split devectorizing of buf/index

* add rule and test

* change comment

.github/workflows/test.yml

@@ -343,6 +343,8 @@ jobs:
       run: |
         python -m mypy --strict-equality --lineprecision-report .
         cat lineprecision.txt
+    - name: Run TYPED=1
+      run: TYPED=1 python -c "import tinygrad"
 
   unittest:
     name: Unit Tests

docs/tinybox.md

@@ -6,7 +6,7 @@ If you don't have a tinybox and you want one, see [tinygrad.org](https://tinygra
 
 ## Welcome
 
-Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, and the green box includes six 4090 GPUs. Whether you bought a red one or a green one, we want you to love it.
+Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, the green box includes six 4090 GPUs, and the green v2 box includes four 5090 GPUs. Whether you bought a red one or a green one, we want you to love it.
 
 We don't have a stupid cloud service, you don't have to create a tiny account to set it up, and we aren't tracking how you use the box. We're just happy you bought one. This petaflop is your petaflop.
 

extra/gemm/amd_uop_matmul.py

@@ -5,7 +5,7 @@ from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import getenv, colored, prod, unwrap
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.shape.view import strides_for_shape
-from tinygrad.codegen.opt.kernel import axis_colors
+from tinygrad.codegen.opt.kernel import axis_colors, Opt, OptOps
 from tinygrad.codegen.opt.swizzler import merge_views, view_left
 
 def to_colored(full_shape, axis_types): return '_'.join([colored(str(s), axis_colors[at]) for s,at in zip(full_shape, axis_types)])
@@ -44,6 +44,21 @@ pm = PatternMatcher([
   (UPat(Ops.VIEW, src=(UPat(Ops.REDUCE_AXIS, src=(UPat.var("src"),), name="r"),), name="view"), swizzle_reduceop),
 ])
 
+def rangeify_kernel3():
+  a = Tensor.empty(N,N)
+  b = Tensor.empty(N,N)
+  c = a@b
+  #c = c.reshape((32,2,16,4,32,2,16,4)).contiguous()
+  with Context(RANGEIFY=1):
+    sink = c.schedule()[-1].ast
+  #print(sink)
+
+  opts  = [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.UPCAST, 0, 2)]
+  opts += [Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.LOCAL, 1, 16), Opt(OptOps.UPCAST, 1, 2)]
+  opts += [Opt(OptOps.UNROLL, 0, 8)]
+
+  return sink.replace(arg=KernelInfo(opts_to_apply=tuple(opts)))
+
 def top_spec_kernel3():
   a = Tensor.empty(N,N)
   b = Tensor.empty(N,N)
@@ -309,10 +324,15 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
 
 if __name__ == "__main__":
   HL = getenv("HL")
-  if HL == 2: hprg = top_spec_kernel3()
+  if HL == 3: hprg = rangeify_kernel3()
+  elif HL == 2: hprg = top_spec_kernel3()
   elif HL == 1: hprg = hl_spec_kernel3()
   else: hprg = hand_spec_kernel3()
-  prg = get_program(hprg, Device.default.renderer)
+  if HL == 3:
+    with Context(RANGEIFY=1, BLOCK_REORDER=0):
+      prg = get_program(hprg, Device.default.renderer)
+  else:
+    prg = get_program(hprg, Device.default.renderer)
   print(prg.src)
   if getenv("SRC"): exit(0)
   hrunner = CompiledRunner(prg)

setup.py

@@ -29,6 +29,7 @@ setup(name='tinygrad',
         'tinygrad.apps',
         'tinygrad.codegen',
         'tinygrad.codegen.opt',
+        'tinygrad.codegen.late',
         'tinygrad.engine',
         'tinygrad.frontend',
         'tinygrad.nn',
@@ -63,6 +64,7 @@ setup(name='tinygrad',
             "pre-commit",
             "ruff",
             "numpy",
+            "typeguard",
         ],
         #'mlperf': ["mlperf-logging @ git+https://github.com/mlperf/[email protected]"],
         'testing_minimal': testing_minimal,

test/test_dtype.py

@@ -414,11 +414,11 @@ class TestDtypeUsage(unittest.TestCase):
         t = Tensor([[1, 2], [3, 4]], dtype=d)
         (t*t).max().item()
 
-@unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), f"no bfloat16 on {Device.DEFAULT}")
+@unittest.skipUnless(is_dtype_supported(dtypes.bfloat16) or Device.DEFAULT == "PYTHON", f"no bfloat16 on {Device.DEFAULT}")
 class TestOpsBFloat16(unittest.TestCase):
   def test_cast(self):
     # TODO: helper_test_op breaks in unrelated part
-    # TODO: wrong output with GPU=1 / PYTHON=1 on mac
+    # TODO: wrong output with GPU=1 on mac
     data = [60000.0, 70000.0, 80000.0]
     np.testing.assert_allclose(Tensor(data).cast("bfloat16").numpy(), torch.tensor(data).type(torch.bfloat16).float().numpy())
 

test/test_pickle.py

@@ -20,7 +20,7 @@ class TestPickle(unittest.TestCase):
     self.assertEqual(pm2.rewrite(sink).key, tt.key)
 
   def test_pickle_main_pattern_matcher(self):
-    from tinygrad.codegen.devectorizer import sym
+    from tinygrad.codegen.late.devectorizer import sym
     ssym = pickle.dumps(sym)
     dsym = pickle.loads(ssym)
     self.assertEqual(dsym.patterns[0][0].location, sym.patterns[0][0].location)

test/test_tensor.py

@@ -415,6 +415,21 @@ class TestTinygrad(unittest.TestCase):
         data = _generate_data(depth)
         np.testing.assert_allclose(Tensor(data).numpy(), np.array(data))
 
+  def test_tensor_list_implicit_cast(self):
+    data = [True, False]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+    data = [-1, 0, 1, 2, 3]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+    data = [-3.5, -2.5, -1.5, 0, 1.5, 2.5, 3.5]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    # NOTE: torch and jax raise OverflowError: Python integer -3 out of bounds for uint8
+    # np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+
   def test_tensor_list_special_values(self):
     if is_dtype_supported(dtypes.float16):
       data = [math.nan, -math.inf, 65504, 65519, 65519.999, 65520, 65520.1]

test/test_uop_graph.py

@@ -6,7 +6,7 @@ from tinygrad.helpers import DEBUG, Context
 from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, track_rewrites, graph_rewrite, GroupOp
 from tinygrad.uop.symbolic import sym
 from tinygrad.codegen import full_rewrite, full_rewrite_to_sink
-from tinygrad.codegen.expander import expander
+from tinygrad.codegen.late.expander import expander
 
 simple_pm = PatternMatcher([
   (UPat.cvar('x', dtypes.int), lambda x: UOp.const(dtypes.float, 1.0) + UOp.const(dtypes.float, 2.0)),

test/test_uops.py

@@ -402,6 +402,14 @@ class TestAssembly(unittest.TestCase):
     self.assertIn(Ops.SHR, ops)
     self.assertNotIn(Ops.IDIV, ops)
 
+  def test_fast_idiv_remove_powers_of_two(self):
+    ridx = UOp.range(dtypes.int, 2**20, 0)
+    uops = to_uops_list([ridx//(7*64)], opts=Device[Device.DEFAULT].renderer)
+    ops = [x.op for x in uops]
+    # this requires shifting out the powers of two before doing fast_idiv
+    # (((ridx0>>6)*18725)>>17) instead of (int)((((long)(ridx0)*1198373)>>29))
+    self.assertNotIn(Ops.CAST, ops)
+
   def test_mulacc_unrolled(self):
     # test that     acc = acc + a0*b0 + a1*b1 + a2*b2 + a3*b3
     # is not        acc = acc + (a0*b0 + a1*b1 + a2*b2 + a3*b3)

test/unit/test_block_reorder.py

@@ -1,7 +1,7 @@
 import unittest, random
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import print_uops, UOp, Ops
-from tinygrad.codegen.linearize import block_reorder
+from tinygrad.codegen.late.linearize import block_reorder
 from tinygrad.renderer.cstyle import OpenCLRenderer
 
 def is_toposorted(lst:list[UOp]):

test/unit/test_dtype.py

@@ -56,6 +56,7 @@ class TestCastConvenienceMethod(unittest.TestCase):
 class TestDtypeTolist(unittest.TestCase):
   def test_bfloat16(self):
     self.assertEqual(Tensor([-60000, 1.5, 3.1, 60000], device="PYTHON", dtype=dtypes.bfloat16).tolist(), [-59904.0, 1.5, 3.09375, 59904.0])
+  def test_fp8(self):
     # 448
     self.assertEqual(Tensor([-30000, 1.5, 3.1, 30000], device="PYTHON", dtype=dtypes.fp8e4m3).tolist(), [-448.0, 1.5, 3.0, 448.0])
     # 57344

test/unit/test_dtype_spec.py

@@ -1,6 +1,6 @@
-import unittest, math, operator, subprocess
+import unittest, math, operator, subprocess, struct
 from tinygrad.tensor import Tensor, dtypes, Device
-from tinygrad.dtype import DType, DTYPES_DICT, truncate, truncate_fp16, truncate_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
+from tinygrad.dtype import DType, DTYPES_DICT, truncate, truncate_fp16, float_to_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv, CI, DEBUG
 from hypothesis import given, settings, strategies as strat
@@ -26,6 +26,9 @@ def _assert_eq(tensor:Tensor, target_dtype:DType, target, tol_target_dtype:float
   except AssertionError as e:
     raise AssertionError(f"\ntensor {tensor.numpy()} dtype {tensor.dtype} does not match target {target} with dtype {target_dtype}") from e
 
+def u32_to_f32(u): return struct.unpack('f', struct.pack('I', u))[0]
+def f32_to_u32(f): return struct.unpack('I', struct.pack('f', f))[0]
+
 class TestHelpers(unittest.TestCase):
   signed_ints = (dtypes.int8, dtypes.int16, dtypes.int32, dtypes.int64)
   uints = (dtypes.uint8, dtypes.uint16, dtypes.uint32, dtypes.uint64)
@@ -102,18 +105,79 @@ class TestHelpers(unittest.TestCase):
     self.assertEqual(truncate_fp16(65504), 65504)
     self.assertEqual(truncate_fp16(65519.999), 65504)
     self.assertEqual(truncate_fp16(65520), math.inf)
+    self.assertEqual(truncate_fp16(1e-8), 0.0)
+    self.assertEqual(truncate_fp16(-65504), -65504)
+    self.assertEqual(truncate_fp16(-65519.999), -65504)
+    self.assertEqual(truncate_fp16(-65520), -math.inf)
+    self.assertTrue(math.isnan(truncate_fp16(math.nan)))
 
-  def test_truncate_bf16(self):
-    self.assertEqual(truncate_bf16(1), 1)
-    self.assertAlmostEqual(truncate_bf16(1.1), 1.09375, places=7)
-    for a in [1234, 23456, -777.777]:
-      self.assertEqual(truncate_bf16(a), torch.tensor([a], dtype=torch.bfloat16).item())
-    # TODO: torch bfloat 1.1 gives 1.1015625 instead of 1.09375
+  def test_float_to_bf16(self):
+    # TODO: fuzz this better
     max_bf16 = torch.finfo(torch.bfloat16).max
-    self.assertEqual(truncate_bf16(max_bf16), max_bf16)
-    self.assertEqual(truncate_bf16(min_bf16:=-max_bf16), min_bf16)
-    self.assertEqual(truncate_bf16(max_bf16 * 1.00001), math.inf)
-    self.assertEqual(truncate_bf16(min_bf16 * 1.00001), -math.inf)
+    for a in [1, 1.1, 1234, 23456, -777.777, max_bf16, max_bf16 * 1.00001, -max_bf16, -max_bf16 * 1.00001, math.inf, -math.inf]:
+      self.assertEqual(float_to_bf16(a), torch.tensor([a], dtype=torch.bfloat16).item())
+    self.assertTrue(math.isnan(float_to_bf16(math.nan)))
+
+  def test_float_to_bf16_nan(self):
+    # In f32, NaN = exp 0xFF and mantissa ≠ 0. Quiet-vs-signaling is bit 22 of the mantissa: 1 = qNaN, 0 = sNaN.
+    # qNaN(+/-), sNaN(+/-) overflow(+/-)
+    patterns = [0x7FC00001, 0xFFC00001, 0x7F800001, 0xFF800001, 0x7FFFFFFF, 0xFFFFFFFF]
+    for u in patterns:
+      x = u32_to_f32(u)
+      y = float_to_bf16(x)
+      t = torch.tensor([x], dtype=torch.bfloat16).item()
+      self.assertTrue(math.isnan(y))
+      self.assertTrue(math.isnan(t))
+
+  def test_float_to_bf16_round(self):
+    # round_to_nearest_even
+    uppers = [0x3f800000, 0x41230000, 0xC1460000] # 1.0, 10.1875, -12.375
+    for upper in uppers:
+      base = upper & 0xFFFF0000
+      base_f32 = u32_to_f32(base)
+      base_f32_round_up = u32_to_f32(base + 0x00010000)
+
+      # low < 0x8000(0.5ULP) -> round down
+      x = u32_to_f32(base | 0x00007000)
+      self.assertEqual(float_to_bf16(x), base_f32)
+      self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32)
+
+      # low > 0x8000(0.5ULP) -> round up
+      x = u32_to_f32(base | 0x0000C000)
+      self.assertEqual(float_to_bf16(x), base_f32_round_up)
+      self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32_round_up)
+
+      # low == 0x8000(0.5ULP) and LSB even -> round down
+      if ((upper >> 16) & 1) == 0:
+        x = u32_to_f32(base | 0x00008000)
+        self.assertEqual(float_to_bf16(x), base_f32)
+        self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32)
+      # low == 0x8000(0.5ULP) and LSB odd -> round up
+      else:
+        x = u32_to_f32(base | 0x00008000)
+        self.assertEqual(float_to_bf16(x), base_f32_round_up)
+        self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32_round_up)
+
+  def test_float_to_bf16_boundary(self):
+    # bf16 max finite: exp=0xFE, faction=0x7F => 0x7F7F0000(f32)
+    # bf16 inf(+/-):   exp=0xFF
+    base = 0x7F7F0000
+    inf_u32 = 0x7F800000
+
+    # low < 0.5ULP
+    x = u32_to_f32(base | 0x00007FFF)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), base)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), base)
+
+    # low > 0.5ULP -> overflows to +inf
+    x = u32_to_f32(base | 0x0000C000)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), inf_u32)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), inf_u32)
+
+    # low == 0.5ULP and LSB odd -> overflows to +inf
+    x = u32_to_f32(base | 0x00008000)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), inf_u32)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), inf_u32)
 
