fixups

.github/actions/setup-tinygrad/action.yml

@@ -302,4 +302,4 @@ runs:
     - name: Install mesa (macOS)
       if: inputs.mesa == 'true' && runner.os == 'macOS'
       shell: bash
-      run: brew install sirhcm/tinymesa/tinymesa_cpu
+      run: brew install sirhcm/tinymesa/tinymesa

.github/workflows/benchmark.yml

@@ -626,11 +626,11 @@ jobs:
     - name: benchmark openpilot 0.9.9 dmonitoring
       run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=25 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=25 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
     - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=7 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/refs/heads/master/selfdrive/modeld/models/driving_policy.onnx
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=7 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/refs/heads/master/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=12 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/refs/heads/master/selfdrive/modeld/models/dmonitoring_model.onnx
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=12 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/refs/heads/master/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: benchmark MobileNetV2 on DSP
       run: |
         # generate quantized weights

autogen_stubs.sh

@@ -520,17 +520,17 @@ generate_mesa() {
   LVP_NIR_OPTIONS=$(./extra/mesa/lvp_nir_options.sh $MESA_SRC)
 
   fixup $BASE/mesa.py
-  patch_dlopen $BASE/mesa.py tinymesa_cpu "(BASE:=os.getenv('MESA_PATH', f\"/usr{'/local/' if helpers.OSX else '/'}lib\"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so')" "f'{BASE}/libtinymesa{EXT}'" "brew_path('tinymesa_cpu')" "brew_path('tinymesa')"
+  patch_dlopen $BASE/mesa.py tinymesa_cpu "(BASE:=os.getenv('MESA_PATH', f\"/usr{'/local/' if helpers.OSX else '/'}lib\"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so')" "f'{BASE}/libtinymesa{EXT}'" "f'{brew_prefix()}/lib/libtinymesa_cpu.dylib'"
   echo "lvp_nir_options = gzip.decompress(base64.b64decode('$LVP_NIR_OPTIONS'))" >> $BASE/mesa.py
   cat <<EOF | sed -i "/import ctypes.*/r /dev/stdin" $BASE/mesa.py
-def brew_path(nm):
-  try: return f"{subprocess.check_output(['brew', '--prefix', nm]).decode().strip()}/lib/lib{nm}.dylib"
-  except Exception: return 'failed'
+def brew_prefix():
+  try: return subprocess.check_output(['brew', '--prefix', 'tinymesa']).decode().strip()
+  except Exception: return ''
 EOF
-  sed -i "/in_dll/s/.*/try: &\nexcept (AttributeError, ValueError): pass/" $BASE/mesa.py
+  sed -i "/in_dll/s/.*/try: &\nexcept AttributeError: pass/" $BASE/mesa.py
   sed -i "s/import ctypes/import ctypes, ctypes.util, os, gzip, base64, subprocess, tinygrad.helpers as helpers/" $BASE/mesa.py
   sed -i "s/ctypes.CDLL('.\+')/(dll := _try_dlopen_tinymesa_cpu())/" $BASE/mesa.py
-  echo "def __getattr__(nm): raise AttributeError('LLVMpipe requires tinymesa_cpu' if 'tinymesa_cpu' not in dll._name else f'attribute {nm} not found') if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa ($TINYMESA_TAG, $MESA_TAG)')" >> $BASE/mesa.py
+  echo "def __getattr__(nm): raise AttributeError() if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa ($TINYMESA_TAG, $MESA_TAG)')" >> $BASE/mesa.py
   sed -i "s/ctypes.glsl_base_type/glsl_base_type/" $BASE/mesa.py
   # bitfield bug in clang2py
   sed -i "s/('fp_fast_math', ctypes.c_bool, 9)/('fp_fast_math', ctypes.c_uint32, 9)/" $BASE/mesa.py

examples/openpilot/compile3.py

@@ -121,12 +121,6 @@ def test_vs_onnx(new_inputs, test_val, onnx_file, tol):
   print("test vs onnx passed")
   return timings
 
-def bench(run, inputs):
-  from extra.bench_log import WallTimeEvent, BenchEvent
-  for _ in range(10):
-    with WallTimeEvent(BenchEvent.STEP):
-      run(**inputs).numpy()
-
 if __name__ == "__main__":
   onnx_file = fetch(OPENPILOT_MODEL)
   inputs, outputs = compile(onnx_file)
@@ -137,5 +131,3 @@ if __name__ == "__main__":
   if not getenv("FLOAT16"):
     test_vs_onnx(inputs, outputs, onnx_file, 1e-4)
 
-  if getenv("BENCHMARK_LOG", ""):
-    bench(pickle_loaded, inputs)

examples/sdv2.py

@@ -99,7 +99,6 @@ if __name__ == "__main__":
   parser.add_argument('--timing',      action='store_true', help="Print timing per step")
   parser.add_argument('--noshow',      action='store_true', help="Don't show the image")
   parser.add_argument('--fp16',        action='store_true', help="Cast the weights to float16")
-  parser.add_argument('--fakeweights', action='store_true', help="Skip loading checkpoints and use fake weights")
   args = parser.parse_args()
 
   N = 1
@@ -113,22 +112,19 @@ if __name__ == "__main__":
 
   model = StableDiffusionV2(**params)
 
-  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    if not args.fakeweights:
-      default_weights_url = 'https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors'
-      weights_fn = args.weights_fn
-      if not weights_fn:
-        weights_url = args.weights_url if args.weights_url else default_weights_url
-        weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))
+  default_weights_url = 'https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors'
+  weights_fn = args.weights_fn
+  if not weights_fn:
+    weights_url = args.weights_url if args.weights_url else default_weights_url
+    weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))
 
-        load_state_dict(model, safe_load(weights_fn), strict=False)
+  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
+    load_state_dict(model, safe_load(weights_fn), strict=False)
 
     if args.fp16:
       for k,v in get_state_dict(model).items():
         if k.startswith("model"):
-          v.replace(v.cast(dtypes.float16))
-
-    Tensor.realize(*get_state_dict(model).values())
+          v.replace(v.cast(dtypes.float16).realize())
 
   c  = { "crossattn": model.cond_stage_model(args.prompt) }
   uc = { "crossattn": model.cond_stage_model("") }

examples/stable_diffusion.py

@@ -263,16 +263,14 @@ if __name__ == "__main__":
   parser.add_argument('--timing', action='store_true', help="Print timing per step")
   parser.add_argument('--seed', type=int, help="Set the random latent seed")
   parser.add_argument('--guidance', type=float, default=7.5, help="Prompt strength")
-  parser.add_argument('--fakeweights', action='store_true', help="Skip loading checkpoints and use fake weights")
   args = parser.parse_args()
 
   model = StableDiffusion()
 
   # load in weights
   with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    if not args.fakeweights:
-      model_bin = fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt')
-      load_state_dict(model, torch_load(model_bin)['state_dict'], verbose=False, strict=False, realize=False)
+    model_bin = fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt')
+    load_state_dict(model, torch_load(model_bin)['state_dict'], verbose=False, strict=False, realize=False)
 
     if args.fp16:
       for k,v in get_state_dict(model).items():

test/external/external_benchmark_schedule.py

@@ -2,8 +2,7 @@ from extra.models.resnet import ResNet50
 from tinygrad import Tensor, nn, Device
 from tinygrad.helpers import Profiling, Timing, getenv
 from tinygrad.uop.ops import Ops
-from tinygrad.codegen import get_rewrites_for_renderer, apply_rewrites
-from tinygrad.codegen.late.control_flow import linearize
+from tinygrad.codegen import get_rewrites_for_renderer, apply_rewrites, rewrites_for_linearizer
 from tinygrad.uop.spec import type_verify
 
 if __name__ == "__main__":
@@ -40,7 +39,7 @@ if __name__ == "__main__":
           with Timing("***** model linearize in "):
             uops_line = []
             for u in rewritten_uops:
-              uops_line.append(linearize(u))
+              uops_line.append(apply_rewrites(u, rewrites_for_linearizer))
           with Timing("***** model verify in    "):
-            for u in uops_line: type_verify(u)
-          print(sum(len(u) for u in uops_line))
+            for u in uops_line: type_verify(u.arg.lst)
+          print(sum(len(u.arg.lst) for u in uops_line))

test/external/speed_v_theoretical.py

@@ -91,11 +91,11 @@ class TestKernelSpeed(unittest.TestCase):
 
   # theoretical is nv_tflops=165, amd_tflops=123
   def test_gemm_4096(self): self._test_matmul(4096, nv_tflops=115, amd_tflops=65)
-  def test_gemm_8192(self): self._test_matmul(8192, nv_tflops=115, amd_tflops=60)
+  def test_gemm_8192(self): self._test_matmul(8192, nv_tflops=125, amd_tflops=60)
 
   # theoretical is nv_gbs=1008, amd_gbs=960
   def test_gemv_16384_4096(self): self._test_matmul(16384, 4096, 1, nv_gbs=840, amd_gbs=750)
-  def test_gemv_4096_16384(self): self._test_matmul(4096, 16384, 1, nv_gbs=820, amd_gbs=750)
+  def test_gemv_4096_16384(self): self._test_matmul(4096, 16384, 1, nv_gbs=830, amd_gbs=750)
 
 if __name__ == '__main__':
   unittest.main()

test/test_linearizer.py

@@ -41,7 +41,7 @@ class TestLinearizer(unittest.TestCase):
   def _test_no_nested_ranges(self, lins, skip=None):
     for l in lins:
       range_in_acc = flatten([[x for x in u.src if x.op is Ops.RANGE] for u in l.uops if u.op is Ops.DEFINE_REG])
-      ranges = [u.op for u in l.uops if (u.op is Ops.RANGE and u in range_in_acc) or (u.op is Ops.END and u.src[0] in range_in_acc)]
+      ranges = [u.op for u in l.uops if (u.op is Ops.RANGE and u in range_in_acc) or (u.op is Ops.ENDRANGE and u.src[0] in range_in_acc)]
       for i,u in enumerate(ranges):
         if skip and i in skip: continue
         assert ranges[i-1] != u, f"multireduce nested the ranges! {ranges[i-1], {u}}"
@@ -205,7 +205,7 @@ class TestLinearizer(unittest.TestCase):
     # the uops graph is DEFINE_REG -> 4x STORE 0.0 -> RANGE -> 4x ALU -> 4x STORE -> ENDRANGE
     uops = get_program(ast, opts=opt).uops
     begin_range = [i for i, x in enumerate(uops) if x.op is Ops.RANGE][-1]
-    end_range = [i for i, x in enumerate(uops) if x.op is Ops.END][0]
+    end_range = [i for i, x in enumerate(uops) if x.op is Ops.ENDRANGE][0]
     for i,u in enumerate(uops): print(i, u.op, [uops.index(s) for s in u.src], u.arg, u.dtype)
     for u in uops:
       if u.op is Ops.STORE and isinstance(dt:=u.src[0].dtype, PtrDType) and dt.addrspace is AddrSpace.REG:
@@ -214,8 +214,8 @@ class TestLinearizer(unittest.TestCase):
         else:
           assert u.src[1].op in GroupOp.ALU
           assert begin_range < uops.index(u) < end_range
-      # children of END are placed after ENDRANGE
-      if any(x.op is Ops.END and x.src[1].op in GroupOp.ALU for x in u.src):
+      # children of STORE are placed after ENDRANGE
+      if any(x.op is Ops.STORE and x.src[1].op in GroupOp.ALU for x in u.src):
         assert end_range < uops.index(u)
 
   def test_grouped_dims(self):
@@ -400,7 +400,7 @@ class TestLinearizer(unittest.TestCase):
     # # check the children's vins
     # TODO: src ALU are not the same, should it?
     # assert barrier.src == tuple(local_stores)
-    assert len([u for u in uops if u.op is Ops.IF and u.src[1] == barrier]) == 1
+    assert len([u for u in uops if u.op is Ops.IF and u.src[-1] == barrier]) == 1
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")

test/test_ops.py

@@ -2602,7 +2602,6 @@ class TestOps(unittest.TestCase):
       lambda x: torch.nn.functional.avg_pool2d(x, kernel_size=(111,28)),
       lambda x: Tensor.avg_pool2d(x, kernel_size=(111,28)), rtol=1e-5)
 
-  @unittest.skipIf(Device.DEFAULT == "AMD" and CI, "remu failure?")
   def test_avg_pool3d_failure(self):
     with Context(NOOPT=0):
       helper_test_op([(1,1,16,16,16)],

test/test_rangeify.py

@@ -1,9 +1,7 @@
 import unittest
-from tinygrad import Tensor, nn, Device
-from tinygrad.helpers import Context, GlobalCounters, CI, getenv
+from tinygrad import Tensor, nn
+from tinygrad.helpers import Context, GlobalCounters, CI, CPU_LVP, getenv
 from tinygrad.uop.ops import graph_rewrite, PatternMatcher, UPat, Ops
-from tinygrad.renderer.ptx import PTXRenderer
-from tinygrad.renderer.nir import NIRRenderer
 
 class TestRangeifyAssign(unittest.TestCase):
   def test_assign_permuted(self):
@@ -42,7 +40,7 @@ elif getenv("BIG") > 0:
 else:
   BS, HEADS, SEQLEN, EMB = 4, 2, 16, 8
 
-@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (NIRRenderer, PTXRenderer)), "broken in LVP and PTX")
+@unittest.skipIf(CPU_LVP, "broken in LVP")
 class TestPcontig(unittest.TestCase):
   def test_flash_attention_bw(self):
     def fa_bw():
@@ -64,7 +62,7 @@ class TestPcontig(unittest.TestCase):
       Tensor.realize(*ret)
       return ret
 
-    with Context(PCONTIG=2, DEBUG=2):
+    with Context(PCONTIG=2, REAL_SUBSTITUTE=1, DEBUG=2):
       grads = fa_bw()
       print(f"{GlobalCounters.global_ops/1e9:.2f} GFLOPS")
 

test/test_schedule.py

@@ -2414,16 +2414,5 @@ class TestUOpBecome(unittest.TestCase):
     b.shrink(((0,4),)).assign(a_view).realize()
     self.assertListEqual(b.tolist(), [0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0])
 
-class TestScheduleMultioutput(unittest.TestCase):
-  def test_simplest_multioutput(self):
-    with Context(MULTIOUTPUT=1):
-      a = Tensor.ones(256, 256).contiguous().realize()
-      r = a.sum(axis=1)
-      b = r+1
-      c = r+2
-      run_schedule(check_schedule([b, c], 1))
-      np.testing.assert_allclose(b.numpy(), 257)
-      np.testing.assert_allclose(c.numpy(), 258)
-
 if __name__ == '__main__':
   unittest.main(verbosity=2)

test/test_tensor_uop.py

@@ -3,7 +3,7 @@ import numpy as np
 import unittest
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.engine.realize import run_schedule
-from tinygrad.uop.ops import UOp
+from tinygrad.uop.ops import Ops, UOp, UPat
 from tinygrad.helpers import SPLIT_REDUCEOP
 
 class TestTensorUOp(unittest.TestCase):
@@ -93,6 +93,7 @@ class TestTensorUOp(unittest.TestCase):
     out.realize()
     self.assertEqual(out.tolist(), Tensor.zeros(4, 8).tolist())
 
+reduce_kernel = UPat(Ops.SINK, src=(UPat(Ops.STORE, allow_any_len=True, src=(UPat(), UPat((Ops.REDUCE_AXIS, Ops.REDUCE))))))
 @unittest.skipUnless(SPLIT_REDUCEOP, "only for SPLIT_REDUCEOP")
 class TestReduceOp(unittest.TestCase):
   def test_no_split_reduce_kernel(self):
@@ -100,18 +101,23 @@ class TestReduceOp(unittest.TestCase):
     a = a.sum()
     sched = a.schedule()
     assert len(sched) == 1
+    assert reduce_kernel.match(sched[0].ast, {})
 
   def test_split_reduce_kernel_dim0(self):
     a = Tensor.rand(256, 255).realize()
     a = a.sum()
     sched = a.schedule()
     assert len(sched) == 2
+    for s in sched:
+      assert reduce_kernel.match(s.ast, {})
 
   def test_split_reduce_kernel_dim1(self):
     a = Tensor.rand(255, 256).realize()
     a = a.sum()
     sched = a.schedule()
     assert len(sched) == 2
+    for s in sched:
+      assert reduce_kernel.match(s.ast, {})
 
 if __name__ == "__main__":
   unittest.main()

test/test_uop_graph.py

@@ -665,6 +665,19 @@ class TestUOpGraph(unittest.TestCase):
     bad_gate = UOp.const(dtypes.int, 1)
     with self.assertRaises(AssertionError): to_uops_list([UOp(Ops.STORE, dtypes.void, (glbl0, idx, UOp.const(dtypes.int, 42), bad_gate))])
 
