int, 0, addrspace=AddrSpace.REG)
+  i = i.after(i[0].store(0))
+
+  inc = i.after(l, r)[0].load() + 1
+  st = i[0].store(inc)
+
+  lend = st.end(l, inc < (r.cast(dtypes.int)+1)*3)
+  i = i.after(lend.end(r))
+
+  return C[0].store(i[0].load()).sink(arg=KernelInfo(name="nested_loop", opts_to_apply=()))
+
+def two_loops_kernel(C:UOp) -> UOp:
+  # two sequential loops on the same counter: ++ until 10, then ++ until 25
+  l1, l2 = UOp.loop(0), UOp.loop(1)
+
+  i = UOp.placeholder((1,), dtypes.int, 0, addrspace=AddrSpace.REG)
+  i = i.after(i[0].store(0))
+
+  inc1 = i.after(l1)[0].load() + 1
+  i = i.after(i[0].store(inc1).end(l1, inc1 < 10))
+
+  inc2 = i.after(l2)[0].load() + 1
+  i = i.after(i[0].store(inc2).end(l2, inc2 < 25))
+
+  return C[0].store(i[0].load()).sink(arg=KernelInfo(name="two_loops", opts_to_apply=()))
+
+def loop_in_loop_kernel(C:UOp) -> UOp:
+  # outer loop while i < 12, inner loop increments until i % 4 == 0 -> 12
+  l1, l2 = UOp.loop(0), UOp.loop(1)
+
+  i = UOp.placeholder((1,), dtypes.int, 0, addrspace=AddrSpace.REG)
+  i = i.after(i[0].store(0))
+
+  inc = i.after(l1, l2)[0].load() + 1
+  st = i[0].store(inc)
+
+  # the outer END closes the inner END, and its cond reloads the register after the inner loop (in scope at the outer level)
+  e2 = st.end(l2, inc % 4 != 0)
+  oc = i.after(e2)[0].load()
+  i = i.after(e2.end(l1, oc < 12))
+
+  return C[0].store(i[0].load()).sink(arg=KernelInfo(name="loop_in_loop", opts_to_apply=()))
+
+@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (NIRRenderer, X86Renderer)), "loops are not supported in LVP and X86")
+class TestWaitLoop(unittest.TestCase):
+  def test_wait_loop(self):
+    c = Tensor.empty(1, dtype=dtypes.int)
+    c = Tensor.custom_kernel(c, fxn=wait_loop_kernel)[0]
+    c.realize()
+    self.assertEqual(c.item(), 10)
+
+  def test_nested_loop_in_range(self):
+    c = Tensor.empty(1, dtype=dtypes.int)
+    c = Tensor.custom_kernel(c, fxn=nested_loop_kernel)[0]
+    c.realize()
+    self.assertEqual(c.item(), 12)
+
+  def test_two_sequential_loops(self):
+    c = Tensor.empty(1, dtype=dtypes.int)
+    c = Tensor.custom_kernel(c, fxn=two_loops_kernel)[0]
+    c.realize()
+    self.assertEqual(c.item(), 25)
+
+  def test_loop_in_loop(self):
+    c = Tensor.empty(1, dtype=dtypes.int)
+    c = Tensor.custom_kernel(c, fxn=loop_in_loop_kernel)[0]
+    c.realize()
+    self.assertEqual(c.item(), 12)
+
+if __name__ == "__main__": unittest.main()

test/null/test_dtype_spec.py

@@ -315,6 +315,19 @@ class TestAutoCastType(unittest.TestCase):
     assert (Tensor.ones(4, 4, dtype=dt) + 2).dtype == (dt if dtypes.is_float(dt) or dtypes.is_int(dt) else dtypes.default_int)
     assert (Tensor.ones(4, 4, dtype=dt) + True).dtype == dt
 
+  @given(strat.sampled_from(core_dtypes))
+  def test_pad_scalar(self, dt):
+    t = Tensor.ones(4, dtype=dt)
+    assert t.pad(((1, 1),), value=2.3).dtype == (dt if dtypes.is_float(dt) else dtypes.default_float)
+    assert t.pad(((1, 1),), value=2).dtype == (dt if dtypes.is_float(dt) or dtypes.is_int(dt) else dtypes.default_int)
+    assert t.pad(((1, 1),), value=True).dtype == dt
+
+  @given(strat.sampled_from(core_dtypes))
+  def test_sort(self, dt):
+    # sort pads with dtype.min/max, a scalar of its own dtype
+    assert Tensor([3, 1, 2], dtype=dt).sort()[0].dtype == dt
+    assert Tensor([3, 1, 2], dtype=dt).sort(descending=True)[0].dtype == dt
+
   @given(strat.sampled_from(dtype_floats))
   def test_int_div_int(self, default_float):
     dtypes.default_float = default_float
@@ -415,6 +428,9 @@ class TestAutoCastType(unittest.TestCase):
     self.check_where_alternate_input_other(3.1, True, dtypes.default_float)
     self.check_where_alternate_input_other(3, 2, dtypes.default_int)
     self.check_where_alternate_input_other(3, True, dtypes.default_int)
+
+  def test_where_non_bool_cond_raises(self):
+    with self.assertRaises(RuntimeError): Tensor([1, 0, 2]).where(1, 0)
     self.check_where_alternate_input_other(False, True, dtypes.bool)
 
   @given(strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes))

test/unit/test_dtype_weak.py

@@ -38,7 +38,6 @@ class TestWeakPromotion(unittest.TestCase):
     self.assertEqual(((t_bool + 1) + t_i8).dtype, dtypes.int8)
     self.assertEqual(((t_bool + 1) + t_u16).dtype, dtypes.uint16)
     self.assertEqual((Tensor(3) + t_i8).dtype, dtypes.int8)
-    self.assertEqual(Tensor([2], dtype=dtypes.uint8).pad(((1, 1),), value=1).dtype, dtypes.uint8)
     # zeros/ones are full with a python fill value, so they are weak too (jnp.zeros pins float32; deliberate divergence)
     self.assertEqual((Tensor.zeros(3) + t_f16).dtype, dtypes.float16)
 
@@ -47,6 +46,7 @@ class TestWeakPromotion(unittest.TestCase):
     self.assertEqual((t_i8 + 1).dtype, dtypes.int8)
     self.assertEqual((t_f16 + 0.5).dtype, dtypes.float16)
     self.assertEqual((t_f32 + t_f16).dtype, dtypes.float32)
+    self.assertEqual(Tensor([2], dtype=dtypes.uint8).pad(((1, 1),), value=1).dtype, dtypes.uint8)
 
