Merge branch 'master' into x86_numel

* switch x86 to use numel instead of dtype.count

* works?

.github/workflows/autogen.yml

@@ -42,7 +42,7 @@ jobs:
       run: |
         find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "comgr.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
         python3 -c "from tinygrad.runtime.autogen import opencl"
-        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv"
+        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv_610, nv"
         python3 -c "from tinygrad.runtime.autogen import comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
         python3 -c "from tinygrad.runtime.autogen.am import *"
         python3 -c "from tinygrad.runtime.autogen.nv_regs import *"

.github/workflows/test.yml

@@ -237,7 +237,7 @@ jobs:
         deps: testing_unit
         llvm: 'true'
     - name: Test SPEC=2
-      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 test/unit test/backend test/opt --ignore test/backend/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" -k "not test_conv2d_ceildiv_edge_case" --splits 2 --group ${{ matrix.group }}
+      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 test/unit test/backend test/opt --ignore test/backend/test_custom_kernel.py --ignore test/unit/test_hashing.py -k "not test_setitem_big" -k "not test_conv2d_ceildiv_edge_case" --splits 2 --group ${{ matrix.group }}
 
   fuzzing:
     name: Fuzzing
@@ -552,7 +552,7 @@ jobs:
           key: ${{ matrix.backend }}-minimal
           deps: testing_unit
           amd: 'true'
-          llvm: ${{ matrix.backend == 'amdllvm' && 'true' }}
+          llvm: 'true'
       - name: Check Device.DEFAULT and print some source
         run: |
           python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"

AGENTS.md

@@ -0,0 +1,5 @@
+# Notes
+
+- Run tests with `-n12` for speed (e.g. `python -m pytest test/null/test_dtype.py -x -q -n12`)
+- Run `python -m mypy tinygrad/` to typecheck
+- Run `python -m ruff check .` to lint

examples/anthropic_challenge.py

@@ -1,196 +0,0 @@
-from tinygrad import Tensor, dtypes, Context, getenv, UOp, fetch
-from tinygrad.uop.ops import Ops, PatternMatcher, UPat
-from tinygrad.uop.symbolic import symbolic
-from tinygrad.codegen import Renderer
-from tinygrad.codegen.opt import Opt, OptOps
-
-# ************************* implementation of the problem ************************
-
-def myhash(a: Tensor) -> Tensor:
-  a = (a + 0x7ED55D16) + (a << 12)
-  a = (a ^ 0xC761C23C) ^ (a >> 19)
-  a = (a + 0x165667B1) + (a << 5)
-  a = (a + 0xD3A2646C) ^ (a << 9)
-  a = (a + 0xFD7046C5) + (a << 3)
-  a = (a ^ 0xB55A4F09) ^ (a >> 16)
-  return a
-
-def select_with_where_tree(values: Tensor, relative_idx: Tensor) -> Tensor:
-  n = values.shape[0]
-  if n == 1: return values[0].expand(relative_idx.shape)
-
-  mid = n // 2
-  left = select_with_where_tree(values[:mid], relative_idx)
-  right = select_with_where_tree(values[mid:], relative_idx - mid)
-
-  go_left = relative_idx < mid
-  return go_left.where(left, right)
-
-def tree_traversal(forest: Tensor, val: Tensor, height: int, rounds: int, where_tree_threshold=3) -> Tensor:
-  # All walkers start at idx=0
-  idx = Tensor.zeros(val.shape, device=val.device, dtype=dtypes.uint32)
-
-  for r in range(rounds):
-    level = r % (height + 1)
-    level_start = (1 << level) - 1
-    level_size = 1 << level
-
-    if level == 0:
-      # At root (level 0), all walkers are at idx=0
-      # No gather needed, just broadcast the root value
-      node_val = forest[0].expand(val.shape)
-      idx = idx * 0  # Reset to 0
-    elif level <= where_tree_threshold:
-      # Small level: use where-tree
-      level_values = forest[level_start : level_start + level_size]
-      relative_idx = (idx - level_start)
-      node_val = select_with_where_tree(level_values, relative_idx)
-    else:
-      # Large level: use gather
-      node_val = forest.gather(0, idx)
-
-    val = myhash(val ^ node_val)
-    idx = (idx << 1) + (1 + (val & 1))
-
-    # No wrap check needed! At round 10 (level becomes 0), we reset idx above.
-
-  return val.contiguous(arg=(Opt(OptOps.UPCAST, 0, 8),))
-
-# ************************* renderer for VLIW machine *************************
-
-def loop_unrolling(sink:UOp):
-  rng = [x for x in sink.toposort() if x.op is Ops.RANGE]
-  if len(rng) == 0: return None
-  print(f"unrolling loop with size {rng[0].vmax+1}")
-  unrolled_sinks = [sink.substitute({rng[0]:rng[0].const_like(i)}).src[0] for i in range(rng[0].vmax+1)]
-  return UOp.sink(*unrolled_sinks, arg=sink.arg)
-
-global_addrs = []
-vliw_prepare = PatternMatcher([
-  # loop unrolling (should be a part of tinygrad)
-  (UPat(Ops.SINK, name="sink"), loop_unrolling),
-  # cast is fake
-  (UPat(Ops.CAST, name="c"), lambda c: c.src[0]),
-  # rewrites to hardcode the addresses in memory
-  (UPat(Ops.PARAM, name="dg"), lambda dg: UOp.const(dtypes.uint, global_addrs[dg.arg])),
-  # INDEX is just plus
-  (UPat(Ops.INDEX, name="i"), lambda i: i.src[0]+i.src[1]),
-])+symbolic
-
-class VLIWRenderer(Renderer):
-  has_local = False  # TODO: this should be the default / cleaned up
-  # this says this backend supports MULACC + more. decompositions uses this
-  code_for_op: dict = {Ops.MULACC: None, Ops.ADD: "+", Ops.MUL: "*",
-                       Ops.XOR: "^", Ops.AND: "&", Ops.OR: "|",
-                       Ops.SHL: "<<", Ops.SHR: ">>", Ops.CMPLT: "<"}
-  # this matcher runs while still in graph form
-  pre_matcher = vliw_prepare
-
-  def render(self, uops:list[UOp]):
-
-    # TODO: this is a minimal renderer. for low cycle count, make it good
-    # to get speed, you need to add VLIW packing
-    # to get under 1536 regs, you need to add a register allocator
-    # we left the fun parts to you
-
-    print(f"rendering with {len(uops)} uops")
-    reg, inst = 0, []
-    r: dict[UOp, int] = {}
-    for u in uops:
-      assert u.dtype.count in (1,8), "dtype count must be 1 or 8"
-
-      # dumb register allocator
-      if u.op not in {Ops.STORE, Ops.SINK, Ops.INDEX}:
-        r[u] = reg
-        reg += u.dtype.count
-
-      # render UOps to instructions
-      match u.op:
-        case Ops.SINK:
-          inst.append({"flow": [("halt",)]})
-        case Ops.CONST:
-          inst.append({"load": [("const", r[u], u.arg)]})
-        case Ops.INDEX:
-          # an INDEX is just an alias to a special register in the vector
-          r[u] = r[u.src[0]] + u.src[1].arg
-        case Ops.STACK:
-          if all(s == u.src[0] for s in u.src):
-            # if all sources are the same, we can broadcast
-            inst.append({"valu": [("vbroadcast", r[u], r[u.src[0]])]})
-          else:
-            # this is a copy into a contiguous chunk of registers
-            inst.extend({"flow": [("add_imm", r[u]+i, r[s], 0)]} for i,s in enumerate(u.src) if r[s] != r[u]+i)
-        case Ops.LOAD:
-          op = "vload" if u.dtype.count > 1 else "load"
-          inst.append({"load": [(op, r[u], r[u.src[0]])]})
-        case Ops.STORE:
-          op = "vstore" if u.src[1].dtype.count > 1 else "store"
-          inst.append({"store": [(op, r[u.src[0]], r[u.src[1]])]})
-        case Ops.MULACC:
-          assert u.dtype.count == 8
-          inst.append({"valu": [("multiply_add", r[u], r[u.src[0]], r[u.src[1]], r[u.src[2]])]})
-        case Ops.WHERE:
-          assert u.dtype.count == 8
-          inst.append({"flow": [("vselect", r[u], r[u.src[0]], r[u.src[1]], r[u.src[2]])]})
-        case _ if u.op in self.code_for_op:
-          cat = "valu" if u.dtype.count > 1 else "alu"
-          inst.append({cat: [(self.code_for_op[u.op], r[u], r[u.src[0]], r[u.src[1]])]})
-        case _:
-          raise NotImplementedError(f"unhandled op {u.op}")
-    return repr(inst)
-
-# ************************* test and render *************************
-
-import sys, types
-PROBLEM_URL = "https://raw.githubusercontent.com/anthropics/original_performance_takehome/refs/heads/main/tests/frozen_problem.py"
-sys.modules["problem"] = problem = types.ModuleType("problem")
-exec(fetch(PROBLEM_URL).read_text(), problem.__dict__)
-
-if __name__ == "__main__":
-  batch_size = getenv("BS", 256)
-  height = 10
-  rounds = getenv("ROUNDS", 16)
-
-  # build problem
-  tree = problem.Tree.generate(height)
-  inp = problem.Input.generate(tree, batch_size, rounds)
-  mem = problem.build_mem_image(tree, inp)
-  global_addrs.extend([mem[6], mem[6], mem[4]])  # output, input, forest
-
-  # *** verify the kernel in tinygrad compared to reference ***
-
-  forest_t = Tensor(tree.values, dtype=dtypes.uint32)
-  val_t = Tensor(inp.values, dtype=dtypes.uint32)
-
-  if getenv("VERIFY", 1):
-    # verify on normal tinygrad device
-    with Context(PCONTIG=2):
-      out = tree_traversal(forest_t, val_t, height, rounds)
-      val_out = out.tolist()
-    problem.reference_kernel(tree, inp)
-    assert val_out == inp.values
-    print("verification passed")
-
-  # *** render to device ***
-
-  from tinygrad.codegen import to_program
-  with Context(PCONTIG=2, SPEC=0):
-    out = tree_traversal(forest_t, val_t, height, rounds)
-    sink = out.schedule_linear().src[-1].src[0]
-    prg = to_program(sink, VLIWRenderer())
-
-  # *** run on Machine and compare ***
-
-  # NOTE: the scratch size needs to be reduced to 1536 when you have a register allocator
-  src = eval(prg.src[2].arg)
-  max_regs = max(t[1] for instr in src for v in instr.values() for t in v if len(t) > 1) + 8
-  print(f"{max_regs:5d} regs used" + ("" if max_regs <= 1536 else "       <-- WARNING: TOO MANY REGISTERS, MUST BE <= 1536"))
-  machine = problem.Machine(mem, src, problem.DebugInfo(scratch_map={}), n_cores=1, trace=False, scratch_size=max_regs)
-  machine.run()
-  print(f"ran for {machine.cycle:5d} cycles" + ("" if machine.cycle <= 1363 else "  <-- EVEN CLAUDE GOT 1363"))
-
-  # compare to reference
-  ref_mem = mem.copy()
-  for _ in problem.reference_kernel2(ref_mem, {}): pass
-  assert machine.mem[mem[6]:mem[6]+mem[2]] == ref_mem[mem[6]:mem[6]+mem[2]]
-  print("compare passed!")

examples/hlb_cifar10.py

@@ -152,24 +152,19 @@ def train_cifar():
 
   # ========== Model ==========
   def whitening(X, kernel_size=hyp['net']['kernel_size']):
-    def _cov(X):
-      return (X.T @ X) / (X.shape[0] - 1)
-
-    def _patches(data, patch_size=(kernel_size,kernel_size)):
+    def _patches(data:Tensor, patch_size=(kernel_size,kernel_size)):
       h, w = patch_size
-      c = data.shape[1]
-      axis = (2, 3)
-      return np.lib.stride_tricks.sliding_window_view(data, window_shape=(h,w), axis=axis).transpose((0,3,2,1,4,5)).reshape((-1,c,h,w))
+      _, c, _, _ = data.shape
+      return data._pool((h, w)).permute(1, 4, 5, 0, 3, 2).reshape(c*h*w, -1)
 
     def _eigens(patches):
-      n,c,h,w = patches.shape
-      Σ = _cov(patches.reshape(n, c*h*w))
-      Λ, V = np.linalg.eigh(Σ, UPLO='U')
-      return np.flip(Λ, 0), np.flip(V.T.reshape(c*h*w, c, h, w), 0)
+      cov = ((patches @ patches.T) / (patches.shape[1] - 1)).numpy()
+      eigvals, eigvecs = np.linalg.eigh(cov, UPLO='U')
+      return np.flip(eigvals, 0), np.flip(eigvecs.T.reshape(patches.shape[0], X.shape[1], kernel_size, kernel_size), 0)
 
     # NOTE: np.linalg.eigh only supports float32 so the whitening layer weights need to be converted to float16 manually
-    Λ, V = _eigens(_patches(X.float().numpy()))
-    W = V/np.sqrt(Λ+1e-2)[:,None,None,None]
+    eigvals, eigvecs = _eigens(_patches(X.float()))
+    W = eigvecs/np.sqrt(eigvals+1e-2)[:,None,None,None]
 
     return Tensor(W.astype(np.float32)).cast(dtypes.default_float).is_param_(False)
 
@@ -223,7 +218,7 @@ def train_cifar():
 
   @TinyJit
   def augmentations(X:Tensor, Y:Tensor):
-    perms = Tensor.randperm(X.shape[0], device=X.device) # We reuse perms for cutmix, because they are expensivne to generate
+    perms = Tensor.randperm(X.shape[0], device=X.device) # We reuse perms for cutmix, because they are expensive to generate
     if getenv("RANDOM_CROP", 1):
       X = random_crop(X, crop_size=32)
     if getenv("RANDOM_FLIP", 1):
@@ -314,6 +309,9 @@ def train_cifar():
   opt_bias     = optim.SGD(params_bias,     lr=0.01, momentum=hyp['opt']['momentum'], nesterov=True, weight_decay=hyp['opt']['bias_decay'])
   opt_non_bias = optim.SGD(params_non_bias, lr=0.01, momentum=hyp['opt']['momentum'], nesterov=True, weight_decay=hyp['opt']['non_bias_decay'])
 
+  # realize model params and optimizer state before JIT to avoid cache misses
+  Tensor.realize(*params_dict.values(), *opt_bias.b, *opt_non_bias.b)
+
   # NOTE taken from the hlb_CIFAR repository, might need to be tuned
   initial_div_factor = hyp['opt']['initial_div_factor']
   final_lr_ratio = hyp['opt']['final_lr_ratio']
@@ -330,9 +328,7 @@ def train_cifar():
       # index 0 for bias and 1 for non-bias
       optimizer.zero_grad()
       loss.backward()
-      optimizer.step()
-      lr_scheduler[0].step()
-      lr_scheduler[1].step()
+      return loss.realize(*optimizer.schedule_step(), *lr_scheduler[0].schedule_step(), *lr_scheduler[1].schedule_step())
     return loss.realize()
 
   train_step_jitted = TinyJit(train_step)

examples/mlperf/model_train.py

@@ -2006,7 +2006,7 @@ def train_stable_diffusion():
     # move to CPU first so more GPU bufs aren't created (can trigger OOM)
     for k,v in ckpt.items(): ckpt[k] = v.detach().to("CPU")
     Tensor.realize(*[v for v in ckpt.values()])
-    for k,v in ckpt.items(): ckpt[k] = v.cast(v.dtype.base).contiguous()
+    for k,v in ckpt.items(): ckpt[k] = v.cast(v.dtype).contiguous()
     Tensor.realize(*[v for v in ckpt.values()])
     return ckpt
 

examples/mlperf/optim.py

@@ -6,6 +6,7 @@ from tinygrad.uop.ops import UOp, Ops
 
 STOCHASTIC_ROUND = getenv("STOCHASTIC_ROUND", 0)
 MASTER_WEIGHTS = getenv("MASTER_WEIGHTS", 0)
+ZERO_OPTIM = getenv("ZERO_OPTIM", 0)
 FP8_AMAX_MARGIN = getenv("FP8_AMAX_MARGIN", 1.1)
 IMMEDIATE_SCALE = getenv("IMMEDIATE_SCALE", 0)
 MXFP8 = getenv("MXFP8", 0)
@@ -25,14 +26,24 @@ class GradAccClipAdamW(Optimizer):
     super().__init__(params, lr, device, fused)
     self.b1, self.b2, self.eps, self.wd = b1, b2, eps, weight_decay
     self.b1_t, self.b2_t = (Tensor.ones((1,), dtype=dtypes.float32, device=self.device) for _ in [b1, b2])
-    self.m = self._new_optim_param()
-    self.v = self._new_optim_param()
+    self.zero = bool(ZERO_OPTIM) and isinstance(self.device, tuple) and not self.fused
+    self.m = [self._zero_shard(x) for x in self._new_optim_param()]
+    self.v = [self._zero_shard(x) for x in self._new_optim_param()]
     self.grad_acc, self.clip_norm = grad_acc, clip_norm
     if MASTER_WEIGHTS and self.params[0].dtype != dtypes.float32:
-      self.master_params:list[Tensor]|None = [p.to(self.device).float().contiguous() for p in self.params]
+      self.master_params:list[Tensor]|None = [self._zero_shard(p.to(self.device).float().contiguous()) for p in self.params]
     else:
       self.master_params = None
 
+  def _zero_shard(self, t:Tensor) -> Tensor:
+    if not self.zero or (t.shape[0] % len(self.device)) != 0: return t
+    return Tensor(t.uop._shard(0, len(self.device)).multi(0)).clone()
+
+  def _zero_gather(self, t:Tensor) -> Tensor:
+    if not isinstance(t.device, tuple) or t.uop.axis != 0: return t
+    n, sz = len(t.device), t.shape[0] // len(t.device)
+    return Tensor.cat(*[t[p*sz:(p+1)*sz] for p in range(n)], dim=0)
+
   def fstep(self, grads:list[Tensor]):
     if self.fused:
       out, extra = self._step([], grads)
@@ -85,6 +96,7 @@ class GradAccClipAdamW(Optimizer):
     up = up.float().shard_like(w) + self.lr.to(w.device) * wd * w.detach()
     new_w = w.detach() - up
     if master is not None: master.assign(new_w)
+    if self.zero: new_w = self._zero_gather(new_w)
     # when master is offloaded to a different device than the param, results are resharded back onto the param's (sharded) device
     offloaded = master is not None and master.device != t.device
     if STOCHASTIC_ROUND and t.dtype == dtypes.bfloat16:

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-2}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export MXFP8=${MXFP8:-1}
+export ZERO_OPTIM=${ZERO_OPTIM:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="gptoss"
+export BASEDIR="/raid/datasets/c4-8b/"
+export EVAL_TARGET=3.34 EVAL_FREQ=12288
+export END_LR="4e-5" WARMUP_STEPS=128 MAX_STEPS=1200000
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-5760}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
+if [ -z "$FULL_LAYERS" ]; then
+  export LAYERS=${LAYERS:-2}
+fi
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-0}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export MXFP8=${MXFP8:-1}
+export ZERO_OPTIM=${ZERO_OPTIM:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="gptoss"
+export BASEDIR="/raid/datasets/c4-8b/"
+export EVAL_TARGET=3.34 EVAL_FREQ=12288
+export END_LR="4e-5" WARMUP_STEPS=128 MAX_STEPS=1200000
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-$RANDOM}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+python3 examples/mlperf/model_train.py

examples/yolov8.py

@@ -193,8 +193,8 @@ class SPPF:
     self.cv1 = Conv_Block(c1, c_, 1, 1, padding=None)
     self.cv2 = Conv_Block(c_ * 4, c2, 1, 1, padding=None)
 
-    # TODO: this pads with 0s, whereas torch function pads with -infinity. This results in a < 2% difference in prediction which does not make a difference visually.
-    self.maxpool = lambda x : x.pad((k // 2, k // 2, k // 2, k // 2)).max_pool2d(kernel_size=k, stride=1)
+    # Pad with -inf to match PyTorch's MaxPool2d behavior.
+    self.maxpool = lambda x : x.pad((k // 2, k // 2, k // 2, k // 2), value=float('-inf')).max_pool2d(kernel_size=k, stride=1)
 
   def __call__(self, x):
     x = self.cv1(x)

extra/amdpci/am_smi.py

@@ -1,14 +1,14 @@
 #!/usr/bin/env python3
 
 import time, mmap, sys, shutil, os, glob, subprocess, argparse, collections
-from tinygrad.helpers import DEBUG, colored, ansilen
+from tinygrad.helpers import DEBUG, NO_COLOR, colored, ansilen
 from tinygrad.runtime.autogen import libc
 from tinygrad.runtime.autogen.am import am
 from tinygrad.runtime.support.hcq import MMIOInterface
 from tinygrad.runtime.support.am.amdev import AMDev, AMMemoryManager, AMPageTableEntry
 from tinygrad.runtime.support.am.ip import AM_SOC, AM_GMC, AM_IH, AM_PSP, AM_SMU, AM_GFX, AM_SDMA
 
-def bold(s): return f"\033[1m{s}\033[0m"
+def bold(s): return s if NO_COLOR else f"\033[1m{s}\033[0m"
 
 def trim(s:str, length:int) -> str:
   if len(s) > length: return s[:length-3] + "..."
@@ -276,7 +276,7 @@ class SMICtx:
     return usage
 
   def draw(self, once):
-    terminal_width, terminal_height = shutil.get_terminal_size()
+    terminal_width, terminal_height = shutil.get_terminal_size(fallback=(231, 24))
     if not once and (self.prev_terminal_width != terminal_width or self.prev_terminal_height != terminal_height):
       os.system('clear')
     self.prev_terminal_width, self.prev_terminal_height = terminal_width, terminal_height

extra/gemm/amd_copy_matmul.py

@@ -46,8 +46,8 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
   # -- GLOBAL -> LOCAL --
   # wmma: spatial outer, k inner (k contiguous for vectorized WMMA tile loads)
   # gemm: k outer, spatial inner
-  A_local = UOp.placeholder((BLOCK_M, BLOCK_K) if use_wmma else (BLOCK_K, BLOCK_M), a.dtype.base, slot=0, addrspace=AddrSpace.LOCAL)
-  B_local = UOp.placeholder((BLOCK_N, BLOCK_K) if use_wmma else (BLOCK_K, BLOCK_N), b.dtype.base, slot=1, addrspace=AddrSpace.LOCAL)
+  A_local = UOp.placeholder((BLOCK_M, BLOCK_K) if use_wmma else (BLOCK_K, BLOCK_M), a.dtype, slot=0, addrspace=AddrSpace.LOCAL)
+  B_local = UOp.placeholder((BLOCK_N, BLOCK_K) if use_wmma else (BLOCK_K, BLOCK_N), b.dtype, slot=1, addrspace=AddrSpace.LOCAL)
 
   a = a.reshape(K // BLOCK_K, BLOCK_K, BLOCK_M)
   b = b.reshape(K // BLOCK_K, BLOCK_K, BLOCK_N)

extra/gemm/cdna_asm_gemm.py

@@ -95,6 +95,7 @@ def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
     elif a.ndim == 2 and a.uop.axis == 1 and b.uop.axis == 0: K //= len(a.device)
     elif a.ndim == 2 and a.uop.axis is None and b.uop.axis == 1: N //= len(a.device)
     elif a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None: batch //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis == 1 and b.uop.axis is None: M //= len(a.device)
     elif a.ndim == 3 and a.uop.axis is None and b.uop.axis == 1: N //= len(a.device)
     elif a.ndim == 3 and a.uop.axis == 2 and b.uop.axis == 0: K //= len(a.device)
     else: return todo(f"sharding mismatch a.ndim={a.ndim} a.uop.axis={a.uop.axis} b.uop.axis={b.uop.axis}")
@@ -116,10 +117,10 @@ def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
   m = UOp.range(M, 1, AxisType.LOOP)
   n = UOp.range(N, 2, AxisType.LOOP)
   k = UOp.range(K, 0, AxisType.REDUCE)
-  mul = (A.flatten().index((m*UOp.const(dtypes.weakint, K)+k))*
-         B.flatten().index((k*UOp.const(dtypes.weakint, N)+n))).cast(dtypes.float32)
-  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype.base)
-  store = C.flatten().index((m*UOp.const(dtypes.weakint, N)+n)).store(red).end(m, n)
+  mul = (A.flatten().index((m*UOp.const(dtypes.index, K)+k))*
+         B.flatten().index((k*UOp.const(dtypes.index, N)+n))).cast(dtypes.float32)
+  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype)
+  store = C.flatten().index((m*UOp.const(dtypes.index, N)+n)).store(red).end(m, n)
   return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))
 
 # ** bf16 A @ B.T kernel in C

extra/gemm/metal_uop_matmul.py

@@ -20,8 +20,8 @@ def hand_spec_tc_cores():
 
   gk = UOp.range(N // 8, 0, AxisType.REDUCE)
 
-  a_tc = UOp.vectorize(*[mat_idx(a, gx, gk, warp, i) for i in range(2)])
-  b_tc = UOp.vectorize(*[mat_idx(b, gk, gy, warp, i) for i in range(2)])
+  a_tc = UOp.stack(*[mat_idx(a, gx, gk, warp, i) for i in range(2)])
+  b_tc = UOp.stack(*[mat_idx(b, gk, gy, warp, i) for i in range(2)])
 
   acc = UOp.placeholder((2,), dtypes.float, slot=0, addrspace=AddrSpace.REG)
   acc = acc[0].set(0.0)
@@ -30,7 +30,7 @@ def hand_spec_tc_cores():
   # TODO: make this simple
   wmma_arg = ('WMMA_8_8_8_float_float', (8, 8, 8), dtypes.float, dtypes.float, 'METAL', 32, (((3, 2),), ((3, 2),), ((3, 2),)), ())
 
-  acc_load = UOp.vectorize(acc.after(gk)[0], acc.after(gk)[1])
+  acc_load = UOp.stack(acc.after(gk)[0], acc.after(gk)[1])
   out = UOp(Ops.WMMA, dtypes.float.vec(2), (a_tc, b_tc, acc_load), arg=wmma_arg)
 
   end_loop = UOp.group(*[acc[i].store(out.index(i)) for i in range(2)]).end(gk)

extra/gemm/mi350x_uop_matmul.py

@@ -192,7 +192,7 @@ acc = UOp.placeholder((4,), dtypes.float, 0, AddrSpace.REG)
 acc = acc[init_l:=UOp.range(4, 1)].set(0.0, end=init_l)
 
 # do the wmma
-acc_load = UOp.vectorize(*[acc.after(K_loop)[i] for i in range(4)])
+acc_load = UOp.stack(*[acc.after(K_loop)[i] for i in range(4)])
 wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
 out = UOp(Ops.WMMA, dtypes.float.vec(4), (A_in, B_in, acc_load), arg=wmma_arg)
 

extra/gemm/mi350x_uop_matmul_2.py

@@ -72,7 +72,7 @@ def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
   K_outer_loop = UOp.range(K//BLOCK_K, 0, AxisType.REDUCE)
 
   # split out the globals into blocks
-  C = C.src[0].cast(dtypes.float.vec(4).ptr(C.ptrdtype.size)).reshape((M//BLOCK_M, BLOCK_M, N//BLOCK_N, BLOCK_N))
+  C = C.src[0].cast(dtypes.float.vec(4)).reshape((M//BLOCK_M, BLOCK_M, N//BLOCK_N, BLOCK_N))
   A = A.reshape((M//BLOCK_M, BLOCK_M, K//BLOCK_K, BLOCK_K))[gx, :, K_outer_loop, :]
   B = B.reshape((K//BLOCK_K, BLOCK_K, N//BLOCK_N, BLOCK_N))[K_outer_loop, :, gy, :]
 

extra/hcq2/hcq2.py

@@ -3,175 +3,25 @@ from typing import cast, Callable, TypeVar, Generic, Any
 import struct, functools, time, collections, itertools
 from dataclasses import replace, dataclass
 from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, dedup, pluralize
-from tinygrad.helpers import to_tuple, round_up, partition, data64_le
+from tinygrad.helpers import to_tuple, round_up, partition, data64_le, panic
 from tinygrad.device import Device, Buffer, BufferSpec, Compiled, LRUAllocator, MultiBuffer
 from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, track_rewrites, GroupOp
 from tinygrad.uop.symbolic import symbolic
 from tinygrad.dtype import dtypes, truncate
 from tinygrad.runtime.support.hcq import MMIOInterface
+from tinygrad.runtime.support.memory import BumpAllocator
 from tinygrad.renderer import Renderer, Estimates
 from tinygrad.engine.realize import to_program, get_call_arg_uops, get_call_name, get_call_outs_ins, estimate_uop, pm_flatten_linear
 from tinygrad.engine.jit import DepsTracker
 
 HCQDeviceType = TypeVar('HCQDeviceType', bound='HCQ2Compiled')
 
-class HCQ2Compiled(Compiled):
-  timestamp_divider: float = 1000.0 # GPU timestamp counter ticks per microsecond; override per device
-
-  def __init__(self, device:str, allocator:'HCQAllocator', compilers:list[type[Renderer]], runtime, can_recover:bool=False, arch=None):
-    self.device_id:int = int(device.split(":")[1]) if ":" in device else 0
-
-    # default pm bufferize
-    self.pm_bufferize = PatternMatcher([
-      (UPat(Ops.PARAM, tag="timeline_signal"), lambda ctx: ctx.timeline_signal()),
-      (UPat(Ops.PARAM, tag="timeline_value"), lambda ctx: ctx.timeline_value()),
-      (UPat(Ops.PARAM, tag="sentinel_signal"), lambda ctx: ctx.timeline_signal("sentinel", (1 << 64) - 1)),
-      (UPat(Ops.PARAM, name="b"), lambda ctx, b:
-        Buffer(ctx.device, b.max_numel(), b.dtype.base, options=BufferSpec(host=False, uncached=True, cpu_access=True, nolru=True))
-        if b.tag is not None else None), # TODO: remove nolru
-    ])
-
-    super().__init__(device, allocator, compilers, lambda *a, **kw: None, None, arch=arch)
-
-  @functools.cache
-  def timeline_signal(self, queue:str|None=None, init_value:int=0) -> Buffer:
-    buf = Buffer(self.device, 1, dtypes.uint64, options=BufferSpec(host=True, uncached=True, cpu_access=True), preallocate=True)
-    buf._buf.cpu_view().mv.cast('Q')[0] = init_value
-    return buf
-
-  @functools.cache
-  def timeline_value(self, queue:str|None=None, init_value:int=1) -> Buffer:
-    buf = Buffer("CPU", 1, dtypes.uint64, preallocate=True)
-    buf.as_memoryview(force_zero_copy=True).cast('Q')[0] = init_value
-    return buf
-
-  def synchronize(self, timeout:int|None=None):
-    if not hasattr(self, 'iface'): return
-    sig = self.timeline_signal()._buf.cpu_view().mv.cast('Q')
-    tl = self.timeline_value().as_memoryview(force_zero_copy=True).cast('Q')
-    st = time.perf_counter()
-    while sig[0] < tl[0] - 1:
-      if time.perf_counter() - st > (timeout or 3000) / 1000: self.on_device_hang()
-
-  def device_props(self) -> dict[str,Any]: return {} # to be overridden if needed. dict keys are backend dependent.
-
-  def count(self) -> int: return self.iface.count if hasattr(self, 'iface') else 1
-
-  def _select_iface(self):
-    assert (v:=getenv(k:=f'{type(self).__name__[:-6].upper()}_IFACE', "")) == "",  \
-      f"{k}={v} is deprecated, use DEV={replace(DEV.target(type(self).__name__[:-6]), interface=v)} instead"
-    assert hasattr(self, "ifaces"), "must have ifaces to select an iface"
-    t = DEV.target(dev:=type(self).__name__[:-6])
-    filtered = select_by_name(self.ifaces, lambda i: i.__name__[:-5], t.interface, f"{dev} has no interface {t.interface!r}")
-    filtered = [i for i in filtered if t.interface.startswith("MOCK") or not i.__name__[:-5].startswith("MOCK")] # never fall back to mock ifaces
-    return select_first_inited([functools.partial(cast(Callable, iface), self, self.device_id) for iface in filtered],
-                               f"No interface for {dev}:{self.device_id} is available")
-
-  def _is_cpu(self) -> bool: return hasattr(self, 'device') and self.device.split(":")[0] == "CPU"
-
-  def finalize(self):
-    try: self.synchronize() # try to finalize the device in any case
-    except RuntimeError as e: print(f"{self.device} synchronization failed before finalizing: {e}")
-
-    # if the device has an interface, call device_fini to clean up resources
-    if hasattr(self, 'iface') and hasattr(self.iface, 'device_fini'): self.iface.device_fini()
-
-class HCQ2Buffer:
-  def __init__(self, va_addr:sint, size:int, meta:Any=None, _base:HCQ2Buffer|None=None, view:MMIOInterface|None=None, owner:HCQ2Compiled|None=None):
-    self.va_addr, self.size, self.meta, self._base, self.view, self.owner = va_addr, size, meta, _base, view, owner
-
-  def offset(self, offset:int=0, size:int|None=None) -> HCQ2Buffer:
-    return HCQ2Buffer(self.va_addr+offset, size or (self.size - offset), owner=self.owner, meta=self.meta,
-      _base=self._base or self, view=(self.view.view(offset=offset, size=size) if self.view is not None else None))
-
-  def cpu_view(self) -> MMIOInterface:
-    assert self.view is not None, "buffer has no cpu_view"
-    return self.view
-
-  @property
-  def base(self) -> HCQ2Buffer: return self._base or self
-
-class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
-  def _map(self, buf:HCQ2Buffer) -> HCQ2Buffer:
-    if not hasattr(self, '_do_map'): raise NotImplementedError("map failed: no method implemented")
-    return self._do_map(buf)
-
-  @suppress_finalizing
-  def _free(self, buf:HCQ2Buffer, options:BufferSpec|None=None):
-    self.dev.synchronize()
-    if options is not None and options.external_ptr is not None: return
-    if hasattr(self, '_do_free'): self._do_free(buf, options)
-
-  def _unmap(self, mb):
-    self.dev.synchronize()
-    self.dev.iface.free(mb)
-
-  def _offset(self, buf, size:int, offset:int) -> HCQ2Buffer: return buf.offset(offset=offset, size=size)
-
-  def _wrap(self, dev:str, sz:int, opaque:HCQ2Buffer) -> Buffer:
-    return Buffer(dev, sz, dtypes.uint8, opaque=opaque, options=BufferSpec(external_ptr=1))
-
-  def _copy(self, dst:Buffer, src:Buffer):
-    from tinygrad.engine.realize import run_linear
-    su = UOp.from_buffer(src)
-    run_linear(UOp(Ops.LINEAR, dtypes.void, (su.copy_to_device(dst.device).call(UOp.from_buffer(dst), su),)), update_stats=False)
-
-  def _copyin(self, dest:HCQ2Buffer, src:memoryview):
-    s = Buffer(self.dev.device, len(src), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
-    s._buf.cpu_view()[:len(src)] = src
-    self._copy(self._wrap(self.dev.device, len(src), dest), s)
-
-  def _copyout(self, dest:memoryview, src:HCQ2Buffer):
-    d = Buffer(self.dev.device, len(dest), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
-    self._copy(d, self._wrap(self.dev.device, len(dest), src))
-    self.dev.synchronize()
-    dest[:] = d._buf.cpu_view()[:len(dest)]
-
-  # def _as_buffer(self, buf): return buf.cpu_view().mv
-
 # *****************
 # 0. helpers
 
 HCQ_DEVS = frozenset(("AMD",))
 HCQ_P2P_DEVS = HCQ_DEVS | frozenset(("CPU",))
-
-def all_devices_in(d:Any, c:frozenset[str]) -> bool: return {x.split(":")[0] for x in to_tuple(d)} <= c
-
-def unwrap_after(uop):
-  while uop.op is Ops.AFTER: uop = uop.src[0]
-  return uop
-
-def make_getaddr(u, device=None):
-  if unwrap_after(u).op not in (Ops.BUFFER, Ops.SLICE, Ops.BINARY, Ops.MSTACK, Ops.MSELECT, Ops.PARAM): return u
-  return UOp(Ops.GETADDR, dtypes.uint64, src=(u,), arg=device or to_tuple(u.device)[0])
-
-def make_ins(op, *srcs):
-  return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)
-
-def make_placeholder(devs, size:int, dtype, name=None, unique=True) -> UOp:
-  return UOp.param(next(UOp.unique_num) if unique else 0, dtype.ptr(size), device=devs).rtag(name or "buf")
-
-def make_patch(buf:UOp, off:sint, val:UOp, dtype=None) -> UOp:
-  return buf.index(UOp.const(dtypes.int, off//buf.dtype.base.itemsize)).store(val.cast(dtype or buf.dtype.base))
-
-def make_cmdbuf(lin, devs):
-  blob, patches = b'', []
-  for s in (s for ins in lin.src for s in ins.src):
-    if s.op is not Ops.CONST: patches.append((len(blob), s))
-    blob += struct.pack(f'<{s.dtype.fmt}', s.arg if s.op is Ops.CONST else 0x0)
-  buf = make_placeholder(devs, len(blob) // 4, dtypes.uint32)
-  return buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *[make_patch(buf, off, s) for off, s in patches])
-
-def make_mstack(uops): return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, tuple(uops))
-
-def make_signal(devs, queue=None, sentinel=False):
-  return make_placeholder(devs, 1, dtypes.uint64, "sentinel_signal" if sentinel else (queue, "timeline_signal") if queue else "timeline_signal", unique=False)
-def make_signal_value(devs, queue=None):
-  return make_placeholder(devs, 1, dtypes.uint64, (queue, "timeline_value") if queue else "timeline_value", unique=False)
-
-def make_submit(*cmds, devs:str|tuple[str, ...], queue:str) -> UOp:
-  return UOp.custom_function("submit_cmdbuf", UOp(Ops.LINEAR, dtypes.void, src=tuple(cmds), arg=(to_tuple(devs), queue)))
-def get_submit(ast:UOp) -> UOp: return next(u for u in ast.toposort() if u.op is Ops.CUSTOM_FUNCTION and u.arg == "submit_cmdbuf")
+HCQ_CACHE_TAGS = frozenset(("program", "systems", "template"))
 