+  def test_switched_range_order(self):
+    glbl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
+    cf = UOp.const(dtypes.float, 0.0)
+    r1 = UOp.range(2, 0)
+    r2 = UOp.range(2, 1)
+    alu = UOp(Ops.MUL, dtypes.int, (r2, r1))
+    store = UOp(Ops.STORE, dtypes.void, (glbl.index(alu), cf))
+    uops = to_uops_list([store])
+    ranges = [x for x in uops if x.op is Ops.RANGE]
+    endranges = [x for x in uops if x.op is Ops.ENDRANGE]
+    # ranges are closed in the right order
+    self.assertEqual(endranges[-1].src[0], ranges[0])
+
 @track_rewrites()
 def expander_rewrite(sink): return graph_rewrite(sink, sym + expander)
 
@@ -832,6 +845,8 @@ class TestIFUOps(unittest.TestCase):
     if_uops = [u for u in sink.toposort() if u.op is Ops.IF]
     self.assertEqual(len(if_uops), 1)
     self.assertEqual(if_uops[0].src[0], gate)
+    for st in sink.src:
+      self.assertEqual(len(st.src), 2)
 
   def test_expand_ifs_one_gate(self):
     gbuf = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
@@ -848,6 +863,8 @@ class TestIFUOps(unittest.TestCase):
     if_uops = [u for u in sink.toposort() if u.op is Ops.IF]
     self.assertEqual(len(if_uops), 1)
     self.assertEqual(if_uops[0].src[0], gate)
+    for st in sink.src:
+      self.assertEqual(len(st.src), 2)
 
   # this will be fixed with the merge gated stores bounty
   @unittest.expectedFailure
@@ -862,6 +879,8 @@ class TestIFUOps(unittest.TestCase):
     if_uops = [u for u in sink.toposort() if u.op is Ops.IF]
     self.assertEqual(len(if_uops), 1)
     self.assertEqual(if_uops[0].src[0], gate)
+    for st in sink.src:
+      self.assertEqual(len(st.src), 2)
 
 class TestUOpTags(unittest.TestCase):
   def test_inc_by_one(self):

test/unit/test_block_reorder.py

@@ -0,0 +1,76 @@
+import unittest, random
+from tinygrad.dtype import dtypes
+from tinygrad.uop.ops import print_uops, UOp, Ops
+from tinygrad.codegen.late.linearize import block_reorder
+from tinygrad.renderer.cstyle import OpenCLRenderer
+
+def is_toposorted(lst:list[UOp]):
+  seen = set()
+  for u in lst:
+    if any(p not in seen for p in u.src): return False
+    seen.add(u)
+  return True
+
+class TestBlockReorder(unittest.TestCase):
+  def _test_randomize(self, golden:list[UOp]):
+    # test random order is always same
+    for _ in range(50):
+      # shuffle and form a valid toposort
+      lst = golden[:]
+      random.shuffle(lst)
+      topolst = []
+      for u in lst:
+        for p in u.toposort():
+          if p not in topolst: topolst.append(p)
+      assert is_toposorted(topolst)
+
+      for x,y in zip(golden, this_order:=block_reorder(topolst)):
+        if x is not y:
+          print_uops(golden)
+          print_uops(this_order)
+        self.assertIs(x, y)
+
+  def _test_render(self, golden:list[UOp]):
+    return OpenCLRenderer().render(golden)
+
+  def test_loads(self):
+    a = UOp(Ops.DEFINE_GLOBAL, dtype=dtypes.float.ptr(), arg=0)
+    b = UOp(Ops.DEFINE_GLOBAL, dtype=dtypes.float.ptr(), arg=1)
+    c = UOp(Ops.DEFINE_GLOBAL, dtype=dtypes.float.ptr(), arg=2)
+    v1 = UOp(Ops.SPECIAL, dtype=dtypes.int, src=(UOp.const(dtypes.int, 4),), arg="gidx0")
+    v2 = UOp(Ops.SPECIAL, dtype=dtypes.int, src=(UOp.const(dtypes.int, 4),), arg="gidx1")
+    v1 = v1*27
+    v2 = v2*4
+    loads = [
+      a.index(v1).load(dtype=dtypes.float),
+      a.index(v1+1).load(dtype=dtypes.float),
+      a.index(v1+2).load(dtype=dtypes.float),
+      a.index(v1+3).load(dtype=dtypes.float),
+      b.index(v2).load(dtype=dtypes.float),
+      b.index(v2+1).load(dtype=dtypes.float),
+      b.index(v2+2).load(dtype=dtypes.float),
+      b.index(v2+3).load(dtype=dtypes.float)]
+    #random.shuffle(loads)
+    sink = c.store(sum(loads)).sink()
+
+    # determine golden order
+    golden = block_reorder(list(sink.toposort()))
+
+    # render for test
+    print(self._test_render(golden))
+    #print_uops(golden)
+
+    # assert the loads are in this order
+    self.assertListEqual([g.src[0].src[1].render() for g in golden if g.op is Ops.LOAD],
+                         ['(gidx1*4)', '((gidx1*4)+1)', '((gidx1*4)+2)', '((gidx1*4)+3)',
+                          '(gidx0*27)', '((gidx0*27)+1)', '((gidx0*27)+2)', '((gidx0*27)+3)'])
+
+    # assert math is after loads
+    first_math = [i for i,g in enumerate(golden) if g.op is Ops.ADD and g.dtype == dtypes.float][0]
+    assert not any(x.op is Ops.LOAD for x in golden[first_math:])
+
+    # confirm the sort is stable
+    self._test_randomize(golden)
+
+if __name__ == '__main__':
+  unittest.main()

test/unit/test_simplify_valid_idx.py

@@ -5,7 +5,7 @@ from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import UOp, Ops
 from tinygrad.uop.symbolic import simplify_valid
 from tinygrad.helpers import Context
-from test.unit.test_uop_symbolic import check_uop_against_string
+from .test_uop_symbolic import check_uop_against_string
 
 def get_gated_load_uop(valid:UOp, idx:UOp):
   return UOp(Ops.LOAD, dtypes.float, (

test/unit/test_viz.py

@@ -148,14 +148,6 @@ class TestViz(BaseTestViz):
     a2 = uop_to_json(a)[id(a)]
     self.assertEqual(ansistrip(a2["label"]), f"CUSTOM\n{TestStruct.__qualname__}(colored_field='xyz12345')")
 
-  def test_colored_label_multiline(self):
-    arg = colored("x", "green")+"\n"+colored("y", "red")+colored("z", "yellow")+colored("ww\nw", "magenta")
-    src = [Tensor.empty(1).uop for _ in range(10)]
-    a = UOp(Ops.CUSTOM, src=tuple(src), arg=arg)
-    exec_rewrite(a, [PatternMatcher([])])
-    a2 = next(get_viz_details(0, 0))["graph"][id(a)]
-    self.assertEqual(ansistrip(a2["label"]), "CUSTOM\nx\nyzww\nw")
-
   def test_inf_loop(self):
     a = UOp.variable('a', 0, 10, dtype=dtypes.int)
     b = a.replace(op=Ops.CONST)

tinygrad/codegen/__init__.py

@@ -14,10 +14,10 @@ from tinygrad.uop.decompositions import get_late_rewrite_patterns
 from tinygrad.codegen.late.expander import migrate_indexing, expander, pm_pre_expander, pm_group_for_reduce
 from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
   ReduceContext, correct_load_store, pm_render
+from tinygrad.codegen.late.linearize import block_create, pm_blockend_merge, block_merge, pm_finalize, BlockContext
 from tinygrad.codegen.opt.postrange import pm_postrange_opt
 from tinygrad.codegen.simplify import pm_simplify_ranges, pm_reduce_simplify, pm_flatten_range, pm_split_ranges
 from tinygrad.schedule.rangeify import pm_add_buffers, rangeify_codegen
-from tinygrad.codegen.late.control_flow import CFGContext, pm_merge_ends, pm_add_control_flow, linearize
 
 @dataclass
 class RewriteStep:
@@ -30,6 +30,12 @@ class RewriteStep:
 
 def apply_rewrites(sink:UOp, rewrites:list[RewriteStep]): return functools.reduce(lambda x,f: f(x), rewrites, sink)
 
+rewrites_for_linearizer = [
+  RewriteStep(block_create, ctx=BlockContext.from_sink, name="Linearizer: Create Blocks", bottom_up=True),
+  RewriteStep(pm_blockend_merge, name="Linearizer: Merge Blockends"),
+  RewriteStep(block_merge, name="Linearizer: Merge Blocks"),
+  RewriteStep(pm_finalize, name="Linearizer: Finalize")]
+
 def get_rewrites_for_renderer(opts:Renderer, optimize:bool=True, linearizer:bool=True) -> list[RewriteStep]:
   # cache with the values of the context vars
   return _get_rewrites_for_renderer(opts, optimize, linearizer, QUANTIZE.value, DEVECTORIZE.value, TRANSCENDENTAL.value)
@@ -95,15 +101,11 @@ def _get_rewrites_for_renderer(opts:Renderer, optimize:bool, linearizer:bool, _Q
   pm_final_rewrite = pm_decomp+pm_render+extra_matcher
   ret.append(RewriteStep(pm_final_rewrite, lambda _: opts.device, name="final rewrite"))
 
-  # this was the linearizer
-  ret.append(RewriteStep(pm_merge_ends, name="merge ends"))
-  ret.append(RewriteStep(pm_add_control_flow, CFGContext, name="add control flow starts", bottom_up=True))
+  # return the list (with optional linearizer)
+  return ret + (rewrites_for_linearizer if linearizer else [])
 
-  # return the list
-  return ret
-
-def full_rewrite_to_sink(sink:UOp, opts:Renderer|None=None, optimize:bool=True) -> UOp:
-  return apply_rewrites(sink, get_rewrites_for_renderer(opts if opts is not None else Renderer(), optimize))
+def full_rewrite_to_sink(sink:UOp, opts:Renderer|None=None, optimize:bool=True, linearizer:bool=False) -> UOp:
+  return apply_rewrites(sink, get_rewrites_for_renderer(opts if opts is not None else Renderer(), optimize, linearizer))
 
 def full_rewrite(sink:UOp, opts:Renderer|None=None) -> list[UOp]:
   """
@@ -117,6 +119,6 @@ def full_rewrite(sink:UOp, opts:Renderer|None=None) -> list[UOp]:
     Linear program in UOps.
   """
 
-  lst = linearize(full_rewrite_to_sink(sink, opts, optimize=sink.tag is None))
+  lst = list(full_rewrite_to_sink(sink, opts, optimize=sink.tag is None, linearizer=True).arg.lst)
   if __debug__: type_verify(lst)
   return lst

tinygrad/codegen/gpudims.py

@@ -1,4 +1,4 @@
-import math, functools, operator
+import math
 from tinygrad.uop.ops import UOp, Ops, sint, PatternMatcher, UPat, KernelInfo, ssimplify, AxisType, sint_to_uop
 from tinygrad.helpers import all_int, dedup, get_contraction
 from tinygrad.dtype import dtypes
@@ -87,15 +87,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
     except ValueError: continue
   return s.substitute(subs)
 
-def add_barrier_and_if(buf:UOp, e:UOp):
-  # TODO: this is not generic
-  local_ranges = [x for x in e.ended_ranges if x.op is Ops.RANGE and x.arg[-1] == AxisType.GROUP_REDUCE]
-  if len(local_ranges) == 0: return None
-  return buf.after(UOp(Ops.IF, dtype=dtypes.void, src=(functools.reduce(operator.and_, [x.eq(0) for x in local_ranges]), e.barrier())))
-
 pm_add_gpudims = PatternMatcher([
   # add gpudims must be last
   (UPat(Ops.SINK, name="s"), add_gpudims),
-  # add barrier and if
-  (UPat(Ops.AFTER, src=(UPat(Ops.DEFINE_LOCAL, name="buf"), UPat(Ops.END, name="e"))), add_barrier_and_if),
 ])

tinygrad/codegen/late/control_flow.py

@@ -1,102 +0,0 @@
-import heapq
-from collections import defaultdict
-from tinygrad.uop.ops import PatternMatcher, UOp, Ops, UPat
-
-def linearize(u:UOp) -> list[UOp]:
-  lst = list(u.toposort())
-  in_this_block = set(lst)
-  local_children: defaultdict[UOp, list[UOp]] = defaultdict(list)
-  in_degree:dict[UOp, int] = {}
-  priorities:dict[UOp, int] = {}
-
-  # get local children and assign priorities
-  # NOTE: this requires the lst be locally toposorted
-  for u in reversed(lst):
-    in_degree[u] = 0
-    for s in u.src:
-      if s in in_this_block:
-        local_children[s].append(u)
-        in_degree[u] += 1
-    # put loads in the beginning of the block and prevent priority inversion. hack for BARRIER grouping too
-    priority = [0] + [priorities[x] for x in local_children[u]]
-    if u.op is Ops.LOAD: priority.append(-1000)
-    if u.op is Ops.BARRIER: priority.append(-1500)
-    # ranges are scheduled as late as possible so anything that can be outside is
-    #if u.op is Ops.RANGE: priority = [2000]
-    # move defines and consts to the top
-    if u.op in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.DEFINE_VAR, Ops.SPECIAL, Ops.CONST}: priority.append(-2000)
-    priorities[u] = min(priority)
-
-  # number the uops in "ideal" order
-  nkey = {u:i for i,u in enumerate(sorted(lst, key=lambda x: (priorities[x],)+x.tuplize))}
-
-  # then force then to be toposorted in as close to the ideal order as possible
-  heapq.heapify(heap:=[(nkey[u],u) for u in lst if in_degree[u] == 0])
-  newlst = []
-  while heap:
-    newlst.append(u:=heapq.heappop(heap)[1])
-    for v in local_children[u]:
-      in_degree[v] -= 1
-      if in_degree[v] == 0: heapq.heappush(heap, (nkey[v],v))
-
-  assert len(newlst) == len(lst), f"len mismatch {len(newlst)} != {len(lst)}"
-  return newlst
-
-class CFGContext:
-  def __init__(self, sink:UOp):
-    # there are 3 relationships between ranges:
-    # nested, meaning endrange y is a dependency of endrange x and range x is a dependency of endrange y
-    # dependent, meaning endrange y is a dependency of endrange x and range x is not a dependency of endrange y
-    # independent, endrange y is not a dependency of endrange x
-    # everything is nested inside the sink
-    deps: dict[UOp, set[UOp]] = {}
-    nesting: dict[UOp, UOp] = {}
-    for u in sink.toposort():
-      deps[u] = set().union(*(deps[s] for s in u.src))
-      if u.op in (Ops.END, Ops.ENDIF, Ops.SINK):
-        nesting |= {x:u for x in deps[u] if x.op in (Ops.END, Ops.ENDIF) and (u.op is Ops.SINK or u.src[0] in deps[x]) and x not in nesting}
-      if u.op in (Ops.RANGE, Ops.END, Ops.IF, Ops.ENDIF): deps[u] |= {u}
-
-    self.edges: dict[UOp, UOp] = {}
-    siblings: dict[UOp, list[UOp]] = {}
-    for k,vv in nesting.items(): siblings.setdefault(vv, []).append(k)
-    for k,v in siblings.items():
-      # range/if that have dependencies on other siblings need to run after them
-      order = sorted(v, key=lambda x: len(deps[x].intersection(v)))
-      zipped = zip(order, order[1:]) if k.op is Ops.SINK else zip([k.src[0]] + order, order)
-      for x,y in zipped:
-        # TODO: is this check correct?
-        if y.src[0] not in x.backward_slice_with_self:
-          self.edges[y.src[0]] = x
-
-pm_add_control_flow = PatternMatcher([
-  (UPat((Ops.RANGE, Ops.IF), name="x"), lambda ctx,x: x.replace(src=x.src+(y,)) if (y:=ctx.edges.get(x)) is not None else None),
-])
-
-def do_merge_ends(s:UOp):
-  # NOTE: this can fail
-  stacked: dict[UOp, list[UOp]] = {}
-  dangling_ifs = []
-  for x in s.toposort():
-    if x.op in {Ops.END, Ops.ENDIF}:
-      assert x.op is not Ops.END or x.arg == 1, "ends must be single ends for linearizer"
-      stacked.setdefault(x.src[0], []).append(x)
-    if x.op is Ops.IF: dangling_ifs.append(x)
-  dangling_ifs = [x for x in dangling_ifs if x not in stacked]
-  replaces = {}
-  for k,v in stacked.items():
-    if len(v) == 1: continue
-    rep = UOp(v[0].op, src=tuple([k] + [y for x in v for y in x.src[1:]]), arg=v[0].arg)
-    for x in v: replaces[x] = rep
-  if not len(replaces) and not len(dangling_ifs): return None
-  ret = s.substitute(replaces)
-  if len(dangling_ifs):
-    assert len(dangling_ifs) == 1, "we only support 1 dangling if"
-    ret = ret.replace(src=(UOp(Ops.ENDIF, src=(dangling_ifs[0], *ret.src)),))
-  return ret
-
-pm_merge_ends = PatternMatcher([
-  # for renderering and linearizing, all ends must end one loop
-  (UPat(Ops.END, name="e"), lambda e: e.replace(src=e.src[e.arg-1:], arg=1).end(ends=e.src[:e.arg-1]) if e.arg > 1 else None),
-  (UPat(Ops.SINK, name="s"), do_merge_ends),
-])