   @unittest.expectedFailure  # TODO: dot of a weak const tensor defers to the other operand once python scalars are weak consts
   def test_dot_defers_weak(self):

test/unit/test_gradient.py

@@ -1,8 +1,8 @@
-import unittest
+import unittest, math
 import numpy as np
 from tinygrad import Tensor
 from tinygrad.dtype import dtypes
-from tinygrad.uop.ops import UOp, KernelInfo
+from tinygrad.uop.ops import UOp, KernelInfo, Ops
 
 class TestTensorGradient(unittest.TestCase):
   def test_example(self):
@@ -98,6 +98,33 @@ class TestTensorGradient(unittest.TestCase):
     x = Tensor.randn(4, 4)
     np.testing.assert_allclose(x.pad(((1,0),(0,0))).gradient(x, gradient=g2)[0].numpy(), np.zeros((4, 4)))
 
+  def test_implicit_broadcast_where_gradient(self):
+    # WHERE with a bare ()-shape branch: the scalar's gradient counts the positions where it is selected
+    cond, x, w = Tensor([True, False, True]), Tensor([1.0, 2.0, 3.0]), Tensor(4.0)
+    dw = Tensor(cond.uop.alu(Ops.WHERE, x.uop, w.uop)).sum().gradient(w)[0]
+    self.assertEqual(dw.shape, ())
+    self.assertEqual(dw.item(), 1.0)
+    dw = Tensor(cond.uop.alu(Ops.WHERE, w.uop, x.uop)).sum().gradient(w)[0]
+    self.assertEqual(dw.item(), 2.0)
+
+  def test_implicit_broadcast_alu_gradient(self):
+    # MUL with a bare ()-shape src, no EXPAND in the graph
+    x, w = Tensor([1.0, 2.0, 3.0]), Tensor(2.0)
+    m = x.uop.alu(Ops.MUL, w.uop)
+    self.assertIs(m.src[1], w.uop)
+    dw = Tensor(m).sum().gradient(w)[0]
+    self.assertEqual(dw.shape, ())
+    self.assertEqual(dw.item(), 6.0)
+
+  def test_implicit_broadcast_intermediate_accumulation(self):
+    # s is used directly and through an implicit broadcast edge, each edge's gradient reduces to s's shape before they sum
+    x, p = Tensor([1.0, 2.0, 3.0]), Tensor(0.5)
+    s = p.sin()
+    z = Tensor(x.uop.alu(Ops.MUL, s.uop)).sum() + s
+    dp = z.gradient(p)[0]
+    self.assertEqual(dp.shape, ())
+    self.assertAlmostEqual(dp.item(), 7*math.cos(0.5), places=5)
+
   def test_bare_const_skipped_by_backward(self):
     Tensor.manual_seed(0)
     w = Tensor(1.0)

test/unit/test_multitensor.py

@@ -63,6 +63,13 @@ class TestMultiTensor(unittest.TestCase):
     np.testing.assert_equal((s + Tensor(UOp.const(dtypes.float, 1.0))).numpy(), [2, 3, 4, 5])
     np.testing.assert_equal((s + Tensor(UOp.const(dtypes.float, 1.0)).reshape((1,)).expand((4,))).numpy(), [2, 3, 4, 5])
 
+  def test_add_rank_expand_shard(self):
+    # a sharded src keeps its own rank under implicit broadcast, its shard axis right-aligns into the output
+    a = Tensor([1.,2.,3.,4.]).shard(devices_2, 0)
+    b = Tensor([[10.,20.,30.,40.]]).shard(devices_2, None)
+    self.assertEqual((a+b).uop.axis, 1)
+    np.testing.assert_equal((a+b).numpy(), [[11.,22.,33.,44.]])
+
   def test_shard_reduce(self):
     self._test_shard_op(lambda t:t.reshape(2, 3).sum(axis=1), [3.,3.], n=6)
     self._test_shard_op(lambda t:t.reshape(2, 3).sum(axis=0), [2.,2.,2.], n=6)
@@ -602,8 +609,8 @@ class TestShrinkMultiTensorShardedAxis(unittest.TestCase):
       np.testing.assert_allclose((a+a).numpy(), (b+b).numpy(), rtol=1e-7, atol=1e-3)
       np.testing.assert_equal((a+1).numpy(), (b+1).numpy())
       np.testing.assert_equal((1+a).numpy(), (1+b).numpy())
-      np.testing.assert_allclose((a.where(a+a, a)).numpy(), (b.where(b+b, b)).numpy(), rtol=1e-7, atol=1e-3)
-      np.testing.assert_allclose((a.where(1, 0)).numpy(), (b.where(1, 0)).numpy(), rtol=1e-7, atol=1e-3)
+      np.testing.assert_allclose((a.bool().where(a+a, a)).numpy(), (b.bool().where(b+b, b)).numpy(), rtol=1e-7, atol=1e-3)
+      np.testing.assert_allclose((a.bool().where(1, 0)).numpy(), (b.bool().where(1, 0)).numpy(), rtol=1e-7, atol=1e-3)
 
       # reduce
       np.testing.assert_allclose(a.max().numpy(), b.max().numpy(), rtol=1e-7, atol=1e-3)

tinygrad/codegen/late/linearizer.py

@@ -2,7 +2,7 @@ import heapq
 from typing import Any
 from collections import defaultdict
 from tinygrad.uop.ops import PatternMatcher, UOp, Ops, UPat, multirange_str
-from tinygrad.dtype import AddrSpace
+from tinygrad.dtype import AddrSpace, dtypes
 from tinygrad.helpers import prod, getenv, TUPLE_ORDER
 
 def linearize(sink:UOp) -> list[UOp]:
@@ -85,9 +85,9 @@ pm_add_control_flow = PatternMatcher([
 ])
 
 def do_split_ends(e:UOp):
-  ret = e.src[0]
-  for r in sorted(UOp.sink(*e.src[1:]).ranges, key=lambda x: x.arg, reverse=True): ret = ret.end(r)
-  return ret
+  ret, backedge = e.src[0], tuple(x for x in e.src[1:] if x.dtype in (dtypes.void, dtypes.bool))
+  for r in sorted(UOp.sink(*[x for x in e.src[1:] if x not in backedge]).ranges, key=lambda x: x.arg, reverse=True): ret = ret.end(r)
+  return ret.end(*backedge) if len(backedge) else ret
 
 pm_split_ends = PatternMatcher([
   # split the ends

tinygrad/codegen/opt/postrange.py

@@ -21,8 +21,9 @@ class Scheduler:
 