   @given(strat.floats(width=32, allow_subnormal=True, allow_nan=True, allow_infinity=True))
   def test_truncate_fp8e4m3(self, x):

test/unit/test_shapetracker.py

@@ -6,7 +6,7 @@ from tinygrad.helpers import prod
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad import Variable
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite
-from tinygrad.codegen.devectorizer import sym
+from tinygrad.codegen.late.devectorizer import sym
 from itertools import product
 
 def shapetracker_getitem(st:ShapeTracker, val:int):

test/unit/test_uop_symbolic.py

@@ -4,7 +4,7 @@ import z3
 
 from tinygrad.dtype import dtypes, ConstType
 from tinygrad.codegen import full_rewrite
-from tinygrad.codegen.devectorizer import sym
+from tinygrad.codegen.late.devectorizer import sym
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad import Variable
@@ -128,6 +128,8 @@ class TestSymbolic(unittest.TestCase):
     b = Variable("b", 0, 8)
     self.helper_test_variable(a+a, 0, 16, "(a*2)")
     self.helper_test_variable((a+b)+b, 0, 24, "(a+(b*2))")
+    self.helper_test_variable((a*3+b)+a, 0, 40, "(b+(a*4))")
+    self.helper_test_variable((a+b)+a*3, 0, 40, "(b+(a*4))")
 
   def test_sub_self(self):
     a = Variable("a", 0, 8)
@@ -448,7 +450,8 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((-Variable("a", 10, 10))%7, -3, -3, "-3")
 
   def test_div_numerator_negative(self):
-    self.helper_test_variable((Variable("idx", 0, 9)*-10)//11, -8, 0, "(((idx*10)//11)*-1)")
+    with Context(CORRECT_DIVMOD_FOLDING=1):
+      self.helper_test_variable((Variable("idx", 0, 9)*-10)//11, -8, 0, "(((idx*10)//11)*-1)")
 
   def test_nest_div_negative_factor(self):
     ridx0=UOp.variable("ridx0", 0, 9)
@@ -637,15 +640,16 @@ class TestSymbolic(unittest.TestCase):
     cond = Variable("x", 0, 3) < 2
     a = Variable("a", 0, 3)
     b = Variable("b", 0, 3)
+    c = Variable("c", 0, 3)
     aa = cond.where(a, a.ufix(0))
     bb = cond.where(b, b.ufix(1))
     self.helper_test_variable(aa, 0, 3, "(a if (x<2) else 0)")
     self.helper_test_variable(bb, 0, 3, "(b if (x<2) else 1)")
     self.helper_test_variable(aa+bb, 0, 6, "((a+b) if (x<2) else 1)")
     self.helper_test_variable(aa.maximum(bb), 0, 3, "(max(a, b) if (x<2) else 1)")
+    self.helper_test_variable((c+aa)+bb, 0, 9, "(c+((a+b) if (x<2) else 1))")
 
     # not combining because it increased total ALU
-    c = Variable("c", 0, 3)
     cc = cond.where(c, c+1)
     self.helper_test_variable(bb+cc, 0, 7, "((b if (x<2) else 1)+(c if (x<2) else (c+1)))")
 

test/unit/test_viz.py

@@ -276,16 +276,15 @@ def load_profile(lst:list[ProfileEvent]) -> dict:
     layout[k] = v = {"shapes":[]}
     event_type, event_count = u("<BI")
     if event_type == 0:
-      v["max_depth"] = u("<B")
       for _ in range(event_count):
-        name, ref, st, dur, depth, cat, _ = u("<IIIfBBI")
-        v["shapes"].append({"name":strings[name], "ref":option(ref), "st":st, "dur":dur, "depth":depth, "cat":option(cat)})
+        name, ref, st, dur, cat, _ = u("<IIIfBI")
+        v["shapes"].append({"name":strings[name], "ref":option(ref), "st":st, "dur":dur, "cat":option(cat)})
     else:
       v["peak"] = u("<Q")[0]
-      v["timestamps"] = list(u(f"<{u('I')[0]}I"))
       for _ in range(event_count):
-        i = u("<I")[0]
-        v["shapes"].append({"x":list(u(f"<{i}I")), "y":list(u(f"<{i}Q")), "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
+        alloc, ts, key = u("<BII")
+        if alloc: v["shapes"].append({"event":"alloc", "ts":ts, "key":key, "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
+        else: v["shapes"].append({"event":"free", "ts":ts, "key":key})
   return {"dur":dur, "peak":global_peak, "layout":layout}
 
 class TestVizProfiler(unittest.TestCase):
@@ -354,7 +353,7 @@ class TestVizProfiler(unittest.TestCase):
     n_events = 1_000
     prof = [ProfileRangeEvent("CPU", name="k_test", st=decimal.Decimal(ts:=i*step), en=decimal.Decimal(ts)+step) for i in range(n_events)]
     sz = len(get_profile(prof))
-    self.assertLessEqual(sz/n_events, 27)
+    self.assertLessEqual(sz/n_events, 26)
 
   # can pack up to 1hr 11 min of trace events
   def test_trace_duration(self):
@@ -377,8 +376,7 @@ class TestVizMemoryLayout(BaseTestViz):
     profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{a.device} Memory"]
     self.assertEqual(ret["peak"], 2)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 2])
-    self.assertEqual(ret["shapes"][1]["x"], [1, 2])
+    self.assertEqual(len(ret["shapes"]), 2)
 
   def test_del_once(self):
     a = _alloc(1)
@@ -387,10 +385,7 @@ class TestVizMemoryLayout(BaseTestViz):
     profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{b.device} Memory"]
     self.assertEqual(ret["peak"], 1)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 2])
-    self.assertEqual(ret["shapes"][1]["x"], [2, 3])
-    self.assertEqual(ret["shapes"][0]["y"], [0, 0])
-    self.assertEqual(ret["shapes"][1]["y"], [0, 0])
+    self.assertEqual(len(ret["shapes"]), 3)
 
   def test_alloc_free(self):
     a = _alloc(1)
@@ -400,12 +395,7 @@ class TestVizMemoryLayout(BaseTestViz):
     profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{c.device} Memory"]
     self.assertEqual(ret["peak"], 2)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 3])
-    self.assertEqual(ret["shapes"][1]["x"], [1, 3, 3, 4])
-    self.assertEqual(ret["shapes"][0]["y"], [0, 0])
-    self.assertEqual(ret["shapes"][1]["y"], [1, 1, 0, 0])
-    self.assertEqual(ret["shapes"][2]["x"], [3, 4])
-    self.assertEqual(ret["shapes"][2]["y"], [1, 1])
+    self.assertEqual(len(ret["shapes"]), 4)
 
 if __name__ == "__main__":
   unittest.main()

tinygrad/codegen/__init__.py

@@ -12,12 +12,13 @@ from tinygrad.codegen.quantize import pm_quant
 from tinygrad.codegen.gpudims import pm_add_gpudims
 from tinygrad.uop.symbolic import sym, symbolic_simple, gep_pushing
 from tinygrad.uop.decompositions import get_late_rewrite_patterns
-from tinygrad.codegen.expander import migrate_indexing, expander
-from tinygrad.codegen.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
+from tinygrad.codegen.late.expander import migrate_indexing, expander
+from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
   ReduceContext, correct_load_store, pm_render
-from tinygrad.codegen.linearize import block_create, pm_blockend_merge, block_merge, pm_finalize, BlockContext
-from tinygrad.codegen.opt import pm_optimize
+from tinygrad.codegen.late.linearize import block_create, pm_blockend_merge, block_merge, pm_finalize, BlockContext
+from tinygrad.codegen.opt import pm_get_optimization, pm_do_optimize
 from tinygrad.codegen.opt.swizzler import view_left, view_right, fix_kernel_ops
+from tinygrad.schedule.rangeify import pm_add_buffers, rangeify_codegen
 
 @dataclass
 class RewriteStep:
@@ -55,7 +56,8 @@ def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVEC
   ret.extend(rewrites_for_views)
 
   # this is kernel.py
-  ret.append(RewriteStep(pm_optimize, ctx=lambda _: opts, name="optimize ast"))
+  ret.append(RewriteStep(pm_get_optimization, ctx=lambda _: opts, name="get optimization"))
+  ret.append(RewriteStep(pm_do_optimize, ctx=lambda _: opts, name="optimize ast"))
 
   if _QUANTIZE and opts.device in {"CPU", "DSP"}: ret.append(RewriteStep(pm_quant, name="quantize"))
   ret.append(RewriteStep(pm_lowerer, get_index, name="lowerer", bottom_up=True))
@@ -69,6 +71,9 @@ def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVEC
   # expand
   ret.append(RewriteStep(sym+expander, name="expander"))
 
+  # add locals
+  ret.append(RewriteStep(pm_add_buffers+rangeify_codegen, name="add local buffers"))
+
   # ** devectorizer (full_graph_rewrite) **
   # remove reduce
   ret.append(RewriteStep(pm_reduce+gep_pushing, lambda _: ReduceContext(), name="remove_reduce"))

tinygrad/codegen/gpudims.py

@@ -1,7 +1,7 @@
-import math
+import math, functools, operator
 from tinygrad.uop.ops import UOp, Ops, sint, PatternMatcher, UPat, KernelInfo, ssimplify, AxisType
 from tinygrad.helpers import all_int, partition, flatten, prod, dedup
-from tinygrad.dtype import dtypes
+from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.shape.view import get_contraction
 from tinygrad.renderer import Renderer
 
@@ -83,7 +83,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
     if r.op is not Ops.RANGE: continue
     try:
       ii = (global_dims+local_dims).index(r.arg[0]%1000)
-      if r.arg[0] < 2000 and ki.axis_types[r.arg[0]%1000] == AxisType.GROUP_REDUCE: continue
+      if r.arg[1] == AxisType.REDUCE: continue
       subs[r] = idxs[ii]
     except ValueError: continue
   return s.substitute(subs)
@@ -91,7 +91,8 @@ def add_gpudims(ctx:Renderer, s:UOp):
 def fix_reduce_unroll(x:UOp):
   reduce_range, reduce_expand = partition(x.src[1:], lambda y: y.op is Ops.RANGE)
   if len(reduce_expand) == 0: return None
-  assert all(x.op is Ops.UNROLL for x in reduce_expand), f"not all UNROLLS in {reduce_expand} for {x.axis_arg}"
+  reduce_expand = [x for x in reduce_expand if x.op is not Ops.CONST]
+  assert all(x.op is Ops.UNROLL for x in reduce_expand), f"not all UNROLLS in {reduce_expand}"
   ret = x.src[0]
   if len(contract_axis:=flatten(x.arg for x in reduce_expand)):
     ret = UOp(Ops.CONTRACT, x.dtype.vec(prod(x[1] for x in contract_axis)), (ret,), tuple(contract_axis), tag=1)
@@ -103,7 +104,24 @@ def fix_store_unroll(x:UOp):
   if len(store_expand) == 0: return None
   return UOp(Ops.CONTRACT, dtypes.void, (x.replace(src=x.src[:2]+tuple(store_range)),), tuple(flatten(x.arg for x in store_expand)), tag=1)
 
+def fix_group_for_reduce(x:UOp):
+  reduce_gfr, reduce_r = partition(x.src[1:], lambda u: u.op is Ops.RANGE and u.arg[1] == AxisType.GROUP_REDUCE)
+  if len(reduce_gfr) == 0: return None
+
+  # NOTE: if there's other locals here, we need them in the buffer too
+  upstream_locals = [u for u in x.toposort() if u.op is Ops.RANGE and u.arg[1] == AxisType.LOCAL]
+
+  # do only the non grouped reduces early
+  ret = x.replace(src=(x.src[0],)+tuple(reduce_r))
+  reduce_loop = [x.replace(arg=(x.arg[0]+100, AxisType.REDUCE)) for x in reduce_gfr]
+  buf = ret.bufferize(*upstream_locals, *reduce_gfr, arg=(AddrSpace.LOCAL, reduce_gfr[0].arg[0])).index(*upstream_locals, *reduce_loop)
+
+  # 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_add_gpudims = PatternMatcher([
+  # add gpudims must be last
   (UPat(Ops.SINK, name="s"), add_gpudims),
   # rewrite UPCAST/UNROLL range to something to be expanded
   (UPat(Ops.RANGE, name="r"),
@@ -112,4 +130,6 @@ pm_add_gpudims = PatternMatcher([
   # fix REDUCEs with UNROLLs
   (UPat(Ops.REDUCE, name="x"), fix_reduce_unroll),
   (UPat(Ops.STORE, name="x"), fix_store_unroll),
+  # fix group for reduce
+  (UPat(Ops.REDUCE, name="x"), fix_group_for_reduce),
 ])

tinygrad/codegen/late/devectorizer.py

@@ -232,17 +232,21 @@ def no_vectorized_alu(alu:UOp):
   alus = tuple(UOp(alu.op, alu.dtype.scalar(), tuple(s.gep(i) for s in alu.src), alu.arg) for i in range(alu.dtype.vcount))
   return UOp(Ops.VECTORIZE, alu.dtype, alus)
 