tinygrad/codegen/late/devectorizer.py

@@ -268,9 +268,9 @@ pm_render = PatternMatcher([
     UPat.var("a")), lambda c,idx,l,a: l.replace(src=(l.src[0], a.cast(l.dtype))+l.src[2:]).cast(a.dtype)),
   (UPat.var("c").where(UPat.var("a"), UPat(Ops.LOAD, src=(UPat().index(UPat.var("idx"), UPat.var("c").logical_not()).or_casted(),),
     allow_any_len=True, name="l").or_casted()), lambda c,idx,l,a: l.replace(src=(l.src[0], a.cast(l.dtype))+l.src[2:]).cast(a.dtype)),
-  # gate any stores that aren't gated with if/endif pairs
+  # gate any stores that aren't gated with ifs
   (UPat(Ops.STORE, src=(UPat(src=(UPat(), UPat(), UPat(dtype=dtypes.bool)), name="idx").or_casted(), UPat()), name="store", allow_any_len=True),
-    lambda store,idx: UOp(Ops.ENDIF, src=(uif:=UOp(Ops.IF, src=(idx.src[2],)), UOp(Ops.STORE, src=store.src[:2]+(uif,)+store.src[2:]))) if \
+    lambda store,idx: UOp(Ops.STORE, dtype=store.dtype, src=store.src[:2]+(UOp(Ops.IF, src=(idx.src[2],)),)+store.src[2:]) if \
       len(store.src) <= 2 or store.src[2].op != Ops.IF else None),
 ])
 
@@ -295,8 +295,8 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
   # if we have a range
   if len(reduce_range) != 0:
     topo = inp.toposort()
-    ended_ranges = flatten([x.src[:x.arg] for x in topo if x.op is Ops.END])
-    input_ranges = tuple([x for x in topo if x.op is Ops.RANGE and x not in reduce_range and x not in ended_ranges])
+    stored_ranges = flatten([x.src[2:] for x in topo if x.op is Ops.STORE])
+    input_ranges = tuple([x for x in topo if x.op is Ops.RANGE and x not in reduce_range and x not in stored_ranges])
     identity = red.const(red.dtype, identity_element(red.arg, red.dtype.scalar()))
     acc = UOp(Ops.DEFINE_REG, red.dtype.ptr(size=1, addrspace=AddrSpace.REG), arg=(ctx.acc_num,))
     acc_init = acc.after(*input_ranges).index(UOp.const(dtypes.int, 0)).store(identity) if len(input_ranges) else \
@@ -305,7 +305,7 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
     ctx.acc_num += 1
   ret = functools.reduce(lambda x,y: x.alu(red.arg, y), lst)
   if len(reduce_range) == 0: return ret
-  return acc.after(acc.index(UOp.const(dtypes.int, 0)).store(ret).end(ends=reduce_range[::-1])).index(UOp.const(dtypes.int, 0)).load()
+  return acc.after(acc.index(UOp.const(dtypes.int, 0)).store(ret, *reduce_range)).index(UOp.const(dtypes.int, 0)).load()
 
 pm_reduce = PatternMatcher([
   # REDUCE -> DEFINE_ACC+ASSIGN

tinygrad/codegen/late/expander.py

@@ -87,7 +87,7 @@ expander = PatternMatcher([
    lambda outer, inner: UOp(Ops.UNROLL, outer.dtype, (inner.src[0],), inner.arg+outer.arg)),
   # do expansion
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.GEP, Ops.WMMA, Ops.LOAD, Ops.STORE, Ops.INDEX, Ops.BUFFERIZE,
-         Ops.VECTORIZE, Ops.IF, Ops.REDUCE, Ops.END), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
+         Ops.VECTORIZE, Ops.IF, Ops.REDUCE), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
   (UPat(Ops.CONTRACT, name="con"), do_contract),
   # BARRIERs aren't actually expanded
   (UPat(Ops.BARRIER, src=(UPat(Ops.UNROLL, name="ex"),)),
@@ -145,7 +145,8 @@ def fix_group_for_reduce(x:UOp):
   reduce_loop = [x.replace(arg=(x.arg[0]+100, AxisType.REDUCE)) for x in reduce_gfr]
   buf = ret.bufferize(*upstream_locals, *reduce_gfr, arg=BufferizeOpts(reduce_gfr[0].arg[0], AddrSpace.LOCAL)).index(*upstream_locals, *reduce_loop)
 
-  # do the final reduce (if/barrier are added in gpudims step)
+  # gate with an if on the store + do the final reduce
+  buf = UOp(Ops.IF, dtype=buf.dtype, src=(functools.reduce(operator.and_, [x.eq(0) for x in reduce_gfr]), buf))
   return buf.reduce(*reduce_loop, arg=x.arg)
 
 pm_pre_expander = PatternMatcher([

tinygrad/codegen/late/linearize.py

@@ -0,0 +1,243 @@
+from __future__ import annotations
+import heapq
+from collections import defaultdict
+from dataclasses import dataclass, replace
+from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp, BottomUpGate
+from tinygrad.helpers import dedup, all_same, flatten, BLOCK_REORDER
+
+# NOTE: any toposort should be valid here, unlike last time this isn't required, it's just for speed
+def block_reorder(lst:list[UOp]) -> list[UOp]:
+  in_this_block = set(lst)
+  local_children: defaultdict[UOp, list[UOp]] = defaultdict(list)
+  in_degree:dict[UOp, int] = {}
+  priorities:dict[UOp, int] = {}
+
+  # get local children and assign priorities
+  # NOTE: this requires the lst be locally toposorted
+  for u in reversed(lst):
+    in_degree[u] = 0
+    for s in u.src:
+      if s in in_this_block:
+        local_children[s].append(u)
+        in_degree[u] += 1
+    # put loads in the beginning of the block and prevent priority inversion. hack for BARRIER grouping too
+    priority = [0] + [priorities[x] for x in local_children[u]]
+    if u.op is Ops.LOAD: priority.append(-1000)
+    if u.op is Ops.BARRIER: priority.append(-1500)
+    priorities[u] = min(priority)
+
+  # number the uops in "ideal" order
+  nkey = {u:i for i,u in enumerate(sorted(lst, key=lambda x: (priorities[x],)+x.tuplize))}
+
+  # then force then to be toposorted in as close to the ideal order as possible
+  heapq.heapify(heap:=[(nkey[u],u) for u in lst if in_degree[u] == 0])
+  newlst = []
+  while heap:
+    newlst.append(u:=heapq.heappop(heap)[1])
+    for v in local_children[u]:
+      in_degree[v] -= 1
+      if in_degree[v] == 0: heapq.heappush(heap, (nkey[v],v))
+
+  assert len(newlst) == len(lst), f"len mismatch {len(newlst)} != {len(lst)}"
+  return newlst
+
+# ***** basic block *****
+
+def disp(y:UOp) -> str:
+  if y.op is Ops.IF: return f'IF{id(y)}'
+  if y.op is Ops.RANGE: return str(y.arg)
+  return "<NONE>"
+
+@dataclass(frozen=True, eq=False)
+class BasicBlock:
+  lst: tuple[UOp, ...]
+  ctx: tuple[UOp, ...] = ()
+  end: UOp|None = None
+  cnt: int = 0
+  child_ctx: tuple[UOp, ...]|None = None
+  def __lt__(self, _:BasicBlock): raise RuntimeError("no comparing basic blocks")
+  def __repr__(self):
+    return f"{(str(disp(self.end))+' ') if self.end is not None else ''}"+f'f{self.cnt} '+\
+           f"{[disp(y) for y in self.ctx]} {[disp(y) for y in self.child_ctx] if self.child_ctx is not None else '-'} "+\
+           f"{len(self.lst)}" + "\n" + '\n'.join([str(x.op) for x in self.lst])
+  def last_ctx(self): return self.child_ctx if self.child_ctx is not None else self.ctx
+
+def _sort_ctx(inp): return tuple(sorted(dedup(inp), key=lambda x: x.tuplize))
+
+# ***** block context *****
+
+@dataclass
+class BlockContext:
+  child_count: dict[UOp, int]
+  block_ctxs: dict[UOp, tuple[UOp, ...]]
+  child_ctxs: dict[UOp, tuple[UOp, ...]]
+  def last_ctx(self, u): return self.child_ctxs.get(u, self.block_ctxs[u])
+  @staticmethod
+  def from_sink(sink:UOp) -> BlockContext:
+    # get children and all block contexts
+    ctx = BlockContext({}, {}, {})
+    for u in sink.toposort(gate=lambda u:u.op is not Ops.SPECIAL):
+      this_block_ctx: list[UOp] = []
+      ctx.child_count[u] = 0
+
+      # get children and accumulate the last_ctx
+      for s in u.src:
+        if s.op is Ops.SPECIAL: continue
+        # NOTE: if a parent appears multiple times in the src, it counts multiple times as a child
+        ctx.child_count[s] += 1
+        this_block_ctx += ctx.last_ctx(s)
+
+      # save the block ctx. SINK never has anything
+      ctx.block_ctxs[u] = _sort_ctx(this_block_ctx) if u.op is not Ops.SINK else ()
+
+      # RANGE/IF add to the next ctx
+      # STORE/ASSIGN subtract from the next ctx
+      if u.op in {Ops.RANGE, Ops.IF}: ctx.child_ctxs[u] = _sort_ctx(ctx.block_ctxs[u] + (u,))
+      elif u.op is Ops.STORE: ctx.child_ctxs[u] = tuple([y for y in ctx.block_ctxs[u] if y not in u.src])
+    return ctx
+
+# ***** make blocks *****
+
+DONT_PLACE_IN_BLOCK = {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.DEFINE_VAR, Ops.SPECIAL, Ops.CONST}
+
+def add_blockends(base_block:UOp, new_ctx:tuple[UOp, ...], current_ctx:tuple[UOp, ...], cnt:int=1) -> UOp:
+  ends_to_add = [z for z in new_ctx if z not in current_ctx]
+  while len(ends_to_add):
+    r:UOp = ends_to_add.pop(-1)
+    new_ctx = tuple([z for z in new_ctx if z is not r])
+    end_uop = UOp(Ops.ENDIF if r.op is Ops.IF else Ops.ENDRANGE, src=(r,))
+    base_block = UOp(Ops.BLOCKEND, src=(base_block,)*cnt, arg=BasicBlock((end_uop,), tuple(new_ctx), end=r, cnt=cnt))
+  return base_block
+
+def make_block_bottom_up(ctx:BlockContext, x:UOp):
+  if x.op is Ops.BLOCKSTART:
+    current_ctx, child_ctx = x.arg
+    lst = list(x.src)
+    child_count = 1
+  else:
+    current_ctx, child_count, child_ctx = ctx.block_ctxs[x], ctx.child_count[x], ctx.child_ctxs.get(x, None)
+    lst = [x]
+
+  # count of times we've seen this block, or a seed for a new block if we can't merge it
+  unmergable: defaultdict[UOp, int] = defaultdict(int)
+  blockseeds = defaultdict(list)
+
+  # add the srcs of this to the frontier
+  # NOTE: things may be in here multiple times, that's okay
+  frontier_nodes = list(flatten(y.src[::-1] for y in lst))
+  while len(frontier_nodes):
+    u = frontier_nodes.pop(0)
+    if u.op not in DONT_PLACE_IN_BLOCK and ctx.child_count[u] == unmergable[u]+1:
+      # count is correct
+      if (newctx:=ctx.block_ctxs[u]) == current_ctx:
+        # block has same context, merge it, and put the srcs on the frontier
+        lst.append(u)
+        frontier_nodes.extend(u.src[::-1])
+      else:
+        # block has different context, add it to blockseeds
+        blockseeds[(newctx, ctx.child_ctxs.get(u, None))].append(u)
+      del unmergable[u]
+    else:
+      # count is incorrect (or it's DONT_PLACE_IN_BLOCK), add it to unmergable
+      unmergable[u] += 1
+
+  # add unmergables to sources
+  srcs = []
+  for u,cnt in unmergable.items(): srcs += [add_blockends(u, ctx.block_ctxs.get(u,()), current_ctx, cnt=cnt)]*cnt
+
+  # add blockseeds, with blockends as needed
+  for (new_ctx, new_child_ctx), v in blockseeds.items():
+    base_block = UOp(Ops.BLOCKSTART, src=tuple(v), arg=(new_ctx, new_child_ctx))
+    srcs.append(add_blockends(base_block, new_ctx, current_ctx))
+
+  lst = lst[::-1]
+  if BLOCK_REORDER: lst = block_reorder(lst)
+  bb = BasicBlock(tuple(lst), ctx=current_ctx, cnt=child_count, child_ctx=child_ctx)
+  return UOp(Ops.BLOCK, src=tuple(srcs), arg=bb)
+
+# we prevent the source of the SPECIAL from being linearized since its not part of the kernel
+def raise_bottom_up_gate(): raise BottomUpGate()
+
+block_create = PatternMatcher([
+  (UPat(GroupOp.All-DONT_PLACE_IN_BLOCK.union({Ops.BLOCK, Ops.BLOCKEND}), name="x"), make_block_bottom_up),
+  (UPat(Ops.SPECIAL), raise_bottom_up_gate)
+])
+
+# ***** blockend merging ****
+
+def merge_blockends(sink:UOp) -> UOp|None:
+  # only run on the final BLOCK with the SINK in it
+  if sink.arg.lst[-1].op is not Ops.SINK: return None
+  # combine matching BLOCKENDS, the keys of this dictionary are the RANGE UOps, values are the BLOCKENDs
+  blockends_to_arg: dict[UOp, list[UOp]] = {}
+  for be in sink.toposort():
+    if be.op is Ops.BLOCKEND: blockends_to_arg.setdefault(be.arg.end, []).append(be)
+  new_forks = {}
+  for k,v in blockends_to_arg.items():
+    # NOTE: if any BLOCKEND is the parent of any other with the same arg, this algo fails
+    if len(v) > 1:
+      bb = BasicBlock(v[0].arg.lst, _sort_ctx(flatten([y.arg.ctx for y in v])), k, cnt=sum(y.arg.cnt for y in v))
+      out = UOp(Ops.BLOCKEND, src=tuple(flatten([x.src for x in v])), arg=bb)
+      # NOTE: bb.ctx != u.arg.ctx can cause problems here
+      for u in v: new_forks[u] = out
+  if len(new_forks) == 0: return None
+  return sink.substitute(new_forks)
+
+pm_blockend_merge = PatternMatcher([(UPat(Ops.BLOCK, name="sink"), merge_blockends)])
+
+# ***** block merging ****
+
+def merge_block(x:UOp):
+  unmergable_blocks, mergable_blocks = [], []
+  mergable_dict: defaultdict[UOp, int] = defaultdict(int)
+  for y in x.src:
+    if y.op is Ops.BLOCK and x.op is Ops.BLOCK and x.arg.ctx == y.arg.ctx: mergable_dict[y] += 1
+    elif y.op is Ops.BLOCK and x.op is Ops.BLOCKEND and x.arg.end in y.arg.ctx: mergable_dict[y] += 1
+    else: unmergable_blocks.append(y)
+  for k,v in mergable_dict.items():
+    if v == k.arg.cnt: mergable_blocks.append(k)
+    else: unmergable_blocks.extend([k]*v)
+  if len(mergable_blocks) == 0: return None
+  del mergable_dict
+
+  # create the block
+  arg = replace(x.arg, lst=tuple(flatten([y.arg.lst for y in mergable_blocks]))+x.arg.lst)
+  return UOp(x.op, src=tuple(flatten([y.src for y in mergable_blocks])+unmergable_blocks), arg=arg)
+
+def remove_blockend(x:UOp):
+  # if there's any remaining blocks that need to go in this BLOCKEND, we don't remove it
+  if any(x.arg.end in y.arg.ctx for y in x.src if y.op in {Ops.BLOCK, Ops.BLOCKEND}): return None
+
+  if (parent_blocks := [y for y in x.src if y.op is Ops.BLOCK and y.arg.child_ctx is not None and x.arg.end in y.arg.child_ctx]):
+    assert all_same(parent_blocks), f"should never have two parent blocks (has {len(parent_blocks)})"
+    parent_block = parent_blocks[0]
+    assert len(parent_blocks) == parent_block.arg.cnt
+    # NOTE: DEFINE_ACC doesn't have to be handled in any special way
+    late_ops = list(x.arg.lst)
+    # NOTE: we have to add a barrier at the start if barrier is used in the range
+    if x.op is Ops.BLOCKEND and any(y.op is Ops.BARRIER for y in late_ops) and late_ops[-1].op is Ops.ENDRANGE:
+      late_ops = [UOp(Ops.BARRIER)] + late_ops
+    # peephole opt, remove any BARRIERs next to each other
+    for i in range(len(late_ops)-1):
+      if late_ops[i].op is Ops.BARRIER and late_ops[i+1].op is Ops.BARRIER: late_ops[i+1] = UOp(Ops.NOOP)
+    arg = BasicBlock(parent_block.arg.lst+tuple(late_ops), tuple([y for y in x.arg.ctx if y is not x.arg.end]), cnt=x.arg.cnt)
+    return UOp(Ops.BLOCK, src=tuple(y for y in x.src if y is not parent_block)+parent_block.src, arg=arg)
+  # else the whole context ended by the blockend is already in this block and we can safely turn it into a block
+  return UOp(Ops.BLOCK, src=x.src, arg=BasicBlock(x.arg.lst, tuple([y for y in x.arg.ctx if y is not x.arg.end]), cnt=x.arg.cnt))
+
+block_merge = PatternMatcher([
+  (UPat((Ops.BLOCK, Ops.BLOCKEND), name="x"), merge_block),
+  (UPat(Ops.BLOCKEND, name="x"), remove_blockend),
+])
+
+# ****** finalize ******
+
+def finalize(sink:UOp) -> UOp:
+  if sink.op is not Ops.BLOCK or not all(x.op in DONT_PLACE_IN_BLOCK for x in sink.src):
+    raise RuntimeError(f"linearize failure {sink.op} {[x.op for x in sink.src if x.op not in DONT_PLACE_IN_BLOCK]}")
+
+  # place the early things
+  lst = sorted(dedup(sink.src), key=lambda x: x.tuplize) + list(sink.arg.lst)
+  return UOp(Ops.BLOCKFINAL, arg=BasicBlock(tuple(lst)))
+
+pm_finalize = PatternMatcher([(UPat(Ops.BLOCK, name="sink"), finalize)])