   @property
   def rngs(self):
-    # always in order by axistype
-    return sorted([u for u in self.ast.backward_slice if u.op is Ops.RANGE and u.vmax > 0], key=lambda x: (axis_to_pos[x.arg[-1]],) + x.arg[0:-1])
+    # always in order by axistype. void RANGEs are loops, not opt axes
+    return sorted([u for u in self.ast.backward_slice if u.op is Ops.RANGE and u.dtype is not dtypes.void and u.vmax > 0],
+                  key=lambda x: (axis_to_pos[x.arg[-1]],) + x.arg[0:-1])
   @property
   def shape_len(self) -> int: return len(self.rngs)
   @property

tinygrad/codegen/simplify.py

@@ -9,7 +9,9 @@ def flatten_range(r:UOp) -> UOp|None:
   off = range_start[r.op]
   rngs = r.src[off:]
   if not len(rngs): return None
-  return r.replace(src=r.src[:off]+tuple(UOp.sink(*rngs).ranges))
+  # ranges in the cond should not be ended
+  backedge = tuple(x for x in rngs if x.dtype in (dtypes.void, dtypes.bool))
+  return r.replace(src=r.src[:off]+tuple(UOp.sink(*[x for x in rngs if x not in backedge]).ranges)+backedge)
 
 pm_flatten_range = PatternMatcher([
   # real ranges only
@@ -19,6 +21,7 @@ pm_flatten_range = PatternMatcher([
 # index/range arithmetic uses FLOORDIV/FLOORMOD prior to late rewrite
 def count_divmod(x:UOp) -> int: return sum(u.op in {Ops.FLOORDIV, Ops.FLOORMOD} for u in x.backward_slice)
 def simplify_merge_adjacent(u:UOp) -> UOp|None:
+  if not all(r.op is Ops.RANGE for r in u.ended_ranges): return None
   reduce_ranges = [x.ranges for x in u.backward_slice_with_self if x.op is Ops.REDUCE]
   # on END we only want to merge adjacent ranges, on REDUCE we want to try all combinations
   for r0, r1 in (zip(u.ended_ranges, u.ended_ranges[1:]) if u.op is Ops.END else itertools.permutations(u.ended_ranges, 2)):

tinygrad/device.py

@@ -100,8 +100,8 @@ class MultiBuffer:
 
 class Buffer:
   profile_events:list[ProfileEvent] = []
-  def __init__(self, device:str, size:int, dtype:DType, opaque:Any=None, options:BufferSpec|None=None, initial_value:bytes|None=None,
-               uop_refcount=0, base:Buffer|None=None, offset:int=0, preallocate=False):
+  def __init__(self, device:str, size:int, dtype:DType, opaque:Any=None, options:BufferSpec|None=None,
+               initial_value:bytes|pickle.PickleBuffer|None=None, uop_refcount=0, base:Buffer|None=None, offset:int=0, preallocate=False):
     assert isinstance(dtype, DType)
     self.device, self.size, self.dtype, self.options, self.offset, self.allocated_views = device, size, dtype, options, offset, 0
     self._bufs: dict[str, Any] = {}
@@ -113,6 +113,7 @@ class Buffer:
       if initial_value is not None:
         self.allocate()
         self.copy_from(Buffer("PYTHON", self.size, self.dtype, opaque=memoryview(bytearray(initial_value))))
+        if isinstance(initial_value, pickle.PickleBuffer): initial_value.release()
     else:
       assert base._base is None, "base can't have a base"
       assert device == base.device, "base must have the same device"
@@ -171,12 +172,13 @@ class Buffer:
       self.allocator.free(self._buf, self.nbytes, self.options)
     elif self._base is not None: self._base.allocated_views -= 1
     self._bufs.clear()
-  def __reduce__(self):
-    buf = None
+  def __reduce_ex__(self, protocol):
+    buf:bytearray|pickle.PickleBuffer|None = None
     if self._base is not None:
       return self.__class__, (self.device, self.size, self.dtype, None, None, None, 0, self.base, self.offset, self.is_allocated())
     if self.device == "NPY": return self.__class__, (self.device, self.size, self.dtype, self._buf, self.options, None, self.uop_refcount)
-    if self.is_allocated(): buf = bytearray(self.as_memoryview())
+    if self.is_allocated():
+      buf = pickle.PickleBuffer(self.as_memoryview()) if protocol >= 5 else bytearray(self.as_memoryview())
     return self.__class__, (self.device, self.size, self.dtype, None, self.options, buf, self.uop_refcount)
   @property
   def trace_num(self) -> int:

tinygrad/mixin/elementwise.py

@@ -6,7 +6,7 @@ from tinygrad.helpers import argfix, polyN
 from tinygrad.mixin.creation import CreationMixin
 
 if TYPE_CHECKING:
-  from tinygrad.uop.ops import UOp
+  from tinygrad.uop.ops import UOp, sint
 
 
 class ElementwiseMixin(CreationMixin):
@@ -18,9 +18,11 @@ class ElementwiseMixin(CreationMixin):
   def ufix(self, x: 'Self|ConstType|UOp') -> Self:
     return x if isinstance(x, type(self)) else self._wrap_uop(self._uop.ufix(x))
 
-  # implemented in OpMixin, broadcasting needs the movement ops
   def _broadcasted(self, y: 'Self|ConstType|UOp', reverse: bool = False) -> tuple[Self, Self]:
-    raise NotImplementedError
+    y = self.ufix(y)
+    x, y = (self, y) if not reverse else (y, self)
+    if x.dtype == y.dtype: return x, y
+    return x.cast(out_dtype := least_upper_dtype(x.dtype, y.dtype)), y.cast(out_dtype)
 
   def _binop(self, op: Ops, x: Self | ConstType, reverse: bool) -> Self:
     lhs, rhs = self._broadcasted(x, reverse)
@@ -414,10 +416,19 @@ class ElementwiseMixin(CreationMixin):
     m = a.maximum(b)
     return ((a-m).exp() + (b-m).exp()).log() + m
 