 @dataclass(frozen=True)
 class HCQInfo:
@@ -183,11 +33,57 @@ class HCQInfo:
   input_idxs:tuple[int, ...] = () # indexes into input_uops used by this call
   inputs:int|None = None
 
+def all_devices_in(d:Any, c:frozenset[str]) -> bool: return {x.split(":")[0] for x in to_tuple(d)} <= c
+
+def unwrap_after(uop):
+  while uop.op is Ops.AFTER: uop = uop.src[0]
+  return uop
+
+def unwrap_mstack(u):
+  return tuple(x for s in u.src for x in unwrap_mstack(s)) if u.op is Ops.MSTACK else (unwrap_mstack(u.src[0]) if u.op in {Ops.MSELECT, Ops.SLICE} else (u,))
+
+def make_getaddr(u, device=None):
+  if unwrap_after(u).op not in (Ops.BUFFER, Ops.SLICE, Ops.BINARY, Ops.MSTACK, Ops.MSELECT, Ops.PARAM): return u
+  return UOp(Ops.GETADDR, dtypes.uint64, src=(u,), arg=device or to_tuple(u.device)[0])
+
+def make_ins(op, *srcs):
+  return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)
+
+def make_placeholder(devs, size:int, dtype, name=None, unique=True) -> UOp:
+  return UOp.param(next(UOp.unique_num) if unique else 0, dtype, shape=(size,), device=devs).rtag(name or "temp")
+
+def make_patch(buf:UOp, off:sint, val:UOp, dtype=None) -> UOp:
+  return buf.index(UOp.const(dtypes.int, off // buf.dtype.itemsize)).store(val.simplify().cast(dtype or buf.dtype))
+
+def make_binary_patch(buf:UOp, blob:bytes) -> UOp:
+  data = UOp(Ops.BITCAST, buf.dtype, (UOp(Ops.BINARY, dtypes.uint8, src=(), arg=blob),))
+  r = UOp.range(len(blob) // buf.dtype.itemsize, 0, dtype=dtypes.int, src=(buf, data))
+  return buf.index(r).store(data.index(r).load()).end(r)
+
+def make_cmdbuf(lin, devs):
+  blob, patches = b'', []
+  for s in (s for ins in lin.src for s in ins.src):
+    if (ssimp:=s.simplify()).op is not Ops.CONST: patches.append((len(blob), ssimp))
+    blob += struct.pack(f'<{ssimp.dtype.fmt}', ssimp.arg if ssimp.op is Ops.CONST else 0x0)
+  cmdbuf = make_placeholder(devs, len(blob) // 4, dtypes.uint32, name="cmdbuf")
+  return cmdbuf.after(make_binary_patch(cmdbuf, blob), *[make_patch(cmdbuf, off, s) for off, s in patches])
+
+def make_mstack(uops): return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, tuple(uops))
+
+def make_signal(devs, queue=None, sentinel=False):
+  return make_placeholder(devs, 1, dtypes.uint64, "sentinel_signal" if sentinel else (queue, "timeline_signal") if queue else "timeline_signal", unique=False)
+def make_signal_value(devs, queue=None):
+  return make_placeholder(devs, 1, dtypes.uint64, (queue, "timeline_value") if queue else "timeline_value", unique=False)
+
+def make_submit(*cmds, devs:str|tuple[str, ...], queue:str) -> UOp:
+  return UOp.custom_function("submit_cmdbuf", UOp(Ops.LINEAR, src=tuple(cmds), arg=(to_tuple(devs), queue)))
+def get_submit(ast:UOp) -> UOp: return next(u for u in ast.toposort() if u.op is Ops.CUSTOM_FUNCTION and u.arg == "submit_cmdbuf")
+
 # *****************
 # 0.1. prep: replace buffers with params
 
 def replace_call_buffers(ctx:list[UOp], call:UOp) -> UOp|None:
-  ctx += [s for s in dedup(call.src[1:]) if s not in ctx and s.op not in (Ops.PARAM, Ops.BIND)]
+  ctx += [s for s in call.src[1:] if s not in ctx and s.op not in (Ops.PARAM, Ops.BIND)]
   return call.replace(src=call.src[:1] + tuple(s if s.op in (Ops.PARAM, Ops.BIND) else s.param_like(ctx.index(s)) for s in call.src[1:]))
 pm_replace_buffers = PatternMatcher([(UPat(Ops.CALL, name="call"), replace_call_buffers)])
 
@@ -199,8 +95,8 @@ def _need_staging(a, b): return all_devices_in(a.device, HCQ_DEVS) and not all_d
 def stage_copy(dst:UOp, src:UOp) -> UOp|None:
   if not (_need_staging(src, dst) or _need_staging(dst, src)): return None
 
-  stage = UOp.new_buffer("CPU", src.max_numel() * src.dtype.base.itemsize, dtypes.uint8)
-  return UOp(Ops.LINEAR, dtypes.void, (src.copy_to_device("CPU").call(stage, src), stage.copy_to_device(dst.device).call(dst, stage)))
+  stage = UOp.new_buffer("CPU", src.max_numel() * src.dtype.itemsize, dtypes.uint8)
+  return UOp(Ops.LINEAR, src=(src.copy_to_device("CPU").call(stage, src), stage.copy_to_device(dst.device).call(dst, stage)))
 pm_insert_copy_staging = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.COPY), UPat(name="dst"), UPat(name="src"))), stage_copy)])
 
 # *****************
@@ -212,8 +108,7 @@ def tag_hcq_call(ctx:itertools.count, call:UOp) -> UOp:
   queue = "COMPUTE:0" if call.src[0].op is Ops.PROGRAM else "COPY:0"
   info = HCQInfo(get_call_name(call, get_call_arg_uops(call)), estimate_uop(call), to_tuple(hcq_devs), queue)
   return call.replace(arg=replace(call.arg, aux=info)).rtag(next(ctx))
-pm_tag_hcq_calls = PatternMatcher([(UPat(Ops.LINEAR, name="linear"),
-  lambda ctx, linear: linear.replace(src=tuple(tag_hcq_call(ctx, s) for s in linear.src)))])
+pm_tag_hcq_calls = PatternMatcher([(UPat(Ops.LINEAR, name="l"), lambda ctx, l: l.replace(src=tuple(tag_hcq_call(ctx, s) for s in l.src)))])
 
 # *****************
 # 2.2. deps tracking
@@ -232,7 +127,7 @@ pm_tag_hcq_calls = PatternMatcher([(UPat(Ops.LINEAR, name="linear"),
 class HCQDepsTracker(DepsTracker):
   @staticmethod
   def _key(buf:Any) -> tuple[Any, int, int]:
-    return (buf.arg.slot, 0, buf.max_numel() * buf.dtype.base.itemsize) if isinstance(buf, UOp) else DepsTracker._key(buf)
+    return (buf.arg.slot, 0, buf.max_numel() * buf.dtype.itemsize) if isinstance(buf, UOp) else DepsTracker._key(buf)
 
 def make_deps(u:UOp, dep_lanes:list[tuple[UOp, int, int]], nlanes:int) -> UOp:
   deps:dict[UOp, list[int|None]] = collections.defaultdict(lambda: [None]*nlanes)
@@ -377,19 +272,20 @@ pm_encode_cmdbufs = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbu
 
 # *****************
 
-def unwrap_mstack(u): return u.src if u.op is Ops.MSTACK else (u,)
+def is_value_known_at_link(val:UOp) -> bool:
+  runtime_reads = [u for u in val.toposort() if u.op in (Ops.LOAD, Ops.INDEX)]
+  addressed_bufs = [b for g in val.toposort() if g.op is Ops.GETADDR for b in unwrap_mstack(g.buf_uop)]
 
-def _is_link_patch(p:UOp, buf:UOp, jit=False) -> bool:
-  if p.op is not Ops.STORE or p.buf_uop is not buf: return False # this is not a patch :(
+  # addr of input params is not known at link time
+  return not runtime_reads and all(b.op is not Ops.PARAM or b.tag is not None for b in addressed_bufs)
 
-  assert all(x.op is Ops.PARAM for x in unwrap_mstack(p.buf_uop))
-  has_loads = any(u.op in (Ops.LOAD, Ops.INDEX) for u in p.src[1].backward_slice)
-  param_is_input = all(x.tag is None and x.op is Ops.PARAM for x in unwrap_mstack(p.src[1].buf_uop))
-
-  return not has_loads and not param_is_input if True else (p.buf_uop.tag in {"program"})
+def is_link_patch(p:UOp, jit:bool) -> bool:
+  store = p.src[0] if (is_binary_patch:=p.op is Ops.END) else p
+  if not jit: return store.buf_uop.tag == "program"
+  return is_binary_patch or (store.op is Ops.STORE and is_value_known_at_link(store.src[1]))
 
 def trim_link_patches(ctx:tuple[bool, list[UOp]], a:UOp) -> UOp|None:
-  links, kept = partition(a.src[1:], lambda p: _is_link_patch(p, a.src[0], jit=ctx[0]))
+  links, kept = partition(a.src[1:], lambda p: is_link_patch(p, ctx[0]))
 
   # keep all patches from the link-time patches' subtrees in the C code
   afters = [u for u in UOp.sink(*links).toposort() if u.op is Ops.AFTER]
@@ -407,26 +303,38 @@ pm_split_patches = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION
 
 # *****************
 
-def _make_getaddrs_sub(call:UOp, gaddrs:list[UOp], name:str):
+def make_addr_table(call:UOp, gaddrs:list[UOp], name:str):
   bare = {g: g.replace(src=(unwrap_after(g.src[0]),)) for g in gaddrs}
 
-  order = sorted(dedup(bare.values()), key=lambda g: (g.buf_uop.arg.slot, to_tuple(g.buf_uop.tag)))
-  b = make_placeholder(call.arg.aux.device, len(order), dtypes.uint64, name)
+  order = sorted(dedup(bare.values()), key=lambda g: ((b:=unwrap_mstack(g.buf_uop)[0]).arg.slot, to_tuple(b.tag)))
+  slots, table = {g:i for i,g in enumerate(order)}, make_placeholder(call.arg.aux.device, len(order), dtypes.uint64, name)
 
-  sub = {g: b.after(*g.src[0].src[1:] if g.src[0].op is Ops.AFTER else ()).index(UOp.const(dtypes.int, order.index(gr))).load() for g,gr in bare.items()}
-  return sub, (b.after(*[make_patch(b, i * b.dtype.base.itemsize, gr) for i,gr in enumerate(order)]),) if order else ()
+  reads = {g: table.after(*g.src[0].src[1:] if g.src[0].op is Ops.AFTER else ()).index(UOp.const(dtypes.int, slots[bare[g]])).load() for g in gaddrs}
+  return reads, (table.after(*[make_patch(table, i * table.dtype.itemsize, addr) for addr, i in slots.items()]),) if slots else ()
 
 def rm_rt_getaddrs(call:UOp) -> UOp|None:
   if not (gaddrs:=[u for u in call.src[0].toposort() if u.op is Ops.GETADDR]): return None
-  inputs, systems = partition(gaddrs, lambda g: all(x.tag is None for x in unwrap_mstack(g.buf_uop)))
+  inputs, internals = partition(gaddrs, lambda g: all(x.op is Ops.PARAM and x.tag is None for x in unwrap_mstack(g.buf_uop)))
+  runtimes, systems = partition(internals, lambda g: any(x.tag in {"program", "kernargs", "cmdbuf"} for x in unwrap_mstack(g.buf_uop)))
 
-  (inpsub, _), (syssub, sysarg) = _make_getaddrs_sub(call, inputs, "inputs"), _make_getaddrs_sub(call, systems, "systems")
-  return call.replace(src=(call.src[0].substitute(inpsub | syssub), *call.src[1:], *sysarg),
+  # exec fills the inputs table with the input addresses every run, so it has no fill patches
+  (input_reads, _), (rt_reads, rt_fills), (sys_reads, sys_fills) = (make_addr_table(call, gs, name) for gs, name in
+    ((inputs, "inputs"), (runtimes, "runtime"), (systems, "systems")))
+  return call.replace(src=(call.src[0].substitute(input_reads | rt_reads | sys_reads), *call.src[1:], *rt_fills, *sys_fills),
                       arg=replace(call.arg, aux=replace(call.arg.aux, input_idxs=tuple(sorted(dedup(g.buf_uop.arg.slot for g in inputs))))))
 pm_rm_rt_getaddrs = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), rm_rt_getaddrs)])
 
 # *****************
 
+def rm_rt_binaries(call:UOp) -> UOp|None:
+  if not (blobs:=[u for u in call.src[0].toposort() if u.op is Ops.BITCAST and u.src[0].op is Ops.BINARY]): return None
+  blob_bufs = {blob: make_placeholder(call.arg.aux.device, blob.max_numel(), blob.dtype, "template") for blob in blobs}
+  fills = [buf.after(make_binary_patch(buf, blob.src[0].arg)) for blob, buf in blob_bufs.items()]
+  return call.replace(src=(call.src[0].substitute(blob_bufs), *call.src[1:], *fills))
+pm_rm_rt_binaries = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), rm_rt_binaries)])
+
+# *****************
+
 def replace_params(call:UOp) -> UOp|None:
   body, variables, param_ops = call.src[0], call.src[0].variables(), {Ops.PARAM, Ops.MSTACK}
   args = dedup([s for u in body.toposort(gate=lambda u: u.op not in param_ops) for s in u.src if s.op in param_ops and s not in variables])
@@ -435,7 +343,7 @@ def replace_params(call:UOp) -> UOp|None:
   by_root = {p.src[0]: p for p in patched}
   c_args = [by_root.get(a, a) for a in args]
 
-  sub = {unwrap_after(u): UOp.param(i, u.dtype, device=u.device) for i,u in enumerate(c_args)} | \
+  sub = {unwrap_after(u): UOp.param(i, u.dtype, shape=unwrap_after(u).shape, device=u.device) for i,u in enumerate(c_args)} | \
         {v: v.replace(arg=replace(v.arg, slot=-1)) for v in variables if v.op is Ops.PARAM}
   info = replace(call.arg.aux, inputs=next((i for i,u in enumerate(c_args) if u.tag == "inputs"), None))
   return call.replace(src=(body.substitute(sub), *c_args, *refhold), arg=replace(call.arg, aux=info)) # TODO: call.after(*refhold)?
@@ -444,35 +352,32 @@ pm_replace_params = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTIO
 # *****************
 
 def resolve_getaddr_slice(bv:UOp, g:UOp) -> UOp:
+  base = bv.src[0].after(*g.src[0].src[1:] if g.src[0].op is Ops.AFTER else ())
   itemsize = bv.src[0].dtype.itemsize if unwrap_after(bv.src[0]).op in (Ops.BUFFER, Ops.SLICE, Ops.MSTACK, Ops.MSELECT) else bv.dtype.itemsize
-  return UOp(Ops.GETADDR, dtypes.uint64, src=(bv.src[0],), arg=g.arg) + UOp.const(dtypes.uint64, bv.src[1].arg * itemsize)
+  return UOp(Ops.GETADDR, dtypes.uint64, src=(base,), arg=g.arg) + UOp.const(dtypes.uint64, bv.src[1].arg * itemsize)
 
 pm_early_simplify = PatternMatcher([
-  # getaddr(slice(base, off)) -> getaddr(base) + byte offset
-  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"),), name="g"), resolve_getaddr_slice),
+  (UPat(Ops.GETADDR, src=(UPat.any(sl:=UPat(Ops.SLICE, name="bv"), sl.after(allow_any_len=True)),), name="g"), resolve_getaddr_slice),
+  (UPat(Ops.INDEX, src=(UPat(Ops.SLICE, name="bv"),), allow_any_len=True, name="x"),
+   lambda bv,x: x.replace(src=(bv.src[0], x.src[1] + bv.src[1].cast(x.src[1].dtype), *x.src[2:]))),
 ])
 
 # *****************
 # 5.3. pack placeholders buffers
 
-# def pack_hcq_placeholders(call:UOp) -> UOp|None:
-#   bufs = [b for b in call.src[0].toposort() if b.op is Ops.PARAM and b.tag in (maxtags:={"scratch"}) | (sumtags:={"program", "kernargs"})]
-
-#   off_per_buf:dict[UOp, int] = {}
-#   size_per_tag:dict[str, int] = {}
-#   for b in bufs:
-#     bsz = b.max_numel()
-#     if b.tag in maxtags: size_per_tag[b.tag] = max(size_per_tag.get(b.tag, 0), bsz)
-#     elif b.tag in sumtags:
-#       off_per_buf[b] = round_up(size_per_tag.get(b.tag, 0), {"program": 0x1000}.get(b.tag, 128))
-#       size_per_tag[b.tag] = off_per_buf[b] + bsz
-
-#   count_per_tag = collections.Counter(b.tag for b in bufs)
-#   ref_bufs = {b.tag:b for b in bufs if count_per_tag[b.tag] > 1}
-#   bases = {tag:UOp.new_buffer(b.device, size_per_tag[tag], b.dtype).rtag(tag) for tag,b in ref_bufs.items()}
-#   subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(dtypes.weakint, off_per_buf.get(b, 0))), b.max_numel()) for b in bufs if b.tag in bases}
-#   return call.replace(src=(call.src[0].substitute(subs, walk=True), *call.src[1:])) if subs else None
-# pm_pack_placeholders = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), pack_hcq_placeholders)])
+def pack_hcq_placeholders(call:UOp) -> UOp|None:
+  bufs = [b for b in call.src[0].toposort() if b.op is Ops.PARAM and b.tag in {"scratch", "kernargs"}]
+  offs, sizes = {}, {}
+  for b in bufs:
+    if b.tag == "scratch": sizes[b.tag] = max(sizes.get(b.tag, 0), b.max_numel())
+    else:
+      offs[b] = round_up(sizes.get(b.tag, 0), 128 // b.dtype.itemsize)
+      sizes[b.tag] = offs[b] + b.max_numel()
+  counts = collections.Counter(b.tag for b in bufs)
+  bases = {b.tag:make_placeholder(b.device, sizes[b.tag], b.dtype, b.tag) for b in bufs if counts[b.tag] > 1}
+  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(dtypes.index, offs.get(b, 0))), b.max_numel()) for b in bufs if b.tag in bases}
+  return call.replace(src=(call.src[0].substitute(subs, walk=True), *call.src[1:])) if subs else None
+pm_pack_placeholders = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), pack_hcq_placeholders)])
 
 # *****************
 # 8. callify hcq programs
@@ -480,13 +385,13 @@ pm_early_simplify = PatternMatcher([
 pm_callify_hcq = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="hcq", src=(UPat(Ops.SINK),), name="cf"),
   lambda cf: cf.replace(src=(to_program(cf.src[0].replace(arg=KernelInfo("hcq_submit"), tag=1), Device["CPU"].renderer),)))])
 
-hcq_compile_cache:dict[bytes, UOp] = {}
+hcq_compile_cache:dict[tuple[bytes, bool], UOp] = {}
 
-@track_rewrites(lambda linear,input_uops,ret: f"HCQ Compile {pluralize('Kernel', len(ret.src))}")
-def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None) -> UOp:
+@track_rewrites(lambda linear,input_uops,jit,ret: f"HCQ Compile {pluralize('Kernel', len(ret.src))}")
+def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None, jit=False) -> UOp:
   if input_uops is not None: linear = graph_rewrite(linear, pm_replace_buffers, ctx=input_uops, walk=True, enter_calls=True, name="replace buffer")
 
-  if (final_linear:=(hcq_compile_cache.get(cache_key:=linear.key))) is None:
+  if (final_linear:=(hcq_compile_cache.get(cache_key:=(linear.key, jit)))) is None:
     # schedule
     linear = linear.substitute(back_map:={s.param_like(i): s for i,s in enumerate(input_uops)} if input_uops is not None else {}, walk=True)
     linear = graph_rewrite(linear, pm_insert_copy_staging + pm_flatten_linear, name="insert copy staging")
@@ -501,14 +406,15 @@ def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None) -> UOp:
     linear = graph_rewrite(linear, pm_add_inner_loads, ctx=(waited:=set()), walk=True, name="add loads", enter_calls=True)
     linear = graph_rewrite(linear, pm_add_inner_stores, ctx=waited, walk=True, name="add stores", enter_calls=True)
     linear = graph_rewrite(linear, pm_encode_cmdbufs, walk=True, name="encode cmdbufs", enter_calls=True)
+    linear = graph_rewrite(linear, pm_pack_placeholders, walk=True, name="pack placeholders")
 
     # pie
-    linear = graph_rewrite(linear, pm_split_patches, walk=True, name="split rt/lt patches")
+    linear = graph_rewrite(linear, pm_split_patches, ctx=jit, walk=True, name="split rt/lt patches")
+    linear = graph_rewrite(linear, pm_early_simplify + symbolic, bottom_up=False, name="simplify packed placeholders", enter_calls=True)
     linear = graph_rewrite(linear, pm_rm_rt_getaddrs, walk=True, name="replace rt getaddrs")
+    linear = graph_rewrite(linear, pm_rm_rt_binaries, walk=True, name="replace rt binaries")
     linear = graph_rewrite(linear, pm_replace_params, walk=True, name="replace with args")
 
-    linear = graph_rewrite(linear, pm_early_simplify + symbolic, bottom_up=False, name="early simplify patches", enter_calls=True)
-
     # and compile it
     final_linear = hcq_compile_cache[cache_key] = graph_rewrite(linear, pm_callify_hcq, name="callify hcq", enter_calls=True)
 
@@ -517,9 +423,9 @@ def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None) -> UOp:
 # *****************
 # 6. bufferize placeholders: replace placeholders with real buffers.
 
-def bufferize_buf(buf:UOp) -> UOp|None:
+def bufferize_buf(ctx:bool, buf:UOp) -> UOp|None:
   if buf.tag is None: return None
-  return make_mstack(tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=dv), "CPU") for dev in to_tuple(buf.device)))
+  return make_mstack(tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=(dv, ctx)), "CPU") for dev in to_tuple(buf.device)))
 pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, name="buf"), bufferize_buf)])
 
 # *****************
@@ -528,13 +434,13 @@ pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, name="buf"), bufferize_buf)])
 def push_stack(op, s): return UOp(Ops.STACK, op.dtype.scalar().vec(len(s.src)),
   tuple(op.replace(dtype=op.dtype.scalar(), src=tuple(x if y is s else y for y in op.src)) for x in s.src))
 
-def fold_blob_store(buf:UOp, blob:UOp) -> UOp:
-  for b in (mb.bufs if isinstance((mb:=buf.buffer), MultiBuffer) else (mb,)): b.ensure_allocated()._buf.cpu_view().mv.cast('B')[:len(blob.arg)] = blob.arg
+def fold_binary(buf:UOp, blob:UOp) -> UOp:
+  for b in (m.bufs if isinstance(m:=buf.buffer, MultiBuffer) else (m,)): b.ensure_allocated()._buf.cpu_view().view(fmt='B')[:len(blob.arg)] = blob.arg
   return UOp(Ops.NOOP)
 
 def fold_const_store(buf:UOp, off:UOp, val:UOp) -> UOp:
   for b, v in zip((bs:=mb.bufs if isinstance((mb:=buf.buffer), MultiBuffer) else (mb,)), val.src if val.op is Ops.STACK else (val,)*len(bs)):
-    struct.pack_into(f'<{v.dtype.fmt}', b.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * buf.dtype.base.itemsize, truncate[v.dtype](v.arg))
+    struct.pack_into(f'<{v.dtype.fmt}', b.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * buf.dtype.itemsize, truncate[v.dtype](v.arg))
   return UOp(Ops.NOOP)
 
 def resolve_getaddr(buf:UOp, g:UOp) -> UOp:
@@ -555,12 +461,149 @@ pm_resolve_patches = PatternMatcher([
   (UPat(Ops.GETADDR, src=(UPat(name="buf"),), name="g"), resolve_getaddr),
 
   # folders
-  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").store(UPat(Ops.BINARY, name="blob")), fold_blob_store),
+  (UPat(name="buf").index(UPat(Ops.RANGE), allow_any_len=True)
+    .store(UPat.any(UPat(Ops.BINARY, name="blob"), UPat(Ops.BINARY, name="blob").bitcast()).index(UPat(Ops.RANGE), allow_any_len=True).load())
+    .end(UPat(Ops.RANGE)), fold_binary),
   (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").index(UPat.cvar("off"))
     .store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))), fold_const_store),
 ])
 
-@track_rewrites(lambda _,ret: f"HCQ Link {pluralize('Kernel', len(ret.src))}")
-def hcq_link(linear:UOp) -> UOp:
-  linear = graph_rewrite(linear, pm_bufferize, bottom_up=True, walk=True, name="bufferize placeholders")
-  return graph_rewrite(linear, pm_resolve_patches + symbolic, bottom_up=False, name="simplify patches")
+pm_assert_no_afters = PatternMatcher([(UPat(Ops.AFTER, name="a"), lambda a: panic(RuntimeError, f"AFTER left at hcq_link: {a.src[0].op}"))])
+
+hcq_link_cache:dict[tuple[bytes, tuple[str, ...]], UOp] = {}
+
+def link_cache_key(a:UOp): return a.key, to_tuple(a.device)
+pm_link_cache = PatternMatcher([(UPat(Ops.AFTER, name="a"), lambda a: hcq_link_cache.get(link_cache_key(a)))])
+
+@track_rewrites(lambda _,jit,ret: f"HCQ Link {pluralize('Kernel', len(ret.src))}")
+def hcq_link(linear:UOp, jit=False) -> UOp:
+  cacheable = {(j,i):a for j,c in enumerate(linear.src) for i,a in enumerate(c.src[1:], 1)
+               if a.op is Ops.AFTER and unwrap_mstack(a.src[0])[0].tag in HCQ_CACHE_TAGS}
+  hits = {a.src[0]:hcq_link_cache[key] for a in cacheable.values() if (key:=link_cache_key(a)) in hcq_link_cache}
+  linear = graph_rewrite(linear, pm_link_cache, name="apply link cache").substitute(hits, walk=True)
+  linear = graph_rewrite(linear, pm_bufferize, ctx=jit, bottom_up=True, walk=True, name="bufferize placeholders")
+  linear = graph_rewrite(linear, pm_resolve_patches + symbolic, bottom_up=False, name="simplify patches")
+  linear = graph_rewrite(linear, pm_assert_no_afters, name="assert no afters")
+  for (j,i),a in cacheable.items(): hcq_link_cache.setdefault(link_cache_key(a), linear.src[j].src[i])
+  return linear
+
+# *****************
+# Device classes
+
+class HCQ2Compiled(Compiled):
+  timestamp_divider: float = 1000.0
+
+  def __init__(self, device:str, allocator:HCQAllocator, compilers:list[type[Renderer]], runtime, can_recover:bool=False, arch=None):
+    self.device_id:int = int(device.split(":")[1]) if ":" in device else 0
+
+    self.pm_bufferize = PatternMatcher([
+      (UPat(Ops.PARAM, tag="timeline_signal"), lambda ctx: ctx[0].timeline_signal()),
+      (UPat(Ops.PARAM, tag="timeline_value"), lambda ctx: ctx[0].timeline_value()),
+      (UPat(Ops.PARAM, tag="sentinel_signal"), lambda ctx: ctx[0].timeline_signal("sentinel", (1 << 64) - 1)),
+      (UPat(Ops.PARAM, name="b"), lambda ctx, b: None if b.tag is None else ctx[0].new_buffer(b, jit=ctx[1]))
+    ])
+
+    super().__init__(device, allocator, compilers, lambda *a, **kw: None, None, arch=arch)
+
+    self.rt_buffer = Buffer(self.device, 64 << 20, dtypes.uint8, options=BufferSpec(uncached=True, cpu_access=True))
+    self.rt_allocator = BumpAllocator(64 << 20, wrap=False)
+
+  def new_buffer(self, b:UOp, jit:bool) -> Buffer:
+    if jit or b.tag in HCQ_CACHE_TAGS: return Buffer(self.device, b.max_numel(), b.dtype, options=BufferSpec(cpu_access=True, nolru=True))
+    return self.rt_buffer.view(b.max_numel(), b.dtype, self.rt_allocator.alloc(b.max_numel() * b.dtype.itemsize, alignment=128))
+
+  @functools.cache
+  def timeline_signal(self, queue:str|None=None, init_value:int=0) -> Buffer:
+    buf = Buffer(self.device, 1, dtypes.uint64, options=BufferSpec(host=True, uncached=True, cpu_access=True), preallocate=True)
+    buf._buf.cpu_view().mv.cast('Q')[0] = init_value
+    return buf
+
+  @functools.cache
+  def timeline_value(self, queue:str|None=None, init_value:int=1) -> Buffer:
+    buf = Buffer("CPU", 1, dtypes.uint64, preallocate=True)
+    buf.as_memoryview(force_zero_copy=True).cast('Q')[0] = init_value
+    return buf
+
+  def synchronize(self, timeout:int|None=None):
+    if not hasattr(self, 'iface'): return
+    sig = self.timeline_signal()._buf.cpu_view().mv.cast('Q')
+    tl = self.timeline_value().as_memoryview(force_zero_copy=True).cast('Q')
+    st = time.perf_counter()
+    while sig[0] < tl[0] - 1:
+      if time.perf_counter() - st > (timeout or 3000) / 1000: self.on_device_hang()
+
+  def device_props(self) -> dict[str,Any]: return {} # to be overridden if needed. dict keys are backend dependent.
+
+  def count(self) -> int: return self.iface.count if hasattr(self, 'iface') else 1
+
+  def _select_iface(self):
+    assert (v:=getenv(k:=f'{type(self).__name__[:-6].upper()}_IFACE', "")) == "",  \
+      f"{k}={v} is deprecated, use DEV={replace(DEV.target(type(self).__name__[:-6]), interface=v)} instead"
+    assert hasattr(self, "ifaces"), "must have ifaces to select an iface"
+    t = DEV.target(dev:=type(self).__name__[:-6])
+    filtered = select_by_name(self.ifaces, lambda i: i.__name__[:-5], t.interface, f"{dev} has no interface {t.interface!r}")
+    filtered = [i for i in filtered if t.interface.startswith("MOCK") or not i.__name__[:-5].startswith("MOCK")] # never fall back to mock ifaces
+    return select_first_inited([functools.partial(cast(Callable, iface), self, self.device_id) for iface in filtered],
+                               f"No interface for {dev}:{self.device_id} is available")
+
+  def _is_cpu(self) -> bool: return hasattr(self, 'device') and self.device.split(":")[0] == "CPU"
+
+  def finalize(self):
+    try: self.synchronize() # try to finalize the device in any case
+    except RuntimeError as e: print(f"{self.device} synchronization failed before finalizing: {e}")
+
+    # if the device has an interface, call device_fini to clean up resources
+    if hasattr(self, 'iface') and hasattr(self.iface, 'device_fini'): self.iface.device_fini()
+
+class HCQ2Buffer:
+  def __init__(self, va_addr:sint, size:int, meta:Any=None, _base:HCQ2Buffer|None=None, view:MMIOInterface|None=None, owner:HCQ2Compiled|None=None):
+    self.va_addr, self.size, self.meta, self._base, self.view, self.owner = va_addr, size, meta, _base, view, owner
+
+  def offset(self, offset:int=0, size:int|None=None) -> HCQ2Buffer:
+    return HCQ2Buffer(self.va_addr+offset, size or (self.size - offset), owner=self.owner, meta=self.meta,
+      _base=self._base or self, view=(self.view.view(offset=offset, size=size) if self.view is not None else None))
+
+  def cpu_view(self) -> MMIOInterface:
+    assert self.view is not None, "buffer has no cpu_view"
+    return self.view
+
+  @property
+  def base(self) -> HCQ2Buffer: return self._base or self
+
+class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
+  def _map(self, buf:HCQ2Buffer) -> HCQ2Buffer:
+    if not hasattr(self, '_do_map'): raise NotImplementedError("map failed: no method implemented")
+    return self._do_map(buf)
+
+  @suppress_finalizing
+  def _free(self, buf:HCQ2Buffer, options:BufferSpec|None=None):
+    if options is not None and options.external_ptr is not None: return
+    self.dev.synchronize()
+    if hasattr(self, '_do_free'): self._do_free(buf, options)
+
+  def _unmap(self, mb):
+    self.dev.synchronize()
+    self.dev.iface.free(mb)
+
+  def _offset(self, buf, size:int, offset:int) -> HCQ2Buffer: return buf.offset(offset=offset, size=size)
+
+  def _wrap(self, dev:str, sz:int, opaque:HCQ2Buffer) -> Buffer:
+    return Buffer(dev, sz, dtypes.uint8, opaque=opaque, options=BufferSpec(external_ptr=1))
+
+  def _copy(self, dst:Buffer, src:Buffer):
+    from tinygrad.engine.realize import run_linear
+    du, su = UOp.from_buffer(dst), UOp.from_buffer(src)
+    run_linear(UOp(Ops.LINEAR, src=(su.param_like(1).copy_to_device(dst.device).call(du, su),)), update_stats=True)
+
+  def _copyin(self, dest:HCQ2Buffer, src:memoryview):
+    s = Buffer(self.dev.device, len(src), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
+    s._buf.cpu_view()[:len(src)] = src
+    self._copy(self._wrap(self.dev.device, len(src), dest), s)
+
+  def _copyout(self, dest:memoryview, src:HCQ2Buffer):
+    d = Buffer(self.dev.device, len(dest), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
+    self._copy(d, self._wrap(self.dev.device, len(dest), src))
+    self.dev.synchronize()
+    dest[:] = d._buf.cpu_view()[:len(dest)]
+
+  # def _as_buffer(self, buf): return buf.cpu_view().mv

extra/hcq2/ops_amd2.py

@@ -4,6 +4,7 @@ import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, co
 assert sys.platform != 'win32'
 from dataclasses import dataclass
 from extra.hcq2.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, encode_kernargs_clike, make_getaddr, make_ins, make_cmdbuf, make_placeholder
+from extra.hcq2.hcq2 import make_binary_patch
 from tinygrad.uop.ops import sint, UOp
 from tinygrad.device import Compiled, BufferSpec, Buffer, Device
 from tinygrad.dtype import dtypes
@@ -176,7 +177,7 @@ class SDMAOps(FastEnum): COPY = auto(); POLL_REGMEM = auto(); FENCE = auto(); TR
 
 def sdma_copy(ctx, call):
   dst, src = call.src[1], call.src[2]
-  sz = src.max_numel() * src.dtype.base.itemsize
+  sz = src.max_numel() * src.dtype.itemsize
   src_addr, dst_addr = make_getaddr(src, ctx.devs), make_getaddr(dst, ctx.devs)
   return UOp(Ops.LINEAR, dtypes.void, tuple([make_ins(SDMAOps.COPY,
      ctx.sdma.SDMA_OP_COPY | ctx.sdma.SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_COPY_LINEAR),
@@ -279,7 +280,7 @@ def amd_build_program(prg:UOp) -> UOp:
       kernargs_alloc_size=desc.kernarg_size + (ctypes.sizeof(hsa.hsa_kernel_dispatch_packet_t) if edp else 0), enable_dispatch_ptr=edp,
       enable_private_segment_sgpr=desc.kernel_code_properties & hsa.AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER)
     buf = make_placeholder(prg.device, len(image), dtypes.uint8, "program")
-    cached = _amd_program_cache[key] = prg.replace(src=(buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=bytes(image)))),), arg=(data, prg.arg))
+    cached = _amd_program_cache[key] = prg.replace(src=(buf.after(make_binary_patch(buf, bytes(image))),), arg=(data, prg.arg))
   return cached
 
 class AMDAllocator(HCQAllocator['AMDDevice']):
@@ -579,7 +580,7 @@ class AMDDevice(HCQ2Compiled):
 