-def no_vectorized_acc(acc:UOp, c:UOp):
-  if acc.dtype.count == 1: return None
-  assert c.arg == 0, "this only supports index 0"
-  new_acc = acc.replace(dtype=acc.dtype.base.scalar().ptr(acc.dtype.count, cast(PtrDType, acc.dtype).addrspace))
-  return UOp(Ops.PTRCAT, acc.dtype, tuple([new_acc.index(UOp.const(dtypes.int, i)) for i in range(acc.dtype.count)]))
+def no_vectorized_buf(buf:UOp):
+  dtype = cast(PtrDType, buf.dtype)
+  return buf.replace(dtype=dtype.base.scalar().ptr(dtype.size*dtype.count, dtype.addrspace)).cast(dtype)
+
+def no_vectorized_index(buf:UOp, cast:UOp, idx:UOp):
+  cnt = cast.dtype.count
+  assert idx.dtype.count == 1, f"idx dtype must be 1 {idx.dtype}"
+  return buf.broadcast(cnt).index(idx.broadcast(cnt)*cnt+UOp.const(dtypes.int.vec(cnt), tuple(range(cnt))))
 
 devectorize = PatternMatcher([
   # no ALU on vectorized dtypes
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST), name="alu"), no_vectorized_alu),
   (UPat(Ops.WMMA, name="wmma"), no_vectorized_wmma),
-  (UPat(Ops.DEFINE_REG, name="acc").index(UPat.cvar("c")), no_vectorized_acc),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf"), no_vectorized_buf),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf").cast(name="cast").index(UPat.var("idx")), no_vectorized_index),
 ])
 
 pm_render = PatternMatcher([

tinygrad/codegen/late/expander.py

@@ -1,7 +1,7 @@
 # this converts a lowerer program into a vectorized program
 
 import functools, itertools, operator
-from tinygrad.dtype import dtypes
+from tinygrad.dtype import dtypes, PtrDType
 from tinygrad.helpers import AMX, dedup, flatten, all_same, prod
 from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, GroupOp
 
@@ -50,9 +50,11 @@ def do_expand(root:UOp):
       if root.op is Ops.IF or src.op is Ops.IF:
         # for the first arg of IF, just pass them through ignoring UNROLLS
         new_srcs.append(src)
-      elif (root.op is Ops.STORE and i >= 2) or (root.op is Ops.REDUCE and i >= 1):
+      elif (root.op is Ops.STORE and i >= 2) or (root.op in {Ops.REDUCE, Ops.BUFFERIZE} and i >= 1):
         # for any range args of STORE/REDUCE, pass them through
         new_srcs.append(src)
+      elif root.op is Ops.INDEX and i >= 1 and not isinstance(root.dtype, PtrDType):
+        new_srcs.append(src)
       elif src.dtype.count > 1:
         # put any input dtype > 1 grouped together
         new_srcs.append(UOp(Ops.CAT, src.dtype.scalar().vec(expand_sz*src.dtype.count), (src,)*expand_sz))
@@ -84,7 +86,7 @@ expander = PatternMatcher([
   (UPat(Ops.UNROLL, name="outer", src=(UPat(Ops.UNROLL, name="inner"),)),
    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,
+  (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), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
   (UPat(Ops.CONTRACT, name="con"), do_contract),
   # BARRIERs aren't actually expanded

tinygrad/codegen/late/linearize.py

@@ -3,7 +3,7 @@ import heapq
 from collections import defaultdict
 from dataclasses import dataclass, replace
 from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp
-from tinygrad.helpers import dedup, all_same, flatten, getenv
+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]:
@@ -150,7 +150,7 @@ def make_block_bottom_up(ctx:BlockContext, x:UOp):
     srcs.append(add_blockends(base_block, new_ctx, current_ctx))
 
   lst = lst[::-1]
-  if getenv("BLOCK_REORDER", 1): lst = block_reorder(lst)
+  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)
 

tinygrad/codegen/lowerer.py

@@ -1,8 +1,6 @@
 # the job of the lowerer is to do indexing
-import functools, operator
-from typing import cast
 from dataclasses import dataclass
-from tinygrad.dtype import dtypes, AddrSpace, PtrDType
+from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import KernelInfo, UOp, Ops, PatternMatcher, UPat, sint_to_uop, AxisType, graph_rewrite
 
 # ***** indexing *****
@@ -18,7 +16,8 @@ def shape_to_idx(s, axis_types, start=0):
 
 def get_index(ast:UOp) -> IndexContext:
   axis_types = ast.arg.axis_types if isinstance(ast.arg, KernelInfo) else ()
-  if len(ast.full_shape) != len(axis_types): axis_types = (AxisType.LOOP,)*len(ast.full_shape)
+  if len(ast.full_shape) != len(axis_types):
+    axis_types = tuple([AxisType.REDUCE if s is not fs else AxisType.LOOP for s,fs in zip(ast.shape, ast.full_shape)])
   return IndexContext(axis_types, [], 0)
 
 # ***** lowering (given index) *****
@@ -49,15 +48,7 @@ def lower_store(ctx: IndexContext, x: UOp, buf: UOp):
 
   stored = subblock(ctx, real_new_idxs, x.src[1])
   used_ranges = [x for x in used_idxs if x.op is Ops.RANGE]
-  ret = buf.index(idx, valid).store(stored, *used_ranges)
-
-  # insert BARRIER if we are ending a LOCAL, IF if we are ending a GROUP_REDUCE
-  if cast(PtrDType, buf.dtype).addrspace == AddrSpace.LOCAL and \
-      any(ctx.axis_types[x.arg[0]%1000] in {AxisType.GROUP_REDUCE, AxisType.LOCAL} for x in used_ranges):
-    ret = ret.barrier()
-    range_gates = [x.eq(0) for x in used_ranges if ctx.axis_types[x.arg[0]%1000] == AxisType.GROUP_REDUCE]
-    if len(range_gates): ret = UOp(Ops.IF, src=(functools.reduce(operator.and_, range_gates), ret))
-  return ret
+  return buf.index(idx, valid).store(stored, *used_ranges)
 
 def fixup_wmma(ctx:IndexContext, x:UOp):
   if x.tag is not None: return None

tinygrad/codegen/opt/__init__.py

@@ -2,7 +2,7 @@
 
 from tinygrad.codegen.opt.kernel import Kernel
 from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops
+from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, KernelInfo
 from tinygrad.helpers import NOOPT, BEAM, USE_TC, getenv
 from tinygrad.renderer import Renderer
 from tinygrad.uop.spec import type_verify
@@ -19,20 +19,28 @@ def get_optimized_ast(ast:UOp, renderer:Renderer) -> UOp:
     The Ops.SINK rooted AST transformed to apply the opts and with a KernelInfo in the arg.
   """
 
+  assert ast.arg is None, "no opt if there's an arg"
   k = Kernel(ast, opts=renderer)
-  if ast.arg is not None and ast.arg.opts_to_apply is not None: k.apply_opts(ast.arg.opts_to_apply)
-  elif not NOOPT:
+  if not NOOPT:
     if not k.apply_tensor_cores(USE_TC.value): k.apply_opts(hand_coded_optimizations(k))
     if BEAM >= 1:
       from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
       kb = Kernel(ast, opts=renderer)
       rawbufs = bufs_from_lin(kb, allocate=False)
       k = beam_search(kb, rawbufs, BEAM.value, bool(getenv("BEAM_ESTIMATE", 1)))
+  return ast.replace(arg=KernelInfo(opts_to_apply=tuple(k.applied_opts)))
+
+pm_get_optimization = PatternMatcher([
+  (UPat(Ops.SINK, name="ast"), lambda ctx,ast: get_optimized_ast(ast, ctx) if ast.arg is None and ast.src[0].st is not None else None),
+])
+
+def apply_opt(ast:UOp, renderer:Renderer):
+  k = Kernel(ast, opts=renderer)
+  k.apply_opts(ast.arg.opts_to_apply)
   ret = k.get_optimized_ast()
   if __debug__: type_verify(list(ret.toposort()))
   return ret
 
-pm_optimize = PatternMatcher([
-  (UPat(Ops.SINK, name="ast"), lambda ctx,ast:
-   get_optimized_ast(ast, ctx) if (ast.arg is None or ast.arg.opts_to_apply is not None) and ast.src[0].st is not None else None),
+pm_do_optimize = PatternMatcher([
+  (UPat(Ops.SINK, name="ast"), lambda ctx,ast: apply_opt(ast, ctx) if ast.arg is not None and ast.arg.opts_to_apply is not None else None),
 ])

tinygrad/codegen/opt/heuristic.py

@@ -28,7 +28,7 @@ def hand_coded_optimizations(k:Kernel) -> list[Opt]:
           return k.applied_opts
 
   # are we grouping? (requires local shape support)
-  if resolve(prod(k.sts[0].shape[i] for i in k.upcastable_dims) <= 2048, False):
+  if resolve(prod(k.output_shape[i] for i in k.upcastable_dims) <= 2048, False):
     for sz in [16]:
       try:
         k.apply_opt(Opt(OptOps.GROUPTOP, 0, sz))
@@ -62,7 +62,7 @@ def hand_coded_optimizations(k:Kernel) -> list[Opt]:
   # potentially do more upcasts of non reduce axes based on a heuristic
   is_dsp = k.opts is not None and k.opts.device == "DSP"
   upcasted_axis: set[int] = set()
-  while resolve(prod(k.sts[0].shape[i] for i in k.upcastable_dims) >= 1024):
+  while resolve(prod(k.output_shape[i] for i in k.upcastable_dims) >= 1024):
     xb_choices = []
     # consider all upcastable axes with 3 or 4 upcast (128 on the DSP)
     for axis, upcast_amount in itertools.product(k.upcastable_dims, ([128] if not len(upcasted_axis) else []) if is_dsp else [3,4]):

tinygrad/codegen/opt/kernel.py

@@ -10,7 +10,7 @@ from tinygrad.uop.spec import type_verify, ast_spec
 from tinygrad.device import Device
 from tinygrad.codegen.opt.tc import TensorCore
 from tinygrad.renderer import Renderer
-from tinygrad.dtype import ImageDType, AddrSpace
+from tinygrad.dtype import ImageDType
 from tinygrad.helpers import all_same, colored, ansilen, dedup, prod, round_up, to_function_name, unwrap, argfix, DEBUG, TC_SELECT, TC_OPT, AMX
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import strides_for_shape, get_contraction
@@ -60,7 +60,7 @@ class Kernel:
 
     self.vars: list[Variable] = self.ast.variables()
     # NOTE: this requires a specific order with the [::-1], this is likely a bug
-    self.bufs: list[UOp] = [x for x in self.ast.toposort() if x.op in GroupOp.Buffer][::-1]
+    self.bufs: list[UOp] = [x for x in self.ast.toposort() if x.op in GroupOp.Buffer and x.st is not None][::-1]
 
     # create new shapetrackers inside this kernel, we will permute them
     self.sts: list[ShapeTracker] = [x.st_arg for x in self.bufs]
@@ -122,7 +122,7 @@ class Kernel:
   @property
   def output_shape(self) -> tuple[sint, ...]: return self.sts[0].shape
   @property
-  def shape_len(self) -> int: return len(self.sts[0].shape)
+  def shape_len(self) -> int: return len(self.full_shape)
 
   def axes_of(self, *axis_type:AxisType) -> list[int]: return [i for i,t in enumerate(self.axis_types) if t in argfix(axis_type)]
   @property
@@ -174,7 +174,7 @@ class Kernel:
   # amount : the amount to take
   # top : if you want to pull that amount from the top
   # insert_at : place to insert the new stuff
-  def shift_to(self, axis:int, amount:int, new_type:AxisType, top:bool=False, insert_at:int|None=None):
+  def shift_to(self, axis:int, amount:int, new_type:AxisType, top:bool=False, insert_at:int|None=None) -> int:
     if insert_at is None: insert_at = self.shape_len
     self.axis_types.insert(insert_at, new_type)
     move_axis = axis if top else axis+1
@@ -183,6 +183,7 @@ class Kernel:
     new_axes = [i for i in range(insert_at) if i != move_axis]+[move_axis]+[i for i in range(insert_at, self.shape_len+1) if i != move_axis]
     self.reshape(new_shape_fxn)
     self.permute(new_axes)
+    return insert_at
 
   # ******************** complex simplifiers ********************
 
@@ -248,7 +249,7 @@ class Kernel:
       return axis
     except IndexError as e: raise KernelOptError from e
 
-  def apply_opt(self, opt:Opt, append_opt:bool=True):
+  def apply_opt(self, opt:Opt, append_opt:bool=True) -> int|None:
     if self.finalized: raise RuntimeError("can't optimize Kernel after it's finalized")
     if self.dont_use_locals: check(opt.op not in {OptOps.LOCAL, OptOps.GROUP, OptOps.GROUPTOP}, "not using locals")
 
@@ -262,7 +263,7 @@ class Kernel:
       check(0 < (use_tensor_cores:=cast(tuple, opt.arg)[2]) <= 2, "use_tensor_cores value is not valid")
       check(self._apply_tc_opt(use_tensor_cores, cast(int, opt.axis), tc_select, tc_opt), "no tensor core available")
       self.applied_opts.append(opt)
-      return
+      return None
 
     axis = self.real_axis(opt.op, opt.axis)
 
@@ -285,28 +286,30 @@ class Kernel:
       smem_sz = amt*acc_sz*upcast_sz*local_sz
       check(smem_sz <= self.opts.shared_max, f"exceeds maximum shared memory size: needs {smem_sz}, max {self.opts.shared_max}")
 