-  def where(self, x: Self | ConstType, y: Self | ConstType) -> Self:
-    ref: Self = x if isinstance(x, type(self)) else y if isinstance(y, type(self)) else \
-      self.cast(least_upper_dtype(dtypes.from_py(x), dtypes.from_py(y)))
-    return self.alu(Ops.WHERE, ref.ufix(x), ref.ufix(y))
+  def where(self, x: 'Self | ConstType | sint', y: 'Self | ConstType | sint') -> Self:
+    """
+    Returns a tensor of elements selected from either `x` or `y`, depending on `self`.
+    `output_i = x_i if self_i else y_i`.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    cond = Tensor([[True, True, False], [True, False, False]])
+    print(cond.where(1, 3).numpy())
+    ```
+    """
+    ref = x if isinstance(x, type(self)) else y if isinstance(y, type(self)) else self
+    x, y = ref.ufix(x)._broadcasted(y)
+    return self.alu(Ops.WHERE, x, y)
 
   def masked_fill(self, mask:Self, value:Self|PyConst) -> Self:
     """
@@ -548,9 +559,7 @@ class ElementwiseMixin(CreationMixin):
     # TODO: int pow
     if not base.is_floating_point() and isinstance(x, ConstType) and not (isinstance(x, int) and x >= 0):
       raise RuntimeError("base needs to be float")
-    ret = base.alu(Ops.POW, exponent)
-    # NOTE: pow(int, float) -> int
-    return ret.round().cast(self.dtype) if not reverse and not dtypes.is_float(self.dtype) and dtypes.is_float(exponent.dtype) else ret
+    return base.alu(Ops.POW, exponent)
 
   def __pow__(self, x: Self | ConstType) -> Self:
     return self.pow(x)

tinygrad/mixin/gradient.py

@@ -1,18 +1,13 @@
 from typing import cast
 import math, dataclasses
-from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, all_metadata
+from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, all_metadata, broadcast_axes
 from tinygrad.helpers import argsort
 from tinygrad.dtype import sum_acc_dtype
 
 def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):
-  def broadcast_to_input(x:UOp) -> UOp: return x._broadcast_to(ret.src[0].shape)
-  if op == Ops.ADD: return (broadcast_to_input(ctx),)
-  if op == Ops.MAX:
-    assert ret.op is Ops.REDUCE, "only works on REDUCE"
-    mask = ret.src[0].eq(broadcast_to_input(ret)).cast(ctx.dtype)
-    count = mask._rop(Ops.ADD, tuple(range(ret.arg[1])))
-    return ((mask/broadcast_to_input(count)) * broadcast_to_input(ctx),)
-  if op == Ops.MUL: return (broadcast_to_input(ctx * ret) / ret.src[0],)
+  if op == Ops.ADD: return (ctx._broadcast_to(ret.src[0].shape),)
+  if op == Ops.MAX: return (((mask:=ret.src[0].eq(ret).cast(ctx.dtype))/mask._rop(Ops.ADD, tuple(range(ret.arg[1])))) * ctx,)
+  if op == Ops.MUL: return (ctx * ret / ret.src[0],)
 
 def _compact_params(body:UOp, all_args:tuple[UOp, ...]) -> tuple[UOp, tuple[UOp, ...]]:
   """Remove unused PARAMs from body and return compacted (body, args)."""
@@ -67,9 +62,7 @@ pm_gradient = PatternMatcher([
   (UPat(Ops.CONTIGUOUS), lambda ctx: (ctx,)),
   (UPat(Ops.CONTIGUOUS_BACKWARD), lambda ctx: (ctx.contiguous(),)),
   (UPat(Ops.RESHAPE, name="ret"), lambda ctx, ret: (ctx.reshape(ret.src[0].shape), None)),
-  (UPat(Ops.EXPAND, name="ret"), lambda ctx, ret:
-    (ctx.cast(sum_acc_dtype(ctx.dtype))._rop(Ops.ADD, tuple(range(len(ret.marg))))
-     .reshape(ret.src[0].shape).cast(ctx.dtype), None)),
+  (UPat(Ops.EXPAND), lambda ctx: (ctx, None)),
   (UPat(Ops.PAD, name="ret"), lambda ctx, ret: (ctx.shrink(tuple([(p[0], s+p[0]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
   (UPat(Ops.SHRINK, name="ret"), lambda ctx, ret: (ctx.pad(tuple([(p[0], s-p[0]-p[1]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
   (UPat(Ops.PERMUTE, name="ret"), lambda ctx, ret: (ctx.permute(argsort(ret.marg)),)),
@@ -119,6 +112,9 @@ def compute_gradient(root:UOp, root_grad:UOp, targets:set[UOp]) -> dict[UOp, UOp
     assert len(lgrads) == len(t0.src), f"got {len(lgrads)} gradient, expected {len(t0.src)}"
     for k,v in zip(t0.src, lgrads):
       if v is None: continue
+      # a shaped edge's gradient is summed to its source's shape
+      if k._shape is not None and v._shape is not None and k._shape != v._shape:
+        v = v.cast(sum_acc_dtype(v.dtype))._rop(Ops.ADD, broadcast_axes(k.shape, v.shape)).reshape(k.shape).cast(v.dtype)
       if k in grads and grads[k].op is not Ops.NOOP:
         if v.op is Ops.TUPLE and grads[k].op is Ops.TUPLE:
           grads[k] = UOp.maketuple(*(p + n if (p.op is not Ops.NOOP and n.op is not Ops.NOOP) else

tinygrad/mixin/op.py

@@ -11,7 +11,7 @@ from tinygrad.helpers import all_int, argfix, argsort, ceildiv, flatten, flat_to
 from tinygrad.helpers import resolve_pool_pads, round_up, IMAGE, FLOAT16, WINO
 
 if TYPE_CHECKING:
-  from tinygrad.uop.ops import sint, UOp
+  from tinygrad.uop.ops import sint
 
 ReductionStr = Literal["mean", "sum", "none"]
 
@@ -110,8 +110,7 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
       consecutive = dims == list(range(dims[0], dims[0] + len(dims)))
       if v is None and len(dims) > 1 and consecutive and all_int(ishp := tuple(x.shape[d] for d in dims)):
         strides = tuple(prod(ishp[i+1:]) for i in range(len(dims)))
-        try: linear_idx = type(self).usum(*[t._broadcast_to(big_shape) * s for t, s in zip(tensors, strides)])
-        except ValueError as err: raise IndexError(f"cannot broadcast indices: {err}") from err
+        linear_idx = type(self).usum(*[t * s for t, s in zip(tensors, strides)])