     # Scratch setup
     self.max_private_segment_size = 0
-    self.pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, tag="scratch", name="b"), lambda ctx, b: ctx.scratch_buffer(b.max_numel()))]) + self.pm_bufferize
+    self.pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, tag="scratch", name="b"), lambda ctx, b: ctx[0].scratch_buffer(b.max_numel()))]) + self.pm_bufferize
 
     self.pmc_enabled:bool = PROFILE > 0 and PMC > 0
     if self.pmc_enabled:
@@ -629,8 +630,8 @@ class AMDDevice(HCQ2Compiled):
     self.pm_bufferize = PatternMatcher([
       (UPat(Ops.PARAM, tag={(qname, name)}), lambda ctx, b=getattr(queue, name): b) for name in ["ring", "write_ptr", "doorbell", "put_value"]
     ] + [
-      (UPat(Ops.PARAM, tag={(qname, "timeline_signal")}), lambda ctx, q=qname: ctx.timeline_signal(q)),
-      (UPat(Ops.PARAM, tag={(qname, "timeline_value")}), lambda ctx, q=qname: ctx.timeline_value(q)),
+      (UPat(Ops.PARAM, tag={(qname, "timeline_signal")}), lambda ctx, q=qname: ctx[0].timeline_signal(q)),
+      (UPat(Ops.PARAM, tag={(qname, "timeline_value")}), lambda ctx, q=qname: ctx[0].timeline_value(q)),
     ]) + self.pm_bufferize
 
     return queue

extra/hip_large_kernel.py

@@ -18,7 +18,7 @@ prg = dev.runtime("write_ones", mbin)
 prg(buf0._buf, global_size=(1,65537,1), local_size=(1,1,1), wait=True)
 
 import numpy as np
-def to_np(buf): return np.frombuffer(buf.as_memoryview().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))
+def to_np(buf): return np.frombuffer(buf.as_memoryview().cast(buf.dtype.fmt), dtype=_to_np_dtype(buf.dtype))
 
 big = to_np(buf0)
 print(big)

extra/llama_kernels/fused_ce/__init__.py

@@ -16,7 +16,7 @@ def _custom_fused_ce_loss_fwd(loss_out:UOp, max_out:UOp, lse_out:UOp, logits:UOp
   row_lse = (logits[b, s, v_lse].cast(dtypes.float) - row_max).exp().reduce(v_lse, arg=Ops.ADD).log() + row_max
 
   v_smooth = UOp.range(vocab, 3, axis_type=AxisType.REDUCE)
-  target = logits[b, s, targets[row].cast(dtypes.weakint)].cast(dtypes.float)
+  target = logits[b, s, targets[row].cast(dtypes.index)].cast(dtypes.float)
   mean_logits = logits[b, s, v_smooth].cast(dtypes.float).reduce(v_smooth, arg=Ops.ADD) / vocab
   loss = row_lse - (1.0 - label_smoothing) * target - label_smoothing * mean_logits
   stores = UOp.group(loss_out[row].store(loss), max_out[row].store(row_max), lse_out[row].store(row_lse))
@@ -32,11 +32,11 @@ def _custom_fused_ce_loss_bwd(d_logits:UOp, logits:UOp, lse:UOp, targets:UOp, sc
   s = row % seq
 
   prob = (logits[b, s, v].cast(dtypes.float) - lse[row]).exp()
-  target = v.eq(targets[row].cast(dtypes.weakint)).where(1.0 - label_smoothing, 0.0)
+  target = v.eq(targets[row].cast(dtypes.index)).where(1.0 - label_smoothing, 0.0)
   smooth = label_smoothing / vocab
   grad = (prob - target - smooth) * scale[0]
 
-  return d_logits[b, s, v].store(grad.cast(d_logits.dtype.base)).end(v, row).sink(arg=KernelInfo(f"fused_ce_loss_bwd_{rows}_{vocab}"))
+  return d_logits[b, s, v].store(grad.cast(d_logits.dtype)).end(v, row).sink(arg=KernelInfo(f"fused_ce_loss_bwd_{rows}_{vocab}"))
 
 def _fused_ce_loss_bwd(gradient:UOp, kernel:UOp, label_smoothing:float):
   # NOTE: forward inputs are (loss_out, max_out, lse_out, logits, targets)

extra/llama_kernels/fused_silu_mul_quantize_mxfp8/__init__.py

@@ -41,7 +41,7 @@ def _custom_silu_mul_quantize_mxfp8(fp8_out:UOp, e8_out:UOp, si_out:UOp, x_w1:UO
   scaled = (act * qscale).maximum(-FP8_MAX).minimum(FP8_MAX)
   e8u8 = e8f.cast(dtypes.uint8)
 
-  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
+  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype)).end(lane)
   e8_store = e8_out.after(fp8_store)[super_idx * PACK + sb].store(e8u8)
   packed = (e8u8.cast(dtypes.uint32) << (sb.cast(dtypes.uint32) * 8)).reduce(sb, arg=Ops.ADD)
   row, col4 = super_idx // sk4, super_idx % sk4
@@ -72,8 +72,8 @@ def _custom_silu_mul_bwd_mxfp8(gx1_out:UOp, gx3_out:UOp, x_w1:UOp, x_w3:UOp, gra
   sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
   s = w1 * sig
   sprime = sig * (1.0 + w1 * (1.0 - sig))
-  gx1 = gx1_out[idx].store((ga * sprime * w3).cast(gx1_out.dtype.base))
-  gx3 = gx3_out.after(gx1)[idx].store((ga * s).cast(gx3_out.dtype.base))
+  gx1 = gx1_out[idx].store((ga * sprime * w3).cast(gx1_out.dtype))
+  gx3 = gx3_out.after(gx1)[idx].store((ga * s).cast(gx3_out.dtype))
   return gx3.end(lane, tid, wg).sink(arg=KernelInfo(f"silu_mul_bwd_mxfp8_{n_elems}", opts_to_apply=()))
 
 def _silu_mul_quantize_mxfp8_bwd(gradient:UOp, kernel:UOp):

extra/llama_kernels/quantize_fp8_delayed/__init__.py

@@ -27,7 +27,7 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_st
   abs_x = (x_f < 0.0).where(-x_f, x_f)
   scaled = (x_f * scale).maximum(-FP8_MAX).minimum(FP8_MAX)
 
-  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
+  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype)).end(lane)
   lane_max = abs_x.reduce(lane, arg=Ops.MAX)
 
   lmax = UOp.placeholder((1,), dtypes.float, slot=1, addrspace=AddrSpace.REG)
@@ -56,7 +56,7 @@ def _custom_quantize_fp8_scalar(fp8_out:UOp, x:UOp, amax_state:UOp) -> UOp:
 
   x_f = x.reshape(n_elems)[i].cast(dtypes.float)
   scale = FP8_MAX / (amax_state[0].cast(dtypes.float) + 1e-8)
-  store = fp8_out.reshape(n_elems)[i].store((x_f * scale).cast(fp8_out.dtype.base))
+  store = fp8_out.reshape(n_elems)[i].store((x_f * scale).cast(fp8_out.dtype))
 
   return store.end(i).sink(arg=KernelInfo(f"quantize_fp8_scalar_{n_elems}"))
 

extra/llama_kernels/quantize_mxfp8_fused/__init__.py

@@ -38,7 +38,7 @@ def _custom_quantize_mxfp8(fp8_out:UOp, e8_out:UOp, si_out:UOp, x:UOp) -> UOp:
   scaled = (x_f * qscale).maximum(-FP8_MAX).minimum(FP8_MAX)
   e8u8 = e8f.cast(dtypes.uint8)
 
-  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
+  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype)).end(lane)
   e8_store = e8_out.after(fp8_store)[super_idx * PACK + sb].store(e8u8)
 
   # pack the 4 e8 of this super-block into one uint32 (little-endian: byte sb), write transposed (sk4, row)

extra/qcom_gpu_driver/qcom_opencl_interop.py

@@ -78,9 +78,7 @@ hexdump(to_mv(cl_buf_desc_ptr, 0x100))
 rawbuf_ptr = to_mv(cl_buf_desc_ptr, 0x100).cast('Q')[20] # offset 0xA0 is a raw gpu pointer.
 
 # create QCOM tensor with the externally managed buffer
-# dtypes.imageh = cl.cl_image_format(cl.CL_RGBA, cl.CL_HALF_FLOAT)
-# dtypes.imagef = cl.cl_image_format(cl.CL_RGBA, cl.CL_FLOAT)
-x = Tensor.from_blob(rawbuf_ptr, (h*w*4,), dtype=dtypes.imagef((h,w)), device='QCOM')
+x = Tensor.from_blob(rawbuf_ptr, (h,w,4), dtype=dtypes.float, device='QCOM')
 y = (x + 1).tolist()
 print(y[:10])
 

extra/thunder/amd/fa.py

@@ -20,8 +20,8 @@ def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
   return _sharded_empty(ref.shape, ref, axis)
 
 @functools.cache
-def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch):
-  def grad(dou:UOp, ker:UOp) -> tuple[None, None, UOp, UOp, UOp]:
+def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink):
+  def grad(dou:UOp, ker:UOp) -> tuple:
     do = Tensor(dou, device=dou.device)
     attn = Tensor(ker.src[1].after(ker), device=ker.src[1].device)
     l_vec = Tensor(ker.src[2].after(ker), device=ker.src[2].device)
@@ -40,25 +40,33 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
 
     dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:3]
 
-    # unshuffle dq: atomic_pk_add_bf16_with_warpid creates a shuffled layout within each 16x128 tile
-    # decompose each tile into (j=4, a=2, b=2, d=4, e=4, k=4, c=2) and permute to (e, k, j, a, d, b, c) = standard row-major
-    dq = dq.reshape(B, H, N//16, 4, 2, 2, 4, 4, 4, 2).permute(0, 1, 2, 7, 8, 3, 4, 6, 5, 9).reshape(B, H, N, D).transpose(1, 2)
+    if D == 64:
+      dq = dq.reshape(B, H, N//16, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2).permute(0, 1, 2, 8, 9, 10, 11, 3, 4, 6, 7, 5, 12).reshape(B, H, N, D).transpose(1, 2)
+    else:
+      dq = dq.reshape(B, H, N//16, 4, 2, 2, D//32, 4, 4, 2).permute(0, 1, 2, 7, 8, 3, 4, 6, 5, 9).reshape(B, H, N, D).transpose(1, 2)
 
     # reduce partial dK/dV across GROUP_SIZE query heads
     dk = dk_partial.reshape(B, GROUP_SIZE, N, H_KV, D).sum(1)
     dv = dv_partial.reshape(B, GROUP_SIZE, N, H_KV, D).sum(1)
 
-    return None, None, dq.uop, dk.uop, dv.uop
+    if not has_sink: return None, None, dq.uop, dk.uop, dv.uop
+    sinks = Tensor(ker.src[6], device=ker.src[6].device)
+    p_sink = (sinks.reshape(1, H, 1, 1) - l_vec).exp()
+    dsink = -(delta_vec.float() * p_sink).sum(axis=(0, 2, 3))
+
+    return None, None, dq.uop, dk.uop, dv.uop, dsink.uop
   return grad
 
 # TODO: remove write_flat once scheduler can remove reshapes between custom_kernel. TestCustomKernel.test_simple_reshape
-def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False):
+def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None):
   assert attn_mask is None, "attn_mask not supported"
   assert is_causal, "only causal attention supported"
 
   B, N, H, D = xq.shape
   H_KV = xk.shape[2]
-  assert D == 128, "only D=128 supported"
+  assert D in (64, 128), "only D=64 or D=128 supported"
+  has_sink = sinks is not None
+  if has_sink: sinks = sinks.float()
 
   num_devices = len(xq.device) if isinstance(xq.device, tuple) else 1
   is_dp = xq.uop.axis == 0
@@ -77,17 +85,18 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
   attn = _sharded_empty((B, N, H * D), xq, axis=shard_axis) if write_flat else _sharded_empty_like(xq, axis=shard_axis)
   l_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
 
-  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch)
+  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink)
 
-  attn, l_vec = Tensor.custom_kernel(attn, l_vec, xq, xk, xv, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D), grad_fxn=grad)[:2]
+  fwd_inputs = (attn, l_vec, xq, xk, xv) + ((sinks,) if has_sink else ())
+  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, has_sink=has_sink), grad_fxn=grad)[:2]
 
   return attn, attn, l_vec
 
 @functools.cache
-def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True):
   code = (pathlib.Path(__file__).parent / "fa_fwd_causal.cpp").read_text()
   compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}"]
 
   Q_BLOCK_SIZE = 32
   NUM_WARPS = 8
@@ -100,7 +109,8 @@ def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:st
   el = q.dtype.itemsize
   mem = (2*B*N*H*D + 2*B*N*H_KV*D) * el + B*H*N * l_vec.dtype.itemsize
   estimates = Estimates(ops=2*B*H*N*N*D, lds=mem, mem=mem)
-  sink = UOp.sink(o.base, l_vec.base, q.base, k.base, v.base,
+  buf_inputs = (o.base, l_vec.base, q.base, k.base, v.base) + ((sinks.base,) if has_sink else ())
+  sink = UOp.sink(*buf_inputs,
                   threadIdx_x, blockIdx_x, blockIdx_y, blockIdx_z,
                   arg=KernelInfo(name="custom_fa_forward", estimates=estimates))
 
@@ -117,7 +127,7 @@ def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:st
 def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
   code = (pathlib.Path(__file__).parent / "fa_bwd_pre.cpp").read_text()
   compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_D={D}"]
 
   DOT_SLICE_QO = 16
   NUM_WARPS = 4
@@ -147,7 +157,7 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
 def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
   code = (pathlib.Path(__file__).parent / "fa_bwd_causal.cpp").read_text()
   compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}"]
 
   BLOCK_SIZE_KV = 256
   NUM_WARPS = 4
@@ -177,7 +187,7 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
 def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
   code = (pathlib.Path(__file__).parent / "fa_bwd_post.cpp").read_text()
   compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_D={D}"]
 
   DOT_SLICE_QO = 16
   NUM_WARPS = 4

extra/thunder/amd/fa_bwd_post.cpp

@@ -12,7 +12,9 @@ constexpr int ATTN_H = 64; // number of query heads
 constexpr int ATTN_N = 1024; // sequence length
 #endif
 
+#ifndef ATTN_D
 constexpr int ATTN_D = 128; // dimension
+#endif
 constexpr int DOT_SLICE_QO = 16;
 
 #define NUM_WARPS 4

extra/thunder/amd/fa_bwd_pre.cpp

@@ -18,7 +18,9 @@ constexpr int GROUP_SIZE = ATTN_H / ATTN_H_KV; // queries per KV head group
 constexpr int ATTN_N = 1024; // sequence length
 #endif
 
+#ifndef ATTN_D
 constexpr int ATTN_D = 128; // dimension
+#endif
 constexpr int STEP_QO = 64; // block size for QO
 constexpr int BLOCK_SIZE_KV = 256; // block size for KV
 constexpr int SLICE_QO = 32;

extra/thunder/amd/fa_fwd_causal.cpp

@@ -18,7 +18,19 @@ constexpr int GROUP_SIZE = ATTN_H / ATTN_H_KV; // queries per KV head group
 constexpr int ATTN_N = 8192; // sequence length
 #endif
 
+#ifndef ATTN_D
 constexpr int ATTN_D = 128; // dimension
+#endif
+#ifndef ATTN_SINK
+#define ATTN_SINK 0
+#endif
+#if ATTN_D == 64
+#define FA_VM2 "1"
+#define FA_VM4 "2"
+#else
+#define FA_VM2 "2"
+#define FA_VM4 "4"
+#endif
 constexpr int Q_BLOCK_SIZE = 32; // q block size
 constexpr int KV_BLOCK_SIZE = 64; // kv block size
 constexpr bool causal = true;
@@ -156,7 +168,11 @@ template<int D> struct attn_globals {
 };
 
 template<int D> __launch_bounds__(NUM_THREADS, 2)
-__global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_ptr, bf16 *V_ptr) {
+__global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_ptr, bf16 *V_ptr
+#if ATTN_SINK
+    , float *Sinks_ptr
+#endif
+    ) {
     _gl_QKVO Og{O_ptr, ATTN_B, ATTN_N, ATTN_H, ATTN_D};
     _gl_QKVO Qg{Q_ptr, ATTN_B, ATTN_N, ATTN_H, ATTN_D};
     _gl_QKVO Kg{K_ptr, ATTN_B, ATTN_N, ATTN_H_KV, ATTN_D};
@@ -233,7 +249,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     load(k_reg, k_smem[0]);
     __builtin_amdgcn_sched_barrier(0);
     asm volatile("s_waitcnt lgkmcnt(0)");
-    asm volatile("s_waitcnt vmcnt(2)");
+    asm volatile("s_waitcnt vmcnt(" FA_VM2 ")");
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
 
@@ -272,7 +288,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     // All warps then collaboratively load in the second slice of V (V1) into shared memory
     G::load<1, false>(v_smem[1], g.Vg, {batch_idx, 1, head_idx_kv, 0}, swizzled_offsets_V);
     asm volatile("s_waitcnt lgkmcnt(0)");
-    asm volatile("s_waitcnt vmcnt(4)");
+    asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
 
@@ -301,7 +317,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
         //      Load V0 into registers
         load(v_reg, v_smem[0]);
         asm volatile("s_waitcnt lgkmcnt(0)");
-        asm volatile("s_waitcnt vmcnt(4)");
+        asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
         __builtin_amdgcn_sched_barrier(0);
         __builtin_amdgcn_s_barrier();
         __builtin_amdgcn_sched_barrier(0);
@@ -332,7 +348,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
         //      Load K2 into registers
         load(k_reg, k_smem[0]);
         asm volatile("s_waitcnt lgkmcnt(0)");
-        asm volatile("s_waitcnt vmcnt(4)");
+        asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
         __builtin_amdgcn_sched_barrier(0);
         __builtin_amdgcn_s_barrier();
         __builtin_amdgcn_sched_barrier(0);
@@ -368,7 +384,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
             }
         }
         asm volatile("s_waitcnt lgkmcnt(0)");
-        asm volatile("s_waitcnt vmcnt(4)");
+        asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
         __builtin_amdgcn_sched_barrier(0);
         __builtin_amdgcn_s_barrier();
         __builtin_amdgcn_sched_barrier(0);
@@ -399,7 +415,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
         //      Load K3 into registers
         load(k_reg, k_smem[1]);
         asm volatile("s_waitcnt lgkmcnt(0)");
-        asm volatile("s_waitcnt vmcnt(4)");
+        asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
         __builtin_amdgcn_sched_barrier(0);
         __builtin_amdgcn_s_barrier();
         __builtin_amdgcn_sched_barrier(0);
@@ -436,7 +452,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
         }
     }
     asm volatile("s_waitcnt lgkmcnt(0)");
-    asm volatile("s_waitcnt vmcnt(4)");
+    asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
     __builtin_amdgcn_sched_barrier(0);
@@ -467,7 +483,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     //      Load K4 into registers
     load(k_reg, k_smem[0]);
     asm volatile("s_waitcnt lgkmcnt(0)");
-    asm volatile("s_waitcnt vmcnt(4)");
+    asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
     __builtin_amdgcn_sched_barrier(0);
@@ -499,7 +515,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
         }
     }
     asm volatile("s_waitcnt lgkmcnt(0)");
-    asm volatile("s_waitcnt vmcnt(2)");
+    asm volatile("s_waitcnt vmcnt(" FA_VM2 ")");
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
     __builtin_amdgcn_sched_barrier(0);
@@ -529,7 +545,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     //      Load K5 into registers
     load(k_reg, k_smem[1]);
     asm volatile("s_waitcnt lgkmcnt(0)");
-    asm volatile("s_waitcnt vmcnt(2)");
+    asm volatile("s_waitcnt vmcnt(" FA_VM2 ")");
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
     __builtin_amdgcn_sched_barrier(0);
@@ -604,6 +620,16 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     // Cluster 12:
     //      A5V5
     mma_AtB(o_reg, v_reg, att_block_bf16_in, o_reg);
+#if ATTN_SINK
+    {
+        const float sink_l2 = Sinks_ptr[head_idx] * 1.44269504089f;
+        typename attn_tile<float, col_l, rt_32x32_s>::row_vec sink_term;
+        mul(sink_term, max_vec, -1.0f);
+        add(sink_term, sink_term, sink_l2);
+        exp2(sink_term, sink_term);
+        add(norm_vec, norm_vec, sink_term);
+    }
+#endif
     div_col(o_reg, o_reg, norm_vec);
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
@@ -625,4 +651,8 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     store(g.L_vec, norm_vec, {batch_idx, head_idx, 0, tile_idx});
 }
 
-template __global__ void attend_ker<ATTN_D>(bf16*, float*, bf16*, bf16*, bf16*);
+template __global__ void attend_ker<ATTN_D>(bf16*, float*, bf16*, bf16*, bf16*
+#if ATTN_SINK
+    , float*
+#endif
+);

extra/thunder/amd/include/utils.cpp

@@ -23,7 +23,7 @@ __device__ inline static void atomic_pk_add_bf16_with_warpid(const GL &dst, cons
     std::uint64_t  as_u64 = static_cast<std::uint64_t>(as_int);
     buffer_resource br = make_buffer_resource(as_u64, buffer_size, 0x00020000);
 
-    int lane_offset = laneid * 2 + warpid * 512;
+    int lane_offset = laneid * 2 + warpid * (RT::rows * RT::cols);
 
     using range_type = ducks::art::get_nth_range_t<typename RT::register_ranges, N * RT::width + M>;
 
@@ -65,7 +65,7 @@ __device__ inline static void atomic_pk_add_bf16_with_warpid(const GL &dst, cons
     std::uint64_t  as_u64 = static_cast<std::uint64_t>(as_int);
     buffer_resource br = make_buffer_resource(as_u64, buffer_size, 0x00020000);
 
-    int lane_offset = laneid * 2 + warpid * 512;
+    int lane_offset = laneid * 2 + warpid * (RT::rows * RT::cols);
 
     auto perform_atomic_pk_add_bf16_with_warpid = [&]<int N, int M>() {
         using range_type = ducks::art::get_nth_range_t<typename RT::register_ranges, N * RT::width + M>;

extra/thunder/tiny/tk/group.py

@@ -2,7 +2,7 @@ import math
 from typing import cast, Callable
 from tinygrad import dtypes
 from tinygrad.uop.ops import AxisType, UOp, Ops
-from tinygrad.dtype import AddrSpace, PtrDType
+from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import prod
 
 from extra.thunder.tiny.tk import WARP_THREADS
@@ -47,8 +47,8 @@ class Group:
     rngs_for_shape = tuple(self.ker.raw_range(dim) for dim in dst.shape)
 
     src_load = src[*rngs_for_shape]
-    if src.dtype.base != dst.dtype.base:
-      src_load = src_load.cast(dst.dtype.base)
+    if src.dtype != dst.dtype:
+      src_load = src_load.cast(dst.dtype)
     dst_store = dst[*rngs_for_shape].store(src_load).end(*rngs_for_shape)
 
     self.ker.push_store(dst_store, dst)
@@ -62,8 +62,8 @@ class Group:
       for width in self.ker.range(src.shape[-2], track=False):
         for inner in self.ker.range(src.shape[-1], track=False):
           src_load = src[height, width, inner]
-          if src.dtype.base != dst.dtype.base:
-            src_load = src_load.cast(dst.dtype.base)
+          if src.dtype != dst.dtype:
+            src_load = src_load.cast(dst.dtype)
           dst_store = dst[width, height, inner].store(src_load).end(height, width, inner)
 
     self.ker.push_store(dst_store, dst)
@@ -84,13 +84,13 @@ class Group:
       for width in self.ker.range(c.shape[-2], track=False):
         for inner in self.ker.range(a.shape[-2], axis_type=AxisType.REDUCE, track=False):
           if a_base_shape.cols == 16:
-            a_in = UOp.vectorize(*[a[height, inner, i] for i in range(4)])
-            b_in = UOp.vectorize(*[b[inner, width, i] for i in range(4)])
+            a_in = UOp.stack(*[a[height, inner, i] for i in range(4)])
+            b_in = UOp.stack(*[b[inner, width, i] for i in range(4)])
           elif a_base_shape.cols == 32:
-            a_in = UOp.vectorize(*[a[height, inner, i] for i in range(8)])
-            b_in = UOp.vectorize(*[b[inner, width, i] for i in range(8)])
+            a_in = UOp.stack(*[a[height, inner, i] for i in range(8)])
+            b_in = UOp.stack(*[b[inner, width, i] for i in range(8)])
           else: raise NotImplementedError(f"mma_AB not implemented for {a_base_shape.cols=}")
-          d_in = UOp.vectorize(*[c[height, width, i] for i in range(4)])
+          d_in = UOp.stack(*[c[height, width, i] for i in range(4)])
 
           out = UOp(Ops.WMMA, dtypes.float32.vec(4), (a_in, b_in, d_in), arg=wmma_arg)
           c_i = [c[height, width, i].store(out.index(i)) for i in range(4)]
@@ -114,13 +114,13 @@ class Group:
       for width in self.ker.range(c.shape[-2], track=False):
         for inner in self.ker.range(a.shape[-2], axis_type=AxisType.REDUCE, track=False):
           if a_base_shape.cols == 16:
-            a_in = UOp.vectorize(*[a[height, inner, i] for i in range(4)])
-            b_in = UOp.vectorize(*[b[width, inner, i] for i in range(4)])
+            a_in = UOp.stack(*[a[height, inner, i] for i in range(4)])
+            b_in = UOp.stack(*[b[width, inner, i] for i in range(4)])
           elif a_base_shape.cols == 32:
-            a_in = UOp.vectorize(*[a[height, inner, i] for i in range(8)])
-            b_in = UOp.vectorize(*[b[width, inner, i] for i in range(8)])
+            a_in = UOp.stack(*[a[height, inner, i] for i in range(8)])
+            b_in = UOp.stack(*[b[width, inner, i] for i in range(8)])
           else: raise NotImplementedError(f"mma_ABt not implemented for {a_base_shape.cols=}")
-          d_in = UOp.vectorize(*[c[height, width, i] for i in range(4)])
+          d_in = UOp.stack(*[c[height, width, i] for i in range(4)])
 
           out = UOp(Ops.WMMA, dtypes.float32.vec(4), (a_in, b_in, d_in), arg=wmma_arg)
           c_i = [c[height, width, i].store(out.index(i)) for i in range(4)]
@@ -144,13 +144,13 @@ class Group:
       for width in self.ker.range(c.shape[-2], track=False):
         for inner in self.ker.range(a.shape[-3], axis_type=AxisType.REDUCE, track=False):
           if a_base_shape.cols == 16:
-            a_in = UOp.vectorize(*[a[inner, height, i] for i in range(4)])
-            b_in = UOp.vectorize(*[b[inner, width, i] for i in range(4)])
+            a_in = UOp.stack(*[a[inner, height, i] for i in range(4)])
+            b_in = UOp.stack(*[b[inner, width, i] for i in range(4)])
           elif a_base_shape.cols == 32:
-            a_in = UOp.vectorize(*[a[inner, height, i] for i in range(8)])
-            b_in = UOp.vectorize(*[b[inner, width, i] for i in range(8)])
+            a_in = UOp.stack(*[a[inner, height, i] for i in range(8)])
+            b_in = UOp.stack(*[b[inner, width, i] for i in range(8)])
           else: raise NotImplementedError(f"mma_AtB not implemented for {a_base_shape.cols=}")
-          d_in = UOp.vectorize(*[c[height, width, i] for i in range(4)])
+          d_in = UOp.stack(*[c[height, width, i] for i in range(4)])
 
           out = UOp(Ops.WMMA, dtypes.float32.vec(4), (a_in, b_in, d_in), arg=wmma_arg)
           c_i = [c[height, width, i].store(out.index(i)) for i in range(4)]
@@ -174,13 +174,13 @@ class Group:
       for width in self.ker.range(c.shape[-2], track=False):
         for inner in self.ker.range(a.shape[-3], axis_type=AxisType.REDUCE, track=False):
           if a_base_shape.cols == 16:
-            a_in = UOp.vectorize(*[a[inner, height, i] for i in range(4)])
-            b_in = UOp.vectorize(*[b[width, inner, i] for i in range(4)])
+            a_in = UOp.stack(*[a[inner, height, i] for i in range(4)])
+            b_in = UOp.stack(*[b[width, inner, i] for i in range(4)])
           elif a_base_shape.cols == 32:
-            a_in = UOp.vectorize(*[a[inner, height, i] for i in range(8)])
-            b_in = UOp.vectorize(*[b[width, inner, i] for i in range(8)])
+            a_in = UOp.stack(*[a[inner, height, i] for i in range(8)])
+            b_in = UOp.stack(*[b[width, inner, i] for i in range(8)])
           else: raise NotImplementedError(f"mma_AtBt not implemented for {a_base_shape.cols=}")
-          d_in = UOp.vectorize(*[c[height, width, i] for i in range(4)])
+          d_in = UOp.stack(*[c[height, width, i] for i in range(4)])
 
           out = UOp(Ops.WMMA, dtypes.float32.vec(4), (a_in, b_in, d_in), arg=wmma_arg)
           c_i = [c[height, width, i].store(out.index(i)) for i in range(4)]
@@ -209,8 +209,8 @@ class Group:
     vec, src = cast(UOp, vec), cast(UOp, src)
     assert self.warps == 1
 
-    red_local = self.ker.alloc((self.group_threads,), src.dtype.base, AddrSpace.LOCAL)
-    red_reg = self.ker.alloc((1,), src.dtype.base, AddrSpace.REG)
+    red_local = self.ker.alloc((self.group_threads,), src.dtype, AddrSpace.LOCAL)
+    red_reg = self.ker.alloc((1,), src.dtype, AddrSpace.REG)
 
     for height in self.ker.range(src.shape[-3], track=False):
       i = self.ker.raw_range(red_reg.size)
@@ -243,8 +243,8 @@ class Group:
     vec, src = cast(UOp, vec), cast(UOp, src)
     assert self.warps == 1
 
-    red_local = self.ker.alloc((self.group_threads,), src.dtype.base, AddrSpace.LOCAL)
-    red_reg = self.ker.alloc((1,), src.dtype.base, AddrSpace.REG)
+    red_local = self.ker.alloc((self.group_threads,), src.dtype, AddrSpace.LOCAL)
+    red_reg = self.ker.alloc((1,), src.dtype, AddrSpace.REG)
 
     for width in self.ker.range(src.shape[-2], track=False):
       i = self.ker.raw_range(red_reg.size)
@@ -277,9 +277,7 @@ class Group:
 
   def load(self, dst:ALL_TILES, src:ALL_TILES, dst_idxs:tuple[UOp|int,...]=(), idxs:tuple[UOp|int,...]=(), axis:int=0):
     dst, src = cast(UOp, dst), cast(UOp, src)
-    assert isinstance(dst.dtype, PtrDType) and isinstance(src.dtype, PtrDType)
-    dst_dtype, src_dtype = dst.dtype, src.dtype
-    if dst_dtype.addrspace == AddrSpace.REG and src_dtype.addrspace == AddrSpace.LOCAL:
+    if dst.addrspace == AddrSpace.REG and src.addrspace == AddrSpace.LOCAL:
       laneid = self.ker.laneid
       rt, st = cast(RT, dst), cast(ST, src)
       elements_per_thread = rt.base_shape.elements_per_thread
@@ -308,11 +306,11 @@ class Group:
             srow, scol = cast(ST, src).swizzle(row, col)
 
             src_load = src[*idxs[:-2], sheight, swidth, srow, scol]
-            if src.dtype.base != dst.dtype.base:
-              src_load = src_load.cast(dst.dtype.base)
+            if src.dtype != dst.dtype:
+              src_load = src_load.cast(dst.dtype)
             dst_store = dst[*dst_idxs, height, width, inner].store(src_load)
             dst_store = dst_store.end(height, width, inner)
-    elif dst_dtype.addrspace == AddrSpace.LOCAL and src_dtype.addrspace == AddrSpace.GLOBAL:
+    elif dst.addrspace == AddrSpace.LOCAL and src.addrspace == AddrSpace.GLOBAL:
       srcf = src.flatten()
       row_stride = prod(src.shape[axis+1:])
 
@@ -342,11 +340,11 @@ class Group:
           src_i += row * row_stride + col
 
           src_load = srcf[src_i]
-          if src.dtype.base != dst.dtype.base:
-            src_load = src_load.cast(dst.dtype.base)
+          if src.dtype != dst.dtype:
+            src_load = src_load.cast(dst.dtype)
           dst_store = dst[*dst_idxs, height, width, srow, scol].store(src_load)
           dst_store = dst_store.end(height, width, outer, inner).barrier()
-    elif dst_dtype.addrspace == AddrSpace.REG and src_dtype.addrspace == AddrSpace.GLOBAL and isinstance(dst, RT):
+    elif dst.addrspace == AddrSpace.REG and src.addrspace == AddrSpace.GLOBAL and isinstance(dst, RT):
       srcf = src.flatten()
       row_stride = prod(src.shape[axis+1:])
 
@@ -376,10 +374,10 @@ class Group:
             src_i += srow * row_stride + scol
 
             src_load = srcf[src_i]
-            if src.dtype.base != dst.dtype.base:
-              src_load = src_load.cast(dst.dtype.base)
+            if src.dtype != dst.dtype:
+              src_load = src_load.cast(dst.dtype)
             dst_store = dst[*dst_idxs, height, width, inner].store(src_load).end(height, width, inner)
-    elif dst_dtype.addrspace == AddrSpace.REG and src_dtype.addrspace == AddrSpace.GLOBAL and isinstance(dst, RV):
+    elif dst.addrspace == AddrSpace.REG and src.addrspace == AddrSpace.GLOBAL and isinstance(dst, RV):
       srcf = src.flatten()
       row_stride = prod(src.shape[axis+1:])
 
@@ -396,20 +394,18 @@ class Group:
         src_i += outer * reductions + (laneid % reductions)
 
         src_load = srcf[src_i]
-        if src.dtype.base != dst.dtype.base:
-          src_load = src_load.cast(dst.dtype.base)
+        if src.dtype != dst.dtype:
+          src_load = src_load.cast(dst.dtype)
         dst_store = dst[outer, 0].store(src_load).end(outer)
     else:
-      raise NotImplementedError(f"load from {src_dtype.addrspace} to {dst_dtype.addrspace} not implemented for {type(dst)=}")
+      raise NotImplementedError(f"load from {src.addrspace} to {dst.addrspace} not implemented for {type(dst)=}")
 
     self.ker.push_store(dst_store, dst)
     return dst.after(dst_store).reshape(dst.shape)
 
   def store(self, dst:ALL_TILES, src:ALL_TILES, idxs:tuple[UOp|int,...]=(), src_idxs:tuple[UOp|int,...]=(), axis:int=0):
     dst, src = cast(UOp, dst), cast(UOp, src)
-    assert isinstance(dst.dtype, PtrDType) and isinstance(src.dtype, PtrDType)
-    dst_dtype, src_dtype = dst.dtype, src.dtype
-    if src_dtype.addrspace == AddrSpace.REG and dst_dtype.addrspace == AddrSpace.LOCAL:
+    if src.addrspace == AddrSpace.REG and dst.addrspace == AddrSpace.LOCAL:
       laneid = self.ker.laneid
       st, rt = cast(ST, dst), cast(RT, src)
       elements_per_thread = rt.base_shape.elements_per_thread
@@ -427,11 +423,11 @@ class Group:
             srow, scol = cast(ST, dst).swizzle(row, col)
 
             src_load = src[*src_idxs, height, width, inner]
-            if src.dtype.base != dst.dtype.base:
-              src_load = src_load.cast(dst.dtype.base)
+            if src.dtype != dst.dtype:
+              src_load = src_load.cast(dst.dtype)
             dst_store = dst[*idxs[:-2], height, width, srow, scol].store(src_load)
             dst_store = dst_store.end(height, width, inner)
-    elif src_dtype.addrspace == AddrSpace.REG and dst_dtype.addrspace == AddrSpace.GLOBAL and isinstance(src, RT):
+    elif src.addrspace == AddrSpace.REG and dst.addrspace == AddrSpace.GLOBAL and isinstance(src, RT):
       dstf = dst.flatten()
       row_stride = prod(dst.shape[axis+1:])
 
@@ -461,10 +457,10 @@ class Group:
             dst_i += srow * row_stride + scol
 
             src_load = src[*src_idxs, height, width, inner]
-            if src.dtype.base != dst.dtype.base:
-              src_load = src_load.cast(dst.dtype.base)
+            if src.dtype != dst.dtype:
+              src_load = src_load.cast(dst.dtype)
             dst_store = dstf[dst_i].store(src_load).end(height, width, inner)
-    elif src_dtype.addrspace == AddrSpace.REG and dst_dtype.addrspace == AddrSpace.GLOBAL and isinstance(src, RV):
+    elif src.addrspace == AddrSpace.REG and dst.addrspace == AddrSpace.GLOBAL and isinstance(src, RV):
       dstf = dst.flatten()
       row_stride = prod(dst.shape[axis+1:])
 