+    new_axis = None
     if opt.op is OptOps.LOCAL:    # cyan
       # NOTE: LLVM/CPU can use locals too, but they are treated the same as globals (still helpful for L1 cache)
       # it's disabled for now since it makes BEAM slow for little gain
       check(self.opts.has_local, "target does not support local")
       check(self.axis_types[axis] is AxisType.GLOBAL, "local is for globals")
-      self.shift_to(axis, amt, AxisType.LOCAL, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL))+1)
+      new_axis = self.shift_to(axis, amt, AxisType.LOCAL, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL))+1)
     elif opt.op in {OptOps.GROUP, OptOps.GROUPTOP}:   # green
       check(self.opts.has_local and self.opts.has_shared, "target does not support local or shared mem")
       check(self.axis_types[axis] is AxisType.REDUCE, "must be reduce axis to group")
       check(not self.tensor_core, "can't group with tensor cores")
       check(len(reduce_axes:=[i for r in self.reduceops for i in r.axis_arg]) == len(set(reduce_axes)), "can't group with parallel reduces")
-      self.shift_to(axis, amt, AxisType.GROUP_REDUCE, top=(opt.op is OptOps.GROUPTOP), insert_at=min(self.axes_of(AxisType.REDUCE)))
+      new_axis = self.shift_to(axis, amt, AxisType.GROUP_REDUCE, top=(opt.op is OptOps.GROUPTOP), insert_at=min(self.axes_of(AxisType.REDUCE)))
     elif opt.op is OptOps.UNROLL:                     # purple
       check(self.axis_types[axis] not in (AxisType.UPCAST, AxisType.UNROLL), "can't upcasted already upcasted")
       check(amt <= 32, "don't unroll more than 32")
-      self.shift_to(axis, amt, AxisType.UNROLL, insert_at=None)
+      new_axis = self.shift_to(axis, amt, AxisType.UNROLL, insert_at=None)
     elif opt.op is OptOps.UPCAST:                     # yellow
       check(axis in self.upcastable_dims, f"{axis=} not in {self.upcastable_dims=}")
       # NOTE: assume the first get_local_axes() LOCAL are for TC
       check(not (self.tensor_core and axis in self.axes_of(AxisType.LOCAL)[:len(self.tensor_core.get_local_axes())]), "can't upcast TC locals")
       check((self.opts is not None and self.opts.device == "DSP") or amt <= 16, "don't upcast more than 16")
-      self.shift_to(axis, amt, AxisType.UPCAST, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP, AxisType.UPCAST))+1)
+      new_axis = self.shift_to(axis, amt, AxisType.UPCAST,
+                               insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP, AxisType.UPCAST))+1)
     elif opt.op is OptOps.NOLOCALS:
       check(self.opts.has_local and not self.dont_use_locals, "NOLOCALS is meaningless if target does not support local or already not using locals")
       check(AxisType.LOCAL not in self.axis_types and self.group_for_reduces == 0, "can't have no locals with locals")
@@ -336,6 +339,7 @@ class Kernel:
     if append_opt: self.applied_opts.append(opt)
     if self.simplify_ones() and self.tensor_core_opts:
       self.tensor_core_opts.fix_axes(axis) # fix up axes in TC opts if required after simplify_ones()
+    return new_axis
 
   def apply_opts(self, opts:Sequence[Opt]) -> Kernel:
     for opt in opts: self.apply_opt(opt)
@@ -460,8 +464,7 @@ class Kernel:
       if op.op is Ops.REDUCE_AXIS:
         reduce_idx = len(self.bufs) + self.reduceops.index(op) * 2
         changed = tuple(i for i in range(self.shape_len) if resolve(self.sts[reduce_idx].shape[i] != self.sts[reduce_idx + 1].shape[i]))
-        axes = tuple(i for i in self.axes_of(AxisType.REDUCE, AxisType.UNROLL) if i in changed)
-        grouped_axes = tuple(i for i in self.axes_of(AxisType.GROUP_REDUCE) if i in changed)
+        axes = tuple(i for i in self.axes_of(AxisType.REDUCE, AxisType.GROUP_REDUCE, AxisType.UNROLL) if i in changed)
         if (tc := self.tensor_core) and self.use_tensor_cores == 1:
           # get reduce/upcast axes for the tensor cores
           tc_reduce_axes = self.shape_str_to_axis([f"r{i}" for i in range(len(tc.get_reduce_axes()))])
@@ -486,23 +489,6 @@ class Kernel:
           return ret.replace(src=(tc_uop,), arg=(Ops.ADD, new_axes)) if (new_axes := tuple(i for i in axes if i not in tc_reduce_axes)) else tc_uop
 
         ret = ret.replace(arg = (op.arg[0], axes))
-        if self.group_for_reduces and grouped_axes:
-          local_axes = tuple([i for i,t in enumerate(self.axis_types) if t in (AxisType.LOCAL, AxisType.UPCAST) or i in grouped_axes])
-          slocal, supcast, sgroup = sorted(self.axes_of(AxisType.LOCAL)), sorted(self.axes_of(AxisType.UPCAST)), sorted(grouped_axes)
-          # NOTE: start with UPCAST at the end so it has stride 1 and can merge
-          base_shape = tuple([self.full_shape[i] for i in slocal] + [self.full_shape[i] for i in sgroup] + [self.full_shape[i] for i in supcast])
-          permute_axes = tuple([local_axes.index(i) for i in slocal+sgroup+supcast])
-          local_shape = tuple([s if i in local_axes else 1 for i,s in enumerate(self.full_shape)])
-          local_src_shape = tuple([self.full_shape[i] if i in self.axes_of(AxisType.GLOBAL) else s for i,s in enumerate(local_shape)])
-          st = ShapeTracker.from_shape(base_shape).permute(permute_axes).reshape(local_shape).expand(local_src_shape)
-          local_size = st.real_size()
-          local_buffer = UOp(Ops.DEFINE_LOCAL, op.dtype.ptr(local_size, addrspace=AddrSpace.LOCAL), (), f"temp{self.reduceops.index(op)}")
-          local_load = local_buffer.view(st).load(local_buffer.view(st).store(ret))
-          grouped_reduce = UOp(Ops.REDUCE_AXIS, op.dtype, (local_load,), arg=(op.arg[0], grouped_axes))
-          if op is self.reduceops[-1]: return grouped_reduce
-          st = ShapeTracker.from_shape(tuple([1 if i in grouped_axes else s for i,s in enumerate(local_shape)]))
-          return local_buffer.view(st).load(local_buffer.view(st).store(grouped_reduce))
-
       return ret
     self.finalized = True
     fixed_ast = fixup_ast(self.ast)

tinygrad/codegen/opt/swizzler.py

@@ -128,7 +128,8 @@ fix_kernel_ops = view_left_through_load+PatternMatcher([
   (UPat(Ops.VIEW, src=(UPat.cvar(),), name="self"),
    lambda self: UOp.where(UOp(Ops.VALID, dtypes.bool, (UOp(Ops.VIEW, arg=self.st),)), self.const_like(self.base.arg), 0)),
   # no ImageDType after index
-  (UPat(GroupOp.All-{Ops.DEFINE_GLOBAL, Ops.VIEW}, name="x"), lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
+  (UPat(GroupOp.All-{Ops.DEFINE_GLOBAL, Ops.VIEW, Ops.INDEX}, name="x"),
+   lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
   # if this kernel also assigns to the loaded buffer, ensure we can index it correctly
   (UPat(Ops.LOAD, src=(UPat.var("glbl").view(name="view"),)), check_load_st),
 ])

tinygrad/codegen/opt/tc.py

@@ -22,6 +22,15 @@ class TensorCore: # D = A * B + C, A is (M x K), B is (K x N), C and D are (M x
   def permutes_for_shape_str(self, shape_str:list[str]) -> tuple[tuple[int, ...], tuple[int, ...]]:
     ret = [[shape_str.index(remap[ss]) if ss in remap else i for i,ss in enumerate(shape_str)] for remap in self._remaps()]
     return tuple(ret[0]), tuple(ret[1])
+  @functools.cache  # pylint: disable=method-cache-max-size-none
+  def base_shape_str(self) -> list[str]:
+    ret = []
+    cnt = {'u': 0, 'l': 0}
+    for opt in self.opts:
+      ret.append(f"{opt[0]}{cnt[opt[0]]}")
+      cnt[opt[0]] += 1
+    # assumes you do the UNROLL after the opts
+    return ret + [f"r{i}" for i in range(len(self.get_reduce_axes()))]
   def get_reduce_axes(self): return [(i, 2) for i in range(int(math.log2(self.dims[2])))]
   def get_upcast_axes(self): return [opt for opt in self.opts if opt[0] == "u"]
   def get_local_axes(self): return [opt for opt in self.opts if opt[0] == "l"]

tinygrad/dtype.py

@@ -108,7 +108,6 @@ class dtypes:
     if isinstance(val, tuple):
       assert len(val) == dtype.count, f"mismatch {val} {dtype}"
       return tuple(dtypes.as_const(x, dtype) for x in val)
-    # TODO: should truncate here
     return int(val) if dtypes.is_int(dtype) else float(val) if dtypes.is_float(dtype) else bool(val)
   @staticmethod
   @functools.cache
@@ -215,15 +214,14 @@ def sum_acc_dtype(dt:DType):
   return least_upper_dtype(dt, to_dtype(getenv("SUM_DTYPE", "float32")))
 
 def truncate_fp16(x):
-  try: return struct.unpack("@e", struct.pack("@e", float(x)))[0]
+  try: return struct.unpack('e', struct.pack('e', float(x)))[0]
   except OverflowError: return math.copysign(math.inf, x)
 
-def truncate_bf16(x):
-  max_bf16 = struct.unpack('f', struct.pack('I', 0x7f7f0000))[0]
-  if abs(x) > max_bf16: return math.copysign(math.inf, x)
-  f32_int = struct.unpack('I', struct.pack('f', x))[0]
-  bf = struct.unpack('f', struct.pack('I', f32_int & 0xFFFF0000))[0]
-  return bf
+def float_to_bf16(x):
+  if not math.isfinite(x): return x
+  u = struct.unpack('I', struct.pack('f', x))[0]
+  u = (u + 0x7FFF + ((u >> 16) & 1)) & 0xFFFF0000
+  return struct.unpack('f', struct.pack('I', u))[0]
 
 # fp8-float conversions based on https://gitlab.com/nvidia/headers/cuda-individual/cudart/-/blob/main/cuda_fp8.hpp
 def float_to_fp8(x: float, dtype: DType) -> int:
@@ -288,7 +286,7 @@ def fp8_to_float(x: int, dtype: DType) -> float:
   return float(float32_val)
 
 truncate: dict[DType, Callable] = {dtypes.bool: bool,
-  dtypes.float16: truncate_fp16, dtypes.bfloat16: truncate_bf16,
+  dtypes.float16: truncate_fp16, dtypes.bfloat16: lambda x: float_to_bf16(float(x)),
   **{fp8: (lambda x, dtype=fp8: fp8_to_float(float_to_fp8(x, dtype), dtype)) for fp8 in dtypes.fp8s},
   dtypes.float32: lambda x: ctypes.c_float(x).value, dtypes.float64: lambda x: ctypes.c_double(x).value,
   dtypes.uint8: lambda x: ctypes.c_uint8(x).value, dtypes.uint16: lambda x: ctypes.c_uint16(x).value,

tinygrad/helpers.py

@@ -135,7 +135,7 @@ FUSE_ARANGE, FUSE_CONV_BW = ContextVar("FUSE_ARANGE", 1), ContextVar("FUSE_CONV_
 SPLIT_REDUCEOP, NO_MEMORY_PLANNER, RING = ContextVar("SPLIT_REDUCEOP", 1), ContextVar("NO_MEMORY_PLANNER", 0), ContextVar("RING", 1)
 PICKLE_BUFFERS, PROFILE, LRU = ContextVar("PICKLE_BUFFERS", 1), ContextVar("PROFILE", getenv("VIZ")), ContextVar("LRU", 1)
 CACHELEVEL, IGNORE_BEAM_CACHE, DEVECTORIZE = ContextVar("CACHELEVEL", 2), ContextVar("IGNORE_BEAM_CACHE", 0), ContextVar("DEVECTORIZE", 1)
-DISABLE_COMPILER_CACHE = ContextVar("DISABLE_COMPILER_CACHE", 0)
+DISABLE_COMPILER_CACHE, BLOCK_REORDER = ContextVar("DISABLE_COMPILER_CACHE", 0), ContextVar("BLOCK_REORDER", 1)
 DONT_REALIZE_EXPAND, DONT_GROUP_REDUCES = ContextVar("DONT_REALIZE_EXPAND", 0), ContextVar("DONT_GROUP_REDUCES", 0)
 QUANTIZE, VALIDATE_WITH_CPU, DISABLE_FAST_IDIV = ContextVar("QUANTIZE", 0), ContextVar("VALIDATE_WITH_CPU", 0), ContextVar("DISABLE_FAST_IDIV", 0)
 CORRECT_DIVMOD_FOLDING, FUSE_OPTIM = ContextVar("CORRECT_DIVMOD_FOLDING", 0), ContextVar("FUSE_OPTIM", 0)

tinygrad/renderer/cstyle.py

@@ -6,7 +6,7 @@ from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat
 from tinygrad.helpers import strip_parens, getenv, prod, dedup, AMX
 from tinygrad.dtype import ImageDType, dtypes, DType, PtrDType, AddrSpace, truncate
 from tinygrad.renderer import Renderer
-from tinygrad.codegen.devectorizer import no_vectorized_alu
+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}];"),

tinygrad/renderer/ptx.py

@@ -119,7 +119,7 @@ string_rewrite = PatternMatcher([
     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:]};"]),
   (UPat(Ops.DEFINE_LOCAL, name="x"),
-   lambda ctx, x: [f".shared .align 16 .b8 {x.arg}[{x.dtype.size*x.dtype.itemsize}];", f"mov.u64 {ctx.r[x]}, {x.arg}[0];"]),
+   lambda ctx, x: [f".shared .align 16 .b8 local{x.arg}[{x.dtype.size*x.dtype.itemsize}];", f"mov.u64 {ctx.r[x]}, local{x.arg}[0];"]),
   (UPat(Ops.IF, name="x"), lambda ctx, x: f"@!{ctx.r[x.src[0]]} bra IF_{ctx.r[x.src[0]][1:]}_{ctx.uops.index(x)};"),
   (UPat(Ops.ENDIF, name="x"), lambda ctx, x: f"IF_{ctx.r[x.src[0].src[0]][1:]}_{ctx.uops.index(x.src[0])}:"),
   (UPat(Ops.WMMA, name="x"), lambda ctx, x: list(render_wmma(ctx, x))),
@@ -215,7 +215,7 @@ class PTXRenderer(Renderer):
                        [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.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_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/ops_amd.py