tinygrad/codegen/opt/postrange.py

@@ -4,8 +4,8 @@ from collections import defaultdict
 from typing import cast, Final
 from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp
 from tinygrad.device import Buffer
-from tinygrad.dtype import dtypes, ImageDType
-from tinygrad.helpers import colored, BEAM, getenv, DEBUG, to_function_name, NOOPT, argsort, round_up, prod, merge_dicts, get_single_element
+from tinygrad.dtype import AddrSpace, dtypes, ImageDType
+from tinygrad.helpers import colored, BEAM, getenv, DEBUG, to_function_name, NOOPT, argsort, round_up, prod, merge_dicts, get_single_element, flatten
 from tinygrad.codegen.opt import axis_colors, Opt, OptOps, KernelOptError, check, axis_letters
 from tinygrad.codegen.simplify import pm_flatten_range
 from tinygrad.renderer import Renderer
@@ -64,8 +64,21 @@ class Scheduler:
     return self.ast.replace(arg=KernelInfo(name=name, applied_opts=tuple(self.applied_opts), dont_use_locals=self.dont_use_locals), tag=1)
 
   def _globalizable_rngs(self) -> list[UOp]:
-    # all ranges that end before any STOREs
-    return [x for x in self.ast.toposort(lambda x: x.op is not Ops.STORE) if x.op is Ops.RANGE and x not in self.ast.ranges]
+    store_rngs = self.ast.src[0].src[2:]
+
+    # filter any not in local stores
+    local_store_rngs = [x.ranges for x in self.ast.toposort() if (x.op is Ops.STORE and x.src[0].ptrdtype.addrspace == AddrSpace.LOCAL) \
+                        or (x.op is Ops.BUFFERIZE and x.arg == AddrSpace.LOCAL)]
+    for ls in local_store_rngs: store_rngs = tuple([x for x in store_rngs if x in ls])
+
+    # filter any not in reduces
+    # TODO: enable this
+    """
+    reduce_rngs = [x.ranges for x in self.ast.toposort() if x.op is Ops.REDUCE]
+    for ls in reduce_rngs: store_rngs = tuple([x for x in store_rngs if x in ls])
+    """
+
+    return [x for x in UOp.sink(*store_rngs).toposort() if x.op is Ops.RANGE and x.arg[-1] == AxisType.LOOP] if store_rngs else []
 
   def convert_loop_to_global(self):
     if not self.opts.has_local: return None
@@ -76,11 +89,11 @@ class Scheduler:
     self.ast = self.ast.substitute(dict(zip(self.rngs, rng)))
 
   def colors(self) -> list[str]:
-    globalizible_rngs = self._globalizable_rngs()
+    store_rngs = flatten([x.src[2:] for x in self.ast.src])
     ret = []
     for x,r in zip(self.axis_types, self.rngs):
       if self.dont_use_locals and x == AxisType.GLOBAL: ret.append("BLUE")
-      elif r not in globalizible_rngs and x == AxisType.LOOP: ret.append("BLACK")
+      elif r not in store_rngs and x == AxisType.LOOP: ret.append("BLACK")
       else: ret.append(axis_colors[x])
     return ret
   def colored_shape(self) -> str: return ' '.join([colored(f'{x.src[0].render():>4s}', color) for x,color in zip(self.rngs, self.colors())])
@@ -335,7 +348,7 @@ def apply_opts(ctx:Renderer, ast:UOp):
   elif not NOOPT and (ast.arg is None or ast.arg.applied_opts == ()):
     from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
     # NOTE: hand_coded_optimizations doesn't support multiblock opts yet
-    if not any(u.op is Ops.AFTER and u.src[0].op is Ops.DEFINE_LOCAL for u in ast.backward_slice):
+    if all(len(u.src) == 1 for u in ast.backward_slice if u.op is Ops.LOAD):
       k = hand_coded_optimizations(k)
   return k.get_optimized_ast(name_override=ast.arg.name if ast.arg is not None and ast.arg.name != "test" else None)
 

tinygrad/codegen/opt/search.py

@@ -156,9 +156,7 @@ def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=Tr
         if lib in seen_libs: continue
         # filter out kernels that use 1000x more compute than the smallest
         least_compute_ops = min(this_compute_ops:=sym_infer(p.estimates.ops, var_vals), least_compute_ops)
-        if least_compute_ops*1000 < this_compute_ops:
-          if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too much compute. {this_compute_ops} when least is {least_compute_ops}")
-          continue
+        if least_compute_ops*1000 < this_compute_ops: continue
         seen_libs.add(lib)
         try: tms = _time_program(p, lib, var_vals, rawbufs, early_stop=beam[0][1]*3 if len(beam) else 1.0,
                                  allow_test_size=allow_test_size, clear_l2=hasattr(dev, 'invalidate_caches'))

tinygrad/codegen/simplify.py

@@ -13,16 +13,14 @@ def flatten_range(r:UOp):
 pm_flatten_range = PatternMatcher([
   # real ranges only
   (UPat((Ops.REDUCE, Ops.STORE), name="r"), flatten_range),
-  # END is only on RANGES. TODO: this is copied from symbolic
-  (UPat(Ops.END, name="e"), lambda e: UOp.end(*e.src[e.arg:], ends=sorted(UOp.sink(*e.src[:e.arg]).ranges, key=lambda x: x.arg))),
 ])
 
 def count_divmod(x:UOp): return len([u for u in x.toposort() if u.op in {Ops.IDIV, Ops.MOD}])
 def simplify_merge_adjacent(u:UOp) -> UOp|None:
   reduce_ranges = [x.ranges for x in u.backward_slice_with_self if x.op is Ops.REDUCE]
-  i = 0
-  while i < len(u.ended_ranges)-1:
-    r0, r1 = u.ended_ranges[i], u.ended_ranges[i+1]
+  i = range_start[u.op]
+  while i < len(u.src)-1:
+    r0, r1 = u.src[i], u.src[i+1]
     # check same type
     if r0.arg[-1] == r1.arg[-1]:
       # check if the ranges to merge are in the same reduces
@@ -41,7 +39,7 @@ def simplify_merge_adjacent(u:UOp) -> UOp|None:
   return u
 
 pm_simplify_ranges = PatternMatcher([
-  (UPat((Ops.END, Ops.REDUCE), name="u"), simplify_merge_adjacent),
+  (UPat((Ops.STORE, Ops.REDUCE), name="u"), simplify_merge_adjacent),
 ])
 
 def mark_range_mod(ctx, r:UOp, c:UOp):
@@ -59,7 +57,7 @@ def do_substitute(ctx, x: UOp):
 
 def dont_sub_ranges_for_image(ctx, x:UOp):
   if isinstance(x.src[0].dtype, ImageDType):
-    for s in x.src[0].ranges: ctx[s] = None
+    for s in x.src[1:]: ctx[s] = None
 
 pm_split_ranges = PatternMatcher([
   (UPat(Ops.RANGE, name="r")%UPat.cvar("c"), mark_range_mod),

tinygrad/device.py

@@ -5,7 +5,7 @@ from typing import Any, Generic, TypeVar, Iterator, Sequence, cast, Generator
 import importlib, inspect, functools, pathlib, os, platform, contextlib, sys, re, atexit, pickle, decimal
 from tinygrad.helpers import CI, OSX, LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored, CPU_LLVM
 from tinygrad.helpers import Context, DISABLE_COMPILER_CACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, dedup
-from tinygrad.helpers import unwrap_class_type, suppress_finalizing
+from tinygrad.helpers import unwrap_class_type
 from tinygrad.dtype import DType, ImageDType, PtrDType, dtypes, _to_np_dtype
 from tinygrad.renderer import Renderer
 
@@ -163,7 +163,6 @@ class Buffer:
     return self._trace_num
   @property
   def nbytes(self): return self.size*self.dtype.itemsize
-  @suppress_finalizing
   def __del__(self): (not hasattr(self, '_buf')) or self.deallocate()
   def __repr__(self):
     return f"<buf real:{self.is_allocated()} device:{self.device} size:{self.size} dtype:{self.dtype}" + \

tinygrad/helpers.py

@@ -170,8 +170,7 @@ SPEC = ContextVar("SPEC", 0)
 # TODO: disable by default due to speed
 IGNORE_OOB = ContextVar("IGNORE_OOB", 1)
 PCONTIG = ContextVar("PCONTIG", 0)  # partial contiguous in rangeify
-MULTIOUTPUT = ContextVar("MULTIOUTPUT", 0)
-DEBUG_RANGEIFY = ContextVar("DEBUG_RANGEIFY", 0)
+REAL_SUBSTITUTE = ContextVar("REAL_SUBSTITUTE", 0)
 
 @dataclass(frozen=True)
 class Metadata:

tinygrad/nn/__init__.py

@@ -36,7 +36,7 @@ class BatchNorm:
     self.weight: Tensor|None = Tensor.ones(sz) if affine else None
     self.bias: Tensor|None = Tensor.zeros(sz) if affine else None
 
-    self.num_batches_tracked = Tensor.zeros(dtype='long' if is_dtype_supported(dtypes.long) else 'int', requires_grad=False)
+    self.num_batches_tracked = Tensor.zeros(1, dtype='long' if is_dtype_supported(dtypes.long) else 'int', requires_grad=False)
     if track_running_stats: self.running_mean, self.running_var = Tensor.zeros(sz, requires_grad=False), Tensor.ones(sz, requires_grad=False)
 
   def calc_stats(self, x:Tensor) -> tuple[Tensor, Tensor]:

tinygrad/renderer/__init__.py

@@ -28,12 +28,9 @@ class Estimates:
     mult_stack: list[sint] = []
     dont_count: set[UOp] = set()
     if ignore_indexing:
-      def range_gate(x): return x.op is not Ops.RANGE
       for u in uops:
         if u.op in {Ops.LOAD, Ops.STORE} and (not isinstance(u.src[0].dtype, PtrDType) or u.src[0].dtype.addrspace != AddrSpace.REG):
-          # if u.src[0] is INDEX, we have to include the buffer since it might be an AFTER
-          dont_count = dont_count.union((UOp.sink(*u.src[0].src[1:]) if u.src[0].op is Ops.INDEX else u.src[0]).toposort(range_gate))
-          # TODO: is this correct? this all needs to be cleaned up
+          dont_count = dont_count.union(u.src[0].toposort())
           if len(u.src) > 2: dont_count = dont_count.union(u.src[2].toposort())
         elif u.op is Ops.IF:
           dont_count = dont_count.union(u.src[0].toposort())
@@ -48,7 +45,7 @@ class Estimates:
         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.ENDRANGE: 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.LOAD and (not isinstance(u.src[0].dtype, PtrDType) or u.src[0].dtype.addrspace != AddrSpace.REG):
         lds += u.dtype.itemsize * mults

tinygrad/renderer/cstyle.py

@@ -11,7 +11,7 @@ from tinygrad.codegen.late.devectorizer import no_vectorized_alu
 base_rewrite = PatternMatcher([
   (UPat(Ops.DEFINE_REG, name="x"), lambda ctx,x: f"{ctx.render_dtype(x.dtype.base)} {ctx[x]}[{x.dtype.size}];"),
   (UPat(Ops.IF, name="x"), lambda ctx,x: f"if ({ctx[x.src[0]]}) {{"),
-  (UPat((Ops.ENDIF, Ops.END)), lambda ctx: "}"),
+  (UPat((Ops.ENDIF, Ops.ENDRANGE)), lambda ctx: "}"),
   (UPat(Ops.WMMA, name="x"), lambda ctx,x: f"__{x.arg[0]}({ctx[x.src[0]]}, {ctx[x.src[1]]}, {ctx[x.src[2]]})"),
   # r method accesses
   (UPat(Ops.RANGE, name="x"),
@@ -173,7 +173,7 @@ class CStyleLanguage(Renderer):
       l = cast(str, self.string_rewrite.rewrite(u, ctx=self))
       assert l is not None, f"failed to render {u.op} {u.dtype} {[(x.op,x.dtype) for x in u.src]} {u.arg}"
 