@@ -481,11 +477,11 @@ class Group:
         dst_i += outer * reductions + (laneid % reductions)
 
         src_load = src[outer, 0]
-        if src.dtype.base != dst.dtype.base:
-          src_load = src_load.cast(dst.dtype.base)
+        if src.dtype != dst.dtype:
+          src_load = src_load.cast(dst.dtype)
         dst_store = dstf[dst_i].store(src_load).end(outer)
     else:
-      raise NotImplementedError(f"store from {src_dtype.addrspace} to {dst_dtype.addrspace} not implemented for {type(src)=}")
+      raise NotImplementedError(f"store from {src.addrspace} to {dst.addrspace} not implemented for {type(src)=}")
 
     self.ker.push_store(dst_store, dst)
     return dst.after(dst_store).reshape(dst.shape)

extra/thunder/tiny/tk/tiles.py

@@ -3,7 +3,7 @@ import functools
 from typing import Callable
 from dataclasses import dataclass
 from tinygrad.dtype import AddrSpace, DType
-from tinygrad.mixin import ElementwiseMixin
+from tinygrad.mixin.elementwise import ElementwiseMixin
 from tinygrad.uop.ops import UOp, Ops
 
 from extra.thunder.tiny.tk import WARP_THREADS
@@ -209,7 +209,7 @@ class ST:
     return cls(uop, rows, cols, layout, base_shape, ker)
 
   def swizzle(self, row, col):
-    swizzled_offset = self.base_shape.swizzle(row, col, self._uop.dtype.base.scalar())
+    swizzled_offset = self.base_shape.swizzle(row, col, self._uop.dtype.scalar())
 
     row = swizzled_offset // self.base_shape.cols
     col = swizzled_offset % self.base_shape.cols

test/amd/test_asm_kernel.py

@@ -16,7 +16,7 @@ from extra.gemm.amd_asm_matmul import Kernel
 
 def custom_add_one(A:UOp) -> UOp:
   A = A.flatten()
-  assert dtypes.is_float(A.dtype.base), f"buffer dtype must be float32, got {A.dtype}"
+  assert dtypes.is_float(A.dtype), f"buffer dtype must be float32, got {A.dtype}"
   threads = UOp.special(A.numel(), "lidx0")
   insts = [
     s_load_b64(s[0:1], s[0:1], soffset=NULL),
@@ -34,9 +34,9 @@ def custom_add_one(A:UOp) -> UOp:
 
 def custom_add_var(A:UOp, B:UOp) -> UOp:
   A,B = A.flatten(), B.flatten()
-  assert A.dtype.base == dtypes.uint32, f"buffer dtype must be uint32, got {A.dtype}"
+  assert A.dtype == dtypes.uint32, f"buffer dtype must be uint32, got {A.dtype}"
   threads = UOp.special(A.numel(), "lidx0")
-  var = UOp.param(2, dtypes.weakint, vmin_vmax=(0, 10), name="var", addrspace=AddrSpace.ALU)
+  var = UOp.param(2, dtypes.index, vmin_vmax=(0, 10), name="var", addrspace=AddrSpace.ALU)
   insts = [
     s_load_b128(s[4:7], s[0:1]),
     s_load_b32(s[8], s[0:1], offset=0x10), # all threads load the same variable

test/backend/test_asm_gemm.py

@@ -9,6 +9,11 @@ from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8, FP8_MAX
 # Use DEV=NULL:HIP:gfx950 to also test the assembly
 def is_cdna4(): return Device[Device.DEFAULT].renderer.target.arch.startswith("gfx950")
 
+def has_hipcc():
+  try: system("hipcc --version")
+  except Exception: return False
+  return True
+
 def run_asm_gemm(a_shape, b_shape, dtype=dtypes.bfloat16, a_shard=None, b_shard=None, gpus:int=1) -> None:
   Tensor.manual_seed(0)
   input_dtype = dtypes.bfloat16 if dtype == FP8_DTYPE else dtype
@@ -20,8 +25,10 @@ def run_asm_gemm(a_shape, b_shape, dtype=dtypes.bfloat16, a_shard=None, b_shard=
   devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(gpus)) if (multi:=gpus>1) else None
 
   if dtype == FP8_DTYPE:
-    a_rand, x_scale, _ = quantize_fp8(a_rand)
-    b_rand, w_scale, _ = quantize_fp8(b_rand)
+    x_scale = Tensor.full((), FP8_MAX, dtype=dtypes.float32, device=devs).contiguous()
+    a_rand, _, _ = quantize_fp8(a_rand.shard(devs, axis=a_shard) if multi else a_rand, amax_state=x_scale)
+    b_rand, w_scale, _ = quantize_fp8(b_rand.T.contiguous())
+    if multi: b_rand, w_scale = b_rand.shard(devs, axis=None if b_shard is None else 1-b_shard), w_scale.to(devs).contiguous()
     grad_amax_state = Tensor.full((), FP8_MAX, dtype=dtypes.float32, device=devs).contiguous()
     with Context(DEBUG=0):
       Tensor.realize(a_rand, x_scale, b_rand, w_scale, grad_amax_state)
@@ -32,24 +39,24 @@ def run_asm_gemm(a_shape, b_shape, dtype=dtypes.bfloat16, a_shard=None, b_shard=
     a_ref, b_ref = a_rand.detach().cast(dtypes.bfloat16), b_rand.detach().cast(dtypes.bfloat16)
   else:
     a_ref, b_ref = a_rand.clone(), b_rand.clone()
-  if multi: a, b = a.shard(devs, axis=a_shard), b.shard(devs, axis=b_shard)
+  if multi and isinstance(a.device, str): a, b = a.shard(devs, axis=a_shard), b.shard(devs, axis=b_shard)
   if dtype == FP8_DTYPE:
-    tst = asm_gemm(a, b, x_scale=x_scale, w_scale=w_scale, grad_amax_state=grad_amax_state)
+    tst = asm_gemm(a, b.T, x_scale=x_scale, w_scale=w_scale, grad_amax_state=grad_amax_state)
   else:
     tst = asm_gemm(a, b)
   tst.sum().backward()
   Tensor.realize(tst, a.grad, b.grad)
 
-  if multi: a_ref, b_ref = a_ref.shard(devs, axis=a_shard), b_ref.shard(devs, axis=b_shard)
+  if multi and isinstance(a_ref.device, str): a_ref, b_ref = a_ref.shard(devs, axis=a_shard), b_ref.shard(devs, axis=b_shard)
   if dtype == FP8_DTYPE:
-    ref = ((a_ref @ b_ref) * x_scale * w_scale).cast(dtypes.bfloat16)
+    ref = ((a_ref @ b_ref.T) * ((x_scale.float() + 1e-8) / FP8_MAX) * w_scale).cast(dtypes.bfloat16)
   else:
     ref = a_ref @ b_ref
   ref.sum().backward()
   Tensor.realize(ref, a_ref.grad, b_ref.grad)
 
   # no validation on the NULL device
-  if a_rand.device.startswith("NULL"): return None
+  if Device.DEFAULT.startswith("NULL"): return None
   atol, rtol = (2e-1, 1e-2) if dtype == dtypes.bfloat16 else (256, 1e-2) if dtype == FP8_DTYPE else (1e-2, 1e-3)
   # allow more rtol for multi because of ALLREDUCE_CAST
   grad_atol, grad_rtol = (16895, 0.125) if dtype == FP8_DTYPE else (atol, 2e-2 if multi else rtol)
@@ -110,21 +117,22 @@ class TestAsmGEMM(unittest.TestCase):
     if not is_cdna4() or not has_hipcc():
       self.skipTest("assembly gemm is only for cdna4")
 
-  def test_tiny(self): verify_asm_gemm(1, 256, 256, 64)
+  def test_tiny(self): verify_asm_gemm(1, 256, 256, 256)
 
   def test_verify_with_numpy(self):
     import numpy as np
-    M, N, K = 256, 256, 64
+    M, N, K = 256, 256, 256
     rng = np.random.default_rng(0)
-    a_np = (rng.random((M, K), dtype=np.float32) - 0.5).astype(np.half)
-    b_np = (rng.random((K, N), dtype=np.float32) - 0.5).astype(np.half)
-    c_np = a_np @ b_np
-    a, b = Tensor(a_np), Tensor(b_np)
+    a_np = (rng.random((M, K), dtype=np.float32) - 0.5).astype(np.float32)
+    b_np = (rng.random((K, N), dtype=np.float32) - 0.5).astype(np.float32)
+    c_np = (a_np.astype(np.float32) @ b_np.astype(np.float32)).astype(np.float32)
+    Tensor.manual_seed(0)
+    a, b = Tensor(a_np).cast(dtypes.bfloat16), Tensor(b_np).cast(dtypes.bfloat16)
     c = asm_gemm(a, b)
     c.realize()
     # no validation on the NULL device
     if a.device.startswith("NULL"): return None
-    np.testing.assert_allclose(c.numpy(), c_np, atol=2e-3, rtol=5e-2)
+    np.testing.assert_allclose(c.numpy(), c_np, atol=2e-1, rtol=1e-2)
 
   def test_unsupported_batch(self):
     with self.assertRaisesRegex(AssertionError, "batch size"):
@@ -141,11 +149,13 @@ class TestAsmGEMM(unittest.TestCase):
       verify_asm_gemm(1, 256, 1000, 256)
 
 # test the Asm GEMM with Llama shapes, only run on the real machine for speed
+
+@unittest.skipUnless(has_hipcc(), "requires hipcc to compile")
 class TestGemmLlama(unittest.TestCase):
-  dtype = dtypes.bfloat16
+  dtype = FP8_DTYPE
 
   def setUp(self):
-    if not is_cdna4() or DEV.interface.startswith("MOCK") or not has_hipcc():
+    if not is_cdna4() or DEV.interface.startswith("MOCK"):
       self.skipTest("very slow on non mi350x")
 
   def test_empty(self): asm_gemm(Tensor.empty(N:=getenv("N", 4096), N, dtype=self.dtype), Tensor.empty(N, N, dtype=self.dtype)).realize()
@@ -172,18 +182,16 @@ class TestGemmLlama(unittest.TestCase):
   def test_gemm_batched(self): verify_asm_gemm(2, 8192, 4096, 4096, dtype=self.dtype)
 
   def test_gemm1(self): verify_asm_gemm(8, 8192, 4096, 14336, dtype=self.dtype, gpus=8)
-  @unittest.skip("disabled, asm in this shape is slower than tinygrad")
   def test_gemm2(self): verify_asm_gemm(8, 8192, 128256, 4096, dtype=self.dtype, gpus=8)
   def test_gemm3(self): verify_asm_gemm(8, 8192, 14336, 4096, dtype=self.dtype, gpus=8)
   def test_gemm4(self): verify_asm_gemm(8, 4096, 14336, 4096, dtype=self.dtype, gpus=8)
   def test_gemm5(self): verify_asm_gemm(8, 4096, 4096, 14336, dtype=self.dtype, gpus=8)
   def test_gemm6(self): verify_asm_gemm(16, 4096, 4096, 14336, dtype=self.dtype, gpus=8)
-  @unittest.skip("disabled, asm in this shape is slower than tinygrad")
   def test_gemm7(self): verify_asm_gemm(1, 8192, 128256, 4096, dtype=self.dtype)
   def test_gemm8(self): verify_asm_gemm(1, 4096, 14336, 8192, dtype=self.dtype)
   def test_gemm9(self): verify_asm_gemm(8, 4096, 14336, 8192, dtype=self.dtype, gpus=8)
   def test_gemm10(self): verify_asm_gemm(1, 4096, 8192, 4096, dtype=self.dtype)
-  def test_gemm_previously_unsupported(self): verify_asm_gemm(8, 1024, 1024, 4096, gpus=8)
+  def test_gemm11(self): verify_asm_gemm(8, 1024, 1024, 4096, dtype=self.dtype, gpus=8)
   def test_k_sharded_1(self): verify_asm_gemm_k_sharded(14336, 4096, 8*8192, dtype=self.dtype, gpus=8)
   def test_k_sharded_2(self): verify_asm_gemm_k_sharded(4096, 14336, 8*8192, dtype=self.dtype, gpus=8)
   def test_k_sharded_3(self): verify_asm_gemm_k_sharded(4096, 4096, 8*8192, dtype=self.dtype, gpus=8)
@@ -203,33 +211,25 @@ class TestGemmLlama(unittest.TestCase):
   def test_tp_k_sharded_w2(self): verify_asm_gemm_k_sharded_3d(1, 8192, 4096, 14336, dtype=self.dtype, gpus=8)
 
   # more shapes: vary M, N, K independently
-  def test_shape_small_square(self): verify_asm_gemm(1, 256, 256, 256)
-  def test_shape_small_rect_m(self): verify_asm_gemm(1, 512, 256, 256)
-  def test_shape_small_rect_n(self): verify_asm_gemm(1, 256, 512, 256)
-  def test_shape_small_rect_k(self): verify_asm_gemm(1, 256, 256, 512)
-  def test_shape_tall(self): verify_asm_gemm(1, 2048, 256, 256)
-  def test_shape_wide(self): verify_asm_gemm(1, 256, 2048, 256)
-  def test_shape_deep(self): verify_asm_gemm(1, 256, 256, 4096)
-  def test_shape_non_square(self): verify_asm_gemm(1, 1024, 2048, 512)
-  def test_shape_batched_small(self): verify_asm_gemm(2, 256, 256, 256)
-  def test_shape_batched_rect(self): verify_asm_gemm(2, 512, 1024, 256)
-  # K edge cases: iters=1,2,3 exercise different loop paths
-  def test_shape_k64(self): verify_asm_gemm(1, 256, 256, 64)
-  def test_shape_k128(self): verify_asm_gemm(1, 256, 256, 128)
-  def test_shape_k192(self): verify_asm_gemm(1, 256, 256, 192)
+  def test_shape_small_square(self): verify_asm_gemm(1, 256, 256, 256, dtype=self.dtype)
+  def test_shape_small_rect_m(self): verify_asm_gemm(1, 512, 256, 256, dtype=self.dtype)
+  def test_shape_small_rect_n(self): verify_asm_gemm(1, 256, 512, 256, dtype=self.dtype)
+  def test_shape_small_rect_k(self): verify_asm_gemm(1, 256, 256, 512, dtype=self.dtype)
+  def test_shape_tall(self): verify_asm_gemm(1, 2048, 256, 256, dtype=self.dtype)
+  def test_shape_wide(self): verify_asm_gemm(1, 256, 2048, 256, dtype=self.dtype)
+  def test_shape_deep(self): verify_asm_gemm(1, 256, 256, 4096, dtype=self.dtype)
+  def test_shape_non_square(self): verify_asm_gemm(1, 1024, 2048, 512, dtype=self.dtype)
+  def test_shape_batched_small(self): verify_asm_gemm(2, 256, 256, 256, dtype=self.dtype)
+  def test_shape_batched_rect(self): verify_asm_gemm(2, 512, 1024, 256, dtype=self.dtype)
+  # K edge cases: change iters to exercise different loop paths, k big enough for hk kernel
+  def test_shape_k256(self): verify_asm_gemm(1, 256, 256, 256, dtype=self.dtype)
+  def test_shape_k512(self): verify_asm_gemm(1, 256, 256, 512, dtype=self.dtype)
+  def test_shape_k768(self): verify_asm_gemm(1, 256, 256, 768, dtype=self.dtype)
 
   def test_llama3_out1(self): verify_asm_gemm(1, 8192, 128256, 4096, dtype=self.dtype)
   def test_llama3_out2(self): verify_asm_gemm(1, 8192, 4096, 128256, dtype=self.dtype)
   def test_llama3_out3(self): verify_asm_gemm(1, 4096, 128256, 8192, dtype=self.dtype)
 
-def has_hipcc():
-  try: system("hipcc --version")
-  except Exception: return False
-  return True
-
-@unittest.skipUnless(has_hipcc(), "FP8 gemm requires hipcc to compile")
-class TestGemmLlamaFP8(TestGemmLlama): dtype = FP8_DTYPE
-
 # mxfp8: 1x32 block scaling along K, e8m0 scales packed iteration-major (K/128, dim) uint32
 def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
   rows, K = x.shape
@@ -327,7 +327,7 @@ def run_mx_prequant(M:int, N:int, K:int) -> None:
     err = ((t.float() - r.float()).abs().mean() / (r.float().abs().mean() + 1e-8)).item()
     assert err < 6e-2, f"{name} prequant vs analytic rel err {err}"
 
-@unittest.skipUnless(has_hipcc(), "MXFP8 gemm requires hipcc to compile")
+@unittest.skipUnless(has_hipcc(), "requires hipcc to compile")
 class TestGemmMXFP8(unittest.TestCase):
   def setUp(self):
     if not is_cdna4() or DEV.interface.startswith("MOCK"): self.skipTest("mxfp8 gemm is only for cdna4")
@@ -368,7 +368,7 @@ def run_atb_gemm(rows, M, N, a_shard=None, b_shard=None, gpus=1, atol=1.0, rtol=
   out = hk_bf16_atb_gemm(a, b)
   np.testing.assert_allclose(out.float().numpy(), ref.numpy(), atol=atol, rtol=rtol)
 
-@unittest.skipUnless(has_hipcc(), "MXFP8 gemm requires hipcc to compile")
+@unittest.skipUnless(has_hipcc(), "requires hipcc to compile")
 class TestHkBf16AtbGemm(unittest.TestCase):
   def setUp(self):
     if not is_cdna4(): self.skipTest("hk bf16 atb gemm is cdna4 only")

test/backend/test_custom_kernel.py

@@ -7,12 +7,12 @@ from tinygrad.uop.ops import KernelInfo, AxisType, Ops
 
 def custom_arange_kernel(C:UOp) -> UOp:
   i = UOp.range(C.shape[0], 0)
-  return C[i].store(i.cast(C.dtype.base)).end(i).sink(arg=KernelInfo(name=f"custom_arange_{C.shape[0]}"))
+  return C[i].store(i.cast(C.dtype)).end(i).sink(arg=KernelInfo(name=f"custom_arange_{C.shape[0]}"))
 
 def custom_eye_kernel(C:UOp) -> UOp:
   i = UOp.range(C.shape[0], 0)
   j = UOp.range(C.shape[1], 1)
-  return C[i, j].store((i.eq(j)).cast(C.dtype.base)).end(i, j).sink(arg=KernelInfo(name=f"custom_eye_{C.numel()}"))
+  return C[i, j].store((i.eq(j)).cast(C.dtype)).end(i, j).sink(arg=KernelInfo(name=f"custom_eye_{C.numel()}"))
 
 def custom_add_one_kernel(B:UOp, A:UOp) -> UOp:
   A,B = A.flatten(), B.flatten()
@@ -57,7 +57,7 @@ def flip_contract_kernel(dest:UOp, src:UOp):
 def slice_sum_kernel(dest:UOp, src:UOp):
   G = UOp.range(src.shape[0], 0)
   slice_src = src[G, :]
-  reg = UOp.placeholder((1,), dest.dtype.base, 0, addrspace=AddrSpace.REG)
+  reg = UOp.placeholder((1,), dest.dtype, 0, addrspace=AddrSpace.REG)
   reg = reg.after(G)[0].set(0)
   R = UOp.range(src.shape[1], 1, AxisType.REDUCE)
   reg = reg[0].set(reg.after(R)[0] + slice_src[R], end=R)
@@ -73,12 +73,12 @@ def simple_qkv_kernel(O:UOp, Q:UOp, K:UOp, V:UOp) -> UOp:
   j = UOp.range(N, 2, axis_type=AxisType.REDUCE)
 
   k_inner = UOp.range(d, 3, axis_type=AxisType.REDUCE)
-  qk_acc = UOp.placeholder((1,), Q.dtype.base, 0, addrspace=AddrSpace.REG)
+  qk_acc = UOp.placeholder((1,), Q.dtype, 0, addrspace=AddrSpace.REG)
   qk_acc = qk_acc.after(i, j)[0].set(0.0)
   qk_acc = qk_acc[0].set(qk_acc.after(k_inner)[0] + Q[i, k_inner] * K[j, k_inner], end=k_inner)
   qk_score = qk_acc[0] / (d ** 0.5)
 
-  out_acc = UOp.placeholder((1,), Q.dtype.base, 1, addrspace=AddrSpace.REG)
+  out_acc = UOp.placeholder((1,), Q.dtype, 1, addrspace=AddrSpace.REG)
   out_acc = out_acc.after(i, d_out)[0].set(0.0)
   out_acc = out_acc[0].set(out_acc.after(j)[0] + qk_score * V[j, d_out], end=j)
 

test/backend/test_dtype.py

@@ -281,7 +281,7 @@ class TestBitCast(unittest.TestCase):
   def test_shape_change_bitcast_exceptions(self):
     with self.assertRaises(RuntimeError):
       # should fail because 3 int8 is 3 bytes but float16 is two and 3 isn't a multiple of 2
-      Tensor.empty((3,), dtype=dtypes.int8).bitcast(dtypes.float16)
+      Tensor.empty((3,), dtype=dtypes.int8).bitcast(dtypes.float16).shape
 
   def test_bitcast_float_to_int32(self):
     a = Tensor([1.,2,3])

test/backend/test_graph.py

@@ -43,9 +43,12 @@ def get_buf_uop(buf:Buffer, cache:dict[Buffer,UOp]) -> UOp:
     buffers[u] = buf
   return cache[buf]
 
+def copy_call(dst:Buffer, src:Buffer, c:dict[Buffer,UOp]) -> UOp:
+  return get_buf_uop(src,c).copy_to_device(dst.device).call(get_buf_uop(dst,c), get_buf_uop(src,c))
+
 def make_graph(graph_cls, calls:list[UOp]):
   linear = compile_linear(UOp(Ops.LINEAR, src=tuple(calls)))
-  cf = UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(linear,), arg="graph")
+  cf = UOp(Ops.CUSTOM_FUNCTION, src=(linear,), arg="graph")
   return graph_cls(cf, [])
 
 def run_schedule(calls:list[UOp]):
@@ -73,8 +76,8 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[3],c), get_buf_uop(b[4],c), metadata=()),
+      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c)),
+      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[3],c), get_buf_uop(b[4],c)),
     ]
 
     zero_bufs([b[0]])
@@ -92,8 +95,8 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(b[1],c), get_buf_uop(b[3],c), get_buf_uop(b[4],c), metadata=()),
+      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c)),
+      get_ast(d0, 2).call(get_buf_uop(b[1],c), get_buf_uop(b[3],c), get_buf_uop(b[4],c)),
     ]
 
     zero_bufs([b[0], b[1]])
@@ -111,8 +114,8 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(b[1],c), get_buf_uop(b[0],c), get_buf_uop(b[4],c), metadata=()),
+      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c)),
+      get_ast(d0, 2).call(get_buf_uop(b[1],c), get_buf_uop(b[0],c), get_buf_uop(b[4],c)),
     ]
 
     zero_bufs([b[0], b[1]])
@@ -131,8 +134,8 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c), metadata=()),
-      UOp(Ops.COPY).call(get_buf_uop(b[3],c), get_buf_uop(b[0],c), metadata=()),
+      get_ast(d0, 2).call(get_buf_uop(b[0],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c)),
+      copy_call(b[3], b[0], c),
     ]
 
     zero_bufs([b[0], b[3]])
@@ -151,8 +154,8 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      UOp(Ops.COPY).call(get_buf_uop(b[1],c), get_buf_uop(b[0],c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(b[3],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c), metadata=()),
+      copy_call(b[1], b[0], c),
+      get_ast(d0, 2).call(get_buf_uop(b[3],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c)),
     ]
 
     zero_bufs([b[1], b[3]])
@@ -169,9 +172,9 @@ class TestGraph(unittest.TestCase):
     b = [make_buffer(d0, fill=True) for _ in range(8)]
     c: dict[Buffer,UOp] = {}
 
-    calls1 = [get_ast(d0, 2).call(get_buf_uop(b[3],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c), metadata=())]
-    calls2 = [get_ast(d0, 2).call(get_buf_uop(b[4],c), get_buf_uop(b[1],c), get_buf_uop(b[3],c), metadata=())]
-    calls3 = [get_ast(d0, 2).call(get_buf_uop(b[5],c), get_buf_uop(b[4],c), get_buf_uop(b[2],c), metadata=())]
+    calls1 = [get_ast(d0, 2).call(get_buf_uop(b[3],c), get_buf_uop(b[1],c), get_buf_uop(b[2],c))]
+    calls2 = [get_ast(d0, 2).call(get_buf_uop(b[4],c), get_buf_uop(b[1],c), get_buf_uop(b[3],c))]
+    calls3 = [get_ast(d0, 2).call(get_buf_uop(b[5],c), get_buf_uop(b[4],c), get_buf_uop(b[2],c))]
 
     out = [b[3], b[4], b[5]]
     zero_bufs(out)
@@ -194,8 +197,8 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      UOp(Ops.COPY).call(get_buf_uop(b1[0],c), get_buf_uop(b0[0],c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(b0[2],c), get_buf_uop(b0[0],c), get_buf_uop(b0[1],c), metadata=()),
+      copy_call(b1[0], b0[0], c),
+      get_ast(d0, 2).call(get_buf_uop(b0[2],c), get_buf_uop(b0[0],c), get_buf_uop(b0[1],c)),
     ]
 
     out = [b1[0], b0[2]]
@@ -219,8 +222,8 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      UOp(Ops.COPY).call(get_buf_uop(b0,c), get_buf_uop(b2,c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(b1,c), get_buf_uop(b0,c), get_buf_uop(b2,c), metadata=()),
+      copy_call(b0, b2, c),
+      get_ast(d0, 2).call(get_buf_uop(b1,c), get_buf_uop(b0,c), get_buf_uop(b2,c)),
     ]
 
     zero_bufs([b0])
@@ -245,9 +248,9 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      UOp(Ops.COPY).call(get_buf_uop(base,c), get_buf_uop(copy_src_full,c), metadata=()),
-      UOp(Ops.COPY).call(get_buf_uop(v_lo,c), get_buf_uop(copy_src_lo,c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(out,c), get_buf_uop(v_hi,c), get_buf_uop(a,c), metadata=()),
+      copy_call(base, copy_src_full, c),
+      copy_call(v_lo, copy_src_lo, c),
+      get_ast(d0, 2).call(get_buf_uop(out,c), get_buf_uop(v_hi,c), get_buf_uop(a,c)),
     ]
 
     zero_bufs([base, out])
@@ -272,9 +275,9 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      UOp(Ops.COPY).call(get_buf_uop(copy_dst,c), get_buf_uop(base,c), metadata=()),
-      UOp(Ops.COPY).call(get_buf_uop(v_lo,c), get_buf_uop(copy_src_lo,c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(v_hi,c), get_buf_uop(a,c), get_buf_uop(b,c), metadata=()),
+      copy_call(copy_dst, base, c),
+      copy_call(v_lo, copy_src_lo, c),
+      get_ast(d0, 2).call(get_buf_uop(v_hi,c), get_buf_uop(a,c), get_buf_uop(b,c)),
     ]
 
     zero_bufs([copy_dst, base])
@@ -299,10 +302,10 @@ class TestGraph(unittest.TestCase):
     c: dict[Buffer,UOp] = {}
 
     calls = [
-      UOp(Ops.COPY).call(get_buf_uop(base,c), get_buf_uop(copy_src_full,c), metadata=()),
-      UOp(Ops.COPY).call(get_buf_uop(v_mid,c), get_buf_uop(copy_src_mid,c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(out1,c), get_buf_uop(v_lo,c), get_buf_uop(a,c), metadata=()),
-      get_ast(d0, 2).call(get_buf_uop(out2,c), get_buf_uop(v_hi,c), get_buf_uop(a,c), metadata=()),
+      copy_call(base, copy_src_full, c),
+      copy_call(v_mid, copy_src_mid, c),
+      get_ast(d0, 2).call(get_buf_uop(out1,c), get_buf_uop(v_lo,c), get_buf_uop(a,c)),
+      get_ast(d0, 2).call(get_buf_uop(out2,c), get_buf_uop(v_hi,c), get_buf_uop(a,c)),
     ]
 
     outs = [base, out1, out2]

test/backend/test_isel.py

@@ -33,93 +33,14 @@ class TestIselX86(unittest.TestCase):
     # both comparisons become the same instruction
     self.assertTrue(n.src[0].src[2] == n.src[1].src[2] and n.src[0].src[2].arg is X86Ops.CMP)
 
-  def test_vmax(self):
-    dt_op = [(dtypes.float32, X86Ops.VMAXSS), (dtypes.float64, X86Ops.VMAXSD),
-             (dtypes.float32.vec(4), X86Ops.VMAXPS), (dtypes.float64.vec(4), X86Ops.VMAXPD)]
-    self._check_op(dt_op, lambda a,b: (a < b).where(b, a))
-
-  def test_vmin(self):
-    dt_op = [(dtypes.float32, X86Ops.VMINSS), (dtypes.float64, X86Ops.VMINSD),
-             (dtypes.float32.vec(4), X86Ops.VMINPS), (dtypes.float64.vec(4), X86Ops.VMINPD)]
-    self._check_op(dt_op, lambda a,b: (a < b).where(a, b))
-
-  def test_vfmadd(self):
-    dt_op = [(dtypes.float32, X86Ops.VFMADD213SS), (dtypes.float64, X86Ops.VFMADD213SD),
-             (dtypes.float32.vec(4), X86Ops.VFMADD213PS), (dtypes.float64.vec(4), X86Ops.VFMADD213PD)]
-    self._check_op(dt_op, lambda a,b,c: a * b + c)
-
-  # don't use fmadd if op being fused (mul) is used multiple times
-  def test_no_vfmadd(self):
-    dt_op = [(dtypes.float32, X86Ops.VADDSS), (dtypes.float64, X86Ops.VADDSD),
-             (dtypes.float32.vec(4), X86Ops.VADDPS), (dtypes.float64.vec(4), X86Ops.VADDPD)]
-    self._check_op(dt_op, lambda a,b: a * b + a * b)
-
-  def test_vpbroadcast(self):
-    a = UOp.variable("a", 0, 0, dtypes.int32)
-    n = self.isel_rewrite(a.broadcast(4))
-    # need to move src from gpr to xmm before broadcasting
-    self.assertTrue(n.arg is X86Ops.VPBROADCASTD and n.src[0].arg is X86Ops.VMOVD)
-    # if we can fuse a load we can skip the move and access memory directly
-    load = UOp.param(0, dtypes.int32, (16,)).index(UOp.const(dtypes.int32, 0)).load()
-    n = self.isel_rewrite(load.broadcast(4))
-    self.assertTrue(n.arg is X86Ops.VPBROADCASTD and len(n.src) == 4)
-
-  def test_vbroadcastss(self):
-    a = UOp.variable("a", 0, 0, dtypes.float32)
-    valid = [UOp.vectorize(a, a, a, a), UOp.vectorize(a, a, a, a, a, a, a, a)]
-    for shuf in valid: self.assertIs(self.isel_rewrite(shuf).arg, X86Ops.VBROADCASTSS)
-
-  def test_vshufps(self):
-    a = UOp.variable("a", 0, 0, dtypes.float32.vec(8))
-    b = UOp.variable("b", 0, 0, dtypes.float32.vec(8))
-    c = UOp.variable("c", 0, 0, dtypes.float32)
-    d = UOp.variable("d", 0, 0, dtypes.float32)
-
-    valid = [UOp.vectorize(c, c, d, d),
-             UOp.vectorize(lane(a, 0), lane(a, 1), c, c),
-             UOp.vectorize(lane(a, 0), lane(a, 1), lane(b, 2), lane(b, 3)),
-             UOp.vectorize(lane(a, 1), lane(a, 2), lane(a, 3), lane(a, 0)),
-             UOp.vectorize(lane(a, 3), lane(a, 2), lane(a, 1), lane(a, 0), lane(a, 7), lane(a, 6), lane(a, 5), lane(a, 4)),
-             UOp.vectorize(lane(a, 0), lane(a, 0), lane(b, 1), lane(b, 1), lane(a, 4), lane(a, 4), lane(b, 5), lane(b, 5))]
-    for shuf in valid: self.assertIs(self.isel_rewrite(shuf).arg, X86Ops.VSHUFPS)
-
-    invalid = [UOp.vectorize(lane(a, 0), lane(a, 1), lane(b, 4), lane(b, 5)),
-               UOp.vectorize(lane(a, 0), lane(a, 5), lane(b, 2), lane(b, 3)),
-               UOp.vectorize(lane(a, 0), lane(a, 0), lane(a, 0), lane(a, 0), lane(a, 4), lane(a, 4), lane(a, 4), lane(a, 5)),
-               UOp.vectorize(lane(a, 0), lane(a, 0), lane(b, 0), lane(b, 0), lane(a, 4), lane(a, 4), lane(b, 4), lane(a, 4))]
-    for shuf in invalid: self.assertIsNot(self.isel_rewrite(shuf).arg, X86Ops.VSHUFPS)
-
-  def test_vshufpd(self):
-    a = UOp.variable("a", 0, 0, dtypes.float64.vec(4))
-    b = UOp.variable("b", 0, 0, dtypes.float64.vec(4))
-    c = UOp.variable("c", 0, 0, dtypes.float64)
-    d = UOp.variable("d", 0, 0, dtypes.float64)
-
-    valid = [UOp.vectorize(c, d),
-             UOp.vectorize(lane(a, 0), c),
-             UOp.vectorize(lane(a, 1), lane(b, 1)),
-             UOp.vectorize(lane(a, 0), lane(b, 1), lane(a, 2), lane(b, 3)),
-             UOp.vectorize(lane(a, 1), lane(a, 1), lane(a, 3), lane(a, 3))]
-    for shuf in valid: self.assertIs(self.isel_rewrite(shuf).arg, X86Ops.VSHUFPD)
-
-    invalid = [UOp.vectorize(c, c, c, c),
-               UOp.vectorize(lane(a, 0), lane(a, 1), lane(b, 2), lane(b, 3)),
-               UOp.vectorize(lane(a, 2), lane(b, 3), lane(a, 2), lane(b, 3)),
-               UOp.vectorize(lane(a, 0), lane(b, 1), lane(a, 0), lane(b, 1))]
-    for shuf in invalid: self.assertIsNot(self.isel_rewrite(shuf).arg, X86Ops.VSHUFPD)
-
   def test_vinsertps(self):
     a = UOp.variable("a", 0, 0, dtypes.float32.vec(4))
     b = UOp.variable("b", 0, 0, dtypes.float32.vec(4))
     c = UOp.variable("c", 0, 0, dtypes.float32.vec(4))
     d = UOp.variable("e", 0, 0, dtypes.float32)
-    # moving 0th element to position 0 does nothing so only 1 vinsertps is generated
-    n = self.isel_rewrite(UOp.vectorize(lane(a, 0), d))
-    self.assertIs(n.arg, X86Ops.VINSERTPS)
-    self.assertIsNot(n.src[0].arg, X86Ops.VINSERTPS)
 
-    valid = [UOp.vectorize(lane(a, 0), lane(b, 1), lane(a, 2), lane(b, 3)),
-             UOp.vectorize(lane(a, 3), lane(b, 2), lane(c, 1), d)]
+    valid = [UOp.stack(lane(a, 0), lane(b, 1), lane(a, 2), lane(b, 3)),
+             UOp.stack(lane(a, 3), lane(b, 2), lane(c, 1), d)]
     for shuf in valid: self.assertIs(self.isel_rewrite(shuf).arg, X86Ops.VINSERTPS)
 
   # complex address is [base + index*scale + displacement]
@@ -130,21 +51,5 @@ class TestIselX86(unittest.TestCase):
     # displacement is the constant in "a" scaled to the buffer element size, dtype is int8 when the value fits otherwise int32
     self.assertTrue(n.src[2].op is Ops.CONST and n.src[2].dtype is dtypes.int8 and n.src[2].arg == 4)
 