@@ -4,10 +4,11 @@ import os, ctypes, ctypes.util, struct, hashlib, functools, importlib, mmap, err
 assert sys.platform != 'win32'
 from dataclasses import dataclass
 from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, HWQueue, CLikeArgsState, HCQSignal, HCQProgram, FileIOInterface
-from tinygrad.runtime.support.hcq import MMIOInterface
+from tinygrad.runtime.support.hcq import MMIOInterface, BumpAllocator
 from tinygrad.uop.ops import sint
 from tinygrad.device import Compiled, DMAFdRef, BufferSpec
 from tinygrad.helpers import getenv, to_mv, round_up, data64_le, all_same, flatten, DEBUG, AMD_LLVM, PROFILE, ProfileEvent, suppress_finalizing
+from tinygrad.helpers import lo32, hi32
 from tinygrad.renderer.cstyle import AMDRenderer
 from tinygrad.renderer.llvmir import AMDLLVMRenderer
 from tinygrad.runtime.autogen import kfd, hsa, pci, sqtt
@@ -24,6 +25,8 @@ EVENT_INDEX_PARTIAL_FLUSH = 4 # based on a comment in nvd.h
 WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
 WAIT_REG_MEM_FUNCTION_GEQ = 5 # >=
+AQL_HDR = (1 << hsa.HSA_PACKET_HEADER_BARRIER) | (hsa.HSA_FENCE_SCOPE_SYSTEM << hsa.HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE) \
+        | (hsa.HSA_FENCE_SCOPE_SYSTEM << hsa.HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE)
 
 class AMDSignal(HCQSignal):
   def __init__(self, *args, **kwargs): super().__init__(*args, **{**kwargs, 'timestamp_divider': 100})
@@ -284,7 +287,7 @@ class AMDComputeQueue(HWQueue):
 
   def wait(self, signal:AMDSignal, value:sint=0):
     self.wait_reg_mem(mem=signal.value_addr, value=value, mask=0xffffffff)
-    if self.dev.xccs > 1: self.xcc_barrier()
+    if self.dev.xccs > 1 and not self.dev.is_aql: self.xcc_barrier()
     return self
 
   def timestamp(self, signal:AMDSignal):
@@ -329,6 +332,41 @@ class AMDComputeQueue(HWQueue):
     dev.compute_queue.put_value += len(cmds)
     dev.compute_queue.signal_doorbell(dev)
 
+class AMDComputeAQLQueue(AMDComputeQueue):
+  def exec(self, prg:AMDProgram, args_state:CLikeArgsState, global_size:tuple[sint, ...], local_size:tuple[sint, ...]):
+    self.bind_args_state(args_state)
+    self._q.append(pkt:=hsa.hsa_kernel_dispatch_packet_t(header=AQL_HDR | (hsa.HSA_PACKET_TYPE_KERNEL_DISPATCH << hsa.HSA_PACKET_HEADER_TYPE),
+      setup=3<<hsa.HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS, private_segment_size=prg.private_segment_size,
+      group_segment_size=prg.group_segment_size, kernel_object=prg.aql_prog_addr, kernarg_address=args_state.buf.va_addr))
+    self.bind_sints_to_mem(*local_size, mem=(pkt_view:=MMIOInterface(addr=ctypes.addressof(pkt), nbytes=ctypes.sizeof(pkt))), fmt='H', offset=4)
+    self.bind_sints_to_mem(*[l * g for l,g in zip(local_size, global_size)], mem=pkt_view, fmt='I', offset=12)
+
+  def bind(self, dev:AMDDevice): pass # not supported
+  def _submit(self, dev:AMDDevice):
+    pm4_batch:list[int] = []
+    aql_bytes = bytes()
+
+    def flush_pm4_batch():
+      nonlocal pm4_batch
+      if not pm4_batch: return bytes()
+      dev.pm4_ibs.cpu_view().view(off:=dev.pm4_ib_alloc.alloc(len(pm4_batch) * 4), fmt='I')[:len(pm4_batch)] = array.array('I', pm4_batch)
+      pkt = [AQL_HDR | (hsa.HSA_PACKET_TYPE_VENDOR_SPECIFIC << hsa.HSA_PACKET_HEADER_TYPE) | (1 << 16),
+        self.pm4.PACKET3(self.pm4.PACKET3_INDIRECT_BUFFER, 2), *data64_le(dev.pm4_ibs.va_addr+off), len(pm4_batch)|self.pm4.INDIRECT_BUFFER_VALID, 10]
+      pm4_batch.clear()
+      return bytes(array.array('I', pkt + [0] * 10))
+
+    for cmd in self._q:
+      if isinstance(cmd, hsa.hsa_kernel_dispatch_packet_t): aql_bytes += flush_pm4_batch() + bytes(cmd)
+      else: pm4_batch.append(cmd)
+    aql_bytes += flush_pm4_batch()
+
+    assert len(aql_bytes) < dev.compute_queue.ring.nbytes, "submit is too large for the queue"
+    cp_bytes = min(len(aql_bytes), (dev.compute_queue.ring.nbytes - (dev.compute_queue.put_value * 64) % dev.compute_queue.ring.nbytes))
+    dev.compute_queue.ring.view(offset=(dev.compute_queue.put_value * 64) % dev.compute_queue.ring.nbytes, fmt='B')[:cp_bytes] = aql_bytes[:cp_bytes]
+    if (tail_bytes:=(len(aql_bytes) - cp_bytes)) > 0: dev.compute_queue.ring.view(offset=0, fmt='B')[:tail_bytes] = aql_bytes[cp_bytes:]
+    dev.compute_queue.put_value += len(aql_bytes) // 64
+    dev.compute_queue.signal_doorbell(dev, doorbell_value=dev.compute_queue.put_value-1)
+
 class AMDCopyQueue(HWQueue):
   def __init__(self, dev, max_copy_size=0x40000000):
     self.dev, self.sdma, self.internal_cmd_sizes, self.max_copy_size = dev, dev.sdma, [], max_copy_size
@@ -427,13 +465,18 @@ class AMDProgram(HCQProgram):
     self.dev, self.name, self.lib = dev, name, lib
 
     image, sections, _ = elf_loader(self.lib)
-    self.lib_gpu = self.dev.allocator.alloc(round_up(image.nbytes, 0x1000), buf_spec:=BufferSpec(cpu_access=True, nolru=True))
-    self.dev.allocator._copyin(self.lib_gpu, image)
-    self.dev.synchronize()
 
     rodata_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".rodata"), -1)
     text_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".text"), -1)
     assert rodata_entry >= 0 and text_entry >= 0, ".text or .rodata section not found"
+
+    # Relo for kernel_code_entry_byte_offset for AMD_LLVM. Comgr doesn't need that, but keep shared code path.
+    image[rodata_entry+0x10:rodata_entry+0x10+8] = struct.pack('<q', text_entry - rodata_entry)
+
+    self.lib_gpu = self.dev.allocator.alloc(round_up(image.nbytes, 0x1000), buf_spec:=BufferSpec(cpu_access=True, nolru=True))
+    self.dev.allocator._copyin(self.lib_gpu, image)
+    self.dev.synchronize()
+
     self.group_segment_size = image[rodata_entry:rodata_entry+4].cast("I")[0]
     self.private_segment_size = image[rodata_entry+4:rodata_entry+8].cast("I")[0]
     self.kernargs_segment_size = image[rodata_entry+8:rodata_entry+12].cast("I")[0]
@@ -451,8 +494,8 @@ class AMDProgram(HCQProgram):
     self.rsrc1: int = code.compute_pgm_rsrc1 | ((1 << 20) if (11,0,0) <= self.dev.target < (12,0,0) else 0)
     self.rsrc2: int = code.compute_pgm_rsrc2 | (lds_size << 15)
     self.rsrc3: int = image[rodata_entry+44:rodata_entry+48].cast("I")[0] # NOTE: kernel descriptor, not in amd_kernel_code_t struct
+    self.aql_prog_addr: int = self.lib_gpu.va_addr + rodata_entry
     self.prog_addr: int = self.lib_gpu.va_addr + rodata_entry + code.kernel_code_entry_byte_offset
-    if code.kernel_code_entry_byte_offset == 0: self.prog_addr = self.lib_gpu.va_addr + text_entry
     # Some programs use hsa_kernel_dispatch_packet_t to read workgroup sizes during execution.
     # The packet is represented as a pointer and set up in SGPRs. Space for the packet is allocated as part of the kernel arguments.
     self.enable_dispatch_ptr: int = code.kernel_code_properties & hsa.AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_DISPATCH_PTR
@@ -501,7 +544,7 @@ class AMDQueueDesc:
     return cls(ring=queues[0].ring, put_value=queues[0].put_value, doorbells=flatten(q.doorbells for q in queues),
                read_ptrs=flatten(q.read_ptrs for q in queues), write_ptrs=flatten(q.write_ptrs for q in queues))
 
-  def signal_doorbell(self, dev):
+  def signal_doorbell(self, dev, doorbell_value:int|None=None):
     for write_ptr in self.write_ptrs: write_ptr[0] = self.put_value
 
     # Ensure all prior writes are visible to the GPU.
@@ -509,7 +552,7 @@ class AMDQueueDesc:
 
     # Flush hdp if queue is in dev mem.
     if dev.is_am() and not dev.is_usb(): dev.iface.dev_impl.gmc.flush_hdp()
-    for doorbell in self.doorbells: doorbell[0] = self.put_value
+    for doorbell in self.doorbells: doorbell[0] = self.put_value if doorbell_value is None else doorbell_value
 
 class KFDIface:
   kfd:FileIOInterface|None = None
@@ -612,12 +655,12 @@ class KFDIface:
     stm = kfd.AMDKFD_IOC_MAP_MEMORY_TO_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(c_gpus), n_devices=1)
     assert stm.n_success == 1
 
-  def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
+  def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
     queue = kfd.AMDKFD_IOC_CREATE_QUEUE(KFDIface.kfd, ring_base_address=ring.va_addr, ring_size=ring.size, gpu_id=self.gpu_id,
       queue_type=queue_type, queue_percentage=kfd.KFD_MAX_QUEUE_PERCENTAGE|(xcc_id<<8), queue_priority=kfd.KFD_MAX_QUEUE_PRIORITY,
       eop_buffer_address=eop_buffer.va_addr if eop_buffer else 0, eop_buffer_size=eop_buffer.size if eop_buffer else 0, ctl_stack_size=ctl_stack_size,
       ctx_save_restore_address=cwsr_buffer.va_addr if cwsr_buffer else 0, ctx_save_restore_size=ctx_save_restore_size,
-      write_pointer_address=gart.va_addr, read_pointer_address=gart.va_addr + 8 * (xcc_id + 1))
+      write_pointer_address=gart.va_addr+wptr, read_pointer_address=gart.va_addr+rptr+8*xcc_id)
 
     if not hasattr(self, 'doorbells'):
       self.doorbells_base = queue.doorbell_offset & (~0x1fff) # doorbell is two pages
@@ -662,18 +705,19 @@ class PCIIface(PCIIfaceBase):
       'max_slots_scratch_cu': self.dev_impl.gc_info.gc_max_scratch_slots_per_cu, 'max_waves_per_simd': self.dev_impl.gc_info.gc_max_waves_per_simd,
       'simd_arrays_per_engine': self.dev_impl.gc_info.gc_num_sa_per_se, 'lds_size_in_kb': self.dev_impl.gc_info.gc_lds_size}
 
-  def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
+  def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
     assert cwsr_buffer is None, "no cwsr buffer for am"
+    assert queue_type != kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL, "no AQL queues for am"
 
     if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA:
-      self.dev_impl.sdma.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr, wptr_addr=gart.va_addr+0x10,
+      self.dev_impl.sdma.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr+rptr, wptr_addr=gart.va_addr+wptr,
                                     doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_sDMA_ENGINE0), pipe=0, queue=0)
     else:
-      self.dev_impl.gfx.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr, wptr_addr=gart.va_addr+0x10,
+      self.dev_impl.gfx.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr+rptr, wptr_addr=gart.va_addr+wptr,
         eop_addr=eop_buffer.va_addr, eop_size=eop_buffer.size, doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_MEC_RING0), pipe=0, queue=0)
 
     return AMDQueueDesc(ring=ring.cpu_view().view(fmt='I'), doorbells=[self.dev_impl.doorbell64.view(doorbell_index * 8, 8, fmt='Q')],
-      read_ptrs=[gart.cpu_view().view(size=8, fmt='Q')], write_ptrs=[gart.cpu_view().view(offset=0x10, size=8, fmt='Q')])
+      read_ptrs=[gart.cpu_view().view(offset=rptr, size=8, fmt='Q')], write_ptrs=[gart.cpu_view().view(offset=wptr, size=8, fmt='Q')])
 
   def sleep(self, timeout):
     if self.pci_dev.irq_poller is not None and (events_cnt:=len(self.pci_dev.irq_poller.poll(timeout))):
@@ -715,9 +759,9 @@ class USBIface(PCIIface):
     return HCQBuffer(am_mapping.va_addr, size, meta=PCIAllocationMeta(am_mapping, has_cpu_mapping=False),
       view=USBMMIOInterface(self.usb, self.bars[0][0] + am_mapping.paddrs[0][0], size, fmt='B') if cpu_access else None, owner=self.dev)
 
-  def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
+  def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
     if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE: self.usb._pci_cacheable += [(ring.cpu_view().addr, ring.size)]
-    return super().create_queue(queue_type, ring, gart, eop_buffer, cwsr_buffer, ctl_stack_size, ctx_save_restore_size, xcc_id)
+    return super().create_queue(queue_type, ring, gart, rptr, wptr, eop_buffer, cwsr_buffer, ctl_stack_size, ctx_save_restore_size, xcc_id)
 
   def sleep(self, timeout): pass
 
@@ -763,7 +807,13 @@ class AMDDevice(HCQCompiled):
     nbio_pad = (0,) if self.target[0] == 9 else ()
     self.nbio = AMDIP(nbio_name, self.iface.ip_versions[am.NBIF_HWIP], {i:nbio_pad+x for i,x in self.iface.ip_offsets[am.NBIF_HWIP].items()})
 