-      if u.op in {Ops.ENDIF, Ops.END}: depth -= 1
+      if u.op in {Ops.ENDIF, Ops.ENDRANGE}: depth -= 1
       if (u.op is not Ops.CAST or u.dtype.vcount == 1) and (u.op in {Ops.CONST, Ops.GEP, Ops.INDEX, Ops.CUSTOMI} or \
         (u.op is Ops.LOAD and u.src[0].ptrdtype.addrspace == AddrSpace.REG) or \
         (u.op is Ops.CAST and isinstance(u.dtype, PtrDType)) or \

tinygrad/renderer/llvmir.py

@@ -108,7 +108,7 @@ base_rewrite = PatternMatcher([
    f"  br label %loop_entry_{range_str(x)}\nloop_entry_{range_str(x)}:\n"
    f"  br label %loop_body_{range_str(x)}\nloop_body_{range_str(x)}:\n"
    f"  {ctx[x]} = phi {ldt(x.dtype)} [ 0, %loop_entry_{range_str(x)} ], [ {ctx[x]}phi, %loop_latch_{range_str(x)} ]"),
-  (UPat(Ops.END, name="x"), lambda ctx,x:
+  (UPat(Ops.ENDRANGE, name="x"), lambda ctx,x:
    f"  br label %loop_latch_{range_str(x.src[0])}\nloop_latch_{range_str(x.src[0])}:\n"
    f"  {ctx[x.src[0]]}phi = add {ldt(x.src[0].dtype)} {ctx[x.src[0]]}, 1\n"
    f"  {ctx[x]} = icmp ult {ldt(x.src[0].dtype)} {ctx[x.src[0]]}phi, {ctx[x.src[0].src[0]]}\n"

tinygrad/renderer/nir.py

@@ -186,7 +186,7 @@ class NIRRenderer(Renderer):
         nstore(self.b, AddrSpace.REG, i, nimm(self.b, 0, u.dtype), u.dtype)
         mesa.nir_push_loop(self.b)
         self.r[u] = nload(self.b, AddrSpace.REG, i, u.dtype)
-      elif u.op == Ops.END:
+      elif u.op == Ops.ENDRANGE:
         nif(self.b, nalu(self.b, "ilt", x:=nalu(self.b, "iadd", self.r[u.src[0]], nimm(self.b, 1, u.src[0].dtype)), self.r[u.src[0].src[0]]),
             functools.partial(nstore, self.b, AddrSpace.REG, ranges.pop(), x, u.src[0].dtype), lambda: njump(self.b, mesa.nir_jump_break))
         mesa.nir_pop_loop(self.b, None)

tinygrad/renderer/ptx.py

@@ -115,7 +115,7 @@ string_rewrite = PatternMatcher([
      if x.dtype.count > 1 else f"ld.{mem_type(x)}.{ctx.mem_types[x.dtype]} {ctx.r[x]}, [{ctx.r[loc]}+0];"),
   (UPat(Ops.DEFINE_REG, src=()), lambda ctx: []),
   (UPat(Ops.RANGE, name="x"), lambda ctx, x: [f"mov.u32 {ctx.r[x]}, 0;", "LOOP_" + f"{ctx.r[x][1:]}:"]),
-  (UPat(Ops.END, name="x", src=(UPat.var("src0"),), allow_any_len=True), lambda ctx, x, src0: [
+  (UPat(Ops.ENDRANGE, name="x", src=(UPat.var("src0"),)), lambda ctx, x, src0: [
     ctx.code_for_op[Ops.ADD](ctx.r[src0], ctx.r[src0], "1", dtypes.int, ctx.types[dtypes.int]),
     ctx.code_for_op[Ops.CMPLT](ctx.r[x], ctx.r[x.src[0]], ctx.r[src0.src[0]], dtypes.int, ctx.types[dtypes.int]),
     f"@{ctx.r[x]} bra LOOP_{ctx.r[src0][1:]};"]),
@@ -219,7 +219,7 @@ class PTXRenderer(Renderer):
                        [ssa("wmma_in", dtype="b32") for _ in range(0, len(r[u.src[1]]), 4 // u.src[0].dtype.scalar().itemsize)],
                        [ssa("wmma_acc", dtype="b32") for _ in range(0, len(r[u.src[2]]), 4 // u.dtype.scalar().itemsize)]]
         r[u] = [ssa("wmma", dtype=self.types[u.dtype.scalar()]) for _ in range(u.dtype.count)]
-      prefix, dtype = {Ops.CAST: ("cast", None), Ops.BITCAST: ("cast", None), Ops.END: ("pred", "pred"), Ops.RANGE: ("ridx", None),
+      prefix, dtype = {Ops.CAST: ("cast", None), Ops.BITCAST: ("cast", None), Ops.ENDRANGE: ("pred", "pred"), Ops.RANGE: ("ridx", None),
         Ops.DEFINE_VAR: ("dat", None), Ops.CONST: ("const", None), Ops.DEFINE_LOCAL: ("local",self.types[dtypes.ulong]),
         Ops.DEFINE_GLOBAL: ("dat", self.types[dtypes.ulong]), **{op: ("alu", None) for op in GroupOp.ALU}}.get(u.op, (None, None))
       if prefix: r[u] = ssa(prefix, u, dtype)

tinygrad/runtime/autogen/mesa.py

@@ -7,14 +7,13 @@
 # LONGDOUBLE_SIZE is: 16
 #
 import ctypes, ctypes.util, os, gzip, base64, subprocess, tinygrad.helpers as helpers
-def brew_path(nm):
-  try: return f"{subprocess.check_output(['brew', '--prefix', nm]).decode().strip()}/lib/lib{nm}.dylib"
-  except Exception: return 'failed'
+def brew_prefix():
+  try: return subprocess.check_output(['brew', '--prefix', 'tinymesa']).decode().strip()
+  except Exception: return ''
 PATHS_TO_TRY = [
   (BASE:=os.getenv('MESA_PATH', f"/usr{'/local/' if helpers.OSX else '/'}lib"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so'),
   f'{BASE}/libtinymesa{EXT}',
-  brew_path('tinymesa_cpu'),
-  brew_path('tinymesa'),
+  f'{brew_prefix()}/lib/libtinymesa_cpu.dylib',
 ]
 def _try_dlopen_tinymesa_cpu():
   library = ctypes.util.find_library("tinymesa_cpu")
@@ -6088,7 +6087,7 @@ struct_nir_op_info._fields_ = [
 
 nir_op_info = struct_nir_op_info
 try: nir_op_infos = (struct_nir_op_info * 489).in_dll(_libraries['libtinymesa_cpu.so'], 'nir_op_infos')
-except (AttributeError, ValueError): pass
+except AttributeError: pass
 try:
     nir_op_is_selection = _libraries['FIXME_STUB'].nir_op_is_selection
     nir_op_is_selection.restype = ctypes.c_bool
@@ -8119,7 +8118,7 @@ c__EA_nir_intrinsic_index_flag = ctypes.c_uint32 # enum
 nir_intrinsic_index_flag = c__EA_nir_intrinsic_index_flag
 nir_intrinsic_index_flag__enumvalues = c__EA_nir_intrinsic_index_flag__enumvalues
 try: nir_intrinsic_index_names = (ctypes.POINTER(ctypes.c_char) * 75).in_dll(_libraries['libtinymesa_cpu.so'], 'nir_intrinsic_index_names')
-except (AttributeError, ValueError): pass
+except AttributeError: pass
 class struct_nir_intrinsic_instr(Structure):
     pass
 
@@ -8243,7 +8242,7 @@ struct_nir_intrinsic_info._fields_ = [
 
 nir_intrinsic_info = struct_nir_intrinsic_info
 try: nir_intrinsic_infos = (struct_nir_intrinsic_info * 732).in_dll(_libraries['libtinymesa_cpu.so'], 'nir_intrinsic_infos')
-except (AttributeError, ValueError): pass
+except AttributeError: pass
 try:
     nir_intrinsic_src_components = _libraries['libtinymesa_cpu.so'].nir_intrinsic_src_components
     nir_intrinsic_src_components.restype = ctypes.c_uint32
@@ -19878,4 +19877,4 @@ __all__ = \
     'union_util_format_description_0', 'util_format_colorspace',
     'util_format_layout', 'va_list']
 lvp_nir_options = gzip.decompress(base64.b64decode('H4sIAAAAAAAAA2NgZGRkYGAAkYxgCsQFsxigwgwQBoxmhCqFq2WEKwIrAEGIkQxoAEMALwCqVsCiGUwLMHA0QPn29nBJkswHANb8YpH4AAAA'))
-def __getattr__(nm): raise AttributeError('LLVMpipe requires tinymesa_cpu' if 'tinymesa_cpu' not in dll._name else f'attribute {nm} not found') if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa (tinymesa-32dc66c, mesa-25.2.4)')
+def __getattr__(nm): raise AttributeError() if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa (tinymesa-32dc66c, mesa-25.2.4)')

tinygrad/runtime/ops_python.py

@@ -52,11 +52,11 @@ class PythonProgram:
       loop_ends: dict[int, int] = {}
       while i < len(self.uops):
         uop, dtype, idp, arg = self.uops[i]
-        void_ops = {Ops.END, Ops.BARRIER, Ops.IF, Ops.ENDIF, Ops.SINK, Ops.NOOP, Ops.STORE}
+        void_ops = {Ops.ENDRANGE, Ops.BARRIER, Ops.IF, Ops.ENDIF, Ops.SINK, Ops.NOOP, Ops.STORE}
         inp = [ul[v] for v in idp if self.uops[v][0] not in void_ops]
         dtp = [dl[v] for v in idp if self.uops[v][0] not in void_ops]
         if getenv("TRACE"): print(i, uop, dtype, arg, inp, dtp)
-        if uop is Ops.END:
+        if uop is Ops.ENDRANGE:
           loop_ends[idp[0]] = i
           i = idp[0]
           continue

tinygrad/runtime/support/compiler_cpu.py

@@ -58,7 +58,7 @@ class LLVMCompiler(Compiler):
         self.diag_msgs.append(msg)
     self.handle_diag = handle_diag
     llvm.LLVMContextSetDiagnosticHandler(llvm.LLVMGetGlobalContext(), handle_diag, None)
-    super().__init__(f"compile_llvm_{processor}_{feats}{'_jit' if self.jit else ''}{'_opt' if opt else ''}")
+    super().__init__(f"compile_llvm_{self.target_arch}{'_jit' if self.jit else ''}{'_opt' if opt else ''}")
 
   def __del__(self): llvm.LLVMDisposePassBuilderOptions(self.pbo)
 

tinygrad/runtime/support/compiler_cuda.py

@@ -69,9 +69,7 @@ class PTXCompiler(Compiler):
   def disassemble(self, lib:bytes): cuda_disassemble(lib, self.arch)
 
 class NVPTXCompiler(PTXCompiler):
-  def __init__(self, arch:str):
-    nvrtc_check(nvrtc.nvJitLinkVersion(ctypes.byref(ctypes.c_uint()), ctypes.byref(ctypes.c_uint())))
-    super().__init__(arch, cache_key="nv_ptx")
+  def __init__(self, arch:str): super().__init__(arch, cache_key="nv_ptx")
   def compile(self, src:str) -> bytes:
     jitlink_check(nvrtc.nvJitLinkCreate(handle := nvrtc.nvJitLinkHandle(), 1, to_char_p_p([f'-arch={self.arch}'.encode()])), handle)
     jitlink_check(nvrtc.nvJitLinkAddData(handle, nvrtc.NVJITLINK_INPUT_PTX, ptxsrc:=super().compile(src), len(ptxsrc), "<null>".encode()), handle)

tinygrad/schedule/indexing.py

@@ -1,10 +1,10 @@
-from typing import Iterator, cast
+from typing import Iterator
 import functools, operator, itertools
 from dataclasses import dataclass, field
 from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, graph_rewrite, sint, AxisType, profile_matches
 from tinygrad.uop.symbolic import symbolic, pm_simplify_valid, pm_drop_and_clauses
-from tinygrad.helpers import argsort, all_same, cpu_profile, PCONTIG, colored, MULTIOUTPUT
+from tinygrad.helpers import argsort, all_same, cpu_profile, PCONTIG, colored
 
 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,
@@ -141,7 +141,6 @@ def apply_movement_op(op:Ops, in_shape:tuple[sint,...], arg:tuple, rngs:tuple[UO
 
 @profile_matches
 def run_rangeify(tsink:UOp, debug:bool=False) -> tuple[UOp, IndexingContext]:
-  if debug: print("**************************")
   rctx = IndexingContext()
 
   # get ops to realize
@@ -253,28 +252,5 @@ def run_rangeify(tsink:UOp, debug:bool=False) -> tuple[UOp, IndexingContext]:
     # assign to the range map. rngs are the input ranges, out_rngs are the output ranges, from the x op.
     rctx.range_map[x] = (rngs, out_rngs)
 
-  if MULTIOUTPUT:
-    # second forward pass to fuse children
-    replaced_ranges = {}
-    for x in tsink.toposort():
-      if x not in rctx.realize_map: continue
-      out_rngs = rctx.range_map[x][1]
-      _realize_axis = cast(list[int], rctx.realize_map[x])
-      consumers = [rctx.range_map[u][0] for u in consumer_map[x] if u in rctx.range_map]
-      if len(consumers) < 2: continue
-      assert all(len(out_rngs) == len(rr) for rr in consumers)
-      for i,c in enumerate(zip(*consumers)):
-        out_rng = out_rngs[i]
-        # check if they are all simple ranges
-        if not all(y.op is Ops.RANGE and y.vmax == out_rng.vmax for y in c): continue
-        for r in c: replaced_ranges[r] = out_rngs[i]
-        _realize_axis.remove(i)
-      if len(_realize_axis) == 0: del rctx.realize_map[x]
-      else: rctx.realize_map[x] = _realize_axis
-
-    # do all the replaces
-    for k,(v0,v1) in rctx.range_map.items():
-      rctx.range_map[k] = (tuple(x.substitute(replaced_ranges) for x in v0), tuple(x.substitute(replaced_ranges) for x in v1))
-
   tsink = graph_rewrite(tsink, pm_apply_rangeify, ctx=rctx, bottom_up=True, name="apply rangeify")
   return tsink, rctx

tinygrad/schedule/rangeify.py

@@ -2,13 +2,11 @@ from typing import cast
 from dataclasses import dataclass, field
 from tinygrad.dtype import dtypes, PtrDType, ImageDType, AddrSpace
 from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, _substitute, ssimplify, KernelInfo
-from tinygrad.uop.ops import track_rewrites, graph_rewrite, identity_element, sint, AxisType, BottomUpGate
+from tinygrad.uop.ops import track_rewrites, graph_rewrite, identity_element, sint, AxisType
 from tinygrad.uop.symbolic import symbolic_flat
-from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, flatten, dedup, all_int, DEBUG, SPLIT_REDUCEOP, \
-  Metadata, DEBUG_RANGEIFY, MULTIOUTPUT
+from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, flatten, dedup, all_int, DEBUG, SPLIT_REDUCEOP, Metadata, REAL_SUBSTITUTE
 from tinygrad.codegen.simplify import pm_flatten_range, pm_reduce_unparented
 from tinygrad.codegen.opt import Opt
-from tinygrad.codegen.late.control_flow import pm_merge_ends
 from tinygrad.schedule.indexing import run_rangeify, BufferizeOpts, ALWAYS_CONTIGUOUS, IndexingContext, apply_movement_op
 
 # creation can recurse a lot
@@ -138,9 +136,6 @@ def cleanup_dead_axes(b:UOp):
     # move the tag to the expand. NOTE: this expand tag might not survive
     return b.replace(src=b.src[0:1]+tuple(new_rng), tag=None).reshape(tuple(reshape)).expand(b.shape).replace(tag=b.tag)
 
-def gate_substitute(ctx, b:UOp) -> None:
-  if not any(r in b.ranges for r in ctx.keys()): raise BottomUpGate()
-pm_gate_substitute = PatternMatcher([(UPat(GroupOp.All, name="b"), gate_substitute)], compiled=False)
 # if a buffer is being stored just for permutes or something, remove it
 # we want to reexpress the indexes of idx2 in terms of the implied b1
 def remove_bufferize(src:UOp, buf:UOp, idx:UOp):
@@ -183,7 +178,11 @@ def remove_bufferize(src:UOp, buf:UOp, idx:UOp):
   # if it makes it here, the bufferize is removed
   # this is the ranges replaced
   # NOTE: if buf src is a const, we don't replace it
-  return src.substitute({k:v for k,v in zip(buf.src[1:], idx.src[1:]) if k.op is not Ops.CONST}, extra_pm=pm_gate_substitute)
+  if REAL_SUBSTITUTE:
+    return src.substitute({k:v for k,v in zip(buf.src[1:], idx.src[1:]) if k.op is not Ops.CONST})
+  else:
+    replaces = flatten([(k,v) for k,v in zip(buf.src[1:], idx.src[1:]) if k.op is not Ops.CONST])
+    return UOp(Ops.SUBSTITUTE, dtype=src.dtype, src=(src, UOp(Ops.NOOP, src=tuple(replaces[0::2])), UOp(Ops.NOOP, src=tuple(replaces[1::2]))))
 
 def pre_bufferize(b:UOp, x:UOp, copy:UOp):
   nb = b.replace(src=(b.src[0].contiguous(),)+b.src[1:])
@@ -212,6 +211,38 @@ pm_cleanups = pm_mops+PatternMatcher([
    lambda s: UOp.const(c.dtype, c.arg) if (c:=s.base).op is Ops.CONST else None),
 ])
 