-  def test_fold_load(self):
-    load1 = UOp.param(0, dtypes.int32, (16,)).index(UOp.const(dtypes.int32, 0)).load()
-    load2 = UOp.param(0, dtypes.int32, (16,)).index(UOp.const(dtypes.int32, 1)).load()
-    n = self.isel_rewrite(load1 + load2)
-    self.assertTrue(len(n.src) == 5)
-
-  # don't fold when used multiple times
-  def test_dont_fold_load(self):
-    load = UOp.param(0, dtypes.int32, (16,)).index(UOp.const(dtypes.int32, 0)).load()
-    # used by multiple users
-    n = self.isel_rewrite(load + 1 + load)
-    self.assertTrue(len(n.src) == 2)
-    # used mutiple times by same user
-    n = self.isel_rewrite(load * load)
-    self.assertTrue(len(n.src) == 2)
-
 if __name__ == "__main__":
   unittest.main()

test/backend/test_linearizer.py

@@ -211,14 +211,14 @@ class TestLinearizer(unittest.TestCase):
         realized_ast = a.schedule_linear().src[-1].src[0]
         program = to_program(replace_opts(realized_ast, []), renderer=Device[Device.DEFAULT].renderer)
         local = [uop for uop in tuple(program.src[1].src) if uop.op is Ops.BUFFER and uop.addrspace in (AddrSpace.LOCAL, AddrSpace.REG)]
-        assert local[0].dtype.base == acc_dtype
+        assert local[0].dtype == acc_dtype
 
   def test_arg_acc_dtype(self):
     def helper_arg_acc_dtype(c: Tensor, expected_dtype:DType):
       realized_ast = c.schedule_linear().src[-1].src[0]
       program = to_program(replace_opts(realized_ast, []), renderer=Device[Device.DEFAULT].renderer)
       local = [uop for uop in tuple(program.src[1].src) if uop.op is Ops.BUFFER and uop.addrspace in (AddrSpace.LOCAL, AddrSpace.REG)]
-      self.assertEqual(local[0].dtype.base, expected_dtype)
+      self.assertEqual(local[0].dtype, expected_dtype)
 
     tests = (
       (dtypes.float16, None, dtypes.float),
@@ -251,9 +251,9 @@ class TestLinearizer(unittest.TestCase):
     for u in uops:
       if u.op is Ops.STORE and u.src[0].addrspace is AddrSpace.REG:
         if uops.index(u) < begin_range:
-          assert u.src[1].op is Ops.CONST or (u.src[1].op is Ops.STACK and all(x.op is Ops.CONST for x in u.src[1].src))
+          assert u.src[1].op is Ops.CONST
         else:
-          assert u.src[1].op is Ops.CONST or (u.src[1].op is Ops.STACK and all(x.op in GroupOp.ALU for x in u.src[1].src))
+          assert u.src[1].op in GroupOp.ALU
           assert begin_range < uops.index(u) < end_range
       # children of END are placed after ENDRANGE
       if any(x.op is Ops.END and x.src[1].op in GroupOp.ALU for x in u.src):

test/backend/test_linearizer_dumb.py

@@ -12,18 +12,18 @@ class TestLinearizerFailure(unittest.TestCase):
   @unittest.skipUnless(Device.DEFAULT == "METAL", "only tested on METAL")
   def test_failure_beam_mnist(self):
     c0 = UOp.param(0, dtypes.uchar, (4014080,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 0, AxisType.GLOBAL)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 784), 1, AxisType.GLOBAL)
-    c3 = UOp.range(UOp.const(dtypes.weakint, 10), 3, AxisType.GLOBAL)
+    c1 = UOp.range(UOp.const(dtypes.index, 512), 0, AxisType.GLOBAL)
+    c2 = UOp.range(UOp.const(dtypes.index, 784), 1, AxisType.GLOBAL)
+    c3 = UOp.range(UOp.const(dtypes.index, 10), 3, AxisType.GLOBAL)
     c4 = UOp.param(1, dtypes.int, (512,))
     c5 = c4.index(c1.valid(UOp.const(dtypes.bool, True)))
-    c6 = UOp.range(UOp.const(dtypes.weakint, 6000), 1004, AxisType.REDUCE)
-    c7 = UOp.range(UOp.const(dtypes.weakint, 3750), 2006, AxisType.REDUCE)
-    c8 = UOp.range(UOp.const(dtypes.weakint, 16), 2007, AxisType.GROUP_REDUCE)
+    c6 = UOp.range(UOp.const(dtypes.index, 6000), 1004, AxisType.REDUCE)
+    c7 = UOp.range(UOp.const(dtypes.index, 3750), 2006, AxisType.REDUCE)
+    c8 = UOp.range(UOp.const(dtypes.index, 16), 2007, AxisType.GROUP_REDUCE)
     c9 = UOp.param(2, dtypes.uchar, (47040000,))
-    c10 = c9.index((((c3*UOp.const(dtypes.weakint, 4704000))+c2)+(c6*UOp.const(dtypes.weakint, 784))).valid(UOp.const(dtypes.bool, True)))
-    c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(dtypes.weakint, 6000))+c6)+((c7*UOp.const(dtypes.weakint, 16))+c8)).alu(Ops.CMPLT, UOp.const(dtypes.weakint, 59999)).where(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(dtypes.int, -1))).where(UOp.const(dtypes.uchar, 0), c10).reduce(c6, arg=Ops.ADD)
-    c12 = c0.index((((c1*UOp.const(dtypes.weakint, 7840))+(c2*UOp.const(dtypes.weakint, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11).end(c1, c2, c3)
+    c10 = c9.index((((c3*UOp.const(dtypes.index, 4704000))+c2)+(c6*UOp.const(dtypes.index, 784))).valid(UOp.const(dtypes.bool, True)))
+    c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(dtypes.index, 6000))+c6)+((c7*UOp.const(dtypes.index, 16))+c8)).alu(Ops.CMPLT, UOp.const(dtypes.index, 59999)).where(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(dtypes.int, -1))).where(UOp.const(dtypes.uchar, 0), c10).reduce(c6, arg=Ops.ADD)
+    c12 = c0.index((((c1*UOp.const(dtypes.index, 7840))+(c2*UOp.const(dtypes.index, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11).end(c1, c2, c3)
     ast = c12.sink(arg=KernelInfo(name='test', axis_types=(), dont_use_locals=False, applied_opts=(Opt(op=OptOps.GROUP, axis=1, arg=16),), opts_to_apply=None))
     _ = to_program(ast, Device["METAL"].renderer)
 

test/backend/test_multitensor.py

@@ -127,6 +127,15 @@ class TestMultiTensor(unittest.TestCase):
     fn = f(n)
     np.testing.assert_allclose(fX.numpy(), fn, rtol=1e-6, atol=1e-6)
 
+  def test_stack(self):
+    X = Tensor.rand(4, 4).shard_(devices_2, 0)
+    Y = Tensor.rand(4, 4).shard_(devices_2, 0)
+    Z = Tensor.rand(4, 4).shard_(devices_2, 1)  # mismatched shard axis gets resharded
+    for dim in (0, 1):
+      np.testing.assert_allclose(Tensor.stack(X, Y, Z, dim=dim).numpy(), np.stack([X.numpy(), Y.numpy(), Z.numpy()], axis=dim))
+    grad = Tensor.stack(X, Y).sum().gradient(X)[0]
+    np.testing.assert_allclose(grad.numpy(), 1)
+
   def test_allreduce_naive(self):
     with Context(RING=0):
       a,b = _test_allreduce(Tensor.rand(256, 256))

test/backend/test_renderer_failures.py

@@ -28,25 +28,25 @@ def _setup_and_test_alu(alu_op:Ops, input_val:ConstType, *alu_src_uops:UOp):
   ld = b.index(idx).load()
   alu = ld.alu(alu_op, *alu_src_uops)
   store = UOp.store(a.index(idx), alu)
-  return _test_uop_result([Tensor([input_val])], UOp(Ops.SINK, dtypes.void, (store,), arg=KernelInfo()))[0]
+  return _test_uop_result([Tensor([input_val])], UOp(Ops.SINK, src=(store,), arg=KernelInfo()))[0]
 
 class TestRendererFailures(unittest.TestCase):
   @unittest.skipIf(not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, PythonRenderer)), "test is for ptx or python renderer")
   def test_gated_store_with_alu(self):
     a = UOp.param(0, dtypes.int, (4,))
-    gate_alu = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
-    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index(lidx0.valid(gate_alu)), UOp.const(dtypes.int, 1)))
-    sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,), arg=KernelInfo())
+    gate_alu = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
+    gated_alu_store = UOp(Ops.STORE, src=(a.index(lidx0.valid(gate_alu)), UOp.const(dtypes.int, 1)))
+    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
     ret = _test_uop_result([], sink, local_size=[4, 1, 1])[0]
     np.testing.assert_equal(ret, [0, 1, 1, 1])
 
   @unittest.skipIf(not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, PythonRenderer)), "test is for ptx or python renderer")
   def test_gated_store_with_alu_2d(self):
     a = UOp.param(0, dtypes.int, (8,))
-    gate_alu_0 = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
-    gate_alu_1 = (lidx1:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 2),), 'lidx1')).ne(0)
-    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(dtypes.int, 1)))
-    sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,), arg=KernelInfo())
+    gate_alu_0 = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
+    gate_alu_1 = (lidx1:=UOp.special(2, 'lidx1')).ne(0)
+    gated_alu_store = UOp(Ops.STORE, src=(a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(dtypes.int, 1)))
+    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
     ret = _test_uop_result([], sink, local_size=[4, 2, 1])[0]
     np.testing.assert_equal(ret, [0, 0, 0, 0, 0, 1, 1, 1])
 
@@ -88,13 +88,13 @@ class TestWGSLFailures(unittest.TestCase):
     a = UOp.param(0, dtypes.int, (4,))
     b = UOp.param(1, dtypes.int, (4,))
     c = UOp.param(2, dtypes.int, (4,))
-    lidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), "lidx0")
+    lidx0 = UOp.special(4, "lidx0")
     gate = lidx0.ne(0)
     alt = c.index(lidx0).load()
     ld = UOp.load(b.index(lidx0.valid(gate)))
     alt_load = gate.where(ld, alt)
     store = UOp.store(a.index(lidx0), alt_load)
-    sink = UOp(Ops.SINK, dtypes.void, (store,), arg=KernelInfo())
+    sink = UOp(Ops.SINK, src=(store,), arg=KernelInfo())
     ret = _test_uop_result([Tensor([0,1,2,3], dtype=dtypes.int), Tensor([4,5,6,7], dtype=dtypes.int)], sink, local_size=[4])[0]
     np.testing.assert_equal(ret, [4,1,2,3])
 
@@ -103,11 +103,11 @@ class TestPTXFailures(unittest.TestCase):
   @unittest.skip("INDEX can only have a gate ALU parent, not an IF")
   def test_gated_store_with_if(self):
     a = UOp.param(0, dtypes.int, (4,))
-    gate_alu = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
+    gate_alu = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
     val = UOp.const(dtypes.int, 1)
-    if_uop = UOp(Ops.IF, dtypes.void, (gate_alu,))
-    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index(lidx0, if_uop), val))
-    sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,), arg=KernelInfo())
+    if_uop = UOp(Ops.IF, src=(gate_alu,))
+    gated_alu_store = UOp(Ops.STORE, src=(a.index(lidx0, if_uop), val))
+    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
     ret = _test_uop_result([], sink, local_size=[4, 1, 1])[0]
     np.testing.assert_equal(ret, [0, 1, 1, 1])
 

test/backend/test_tensor_variable.py

@@ -1,6 +1,7 @@
 import unittest
 import numpy as np
-from tinygrad import Tensor, Variable
+from tinygrad import Tensor, Variable, dtypes
+from tinygrad.helpers import CHECK_OOB
 
 class TestTensorVariable(unittest.TestCase):
   def test_add_tvar(self):
@@ -8,6 +9,50 @@ class TestTensorVariable(unittest.TestCase):
     ret = (Tensor(vv) + 3).item()
     assert ret == 4
 
+  def test_variable_mul_tensor(self):
+    vv = Variable("a", 1, 10).bind(2)
+    t = Tensor.ones(3, dtype=dtypes.int8)
+    self.assertListEqual((t * vv).tolist(), [2, 2, 2])
+    # TODO: fix
+    try:
+      self.assertListEqual((vv * t).tolist(), [2, 2, 2])
+    except RuntimeError: pass
+
+  def test_large_range_variable(self):
+    vv = Variable("b", 0, 2**40).bind(2**35)
+    # TODO: pm_lower_index_dtype lowers ALU PARAM to int32 unconditionally
+    try:
+      self.assertEqual(Tensor(vv).item(), 2**35)
+    except AssertionError:
+      pass
+
+  def test_variable_tensor_dtype_arg(self):
+    vv = Variable("a", 1, 10).bind(2)
+    t = Tensor(vv, dtype=dtypes.float32)
+    self.assertEqual(t.dtype, dtypes.float32)
+    self.assertEqual(t.item(), 2.0)
+
+  def test_unbound_variable_tensor(self):
+    # an unbound variable schedules fine, but can't execute
+    with self.assertRaisesRegex(RuntimeError, "unbound"): Tensor(Variable("u", 1, 10)).item()
+    with self.assertRaisesRegex(RuntimeError, "unbound"): (Tensor(Variable("u", 1, 10)) + 1).item()
+    # bound variables in an expression are fine
+    self.assertEqual(Tensor(Variable("u", 1, 10).bind(2) + 1).item(), 3)
+
+  def test_shrink_beyond_buffer_variable(self):
+    # TODO: shrink by a variable whose vmax exceeds the dim should fail at build, today only CHECK_OOB=1 rejects it
+    t = Tensor.ones(3).contiguous()[:Variable("a", 1, 10).bind(5)]
+    if CHECK_OOB: self.assertRaises(RuntimeError, t.sum().item)
+    else: t.sum().item()  # silent OOB: reads 2 elements past the buffer, result depends on the allocator
+
+  def test_symbolic_shape_mul_variable_tensor(self):
+    # NOTE: the buffer dim must cover the variable's vmax
+    vv = Variable("a", 1, 10).bind(2)
+    self.assertEqual((Tensor.ones(10).contiguous()[:vv] * Tensor(vv)).sum().item(), 4.0)
+    # a vmin=0 symbolic dim broadcasts too
+    v0 = Variable("z", 0, 10).bind(2)
+    self.assertEqual((Tensor.ones(10).contiguous()[:v0] * Tensor(v0)).sum().item(), 4.0)
+
   def test_inner_tvar_node(self):
     vv = Variable("w", 0, 10).bind(2)
     ret = Tensor(vv * 4).item()

test/backend/test_uops.py

@@ -230,8 +230,8 @@ class TestAssembly(unittest.TestCase):
     c1 = UOp.const(dtypes.int, 2)
     c2 = UOp.const(dtypes.int, 3)
     l1 = g1.index(c1)
-    a1 = UOp(Ops.MUL, dtypes.int, (l1, c1))
-    a2 = UOp(Ops.MUL, dtypes.int, (l1, c2))
+    a1 = UOp(Ops.MUL, src=(l1, c1))
+    a2 = UOp(Ops.MUL, src=(l1, c2))
     uops = to_uops_list([out.index(UOp.const(dtypes.int, 0)).store(a1), out.index(UOp.const(dtypes.int, 1)).store(a2)],
                         ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)

test/external/external_benchmark_op_conv.py

@@ -12,7 +12,7 @@ from tinygrad.dtype import Invalid
 # PYTHONPATH="." DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
 
 def vision_conv_143():
-  c0 = UOp.param(0, dtypes.imageh((16, 1024, 4)))
+  c0 = UOp.param(0, dtypes.half, shape=(16, 1024, 4))
   c2 = UOp.range(32, 3, AxisType.LOOP)
   c5 = UOp.range(128, 4, AxisType.LOOP)
   c8 = UOp.range(16, 2, AxisType.LOOP)
@@ -22,11 +22,11 @@ def vision_conv_143():
   c26 = UOp.range(7, 1, AxisType.REDUCE)
   c27 = c2*2+c26
   c32 = ((c27<3)!=True)&(c27<67)
-  c34 = UOp.param(1, dtypes.imageh((32, 1024, 4)))
+  c34 = UOp.param(1, dtypes.half, shape=(32, 1024, 4))
   c38 = c5//2
-  c45 = (c32&c24).where((c27*64+c38+c17*4096+-12480), UOp.const(dtypes.weakint, Invalid))
+  c45 = (c32&c24).where((c27*64+c38+c17*4096+-12480), UOp.const(dtypes.index, Invalid))
   c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
-  c49 = UOp.param(2, dtypes.imageh((64, 49, 4)))
+  c49 = UOp.param(2, dtypes.half, shape=(64, 49, 4))
   c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
   c63 = UOp.param(3, dtypes.float, (128,))
   c65 = c61.reduce(c16, c26, arg=Ops.ADD)+c63.index(c5)
@@ -38,7 +38,7 @@ def vision_conv_143():
   return c67.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))
 
 def vision_conv_153():
-  c0 = UOp.param(0, dtypes.imageh((8, 1024, 4)))
+  c0 = UOp.param(0, dtypes.half, shape=(8, 1024, 4))
   c2 = UOp.range(16, 3, AxisType.LOOP)
   c5 = UOp.range(256, 4, AxisType.LOOP)
   c8 = UOp.range(8, 2, AxisType.LOOP)
@@ -48,11 +48,11 @@ def vision_conv_153():
   c26 = UOp.range(7, 1, AxisType.REDUCE)
   c27 = c2*2+c26
   c32 = ((c27<3)!=True)&(c27<35)
-  c34 = UOp.param(1, dtypes.imageh((16, 1024, 4)))
+  c34 = UOp.param(1, dtypes.half, shape=(16, 1024, 4))
   c38 = c5//2
-  c45 = (c32&c24).where((c27*128+c38+c17*4096+-12672), UOp.const(dtypes.weakint, Invalid))
+  c45 = (c32&c24).where((c27*128+c38+c17*4096+-12672), UOp.const(dtypes.index, Invalid))
   c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
-  c49 = UOp.param(2, dtypes.imageh((128, 49, 4)))
+  c49 = UOp.param(2, dtypes.half, shape=(128, 49, 4))
   c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
   c63 = UOp.param(3, dtypes.float, (256,))
   c65 = c61.reduce(c16, c26, arg=Ops.ADD)+c63.index(c5)
@@ -64,14 +64,14 @@ def vision_conv_153():
   return c67.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))
 
 def dm_conv_172():
-  c0 = UOp.param(0, dtypes.imageh((1, 240, 4)))
+  c0 = UOp.param(0, dtypes.half, shape=(1, 240, 4))
   c2 = UOp.range(960, 4, AxisType.LOOP)
-  c5 = UOp.param(1, dtypes.imageh((8, 384, 4)))
+  c5 = UOp.param(1, dtypes.half, shape=(8, 384, 4))
   c7 = UOp.range(32, 0, AxisType.REDUCE)
   c10 = UOp.range(4, 1, AxisType.REDUCE)
   c13 = UOp.range(12, 3, AxisType.REDUCE)
   c18 = UOp.range(8, 2, AxisType.REDUCE)
-  c23 = UOp.param(2, dtypes.imageh((240, 128, 4)))
+  c23 = UOp.param(2, dtypes.half, shape=(240, 128, 4))
   c35 = c5.index((c7*4+c10+c13*128+c18*1536))*c23.index((c10*4+c2%4+c7*16+c2//4*512))
   c37 = UOp.param(3, dtypes.float, (960,))
   c39 = c35.reduce(c7, c10, arg=Ops.ADD)+c37.index(c2)

test/external/external_test_yolov8.py

@@ -67,7 +67,7 @@ class TestYOLOv8(unittest.TestCase):
     # currently rtol is 0.025 because there is a 1-2% difference in our predictions
     # because of the zero padding in SPPF module (line 280) maxpooling layers rather than the -infinity in torch.
     # This difference does not make a difference "visually".
-    np.testing.assert_allclose(onnx_output, tiny_output, atol=5e-4, rtol=0.025)
+    np.testing.assert_allclose(onnx_output, tiny_output, atol=5e-4, rtol=0.01)
 
 if __name__ == '__main__':
   unittest.main()

test/external/fuzz_symbolic.py

@@ -40,7 +40,7 @@ def random_int_expr(depth=10):
 def random_bool_expr(depth=10, expr1=None):
   if depth == 0: return True
   if expr1 is None: expr1 = random_int_expr(depth-1)
-  expr2 = random.choice([random_or_sub_expression_int(depth-1, expr1), UOp.const(dtypes.int, random.randint(-10, 10))])
+  expr2 = random.choice([random_or_sub_expression_int(depth-1, expr1), UOp.const(dtypes.index, random.randint(-10, 10))])
   return random.choice(comp_ops)(expr1, expr2)
 
 
@@ -82,8 +82,8 @@ if __name__ == "__main__":
             f"v2=Variable(\"{u2.arg[0]}\", {u2.arg[1]}, {u2.arg[2]})\n" +\
             f"v3=Variable(\"{u3.arg[0]}\", {u3.arg[1]}, {u3.arg[2]})\n" +\
             f"expr = {expr}\n" +\
-            f"v1_val, v2_val, v3_val = UOp.const(dtypes.int, {n1.as_long()}), UOp.const(dtypes.int, {n2.as_long()})," +\
-                f"UOp.const(dtypes.int, {n3.as_long()})\n" +\
+            f"v1_val, v2_val, v3_val = UOp.const(dtypes.index, {n1.as_long()}), UOp.const(dtypes.index, {n2.as_long()})," +\
+                f"UOp.const(dtypes.index, {n3.as_long()})\n" +\
             "num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()\n" +\
             "rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()\n" +\
             "assert num==rn, f\"{num} != {rn}\"\n"

test/mockgpu/amd/pcode.py

@@ -77,7 +77,7 @@ def _val_to_bits(val):
   return val if val.dtype == dtypes.uint32 else val.cast(dtypes.uint32)
 
 def _floor(x):
-  t = UOp(Ops.TRUNC, x.dtype, (x,))
+  t = UOp(Ops.TRUNC, src=(x,))
   return ((x < _const(x.dtype, 0)) & x.ne(t)).where(t - _const(x.dtype, 1), t)
 def _f16_extract(v): return (v & _u32(0xFFFF)).cast(dtypes.uint16).bitcast(dtypes.half) if v.dtype == dtypes.uint32 else v
 
@@ -156,10 +156,10 @@ def _trig_reduce(x, phase=0.0):
     turns, two_pi = match
     if phase: turns = turns + _const(turns.dtype, phase)
     n = _floor(turns + _const(turns.dtype, 0.5))
-    return UOp(Ops.SIN, turns.dtype, ((turns - n) * _const(turns.dtype, two_pi),))
+    return UOp(Ops.SIN, src=((turns - n) * _const(turns.dtype, two_pi),))
   if phase: x = x + _const(x.dtype, phase * 6.283185307179586)
   n = _floor(x * _const(x.dtype, 0.15915494309189535) + _const(x.dtype, 0.5))
-  return UOp(Ops.SIN, x.dtype, (x - n * _const(x.dtype, 6.283185307179586),))
+  return UOp(Ops.SIN, src=(x - n * _const(x.dtype, 6.283185307179586),))
 
 def _signext(val: UOp) -> UOp:
   for bits, mask, ext in [(4, 0xF, 0xFFFFFFF0), (8, 0xFF, 0xFFFFFF00), (16, 0xFFFF, 0xFFFF0000)]:
@@ -184,7 +184,7 @@ def _abs(val: UOp) -> UOp:
 
 def _f_to_u(f, dt):
   clamped = (f < _const(f.dtype, 0.0)).where(_const(f.dtype, 0.0), f)
-  truncated = UOp(Ops.TRUNC, f.dtype, (clamped,))
+  truncated = UOp(Ops.TRUNC, src=(clamped,))
   return (truncated >= _const(f.dtype, 2**(dt.itemsize*8))).where(_const(dt, dt.max), truncated.cast(dt))
 
 def _cvt_quiet(val: UOp) -> UOp:
@@ -230,7 +230,7 @@ def _ldexp(val: UOp, exp: UOp) -> UOp:
   if val.dtype == dtypes.uint32: val = val.bitcast(dtypes.float32)
   elif val.dtype == dtypes.uint64: val = val.bitcast(dtypes.float64)
   if exp.dtype in (dtypes.uint32, dtypes.uint64): exp = exp.cast(dtypes.int if exp.dtype == dtypes.uint32 else dtypes.int64)
-  return val * UOp(Ops.EXP2, val.dtype, (exp.cast(val.dtype),))
+  return val * UOp(Ops.EXP2, src=(exp.cast(val.dtype),))
 
 def _frexp_mant(val: UOp) -> UOp:
   val = val.bitcast(dtypes.float32) if val.dtype == dtypes.uint32 else val.bitcast(dtypes.float64) if val.dtype == dtypes.uint64 else val
@@ -288,20 +288,20 @@ def _sad_u8(a: UOp, b: UOp, acc: UOp, masked: bool = False) -> UOp:
   return result
 
 _FUNCS: dict[str, Callable[..., UOp]] = {
-  'sqrt': lambda a: UOp(Ops.SQRT, a.dtype, (a,)), 'trunc': lambda a: UOp(Ops.TRUNC, a.dtype, (a,)),
-  'log2': lambda a: UOp(Ops.LOG2, a.dtype, (a,)), 'sin': lambda a: _trig_reduce(a),
+  'sqrt': lambda a: UOp(Ops.SQRT, src=(a,)), 'trunc': lambda a: UOp(Ops.TRUNC, src=(a,)),
+  'log2': lambda a: UOp(Ops.LOG2, src=(a,)), 'sin': lambda a: _trig_reduce(a),
   'cos': lambda a: _trig_reduce(a, 0.25), 'floor': _floor, 'fract': lambda a: a - _floor(a),
   'signext': _signext, 'abs': _abs,
-  'isEven': lambda a: (UOp(Ops.TRUNC, a.dtype, (a,)).cast(dtypes.int) & _const(dtypes.int, 1)).eq(_const(dtypes.int, 0)),
-  'max': lambda a, b: UOp(Ops.MAX, a.dtype, (a, b)),
-  'min': lambda a, b: UOp(Ops.MAX, a.dtype, (a.neg(), b.neg())).neg(),
-  'pow': lambda a, b: UOp(Ops.EXP2, dtypes.float32, (b.bitcast(dtypes.float32),)),
+  'isEven': lambda a: (UOp(Ops.TRUNC, src=(a,)).cast(dtypes.int) & _const(dtypes.int, 1)).eq(_const(dtypes.int, 0)),
+  'max': lambda a, b: UOp(Ops.MAX, src=(a, b)),
+  'min': lambda a, b: UOp(Ops.MAX, src=(a.neg(), b.neg())).neg(),
+  'pow': lambda a, b: UOp(Ops.EXP2, src=(b.bitcast(dtypes.float32),)),
   'fma': lambda a, b, c: a * b + c,
   'i32_to_f32': lambda a: a.cast(dtypes.int).cast(dtypes.float32),
   'u32_to_f32': lambda a: a.cast(dtypes.uint32).cast(dtypes.float32),
-  'f32_to_i32': lambda a: UOp(Ops.TRUNC, dtypes.float32, (a.bitcast(dtypes.float32),)).cast(dtypes.int),
+  'f32_to_i32': lambda a: UOp(Ops.TRUNC, src=(a.bitcast(dtypes.float32),)).cast(dtypes.int),
   'f32_to_u32': lambda a: _f_to_u(a.bitcast(dtypes.float32), dtypes.uint32),
-  'f64_to_i32': lambda a: UOp(Ops.TRUNC, dtypes.float64, (a.bitcast(dtypes.float64),)).cast(dtypes.int),
+  'f64_to_i32': lambda a: UOp(Ops.TRUNC, src=(a.bitcast(dtypes.float64),)).cast(dtypes.int),
   'f64_to_u32': lambda a: _f_to_u(a.bitcast(dtypes.float64), dtypes.uint32),
   'f16_to_f32': lambda a: _f16_extract(a).cast(dtypes.float32),
   'f32_to_f16': lambda a: a.cast(dtypes.half),
@@ -309,8 +309,8 @@ _FUNCS: dict[str, Callable[..., UOp]] = {
   'f64_to_f32': lambda a: a.bitcast(dtypes.float64).cast(dtypes.float32),
   'i32_to_f64': lambda a: a.cast(dtypes.int).cast(dtypes.float64),
   'u32_to_f64': lambda a: a.cast(dtypes.uint32).cast(dtypes.float64),
-  'f16_to_i16': lambda a: UOp(Ops.TRUNC, dtypes.half, (_f16_extract(a),)).cast(dtypes.int16),
-  'f16_to_u16': lambda a: UOp(Ops.TRUNC, dtypes.half, (_f16_extract(a),)).cast(dtypes.uint16),
+  'f16_to_i16': lambda a: UOp(Ops.TRUNC, src=(_f16_extract(a),)).cast(dtypes.int16),
+  'f16_to_u16': lambda a: UOp(Ops.TRUNC, src=(_f16_extract(a),)).cast(dtypes.uint16),
   'i16_to_f16': lambda a: a.cast(dtypes.int16).cast(dtypes.half),
   'u16_to_f16': lambda a: a.cast(dtypes.uint16).cast(dtypes.half),
   'bf16_to_f32': lambda a: (((a.cast(dtypes.uint32) if a.dtype != dtypes.uint32 else a) & _u32(0xFFFF)) << _u32(16)).bitcast(dtypes.float32),
@@ -343,7 +343,7 @@ _FUNCS: dict[str, Callable[..., UOp]] = {
   'i8_to_i32': lambda a: _signext(a.cast(dtypes.uint32) & _u32(0xFF)),
   'i4_to_i32': lambda a: _signext_4bit(a.cast(dtypes.uint32) & _u32(0xF)),
   # Float to int16 conversions
-  'v_cvt_i16_f32': lambda a: UOp(Ops.TRUNC, dtypes.float32, (a.bitcast(dtypes.float32),)).cast(dtypes.int16),
+  'v_cvt_i16_f32': lambda a: UOp(Ops.TRUNC, src=(a.bitcast(dtypes.float32),)).cast(dtypes.int16),
   'v_cvt_u16_f32': lambda a: _f_to_u(a.bitcast(dtypes.float32), dtypes.uint16),
   # SAD (Sum of Absolute Differences) - sum |a_i - b_i| for 4 bytes + accumulator
   'v_sad_u8': lambda a, b, c: _sad_u8(a, b, c),
@@ -492,7 +492,10 @@ class Parser:
       case '>=' | '<=' | '>' | '<' | '<>':
         ops = {'>=':(lambda a,b:a>=b),'<=':(lambda a,b:a<=b),'>':(lambda a,b:a>b),'<':(lambda a,b:a<b),'<>':(lambda a,b:a.ne(b))}
         return self._cmp_nan(left, right, ops[op])
-      case '>>' | '<<': return (left >> right) if op == '>>' else (left << right)
+      case '>>' | '<<':
+        if not dtypes.is_int(left.dtype): left = left.cast(dtypes.uint32)
+        if not dtypes.is_int(right.dtype): right = right.cast(dtypes.uint32)
+        return (left >> right) if op == '>>' else (left << right)
       case '+' | '-':
         if op == '-' and left.op == Ops.CONST and right.op == Ops.CONST: return _const(left.dtype, left.arg - right.arg)
         return (left + right) if op == '+' else (left - right)
@@ -504,7 +507,7 @@ class Parser:
           left, right = left.cast(pdt), right.cast(pdt)
         if op == '*': return left * right
         return (left // right) if dtypes.is_int(left.dtype) else (left / right)
-      case '**': return UOp(Ops.EXP2, left.dtype, (right.cast(left.dtype),)) if left.op == Ops.CONST and left.arg == 2.0 else left
+      case '**': return UOp(Ops.EXP2, src=(right.cast(left.dtype),)) if left.op == Ops.CONST and left.arg == 2.0 else left
 
   _PREC = [('||',), ('&&',), ('|',), ('^',), ('&',), ('==', '!=', '<>'), ('>=', '<=', '>', '<'), ('>>', '<<'), ('+', '-'), ('*', '/'), ('**',)]
 
@@ -829,7 +832,7 @@ class Parser:
     adt = dtypes.uint64 if addr.dtype == dtypes.uint64 else dtypes.uint32
     active = self.vars.get('_active')
     def mindex(idx:UOp): return mem.index(idx.valid(active) if active is not None else idx)
-    byte_mem = mem.dtype.base == dtypes.uint8
+    byte_mem = mem.dtype == dtypes.uint8
     if byte_mem:
       idx = addr
       if dt in (dtypes.uint64, dtypes.int64, dtypes.float64):
@@ -1304,7 +1307,7 @@ def parse_block(lines: list[str], start: int, env: dict[str, VarVal], funcs: dic
         result = else_branch[0]
         for c, rv in reversed(conditions):
           if isinstance(rv, UOp) and isinstance(result, UOp):
-            if rv.dtype != result.dtype and rv.dtype.itemsize == result.dtype.itemsize: result = result.cast(rv.dtype)
+            if rv.dtype != result.dtype: result = result.cast(rv.dtype)
             result = c.where(rv, result)
         return i, block_assigns, result
       # If statically true, use that branch directly; otherwise merge with WHERE
@@ -1325,7 +1328,7 @@ def parse_block(lines: list[str], start: int, env: dict[str, VarVal], funcs: dic
             if isinstance(ba, dict) and var in ba:
               tv = ba[var]
               if isinstance(tv, UOp) and isinstance(res, UOp):
-                res = cond.where(tv, res.cast(tv.dtype) if tv.dtype != res.dtype and tv.dtype.itemsize == res.dtype.itemsize else res)
+                res = cond.where(tv, res.cast(tv.dtype) if tv.dtype != res.dtype else res)
           block_assigns[var] = env[var] = res
         # Merge side effects from branches with conditions
         if assigns is not None:

test/null/test_dtype.py

@@ -1,25 +1,10 @@
 import unittest, pickle
 from tinygrad.tensor import Tensor
-from tinygrad.dtype import dtypes, DType, ImageDType, to_dtype, Invalid, InvalidType
-
-class TestImageDType(unittest.TestCase):
-  def test_image_scalar(self):
-    assert dtypes.imagef((10,10)).base.scalar() == dtypes.float32
-    assert dtypes.imageh((10,10)).base.scalar() == dtypes.float32
-  def test_image_vec(self):
-    assert dtypes.imagef((10,10)).base.vec(4) == dtypes.float32.vec(4)
-    assert dtypes.imageh((10,10)).base.vec(4) == dtypes.float32.vec(4)
+from tinygrad.dtype import dtypes, DType, to_dtype, Invalid, InvalidType
 
 class TestEqStrDType(unittest.TestCase):
-  def test_image_ne(self):
-    if ImageDType is None: raise unittest.SkipTest("no ImageDType support")
-    assert dtypes.float == dtypes.float32, "float doesn't match?"
-    assert dtypes.imagef((1,2,4)) != dtypes.imageh((1,2,4)), "different image dtype doesn't match"
-    assert dtypes.imageh((1,2,4)) != dtypes.imageh((1,4,2)), "different shape doesn't match"
-    assert dtypes.imageh((1,2,4)) == dtypes.imageh((1,2,4)), "same shape matches"
-    assert isinstance(dtypes.imageh((1,2,4)), ImageDType)
   def test_strs(self):
-    self.assertEqual(str(dtypes.imagef((1,2,4))), "dtypes.imagef((1, 2, 4))")
+    self.assertEqual(str(dtypes.float32), "dtypes.float")
 
 class TestToDtype(unittest.TestCase):
   def test_dtype_to_dtype(self):

test/null/test_dtype_spec.py

@@ -239,6 +239,22 @@ class TestTypePromotion(unittest.TestCase):
     assert least_upper_dtype(dtypes.weakint, dtypes.float32) == dtypes.float32
     assert least_upper_dtype(dtypes.weakint, dtypes.float64) == dtypes.float64
 
+  def test_weakfloat_promo(self):
+    # weakfloat is a float, but like weakint it is not one of dtypes.floats
+    assert dtypes.is_float(dtypes.weakfloat) and dtypes.weakfloat not in dtypes.floats
+    # weakfloat with itself is weakfloat
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.weakfloat) == dtypes.weakfloat
+    # weakfloat is above bool, weakint and any concrete int (they defer up to it)
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.bool) == dtypes.weakfloat
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.weakint) == dtypes.weakfloat
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.int32) == dtypes.weakfloat
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.uint64) == dtypes.weakfloat
+    # weakfloat defers to any concrete float type
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.fp8e4m3) == dtypes.fp8e4m3
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.float16) == dtypes.float16
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.float32) == dtypes.float32
+    assert least_upper_dtype(dtypes.weakfloat, dtypes.float64) == dtypes.float64
+
 class TestTypeSpec(unittest.TestCase):
   def setUp(self):
     self.old_default_int, self.old_default_float = dtypes.default_int, dtypes.default_float

test/null/test_gpudims.py

@@ -24,10 +24,10 @@ class TestGroupedDims(unittest.TestCase):
     total = math.prod(dims)
     specials = sorted(dedup(flatten([[y for y in x.toposort() if y.op is Ops.SPECIAL] for x in idxs])), key=lambda u: u.arg)
     # build flat index and primed flat (same expression with renamed SPECIALs)
-    flat = UOp.const(dtypes.weakint, 0)
+    flat = UOp.const(dtypes.index, 0)
     for i, idx in enumerate(idxs):
       flat = flat + idx * int(math.prod(dims[i+1:]))
-    flat_p = flat.substitute({s: UOp(Ops.SPECIAL, s.dtype, s.src, s.arg+"_p") for s in specials})
+    flat_p = flat.substitute({s: UOp(Ops.SPECIAL, src=s.src, arg=s.arg+"_p") for s in specials})
     solver = z3.Solver()
     [z3_flat, z3_flat_p] = uops_to_z3(solver, flat, flat_p)
     # bounds

test/null/test_graph_rewrite.py

@@ -107,28 +107,28 @@ class TestFoldingAndReduction(unittest.TestCase):
 class TestModuloAndDivisionFolding(unittest.TestCase):
   def test_full_graph_rewrite_modulo_folding_with_define_var(self):
     # index dtype because div-mod rules only work on index
-    x_var_uop = UOp.variable('x', 0, 100).cast(dtypes.weakint)
+    x_var_uop = UOp.variable('x', 0, 100).cast(dtypes.index)
     optimized_mod_uop = apply_rewrite(((x_var_uop * 4) + 2) % 4)
     self.assertEqual(optimized_mod_uop.op, Ops.CONST)
     self.assertEqual(optimized_mod_uop.arg, 2)
 
   def test_full_graph_rewrite_division_folding_with_define_var(self):
     # index dtype because div-mod rules only work on index
-    n_var_uop = UOp.variable('n', 1, 1000).cast(dtypes.weakint)
+    n_var_uop = UOp.variable('n', 1, 1000).cast(dtypes.index)
     optimized_div_uop = apply_rewrite((n_var_uop * 6) // 3)
     self.assertEqual(optimized_div_uop.op, Ops.MUL)
     self.assertEqual(optimized_div_uop.src[1].arg, 2)
 
   def test_full_graph_rewrite_complex_mod_div_folding(self):
     # index dtype because div-mod rules only work on index
-    k_var_uop = UOp.variable('k', 0, 50).cast(dtypes.weakint)
+    k_var_uop = UOp.variable('k', 0, 50).cast(dtypes.index)
     optimized_div_uop = apply_rewrite(((k_var_uop * 12 + 8) % 6) // 2)
     self.assertEqual(optimized_div_uop.op, Ops.CONST)
     self.assertEqual(optimized_div_uop.arg, 1)
 
   def test_graph_rewrite_div_folding_bug(self):
-    lhs = UOp(Ops.ADD, dtypes.int.vec(4), src=(
-      UOp(Ops.STACK, dtypes.int.vec(4), arg=None, src=(UOp(Ops.SPECIAL, dtypes.int, arg='lidx0', src=(UOp.const(dtypes.int, 32),)),)*4),
+    lhs = UOp(Ops.ADD, src=(
+      UOp(Ops.STACK, arg=None, src=(UOp(Ops.SPECIAL, src=(UOp.const(dtypes.int, 32),), arg='lidx0'),)*4),
       UOp.const(dtypes.int, (0, 256, 512, 768))))
     rhs = UOp.const(dtypes.int, (2,)*4)
     unopt = lhs<rhs
@@ -140,7 +140,7 @@ class TestModuloAndDivisionFolding(unittest.TestCase):
   def test_full_graph_rewrite_modulo_large_divisor(self):
     # index dtype because div-mod rules only work on index
     x_var_uop = UOp.variable('x', 1, 5)
-    self.assertIs(apply_rewrite(x_var_uop.cast(dtypes.weakint) % 10).render(simplify=False), x_var_uop.render(simplify=False))
+    self.assertIs(apply_rewrite(x_var_uop.cast(dtypes.index) % 10).render(simplify=False), x_var_uop.render(simplify=False))
 
   def test_full_graph_rewrite_division_with_remainder(self):
     x_var_uop = UOp.variable('x', 7, 9)
@@ -188,7 +188,7 @@ class TestGEPAndVectorizeRewrite(unittest.TestCase):
   def test_gep_tuple_extraction(self):
     # GEP on a vector dtype to extract multiple elements as a vector
     base_vector = UOp.const(dtypes.float32, (1.0, 2.0, 3.0, 4.0))
-    self.assertEqual(list(apply_rewrite_values(UOp.vectorize(*[base_vector.index(i) for i in (2, 3)]))), [3.0, 4.0])
+    self.assertEqual(list(apply_rewrite_values(UOp.stack(*[base_vector.index(i) for i in (2, 3)]))), [3.0, 4.0])
 
   def test_gep_on_const_stack(self):
     # GEP on a const STACK to extract a single element
@@ -198,12 +198,12 @@ class TestGEPAndVectorizeRewrite(unittest.TestCase):
   def test_gep_tuple_on_const_stack(self):
     # GEP on a const STACK using a tuple to extract multiple elements
     const_stack = UOp.const(dtypes.float32, (7.0, 8.0, 9.0, 10.0))
-    self.assertEqual(list(apply_rewrite_values(UOp.vectorize(*[const_stack.index(i) for i in (1, 3)]))), [8.0, 10.0])
+    self.assertEqual(list(apply_rewrite_values(UOp.stack(*[const_stack.index(i) for i in (1, 3)]))), [8.0, 10.0])
 
   def test_vectorize_multiple_elements(self):
     # Vectorizing multiple elements using GEP
     base_vector = UOp.const(dtypes.float32, (5.0, 10.0, 15.0, 20.0))
-    vectorized_uop = UOp(Ops.STACK, dtypes.float32, src=tuple(base_vector.index(i) for i in range(4)))
+    vectorized_uop = UOp(Ops.STACK, src=tuple(base_vector.index(i) for i in range(4)))
     self.assertEqual(list(apply_rewrite_values(vectorized_uop)), [5.0, 10.0, 15.0, 20.0])
 