-    self.compute_queue = self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE, 0x2000 if self.is_usb() else (16 << 20), eop_buffer_size=0x1000,
+    self.is_aql = getenv("AMD_AQL", self.xccs > 1)
+    if self.is_aql:
+      self.pm4_ibs = self.iface.alloc(0x2000 if self.is_usb() else (16 << 20), uncached=True, cpu_access=True)
+      self.pm4_ib_alloc = BumpAllocator(self.pm4_ibs.size, wrap=True)
+
+    self.compute_queue = self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL if self.is_aql else kfd.KFD_IOC_QUEUE_TYPE_COMPUTE,
+      0x2000 if self.is_usb() else (16 << 20), eop_buffer_size=0x1000,
       ctx_save_restore_size=0 if self.is_am() else wg_data_size + ctl_stack_size, ctl_stack_size=ctl_stack_size, debug_memory_size=debug_memory_size)
 
     max_copy_size = 0x40000000 if self.iface.ip_versions[am.SDMA0_HWIP][0] >= 5 else 0x400000
@@ -771,7 +821,8 @@ class AMDDevice(HCQCompiled):
 
     super().__init__(device, AMDAllocator(self), AMDLLVMRenderer(self.arch) if AMD_LLVM else AMDRenderer(self.arch),
                      AMDLLVMCompiler(self.arch) if AMD_LLVM else HIPCompiler(self.arch), functools.partial(AMDProgram, self),
-                     AMDSignal, functools.partial(AMDComputeQueue, self), functools.partial(AMDCopyQueue, self, max_copy_size=max_copy_size),
+                     AMDSignal, functools.partial(AMDComputeAQLQueue if self.is_aql else AMDComputeQueue, self),
+                     functools.partial(AMDCopyQueue, self, max_copy_size=max_copy_size),
                      kernargs_size=(8 << 10) if self.is_usb() else (16 << 20), sigalloc_size=0x100 if self.is_usb() else 0x1000)
 
     # Scratch setup
@@ -780,10 +831,10 @@ class AMDDevice(HCQCompiled):
 
     # XCC setup
     self.xcc_sync: tuple[AMDSignal, AMDSignal]|None = None
-    if self.xccs > 1:
+    if self.xccs > 1 and not self.is_aql:
       self.xcc_sync_area = self.allocator.alloc(0x1000, BufferSpec(nolru=True, cpu_access=True))
       self.xcc_sync = (AMDSignal(base_buf=self.xcc_sync_area), AMDSignal(base_buf=self.xcc_sync_area.offset(256)))
-      AMDComputeQueue(self).xcc_config().submit(self)
+      cast(AMDComputeQueue, self.hw_compute_queue_t()).xcc_config().submit(self)
 
     # SQTT is disabled by default because of runtime overhead and big file sizes (~200mb to Tensor.full() two 4096x4096 tensors and matmul them)
     self.sqtt_enabled = PROFILE and bool(getenv("SQTT", 0))
@@ -798,18 +849,26 @@ class AMDDevice(HCQCompiled):
       self.sqtt_buffers = [self.allocator.alloc(SQTT_BUFFER_SIZE*1024*1024, BufferSpec(cpu_access=True, nolru=True)) for _ in range(SQTT_NUM)]
       self.sqtt_itrace_se_mask = getenv("SQTT_ITRACE_SE_MASK", 2) # -1 enable all, 0 disable all, >0 bitmask for where to enable instruction tracing
       self.cmd_id = 0
-      AMDComputeQueue(self).sqtt_start(self.sqtt_buffers, self.sqtt_itrace_se_mask).submit(self)
+      cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_start(self.sqtt_buffers, self.sqtt_itrace_se_mask).submit(self)
 
   def create_queue(self, queue_type, ring_size, ctx_save_restore_size=0, eop_buffer_size=0, ctl_stack_size=0, debug_memory_size=0):
     ring = self.iface.alloc(ring_size, uncached=True, cpu_access=True)
     gart = self.iface.alloc(0x100, uncached=True, cpu_access=True)
 
+    if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL:
+      aql_desc = hsa.amd_queue_t(queue_properties=hsa.AMD_QUEUE_PROPERTIES_IS_PTR64 | hsa.AMD_QUEUE_PROPERTIES_ENABLE_PROFILING,
+        read_dispatch_id_field_base_byte_offset=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
+        max_cu_id=self.max_cu_id, max_wave_id=self.max_wave_id)
+      gart.cpu_view().view(fmt='B')[:ctypes.sizeof(aql_desc)] = bytes(aql_desc)
+      self.aql_desc = hsa.amd_queue_t.from_address(gart.va_addr)
+
     cwsr_buffer_size = round_up((ctx_save_restore_size + debug_memory_size) * self.iface.props.get('num_xcc', 1), mmap.PAGESIZE)
     cwsr_buffer = self.iface.alloc(cwsr_buffer_size) if ctx_save_restore_size else None
     eop_buffer = self.iface.alloc(eop_buffer_size) if eop_buffer_size else None
 
-    return AMDQueueDesc.multi(*(self.iface.create_queue(queue_type, ring, gart, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer, xcc_id=xcc_id,
-                                                            ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size)
+    return AMDQueueDesc.multi(*(self.iface.create_queue(queue_type, ring, gart, rptr=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
+                                wptr=getattr(hsa.amd_queue_t, 'write_dispatch_id').offset, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer,
+                                xcc_id=xcc_id, ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size)
                                 for xcc_id in range(self.xccs if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE else 1)))
 
   def _ensure_has_local_memory(self, required):
@@ -828,8 +887,16 @@ class AMDDevice(HCQCompiled):
       self.tmpring_size = waves << 12 | wavesize
       self.max_private_segment_size = required
 
+      if hasattr(self, 'aql_desc'):
+        self.aql_desc.scratch_backing_memory_location = self.scratch.va_addr
+        self.aql_desc.scratch_backing_memory_byte_size = self.scratch.size
+        self.aql_desc.scratch_wave64_lane_byte_size = self.max_private_segment_size * (self.aql_desc.max_wave_id + 1) // 64
+        self.aql_desc.scratch_resource_descriptor[:] = [lo32(self.scratch.va_addr), hi32(self.scratch.va_addr) | (1 << 30), lo32(self.scratch.size),
+          0x20814fac] # FORMAT=BUF_FORMAT_32_UINT,OOB_SELECT=2,ADD_TID_ENABLE=1,TYPE=SQ_RSRC_BUF,SQ_SELs
+        self.aql_desc.compute_tmpring_size = self.tmpring_size
+
   def invalidate_caches(self):
-    AMDComputeQueue(self).memory_barrier().signal(self.timeline_signal, self.next_timeline()).submit(self)
+    self.hw_compute_queue_t().memory_barrier().signal(self.timeline_signal, self.next_timeline()).submit(self)
     self.synchronize()
 
   def on_device_hang(self): self.iface.on_device_hang()
@@ -838,7 +905,8 @@ class AMDDevice(HCQCompiled):
     if self.sqtt_enabled:
       wptrs_buf = self.allocator.alloc(round_up(len(self.sqtt_buffers), 0x1000), BufferSpec(cpu_access=True, nolru=True))
       wptrs = to_mv(wptrs_buf.va_addr, wptrs_buf.size)
-      AMDComputeQueue(self).sqtt_stop(len(self.sqtt_buffers), wptrs_buf).signal(self.timeline_signal, self.next_timeline()).submit(self)
+      cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_stop(len(self.sqtt_buffers), wptrs_buf) \
+                                                      .signal(self.timeline_signal, self.next_timeline()).submit(self)
       self.synchronize()
       if DEBUG>=2: print('Saving SQTT in profile...')
       for i,buf0 in enumerate(self.sqtt_buffers):

tinygrad/schedule/rangeify.py

@@ -1,12 +1,12 @@
 from typing import Any
 from dataclasses import dataclass, field
-from tinygrad.dtype import dtypes, PtrDType
-from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, RewriteNotReady, _substitute, AxisType
+from tinygrad.dtype import dtypes, PtrDType, ImageDType, AddrSpace
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, RewriteNotReady, _substitute
 from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, colored, RANGEIFY
 from tinygrad.schedule.multi import multi_pm
 
 from tinygrad.schedule.kernelize import Kernel
-from tinygrad.uop.ops import track_rewrites, graph_rewrite_map, graph_rewrite, KernelInfo, identity_element, sint
+from tinygrad.uop.ops import track_rewrites, graph_rewrite_map, graph_rewrite, KernelInfo, identity_element, sint, AxisType
 
 # 0. do some cleanup rewrites, mostly copied from the old stuff
 
@@ -332,13 +332,15 @@ pm_cleanups = double_reshape+pm_mops+PatternMatcher([
 def bufferize_to_store(x:UOp):
   rngs = x.src[1:]
   shape = tuple([int(r.vmax+1) for r in rngs])
-  sdtype = x.dtype.ptr(size=prod(shape))
-  assert prod(shape) > 0, f"no zero sized buffers {shape}"
+  size = prod(shape)
+  assert size > 0, f"no zero sized buffers {shape}"
+  sdtype = x.dtype.ptr(size=size, addrspace=AddrSpace.GLOBAL if not isinstance(x.arg, tuple) else x.arg[0])
   if x.src[0].op is Ops.ASSIGN:
     assign_target, assign_src = x.src[0].src
     assert assign_target.op is Ops.INDEX
     return assign_target.replace(dtype=sdtype).store(assign_src, *rngs, dtype=sdtype)
-  buf = UOp.new_buffer(x.arg, prod(shape), x.dtype)
+  if sdtype.addrspace == AddrSpace.GLOBAL: buf = UOp.new_buffer(x.arg, size, x.dtype)
+  else: buf = UOp(Ops.DEFINE_LOCAL, sdtype, arg=x.arg[1])
   return buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0], *rngs, dtype=sdtype).forced_reshape(shape, dtype=x.dtype)
 
 pm_add_buffers = pm_mops+PatternMatcher([
@@ -380,28 +382,34 @@ to_define_global = PatternMatcher([
   (UPat(Ops.BIND, name="b"), unbind_kernel),
   (UPat((Ops.ASSIGN, Ops.MSTACK, Ops.MSELECT), name="assign"), handle_assign),
 
-  # add loads to non ptr indexes
-  # TODO: this can be moved into codegen?
-  (UPat((Ops.DEFINE_GLOBAL, Ops.STORE), name="dg").f(Ops.INDEX, name="idx", allow_any_len=True),
-   lambda dg,idx: idx.replace(dtype=dg.dtype, arg=None).load() if not isinstance(idx.dtype, PtrDType) else None),
-
-  # 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)),
-
   # HACK in case any CONSTs were replaced
   # this is only needed if you are using symbolic
   #(UPat(Ops.CONST, name="c"), lambda c: c.replace(src=()) if len(c.src) else None),
 ])
 
+rangeify_codegen = PatternMatcher([
+  # add loads to non ptr indexes
+  # TODO: this can be moved into codegen?
+  (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 split_store(x:UOp):
   if len(x.ranges): return None
   ctx = LocalAddBufferContext()
-  ret = graph_rewrite(x, to_define_global, ctx=ctx, name="kernel split", bottom_up=True)
+  ret = graph_rewrite(x, to_define_global+rangeify_codegen, ctx=ctx, name="kernel split", bottom_up=True)
 
-  store_rngs = ret.src[2:]
+  # get name
   rng = sorted([u for u in ret.toposort() if u.op is Ops.RANGE], key=lambda x: x.arg)
-  name = "k"+colored('_', 'BLACK').join(['']+[colored(s.src[0].render(), "WHITE" if s in store_rngs else "red") for s in rng])
+  name = "k"+colored('_', 'BLACK').join(['']+[colored(s.src[0].render(), "WHITE" if s in ret.src[2:] else "red") for s in rng])
 
   # NOTE: the hack for COPY is here
   ret = ret.sink(arg=KernelInfo(name=name)) if ret.src[1].op is not Ops.COPY else ret.src[1]

tinygrad/tensor.py

@@ -68,7 +68,7 @@ def _frompy(x:list|tuple|bytes, dtype:DType) -> UOp:
     ret = UOp.new_buffer("PYTHON", prod(shape:=get_shape(x)), dtype).reshape(shape)
     assert dtype.fmt is not None, f"{dtype=} has None fmt"
     truncate_function = truncate[dtype]
-    data = struct.pack(f"@{ret.size}{dtype.fmt}", *[truncate_function(xi) for xi in fully_flatten(x)])
+    data = struct.pack(f"{ret.size}{dtype.fmt}", *[truncate_function(dtypes.as_const(xi, dtype)) for xi in fully_flatten(x)])
   # fake realize
   ret.buffer.allocate(memoryview(data if Device.DEFAULT != "PYTHON" else bytearray(data)))
   return ret

tinygrad/uop/decompositions.py

@@ -280,7 +280,7 @@ def magicgu(vmax:int, d:int) -> tuple[int,int]:
       return m, s
   assert False
 
-def fast_idiv(device: str, x: UOp, d: int) -> UOp|None:
+def fast_idiv(device: str, x: UOp, d: int, dont_cast=False) -> UOp|None:
   # If d is a power of two this is not valid for signed ints!
   is_unsigned = True if x.vmin>=0 or x.dtype in dtypes.uints else False
   assert d>0, "Sign should have been taken out of divisor"
@@ -288,6 +288,10 @@ def fast_idiv(device: str, x: UOp, d: int) -> UOp|None:
   m,s = magicgu(max(vmax, abs(vmin)), d)
   if m*vmin >= dtypes.min(x.dtype) and m*vmax <= dtypes.max(x.dtype):
     return ((x*m) >> s) if is_unsigned else ((x*m) >> s) + (x<0).where(x.ufix(1), 0)
+  # before we try casting to a larger dtype (slow), we see if there are powers of two in d we can shift to make x smaller
+  if (largest_factor_of_two_in_d := (d & -d)) > 1:
+    if (ret:=fast_idiv(device, x//largest_factor_of_two_in_d, d//largest_factor_of_two_in_d, dont_cast=True)) is not None: return ret
+  if dont_cast: return None
   # promo_lattice needs to return an unsigned type if the type is unsigned
   if dtypes.is_int(next_dtype := promo_lattice[x.dtype][-1]) and is_dtype_supported(next_dtype, None if device=='' else device):
     if m*vmin >= dtypes.min(next_dtype) and m*vmax <= dtypes.max(next_dtype):
@@ -329,6 +333,8 @@ def get_late_rewrite_patterns(ops:tuple[Ops, ...], force_transcendental=False):
   if Ops.SQRT not in ops: pat.append((UPat(Ops.SQRT, src=UPat.var("d")), lambda d: xpow(d, d.const_like(0.5))))
   # rewrite MOD to AND (which should always be supported, but not for generic in tests): x % (2**y) -> x & (2**y-1)
   if Ops.AND in ops: pat += [(UPat.var("x", dtypes.ints)%UPat.cvar("c"), lambda x,c: x & (c.arg-1) if c.arg in powers_of_two else None)]
+  if Ops.OR in ops: pat += [(UPat.var("x", dtypes.bool).logical_not()&UPat.var("y", dtypes.bool).logical_not(),
+    lambda x,y: (x | y).logical_not())]
   # rewrite MUL/IDIV to SHL+SHR: x*(2**y) -> shl(x,y) and x//(2**y) -> shr(x,y)
   if Ops.SHL in ops: pat += [(UPat.var("x", dtypes.ints)*UPat.cvar("c"), lambda c,x: x << v if (v:=powers_of_two.get(c.arg, 0)) else None)]
   if Ops.SHR in ops:

tinygrad/uop/ops.py

@@ -681,7 +681,8 @@ class UPat(MathTrait):
   def var(name:str|None=None, dtype:DType|tuple[DType, ...]|None=None): return UPat(dtype=dtype, name=name)
   @staticmethod
   @functools.cache
-  def cvar(name:str|None=None, dtype:DType|None=None, vec=True): return UPat((Ops.CONST,Ops.VCONST) if vec else Ops.CONST, dtype, name=name)
+  def cvar(name:str|None=None, dtype:DType|tuple[DType, ...]|None=None, vec=True):
+    return UPat((Ops.CONST,Ops.VCONST) if vec else Ops.CONST, dtype, name=name)
   @staticmethod
   def const(dtype:DType|tuple[DType, ...]|None, b:ConstType): return UPat(Ops.CONST, dtype=dtype, arg=b)
 

tinygrad/uop/symbolic.py

@@ -1,5 +1,5 @@
 # all of symbolic lives here now
-from typing import Any, cast
+from typing import cast
 import math, operator, struct, functools
 from collections import defaultdict
 from tinygrad.uop.ops import Ops, PatternMatcher, UPat, UOp, GroupOp, exec_alu
@@ -19,7 +19,7 @@ def simplify_pow(x:UOp, c:UOp) -> UOp|None:
 def fold_bitcast(root:UOp, c:UOp) -> UOp|None:
   if (from_fmt:=c.dtype.scalar().fmt) is None or (to_fmt:=root.dtype.scalar().fmt) is None: return None
   if c.dtype.itemsize != root.dtype.itemsize: return None
-  def convert(v:Any): return struct.unpack(to_fmt, struct.pack(from_fmt, v))[0]
+  def convert(v:ConstType): return struct.unpack(to_fmt, struct.pack(from_fmt, v))[0]
   return root.const_like(convert(c.arg) if root.dtype.count == 1 else tuple(map(convert, c.arg)))
 
 symbolic_simple = PatternMatcher([
@@ -183,10 +183,9 @@ def fold_divmod_congruence(d: UOp, x: UOp, y: UOp) -> UOp|None:
   terms, factors = zip(*[(u.divides(f:=u.const_factor()),f) for u in split_uop(x, Ops.ADD)])
   # a//c = (a-a%c)/c, if we can fold a%c, we can fold a//c
   rems = [min((r:=f%c), r-c, key=abs) for f in factors]
-  if (rem:=sum(r*v for r,v in zip(rems,terms))+const%c).vmin//c==rem.vmax//c and all(f > 0 for f in factors):
-    if d.op is Ops.MOD: return rem - rem.vmin//c*c
-    return sum((f-r)//c * v for f,r,v in zip(factors,rems,terms)) + (const-const%c+rem.vmin//c*c)//c
-  return None
+  if (rem:=sum(r*v for r,v in zip(rems,terms))+const%c).vmin//c!=rem.vmax//c: return None
+  if d.op is Ops.MOD: return rem - rem.vmin//c*c
+  return sum((f-r)//c * v for f,r,v in zip(factors,rems,terms)) + (const-const%c+rem.vmin//c*c)//c
 
 def divide_by_gcd(d: UOp, x: UOp, y: UOp) -> UOp|None:
   # x//y -> (x//gcd)//(y//gcd) or x%y -> gcd*(x//gcd)%(y//gcd)
@@ -280,6 +279,7 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
   ((UPat.var("y") + UPat.var("x") * UPat.cvar("c0")) + UPat.var("x") * UPat.cvar("c1"), lambda x,y,c0,c1: y+x*(c0+c1)),
   (UPat.var("x") + UPat.var("x") * UPat.cvar("c"), lambda x,c: x*(c+1)), # (x+x*c)-> x*(c+1)
   ((UPat.var("y") + UPat.var("x")) + UPat.var("x") * UPat.cvar("c"), lambda x,y,c: y+x*(c+1)),
+  ((UPat.var("y") + UPat.var("x") * UPat.cvar("c")) + UPat.var("x"), lambda x,y,c: y+x*(c+1)),
   (UPat.var("x") + UPat.var("x"), lambda x: x*2), # (x+x)-> x*2
   ((UPat.var("y") + UPat.var("x")) + UPat.var("x"), lambda y,x: y+x*2),
   ((UPat.var("x") / UPat.var("x2")) / UPat.var("x3"), lambda x,x2,x3: x/(x2*x3) if x2 is not x3 else None), # (x/x2)/x3 -> x/(x2*x3)
@@ -291,6 +291,9 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
   # alu of two where with same conds can combine, only do if true branch or false branch is const
   (UPat(GroupOp.Binary, name="alu", src=(UPat.var("c").where(UPat.var("t"), UPat.var("f")), UPat.var("c").where(UPat.var("tt"), UPat.var("ff")))), \
    lambda alu,c,t,tt,f,ff: c.where(t.alu(alu.op, tt), f.alu(alu.op, ff)) if t.op == tt.op == Ops.CONST or f.op == ff.op == Ops.CONST else None),
+  # if its a plus we add the associative variation too
+  ((UPat.var("y")+UPat.var("c").where(UPat.var("t"), UPat.var("f"))) + UPat.var("c").where(UPat.var("tt"), UPat.var("ff")), \
+   lambda y,c,t,tt,f,ff: y+c.where(t+tt, f+ff) if t.op == tt.op == Ops.CONST or f.op == ff.op == Ops.CONST else None),
   # ALU/variable min==max -> CONST (slow!)
   (UPat(GroupOp.ALU|{Ops.DEFINE_VAR, Ops.SPECIAL, Ops.RANGE}, name="x"), lambda x: x.const_like(x.vmin) if x.vmin == x.vmax else None),
   # max folding

tinygrad/viz/index.html

@@ -75,9 +75,14 @@
   g.tag circle {
     fill: #FFD700;
     stroke: #B8860B;
+  }
+  g.port circle {
+    fill: #b3dcc2;
+  }
+  g.tag circle, #edge-labels circle {
     stroke-width: 0.8;
   }
-  g.tag text {
+  g.tag text, #edge-labels text {
     text-anchor: middle;
     font-size: 6px;
     fill: #08090e;
@@ -85,11 +90,30 @@
   .label :is(text, p) {
     font-weight: 350;
   }
+  rect.node {
+    stroke-width: 1.4;
+    stroke: #4a4b57;
+  }
+  rect.overlay {
+    fill: rgba(26, 27, 38, 0.5);
+  }
   .edgePath {
     stroke: #4a4b57;
     fill: none;
     stroke-width: 1.4px;
   }
+  .highlight rect, .edgePath.highlight, g.port circle {
+    stroke: #89C9A2;
+  }
+  #edge-labels g.port.highlight {
+    display: block
+  }
+  #edge-labels g.port {
+    display: none
+  }
+  #arrowhead {
+    fill: #4a4b57;
+  }
   .main-container {
     display: flex;
     width: 100%;
@@ -331,7 +355,7 @@
         </g>
         <defs>
           <marker id="arrowhead" viewBox="0 -5 10 10" refX="10" refY="0" markerWidth="6" markerHeight="6" orient="auto">
-            <path d="M0,-5L10,0L0,5" fill="#4a4b57"></path>
+            <path d="M0,-5L10,0L0,5" fill="context-stroke"></path>
           </marker>
         </defs>
       </svg>

tinygrad/viz/js/index.js

@@ -56,11 +56,23 @@ async function renderDag(graph, additions, recenter=false) {
     const g = dagre.graphlib.json.read(e.data);
     // draw nodes
     const STROKE_WIDTH = 1.4;
+    d3.select("#graph-svg").on("click", () => d3.selectAll(".highlight").classed("highlight", false));
     const nodes = d3.select("#nodes").selectAll("g").data(g.nodes().map(id => g.node(id)), d => d).join("g")
-      .attr("transform", d => `translate(${d.x},${d.y})`).classed("clickable", d => d.ref != null)
-      .on("click", (_,d) => setCtxWithHistory(d.ref));
+      .attr("transform", d => `translate(${d.x},${d.y})`).classed("clickable", d => d.ref != null).on("click", (e,d) => {
+        if (d.ref != null) return setCtxWithHistory(d.ref);
+        const parents = g.predecessors(d.id);
+        if (parents == null) return;
+        const src = [...parents, d.id];
+        nodes.classed("highlight", n => src.includes(n.id));
+        d3.select("#edges").selectAll("path.edgePath").classed("highlight", e => src.includes(e.v) && e.w===d.id);
+        d3.select("#edge-labels").selectAll("g.port").classed("highlight", (_, i, nodes) => {
+          const [v, w] = nodes[i].id.split("-");
+          return src.includes(v) && w===d.id;
+        });
+        e.stopPropagation();
+      });
     nodes.selectAll("rect").data(d => [d]).join("rect").attr("width", d => d.width).attr("height", d => d.height).attr("fill", d => d.color)
-      .attr("x", d => -d.width/2).attr("y", d => -d.height/2).attr("style", d => d.style ?? `stroke:#4a4b57; stroke-width:${STROKE_WIDTH}px;`);
+      .attr("x", d => -d.width/2).attr("y", d => -d.height/2).attr("class", d => d.className ?? "node");
     nodes.selectAll("g.label").data(d => [d]).join("g").attr("class", "label").attr("transform", d => {
       const x = (d.width-d.padding*2)/2;
       const y = (d.height-d.padding*2)/2+STROKE_WIDTH;
@@ -79,15 +91,15 @@ async function renderDag(graph, additions, recenter=false) {
     addTags(nodes.selectAll("g.tag").data(d => d.tag != null ? [d] : []).join("g").attr("class", "tag")
       .attr("transform", d => `translate(${-d.width/2+8}, ${-d.height/2+8})`).datum(e => e.tag));
     // draw edges
-    const line = d3.line().x(d => d.x).y(d => d.y).curve(d3.curveBasis);
-    d3.select("#edges").selectAll("path.edgePath").data(g.edges()).join("path").attr("class", "edgePath").attr("d", (e) => {
+    const line = d3.line().x(d => d.x).y(d => d.y).curve(d3.curveBasis), edges = g.edges();
+    d3.select("#edges").selectAll("path.edgePath").data(edges).join("path").attr("class", "edgePath").attr("d", (e) => {
       const edge = g.edge(e);
       const points = edge.points.slice(1, edge.points.length-1);
       points.unshift(intersectRect(g.node(e.v), points[0]));
       points.push(intersectRect(g.node(e.w), points[points.length-1]));
       return line(points);
     }).attr("marker-end", "url(#arrowhead)");
-    addTags(d3.select("#edge-labels").selectAll("g").data(g.edges().filter(e => g.edge(e).label != null)).join("g").attr("transform", (e) => {
+    addTags(d3.select("#edge-labels").selectAll("g").data(edges).join("g").attr("transform", (e) => {
       // get a point near the end
       const [p1, p2] = g.edge(e).points.slice(-2);
       const dx = p2.x-p1.x;
@@ -101,7 +113,7 @@ async function renderDag(graph, additions, recenter=false) {
       const x = p2.x - ux * offset;
       const y = p2.y - uy * offset;
       return `translate(${x}, ${y})`
-    }).attr("class", "tag").datum(e => g.edge(e).label));
+    }).attr("class", e => g.edge(e).label.type).attr("id", e => `${e.v}-${e.w}`).datum(e => g.edge(e).label.text));
     if (recenter) document.getElementById("zoom-to-fit-btn").click();
   };
 