+def second_stage_removal(src:UOp, buf:UOp):
+  if buf.arg.addrspace != AddrSpace.GLOBAL: return None
+  # if it's user contiguous, we never remove it
+  if src.op in ALWAYS_RUN_OPS: return None
+  accessed_buffers: list[UOp] = []
+  def red_gate(x:UOp):
+    if x.op is Ops.BUFFERIZE and x.arg.addrspace == AddrSpace.GLOBAL:
+      accessed_buffers.append(x)
+      return False
+    if x.op is Ops.BUFFER:
+      accessed_buffers.append(x)
+    return True
+  src.toposort(gate=red_gate)
+  del red_gate
+  accessed_buffers = dedup(accessed_buffers)
+  out_in_ratio = prod(buf.shape) / sum([x.size for x in accessed_buffers])
+  print(f"ratio {out_in_ratio:.2f} {buf.shape} from {[x.size for x in accessed_buffers]}")
+  if out_in_ratio > 10: return buf.replace(op=Ops.REMOVE, arg=None)
+
+pm_cleanups_2 = PatternMatcher([
+  (UPat.var("src").f(Ops.BUFFERIZE, allow_any_len=True, name="buf"), second_stage_removal),
+])
+
+def do_remove(rem:UOp, idx:UOp):
+  assert len(rem.src) == len(idx.src), f"remove on wrong bufferize, {len(rem.src)} != {len(idx.src)}"
+  return rem.src[0].substitute({k:v for k,v in zip(rem.src[1:], idx.src[1:]) if k.op is not Ops.CONST})
+
+pm_cleanup_remove = PatternMatcher([
+  # NOTE: this is better than substitute
+  (UPat(Ops.REMOVE, name="rem").f(Ops.INDEX, allow_any_len=True, name="idx"), do_remove),
+])
+
 def late_buffer_view(t:UOp, b:UOp):
   if isinstance(b.device, str) and (b.device.startswith("DISK") or b.device.startswith("TINYFS")):
     rngs = b.src[1:]
@@ -278,7 +309,7 @@ def bufferize_to_store(x:UOp):
     assert assign_target.op is Ops.INDEX, f"{assign_target.op} is not index"
     # in assign, this is the buffer size, not the bufferize size
     # TODO: assign_mops here
-    do_store = assign_target.replace(dtype=sdtype).store(assign_src).replace(tag=x.tag).end(ends=[x for x in rngs if x.op is Ops.RANGE])
+    do_store = assign_target.replace(dtype=sdtype).store(assign_src, *rngs).replace(tag=x.tag)
     ret = assign_target.src[0].after(do_store)
     mops = []
     walk = assign_mops
@@ -291,7 +322,7 @@ def bufferize_to_store(x:UOp):
   # NOTE: the DEFINE_LOCAL needs to be disambiguated here
   if sdtype.addrspace == AddrSpace.GLOBAL:
     buf = UOp.new_buffer(x.arg.device, size, x.dtype)
-    do_store = buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0]).replace(tag=x.tag).end(ends=[x for x in rngs if x.op is Ops.RANGE])
+    do_store = buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0], *rngs).replace(tag=x.tag)
     ret = buf.after(do_store).forced_reshape(shape)
     # TODO: is this right? what if it's offset
     if any(r.op is Ops.RANGE and r.src[0].op is not Ops.CONST for r in rngs):
@@ -303,8 +334,9 @@ def bufferize_to_store(x:UOp):
   tag = x.arg.device
   if tag is None: tag = UOp.unique().arg # TODO: hack
   buf = UOp(Ops.DEFINE_LOCAL, sdtype, arg=tag)
-  do_store = buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0]).end(ends=[x for x in rngs if x.op is Ops.RANGE])
-  return buf.after(do_store).reshape(shape)
+  # store has the other dtype here
+  # TODO: use after here?
+  return buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0], *rngs, dtype=sdtype).reshape(shape)
 
 pm_add_buffers = pm_mops+to_bufferview+PatternMatcher([
   (UPat(Ops.BUFFERIZE, name="x"), bufferize_to_store),
@@ -337,7 +369,6 @@ def unbind_kernel(ctx:LocalAddBufferContext, b:UOp):
   return b.src[0]
 
 def handle_after(ctx:LocalAddBufferContext, after:UOp):
-  if isinstance(after.dtype, PtrDType) and after.ptrdtype.addrspace == AddrSpace.LOCAL: return None
   buf = after.as_buf()
   # HACK to put the buffer in the MAP instead of MSTACK/MSELECT
   if buf.op in {Ops.MSTACK, Ops.MSELECT}: buf = buf.src[0]
@@ -390,9 +421,16 @@ rangeify_codegen = PatternMatcher([
 
   # add loads to non ptr indexes
   # TODO: this can be moved into codegen?
-  (UPat.any(UPat(Ops.DEFINE_GLOBAL, name="dg"), UPat(Ops.DEFINE_LOCAL).f(Ops.AFTER, allow_any_len=True, name="dg"))
-   .f(Ops.INDEX, name="idx", allow_any_len=True),
-    lambda dg,idx: None if isinstance(idx.dtype, (PtrDType, ImageDType)) else idx.replace(dtype=dg.dtype, arg=None).load()),
+  (UPat((Ops.DEFINE_GLOBAL, Ops.STORE), name="dg").f(Ops.INDEX, name="idx", allow_any_len=True),
+   lambda dg,idx: None if isinstance(idx.dtype, (PtrDType, ImageDType)) else idx.replace(dtype=dg.dtype, arg=None).load()),
+
+  # TODO: this can be moved into codegen
+  (UPat(Ops.STORE, name="store").f(Ops.INDEX, allow_any_len=True, name="idx").f(Ops.LOAD),
+    lambda store,idx: idx.replace(src=(store.as_buf(),)+idx.src[1:]).load(store if idx.dtype.addrspace != AddrSpace.LOCAL else store.barrier())),
+
+  # TODO: hack for group for reduce
+  (UPat(Ops.IF, src=(UPat.var("gate"), UPat(Ops.LOAD, src=(UPat.var("src"), UPat.var("barrier"))),)),
+   lambda src, barrier, gate: src.load(UOp(Ops.IF, src=(gate, barrier)))),
 ])
 
 def remove_metadata_tags(ctx:LocalAddBufferContext, x:UOp):
@@ -415,6 +453,7 @@ class Kernel:
 
 def split_store(ctx:list[UOp], x:UOp) -> UOp|None:
   if len(x.ranges): return None
+  if x.src[0].ptrdtype.addrspace is AddrSpace.LOCAL: return None
 
   # local kernel rewrite
   lctx = LocalAddBufferContext()
@@ -424,14 +463,8 @@ def split_store(ctx:list[UOp], x:UOp) -> UOp|None:
   metadatas = [ctx[y].metadata for y in lctx.parent_tags]
 
   # NOTE: the hack for COPY is here
-  for u in ret.toposort():
-    # TODO: this can be wrong if there's multiple of these
-    if u.op in {Ops.COPY, Ops.BUFFER_VIEW}:
-      ret = u
-      break
-  else:
-    ret = ret.sink(arg=KernelInfo(opts_to_apply=lctx.opts) if lctx.opts is not None else None)
-
+  ret = ret.sink(arg=KernelInfo(opts_to_apply=lctx.opts) if lctx.opts is not None else None) \
+    if ret.src[1].op not in {Ops.COPY, Ops.BUFFER_VIEW} else ret.src[1]
   kernel_arg = Kernel(ret,tuple(dedup(flatten([x for x in metadatas if x is not None])))[::-1])
   kernel = UOp(Ops.KERNEL, src=tuple(lctx.map.values())+tuple(lctx.vars.keys()), arg=kernel_arg)
   if ret.op is Ops.SINK and not all_same([x.device for x in kernel.src if x.op is not Ops.BIND]):
@@ -439,7 +472,7 @@ def split_store(ctx:list[UOp], x:UOp) -> UOp|None:
   return kernel
 
 split_kernels = PatternMatcher([
-  (UPat((Ops.STORE, Ops.END), name="x"), split_store),
+  (UPat(Ops.STORE, name="x"), split_store),
 ])
 
 def tag_uop(ctx:list[UOp], x:UOp):
@@ -470,6 +503,25 @@ replace_contiguous = PatternMatcher([
   (UPat(GroupOp.ALU, name="alu"), lambda ctx,alu: alu.replace(src=new_src) if (new_src:=tuple(ctx.get(s, s) for s in alu.src)) != alu.src else None),
 ])
 
+def do_sub_recurse(s:UOp):
+  x,keys,values = s.src[0], s.src[1].src, s.src[2].src
+  # SUBSTITUTE applied to SUBSTITUTE runs the child SUB on the parents. though this is probably wrong in the generic case
+  if x.op is Ops.SUBSTITUTE:
+    sub_k = UOp(Ops.SUBSTITUTE, src=(x.src[1],)+s.src[1:])
+    sub_v = UOp(Ops.SUBSTITUTE, src=(x.src[2],)+s.src[1:])
+    return UOp(Ops.SUBSTITUTE, dtype=x.dtype, src=(x.src[0], sub_k, sub_v))
+  # here we actually do the SUBSTITUTE
+  if x in keys: return values[keys.index(x)]
+  # we filter any keys where the ranges don't overlap. this keeps the algorithm O(output graph size)
+  x_ranges = x.ranges
+  new_kv = {k:v for k,v in zip(keys,values) if any(r in x_ranges for r in k.ranges)}
+  # if there's no SUBSTITUTEs left, we can just return x
+  if len(new_kv) == 0: return x
+  # then we add SUBSTITUTE to all parents
+  uop_keys, uop_values = UOp(Ops.NOOP, src=tuple(new_kv.keys())), UOp(Ops.NOOP, src=tuple(new_kv.values()))
+  return x.replace(src=tuple([UOp(Ops.SUBSTITUTE, dtype=y.dtype, src=(y,uop_keys,uop_values)) for y in x.src]))
+pm_substitute_recurse = PatternMatcher([(UPat(Ops.SUBSTITUTE, src=(UPat(), UPat(Ops.NOOP), UPat(Ops.NOOP)), name="s"), do_sub_recurse)])
+
 @track_rewrites(lambda _,ret: f"Schedule {pluralize('Kernel', len([u for u in UOp.sink(*ret.values()).toposort() if u.op is Ops.KERNEL]))}", True)
 def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
   if getenv("VIZ"): graph_rewrite(sink, PatternMatcher([]), name="View Input Graph")
@@ -479,11 +531,15 @@ def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
   tsink = graph_rewrite(tsink, earliest_rewrites+replace_contiguous, ctx={}, name="earliest rewrites")
 
   # convert movement ops to ranges
-  tsink, rctx = run_rangeify(tsink, DEBUG_RANGEIFY)
+  tsink, rctx = run_rangeify(tsink, getenv("DEBUG_RANGEIFY", 0))
 
   # NOTE: sym (vs symbolic_simple) breaks things here because ranges with len 1 aren't handled right
   tsink = graph_rewrite(tsink, symbolic_flat+pm_reduce_unparented, name="symbolic")  # this supports const folding
   tsink = graph_rewrite(tsink, pm_cleanups, bottom_up=True, name="remove costly buffers")
+  # TODO: can you substitute and remove costly buffers at the same time?
+  tsink = graph_rewrite(tsink, pm_substitute_recurse, bottom_up=True, name="run substitutes")
+  tsink = graph_rewrite(tsink, pm_cleanups_2, name="remove (more) costly buffers")
+  tsink = graph_rewrite(tsink, pm_cleanup_remove, name="actually remove")
   tsink = graph_rewrite(tsink, pm_limit_bufs, ctx=rctx, name="limit buffers")
 
   # rebuild the sink with all the BUFFERIZEs with tags, this is what's ending up in the tensor graph
@@ -496,7 +552,6 @@ def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
 
   # bufferize -> store
   tsink = graph_rewrite(tsink, pm_add_buffers, bottom_up=True, name="bufferize to store")
-  if MULTIOUTPUT: tsink = graph_rewrite(tsink, pm_merge_ends, name="merge end ranges")
   tsink = graph_rewrite(tsink, split_kernels, ctx=uop_list, name="split kernels")
 
   # if a kernel depends on a buffer, and that buffer is later assigned to, make the assign depend on the kernel's assign

tinygrad/uop/__init__.py

@@ -20,10 +20,15 @@ class Ops(FastEnum):
 
   # create buffer
   BUFFERIZE = auto()
+  REMOVE = auto()
+  SUBSTITUTE = auto()
 
   # ops that adjust the behavior of the scheduler
   CONTIGUOUS = auto(); CONTIGUOUS_BACKWARD = auto(); DETACH = auto(); FUSE = auto() # noqa: E702
 
+  # blocks in linearizer (only used there)
+  BLOCK = auto(); BLOCKSTART = auto(); BLOCKEND = auto(); BLOCKFINAL = auto() # noqa: E702
+
   # movement ops! these only exist in the tensor graph
   RESHAPE = auto(); PERMUTE = auto(); EXPAND = auto(); PAD = auto(); SHRINK = auto(); FLIP = auto() # noqa: E702
   MULTI = auto()  # MULTI is really a movement op
@@ -66,7 +71,7 @@ class Ops(FastEnum):
   WHERE = auto(); MULACC = auto() # noqa: E702
 
   # control flow ops
-  BARRIER = auto(); RANGE = auto(); IF = auto(); END = auto(); ENDIF = auto() # noqa: E702
+  BARRIER = auto(); RANGE = auto(); IF = auto(); ENDRANGE = auto(); ENDIF = auto() # noqa: E702
 
   # consts. VCONST is a vectorized const
   VCONST = auto(); CONST = auto() # noqa: E702
@@ -90,6 +95,7 @@ class GroupOp:
   Movement = {Ops.RESHAPE, Ops.EXPAND, Ops.PERMUTE, Ops.PAD, Ops.SHRINK, Ops.FLIP}
 
   Buffer = {Ops.LOAD, Ops.STORE, Ops.CONST, Ops.DEFINE_VAR}
+  Block = {Ops.BLOCK, Ops.BLOCKEND, Ops.BLOCKSTART}
 
   # BinaryOps that can be flipped
   Commutative = {Ops.ADD, Ops.MUL, Ops.MAX, Ops.CMPNE, Ops.CMPEQ, Ops.XOR, Ops.AND, Ops.OR}

tinygrad/uop/ops.py

@@ -17,7 +17,7 @@ class AxisType(Enum):
   GLOBAL = auto(); WARP = auto(); LOCAL = auto(); LOOP = auto(); GROUP_REDUCE = auto(); REDUCE = auto(); UPCAST = auto(); UNROLL = auto() # noqa: E702
   THREAD = auto()
 
-range_start = {Ops.BUFFERIZE: 1, Ops.REDUCE: 1, Ops.STORE: 2, Ops.WMMA: 3}
+range_start = {Ops.BUFFERIZE: 1, Ops.REMOVE: 1, Ops.REDUCE: 1, Ops.STORE: 2, Ops.WMMA: 3}
 
 # https://en.wikipedia.org/wiki/Identity_element
 def identity_element(op:Ops, dt:DType) -> ConstType: return dtypes.as_const({Ops.ADD:0, Ops.MUL:1, Ops.MAX:dtypes.min(dt)}[op], dt)
@@ -169,7 +169,7 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
 
   @property
   def ptrdtype(self) -> PtrDType:
-    if not isinstance(self.dtype, PtrDType): raise RuntimeError(f"ptrdtype called on UOp with type {self.dtype}")
+    if not isinstance(self.dtype, PtrDType): raise RuntimeError("ptrdtype called on UOp without PtrDType")
     return self.dtype
 
   # *** uop shape stuff ***
@@ -179,18 +179,18 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     match self.op:
       # late ops don't have shape
       case Ops.UNIQUE | Ops.DEVICE | Ops.RANGE | Ops.INDEX | Ops.LOAD | Ops.IF | Ops.BARRIER | Ops.CUSTOM | Ops.CUSTOMI | \
-           Ops.VECTORIZE | Ops.VCONST | Ops.GEP | Ops.SPECIAL | Ops.UNROLL | Ops.PRECAST:
+           Ops.VECTORIZE | Ops.VCONST | Ops.SUBSTITUTE | Ops.GEP | Ops.SPECIAL | Ops.UNROLL | Ops.PRECAST:
         return None
 
       # some ops init the shape
       case Ops.CONST | Ops.DEFINE_VAR | Ops.BIND: return () if self._device is not None else None
       case Ops.BUFFER: return (self.arg,)
       case Ops.BUFFER_VIEW: return (self.arg[0],)
-      case Ops.BUFFERIZE: return tuple([int(r.vmax+1) for r in self.src[1:]])
+      case Ops.BUFFERIZE | Ops.REMOVE: return tuple([int(r.vmax+1) for r in self.src[1:]])
       case Ops.DEFINE_GLOBAL | Ops.DEFINE_LOCAL | Ops.DEFINE_REG: return (self.ptrdtype.size,)
 