 

test/null/test_linearizer_failures.py

@@ -8,12 +8,12 @@ from tinygrad.codegen import to_program
 class TestLinearizerFailures(unittest.TestCase):
   def test_fail_1(self):
     c0 = UOp.param(0, dtypes.float, (64,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 2), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 32), 2, AxisType.LOOP)
-    c3 = ((c1*UOp.const(dtypes.weakint, 32))+c2)
+    c1 = UOp.range(UOp.const(dtypes.index, 2), 1, AxisType.LOOP)
+    c2 = UOp.range(UOp.const(dtypes.index, 32), 2, AxisType.LOOP)
+    c3 = ((c1*UOp.const(dtypes.index, 32))+c2)
     c4 = UOp.param(1, dtypes.float, (163840,))
-    c5 = UOp.range(UOp.const(dtypes.weakint, 2560), 0, AxisType.REDUCE)
-    c6 = c4.index(((((((c5//UOp.const(dtypes.weakint, 8))%UOp.const(dtypes.weakint, 8))*UOp.const(dtypes.weakint, 8))+(c5%UOp.const(dtypes.weakint, 8)))+(((c2*UOp.const(dtypes.weakint, 40))+(c5//UOp.const(dtypes.weakint, 64)))*UOp.const(dtypes.weakint, 64)))+(c1*UOp.const(dtypes.weakint, 81920))))
+    c5 = UOp.range(UOp.const(dtypes.index, 2560), 0, AxisType.REDUCE)
+    c6 = c4.index(((((((c5//UOp.const(dtypes.index, 8))%UOp.const(dtypes.index, 8))*UOp.const(dtypes.index, 8))+(c5%UOp.const(dtypes.index, 8)))+(((c2*UOp.const(dtypes.index, 40))+(c5//UOp.const(dtypes.index, 64)))*UOp.const(dtypes.index, 64)))+(c1*UOp.const(dtypes.index, 81920))))
     c7 = UOp.param(2, dtypes.float, (64,))
     c8 = c7.index(c3)
     c9 = ((((c6+(c8*UOp.const(dtypes.float, -1.0)))*(c6+(c8*UOp.const(dtypes.float, -1.0)))).reduce(c5, arg=Ops.ADD)*UOp.const(dtypes.float, 0.000390625))+UOp.const(dtypes.float, 1e-05)).sqrt().reciprocal()

test/null/test_memory_planner.py

@@ -20,7 +20,11 @@ def _make_linear(buffer_lists, copies=None):
   calls = []
   for bufs in buffer_lists:
     is_copy = len(bufs) == 2 and frozenset((id(bufs[0]), id(bufs[1]))) in copy_pairs
-    calls.append(UOp(Ops.CALL, dtypes.void, (UOp(Ops.COPY if is_copy else Ops.SINK), *bufs)))
+    if is_copy:
+      src0 = bufs[0].copy_to_device(bufs[1].device)
+    else:
+      src0 = UOp(Ops.SINK, src=tuple(bufs))
+    calls.append(UOp(Ops.CALL, src=(src0, *bufs)))
   return UOp(Ops.LINEAR, src=tuple(calls))
 
 def _get_arena(buf, linear, result):

test/null/test_pattern_matcher.py

@@ -62,16 +62,16 @@ class TestPatternMatcher(unittest.TestCase):
   def test_uop(self):
     matcher = PatternMatcher([(UPat(Ops.CONST, name="x"), lambda x: x.rtag())])
     c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp(Ops.ADD, dtypes.float, (c1, c1))
+    c2 = UOp(Ops.ADD, src=(c1, c1))
     self.assertEqual(matcher.rewrite(c1), c1.rtag())
     self.assertEqual(matcher.rewrite(c2), None)
 
   def test_uop_set(self):
     matcher = PatternMatcher([(UPat((Ops.CONST, Ops.CAST), name="x"), lambda x: x.rtag())])
     c1 = UOp.const(dtypes.bool, False)
-    c2 = UOp(Ops.CAST, dtypes.int, (c1,))
+    c2 = UOp(Ops.CAST, arg=dtypes.int, src=(c1,))
     c3 = UOp.const(dtypes.float, 1.0)
-    c4 = UOp(Ops.ADD, dtypes.float, (c3, c3))
+    c4 = UOp(Ops.ADD, src=(c3, c3))
     self.assertEqual(matcher.rewrite(c1), c1.rtag())
     self.assertEqual(matcher.rewrite(c2), c2.rtag())
     self.assertEqual(matcher.rewrite(c4), None)
@@ -84,8 +84,8 @@ class TestPatternMatcher(unittest.TestCase):
     ])
     c1 = UOp.const(dtypes.float, 0.0)
     c2 = UOp.const(dtypes.bool, False)
-    c3 = UOp(Ops.MAX, dtypes.float, (c1, c1))
-    c4 = UOp(Ops.MUL, dtypes.float, (c1, c1))
+    c3 = UOp(Ops.MAX, src=(c1, c1))
+    c4 = UOp(Ops.MUL, src=(c1, c1))
     c5 = UOp.const(dtypes.int, -1)
     self.assertEqual(matcher.rewrite(c1), c1.rtag())
     self.assertEqual(matcher.rewrite(c2), c2.rtag())
@@ -101,11 +101,11 @@ class TestPatternMatcher(unittest.TestCase):
     y1 = UOp.const(dtypes.int, 1)
     y2 = UOp.const(dtypes.int, 2)
     y3 = UOp.const(dtypes.int, -1)
-    c1 = UOp(Ops.MUL, dtypes.int, (y1, y2))
-    c2 = UOp(Ops.MUL, dtypes.int, (y2, y2))
-    c3 = UOp(Ops.MUL, dtypes.int, (y3, y2))
-    c4 = UOp(Ops.MUL, dtypes.int, (y2, y1))
-    c5 = UOp(Ops.MUL, dtypes.int, (y2, y3))
+    c1 = UOp(Ops.MUL, src=(y1, y2))
+    c2 = UOp(Ops.MUL, src=(y2, y2))
+    c3 = UOp(Ops.MUL, src=(y3, y2))
+    c4 = UOp(Ops.MUL, src=(y2, y1))
+    c5 = UOp(Ops.MUL, src=(y2, y3))
     self.assertEqual(matcher.rewrite(c1), c1.rtag())
     self.assertEqual(matcher.rewrite(c2), None)
     self.assertEqual(matcher.rewrite(c3), c3.rtag())
@@ -116,8 +116,8 @@ class TestPatternMatcher(unittest.TestCase):
     matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=(UPat(Ops.CONST, name="y"), UPat(Ops.CONST, name="y"))), lambda x, y: x.rtag())])
     y1 = UOp.const(dtypes.float, 1.0)
     y2 = UOp.const(dtypes.float, 1.0)
-    c1 = UOp(Ops.ADD, dtypes.float, (y1, y1))
-    c2 = UOp(Ops.ADD, dtypes.float, (y1, y2))
+    c1 = UOp(Ops.ADD, src=(y1, y1))
+    c2 = UOp(Ops.ADD, src=(y1, y2))
     self.assertEqual(matcher.rewrite(c1), c1.rtag())
     self.assertEqual(matcher.rewrite(c2), c1.rtag())
 
@@ -143,14 +143,14 @@ class TestPatternMatcher(unittest.TestCase):
     matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=(UPat(Ops.CONST), UPat(Ops.CONST))), lambda x: x.rtag())])
     c1 = UOp.const(dtypes.float, 1.0)
     c2 = UOp.const(dtypes.float, 2.0)
-    c3 = UOp(Ops.ADD, dtypes.float, (c1,c2))
+    c3 = UOp(Ops.ADD, src=(c1,c2))
     self.assertEqual(matcher.rewrite(c3), c3.rtag())
     self.assertEqual(matcher.rewrite(c2), None)
     # that CONST/ALU -> ALU/CONST rewrite is now instant
     """
     matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=(UPat(Ops.CONST), UPat(GroupOp.ALU))), lambda x: x)])
-    c4 = UOp(Ops.ADD, dtypes.float, (c1,c3))
-    c5 = UOp(Ops.ADD, dtypes.float, (c3,c1))
+    c4 = UOp(Ops.ADD, src=(c1,c3))
+    c5 = UOp(Ops.ADD, src=(c3,c1))
     self.assertEqual(matcher.rewrite(c3), None)
     self.assertEqual(matcher.rewrite(c4), c4)
     self.assertEqual(matcher.rewrite(c5), None)
@@ -160,10 +160,10 @@ class TestPatternMatcher(unittest.TestCase):
     matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=[UPat(Ops.CONST), UPat(GroupOp.ALU)]), lambda x: x.rtag())])
     c1 = UOp.const(dtypes.float, 1.0)
     c2 = UOp.const(dtypes.float, 2.0)
-    c3 = UOp(Ops.ADD, dtypes.float, (c1,c2))
-    c4 = UOp(Ops.ADD, dtypes.float, (c3,c2))
-    c5 = UOp(Ops.ADD, dtypes.float, (c2,c3))
-    c6 = UOp(Ops.ADD, dtypes.float, (c3,c4))
+    c3 = UOp(Ops.ADD, src=(c1,c2))
+    c4 = UOp(Ops.ADD, src=(c3,c2))
+    c5 = UOp(Ops.ADD, src=(c2,c3))
+    c6 = UOp(Ops.ADD, src=(c3,c4))
     self.assertEqual(matcher.rewrite(c3), None)
     self.assertEqual(matcher.rewrite(c4), c4.rtag())
     self.assertEqual(matcher.rewrite(c5), c5.rtag())
@@ -173,8 +173,8 @@ class TestPatternMatcher(unittest.TestCase):
     matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=UPat(Ops.CONST)), lambda x: x.rtag())])
     c1 = UOp.const(dtypes.float, 1.0)
     c2 = UOp.const(dtypes.float, 2.0)
-    c3 = UOp(Ops.ADD, dtypes.float, (c1,c2))
-    c4 = UOp(Ops.ADD, dtypes.float, (c2,c3))
+    c3 = UOp(Ops.ADD, src=(c1,c2))
+    c4 = UOp(Ops.ADD, src=(c2,c3))
     self.assertEqual(matcher.rewrite(c3), c3.rtag())
     self.assertEqual(matcher.rewrite(c4), None)
 
@@ -183,9 +183,9 @@ class TestPatternMatcher(unittest.TestCase):
     c1 = UOp.const(dtypes.float, 1.0)
     c2 = UOp.const(dtypes.float, 2.0)
     c3 = UOp.const(dtypes.float, 3.0)
-    c4 = UOp(Ops.EXP2, dtypes.float, (c1,))
-    c5 = UOp(Ops.ADD, dtypes.float, (c1,c2))
-    c6 = UOp(Ops.MULACC, dtypes.float, (c1,c2,c3))
+    c4 = UOp(Ops.EXP2, src=(c1,))
+    c5 = UOp(Ops.ADD, src=(c1,c2))
+    c6 = UOp(Ops.MULACC, src=(c1,c2,c3))
     self.assertEqual(matcher.rewrite(c4), None)
     self.assertEqual(matcher.rewrite(c5), None)
     self.assertEqual(matcher.rewrite(c6), c6.rtag())

test/null/test_simplify_valid_idx.py

@@ -14,16 +14,16 @@ def simplify_valid_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move
 def simplify_image_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move_where_on_load+indexing_simplify, name="simplify_image_idx")
 
 def get_gated_load_uop(valid:UOp, idx:UOp):
-  return UOp(Ops.LOAD, dtypes.float, (
+  return UOp(Ops.LOAD, src=(
     UOp.param(0, dtypes.float, (1024,)).index(idx.valid(valid)),
   ))
 
 def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UOp]):
-  return UOp(Ops.LOAD, dtypes.float, (
-    UOp.param(0, dtypes.imagef(image_shape)).index(idx[1].valid(valid), idx[0].valid(valid)),
+  return UOp(Ops.LOAD, src=(
+    UOp.param(0, dtypes.float, image_shape).index(idx[1].valid(valid), idx[0].valid(valid)),
   ))
 
-def Special(expr, nmax): return UOp(Ops.SPECIAL, dtypes.weakint, (UOp.const(dtypes.weakint, nmax),), expr)
+def Special(expr, nmax): return UOp(Ops.SPECIAL, src=(UOp.const(dtypes.index, nmax),), arg=expr)
 def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
 def Range(n, nmax): return UOp.range(nmax, n)
 
@@ -455,7 +455,7 @@ class TestImageSimplification(unittest.TestCase):
     A1 = lidx0*32 + r0*32 + lidx1*4 - 99
     valid = ((lidx1 < 1).ne(True)) & ((lidx0 + r0) < 3).ne(True) & ((lidx0 + r0) < 19)
     alu0 = gidx0 + (A1 % 32)*32 + (A1 // 32 % 16)*1024
-    load = get_load_image_uop((1, 16384, 4), valid, (alu0, UOp.const(dtypes.weakint, 0)))
+    load = get_load_image_uop((1, 16384, 4), valid, (alu0, UOp.const(dtypes.index, 0)))
     try:
       self.check(load, None, "(gidx0+lidx0*1024+r0*1024+lidx1*128+-3168)", "0")
     except AssertionError:
@@ -474,7 +474,7 @@ class TestImageSimplification(unittest.TestCase):
     A1 = lidx0*16 + r0*16 + lidx1*4 - 51
     valid = ((lidx1 < 1).ne(True)) & ((lidx0 + r0) < 3).ne(True) & ((lidx0 + r0) < 11)
     alu0 = lidx2 + gidx0*4 + (A1 % 16)*64 + (A1 // 16 % 8)*1024
-    load = get_load_image_uop((1, 8192, 4), valid, (alu0, UOp.const(dtypes.weakint, 0)))
+    load = get_load_image_uop((1, 8192, 4), valid, (alu0, UOp.const(dtypes.index, 0)))
     try:
       self.check(load, None, "(lidx2+gidx0*4+lidx0*1024+r0*1024+lidx1*256+-3264)", "0")
     except AssertionError:
@@ -488,7 +488,7 @@ class TestImageSimplification(unittest.TestCase):
     gidx0 = Special("gidx0", 1064)
     r12 = Range(12, 3)
     valid = ((gidx0 < 645).ne(True)) & (gidx0 < 653)
-    idx = (r12*4 + (gidx0+3)%4 + (gidx0+3)//4*24 - 3888, UOp.const(dtypes.weakint, 0))
+    idx = (r12*4 + (gidx0+3)%4 + (gidx0+3)//4*24 - 3888, UOp.const(dtypes.index, 0))
     load = get_load_image_uop((1, 48, 4), valid, idx)
     self.check(load, None, "(r12*4+(gidx0+3)%4+(gidx0+3)//4*24+-3888)", "0")
 
@@ -499,9 +499,9 @@ class TestDropTrueGate(unittest.TestCase):
     from tinygrad.uop.ops import graph_rewrite
     from tinygrad.uop.symbolic import sym
     buf = UOp.param(0, dtypes.int, (1,))
-    idx = UOp.const(dtypes.weakint, 0)
+    idx = UOp.const(dtypes.index, 0)
     true_gate = UOp.const(dtypes.bool, True)
-    index_with_gate = UOp(Ops.INDEX, dtypes.int, (buf, idx.valid(true_gate)))
+    index_with_gate = UOp(Ops.INDEX, src=(buf, idx.valid(true_gate)))
     # apply the optimization
     result = graph_rewrite(index_with_gate, sym+indexing_simplify)
     # the True valid should be dropped (INDEX should only have 2 sources)
@@ -516,7 +516,7 @@ class TestRangeShrink(unittest.TestCase):
   def test_range_shrink_single_guard(self):
     # range 0..203 guarded by r < 4 everywhere -> shrink to 0..3
     r = Range(0, 204)
-    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
+    load = get_gated_load_uop(r < UOp.const(dtypes.index, 4), r)
     ranges = self.get_ranges(load.sink())
     self.assertEqual(len(ranges), 1)
     self.assertEqual(ranges[0].src[0].arg, 4)
@@ -524,8 +524,8 @@ class TestRangeShrink(unittest.TestCase):
   def test_range_shrink_picks_max_guard(self):
     # two loads guard the same range with r < 4 and r < 8 -> shrink to max(4, 8) = 8
     r = Range(0, 204)
-    load1 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
-    load2 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 8), r)
+    load1 = get_gated_load_uop(r < UOp.const(dtypes.index, 4), r)
+    load2 = get_gated_load_uop(r < UOp.const(dtypes.index, 8), r)
     ranges = self.get_ranges(UOp.sink(load1, load2))
     self.assertEqual(len(ranges), 1)
     self.assertEqual(ranges[0].src[0].arg, 8)
@@ -533,7 +533,7 @@ class TestRangeShrink(unittest.TestCase):
   def test_range_no_shrink_guard_ge_max(self):
     # guard r < 300 with range max 204 -> no shrink (guard doesn't constrain)
     r = Range(0, 204)
-    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 300), r)
+    load = get_gated_load_uop(r < UOp.const(dtypes.index, 300), r)
     ranges = self.get_ranges(load.sink())
     self.assertEqual(len(ranges), 1)
     self.assertEqual(ranges[0].src[0].arg, 204)
@@ -541,8 +541,8 @@ class TestRangeShrink(unittest.TestCase):
   def test_range_no_shrink_when_unguarded_elsewhere(self):
     # one load guards r < 4, but another load uses r without a gate -> no shrink
     r = Range(0, 204)
-    load1 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
-    load2 = UOp(Ops.LOAD, dtypes.float, (UOp.param(1, dtypes.float, (204,)).index(r),))
+    load1 = get_gated_load_uop(r < UOp.const(dtypes.index, 4), r)
+    load2 = UOp(Ops.LOAD, src=(UOp.param(1, dtypes.float, (204,)).index(r),))
     ranges = self.get_ranges(UOp.sink(load1, load2))
     self.assertEqual(len(ranges), 1)
     self.assertEqual(ranges[0].src[0].arg, 204)
@@ -550,7 +550,7 @@ class TestRangeShrink(unittest.TestCase):
   def test_range_no_shrink_when_used_in_reduce(self):
     # range used in both a gated load AND directly in the reduce expression -> no shrink
     r = Range(0, 204)
-    gated_load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
+    gated_load = get_gated_load_uop(r < UOp.const(dtypes.index, 4), r)
     red = (r.cast(dtypes.float) + gated_load).reduce(r, arg=Ops.ADD)
     ranges = self.get_ranges(red.sink())
     self.assertEqual(len(ranges), 1)
@@ -559,7 +559,7 @@ class TestRangeShrink(unittest.TestCase):
   def test_range_shrink_to_single_iteration(self):
     # guard r < 1 shrinks range to 1 -> single iteration, range eliminated entirely
     r = Range(0, 204)
-    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 1), r)
+    load = get_gated_load_uop(r < UOp.const(dtypes.index, 1), r)
     ranges = self.get_ranges(load.sink())
     self.assertEqual(len(ranges), 0)
 

test/null/test_symbolic_tensor.py

@@ -1,6 +1,23 @@
 import unittest
 from tinygrad import Variable
 from tinygrad.tensor import Tensor
+from tinygrad.uop.ops import _broadcast_shape
+
+class TestBroadcastShape(unittest.TestCase):
+  def test_symbolic(self):
+    v = Variable("v", 1, 10)
+    self.assertEqual(_broadcast_shape((v,), (1,)), (v,))
+    self.assertEqual(_broadcast_shape((v,), ()), (v,))
+    self.assertEqual(_broadcast_shape((v,), (v,)), (v,))
+    with self.assertRaises(IndexError): _broadcast_shape((v,), (5,))
+
+  def test_symbolic_vmin_zero(self):
+    # a symbolic dim that may be 0 still broadcasts against 1 to itself
+    v0 = Variable("v0", 0, 10)
+    self.assertEqual(_broadcast_shape((v0,), (1,)), (v0,))
+    self.assertEqual(_broadcast_shape((v0,), ()), (v0,))
+    self.assertEqual(_broadcast_shape((3, v0), (3, 1)), (3, v0))
+    with self.assertRaises(IndexError): _broadcast_shape((v0,), (5,))
 
 class TestSymbolic(unittest.TestCase):
   def assert_tuple_equal(self, x, y):

test/null/test_tensor.py

@@ -69,7 +69,7 @@ class TestIdxUpcast(unittest.TestCase):
     if not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)):
       assert idx.op is Ops.INDEX
       idx_val = idx.src[1]
-      self.assertFalse(idx_val.overflows(idx_val.dtype.base.scalar()))
+      self.assertFalse(idx_val.overflows(idx_val.dtype.scalar()))
 
   # use expand to generate kernel that uses large idx
   def do_op_then_assert(self, dtype: DType, dim1, dim2, dim3):

test/null/test_tensor_uop_mixin.py

@@ -222,6 +222,10 @@ class TestTensorUOpBitcast(unittest.TestCase):
     t = _t(4)
     self.assertIs(t.bitcast("uint32").uop, t.uop.bitcast("uint32"))
     self.assertIs(t.uop.bitcast("uint32").dtype, dtypes.uint32)
+  def test_bitcast_same_and_diff_size(self):
+    _check(self, _t(4).float(), lambda x: x.bitcast(dtypes.uint32))              # same size
+    _check(self, _t(4).cast(dtypes.uint8), lambda x: x.bitcast(dtypes.uint16))   # widen: uint8[4] -> uint16[2]
+    _check(self, _t(4).cast(dtypes.uint16), lambda x: x.bitcast(dtypes.uint8))   # narrow: uint16[4] -> uint8[8]
 
 class TestTensorUOpRand(unittest.TestCase):
   def test_random_bits(self):
@@ -372,6 +376,13 @@ class TestTensorUOpStack(unittest.TestCase):
   def test_stack_dim1(self):     _check(self, _t(2, 3), lambda x: x.stack(x, dim=1))
   def test_stack_3tensors(self): _check(self, _t(2, 3), lambda x: x.stack(x, x, dim=0))
   def test_stack_new_last(self): _check(self, _t(2, 3), lambda x: x.stack(x, dim=-1))
+  def test_stack_mixed_dtype(self):
+    w = _t(2, 3).float()
+    _check(self, _t(2, 3), lambda x: x.stack(w if isinstance(x, Tensor) else w.uop))
+    self.assertIs(_t(2, 3).uop.stack(w.uop).dtype, dtypes.float32)
+  def test_stack_index_dtype(self):
+    # index is outside the promotion lattice, equal dtypes bypass promotion
+    self.assertEqual(UOp.const(dtypes.index, 1).stack(UOp.const(dtypes.index, 2)).shape, (2,))
 
 class TestTensorUOpConv2d(unittest.TestCase):
   def test_conv2d_basic(self):
@@ -420,6 +431,10 @@ class TestTensorUOpConv2d(unittest.TestCase):
     w = _t(1, 1, 2, 2).float()
     _check(self, _t(1, 1, 3, 3).float(), lambda x: x.conv_transpose2d(w if isinstance(x, Tensor) else w.uop, stride=2))
 
+class TestTensorUOpHashing(unittest.TestCase):
+  def test_keccak_sha3_256(self):  _check(self, _t(8).cast(dtypes.uint8), lambda x: x.keccak())
+  def test_keccak_shake_128(self): _check(self, _t(8).cast(dtypes.uint8), lambda x: x.keccak("shake_128"))
+
 class TestTensorUOpEinsum(unittest.TestCase):
   def test_einsum_dot(self):       _check(self, _t(2, 3), lambda x: type(x).einsum("ij,ij->", x, x))
   def test_einsum_transpose(self): _check(self, _t(2, 3), lambda x: type(x).einsum("ij->ji", x))

test/null/test_tinyfs.py

@@ -1,27 +0,0 @@
-import unittest
-from tinygrad import Tensor
-from tinygrad.nn.state import fs_store, fs_load
-
-class TestLoadStore(unittest.TestCase):
-  def test_load_shape(self):
-    t = fs_load(Tensor(bytes(16)), 1024)
-    assert t.shape == (1024,), t.shape
-    t.schedule_linear()
-
-  def test_store_shape(self):
-    t = fs_store(Tensor.zeros(1024))
-    assert t.shape == (16,), t.shape
-    t.schedule_linear()
-
-  def test_load_large_shape(self):
-    t = fs_load(Tensor(bytes(16)), 10_000_000)
-    assert t.shape == (10_000_000,), t.shape
-    t.schedule_linear()
-
-  def test_store_large_shape(self):
-    t = fs_store(Tensor.zeros(10_000_000))
-    assert t.shape == (16,), t.shape
-    t.schedule_linear()
-
-if __name__ == "__main__":
-  unittest.main()

test/null/test_uop_graph.py

@@ -202,7 +202,7 @@ class TestUOpGraph(unittest.TestCase):
 
   def test_where_same_fold(self):
     v = UOp.variable('tmp', 0, 1)
-    c0 = UOp.const(dtypes.weakint, 0)
+    c0 = UOp.const(dtypes.index, 0)
     vc = v != c0
     c1 = UOp.const(dtypes.float, 1.0)
     out = vc.where(c1, c1)
@@ -255,8 +255,8 @@ class TestUOpGraph(unittest.TestCase):
     ld = d0.load(idx, dtype=dtypes.float.vec(2))
     vec = UOp(Ops.STACK, dtypes.float.vec(2), (ld,))
     x = vec.index(0)
-    alu = UOp(Ops.SQRT, dtypes.float, (x, ))
-    out = UOp(Ops.STORE, dtypes.void, (d0, idx, alu))
+    alu = UOp(Ops.SQRT, src=(x, ))
+    out = UOp(Ops.STORE, src=(d0, idx, alu))
     uops = to_uops_list([out])
     self.assertEqual(len([x for x in uops if x.op is Ops.STACK]), 0)
 
@@ -303,7 +303,7 @@ class TestUOpGraph(unittest.TestCase):
   def test_gep_vec_const_fold(self):
     for vec_size in [2, 4, 8]:
       consts = [UOp.const(dtypes.float, float(i)) for i in range(vec_size)]
-      vec = UOp(Ops.STACK, dtypes.float, tuple(consts))
+      vec = UOp(Ops.STACK, src=tuple(consts))
       with Context(SPEC=0):
         uops = to_uops_list([vec.index(i) for i in range(vec_size)])
         for uop, const in zip(uops, consts):
@@ -315,7 +315,7 @@ class TestUOpGraph(unittest.TestCase):
       vec = UOp(Ops.STACK, dtypes.half.vec(i), tuple(UOp.const(dtypes.half, 0.0) for _ in range(i)))
       var = UOp.variable("var", 0, 1, dtypes.half.vec(i))
       acc = UOp.variable('acc', 0, 1, dtypes.half.vec(i))
-      wmma = UOp(Ops.WMMA, dtypes.half.vec(i), (vec, var, acc))
+      wmma = UOp(Ops.WMMA, src=(vec, var, acc))
       uops = to_uops_list([wmma])
       self.assertEqual(uops[0], acc)
       self.assertEqual(len(uops), 2)  # +1 for SINK
@@ -324,7 +324,7 @@ class TestUOpGraph(unittest.TestCase):
       var = UOp.variable("var", 0, 1, dtypes.half.vec(i))
       vec = UOp(Ops.STACK, dtypes.half.vec(i), tuple(UOp.const(dtypes.half, 0.0) for _ in range(i)))
       acc = UOp.variable('acc', 0, 1, dtypes.half.vec(i))
-      wmma = UOp(Ops.WMMA, dtypes.half.vec(i), (var, vec, acc))
+      wmma = UOp(Ops.WMMA, src=(var, vec, acc))
       uops = to_uops_list([wmma])
       self.assertEqual(uops[0], acc)
       self.assertEqual(len(uops), 2)  # +1 for SINK
@@ -337,7 +337,7 @@ class TestUOpGraph(unittest.TestCase):
                 tuple(UOp.variable(f'tmp{j}', 0, 1, dtypes.half) for j in range(i//2)))
       var = UOp.variable(f'tmp{i}', 0, 1, dtypes.half.vec(i))
       acc = UOp.variable('acc', 0, 1, dtypes.half.vec(i))
-      wmma = UOp(Ops.WMMA, dtypes.half.vec(i), (vec, var, acc))
+      wmma = UOp(Ops.WMMA, src=(vec, var, acc))
       uops = to_uops_list([wmma])
       self.assertEqual(uops[-2], wmma)  # -2 to skip SINK
 
@@ -347,7 +347,7 @@ class TestUOpGraph(unittest.TestCase):
                 tuple(UOp.const(dtypes.half, 0.0) for _ in range(i//2)) +
                 tuple(UOp.variable(f'tmp{j}', 0, 1, dtypes.half) for j in range(i//2)))
       acc = UOp.variable('acc', 0, 1, dtypes.half.vec(i))
-      wmma = UOp(Ops.WMMA, dtypes.half.vec(i), (var, vec, acc))
+      wmma = UOp(Ops.WMMA, src=(var, vec, acc))
       uops = to_uops_list([wmma])
       self.assertEqual(uops[-2], wmma)  # -2 to skip SINK
 
@@ -356,7 +356,7 @@ class TestUOpGraph(unittest.TestCase):
                 tuple(UOp.const(dtypes.half, 1.0 if j == 0 else 0.0) for j in range(i)))
       var = UOp.variable(f'tmp{i}', 0, 1, dtypes.half.vec(i))
       acc = UOp.variable('acc', 0, 1, dtypes.half.vec(i))
-      wmma = UOp(Ops.WMMA, dtypes.half.vec(i), (vec, var, acc))
+      wmma = UOp(Ops.WMMA, src=(vec, var, acc))
       uops = to_uops_list([wmma])
       self.assertEqual(uops[-2], wmma)  # -2 to skip SINK
 
@@ -365,7 +365,7 @@ class TestUOpGraph(unittest.TestCase):
       vec = UOp(Ops.STACK, dtypes.half.vec(i),
                 tuple(UOp.const(dtypes.half, 1.0 if j == 0 else 0.0) for j in range(i)))
       acc = UOp.variable('acc', 0, 1, dtypes.half.vec(i))
-      wmma = UOp(Ops.WMMA, dtypes.half.vec(i), (var, vec, acc))
+      wmma = UOp(Ops.WMMA, src=(var, vec, acc))
       uops = to_uops_list([wmma])
       self.assertEqual(uops[-2], wmma)  # -2 to skip SINK
 
@@ -484,16 +484,16 @@ class TestUOpGraph(unittest.TestCase):
     # mnist indexing with split reduceop
     # Make sure we are not doign math on the loaded index, which would promote it to long
     c0 = UOp.param(0, dtypes.uchar, (128000,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 250), 2, AxisType.LOOP)
+    c1 = UOp.range(UOp.const(dtypes.index, 512), 1, AxisType.LOOP)
+    c2 = UOp.range(UOp.const(dtypes.index, 250), 2, AxisType.LOOP)
     c3 = UOp.param(1, dtypes.int, (512,))
     c4 = c3.index(c1)
-    c5 = UOp.range(UOp.const(dtypes.weakint, 240), 0, AxisType.REDUCE)
-    c6 = ((c2*UOp.const(dtypes.weakint, 240))+c5)
+    c5 = UOp.range(UOp.const(dtypes.index, 240), 0, AxisType.REDUCE)
+    c6 = ((c2*UOp.const(dtypes.index, 240))+c5)
     c7 = UOp.param(2, dtypes.uchar, (60000,))
     c8 = c7.index(c6)
     c9 = ((c4<0).where((c4+60000), c4)!=c6.cast(dtypes.int)).where(0, c8.cast(dtypes.uint).cast(dtypes.uchar)).reduce(c5, arg=Ops.ADD)
-    c10 = c0.index(((c1*UOp.const(dtypes.weakint, 250))+c2)).store(c9).end(c1, c2)
+    c10 = c0.index(((c1*UOp.const(dtypes.index, 250))+c2)).store(c9).end(c1, c2)
     uops = to_uops_list([c10])
     for u in uops:
       self.assertNotEqual(u.dtype, dtypes.long)
@@ -501,19 +501,19 @@ class TestUOpGraph(unittest.TestCase):
   def test_load_idx_no_math_on_loaded(self):
     # test the (x+y)<c pattern where x has loads - we shouldn't do math on loaded indices
     c0 = UOp.param(0, dtypes.uchar, (128000,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 250), 2, AxisType.LOOP)
+    c1 = UOp.range(UOp.const(dtypes.index, 512), 1, AxisType.LOOP)
+    c2 = UOp.range(UOp.const(dtypes.index, 250), 2, AxisType.LOOP)
     c3 = UOp.param(1, dtypes.int, (512,))
     c4 = c3.index(c1)  # c4 is a load
-    c5 = UOp.range(UOp.const(dtypes.weakint, 240), 0, AxisType.REDUCE)
-    c6 = ((c2*UOp.const(dtypes.weakint, 240))+c5)
+    c5 = UOp.range(UOp.const(dtypes.index, 240), 0, AxisType.REDUCE)
+    c6 = ((c2*UOp.const(dtypes.index, 240))+c5)
     c7 = UOp.param(2, dtypes.uchar, (60000,))
     c8 = c7.index(c6)
     # (loaded + range) < const pattern - loaded value shouldn't be promoted to long
-    loaded_idx = c4.cast(dtypes.weakint)
-    comparison = (loaded_idx + c5) < UOp.const(dtypes.weakint, 60000)
+    loaded_idx = c4.cast(dtypes.index)
+    comparison = (loaded_idx + c5) < UOp.const(dtypes.index, 60000)
     c9 = comparison.where(c8.cast(dtypes.uint).cast(dtypes.uchar), 0).reduce(c5, arg=Ops.ADD)
-    c10 = c0.index(((c1*UOp.const(dtypes.weakint, 250))+c2)).store(c9).end(c1, c2)
+    c10 = c0.index(((c1*UOp.const(dtypes.index, 250))+c2)).store(c9).end(c1, c2)
     uops = to_uops_list([c10])
     for u in uops:
       self.assertNotEqual(u.dtype, dtypes.long)
@@ -532,7 +532,7 @@ class TestUOpGraph(unittest.TestCase):
   def test_fold_gated_load_local(self):
     glbl0 = UOp.param(0, dtypes.int, (16,))
     smem = UOp.placeholder((18,), dtypes.int, slot=0, addrspace=AddrSpace.LOCAL)
-    lidx = UOp.special(16, "lidx0", dtypes.int)
+    lidx = UOp.special(16, "lidx0")
     st = smem.index(lidx).store(glbl0.index(lidx).load())
     barrier = st.barrier()
     ld0 = smem.after(barrier).index(UOp.invalid())
@@ -557,7 +557,7 @@ class TestUOpGraph(unittest.TestCase):
     glbl0 = UOp.param(0, dtypes.int, (1,))
     idx = UOp.const(dtypes.int, 0)
     bad_gate = UOp.const(dtypes.int, 1)
-    with self.assertRaises(AssertionError): to_uops_list([UOp(Ops.STORE, dtypes.void, (glbl0, idx, UOp.const(dtypes.int, 42), bad_gate))])
+    with self.assertRaises(AssertionError): to_uops_list([UOp(Ops.STORE, src=(glbl0, idx, UOp.const(dtypes.int, 42), bad_gate))])
 
   def test_after_end(self):
     r = UOp.range(10, 0)
@@ -617,7 +617,7 @@ class TestConstBufferize(unittest.TestCase):
     from tinygrad.schedule.rangeify import pm_const_buffer_folding, BufferizeOpts
     c = UOp.const(dtypes.float, 42.0)
     r1 = UOp.range(3, 0)
-    bufferize_with_range = UOp(Ops.STAGE, dtypes.float, (c, r1), arg=BufferizeOpts(device="CPU"))
+    bufferize_with_range = UOp(Ops.STAGE, src=(c, r1), arg=BufferizeOpts(device="CPU"))
     self.assertEqual(len(bufferize_with_range.src), 2)  # const + 1 range
 
     result = graph_rewrite(bufferize_with_range, pm_const_buffer_folding, name='test')
@@ -632,7 +632,7 @@ class TestConstBufferize(unittest.TestCase):
     c = UOp.const(dtypes.float, 3.14)
     r1 = UOp.range(3, 0)
     r2 = UOp.range(4, 1)
-    bufferize_with_ranges = UOp(Ops.STAGE, dtypes.float, (c, r1, r2), arg=BufferizeOpts(device="CPU"))
+    bufferize_with_ranges = UOp(Ops.STAGE, src=(c, r1, r2), arg=BufferizeOpts(device="CPU"))
     self.assertEqual(len(bufferize_with_ranges.src), 3)  # const + 2 ranges
 
     result = graph_rewrite(bufferize_with_ranges, pm_const_buffer_folding, name='test')

test/null/test_uop_symbolic.py

@@ -11,13 +11,13 @@ from tinygrad.uop.validate import uops_to_z3
 def check_uop_against_string(self, v:UOp, s:str):
   sym_vars = {v.render():v for v in v.toposort() if v.op in (Ops.RANGE, Ops.SPECIAL, Ops.PARAM)}
   s_eval = eval(s, sym_vars)
-  if isinstance(s_eval, int) and v.dtype==dtypes.weakint: s_eval = UOp.const(dtypes.weakint, s_eval)
+  if isinstance(s_eval, int) and v.dtype==dtypes.index: s_eval = UOp.const(dtypes.index, s_eval)
   elif isinstance(s_eval, (bool, int, float)): s_eval = UOp.const(dtypes.from_py(s_eval), s_eval)
   s_eval = graph_rewrite(s_eval, commutative, name="cannonicalize eval")
   self.assertIs(s_eval, v, f"eval did not match simplified: {s_eval} != {v.render()} for {s}")
 