         valid = type(self).uprod(*[(t >= 0) & (t < s) for t, s in zip(tensors, ishp)])
         pre, post = x.shape[:dims[0]], x.shape[dims[-1]+1:]
         x = x.reshape(pre + (prod(ishp),) + post)[tuple([slice(None)] * len(pre)) + (valid.where(linear_idx, 0),)]
@@ -285,8 +284,7 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
     pads = tuple((smax(pB,0), smax(pA,0)) for pB,pA in pX) if has_neg else pX
     base = MovementMixin.pad(X, pads)
     if value == 0: return base
-    if value is not Invalid: base = base.cast(least_upper_dtype(base.dtype, dtypes.from_py(value)))
-    return MovementMixin.pad(X.const_like(1).cast(dtypes.bool), pads).where(base, base.const_like(value))
+    return MovementMixin.pad(X.const_like(1).cast(dtypes.bool), pads).where(base, value)
 
   def _pad_circular(self, pX:tuple[tuple[sint, sint], ...]) -> Self:
     # shrink first for negative pads, then wrap the non-negative remainder
@@ -357,14 +355,6 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
     if mode in {"reflect", "replicate"}: return self._pad_reflect_replicate(pX, mode)
     raise NotImplementedError(f"{mode=} is not supported")
 
-  def _broadcasted(self, y:Self|ConstType|UOp, reverse:bool=False) -> tuple[Self, Self]:
-    if not isinstance(y, type(self)): y = self.ufix(y)
-    x, y = (self, y) if not reverse else (y, self)
-    out_shape = _broadcast_shape(x.shape, y.shape)
-    x, y = x._broadcast_to(out_shape), y._broadcast_to(out_shape)
-    if x.dtype == y.dtype: return x, y
-    return x.cast(out_dtype := least_upper_dtype(x.dtype, y.dtype)), y.cast(out_dtype)
-
   def dot(self, w:Self, dtype:DTypeLike|None=None) -> Self:
     """
     Performs dot product between two tensors.

tinygrad/mixin/rand.py

@@ -11,7 +11,7 @@ class RandMixin(OpMixin):
   @staticmethod
   def _threefry_random_bits(key, counts0, counts1):
     x = (counts1.cast(dtypes.uint64) << 32) | counts0.cast(dtypes.uint64)
-    x = x.threefry((key[1]._broadcast_to(x.shape).cast(dtypes.uint64) << 32) | key[0]._broadcast_to(x.shape).cast(dtypes.uint64))
+    x = x.threefry((key[1].cast(dtypes.uint64) << 32) | key[0].cast(dtypes.uint64))
     return (x & 0xffffffff).cast(dtypes.uint32).cat(((x >> 32) & 0xffffffff).cast(dtypes.uint32))
 
   @classmethod

tinygrad/nn/onnx.py

@@ -617,6 +617,8 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
   def Add(x:Tensor,y:Tensor, broadcast=None, axis=None): return x + y
   def Sub(x:Tensor|int,y:Tensor): return x - y # some test has input as int
   def Div(x:Tensor,y:Tensor): return x.div(y, rounding_mode='trunc' if dtypes.is_int(x.dtype) else None)
+  # ONNX Pow is (T, T1) -> T, the output takes the base dtype while Tensor.pow promotes base and exponent
+  def Pow(x:Tensor,y:Tensor): return x.pow(y).round().cast(x.dtype) if dtypes.is_int(x.dtype) else x.pow(y)
   def Less(x:Tensor,y:Tensor): return x < y
   def LessOrEqual(x:Tensor,y:Tensor): return x <= y
   def Greater(x:Tensor,y:Tensor): return x > y
@@ -1297,7 +1299,7 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
 
   return {
     # Tensor ops
-    **{op: getattr(Tensor, op.lower()) for op in ("Neg", "Reciprocal", "Pow", "Sqrt", "Sign", "Abs", "Exp", "Log", "Mish", "Sin", "Cos", "Tan",
+    **{op: getattr(Tensor, op.lower()) for op in ("Neg", "Reciprocal", "Sqrt", "Sign", "Abs", "Exp", "Log", "Mish", "Sin", "Cos", "Tan",
     "Asin", "Acos", "Atan", "Relu", "Sigmoid", "MatMul", "Floor", "Ceil", "IsNaN", "Softplus", "HardSwish", "Where", "Mul", "Sinh", "Cosh",
     "Tanh", "Softsign", "Asinh", "Acosh", "Atanh", "Elu", "Celu", "Selu", "Round", "Erf")},
     # Implemented ops

tinygrad/renderer/__init__.py

@@ -39,9 +39,10 @@ class Estimates:
           mem[(buf, u.op)] = smin(accessed, buf.max_numel() * buf.dtype.scalar().itemsize)
       if u.op is Ops.RANGE:
         mult_stack.append(mults)
-        mults *= cast(sint, u.src[0].ssimplify())
-        # SPECIAL are already counted in mults
-        mults = mults.substitute({x:x.const_like(0) for x in mults.toposort() if x.op is Ops.SPECIAL}) if isinstance(mults, UOp) else mults
+        if u.dtype is not dtypes.void:  # unbounded loop, unknown trip count
+          mults *= cast(sint, u.src[0].ssimplify())
+          # SPECIAL are already counted in mults
+          mults = mults.substitute({x:x.const_like(0) for x in mults.toposort() if x.op is Ops.SPECIAL}) if isinstance(mults, UOp) else mults
       elif u.op is Ops.END: mults = mult_stack.pop(-1)
       elif u.op is Ops.SPECIAL: mults *= cast(sint, u.src[0].ssimplify()) # NOTE: we don't push to the mult_stack here, you can't end these
       elif u.op is Ops.PARAM and u.arg.addrspace == AddrSpace.ALU and u.expr == 'core_id': mults *= int(u.vmax) + 1

tinygrad/renderer/cstyle.py

@@ -12,9 +12,11 @@ base_rewrite = PatternMatcher([
   # local/reg buffers
   (UPat(Ops.BUFFER, name="x"), lambda ctx,x: ctx.render_buffer(x)),
 
-  # range/if/endif
+  # range/loop/if/endif
+  (UPat(Ops.RANGE, dtypes.void, name="x"), lambda ctx,x: "for (;;) {"),
   (UPat(Ops.RANGE, name="x"),
    lambda ctx,x: f"for ({ctx.render_dtype(x.dtype)} {ctx[x]} = 0; {ctx[x]} < {ctx[x.src[0]]}; {ctx[x]}++) {{"),
+  (UPat(Ops.END, src=(UPat(), UPat(Ops.RANGE), UPat(name="c", dtype=dtypes.bool))), lambda ctx,c: f"  if (!({ctx[c]})) {{ break; }}\n}}"),