       # passthrough ops
-      case Ops.REDUCE | Ops.MSTACK | Ops.MSELECT | Ops.DETACH | Ops.CONTIGUOUS | Ops.CONTIGUOUS_BACKWARD | Ops.FUSE | Ops.AFTER | Ops.END:
+      case Ops.REDUCE | Ops.MSTACK | Ops.MSELECT | Ops.DETACH | Ops.CONTIGUOUS | Ops.CONTIGUOUS_BACKWARD | Ops.FUSE | Ops.AFTER:
         return self.src[0]._shape
 
       # ops with custom handling
@@ -276,10 +276,6 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
       for s in self.src[:range_start[self.op]]: ret.update(s.ranges)
       for s in UOp.sink(*self.src[range_start[self.op]:]).ranges:
         if s in ret: del ret[s]
-    elif self.op is Ops.END:
-      for s in self.src[self.arg:]: ret.update(s.ranges)
-      for s in UOp.sink(*self.src[:self.arg]).ranges:
-        if s in ret: del ret[s]
     else:
       for s in self.src: ret.update(s.ranges)
     return ret
@@ -289,13 +285,6 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     if self.op is Ops.RANGE: return {self:None}
     return self._ranges
 
-  @functools.cached_property
-  def ended_ranges(self):
-    match self.op:
-      case Ops.REDUCE: return self.src[1:]
-      case Ops.END: return self.src[:self.arg]
-      case _: raise RuntimeError(f"{self.op} doesn't end ranges")
-
   # *** uop evaluation ***
 
   def simplify(self, tracked=False, full_symbolic=True):
@@ -361,11 +350,6 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     return UOp(Ops.GEP, self.dtype.scalar().vec(len(i)) if len(i) > 1 else self.dtype.scalar(), (self,), i)
   def load(self, *src:UOp, **kwargs): return UOp(Ops.LOAD, dtype=kwargs.pop("dtype", self.dtype.base), src=(self,)+src, **kwargs)
   def store(self, *src:UOp, **kwargs): return UOp(Ops.STORE, kwargs.pop("dtype", dtypes.void), (self,)+src, **kwargs)
-  def end(self, *src:UOp, ends:Sequence[UOp]):
-    if len(ends) == 0:
-      if len(src): return UOp(Ops.NOOP, src=(self,*src))
-      return self
-    return UOp(Ops.END, src=(*ends, self, *src), arg=len(ends))
   def after(self, *src:UOp): return UOp(Ops.AFTER, self.dtype, (self,)+src)
   def assign(self, x:UOp): return UOp(Ops.ASSIGN, self.dtype, (self, x))
   def barrier(self, *src:UOp): return UOp(Ops.BARRIER, src=(self,)+src)
@@ -857,6 +841,7 @@ class UPat(MathTrait):
   def reduce(self, *src:UPat, **kwargs): return UPat(Ops.REDUCE, self.dtype, src=(self,)+src, **kwargs)
   def fuse(self): return self.alu(Ops.FUSE)
   def broadcast(self, **kwargs): return UPat(Ops.VECTORIZE, self.dtype, src=self, **kwargs)
+  def or_broadcasted(self, **kwargs): return UPat.any(self, self.broadcast(**kwargs))
   def contiguous(self, *args, **kwargs): return UPat(Ops.CONTIGUOUS, dtype=self.dtype, src=(self,)+args, **kwargs)
 
   def const_like(self, b:ConstLike): return UPat.const(self.dtype, cast(ConstType, b))
@@ -1111,7 +1096,7 @@ class RewriteContext:
               new_n, test_n = test_n, self.cached_bpm_rewrite(test_n)
           except BottomUpGate:
             # if the bpm matching raised a gate, we are done with this node and dont continue down the srcs
-            self.replace[n] = unwrap(test_n)
+            self.replace[n] = new_n
             continue
         stack.append((n, 1, new_n))
         for x in reversed(new_n.src):
@@ -1193,6 +1178,8 @@ pm_lower_index_dtype = PatternMatcher([
     lambda s: s.replace(src=s.src[:2]+tuple(u.src[0] for u in s.src[2:]))),
   # TODO: this is only triggering if they are all casts, correct?
   (UPat((Ops.SINK, Ops.NOOP), src=UPat().cast(dtypes.index), name="n"), lambda n: n.replace(src=tuple(s.src[0] for s in n.src))),
+  # TODO: this should be more general
+  (UPat(Ops.AFTER, name="x"), lambda x: x.replace(src=tuple(y.src[0] if y.op is Ops.CAST and y.dtype.scalar()==dtypes.index else y for y in x.src))),
 ])
 def _index_to_concrete_int(u:UOp): return graph_rewrite(u.sink(), pm_lower_index_dtype).src[0]
 

tinygrad/uop/spec.py

@@ -159,9 +159,8 @@ spec = PatternMatcher([
   (UPat(Ops.DEFINE_REG, src=()), lambda: True),
   (UPat(Ops.DEFINE_VAR, name="x"), lambda x: isinstance(x.arg[1], int) and isinstance(x.arg[2], int)),
 
-  (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.RANGE, src=(UPat.var("x"),), 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.SPECIAL, src=(UPat.var("x"),), name="s"), lambda s,x: s.dtype == x.dtype == dtypes.int32 and isinstance(s.arg, str)),
 
   (UPat(Ops.CONST, src=(), name="x"), lambda x: type(x.arg) is type(dtypes.as_const(x.arg, x.dtype))),
@@ -181,11 +180,15 @@ spec = PatternMatcher([
   (UPat(Ops.INDEX, src=(UPat(GroupOp.Defines).or_after(), UPat())), lambda: True),
   (UPat(Ops.INDEX, src=(UPat(GroupOp.Defines).or_after(), UPat(), UPat(dtype=dtypes.bool))), lambda: True),
 
+  # LOAD on STORE
+  (UPat(Ops.LOAD, src=(UPat(Ops.STORE),), allow_any_len=True), lambda: True),
+
   # LOAD takes a <bufidx, alt?, barrier?>
   (UPat(Ops.LOAD, src=(index_pat, UPat(Ops.IF, name="cond")), allow_any_len=True), lambda idx,cond: validate_index(idx,cond.src[0])),
   (UPat(Ops.LOAD, src=(index_pat,), allow_any_len=True), validate_index),
 
-  # STORE takes a <bufidx, val, ranges...>
+  # STORE takes a <bufidx, val, gate?>
+  (UPat(Ops.STORE, src=(index_pat, UPat(name="val"), UPat(Ops.IF, name="gate")), allow_any_len=True), validate_store),
   (UPat(Ops.STORE, src=(index_pat, UPat(name="val")), allow_any_len=True), validate_store),
 
   # most ALUs have all matching dtypes, except CMPLT, CMPNE, and WHERE
@@ -196,7 +199,7 @@ spec = PatternMatcher([
   (UPat((Ops.IDIV, Ops.MOD), name="x"), lambda x: None if dtypes.is_int(x.dtype) else False),
   (UPat(GroupOp.ALU, name="x"), lambda x: all(x.dtype.base == y.dtype.base for y in x.src)),
 
-  (UPat(Ops.END, dtype=dtypes.void), lambda: True),
+  (UPat(Ops.ENDRANGE, dtype=dtypes.void, src=(UPat(Ops.RANGE),)), lambda: True),
 
   # WMMA has a <a, b, acc>
   (UPat(Ops.WMMA, src=(UPat(), UPat(), UPat()), name="x"), lambda x: isinstance(x.arg, tuple) and len(x.arg) == 8),
@@ -204,8 +207,9 @@ spec = PatternMatcher([
   (UPat(Ops.UNROLL, name="x"), lambda x: x.src[0].dtype.count == prod(y[1] for y in x.arg)),
 
   # if has a <gate, barrier?>
-  (UPat(Ops.IF, dtype=dtypes.void, src=(UPat(),), allow_any_len=True), lambda: True),
-  (UPat(Ops.ENDIF, dtype=dtypes.void, src=(UPat(Ops.IF),), allow_any_len=True), lambda: True),
+  (UPat(Ops.IF, dtype=dtypes.void, src=(UPat(),)), lambda: True),
+  (UPat(Ops.IF, dtype=dtypes.void, src=(UPat(), UPat(Ops.BARRIER))), lambda: True),
+  (UPat(Ops.ENDIF, dtype=dtypes.void, src=(UPat(Ops.IF),)), lambda: True),
 
   (UPat(Ops.REDUCE_AXIS, name="x"), lambda x: isinstance(x.arg, tuple) and len(x.arg) >= 2 and x.arg[0] in {Ops.ADD, Ops.MUL, Ops.MAX}),
   (UPat(Ops.GEP, src=(UPat.var("src"),), name="gep"), lambda gep,src: gep.dtype == src.dtype.scalar()),
@@ -236,6 +240,9 @@ full_spec = PatternMatcher([
   # SENTINEL should never be in the graph
   (UPat(Ops.SENTINEL), lambda: False),
 
+  # allow any SUBSTITUTE
+  (UPat(Ops.SUBSTITUTE), lambda: True),
+
   # Invalid must have type Index
   (UPat(Ops.CONST, arg=Invalid, name="x"), lambda x: x.dtype.scalar() == dtypes.index),
   # where on index in rhs position is fine
@@ -268,7 +275,7 @@ full_spec = PatternMatcher([
   (UPat(Ops.INDEX, src=(UPat((Ops.VECTORIZE, Ops.CAST)), UPat())), lambda: True),
 
   # linearizer: outputs + intermediate KERNELs
-  (UPat(Ops.KERNEL, dtype=dtypes.void), lambda: True),
+  (UPat((Ops.BLOCKSTART, Ops.BLOCK, Ops.BLOCKFINAL, Ops.BLOCKEND, Ops.KERNEL), dtype=dtypes.void), lambda: True),
 
   # allow index dtype on a restricted set of UOps
   (UPat((Ops.ADD, Ops.MUL, Ops.MOD, Ops.IDIV, Ops.MAX, Ops.WHERE,

tinygrad/uop/symbolic.py

@@ -377,13 +377,6 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
   (UPat(GroupOp.Binary, src=(UPat.var("x", dtypes.long), UPat.var("y", dtypes.long)), name="u"), lambda u,x,y:
     x.cast(dtypes.int).alu(u.op, y.cast(dtypes.int)).cast(u.dtype) if not any(v.overflows(dtypes.int) for v in (u,x,y)) else None),
   ((UPat.var("x", dtypes.index) + UPat.cvar("c")).cast(dtypes.sints, name="cast"), lambda x,c,cast:x.cast(cast.dtype)+c.cast(cast.dtype)),
-  # only RANGE/IF/STORE/KERNEL have side effects
-  (UPat(Ops.AFTER, name="x"), lambda x: x.replace(src=(x.src[0],)+
-    tuple(flatten([(y,) if y.op in {Ops.RANGE, Ops.IF, Ops.STORE, Ops.KERNEL, Ops.BARRIER, Ops.END} else y.src for y in x.src[1:]])))),
-  # after with 1 src is just src[0]
-  (UPat(Ops.AFTER, src=(UPat.var("s"),)), lambda s: s),
-  # END is only on RANGES
-  (UPat(Ops.END, name="e"), lambda e: UOp.end(*e.src[e.arg:], ends=sorted(UOp.sink(*e.src[:e.arg]).ranges, key=lambda x: x.arg))),
 ])+gep_pushing
 
 symbolic_flat = symbolic+PatternMatcher([

tinygrad/viz/index.html

@@ -2,7 +2,6 @@
 <html>
 <head>
   <title>tinygrad viz</title>
-  <meta charset="utf-8" />
   <meta name="viewport" content="width=device-width, initial-scale=1.0">
   <link rel="icon" href="data:;base64,iVBORw0KGgo=">
   <script src="assets/d3js.org/d3.v7.min.js" charset="utf-8"></script>
@@ -39,8 +38,6 @@
   ::-webkit-scrollbar-thumb { background: #686977; }
   a {
     color: #4a90e2;
-    text-decoration: underline;
-    cursor: pointer;
   }
   ul {
     padding: 0;
@@ -55,13 +52,6 @@
   }
   ul > ul {
     display: none;
-    margin-left: 6px;
-  }
-  ul.has-children > p::before {
-    content:"▸ ";
-  }
-  ul.has-children.expanded > p::before {
-    content:"▾ ";
   }
   ul.expanded > ul {
     display: block;
@@ -151,7 +141,7 @@
   .metadata > * + *, .rewrite-container > * + *, .ctx-list > * + * {
     margin-top: 12px;
   }
-  ul > * + * {
+  .ctx-list > ul > * + * {
     margin-top: 4px;
   }
   .graph {

tinygrad/viz/js/index.js

@@ -78,7 +78,7 @@ function renderDag(graph, additions, recenter) {
         if (parents == null && children == null) return;
         const src = [...parents, ...children, d.id];
         nodes.classed("highlight", n => src.includes(n.id)).classed("child", n => children.includes(n.id));
-        const matchEdge = (v, w) => (v===d.id && children.includes(w)) ? "highlight child " : (parents.includes(v) && w===d.id) ? "highlight " : "";
+        const matchEdge = (v, w) => (v===d.id && children.includes(w)) ? "highlight child "  : (parents.includes(v) && w===d.id) ? "highlight " : "";
         d3.select("#edges").selectAll("path.edgePath").attr("class", e => matchEdge(e.v, e.w)+"edgePath");
         d3.select("#edge-labels").selectAll("g.port").attr("class",  (_, i, n) => matchEdge(...n[i].id.split("-"))+"port");
         e.stopPropagation();
@@ -92,9 +92,10 @@ function renderDag(graph, additions, recenter) {
     }).selectAll("text").data(d => {
       const ret = [[]];
       for (const { st, color } of parseColors(d.label, defaultColor="initial")) {
-        const lines = st.split("\n");
-        ret.at(-1).push({ st:lines[0], color });
-        for (let i=1; i<lines.length; i++) ret.push([{ st:lines[i], color }]);
+        for (const [i, l] of st.split("\n").entries()) {
+          if (i > 0) ret.push([]);
+          ret.at(-1).push({ st:l, color });
+        }
       }
       return [ret];
     }).join("text").selectAll("tspan").data(d => d).join("tspan").attr("x", "0").attr("dy", 14).selectAll("tspan").data(d => d).join("tspan")
@@ -173,16 +174,10 @@ function tabulate(rows) {
   return root;
 }
 