-def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.weakint): return UOp.variable(name,min_val,max_val,dtype)
-def uconst(val): return UOp.const(dtypes.weakint, val)
+def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.index): return UOp.variable(name,min_val,max_val,dtype)
+def uconst(val): return UOp.const(dtypes.index, val)
 def usum(ops): return functools.reduce(lambda x,y: x+y, ops)
 def uand(ops): return functools.reduce(lambda x,y: x*y, ops)
 
@@ -247,12 +247,12 @@ class TestSymbolic(unittest.TestCase):
     self.assertEqual((Variable("x", -10, 0)%Variable("y", 1, 10))._min_max, (0, 9))
 
   def test_range_div_its_symbolic_bound(self):
-    a = Variable("a", 1, 10, dtypes.weakint)
+    a = Variable("a", 1, 10, dtypes.index)
     ridx0 = UOp.range(a+2, 0)
     self.helper_test_variable(ridx0//(a+2), 0, 0, "0")
 
   def test_range_mod_its_symbolic_bound(self):
-    a = Variable("a", 1, 10, dtypes.weakint)
+    a = Variable("a", 1, 10, dtypes.index)
     ridx = UOp.range(a+2, 0)
     self.helper_test_variable(ridx%(a+2), 0, 11, "r0")
 
@@ -918,9 +918,9 @@ class TestSymbolic(unittest.TestCase):
     # CAST(bool -> int) != c (c not in {0,1})  ->  always True (CAST is 0 or 1)
     self.helper_test_variable(cond.cast(dtypes.int).ne(2), 1, 1, "True")
     self.helper_test_variable(cond.cast(dtypes.int).ne(-1), 1, 1, "True")
-    # CAST(bool -> weakint) folds too
-    self.helper_test_variable(cond.cast(dtypes.weakint).ne(0), 0, 1, "(a<2)")
-    self.helper_test_variable(cond.cast(dtypes.weakint).ne(1), 0, 1, "((a<2)!=True)")
+    # CAST(bool -> index) folds too
+    self.helper_test_variable(cond.cast(dtypes.index).ne(0), 0, 1, "(a<2)")
+    self.helper_test_variable(cond.cast(dtypes.index).ne(1), 0, 1, "((a<2)!=True)")
 
   def test_where_removal(self):
     cond = Variable("a", 0, 3) < 2
@@ -977,7 +977,7 @@ class TestSymbolic(unittest.TestCase):
 
     # TODO: copied from render, render does not support cast
     glbl = UOp.param(0, dtypes.int, (1,))
-    uops = get_uops(UOp(Ops.STORE, dtypes.void, (glbl.index(UOp.const(dtypes.int, 0)), expr)).sink())
+    uops = get_uops(UOp(Ops.STORE, src=(glbl.index(UOp.const(dtypes.int, 0)), expr)).sink())
     rewritten_uop = [uop for uop in uops if uop.op is Ops.STORE][0].src[1]
 
     # the vars are now scalar PARAMs
@@ -1021,7 +1021,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((numerator//denominator)<=0, 1, 1, "True")
 
   def test_symbolic_range_doesnt_collapse(self):
-    r0 = UOp.range((Variable("a", 1, 10)<5).cast(dtypes.weakint), 0)
+    r0 = UOp.range((Variable("a", 1, 10)<5).cast(dtypes.index), 0)
     self.helper_test_variable(r0, 0, 0, "r0")
 
   def test_const_reciprocal(self):
@@ -1289,16 +1289,16 @@ class TestInvalidIndex(unittest.TestCase):
     self.assertIs((UOp.invalid()<Variable("a",0,10)).simplify().dtype, dtypes.bool)
 
   def test_alu_invalid_vconst(self):
-    c1 = UOp.const(dtypes.weakint, (1, 1, Invalid, Invalid))
-    c2 = UOp.const(dtypes.weakint, (1, Invalid, 1, 1))
-    self.assertIs((c1+c2).simplify(), UOp.const(dtypes.weakint, (2, Invalid, Invalid, Invalid)))
+    c1 = UOp.const(dtypes.index, (1, 1, Invalid, Invalid))
+    c2 = UOp.const(dtypes.index, (1, Invalid, 1, 1))
+    self.assertIs((c1+c2).simplify(), UOp.const(dtypes.index, (2, Invalid, Invalid, Invalid)))
 
 class TestStoreLoadFolding(unittest.TestCase):
   """Tests for store(index, load(index)) -> NOOP rule. This rule matches patterns that EMERGE during simplification."""
   def test_store_load_folding(self):
     # store(idx, load(idx)) -> NOOP, including emergent patterns like store(idx, load(idx) + 0)
     buf = UOp.param(0, dtypes.int, (1,))
-    index = buf.index(UOp.const(dtypes.weakint, 0))
+    index = buf.index(UOp.const(dtypes.index, 0))
     # Direct: store(idx, load(idx)) -> NOOP
     self.assertEqual(graph_rewrite(index.store(index.load()), sym).op, Ops.NOOP)
     # Emergent: store(idx, load(idx) + 0) -> store(idx, load(idx)) -> NOOP
@@ -1317,7 +1317,7 @@ class TestMoveWhereOnLoad(unittest.TestCase):
     cond = (a < 4) & (r < 2)
     valid = (a < 2)  # pre-existing valid on the load (to pass can_move check for the r-only clause)
     idx = buf.index(a.valid(valid))
-    expr = cond.where(idx, 0)
+    expr = cond.where(idx, idx.const_like(0))
     out = graph_rewrite(expr, pm_move_where_on_load)
     # any WHERE in the rewritten graph must have matched-dtype branches
     for u in out.toposort():
@@ -1355,10 +1355,10 @@ class TestGatedUopGivenValid(unittest.TestCase):
 
     idx0 = (r0 + uconst(-1)) // uconst(3)
     idx1 = r0 % uconst(3)
-    idx:UOp = (r0 < 3).where(UOp(Ops.STACK, dtypes.weakint.vec(2), (idx0, idx1)), UOp.invalid())
+    idx:UOp = (r0 < 3).where(UOp(Ops.STACK, src=(idx0, idx1)), UOp.invalid())
     idx = graph_rewrite(idx, pm_simplify_valid)
     # independent simplification: (r0-1)//3 -> (r0+2)//3 - 1, and r0%3 -> r0 when r0 in [0,2]
-    expected_vec = UOp(Ops.STACK, dtypes.weakint.vec(2), ((r0 + uconst(2)) // uconst(3) + uconst(-1), r0))
+    expected_vec = UOp(Ops.STACK, src=((r0 + uconst(2)) // uconst(3) + uconst(-1), r0))
     self.assertEqual(idx, (r0 < 3).where(expected_vec, UOp.invalid()))
 
 class TestRangeSplitting(unittest.TestCase):
@@ -1369,8 +1369,8 @@ class TestRangeSplitting(unittest.TestCase):
     # create a simple expression using the range with mod: store range%2 to a buffer
     buf = UOp.param(0, dtypes.int, (1,))
     val = (r0 % uconst(2)).cast(dtypes.int)
-    store = UOp(Ops.STORE, dtypes.void, (buf.index(uconst(0)), val))
-    sink = UOp(Ops.SINK, dtypes.void, (UOp(Ops.END, dtypes.void, (store, r0)),))
+    store = UOp(Ops.STORE, src=(buf.index(uconst(0)), val))
+    sink = UOp(Ops.SINK, src=(UOp(Ops.END, src=(store, r0)),))
     # count RANGEs before
     ranges_before = len([u for u in sink.toposort() if u.op is Ops.RANGE])
     # apply the range splitting optimization

test/null/test_uop_vmin_vmax.py

@@ -75,7 +75,7 @@ class TestVminVmaxProperties(unittest.TestCase):
     self.assertEqual(uop.vmax, 8)
 
   def test_vmin_vmax_variable_inside_special(self):
-    uop = UOp(Ops.SPECIAL, dtypes.int, arg='gidx0', src=(UOp.variable('i', 1, 10, dtypes.int),))
+    uop = UOp(Ops.SPECIAL, arg='gidx0', src=(UOp.variable('i', 1, 10, dtypes.int),))
     self.assertEqual(uop.vmin, 0)
     self.assertEqual(uop.vmax, 9)
 
@@ -160,7 +160,7 @@ class TestVminVmaxProperties(unittest.TestCase):
     self.assertNotEqual(i.vmin, i.vmax)
 
   def test_vmin_vmax_invalid_vconst(self):
-    x = UOp.const(dtypes.weakint, (0, 4, Invalid, Invalid))
+    x = UOp.const(dtypes.index, (0, 4, Invalid, Invalid))
     self.assertLess(x.vmin, 0)
     self.assertGreater(x.vmax, 4)
 
@@ -318,7 +318,7 @@ class TestVminVmaxVConst(unittest.TestCase):
     # vmin and vmax for a vector constant of bool values
     d1 = UOp.param(1, dtypes.int, (1,))
     idx = UOp.const(dtypes.int, 0)
-    val = UOp(Ops.LOAD, dtypes.int, (d1.index(idx),))
+    val = UOp(Ops.LOAD, src=(d1.index(idx),))
     uop = (val // 32)
     self.assertEqual(uop.vmin, -67108864)
     self.assertEqual(uop.vmax, 67108863)

test/null/test_uops.py

@@ -111,11 +111,11 @@ class TestExecALU(unittest.TestCase):
 class TestGatedStoreRewrite(unittest.TestCase):
   def test_tiny_gate_store(self):
     gmem = UOp.param(0, dtypes.float, (8,))
-    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
-    gate = gidx0<UOp.const(dtypes.int, 1)
-    idx = UOp(Ops.INDEX, dtypes.float, (gmem, (gidx0 * UOp.const(dtypes.int, 2)).valid(gate)))
+    gidx0 = UOp.special(4, 'gidx0')
+    gate = gidx0<UOp.const(dtypes.index, 1)
+    idx = UOp(Ops.INDEX, src=(gmem, (gidx0 * UOp.const(dtypes.index, 2)).valid(gate)))
     val = UOp.const(dtypes.float, 42.0)
-    store = UOp(Ops.STORE, dtypes.void, (idx, val))
+    store = UOp(Ops.STORE, src=(idx, val))
     uops = to_uops_list([store])
     if_uop = next(u for u in uops if u.op is Ops.IF)
     endif = next(u for u in uops if u.op is Ops.ENDIF)
@@ -128,10 +128,10 @@ class TestGatedStoreRewrite(unittest.TestCase):
   def test_gate_some_stores(self):
     gmem0 = UOp.param(0, dtypes.float, (8,))
     gmem1 = UOp.param(1, dtypes.float, (8,))
-    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
-    idx = gidx0 * UOp.const(dtypes.int, 2)
-    idx0 = UOp(Ops.INDEX, dtypes.float, (gmem0, idx.valid(gidx0<UOp.const(dtypes.int, 1))))
-    idx1 = UOp(Ops.INDEX, dtypes.float, (gmem1, idx))
+    gidx0 = UOp.special(4, 'gidx0')
+    idx = gidx0 * UOp.const(dtypes.index, 2)
+    idx0 = UOp(Ops.INDEX, src=(gmem0, idx.valid(gidx0<UOp.const(dtypes.index, 1))))
+    idx1 = UOp(Ops.INDEX, src=(gmem1, idx))
     val = UOp.const(dtypes.float, 42.0)
     stores = [UOp.store(idx0, val), UOp.store(idx1, val)]
     uops = to_uops_list(stores)
@@ -148,11 +148,11 @@ class TestGatedStoreRewrite(unittest.TestCase):
   def test_merge_ifs_alt(self):
     gmem0 = UOp.param(0, dtypes.float, (8,))
     gmem1 = UOp.param(1, dtypes.float, (8,))
-    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
-    idx = gidx0*UOp.const(dtypes.int, 2)
-    gate = gidx0<UOp.const(dtypes.int, 1)
-    idx0 = UOp(Ops.INDEX, dtypes.float, (gmem0, idx.valid(gate)))
-    idx1 = UOp(Ops.INDEX, dtypes.float, (gmem1, idx.valid(gate)))
+    gidx0 = UOp.special(4, 'gidx0')
+    idx = gidx0*UOp.const(dtypes.index, 2)
+    gate = gidx0<UOp.const(dtypes.index, 1)
+    idx0 = UOp(Ops.INDEX, src=(gmem0, idx.valid(gate)))
+    idx1 = UOp(Ops.INDEX, src=(gmem1, idx.valid(gate)))
     val = UOp.const(dtypes.float, 42.0)
     stores = [UOp.store(idx0, val), UOp.store(idx1, val)]
     uops = to_uops_list(stores)
@@ -210,14 +210,14 @@ class TestFastIdiv(unittest.TestCase):
     g = UOp.param(0, dtypes.uint32, (4,))
     c = UOp.const(dtypes.uint, 3)
     l = g.index(c)
-    a = UOp(Ops.CDIV, dtypes.uint, (l, c))
+    a = UOp(Ops.CDIV, src=(l, c))
     uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
     self.assertIn(Ops.SHR, ops)
     self.assertNotIn(Ops.CDIV, ops)
 
-    b = UOp(Ops.CMOD, dtypes.uint, (l, c))
+    b = UOp(Ops.CMOD, src=(l, c))
     uops = to_uops_list([b], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
@@ -244,8 +244,8 @@ class TestFastIdiv(unittest.TestCase):
     # This will be possible with a slightly different method for fast_idiv
     g = UOp.param(0, dtypes.uint32, (8,))
     c = UOp.const(dtypes.uint, 7)
-    l = UOp(Ops.LOAD, dtypes.uint, (g.index(c),))
-    a = UOp(Ops.CDIV, dtypes.uint, (l, c))
+    l = UOp(Ops.LOAD, src=(g.index(c),))
+    a = UOp(Ops.CDIV, src=(l, c))
     uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
@@ -256,7 +256,7 @@ class TestFastIdiv(unittest.TestCase):
     g = UOp.param(0, dtypes.uint32, (4,))
     c = UOp.const(dtypes.uint, 3)
     l = g.index(c)
-    a = UOp(Ops.CDIV, dtypes.uint, (l, c))
+    a = UOp(Ops.CDIV, src=(l, c))
     with Context(DISABLE_FAST_IDIV=1):
       uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
     ops = [x.op for x in uops]
@@ -269,8 +269,8 @@ class TestUOpMethod(unittest.TestCase):
     a = UOp.const(dtypes.float, 2.0)
     b = UOp.const(dtypes.float, 3.0)
 
-    add = UOp(Ops.ADD, dtypes.float, (a, b))
-    mul = UOp(Ops.MUL, dtypes.float, (a, b))
+    add = UOp(Ops.ADD, src=(a, b))
+    mul = UOp(Ops.MUL, src=(a, b))
     assert (add < mul) or (mul < add), "add and mul with same src should have an order"
 
   def test_uop_variables(self):
@@ -282,7 +282,7 @@ class TestUOpMethod(unittest.TestCase):
     self.assertEqual(list(var_vals)[0], a.expr)
 
   def test_const_factor(self):
-    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 8),), 'gidx0')
+    gidx0 = UOp(Ops.SPECIAL, src=(UOp.const(dtypes.int, 8),), arg='gidx0')
     self.assertEqual(UOp.const(dtypes.int, 17).const_factor(), 17)
     self.assertEqual(gidx0.const_factor(), 1)
     self.assertEqual((gidx0*3).const_factor(), 3)
@@ -315,7 +315,7 @@ class TestUOpStr(unittest.TestCase):
     assert str(eval(str(a))) == str(a)
 
   def test_vectorized_str(self):
-    vec = UOp(Ops.STACK, dtypes.int, tuple(UOp.const(dtypes.int, x) for x in range(4)))
+    vec = UOp(Ops.STACK, src=tuple(UOp.const(dtypes.int, x) for x in range(4)))
     assert str(eval(str(vec))) == str(vec)
 
   def test_reduceop_arg(self):
@@ -344,10 +344,10 @@ class TestUopsObject(unittest.TestCase):
 
 class TestUOpRender(unittest.TestCase):
   def test_render_vectorize_empty(self):
-    u = UOp(Ops.STACK, dtype=dtypes.int, src=())
+    u = UOp(Ops.STACK, dtype=dtypes.void, src=())
     self.assertEqual(u.render(simplify=False), "{}")
   def test_render_vectorize_empty_simplified(self):
-    u = UOp(Ops.STACK, dtype=dtypes.int, src=())
+    u = UOp(Ops.STACK, dtype=dtypes.void, src=())
     self.assertEqual(u.render(), "{}")
   def test_render_vectorize_same(self):
     u = UOp(Ops.STACK, dtype=dtypes.int, src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0)))

test/null/test_uops_stats.py

@@ -144,8 +144,8 @@ class TestUOpsStats(unittest.TestCase):
     u1 = globl.index(o1)
     u2 = globl.index(o2)
     u3 = UOp(Ops.CONST, dtypes.int, tuple(), 3)
-    u4 = UOp(Ops.MUL, dtypes.int, (u1,u2))
-    u5 = UOp(Ops.ADD, dtypes.int, (u4,u3))
+    u4 = UOp(Ops.MUL, src=(u1,u2))
+    u5 = UOp(Ops.ADD, src=(u4,u3))
     uops = tuple(u5.toposort())
 
     globl = UOp.param(0, dtypes.int, (3,))
@@ -154,7 +154,7 @@ class TestUOpsStats(unittest.TestCase):
     u1 = globl.index(o1)
     u2 = globl.index(o2)
     u3 = UOp(Ops.CONST, dtypes.int, tuple(), 3)
-    u4 = UOp(Ops.MULACC, dtypes.int, (u1,u2,u3))
+    u4 = UOp(Ops.MULACC, src=(u1,u2,u3))
     uops_fma = tuple(u4.toposort())
 
     self.assertEqual(flops_mem(uops), flops_mem(uops_fma))

test/null/test_validate_oob.py

@@ -126,7 +126,7 @@ class TestValidateOOB(unittest.TestCase):
       buf0 = UOp.param(0, dtypes.int, (16,))
       buf1 = UOp.param(1, dtypes.int, (64,))
       r = UOp.range(42, 0, AxisType.GLOBAL)
-      ld0 = buf0.index(r.valid(r < 8)).load(dtype=dtypes.int).cast(dtypes.weakint)
+      ld0 = buf0.index(r.valid(r < 8)).load(dtype=dtypes.int).cast(dtypes.index)
       to_uops_list([buf1.index((ld0 * 2).valid((ld0 >= 0) & (ld0 < 32))).load(dtype=dtypes.int)])  # valid
       with self.assertRaises(RuntimeError):
         to_uops_list([buf1.index((ld0 * 2).valid((ld0 >= 0) & (ld0 < 64))).load(dtype=dtypes.int)])  # oob
@@ -135,7 +135,7 @@ class TestValidateOOB(unittest.TestCase):
     with Context(CHECK_OOB=1, SPEC=2):
       buf_bool = UOp.param(0, dtypes.bool, (16,))
       buf_int = UOp.param(1, dtypes.int, (8,))
-      gidx = UOp(Ops.SPECIAL, dtypes.weakint, (UOp.const(dtypes.weakint, 16),), "gidx0")
+      gidx = UOp(Ops.SPECIAL, src=(UOp.const(dtypes.index, 16),), arg="gidx0")
       ld_bool = buf_bool.index(gidx).load()
       with self.assertRaises(RuntimeError):
         to_uops_list([buf_int.index(gidx.valid(ld_bool)).load()])  # gidx 0..15, buf_int size 8
@@ -149,21 +149,21 @@ class TestValidateOOB(unittest.TestCase):
       sbuf = UOp.placeholder((8,), dtypes.uint, slot=0, addrspace=AddrSpace.LOCAL)
 
       # Define indices, valids and barrier
-      gidx = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 416),), "gidx0")
-      lidx = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 10),), "lidx0")
+      gidx = UOp(Ops.SPECIAL, src=(UOp.const(dtypes.int, 416),), arg="gidx0")
+      lidx = UOp(Ops.SPECIAL, src=(UOp.const(dtypes.int, 10),), arg="lidx0")
 
       gate = (gidx<400) & (lidx<8)
 
       local_store = sbuf.index(lidx.valid(lidx<8)).store(UOp.const(dtypes.uint, 1))
 
-      barrier = UOp(Ops.BARRIER, dtypes.void, (local_store,))
-      if_barrier = UOp(Ops.IF, dtypes.void, (gate, barrier))
+      barrier = UOp(Ops.BARRIER, src=(local_store,))
+      if_barrier = UOp(Ops.IF, src=(gate, barrier))
 
       # Load from local memory (after the IF/barrier)
-      local_load = UOp(Ops.LOAD, dtypes.uint, (sbuf.index(lidx), if_barrier))
+      local_load = UOp(Ops.LOAD, src=(sbuf.index(lidx), if_barrier))
 
       # Store to global memory
-      global_store = UOp(Ops.STORE, dtypes.void, (gbuf.index(gidx), local_load))
+      global_store = UOp(Ops.STORE, src=(gbuf.index(gidx), local_load))
       to_uops_list([global_store])
 
   @unittest.skip("Bool load is not supported yet")
@@ -172,7 +172,7 @@ class TestValidateOOB(unittest.TestCase):
       glbl0 = UOp.param(0, dtypes.int, (16,))
       mask = UOp.param(0, dtypes.bool, (16,))
       ridx = UOp.range(20, 0)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(ridx, ridx<16&mask))))
+      ld0 = UOp(Ops.LOAD, src=(glbl0.index(UOp.const(ridx, ridx<16&mask))))
       to_uops_list([ld0])
 
 if __name__ == "__main__":

test/null/test_viz.py

@@ -268,12 +268,12 @@ class TestViz(unittest.TestCase):
   def test_stack_movement_not_folded_unless_all_const(self):
     a = UOp.variable("a", 0, 10, dtype=dtypes.int)
     c = UOp.const(dtypes.int, 1)
-    stack = a.vectorize(c)
+    stack = a.stack(c)
     reshaped = stack.reshape((1, 2))
     graph = uop_to_json(VizData(), reshaped)
     self.assertFalse(graph[id(stack)]["exclude"])
 
-    const_stack = c.vectorize(UOp.const(dtypes.int, 2))
+    const_stack = c.stack(UOp.const(dtypes.int, 2))
     const_reshaped = const_stack.reshape((1, 2))
     const_graph = uop_to_json(VizData(), const_reshaped)
     self.assertTrue(const_graph[id(const_stack)]["exclude"])

test/unit/test_disk_tensor.py

@@ -3,7 +3,7 @@ import numpy as np
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.dtype import DType, DTYPES_DICT
 from tinygrad.nn.state import safe_load, safe_save, get_state_dict, torch_load
-from tinygrad.helpers import Timing, fetch, OSX, dedup
+from tinygrad.helpers import Timing, fetch, OSX, dedup, Context
 from test.helpers import slow
 
 class TempDirTestCase(unittest.TestCase):
@@ -88,7 +88,7 @@ class TestRawDiskBuffer(unittest.TestCase):
     # Those two should be moved to test_dtype.py:test_shape_change_bitcast after bitcast works on non-disk
     with self.assertRaises(RuntimeError):
       # should fail because 3 int8 is 3 bytes but float16 is two and 3 isn't a multiple of 2
-      Tensor.empty((3,), dtype=dtypes.int8, device=f"DISK:{tmp}").bitcast(dtypes.float16)
+      Tensor.empty((3,), dtype=dtypes.int8, device=f"DISK:{tmp}").bitcast(dtypes.float16).shape
 
     pathlib.Path(tmp).unlink()
 
@@ -410,6 +410,13 @@ class TestDiskTensor(TempDirTestCase):
       on_dev = t.to(Device.DEFAULT).realize()
       np.testing.assert_equal(on_dev.numpy(), t.numpy())
 
+  def test_shard_copy_from_disk_slice(self):
+    fn = pathlib.Path(self.tmp("dt_shard_copy_from_disk_slice"))
+    fn.write_bytes(bytes(range(32)))
+    with Context(CACHELEVEL=0):
+      t = Tensor.empty(8, 4, device=f"disk:{fn}", dtype=dtypes.uint8)[0:4].shard(("CPU:0", "CPU:1"), axis=0).realize()
+      np.testing.assert_equal(t.to("CPU").numpy(), np.arange(16, dtype=np.uint8).reshape(4, 4))
+
   @slow
   def test_copy_from_disk_huge(self):
     fn = pathlib.Path(self.tmp("dt_copy_from_disk_huge"))

test/unit/test_metal_graph.py

@@ -17,7 +17,7 @@ class TestMetalGraph(unittest.TestCase):
       buf.op = Ops.SLICE
       src = MagicMock()
       src.dtype = dtypes.uint8
-      buf.src = (src, UOp.const(dtypes.weakint, offset))
+      buf.src = (src, UOp.const(dtypes.index, offset))
       buf.dtype = dtypes.uint8
     else:
       buf.op = Ops.BUFFER

test/unit/test_multitensor.py

@@ -78,8 +78,9 @@ class TestMultiTensor(unittest.TestCase):
     self.assertEqual(Y.device, devices_2)
     np.testing.assert_equal(X.numpy(), Y.numpy())
 
-    with self.assertRaises(AssertionError):
-      _ = Tensor(X.uop, dtype=dtypes.float)
+    Z = Tensor(X.uop, dtype=dtypes.float)
+    self.assertEqual(Z.dtype, dtypes.float)
+    np.testing.assert_equal(Z.numpy(), [1.0, 2.0])
 
   def test_sharded_arange(self):
     sharded_arange = Tensor.arange(1000).clone().shard(devices_2, 0)

tinygrad/callify.py

@@ -1,5 +1,5 @@
 from dataclasses import dataclass, field
-from tinygrad.dtype import dtypes, AddrSpace, PtrDType, ImageDType
+from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.uop.ops import UOp, UPat, PatternMatcher, Ops, GroupOp, ParamArg, graph_rewrite, track_rewrites
 from tinygrad.helpers import VIZ, pluralize, all_int
 
@@ -50,10 +50,10 @@ def replace_contig_with_store_after(u:UOp):
 def replace_store_after_with_contig(u:UOp, src:UOp):
   assigned_to = u
   while assigned_to.op in {Ops.BITCAST, Ops.AFTER, Ops.MULTI}: assigned_to = assigned_to.src[0].base
-  if assigned_to.op is not Ops.BUFFER: return src.contiguous(tag=u.tag)
+  if assigned_to.op not in {Ops.BUFFER, Ops.SLICE}: return src.contiguous(tag=u.tag)
 
 def _make_buffer_view(src:UOp) -> UOp|None:
-  """If movement ops on src collapse to a contiguous range, return SLICE.reshape(src.shape). Otherwise None."""
+  """If movement ops on src collapse to a contiguous range, return SLICE. Otherwise None."""
   if (offset := src.contiguous_view_offset()) is None: return None
   buf = src.base
   if buf.op is Ops.SLICE:
@@ -61,25 +61,26 @@ def _make_buffer_view(src:UOp) -> UOp|None:
     buf = buf.src[0]
     if byte_offset % buf.dtype.itemsize != 0: return None
     offset = byte_offset // buf.dtype.itemsize
-  return UOp(Ops.SLICE, src.dtype, (buf, UOp.const(dtypes.weakint, offset)), src.numel()).reshape(src.shape)
+  return UOp(Ops.SLICE, src.dtype, (buf, UOp.const(dtypes.index, offset)), src.numel())
 
 def contiguous_mops_to_view(c:UOp, src:UOp):
-  """CONTIGUOUS(MOPS(BUFFER)) → CONTIGUOUS(SLICE) when movement ops collapse to a contiguous range."""
+  """MOPS(BUFFER) → SLICE when movement ops collapse to a contiguous range."""
   buf = src.base
   if buf.op not in {Ops.BUFFER, Ops.SLICE, Ops.MULTI}: return None
-  if src.op is Ops.RESHAPE and src.src[0].op in {Ops.BUFFER, Ops.SLICE}: return None
+  if src.op is Ops.RESHAPE and src.src[0].op in {Ops.BUFFER, Ops.SLICE} and c.op is not Ops.BITCAST: return None
+  if c.op is not Ops.BITCAST and src.op is Ops.BUFFER: return None
 
   # no symbolic shape
   if not all_int(c.shape): return None
 
   # check if view is supported
   from tinygrad.device import Device
-  devs = (c.device,) if isinstance(c.device, str) else c.device
+  devs = (src.device,) if isinstance(src.device, str) else src.device
   if not all(hasattr(Device[d].allocator, "_offset") for d in devs): return None
 
-  # NOTE: this contiguous is removed because this SLICE/RESHAPE has_buffer_identity
   if buf.op is not Ops.MULTI and (view := _make_buffer_view(src)) is not None:
-    return view.contiguous(tag=c.tag)
+    view = (view.replace(dtype=c.dtype, arg=c.numel()) if c.op is Ops.BITCAST else view).reshape(c.shape)
+    return c.replace(src=(view,)) if c.op is Ops.COPY else view
 
   # for MULTI tensors, use multi_pm to resolve per-shard movement ops, then create SLICE on the resolved result
   if not isinstance(c.device, str):
@@ -87,7 +88,7 @@ def contiguous_mops_to_view(c:UOp, src:UOp):
     resolved = graph_rewrite(src, multi_pm, name="multi_buffer_view")
     if resolved.op is not Ops.MULTI: return None
     if (view := _make_buffer_view(resolved.src[0])) is None: return None
-    return view.multi(resolved.arg).contiguous(tag=c.tag)
+    return view.reshape(resolved.src[0].shape).multi(resolved.arg).contiguous(tag=c.tag)
 
   return None
 
@@ -125,7 +126,7 @@ def transform_precompiled_call(c:UOp) -> UOp|None:
   fxn = UOp.sink(*(x.substitute(subs) for x in items))
 
   # body switches from TUPLE to SINK, so the node becomes an opaque CALL (not FUNCTION)
-  new_call = UOp(Ops.CALL, c.dtype, (fxn, *input_buffers, *outs), c.arg)
+  new_call = UOp(Ops.CALL, src=(fxn, *input_buffers, *outs), arg=c.arg)
   rets = tuple(o.after(new_call) for o in outs)
 
   # if the CALL has symbolic shapes, shrink the max-sized output to the actual symbolic shape
@@ -142,8 +143,16 @@ pm_early_transform_tensor_graph = PatternMatcher([
   # resolve TUPLE+GETTUPLE (for precompiled calls)
   (UPat(Ops.GETTUPLE, src=(UPat(Ops.TUPLE, name="t"),), name="g"), lambda g,t: t.src[g.arg]),
 
-  # CONTIGUOUS(MOPS(BUFFER/SLICE)) → CONTIGUOUS(SLICE) when movement ops collapse to contiguous range
-  (UPat(Ops.CONTIGUOUS, src=(UPat(GroupOp.Movement, name="src"),), name="c"), contiguous_mops_to_view),
+  # fold MOPS+BITCAST over BUFFER/SLICE into SLICE when movement ops collapse to contiguous range
+  (UPat((Ops.BITCAST, Ops.COPY, Ops.CONTIGUOUS), src=(UPat(GroupOp.Movement|{Ops.BUFFER}, name="src"),), name="c"), contiguous_mops_to_view),
+
+  # remove contiguous on movement ops before a copy on disk
+  (UPat(GroupOp.Movement-{Ops.SHRINK, Ops.RESHAPE}, name="x").f(Ops.CONTIGUOUS).f(Ops.COPY, allow_any_len=True, name="copy"), lambda x,copy:
+   copy.replace(src=(x,)+copy.src[1:], tag=None) if isinstance(x.device, str) and x.device.startswith("DISK") else None),
+  # push copy past movement ops to disk
+  (UPat(GroupOp.Movement-{Ops.SHRINK, Ops.RESHAPE}, name="x").f(Ops.COPY, name="copy"), lambda x,copy:
+   x.replace(src=(copy.replace(src=(x.src[0],)+copy.src[1:], tag=None),)+x.src[1:]) \
+   if isinstance(x.device, str) and x.device.startswith("DISK") else None),
 
   # add CONTIGUOUS to tagged UOps
   (UPat(GroupOp.All-{Ops.CONTIGUOUS, Ops.AFTER, Ops.STORE}, name="x"),
@@ -177,7 +186,7 @@ def replace_input_buffer(ctx:AllocCtx, b:UOp):
   ctx.replacements.append(b)
   return UOp.param(len(ctx.replacements)-1, b.dtype, b.shape, b.device,
                    b._min_max if b.op is Ops.BIND else None, b.src[0].expr if b.op is Ops.BIND else None,
-                   b.addrspace if isinstance(b.dtype, (PtrDType, ImageDType)) else AddrSpace.GLOBAL)
+                   b.addrspace if b.addrspace is not None else AddrSpace.GLOBAL)
 
 pm_finalize_call = PatternMatcher([
   (UPat(Ops.AFTER, name="x"), finalize_after),
@@ -189,7 +198,7 @@ pm_replace_buf = PatternMatcher([
   (UPat(Ops.BUFFER, src=(UPat(),), name="b"), lambda ctx,b:
    replace_input_buffer(ctx, b) if isinstance(b.arg, ParamArg) and b.addrspace is AddrSpace.GLOBAL else None),
   # replace SLICE with PARAM. this rewrite is bottom up so BUFFERs we don't need won't be in the input
-  (UPat(Ops.SLICE, src=(UPat(Ops.BUFFER), UPat(Ops.CONST, dtype=dtypes.weakint)), name="b"), replace_input_buffer),
+  (UPat(Ops.SLICE, src=(UPat(Ops.BUFFER), UPat(Ops.CONST, dtype=dtypes.index)), name="b"), replace_input_buffer),
   # strip value from BIND for cache key normalization, so different values hit same cache
   (UPat(Ops.BIND, src=(UPat(Ops.PARAM), UPat(Ops.CONST)), name="b"), replace_input_buffer),
 ])

tinygrad/codegen/__init__.py

@@ -13,6 +13,7 @@ from tinygrad.dtype import dtypes, AddrSpace
 # import all pattern matchers here
 from tinygrad.codegen.gpudims import pm_add_gpudims
 from tinygrad.uop.symbolic import sym, symbolic_simple, symbolic, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
+from tinygrad.uop.movement import mop_cleanup
 from tinygrad.codegen.decomp.dtype import pm_dtype_decomps
 from tinygrad.codegen.decomp.op import get_late_rewrite_patterns, get_simplifying_rewrite_patterns
 from tinygrad.codegen.decomp.transcendental import get_transcendental_patterns
@@ -20,7 +21,7 @@ from tinygrad.codegen.late.coalese import indexing_simplify
 from tinygrad.codegen.opt.postrange import apply_opts
 from tinygrad.codegen.late.gater import pm_move_gates_from_index
 from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse
-from tinygrad.schedule.rangeify import pm_mops, pm_syntactic_sugar, mop_cleanup
+from tinygrad.schedule.rangeify import pm_mops
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize
 from tinygrad.codegen.late.regalloc import LinearScanRegallocContext, pm_regalloc_rewrite
 from tinygrad.codegen.late.coalese import memory_coalesing, pm_simplify_add_image
@@ -37,8 +38,9 @@ pm_number_params = PatternMatcher([
   (UPat(Ops.PARAM, name="x"), do_number_param),
 ])
 