   (UPat(Ops.IF, name="x"), lambda ctx,x: f"if ({ctx[x.src[0]]}) {{"),
   (UPat((Ops.ENDIF, Ops.END)), lambda ctx: "}"),
 
@@ -227,14 +229,14 @@ class CStyleLanguage(Renderer):
 
       if u.op in {Ops.ENDIF, Ops.END}: depth -= 1
       if (u.op is not Ops.CAST or u.max_numel() == 1) and (u.op in {Ops.CONST, Ops.INDEX, Ops.SHRINK, Ops.CUSTOMI} or \
-        (u.op is Ops.LOAD and u.src[0].addrspace == AddrSpace.REG) or \
+        (u.op is Ops.LOAD and u.src[0].addrspace == AddrSpace.REG and child_count[u] == 1) or \
         (u.op is Ops.CAST and u.addrspace in (AddrSpace.GLOBAL, AddrSpace.LOCAL)) or \
         (u.op in {Ops.STACK, *(GroupOp.ALU-{Ops.WHERE}), Ops.CAST, Ops.BITCAST} and child_count[u] == 1 and not getenv("EXPAND_SSA"))):
         r[u] = l
       else:
         if u.op not in {Ops.RANGE, Ops.STORE, Ops.BUFFER} and u.dtype != dtypes.void:
           l = f"{self.render_type(u)} {r[u]} = {l}" + (";" if u.op is not Ops.SPECIAL else "")
-        kernel.append("  "*depth + l)
+        kernel.append("\n".join("  "*depth + line for line in l.split("\n")))
         if prefix: c[prefix] += 1  # if it was used, increment
       if u.op in {Ops.IF, Ops.RANGE}: depth += 1
     del self.r
@@ -595,3 +597,8 @@ class QCOMCLRenderer(OpenCLRenderer):
   def supported_dtypes(self):
     return {d for d in Renderer.supported_dtypes(self)
             if (d != dtypes.float16 or (bool(IMAGE) and bool(FLOAT16))) and d not in dtypes.fp8s+(dtypes.bfloat16,dtypes.double)}
+
+  # QCOM's load vectorizer emits invalid IR for vectorized bool loads ("Range types must match load type"), type bool buffers as uchar
+  def _render_dtype(self, dtype:DType, sz:int=1, addrspace=AddrSpace.ALU, mutable=True, override_ptr=False, shape=None):
+    if dtype == dtypes.bool and addrspace == AddrSpace.GLOBAL: dtype = dtypes.uint8
+    return super()._render_dtype(dtype, sz, addrspace, mutable, override_ptr, shape)

tinygrad/renderer/llvmir.py

@@ -96,6 +96,11 @@ base_rewrite = PatternMatcher([
   (UPat(Ops.WHERE, name="x"), lambda ctx,x:
    f"  {ctx[x]} = select {ldt(x.src[0].dtype)} {ctx[x.src[0]]}, {ldt(x.src[1].dtype)} {ctx[x.src[1]]}, {ldt(x.src[2].dtype)} {ctx[x.src[2]]}"),
 
+  # loop (a RANGE with no src is an unbounded loop header)
+  (UPat(Ops.RANGE, dtypes.void, name="l"), lambda ctx,l: f"  br label %loop_{ctx[l][1:]}\nloop_{ctx[l][1:]}:"),
+  (UPat(Ops.END, src=(UPat(), UPat(Ops.RANGE, dtypes.void, name="l"), UPat(name="c"))), lambda ctx,l,c:
+    f"  br i1 {ctx[c]}, label %loop_{ctx[l][1:]}, label %loop_exit_{ctx[l][1:]}\nloop_exit_{ctx[l][1:]}:"),
+
   # range
   (UPat(Ops.RANGE, name="r"), lambda ctx,r:
    f"  br label %loop_entry_{range_str(r)}\n"

tinygrad/renderer/ptx.py

@@ -116,6 +116,9 @@ string_rewrite = PatternMatcher([
   # simple
   (UPat(Ops.BUFFER, name="x"), lambda ctx, x: [] if x.addrspace == AddrSpace.REG else [
     f".shared .align 16 .b8 local{x.arg.slot}[{x.max_numel()*x.dtype.itemsize}];", f"mov.u64 {ctx.r[x]}, local{x.arg.slot}[0];"]),
+  (UPat(Ops.RANGE, dtypes.void, name="l"), lambda ctx, l: f"WAITLOOP_{ctx.uops.index(l)}:"),
+  (UPat(Ops.END, src=(UPat(), UPat(Ops.RANGE, dtypes.void, name="l"), UPat(name="c"))), lambda ctx, l, c:
+    f"@{ctx.r[c]} bra WAITLOOP_{ctx.uops.index(l)};"),
   (UPat(Ops.RANGE, name="r"), lambda ctx, r: [
     f"mov.u32 {ctx.r[r]}, -1;",
     f"bra END_{ctx.r[r][1:]};",
@@ -211,6 +214,7 @@ class PTXRenderer(Renderer):
       prefix, dtype = {Ops.CAST: ("cast", None), Ops.BITCAST: ("cast", None), Ops.END: ("pred", "pred"), Ops.RANGE: ("ridx", None),
         Ops.CONST: ("const", None), Ops.BUFFER: ("local", "u64"), Ops.INDEX: ("bidx", "u64"), Ops.SHRINK: ("bidx", "u64"),
         Ops.PARAM: ("dat", "u64" if u.addrspace is AddrSpace.GLOBAL else None), **{op: ("alu", None) for op in GroupOp.ALU}}.get(u.op, (None, None))
+      if u.op is Ops.RANGE and u.dtype == dtypes.void: prefix = None  # loop headers don't have a register
       if prefix: r[u] = ssa(prefix, u, dtype)
 
       l: str|list[str]|None = string_rewrite.rewrite(u, ctx=self)

tinygrad/runtime/ops_python.py

@@ -48,7 +48,6 @@ class PythonProgram:
     st = time.perf_counter()
     warp = list(itertools.product(*[range(x) for x in local_size[::-1]]))
     warp_size = len(warp)
-    void_ops = {Ops.END, Ops.BARRIER, Ops.IF, Ops.ENDIF, Ops.SINK, Ops.NOOP, Ops.GROUP, Ops.STORE}
     for idxs in itertools.product(*[range(x) for x in global_size[::-1]]):
       values: dict[UOp, Any] = {}
       pbufs: list[memoryview] = list(bufs)
@@ -57,11 +56,15 @@ class PythonProgram:
       i = 0
       while i < len(self.uops):
         u = self.uops[i]
-        src_values = [values[v] for v in u.src if v.op not in void_ops]
-        src_dtypes = [v.dtype for v in u.src if v.op not in void_ops]
+        src_values = [values[v] for v in u.src if v.dtype is not dtypes.void]
+        src_dtypes = [v.dtype for v in u.src if v.dtype is not dtypes.void]
         if getenv("TRACE"): print(i, u.op, u.dtype, u.arg, src_values, src_dtypes)
         if u.op is Ops.END:
-          i = self.uop_to_index[u.src[1]]
+          if len(u.src) == 3:
+            # conditional backedge on a loop: jump back while the condition is true
+            if values[u.src[2]][0]: i = self.uop_to_index[u.src[1]]
+            else: i += 1
+          else: i = self.uop_to_index[u.src[1]]
           continue
         if u.op is Ops.IF:
           exec_masks.append([x and y for x,y in zip(exec_masks[-1], src_values[0])])
@@ -71,7 +74,7 @@ class PythonProgram:
           exec_masks.pop()
           i += 1
           continue
-        if u.op in (Ops.BARRIER, Ops.SINK, Ops.NOOP, Ops.GROUP):
+        if u.op in (Ops.BARRIER, Ops.SINK, Ops.NOOP, Ops.GROUP) or (u.op is Ops.RANGE and u.dtype == dtypes.void):
           # in the python emulator, the warp is always in sync
           i += 1
           continue

tinygrad/schedule/indexing.py

@@ -196,6 +196,12 @@ def run_rangeify(tsink:UOp, debug:bool=False) -> tuple[UOp, IndexingContext]:
 
     if x.dtype == dtypes.weakint: continue  # TODO: why do I need this?
     ending_ranges[x] = sum([ending_ranges.get(u, []) for u in consumer_map[x]], [])
+    # ranges the consumers iterate that this node broadcasts over
+    ended = [rctx.range_map[c][0][i] for c in consumer_map[x] if c in rctx.range_map and c.op in GroupOp.Broadcastable
+             for i in broadcast_axes(x.shape, c.shape)]
+    broadcast_ending_ranges = list(UOp.sink(*ended).ranges)
+    # fusion decision: REDUCE before the broadcast
+    if x.op is Ops.REDUCE: ending_ranges[x] += broadcast_ending_ranges
 
     # *** the ranges on the output are
     #  1. new if this op is realized
@@ -254,6 +260,7 @@ def run_rangeify(tsink:UOp, debug:bool=False) -> tuple[UOp, IndexingContext]:
       if len(_realize_axis):
         rctx.realize_map[x] = _realize_axis
         out_rngs = tuple([(rctx.new_range(x.shape[i]) if i in _realize_axis else r) for i,r in enumerate(out_rngs)])
+    ending_ranges[x] += broadcast_ending_ranges
 
     # TODO: some ops don't have shape, enable this after the `.st` property is removed
     #assert len(out_rngs) == len(x.shape), \

tinygrad/schedule/multi.py

@@ -1,5 +1,5 @@
 from tinygrad.helpers import all_same, prod, getenv, ALLREDUCE_CAST
-from tinygrad.uop.ops import Ops, UOp, PatternMatcher, UPat, GroupOp, graph_rewrite
+from tinygrad.uop.ops import Ops, UOp, PatternMatcher, UPat, GroupOp, graph_rewrite, broadcast_axes, _broadcast_shape
 from tinygrad.dtype import dtypes
 from tinygrad.schedule.allreduce import handle_allreduce
 
@@ -47,20 +47,17 @@ def shard_srcs(msrcs:tuple[UOp, ...], axis:int) -> list[UOp]:
   assert all_same(devices), f"all buffers must have the same device {devices}"
   dcount = len(devices[0])
 
+  out_shape = _broadcast_shape(*[x.shape for x in msrcs])
   srcs:list[UOp] = []
   for mlb in msrcs:
-    if mlb.axis is None:
-      # no axis, shard it
-      assert mlb.op is not Ops.MULTI
-      srcs.append(mlb._shard(axis, dcount))
+    src_axis = axis - (len(out_shape)-len(mlb.shape))
+    if mlb.axis == src_axis:
+      # same axis, just copy through
+      srcs.append(mlb.src[0])
     else:
-      assert mlb.op is Ops.MULTI
-      if mlb.axis == axis:
-        # same axis, just copy through
-        srcs.append(mlb.src[0])
-      else:
-        # axis mismatch, copy to all devices, and shard it correctly
-        srcs.append(copy_multi(mlb, mlb.device)._shard(axis, dcount))
+      # otherwise every device gets the full copy, sharded iff this src has the axis (broadcast srcs stay whole)
+      full = mlb if mlb.axis is None else copy_multi(mlb, mlb.device)
+      srcs.append(full if axis in broadcast_axes(mlb.shape, out_shape) else full._shard(src_axis, dcount))
   return srcs
 
 def alu_multi(root:UOp):

tinygrad/tensor.py

@@ -6,7 +6,7 @@ if TYPE_CHECKING: import numpy
 from tinygrad.dtype import DType, DTypeLike, dtypes, ConstType, least_upper_dtype, to_dtype, strong_dtype, _from_np_dtype, _to_np_dtype, PyConst
 from tinygrad.helpers import all_int, getenv, fully_flatten, fetch, Metadata, TRACEMETA, TracingKey
 from tinygrad.helpers import cpu_profile, suppress_finalizing, disable_gc
-from tinygrad.uop.ops import UOp, Ops, sint, all_metadata, _index_to_concrete_int, Variable, _broadcast_shape
+from tinygrad.uop.ops import UOp, Ops, sint, all_metadata, _index_to_concrete_int, Variable
 from tinygrad.mixin.rand import RandMixin
 from tinygrad.schedule import create_linear_with_vars
 from tinygrad.device import Buffer, canonicalize_device
@@ -489,32 +489,6 @@ class Tensor(RandMixin):
   def __delitem__(self, indices) -> None:
     raise TypeError("Tensor does not support deleting items")
 
-  # ***** broadcasted elementwise ops *****
-
-  def where(self:Tensor, x:Tensor|ConstType|sint, y:Tensor|ConstType|sint) -> Tensor:
-    """
-    Returns a tensor of elements selected from either `x` or `y`, depending on `self`.
-    `output_i = x_i if self_i else y_i`.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    cond = Tensor([[True, True, False], [True, False, False]])
-    print(cond.where(1, 3).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    Tensor.manual_seed(42)
-    cond = Tensor.randn(2, 3)
-    print(cond.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print((cond > 0).where(cond, -float("inf")).numpy())
-    ```
-    """
-    if isinstance(x, Tensor): x, y = x._broadcasted(y)
-    elif isinstance(y, Tensor): y, x = y._broadcasted(x)
-    else: x, y = self.ufix(x)._broadcasted(y)
-    out_shape = _broadcast_shape(self.shape, x.shape)
-    return self.cast(dtypes.bool)._broadcast_to(out_shape)._apply_uop(UOp.where, x._broadcast_to(out_shape), y._broadcast_to(out_shape))
-
   # ***** op wrappers *****
 