-var data, focusedDevice, focusedShape, canvasZoom, zoomLevel = d3.zoomIdentity, shapeMetadata = new Map();
-function focusShape(shape) {
-  saveToHistory({ shape:focusedShape });
-  focusedShape = shape?.key; d3.select("#timeline").call(canvasZoom.transform, zoomLevel);
-  return document.querySelector(".metadata").replaceChildren(shapeMetadata.get(focusedShape) ?? "");
-}
-
+var data, focusedDevice, focusedShape, canvasZoom, zoomLevel = d3.zoomIdentity;
 async function renderProfiler() {
   displayGraph("profiler");
-  d3.select(".metadata").node().replaceChildren(shapeMetadata.get(focusedShape) ?? "");
+  d3.select(".metadata").node().replaceChildren(focusedShape?.html ?? "");
   // layout once!
   if (data != null) return updateProgress({ start:false });
   const profiler = d3.select(".profiler").html("");
@@ -248,17 +243,15 @@ async function renderProfiler() {
         }
         const html = document.createElement("div");
         html.appendChild(tabulate([["Name", colored(e.name)], ["Duration", formatTime(e.dur)], ["Start Time", formatTime(e.st)]]).node());
-        const argsDiv = document.createElement("div"); argsDiv.id = "args"; html.appendChild(document.createElement("br")); html.appendChild(argsDiv);
         if (e.info != null) html.appendChild(document.createElement("p")).innerText = "\n"+e.info;
         if (shapeRef != null) {
-          const a = html.appendChild(document.createElement("a"));
-          a.innerText = "\nView codegen rewrite";
-          a.onclick = () => setCtxWithHistory(shapeRef.ctx, shapeRef.step);
+          const p = html.appendChild(document.createElement("p"));
+          p.innerText = "\nView Codegen Rewrite"; p.style.cursor = "pointer";
+          p.onclick = () => setCtxWithHistory(shapeRef.ctx, shapeRef.step);
         }
         // tiny device events go straight to the rewrite rule
         const key = k.startsWith("TINY") ? null : `${k}-${j}`;
-        if (key != null) shapeMetadata.set(key, html);
-        const arg = { tooltipText:colored(e.name).outerHTML+"\n"+formatTime(e.dur)+(e.info != null ? "\n"+e.info : ""), key, ...shapeRef };
+        const arg = { tooltipText:colored(e.name).outerHTML+"\n"+formatTime(e.dur)+(e.info != null ? "\n"+e.info : ""), html, key, ...shapeRef };
         if (e.key != null) shapeMap.set(e.key, arg);
         // offset y by depth
         shapes.push({x:e.st, y:levelHeight*depth, width:e.dur, height:levelHeight, arg, label, fillColor });
@@ -302,30 +295,17 @@ async function renderProfiler() {
         const rows = [["DType", dtype], ["Len", formatUnit(sz)], ["Size", formatUnit(nbytes, "B")], ["Lifetime", formatTime(dur)]];
         if (users != null) rows.push(["Users", users.length]);
         const info = html.appendChild(tabulate(rows).node());
-        const arg = {tooltipText:info.outerHTML, key:`${k}-${num}`};
         for (let u=0; u<users?.length; u++) {
           const p = html.appendChild(document.createElement("p")); p.style.marginTop = "4px";
-          const { repr, num, mode, shape } = users[u];
-          const bufInfo = `${mode == 2 ? 'read+write' : mode == 1 ? 'write' : 'read'}@data${num}`
-          p.appendChild(colored(`[${u}] ${repr} ${bufInfo}`));
+          const { repr, num, mode, shape } = users[u]; p.appendChild(colored(`[${u}] ${repr} ${mode == 2 ? 'read+write' : mode == 1 ? 'write' : 'read'}@data${num}`));
           const metadata = shape?.tooltipText?.split("\n").at(-1);
           if (metadata != null) p.appendChild(document.createElement("span")).innerText = "\n"+metadata;
           if (shape != null) {
             p.style.cursor = "pointer";
             p.onclick = () => focusShape(shape);
-            const args = shapeMetadata.get(shape.key).querySelector("#args");
-            const bufArg = d3.create("p").text(`${bufInfo} ${rows[2][1]}`).style("cursor", "pointer").style("margin-top", "4px").on("click", () => {
-              const device = document.getElementById(k);
-              if (!isExpanded(device)) device.click();
-              focusShape(arg);
-            }).node();
-            bufArg.dataset.num = num;
-            let before = null;
-            for (const c of args.children) { if (+c.dataset.num > num) { before = c; break; } }
-            args.insertBefore(bufArg, before);
           }
         }
-        shapeMetadata.set(arg.key, html)
+        const arg = {tooltipText:info.outerHTML, html, key:`${k}-${num}`};
         shapes.push({ x, y0:y.map(yscale), y1:y.map(y0 => yscale(y0+nbytes)), arg, fillColor:cycleColors(colorScheme.BUFFER, shapes.length) });
       }
       // generic polygon merger
@@ -358,7 +338,6 @@ async function renderProfiler() {
           else if (tid === focusedDevice) { track.shapes = track.views[0]; offset += rescaleTrack(track, tid, 1/track.scaleFactor); }
         }
         data.axes.y = newFocus != null ? { domain:[0, (t=data.tracks.get(newFocus)).peak], range:[t.offsetY+t.height, t.offsetY], fmt:"B" } : null;
-        toggleCls(document.getElementById(focusedDevice), document.getElementById(newFocus), "expanded");
         focusedDevice = newFocus;
         return resize();
       });
@@ -417,7 +396,7 @@ async function renderProfiler() {
             lw += e.label[li].width;
           }
         }
-        if (focusedShape != null && e.arg?.key === focusedShape) { paths.push([p, pcolor]); }
+        if (focusedShape?.key && e.arg?.key === focusedShape.key) { paths.push([p, pcolor]); }
       }
     }
     // draw axes
@@ -484,11 +463,15 @@ async function renderProfiler() {
     }
   }
 
+  function focusShape(shape) {
+    focusedShape = shape; render(zoomLevel);
+    return document.querySelector(".metadata").replaceChildren(shape?.html ?? "");
+  }
   canvas.addEventListener("click", e => {
     e.preventDefault();
     const foundRect = findRectAtPosition(e.clientX, e.clientY);
     if (foundRect?.step != null && foundRect?.key == null) { return setCtxWithHistory(foundRect.ctx, foundRect.step); }
-    if (foundRect?.key != focusedShape) { focusShape(foundRect); }
+    if (foundRect?.key != focusedShape?.key) { focusShape(foundRect); }
   });
 
   canvas.addEventListener("mousemove", e => {
@@ -547,10 +530,10 @@ function codeBlock(st, language, { loc, wrap }={}) {
   return ret;
 }
 
-function toggleCls(prev, next, cls, value) {
-  prev?.classList.remove(cls);
-  next?.classList.toggle(cls, value ?? true);
-  requestAnimationFrame(() => next?.scrollIntoView({ behavior: "auto", block: "nearest" }));
+function setActive(e) {
+  if (e == null) return;
+  e.classList.add("active");
+  requestAnimationFrame(() => e.scrollIntoView({ behavior: "auto", block: "nearest" }));
 }
 
 // ** hljs extra definitions for UOps and float4
@@ -580,41 +563,31 @@ const evtSources = [];
 // context: collection of steps
 const state = {currentCtx:-1, currentStep:0, currentRewrite:0, expandSteps:false};
 function setState(ns) {
-  const { ctx:prevCtx, step:prevStep } = select(state.currentCtx, state.currentStep);
+  const { currentCtx:prevCtx, currentStep:prevStep } = state;
   Object.assign(state, ns);
   // update element styles if needed
-  const { ctx, step } = select(state.currentCtx, state.currentStep);
-  toggleCls(prevCtx, ctx, "expanded", state.expandSteps);
-  if (ctx?.id !== prevCtx?.id) toggleCls(prevCtx, ctx, "active");
-  if (ctx?.id !== prevCtx?.id || step?.id !== prevStep?.id) {
-    toggleCls(prevStep, step, "active");
-    // walk the tree back until all parents expanded so that the child is visible
-    let e = step;
-    while (e?.parentElement?.id.startsWith("step")) {
-      e.parentElement.classList.add("expanded");
-      e = e.parentElement;
-    }
+  document.getElementById(`ctx-${state.currentCtx}`)?.classList.toggle("expanded", state.expandSteps);
+  if (state.currentCtx !== prevCtx) {
+    document.getElementById(`ctx-${prevCtx}`)?.classList.remove("active", "expanded");
+    setActive(document.getElementById(`ctx-${state.currentCtx}`));
+  }
+  if (state.currentCtx !== prevCtx || state.currentStep !== prevStep) {
+    document.getElementById(`step-${prevCtx}-${prevStep}`)?.classList.remove("active");
+    setActive(document.getElementById(`step-${state.currentCtx}-${state.currentStep}`));
   }
   // re-render
   main();
 }
 
-const getSubrewrites = (ul) => ul.querySelectorAll(":scope > ul");
-
-function saveToHistory(ns) {
-  // NOTE: browser does a structured clone, passing a mutable object is safe.
-  history.replaceState(ns, "");
-  history.pushState(ns, "");
-}
-
 // set a new context and keep the old one in browser history
 function setCtxWithHistory(newCtx, step=0) {
-  saveToHistory(state);
+  // NOTE: browser does a structured clone, passing a mutable object is safe.
+  history.replaceState(state, "");
+  history.pushState(state, "");
   setState({ expandSteps:true, currentCtx:newCtx+1, currentStep:step, currentRewrite:0 });
 }
 
 window.addEventListener("popstate", (e) => {
-  if (e.state?.shape != null) return focusShape({ key:e.state?.shape });
   if (e.state != null) setState(e.state);
 });
 
@@ -632,25 +605,16 @@ async function main() {
       p.onclick = () => {
         setState(i === state.currentCtx ? { expandSteps:!state.expandSteps } : { expandSteps:true, currentCtx:i, currentStep:0, currentRewrite:0 });
       }
-      const stack = []; let list = ul;
       for (const [j,u] of steps.entries()) {
-        while (stack.length && stack.at(-1).depth >= u.depth) stack.pop();
-        const list = stack.length > 0 ? stack.at(-1).li : ul;
-        u.li = list.appendChild(document.createElement("ul"));
-        u.li.id = `step-${i}-${j}`;
-        const p = u.li.appendChild(document.createElement("p"));
+        const inner = ul.appendChild(document.createElement("ul"));
+        inner.id = `step-${i}-${j}`;
+        const p = inner.appendChild(document.createElement("p"));
         p.innerText = `${u.name}`+(u.match_count ? ` - ${u.match_count}` : '');
-        p.onclick = (e) => {
+        inner.style.marginLeft = `${8*u.depth}px`;
+        inner.onclick = (e) => {
           e.stopPropagation();
-          const subrewrites = getSubrewrites(e.currentTarget.parentElement);
-          if (subrewrites.length) { e.currentTarget.parentElement.classList.toggle("expanded"); }
-          setState({ currentStep:j, currentCtx:i });
+          setState({ currentStep:j, currentCtx:i, currentRewrite:0 });
         }
-        stack.push(u);
-      }
-      for (const l of ul.querySelectorAll("ul > ul > p")) {
-        const subrewrites = getSubrewrites(l.parentElement);
-        if (subrewrites.length > 0) { l.innerText += ` (${subrewrites.length})`; l.parentElement.classList.add("has-children"); }
       }
     }
     return setState({ currentCtx:-1 });
@@ -800,32 +764,22 @@ appendResizer(document.querySelector(".metadata-parent"), { minWidth: 20, maxWid
 
 // **** keyboard shortcuts
 
-const select = (ctx, step) => ({ ctx:document.getElementById(`ctx-${ctx}`), step:document.getElementById(`step-${ctx}-${step}`) });
-const deselect = (element) => {
-  const parts = element?.id.split("-").map(Number);
-  return element?.id.startsWith("ctx") ? { ctx:parts[1], step:null } : element?.id.startsWith("step") ? {ctx:parts[1], step:parts[2]} : {};
-}
-const isExpanded = (el) => el?.classList.contains("expanded");
-
 document.addEventListener("keydown", (event) => {
   const { currentCtx, currentStep, currentRewrite, expandSteps } = state;
   // up and down change the step or context from the list
   const changeStep = expandSteps && ctxs[currentCtx].steps?.length;
-  const { step, ctx } = select(currentCtx, currentStep);
   if (event.key == "ArrowUp") {
     event.preventDefault();
     if (changeStep) {
-      let prev = deselect(step.previousElementSibling);
-      if (prev.step == null && isExpanded(step.parentElement)) prev = deselect(step.parentElement);
-      return prev.step != null && !isExpanded(step) && setState({ currentRewrite:0, currentStep:prev.step });
+      return setState({ currentRewrite:0, currentStep:Math.max(0, currentStep-1) });
     }
     return setState({ currentStep:0, currentRewrite:0, currentCtx:Math.max(0, currentCtx-1), expandSteps:false });
   }
   if (event.key == "ArrowDown") {
     event.preventDefault();
     if (changeStep) {
-      const next = deselect(isExpanded(step) ? step.children[1] : step.nextElementSibling);
-      return next.step != null && setState({ currentRewrite:0, currentStep:next.step });
+      const totalUOps = ctxs[currentCtx].steps.length-1;
+      return setState({ currentRewrite:0, currentStep:Math.min(totalUOps, currentStep+1) });
     }
     return setState({ currentStep:0, currentRewrite:0, currentCtx:Math.min(ctxs.length-1, currentCtx+1), expandSteps:false });
   }
@@ -835,7 +789,6 @@ document.addEventListener("keydown", (event) => {
     if (currentCtx === -1) {
       return setState({ currentCtx:0, expandSteps:true });
     }
-    if (expandSteps && getSubrewrites(step).length) return step.children[0].click();
     return setState({ expandSteps:!expandSteps });
   }
   // left and right go through rewrites in a single UOp

tinygrad/viz/serve.py

@@ -17,10 +17,10 @@ uops_colors = {Ops.LOAD: "#ffc0c0", Ops.STORE: "#87CEEB", Ops.CONST: "#e0e0e0",
                Ops.DEFINE_GLOBAL: "#ffe0b0", Ops.DEFINE_LOCAL: "#ffe0d0", Ops.DEFINE_REG: "#f0ffe0", Ops.REDUCE_AXIS: "#FF6B6B",
                Ops.RANGE: "#c8a0e0", Ops.ASSIGN: "#909090", Ops.BARRIER: "#ff8080", Ops.IF: "#c8b0c0", Ops.SPECIAL: "#c0c0ff",
                Ops.INDEX: "#e8ffa0", Ops.WMMA: "#efefc0", Ops.MULTI: "#f6ccff", Ops.KERNEL: "#3e7f55",
-               **{x:"#D8F9E4" for x in GroupOp.Movement}, **{x:"#ffffc0" for x in GroupOp.ALU}, Ops.THREEFRY:"#ffff80",
-               Ops.BUFFER_VIEW: "#E5EAFF", Ops.BUFFER: "#B0BDFF", Ops.COPY: "#a040a0", Ops.FUSE: "#FFa500",
+               **{x:"#D8F9E4" for x in GroupOp.Movement}, **{x:"#ffffc0" for x in GroupOp.ALU}, Ops.THREEFRY:"#ffff80", Ops.BUFFER_VIEW: "#E5EAFF",
+               Ops.BLOCK: "#C4A484", Ops.BLOCKEND: "#C4A4A4", Ops.BUFFER: "#B0BDFF", Ops.COPY: "#a040a0", Ops.FUSE: "#FFa500",
                Ops.ALLREDUCE: "#ff40a0", Ops.MSELECT: "#d040a0", Ops.MSTACK: "#d040a0", Ops.CONTIGUOUS: "#FFC14D",
-               Ops.BUFFERIZE: "#FF991C", Ops.REWRITE_ERROR: "#ff2e2e", Ops.AFTER: "#8A7866", Ops.END: "#524C46"}
+               Ops.BUFFERIZE: "#FF991C", Ops.REWRITE_ERROR: "#ff2e2e", Ops.SUBSTITUTE: "#ffff00", Ops.AFTER: "#8A7866"}
 
 # VIZ API
 
@@ -33,8 +33,8 @@ def get_rewrites(t:RewriteTrace) -> list[dict]:
     steps = [{"name":s.name, "loc":s.loc, "match_count":len(s.matches), "code_line":printable(s.loc),
               "query":f"/ctxs?ctx={i}&idx={j}", "depth":s.depth} for j,s in enumerate(v)]
     if isinstance(k.ret, ProgramSpec):
-      steps.append({"name":"View Program", "query":f"/render?ctx={i}&fmt=src", "depth":0})
-      steps.append({"name":"View Disassembly", "query":f"/render?ctx={i}&fmt=asm", "depth":0})
+      steps.append({"name":"View Program", "query":f"/render?ctx={i}&fmt=src"})
+      steps.append({"name":"View Disassembly", "query":f"/render?ctx={i}&fmt=asm"})
     for key in k.keys: ref_map[key] = i
     ret.append({"name":k.display_name, "steps":steps})
   return ret
@@ -71,7 +71,7 @@ def uop_to_json(x:UOp, ignore_indexing=False) -> dict[int, dict]:
     if u.op in GroupOp.Movement: argst = (mask_to_str if u.op in {Ops.SHRINK, Ops.PAD} else shape_to_str)(u.marg)
     label = f"{str(u.op).split('.')[1]}{(chr(10)+word_wrap(argst.replace(':', ''))) if u.arg is not None else ''}"
     if u.dtype != dtypes.void: label += f"\n{u.dtype}"
-    for idx,x in enumerate(u.src[:1] if u.op in {Ops.BUFFERIZE, Ops.INDEX} else (u.src if u.op is not Ops.END else [])):
+    for idx,x in enumerate(u.src[:1] if u.op in {Ops.BUFFERIZE, Ops.INDEX} else u.src):
       if x in excluded:
         arg = f"{x.arg:g}" if x.op is Ops.CONST and dtypes.is_float(x.dtype) else f"{x.arg}"
         label += f"\n{x.op.name}{idx} {arg}" + (f" {x.src[0].op}" if len(x.src) else "")
@@ -82,8 +82,6 @@ def uop_to_json(x:UOp, ignore_indexing=False) -> dict[int, dict]:
         label += f"\n{shape_to_str(u.shape)}"
       if u.op in {Ops.INDEX, Ops.BUFFERIZE}:
         label += f"\n{u.render()}"
-      if u.op is Ops.END:
-        label += "\n"+' '.join([f"{colored(u.src[i].arg[0], axis_colors[u.src[i].arg[-1]])}({u.src[i].vmax+1})" for i in range(u.arg)])
     except Exception:
       label += "\n<ISSUE GETTING LABEL>"
     if (ref:=ref_map.get(u.arg.ast) if u.op is Ops.KERNEL else None) is not None: label += f"\ncodegen@{ctxs[ref]['name']}"