@@ -161,7 +173,7 @@ async function renderProfiler() {
   const optional = (i) => i === 0 ? null : i-1;
   const dur = u32(), peak = u64(), indexLen = u32(), layoutsLen = u32();
   const textDecoder = new TextDecoder("utf-8");
-  const { strings, dtypes }  = JSON.parse(textDecoder.decode(new Uint8Array(buf, offset, indexLen))); offset += indexLen;
+  const { strings, dtypeSize }  = JSON.parse(textDecoder.decode(new Uint8Array(buf, offset, indexLen))); offset += indexLen;
   // place devices on the y axis and set vertical positions
   const [tickSize, padding] = [10, 8];
   const deviceList = profiler.append("div").attr("id", "device-list").style("padding-top", tickSize+padding+"px");
@@ -180,19 +192,26 @@ async function renderProfiler() {
     const div = deviceList.append("div").attr("id", k).text(k).style("padding", padding+"px");
     const { y:baseY, height:baseHeight } = rect(div.node());
     const offsetY = baseY-canvasTop+padding/2;
+    const shapes = [];
     const EventTypes = {TIMELINE:0, MEMORY:1};
     const eventType = u8(), eventsLen = u32();
     if (eventType === EventTypes.TIMELINE) {
-      const maxDepth = u8();
       const levelHeight = baseHeight-padding;
-      const shapes = [];
+      const levels = [];
       data.tracks.set(k, { shapes, offsetY });
       let colorKey, ref;
       for (let j=0; j<eventsLen; j++) {
-        const e = {name:strings[u32()], ref:optional(u32()), st:u32(), dur:f32(), depth:u8(), cat:optional(u8()), info:strings[u32()] || null};
-        if (e.depth === 0) colorKey = e.cat ?? e.name;
+        const e = {name:strings[u32()], ref:optional(u32()), st:u32(), dur:f32(), cat:optional(u8()), info:strings[u32()] || null};
+        // find a free level to put the event
+        let depth = levels.findIndex(levelEt => e.st >= levelEt);
+        const et = e.st+Math.trunc(e.dur);
+        if (depth === -1) {
+          depth = levels.length;
+          levels.push(et);
+        } else levels[depth] = et;
+        if (depth === 0) colorKey = e.cat ?? e.name;
         if (!colorMap.has(colorKey)) colorMap.set(colorKey, cycleColors(colorScheme[k] ?? colorScheme.DEFAULT, colorMap.size));
-        const fillColor = d3.color(colorMap.get(colorKey)).brighter(e.depth).toString();
+        const fillColor = d3.color(colorMap.get(colorKey)).brighter(depth).toString();
         const label = parseColors(e.name).map(({ color, st }) => ({ color, st, width:ctx.measureText(st).width }));
         if (e.ref != null) ref = {ctx:e.ref, step:0};
         else if (ref != null) {
@@ -202,22 +221,47 @@ async function renderProfiler() {
         }
         const arg = { tooltipText:formatTime(e.dur)+(e.info != null ? "\n"+e.info : ""), ...ref };
         // offset y by depth
-        shapes.push({x:e.st, y:levelHeight*e.depth, width:e.dur, height:levelHeight, arg, label, fillColor });
+        shapes.push({x:e.st, y:levelHeight*depth, width:e.dur, height:levelHeight, arg, label, fillColor });
       }
-      div.style("height", levelHeight*maxDepth+padding+"px").style("pointerEvents", "none");
+      div.style("height", levelHeight*levels.length+padding+"px").style("pointerEvents", "none");
     } else {
       const peak = u64();
       const height = heightScale(peak);
       const yscale = d3.scaleLinear().domain([0, peak]).range([height, 0]);
-      const timestamps = Array.from({length:u32()}, u32);
-      const shapes = [];
+      let x = 0, y = 0;
+      const buf_shapes = new Map(), temp = new Map();
+      const timestamps = [];
       for (let j=0; j<eventsLen; j++) {
-        const length = u32();
-        const x = Array.from({ length }, () => timestamps[u32()]);
-        const e = {y:Array.from({ length }, u64), arg:{dtype:strings[u32()], sz:u64()}};
-        const nbytes = dtypes[e.arg.dtype]*e.arg.sz;
-        const arg = {tooltipText:`${e.arg.dtype} len:${formatUnit(e.arg.sz)}\n${formatUnit(e.arg.nbytes, "B")}`};
-        shapes.push({ x, y0:e.y.map(yscale), y1:e.y.map(y => yscale(y+nbytes)), arg, fillColor:cycleColors(colorScheme.BUFFER, j) });
+        const alloc = u8(), ts = u32(), key = u32();
+        if (alloc) {
+          const dtype = strings[u32()], sz = u64(), nbytes = dtypeSize[dtype]*sz;
+          const shape = {x:[x], y:[y], dtype, sz, nbytes, key};
+          buf_shapes.set(key, shape); temp.set(key, shape);
+          timestamps.push(ts);
+          x += 1; y += nbytes;
+        } else {
+          const free = buf_shapes.get(key);
+          timestamps.push(ts);
+          x += 1; y -= free.nbytes;
+          free.x.push(x);
+          free.y.push(free.y.at(-1));
+          temp.delete(key);
+          for (const [k, v] of temp) {
+            if (k <= key) continue;
+            v.x.push(x, x);
+            v.y.push(v.y.at(-1), v.y.at(-1)-free.nbytes);
+          }
+        }
+      }
+      for (const [_, v] of temp) {
+        v.x.push(x);
+        v.y.push(v.y.at(-1));
+      }
+      timestamps.push(dur);
+      for (const [_, {dtype, sz, nbytes, y, x:steps}] of buf_shapes) {
+        const x = steps.map(s => timestamps[s]);
+        const arg = {tooltipText:`${dtype} len:${formatUnit(sz)}\n${formatUnit(nbytes, "B")}`};
+        shapes.push({ x, y0:y.map(yscale), y1:y.map(y0 => yscale(y0+nbytes)), arg, fillColor:cycleColors(colorScheme.BUFFER, shapes.length) });
       }
       data.tracks.set(k, { shapes, offsetY, height, peak, scaleFactor:maxheight*4/height });
       div.style("height", height+padding+"px").style("cursor", "pointer").on("click", (e) => {
@@ -479,7 +523,6 @@ function setState(ns) {
 
 // set a new context and keep the old one in browser history
 function setCtxWithHistory(newCtx, step=0) {
-  if (newCtx == null) return;
   // NOTE: browser does a structured clone, passing a mutable object is safe.
   history.replaceState(state, "");
   history.pushState(state, "");

tinygrad/viz/js/worker.js

@@ -8,7 +8,7 @@ onmessage = (e) => {
   const { graph, additions, ctxs } = e.data;
   const g = new dagre.graphlib.Graph({ compound: true });
   g.setGraph({ rankdir: "LR" }).setDefaultEdgeLabel(function() { return {}; });
-  if (additions.length !== 0) g.setNode("addition", {label:"", style:"fill: rgba(26, 27, 38, 0.5);", padding:0});
+  if (additions.length !== 0) g.setNode("addition", {label:"", className:"overlay", padding:0});
   for (let [k, {label, src, ref, ...rest }] of Object.entries(graph)) {
     // adjust node dims by label size (excluding escape codes) + add padding
     let [width, height] = [0, 0];
@@ -16,11 +16,11 @@ onmessage = (e) => {
       width = Math.max(width, ctx.measureText(line).width);
       height += LINE_HEIGHT;
     }
-    g.setNode(k, {width:width+NODE_PADDING*2, height:height+NODE_PADDING*2, padding:NODE_PADDING, label, ref, ...rest});
+    g.setNode(k, {width:width+NODE_PADDING*2, height:height+NODE_PADDING*2, padding:NODE_PADDING, label, ref, id:k, ...rest});
     // add edges
     const edgeCounts = {}
-    for (const s of src) edgeCounts[s] = (edgeCounts[s] || 0)+1;
-    for (const s of src) g.setEdge(s, k, { label: edgeCounts[s] > 1 ? edgeCounts[s] : null });
+    for (const [_, s] of src) edgeCounts[s] = (edgeCounts[s] || 0)+1;
+    for (const [port, s] of src) g.setEdge(s, k, { label: edgeCounts[s] > 1 ? {type:"tag", text:edgeCounts[s]} : {type:"port", text:port}});
     if (additions.includes(parseInt(k))) g.setParent(k, "addition");
   }
   dagre.layout(g);

tinygrad/viz/serve.py

@@ -85,7 +85,7 @@ def uop_to_json(x:UOp) -> dict[int, dict]:
     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']}"
     # NOTE: kernel already has metadata in arg
     if TRACEMETA >= 2 and u.metadata is not None and u.op is not Ops.KERNEL: label += "\n"+repr(u.metadata)
-    graph[id(u)] = {"label":label, "src":[id(x) for x in u.src if x not in excluded], "color":uops_colors.get(u.op, "#ffffff"),
+    graph[id(u)] = {"label":label, "src":[(i,id(x)) for i,x in enumerate(u.src) if x not in excluded], "color":uops_colors.get(u.op, "#ffffff"),
                     "ref":ref, "tag":u.tag}
   return graph
 
@@ -131,19 +131,14 @@ def flatten_events(profile:list[ProfileEvent]) -> Generator[tuple[Decimal, Decim
       yield (st:=min(cpu_ts)), (et:=max(cpu_ts)), ProfileRangeEvent(f"{e.ents[0].device.split(':')[0]} Graph", f"batched {len(e.ents)}", st, et)
       for i,ent in enumerate(e.ents): yield (cpu_ts[i*2], cpu_ts[i*2+1], ent)
 
-# timeline layout stacks events in a contiguous block. When a late starter finishes late, there is whitespace in the higher levels.
-def timeline_layout(events:list[tuple[int, int, float, DevEvent]], start_ts:int, scache:dict[str, int]) -> bytes|None:
-  shapes:list[bytes] = []
-  levels:list[int] = []
+# normalize event timestamps and attach kernel metadata
+def timeline_layout(dev_events:list[tuple[int, int, float, DevEvent]], start_ts:int, scache:dict[str, int]) -> bytes|None:
+  events:list[bytes] = []
   exec_points:dict[str, dict] = {}
   category_enum:dict[str, int] = {}
-  for st,et,dur,e in events:
+  for st,et,dur,e in dev_events:
     if isinstance(e, ProfilePointEvent) and e.name == "exec": exec_points[e.key] = e.arg
     if dur == 0: continue
-    # find a free level to put the event
-    depth = next((i for i,level_et in enumerate(levels) if st>=level_et), len(levels))
-    if depth < len(levels): levels[depth] = et
-    else: levels.append(et)
     name, cat, info = e.name, None, None
     if (ref:=ref_map.get(name)) is not None:
       name = ctxs[ref]["name"]
@@ -153,43 +148,28 @@ def timeline_layout(events:list[tuple[int, int, float, DevEvent]], start_ts:int,
     elif isinstance(e.name, TracingKey):
       name, cat = e.name.display_name, e.name.cat
       ref = next((v for k in e.name.keys if (v:=ref_map.get(k)) is not None), None)
-    shapes.append(struct.pack("<IIIfBBI", enum_str(name,scache), option(ref), st-start_ts, dur, depth,
-                              option(None if cat is None else enum_str(cat, category_enum)), enum_str(info or "",scache)))
-  return struct.pack("<BIB", 0, len(shapes), len(levels))+b"".join(shapes) if shapes else None
+    events.append(struct.pack("<IIIfBI", enum_str(name, scache), option(ref), st-start_ts, dur,
+                              option(None if cat is None else enum_str(cat, category_enum)), enum_str(info or "", scache)))
+  return struct.pack("<BI", 0, len(events))+b"".join(events) if events else None
 
-def mem_layout(events:list[tuple[int, int, float, DevEvent]], start_ts:int, end_ts:int, peaks:list[int], dtypes_map:dict[str, int],
+def mem_layout(events:list[tuple[int, int, float, DevEvent]], start_ts:int, end_ts:int, peaks:list[int], dtype_size:dict[str, int],
                scache:dict[str, int]) -> bytes|None:
-  step, peak, mem = 0, 0, 0
-  shps:dict[int, dict] = {}
-  temp:dict[int, dict] = {}
-  timestamps:list[int] = []
+  peak, mem = 0, 0
+  temp:dict[int, int] = {}
+  bufs:list[bytes] = []
   for st,_,_,e in events:
     if not isinstance(e, ProfilePointEvent): continue
     if e.name == "alloc":
-      shps[e.key] = temp[e.key] = {"x":[step], "y":[mem], "arg":{"dtype":e.arg["dtype"].name, "sz":e.arg["sz"]}}
-      dtypes_map.setdefault(e.arg["dtype"].name, e.arg["dtype"].itemsize)
-      timestamps.append(int(e.ts)-start_ts)
-      step += 1
-      mem += e.arg["sz"]*e.arg["dtype"].itemsize
+      bufs.append(struct.pack("<BIIIQ", 1, int(e.ts)-start_ts, e.key, enum_str(e.arg["dtype"].name, scache), e.arg["sz"]))
+      dtype_size.setdefault(e.arg["dtype"].name, e.arg["dtype"].itemsize)
+      temp[e.key] = nbytes = e.arg["sz"]*e.arg["dtype"].itemsize
+      mem += nbytes
       if mem > peak: peak = mem
     if e.name == "free":
-      timestamps.append(int(e.ts)-start_ts)
-      step += 1
-      mem -= (free_nbytes:=(removed:=temp.pop(e.key))["arg"]["sz"]*dtypes_map[removed["arg"]["dtype"]])
-      removed["x"].append(step)
-      removed["y"].append(removed["y"][-1])
-      for k,v in temp.items():
-        if k > e.key:
-          v["x"] += [step, step]
-          v["y"] += [v["y"][-1], v["y"][-1]-free_nbytes]
-  for v in temp.values():
-    v["x"].append(step)
-    v["y"].append(v["y"][-1])
-  timestamps.append(end_ts-start_ts)
+      bufs.append(struct.pack("<BII", 0, int(e.ts)-start_ts, e.key))
+      mem -= temp.pop(e.key)
   peaks.append(peak)
-  bufs = [struct.pack("<I"+str(i:=len(v['x']))+f"I{i}QIQ", i, *v["x"], *v["y"], enum_str(v["arg"]["dtype"], scache),
-                      v["arg"]["sz"]) for v in shps.values()]
-  return struct.pack("<BIQI", 1, len(shps), peak, len(timestamps))+struct.pack(f"<{len(timestamps)}I", *timestamps)+b"".join(bufs) if bufs else None
+  return struct.pack("<BIQ", 1, len(bufs), peak)+b"".join(bufs) if bufs else None
 
 def get_profile(profile:list[ProfileEvent]) -> bytes|None:
   # start by getting the time diffs
@@ -208,13 +188,13 @@ def get_profile(profile:list[ProfileEvent]) -> bytes|None:
   layout:dict[str, bytes|None] = {}
   scache:dict[str, int] = {}
   peaks:list[int] = []
-  dtypes_map:dict[str, int] = {}
+  dtype_size:dict[str, int] = {}
   for k,v in dev_events.items():
     v.sort(key=lambda e:e[0])
     layout[k] = timeline_layout(v, start_ts, scache)
-    layout[f"{k} Memory"] = mem_layout(v, start_ts, unwrap(end_ts), peaks, dtypes_map, scache)
+    layout[f"{k} Memory"] = mem_layout(v, start_ts, unwrap(end_ts), peaks, dtype_size, scache)
   ret = [b"".join([struct.pack("<B", len(k)), k.encode(), v]) for k,v in layout.items() if v is not None]
-  index = json.dumps({"strings":list(scache), "dtypes":dtypes_map}).encode()
+  index = json.dumps({"strings":list(scache), "dtypeSize":dtype_size}).encode()
   return struct.pack("<IQII", unwrap(end_ts)-start_ts, max(peaks,default=0), len(index), len(ret))+index+b"".join(ret)
 
 def get_runtime_stats(key) -> list[dict]:
@@ -277,7 +257,7 @@ class Handler(BaseHTTPRequestHandler):
       if url.path == "/disasm": ret, content_type = get_disassembly(**query), "application/json"
       else: return self.stream_json(get_details(contexts[1][int(query["ctx"][0])][int(query["idx"][0])]))
     elif url.path == "/ctxs": ret, content_type = json.dumps(ctxs).encode(), "application/json"
-    elif url.path == "/get_profile" and profile_ret is not None: ret, content_type = profile_ret, "application/json"
+    elif url.path == "/get_profile" and profile_ret is not None: ret, content_type = profile_ret, "application/octet-stream"
     else: status_code = 404
 
     # send response