   # unlike Tensors, UOps are immutable, so these don't go in mixin

tinygrad/uop/ops.py

@@ -113,7 +113,7 @@ def dtype_from_uop(op:Ops, src:tuple[UOp,...], arg:Any) -> DType|None:
   match op:
     case Ops.STORE | Ops.CALL | Ops.LINEAR | Ops.SINK | Ops.PROGRAM | Ops.SOURCE | \
          Ops.END | Ops.BARRIER | Ops.GROUP | Ops.IF | Ops.ENDIF | \
-         Ops.TUPLE | Ops.FUNCTION | Ops.CUSTOM_FUNCTION | Ops.WAIT | Ops.REWRITE_ERROR:
+         Ops.TUPLE | Ops.FUNCTION | Ops.CUSTOM_FUNCTION | Ops.REWRITE_ERROR:
       # always void
       return dtypes.void
     case Ops.CUSTOM | Ops.CUSTOMI | Ops.INS | Ops.PYLITERAL:
@@ -131,7 +131,7 @@ def dtype_from_uop(op:Ops, src:tuple[UOp,...], arg:Any) -> DType|None:
     case Ops.SIN | Ops.LOG2 | Ops.EXP2 | Ops.SQRT | Ops.RECIPROCAL:
       return least_upper_float(src[0].dtype)
     case Ops.WHERE:
-      assert src[0].dtype == dtypes.bool, f"where first arg isn't bool, it's {src[0].dtype}"
+      if src[0].dtype != dtypes.bool: raise RuntimeError(f"where cond must be bool, got {src[0].dtype}")
       return promo_dtype(src[1:])
     case Ops.STACK:
       if len(src) == 0: return dtypes.void
@@ -563,8 +563,8 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
   def ufix(self, x):
     if isinstance(x, UOp): return x
     # float self keeps its dtype for any scalar, int self only for int/Invalid scalars
-    if dtypes.is_float(self.dtype) or (dtypes.is_int(self.dtype) and isinstance(x, (int, InvalidType))): return self.const_like(x)
-    return self.const_like(x, dtypes.from_py(x))
+    dtype = self.dtype if dtypes.is_float(self.dtype) or (dtypes.is_int(self.dtype) and isinstance(x, (int, InvalidType))) else dtypes.from_py(x)
+    return UOp.const(dtype, x)
   def broadcast(self, count:int):
     if count == 1: return self
     return UOp(Ops.STACK, src=(self,)*count)
@@ -575,19 +575,12 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
   def end(self, *src:UOp): return UOp(Ops.END, src=(self,)+src) if len(src) else self
   def after(self, *src:UOp, **kwargs): return UOp(Ops.AFTER, src=(self,)+src, **kwargs) if len(src) else self
   def barrier(self, *src:UOp): return UOp(Ops.BARRIER, src=(self,)+src)
-  def wait(self, **kwargs): return UOp(Ops.WAIT, src=(self,), **kwargs)
   def ins(self, arg, **kwargs): return UOp(Ops.INS, kwargs.pop("dtype", self.dtype), kwargs.pop("src", self.src), arg, kwargs.pop("tag", self.tag))
   def contract(self, *rngs:UOp):
     assert all(x.arg[-1] == AxisType.UPCAST for x in rngs), "all contract ranges must be upcast"
     return UOp.stack(*[self.substitute(dict(zip(rngs, [r.const_like(i) for r,i in zip(rngs, idx)])))
                            for idx in itertools.product(*[range(int(r.vmax)+1) for r in rngs])])
-  def alu(self, op, *src:UOp, **kwargs):
-    all_srcs = (self, *src)
-    # broadcast shaped operands to a common shape (None and () are falsy, so only real shapes participate)
-    if (shapes := [s for x in all_srcs if (s:=x._shape)]) and not all_same(shapes):
-      out_shape = _broadcast_shape(*shapes)
-      all_srcs = tuple(x._broadcast_to(out_shape) if x._shape else x for x in all_srcs)
-    return UOp(op, src=all_srcs, **kwargs)
+  def alu(self, op, *src:UOp, **kwargs): return UOp(op, src=(self, *src), **kwargs)
   @staticmethod
   def const(dtype:DType, b:ConstLike, shape:tuple[sint, ...]|None=None):
     if isinstance(b, UOp): return b.cast(dtype)
@@ -602,6 +595,8 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
   def range(end:sint, axis_id, axis_type=AxisType.LOOP, *arg, dtype=dtypes.weakint, src=(), **kwargs):
     return UOp(Ops.RANGE, src=(sint_to_uop(end, dtype),)+src, arg=(axis_id, axis_type)+arg, **kwargs)
   @staticmethod
+  def loop(axis_id:int, *arg): return UOp(Ops.RANGE, dtypes.void, src=(UOp(Ops.NOOP),), arg=(axis_id, AxisType.LOOP)+arg)
+  @staticmethod
   def special(end:sint, name:str, dtype=dtypes.weakint): return UOp(Ops.SPECIAL, src=(sint_to_uop(end, dtype),), arg=name)
   @staticmethod
   def wmma(a:UOp, b:UOp, acc:UOp, dims:tuple[int, int, int], device:str, threads:int, tc_upcast_axes=None):
@@ -667,10 +662,9 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
       in_tuple = self.src[0].src[0] if self.src[0].op is Ops.FUNCTION else self.src[0]
       return in_tuple.src[self.arg].axis if in_tuple.op is Ops.TUPLE else None
     if self.op is Ops.PARAM: return self.arg.axis
-    # NOTE: they all have to share an axis, we always choose [-1]
-    if self.op in GroupOp.ALU: return axes[-1] if (axes := dedup([x.axis for x in self.src if x.axis is not None])) else None
-    # STACK adds a leading axis
-    if self.op is Ops.STACK: return axes[-1]+1 if (axes := dedup([x.axis for x in self.src if x.axis is not None])) else None
+    # NOTE: they all have to share an axis, we always choose [-1]. src axes are right-aligned into the output shape
+    if self.op in GroupOp.ALU.union({Ops.STACK}):
+      return axes[-1] if (axes := dedup([x.axis+len(self.shape)-len(x.shape) for x in self.src if x.axis is not None])) else None
     if len(self.src) == 0: return None
     src_axis = self.src[0].axis
     if self.op is Ops.SHRINK and src_axis is not None and self.marg[src_axis] != (0, self.src[0].shape[src_axis]):
@@ -1031,7 +1025,7 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
     if self.op is Ops.WHERE and dtypes.is_int(self.dtype): return min(self.src[1].vmin, self.src[2].vmin), max(self.src[1].vmax, self.src[2].vmax)
     # NOTE: returned UOp is assumed to be CONST
     if self.op is Ops.PARAM and self.arg.vmin_vmax is not None: return self.arg.vmin_vmax
-    if self.op in (Ops.RANGE, Ops.SPECIAL): return 0, (self.src[0]-1).vmax
+    if self.op in (Ops.RANGE, Ops.SPECIAL) and self.dtype is not dtypes.void: return 0, (self.src[0]-1).vmax
     if self.op is Ops.BIND: return self.src[0]._min_max # ignore the bound value
     if self.op is Ops.STACK: return min(x.vmin for x in self.src), max(x.vmax for x in self.src)
     if self.op is Ops.CONST and self.arg is not Invalid: return self.arg, self.arg

tinygrad/uop/render.py

@@ -34,6 +34,7 @@ def strip_binary_parens(x:UOp, left:str, right:str, code_for_op) -> str:
 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.arg)),
   (UPat(Ops.CAST, name="x"), lambda ctx,x: f"({str(x.dtype)[7:]})({ctx[x.src[0]]})"),

tinygrad/uop/spec.py

@@ -74,12 +74,15 @@ spec_shared = PatternMatcher([
   # CAST
   (UPat((Ops.BITCAST, Ops.CAST), src=(UPat(),), name="x"), lambda x: isinstance(x.arg, DType)),
 
-  # RANGE can be in the big graph now
+  # 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:
     rng.dtype == x.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, src=(UPat(),), allow_any_len=True, name="x"), lambda x: all(dtypes.is_int(y.dtype) for y in x.src[1:]) or None),
-  (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:])),
+  # 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)),
@@ -103,9 +106,6 @@ spec_shared = PatternMatcher([
   # BARRIER (on any length). TODO: this should only be in spec_program
   (UPat(Ops.BARRIER, dtypes.void), lambda: True),
 
-  # WAIT until a condition evaluates to true.
-  (UPat(Ops.WAIT, dtypes.void, src=(UPat(dtype=dtypes.bool),)), lambda: True),
-
   # assembly instruction
   (UPat(Ops.INS), lambda: True),