-pm_no_weakints = PatternMatcher([
-  (UPat(GroupOp.All, dtype=dtypes.weakint, name="x"), lambda x: x.replace(dtype=dtypes.int))
+pm_no_index = PatternMatcher([
+  (UPat(GroupOp.ALU.union({Ops.CONST}), dtype=dtypes.index, name="x"), lambda x: x.replace(dtype=dtypes.int)),
+  (UPat(Ops.CAST, dtype=dtypes.index, src=(UPat.var("x"),)), lambda x: x.cast(dtypes.int)),
 ])
 
 def build_range_map(sink:UOp) -> dict[int, int]:
@@ -107,13 +109,13 @@ def broadcast_and_devec_wmma(b:UOp):
                   for u,shp in zip(b.src, shaped_aligned)]
   src = []
   for idx in itertools.product(*[range(i) for i in b.shape[:-1]]):
-    idx_c = [UOp.const(dtypes.weakint, i) for i in idx]
+    idx_c = [UOp.const(dtypes.index, i) for i in idx]
     src.append(b.replace(src=tuple([x.index(*idx_c) for x in src_reshaped])))
-  return UOp.vectorize(*src).reshape(b.shape)
+  return UOp.stack(*src).reshape(b.shape)
 
 pm_wmma_add = PatternMatcher([
   (UPat(Ops.WMMA, name="wmma") + UPat.var("add"),
-   lambda add, wmma: UOp(wmma.op, wmma.dtype, (wmma.src[0], wmma.src[1], wmma.src[2]+add), wmma.arg)),
+   lambda add, wmma: UOp(wmma.op, src=(wmma.src[0], wmma.src[1], wmma.src[2]+add), arg=wmma.arg)),
   # push permute/reshape to the other side of the add
   (UPat(Ops.PERMUTE, src=(UPat(Ops.WMMA, name="wmma"),), name="permute") + UPat.var("add"),
     lambda wmma,permute,add: (wmma + add.permute(argsort(permute.arg))).permute(permute.arg)),
@@ -132,9 +134,9 @@ def do_devectorize(b:UOp):
   if not all_same([x.shape for x in b.src]): return None
   src = []
   for idx in itertools.product(*[range(x) for x in b.shape]):
-    idx_c = [UOp.const(dtypes.weakint, i) for i in idx]
+    idx_c = [UOp.const(dtypes.index, i) for i in idx]
     src.append(b.replace(src=tuple([x.index(*idx_c) for x in b.src])))
-  return UOp.vectorize(*src).reshape(b.shape) if b.op is not Ops.STORE else UOp.group(*src)
+  return UOp.stack(*src).reshape(b.shape) if b.op is not Ops.STORE else UOp.group(*src)
 
 def do_stack_wmma(u:UOp):
   if all(x.op in (Ops.STACK, Ops.WMMA) for x in u.src): return None
@@ -142,7 +144,7 @@ def do_stack_wmma(u:UOp):
   src = []
   for b in u.src:
     if b.op != Ops.STACK:
-      src.append(UOp._stack(*[b.index(UOp.const(dtypes.weakint, i)) for i in range(b.max_numel())]))
+      src.append(UOp.stack(*[b.index(UOp.const(dtypes.index, i)) for i in range(b.max_numel())]))
     else:
       src.append(b)
   return u.replace(src=tuple(src))
@@ -152,29 +154,26 @@ ew_devectorizer = PatternMatcher([
   (UPat(GroupOp.Elementwise, name="b"), do_devectorize),
 ])
 
-devectorizer2 = pm_mops+PatternMatcher([
+devectorizer2 = mop_cleanup+pm_mops+PatternMatcher([
   # unpack broadcasting
   (UPat(GroupOp.Elementwise|{Ops.LOAD,Ops.STORE}, name="b"), do_devectorize),
-  # const INDEX into STACK is src (this is symbolic)
-  (UPat(Ops.INDEX, src=(UPat(Ops.STACK, name="a"), UPat.cvar("i")), name="idx", allow_any_len=True),
-   lambda a,i,idx: a.src[i.arg].index(*idx.src[2:])),
-  # INDEX without src is nothing
+  # INDEX without src is nothing (TODO: this should be in mop_cleanup)
   (UPat(Ops.INDEX, src=(UPat.var('x'),)), lambda x: x),
   # unpack WMMA
   (UPat(Ops.WMMA, name="u"), do_stack_wmma),
   # stacked INDEX is many INDEX
   (UPat(Ops.INDEX, src=(UPat((Ops.PARAM, Ops.BUFFER), name="b"), UPat(Ops.STACK, name="s"))),
-   lambda b,s: UOp.vectorize(*[b.index(u) for u in s.src])),
+   lambda b,s: UOp.stack(*[b.index(u) for u in s.src])),
   # INDEX into RESHAPE moves the RESHAPE
   (UPat(Ops.INDEX, src=(UPat((Ops.PARAM, Ops.BUFFER), name="b"), UPat(Ops.RESHAPE, name="s"))),
    lambda b,s: b.index(s.src[0]).reshape(s.shape)),
   # RESHAPE a void is removed (hack for AFTER)
   (UPat(Ops.RESHAPE, dtype=dtypes.void, name="x"), lambda x: x.src[0]),
   # reshape of a single element shaped value to scalar is an index
-  (UPat(Ops.RESHAPE, name="x"), lambda x: x.src[0].index(UOp.const(dtypes.weakint, 0)) if x.marg == () and x.src[0].shape == (1,) else None),
+  (UPat(Ops.RESHAPE, name="x"), lambda x: x.src[0].index(UOp.const(dtypes.index, 0)) if x.marg == () and x.src[0].shape == (1,) else None),
   # EXPAND on scalar -> STACK
   (UPat(Ops.EXPAND, src=(UPat.var("x"), UPat()), name="out"),
-   lambda x,out: UOp.vectorize(*([x]*out.max_numel())) if x.shape == () and out.shape == (out.max_numel(),) else None),
+   lambda x,out: UOp.stack(*([x]*out.max_numel())) if x.shape == () and out.shape == (out.max_numel(),) else None),
   # INDEX on INDEX is INDEX
   (UPat(Ops.INDEX, src=(UPat(Ops.INDEX, name="idx1", allow_any_len=True),), allow_any_len=True, name="idx2"),
    lambda idx1, idx2: idx1.src[0].index(*idx1.src[1:], *idx2.src[1:])),
@@ -196,11 +195,6 @@ def fix_group_for_reduce(x:UOp):
   # NOTE: we remove all horizontal reduces here, they remain in the first reduce
   return buf.reduce(*reduce_loop, arg=(x.arg[0], 0))
 
-pm_group_for_reduce = PatternMatcher([
-  # fix group for reduce
-  (UPat(Ops.REDUCE, name="x"), fix_group_for_reduce),
-])
-
 @dataclass
 class ReduceContext:
   acc_num: int = 0
@@ -231,7 +225,7 @@ def reduce_ranges_to_acc(ctx:ReduceContext, r:UOp):
   topo = r.src[0].toposort()
   ended_ranges = flatten([x.ended_ranges for x in topo if x.op is Ops.END])
   input_ranges = tuple(x for x in topo if x.op is Ops.RANGE and x not in r.src[1:] and x not in ended_ranges)
-  acc_init = acc.after(*input_ranges).store(identity_element(r.arg[0], r.dtype.scalar()))
+  acc_init = acc.after(*input_ranges).store(identity_element(r.arg[0], r.dtype))
   acc_initted = acc.after(acc_init, *r.src[1:])
   inp = r.src[0].reduce(arg=r.arg) if r.arg[1] else r.src[0]
   acc_out = acc_initted.store(acc_initted.alu(r.arg[0], inp)).end(*r.src[1:]).rtag("mergeable")
@@ -243,13 +237,16 @@ def expand_horizontal_reduce(r:UOp):
   return functools.reduce(lambda x,y: x.alu(r.arg[0], y), vals)
 
 pm_reduce_local = pm_wmma_add+PatternMatcher([
+  # fix group for reduce
+  (UPat(Ops.REDUCE, name="x"), fix_group_for_reduce),
+  # remove reduces
   (UPat(Ops.REDUCE, src=(UPat(), UPat()), allow_any_len=True, name="r"), reduce_ranges_to_acc),
   (UPat(Ops.REDUCE, src=(UPat(),), name="r"), expand_horizontal_reduce),
   (UPat(Ops.SINK, name="sink"), merge_reduce_ends),
 ])+pm_clean_up_group_sink
 
 def maybe_load(u:UOp): return u.load() if u.addrspace in (AddrSpace.GLOBAL, AddrSpace.LOCAL, AddrSpace.REG) else u
-pm_move_regs = PatternMatcher([
+pm_add_loads = PatternMatcher([
   # BITCAST?
   (UPat(GroupOp.Elementwise|{Ops.REDUCE,Ops.WMMA,Ops.STACK}, name="x"), lambda x: x.replace(src=tuple([maybe_load(u) for u in x.src]))),
   (UPat(Ops.STORE, name="x"), lambda x: x.replace(src=(x.src[0], maybe_load(x.src[1]))+x.src[2:])),
@@ -269,7 +266,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   if SPEC: type_verify(ast, spec_tensor)
 
   # preprocess
-  sink = graph_rewrite(ast, pm_mops+pm_syntactic_sugar, ctx=itertools.count(1000), name="early movement ops", bottom_up=True)
+  sink = graph_rewrite(ast, pm_mops, name="early movement ops", bottom_up=True)
 
   # first we optimize
   if optimize:
@@ -293,24 +290,19 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
 
   # expand
   sink = graph_rewrite(sink, expander2, ctx=build_range_map(sink), name="expander")
-  sink = graph_rewrite(sink, pm_group_for_reduce, name="group for reduce")
+
+  # remove reduce
+  sink = graph_rewrite(sink, mop_cleanup+pm_reduce_local, ctx=ReduceContext(), name="remove reduces")
 
   # add locals
   sink = graph_rewrite(sink, pm_add_local_buffers, ctx=itertools.count(0), name="add local buffers")
 
-  # ** devectorizer (full_graph_rewrite) **
-  # remove reduce
-  sink = graph_rewrite(sink, mop_cleanup+pm_reduce_local, ctx=ReduceContext(), name="remove_reduce")
-
   # add gpu dims (late). this works after devectorize, but it's faster here
   sink = graph_rewrite(sink, pm_add_gpudims, ctx=ren, name="add gpudims")
 
   # **** optimizations are done, now we lower to actual code ****
 
-  sink = graph_rewrite(sink, symbolic_simple+unbroadcast, name="*** unbroadcast")
-
-  # add loads and remove invalids
-  sink = graph_rewrite(sink, pm_move_regs, name="** add loads")
+  sink = graph_rewrite(sink, symbolic_simple+unbroadcast+pm_add_loads, name="*** unbroadcast / add loads")
 
   # devectorize
   sink = graph_rewrite(sink, symbolic_simple+devectorizer2, ctx=ren, name="devectorize2")
@@ -337,9 +329,6 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
 
   # **** decomps ****
 
-  # optional pre matcher
-  if ren.pre_matcher is not None: sink = graph_rewrite(sink, ren.pre_matcher, name="pre_matcher")
-
   # floordiv+mod / dtype decomp (early)
   supported_ops = tuple(ren.code_for_op.keys())
   pm_decomp = symbolic_simple+get_simplifying_rewrite_patterns(supported_ops)
@@ -355,7 +344,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
 
   # final rules for the renderer (without sym)
   extra_matcher = ren.extra_matcher if ren.extra_matcher is not None else PatternMatcher([])
-  pm_final_rewrite = pm_decomp+extra_matcher+pm_split_ends+pm_no_weakints
+  pm_final_rewrite = pm_decomp+extra_matcher+pm_split_ends+pm_no_index
   sink = graph_rewrite(sink, pm_final_rewrite+pm_remove_invalid, ctx=ren, name="final rewrite")
 
   # this was the linearizer

tinygrad/codegen/decomp/dtype.py

@@ -1,7 +1,8 @@
+from dataclasses import replace
 from tinygrad.dtype import dtypes, DType, truncate
-from tinygrad.helpers import flatten, DEBUG, EMULATED_DTYPES
+from tinygrad.helpers import flatten, DEBUG, EMULATED_DTYPES, Context, SPEC
 from tinygrad.uop import GroupOp
-from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, graph_rewrite
+from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, graph_rewrite, ParamArg
 from tinygrad.renderer import Renderer
 from tinygrad.codegen.decomp.transcendental import exponent_bias, shl, shr
 
@@ -112,7 +113,7 @@ def f2f_clamp(val:UOp, dt:DType, sat=True) -> UOp:
 
 def f2f_load(x: UOp, fr:DType, to:DType) -> UOp:
   if (n:=x.max_numel()) == 1: return f2f(x.replace(dtype=f2f_dt[fr]), fr, to)
-  return UOp(Ops.STACK, to, tuple(f2f(x.replace(dtype=f2f_dt[fr], src=(reindex(x.src[0], i, 1),)), fr, to) for i in range(n)))
+  return UOp(Ops.STACK, src=tuple(f2f(x.replace(dtype=f2f_dt[fr], src=(reindex(x.src[0], i, 1),)), fr, to) for i in range(n)))
 
 def f2f_store(st, idx, val, fr:DType, to:DType):
   if (n:=val.max_numel()) == 1: return st.replace(src=(idx, f2f(val.bitcast(f2f_dt[to]), to, fr)))
@@ -120,7 +121,7 @@ def f2f_store(st, idx, val, fr:DType, to:DType):
 
 pm_long_decomp = PatternMatcher([
   (UPat(GroupOp.Defines, src=(UPat.var("sz"),), name="x"), lambda x,sz:
-   x.replace(dtype=l2i_dt[x.dtype], src=(sz*2,)) if x.dtype in l2i_dt else None),
+   x.replace(dtype=l2i_dt[x.dtype], arg=replace(x.arg, dtype=l2i_dt[x.dtype]), src=(sz*2,)) if x.dtype in l2i_dt else None),
   (UPat(Ops.INDEX, tuple(l2i_dt.keys()), name='x'), lambda x: reindex(x, x.tag).replace(dtype=l2i_dt[x.dtype]) if x.tag is not None else None),
   (UPat(Ops.STORE, src=(UPat.var('idx'), UPat.var('val', tuple(l2i_dt.keys()))), name='st'), lambda st,idx,val:
    st.replace(src=(idx.rtag(0), val.rtag(0))).group(st.replace(src=(idx.rtag(1), val.rtag(1)))) if val.tag is None else None),
@@ -144,7 +145,7 @@ pm_long_decomp = PatternMatcher([
 # float decomposition patterns - ctx is (fr, to) tuple
 pm_float_decomp = PatternMatcher([
   (UPat((*GroupOp.Defines, Ops.INDEX, Ops.SHRINK), name="x"), lambda ctx,x:
-   x.replace(dtype=f2f_dt[ctx[0]], tag=ctx[0])
+   x.replace(dtype=f2f_dt[ctx[0]], arg=replace(x.arg, dtype=f2f_dt[ctx[0]]) if isinstance(x.arg, ParamArg) else x.arg, tag=ctx[0])
    if x.dtype == ctx[0] and (x.op is not Ops.INDEX or x.src[0].op not in {Ops.LOAD, Ops.STACK}) else None),
   (UPat(Ops.LOAD, dtypes.floats, name="x"), lambda ctx,x: f2f_load(x, *ctx) if x.dtype == ctx[0] else None),
   # bitcasted load should just replace load
@@ -169,10 +170,13 @@ pm_float_decomp = PatternMatcher([
 
 def do_dtype_decomps(sink:UOp, ctx:tuple[set[DType], Renderer]) -> UOp:
   def _should_emulate(dt): return dt in EMULATED_DTYPES.tolist(dtypes) or dt not in ctx[1].supported_dtypes()
-  for fr in sorted(filter(_should_emulate, ctx[0])):
-    to = dtypes.int if fr == dtypes.long else dtypes.half if not _should_emulate(dtypes.half) and fr in dtypes.fp8s else dtypes.float
-    if DEBUG >= 2: print(f"emulating {fr} as {to}")
-    sink = graph_rewrite(sink, pm_float_decomp if fr in dtypes.floats else pm_long_decomp, name=f"decomp {fr} -> {to}", ctx=(fr, to), bottom_up=True)
+  # NOTE: dtype decomp creates intermediate UOps that don't follow the spec (e.g. half LOAD on ushort BUFFER)
+  with Context(SPEC=min(SPEC.value, 1)):
+    for fr in sorted(filter(_should_emulate, ctx[0])):
+      to = dtypes.int if fr == dtypes.long else dtypes.half if not _should_emulate(dtypes.half) and fr in dtypes.fp8s else dtypes.float
+      if DEBUG >= 2: print(f"emulating {fr} as {to}")
+      pm = pm_float_decomp if fr in dtypes.floats else pm_long_decomp
+      sink = graph_rewrite(sink, pm, name=f"decomp {fr} -> {to}", ctx=(fr, to), bottom_up=True)
   ctx[0].clear()
   return sink
 

tinygrad/codegen/gpudims.py

@@ -1,6 +1,6 @@
 import math
 from tinygrad.uop.ops import UOp, Ops, sint, PatternMatcher, UPat, KernelInfo, ssimplify, AxisType
-from tinygrad.dtype import dtypes, AddrSpace, Invalid
+from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.renderer import Renderer
 
 def _dim_max(d:sint) -> int: return d if isinstance(d, int) else int(d.vmax)
@@ -57,7 +57,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
 
   # get the idxs
   ki: KernelInfo = s.arg
-  if ctx.has_threads: idxs = [UOp.variable("core_id", 0, int(global_shape[0])-1, dtypes.int).cast(dtypes.weakint)]
+  if ctx.has_threads: idxs = [UOp.variable("core_id", 0, int(global_shape[0])-1, dtypes.int).cast(dtypes.index)]
   elif ki.dont_use_locals:
     assert not local_dims, "can't use locals if there's no local dims"
     idxs = get_grouped_dims("idx", global_shape, ctx.global_max, reverse=True)
@@ -78,7 +78,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
       if len(missing_locals):
         assert len(idx.src) == 2, "index has 2 sources"
         mask: UOp = UOp.uprod(*[x.eq(0) for x in missing_locals])
-        subs[idx] = idx.replace(src=(idx.src[0], mask.broadcast(idx.src[1].dtype.count).where(idx.src[1], Invalid)))
+        subs[idx] = idx.replace(src=(idx.src[0], idx.src[1].valid(mask)))
     if r.op is not Ops.RANGE: continue
     try:
       ii = (global_dims+local_dims).index(r.arg[0:-1])

tinygrad/codegen/late/coalese.py

@@ -1,10 +1,9 @@
-from typing import Any
 import itertools, functools
 from collections import defaultdict
-from tinygrad.dtype import dtypes, Invalid, ImageDType, DType
-from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp
+from tinygrad.dtype import dtypes, AddrSpace, Invalid, DType
+from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp, shape_to_shape_arg
 from tinygrad.uop.symbolic import uop_given_valid, parse_valid, invalid_gate
-from tinygrad.helpers import getenv, IMAGE, OSX, ceildiv
+from tinygrad.helpers import getenv, IMAGE, OSX, ceildiv, is_image_shape
 from tinygrad.renderer import Renderer
 
 # ***** image load valid simplification *****
@@ -40,15 +39,17 @@ def simplify_valid_load(buf:UOp, start_idx:UOp, valid:UOp) -> UOp|None:
   return None if idx is start_idx or idx is start_idx.simplify() else buf.index(idx.valid(valid))
 
 def simplify_valid_image_load(buf:UOp, idx_y:UOp, idx_x:UOp, valid:UOp) -> UOp|None:
-  if not isinstance(buf.dtype, ImageDType): return None
-  start_idx = idx_x._stack(idx_y)
+  if not is_image_shape(buf._shape): return None
+  if idx_x.dtype != idx_y.dtype: idx_x, idx_y = idx_x.cast(dtypes.int), idx_y.cast(dtypes.int)
+  start_idx = idx_x.stack(idx_y)
   idx = uop_given_valid(valid, start_idx)
-  drop_stmt = _drop_valid_stmts(valid, idx, buf.dtype.shape[0], buf.dtype.shape[1])
+  drop_stmt = _drop_valid_stmts(valid, idx, buf._shape[0], buf._shape[1])
 
   if not drop_stmt and idx is start_idx: return None
   new_valid = UOp.uprod(*ss) if (ss:=[s for s in valid.split_uop(Ops.AND) if s not in drop_stmt]) else None
   idx_y, idx_x = idx.index(1), idx.index(0)
-  return buf.index(idx_y.valid(new_valid), idx_x.valid(new_valid)) if new_valid is not None else buf.index(idx_y, idx_x)
+  if new_valid is not None: return buf.index(idx_y.valid(new_valid), idx_x.valid(new_valid), dtype=dtypes.float)
+  return buf.index(idx_y, idx_x, dtype=dtypes.float)
 
 indexing_simplify = PatternMatcher([
   # image load valid idx simplification
@@ -69,12 +70,11 @@ def image_valid_dims(base:DType, size:int, arch:str) -> list[tuple[int,int]]:
 def transform_to_image(ctx, buf:UOp, x:UOp) -> UOp|None:
   shapes, ren = ctx
   if not IMAGE or ren.target.device not in {"QCOM", "CL", "PYTHON", "NULL"}: return None
-  valid = UOp.const(dtypes.bool, True)
-  if x.op == Ops.WHERE and x.src[2].op == Ops.CONST and x.src[2].arg == Invalid: valid,x,_= x.src
+  valid, x = x.get_valid(), x.get_idx()
   # search for dims that drop the most valid statements
   best_drop, cands = -1, []
   for ch, cw in [shapes[buf.arg.slot]] if buf.arg.slot in shapes else image_valid_dims(buf.dtype, buf.max_numel(), ren.target.arch):
-    cidx = uop_given_valid(valid, ((x//4)%cw)._stack(x//(4*cw)))
+    cidx = uop_given_valid(valid, ((x//4)%cw).stack(x//(4*cw)))
     dropped = len(_drop_valid_stmts(valid, cidx, ch, cw))
     if dropped > best_drop: best_drop, cands = dropped, [(ch, cw, cidx)]
     elif dropped == best_drop: cands.append((ch, cw, cidx))
@@ -82,13 +82,12 @@ def transform_to_image(ctx, buf:UOp, x:UOp) -> UOp|None:
   if len(cands) == 0: return None
   # and tiebreak with indexing complexity (ie. number of nodes)
   h, w, cidx = cands[0] if len(cands) == 1 else min(cands, key=lambda cand: len(cand[2].index(1).simplify().backward_slice))
-  buf = buf.replace(dtype=(dtypes.imageh if buf.dtype.itemsize == 2 else dtypes.imagef)((h, w, 4)))
+  buf = buf.replace(src=(shape_to_shape_arg((h, w, 4)),))
   shapes[buf.arg.slot] = (h, w)
   if valid.op is not Ops.CONST or valid.arg is not True:
-    return buf.index(valid.where(cidx.src[1], cidx.src[1].const_like(Invalid)),
-                     valid.where(cidx.src[0], cidx.src[0].const_like(Invalid)))
+    return buf.index(cidx.src[1].valid(valid), cidx.src[0].valid(valid), dtype=dtypes.float)
   else:
-    return buf.index(cidx.src[1], cidx.src[0])
+    return buf.index(cidx.src[1], cidx.src[0], dtype=dtypes.float)
 
 pm_simplify_add_image = PatternMatcher([
   (UPat(Ops.SHRINK, src=(UPat(Ops.PARAM, name="buf"), UPat(name="x"), UPat(arg=4))), transform_to_image),
@@ -102,15 +101,16 @@ def memory_coalesing(sink:UOp, ctx:Renderer) -> UOp:
   if getenv("DMC"): return sink
 
   # collect
-  memory: defaultdict[tuple[Ops, UOp, Any, Any], dict[int, list[UOp]]]  = defaultdict(dict)
+  memory: defaultdict[tuple[Ops, UOp, UOp|str, UOp], dict[int, list[UOp]]] = defaultdict(dict)
   for u in sink.toposort():
     # TODO: this should handle images too, it's just memory coalesing
     if u.op in {Ops.LOAD, Ops.STORE}:
       assert len(u.src) == (2 if u.op is Ops.STORE else 1), "memory coalesing does not support gated loads/stores"
       assert u.src[0].op is Ops.INDEX, f"memory coalesing should be on INDEX, not {u.src[0].op}"
       buf, idx_u = u.src[0].src
-      idx: Any = idx_u.src[1] if idx_u.op is Ops.WHERE and idx_u.src[2].arg is Invalid else idx_u
-      valid: Any = idx_u.src[0] if idx_u.op is Ops.WHERE and idx_u.src[2].arg is Invalid else None
+      if buf.addrspace == AddrSpace.REG: continue
+      idx, valid = idx_u.get_idx(), idx_u.get_valid()
+      root_src: UOp|str
       if idx.op is Ops.ADD and idx.src[1].op is Ops.CONST: root_src, arg = idx.src[0], idx.src[1].arg
       elif idx.op is Ops.ADD and idx.src[0].op is Ops.CONST: root_src, arg = idx.src[1], idx.src[0].arg
       elif idx.op is Ops.CONST and idx.arg is Invalid: root_src, arg = "INVALID", 0
@@ -127,9 +127,11 @@ def memory_coalesing(sink:UOp, ctx:Renderer) -> UOp:
     if ctx is not None and ctx.target.device == "DSP":
       lengths = [128,64,32,16,8,4]
       must_divide = False
-    elif buf.dtype not in (dtypes.float, dtypes.half, *dtypes.fp8s) and not isinstance(buf.dtype, ImageDType):
+    elif buf.dtype not in (dtypes.float, dtypes.half, *dtypes.fp8s) and not is_image_shape(buf._shape):
       pass
-    elif isinstance(buf.dtype, ImageDType):
+    elif buf.addrspace == AddrSpace.REG:
+      pass
+    elif is_image_shape(buf._shape):
       lengths = [4]
     elif ctx is not None and ctx.supports_float4:
       # TODO: a better way to get this than ctx
@@ -139,24 +141,24 @@ def memory_coalesing(sink:UOp, ctx:Renderer) -> UOp:
     grouped_offsets = [[x for _,x in group] for _,group in itertools.groupby(enumerate(sorted(offsets.keys())), lambda x: x[1]-x[0])]
     for full_grp in grouped_offsets:
       while len(full_grp):
-        offset = (base+full_grp[0]) if isinstance(base, UOp) else UOp.const(dtypes.weakint, full_grp[0])
+        offset = (base+full_grp[0]) if isinstance(base, UOp) else UOp.const(dtypes.index, full_grp[0])
         length = [l for l in lengths if l <= len(full_grp) and (not must_divide or offset.divides(l) is not None)][0]
         grp = full_grp[:length]
         # NOTE: we apply the valid again after we determine the length
-        offset = valid.where(offset, UOp(Ops.CONST, offset.dtype, arg=Invalid)) if valid is not None else offset
-        idx = UOp(Ops.SHRINK, dtype=buf.dtype, src=(buf, offset, UOp.const(dtypes.weakint, len(grp)))) if len(grp) > 1 else buf.index(offset)
+        offset = offset.valid(valid) if valid is not None else offset
+        idx = UOp(Ops.SHRINK, src=(buf, offset, UOp.const(dtypes.index, len(grp)))) if len(grp) > 1 else buf.index(offset)
         if op == Ops.STORE:
           datas = []
           for i,g in enumerate(grp):
             assert len(offsets[g]) == 1, f"attempting multiple stores: {len(offsets[g])}"
             datas.append(offsets[g][0].src[1])
-          store = idx.store(UOp._stack(*datas) if len(datas) > 1 else datas[0])
+          store = idx.store(UOp.stack(*datas) if len(datas) > 1 else datas[0])
           for i,g in enumerate(grp): replacements[offsets[g][0]] = store
         else:
           ld = idx.load()
           for i,g in enumerate(grp):
             for oo in offsets[g]:
-              replacements[oo] = ld.index(UOp.const(dtypes.weakint, i)) if len(grp) > 1 else ld
+              replacements[oo] = ld.index(UOp.const(dtypes.index, i)) if len(grp) > 1 else ld
         full_grp = full_grp[length:]
 
   # apply

tinygrad/codegen/late/gater.py

@@ -6,10 +6,10 @@ pm_move_gates_from_index = PatternMatcher([
   # for image idx (must be first)
   (UPat.var("buf").index(UPat.var("gate").where(UPat.var("idx_y"), UPat(arg=Invalid)),
                          UPat.var("gate").where(UPat.var("idx_x"), UPat(arg=Invalid))).load(name="l"),
-   lambda buf,gate,idx_y,idx_x,l: buf.index(idx_y, idx_x).load(l.vconst_like(0), gate)),
+   lambda buf,gate,idx_y,idx_x,l: buf.index(idx_y, idx_x, dtype=dtypes.float).load(l.vconst_like(0), gate)),
   (UPat.var("buf").index(UPat.var("gate").where(UPat.var("idx_y"), UPat(arg=Invalid)),
                          UPat.var("gate").where(UPat.var("idx_x"), UPat(arg=Invalid))).store(UPat.var("data")),
-   lambda buf,gate,idx_y,idx_x,data: buf.index(idx_y, idx_x).store(data, gate)),
+   lambda buf,gate,idx_y,idx_x,data: buf.index(idx_y, idx_x, dtype=dtypes.float).store(data, gate)),
 
   # here we create the alt value for load to be 0s and remove the where Invalid
   (UPat((Ops.INDEX, Ops.SHRINK), src=(UPat(), UPat.var("gate").where(UPat.var("idx"), UPat(arg=Invalid)),), name="mop", allow_any_len=True) \

tinygrad/codegen/late/regalloc.py

@@ -49,9 +49,8 @@ class LinearScanRegallocContext:
     # assign register to spilled virtual and record load to be emitted before current uop, also assign it a stack slot
     def fill(v:Register, i:int, cons:tuple[Register, ...]|None=None) -> Register:
       if v not in self.spills:
-        # the value of a BUFFER is its 64bit address
-        dt = self.vdef(v).dtype
-        sz = 8 if self.vdef(v).op is Ops.BUFFER else dt.itemsize
+        # the value of a BUFFER is its 64bit address, XMM registers need 16 bytes
+        sz = 16 if v.cons[0].size == 16 else (8 if self.vdef(v).op is Ops.BUFFER else self.vdef(v).dtype.itemsize)
         offset = self.stack_size + (sz - self.stack_size % sz) % sz
         self.spills[v] = UOp.const(dtypes.int32, offset)
         self.stack_size = offset + sz
@@ -127,8 +126,8 @@ def regalloc_rewrite(ctx:LinearScanRegallocContext, x:UOp):
   if ctx.stack_size > 0:
     sp = ctx.ren.stack_pointer()
     offset = UOp(Ops.CONST, sp.dtype, arg=ctx.stack_size)
-    if i == 0: before = [ctx.ren.isel_matcher.rewrite(UOp(Ops.SUB, sp.dtype, (sp, offset), tag=sp.tag))] + before
-    elif i == len(ctx.uops) - 2: before += [ctx.ren.isel_matcher.rewrite(UOp(Ops.ADD, sp.dtype, (sp, offset), tag=sp.tag))]
+    if i == 0: before = [ctx.ren.isel_matcher.rewrite(UOp(Ops.SUB, src=(sp, offset), tag=sp.tag))] + before
+    elif i == len(ctx.uops) - 2: before += [ctx.ren.isel_matcher.rewrite(UOp(Ops.ADD, src=(sp, offset), tag=sp.tag))]
 
   return nx, before + [nx] + after
 

tinygrad/codegen/opt/heuristic.py

@@ -50,7 +50,7 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
   # upcast float4 images, this must be early so we don't accidentally add locals before the upcast
   if IMAGE:
     for buf_index,buf in enumerate(k.bufs):
-      if image_valid_dims(buf.src[0].dtype.base, buf.src[0].max_numel(), k.ren.target.arch):
+      if image_valid_dims(buf.src[0].dtype, buf.src[0].max_numel(), k.ren.target.arch):
         # part of is_expanded
         unit_stride_axes_mul_4 = [k.rngs.index(c) for c in k.bufs[buf_index].src[1].get_idx().split_uop(Ops.ADD) if
           c.op is Ops.RANGE and (c.vmax+1)%4 == 0]

tinygrad/codegen/opt/postrange.py

@@ -196,7 +196,7 @@ class Scheduler:
       store_targets = {s.src[0] for s in self.ast.backward_slice_with_self if s.op is Ops.STORE}
       for b in self.bufs:
         if rng in (i:=b.src[1].get_idx()).backward_slice_with_self:
-          nb = b.replace(src=(b.src[0],(valid&b.src[1].get_valid()).where(i, UOp.invalid())))
+          nb = b.replace(src=(b.src[0], i.valid(valid&b.src[1].get_valid())))
           replaces[b] = nb if b in store_targets else valid.where(nb, UOp.const(b.dtype, Invalid))
       self.ast = self.ast.substitute(replaces, f"padto {rng.arg[:-1]} {opt.arg}")
     elif opt.op is OptOps.SWAP:
@@ -228,7 +228,7 @@ class Scheduler:
         raise KernelOptError(f"invalid tensor core choice {tc_select}")
       for tc in tensor_cores:
         if self.ren.target.device in ("CUDA", "NV") and tc.dtype_in == dtypes.float and not ALLOW_TF32: continue
-        if tc.dtype_in == in0.dtype.scalar() and tc.dtype_in == in1.dtype.scalar() and tc.dtype_out == reduceop.dtype.scalar():
+        if tc.dtype_in == in0.dtype and tc.dtype_in == in1.dtype and tc.dtype_out == reduceop.dtype:
           # tensor cores have three ranges. X, Y, and REDUCE
           in0_ranges = sorted([u for u in in0.ranges if u not in in1.ranges], key=lambda x: x.arg[0], reverse=True)
           in1_ranges = sorted([u for u in in1.ranges if u not in in0.ranges], key=lambda x: x.arg[0], reverse=True)
@@ -302,12 +302,12 @@ class Scheduler:
             # do the reduce_axes always disappear? i think they don't
             # they need to be moved into the WMMA srcs
             wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.ren.target.device, tc.threads, tc_upcast_axes, ()) #, tc_reduce_axes)
-            tc_uop = UOp(Ops.WMMA, dtype=tc.dtype_out, src=(
+            tc_uop = UOp(Ops.WMMA, src=(
               srcs[0], srcs[1], UOp.const(tc.dtype_out, (0.0,)*tc.elements_per_thread[2])), arg=wmma_arg, tag=1)
 
             # preserve extra reduces
             reduce_ranges = [x for x in UOp.sink(*reduceop.src[1:]).toposort() if x.op is Ops.RANGE and x.arg[0] not in tc_reduce_axes]
-            if len(reduce_ranges): tc_uop = UOp(Ops.REDUCE, tc_uop.dtype, (tc_uop,)+tuple(reduce_ranges), (Ops.ADD, 0))
+            if len(reduce_ranges): tc_uop = UOp(Ops.REDUCE, src=(tc_uop,)+tuple(reduce_ranges), arg=(Ops.ADD, 0))
             self.ast = self.ast.substitute({reduceop: tc_uop})
           self.tensor_core = tc
           return axes
@@ -319,7 +319,7 @@ class Scheduler:
   @property
   def reduceop(self) -> UOp|None:
     if not (red := self.reduceops): return None
-    return UOp(Ops.REDUCE, red[0].dtype, red[0].src, red[0].arg)
+    return UOp(Ops.REDUCE, src=red[0].src, arg=red[0].arg)
   @property
   def bufs(self) -> list[UOp]: return [x for x in self.ast.toposort() if x.op is Ops.INDEX][::-1]
   @property
@@ -332,7 +332,7 @@ class Scheduler:
 
 def bufs_from_ast(ast:UOp, dname:str) -> list[Buffer]:
   glbls = sorted([x for x in ast.backward_slice if x.op is Ops.PARAM and x.arg.slot >= 0], key=lambda x: x.arg.slot)
-  return [Buffer(dname, x.max_numel(), x.dtype.base) for x in glbls]
+  return [Buffer(dname, x.max_numel(), x.dtype) for x in glbls]
 
 def apply_opts(ast:UOp, ren:Renderer, beam:int=0) -> UOp:
   if ast.tag is not None: return ast

tinygrad/codegen/simplify.py

@@ -9,8 +9,7 @@ def flatten_range(r:UOp) -> UOp|None:
   off = range_start[r.op]
   rngs = r.src[off:]
   if not len(rngs): return None
-  new_rngs = [x for x in UOp.sink(*rngs).toposort() if x.op is Ops.RANGE]
-  return r.replace(src=r.src[:off]+tuple(new_rngs))
+  return r.replace(src=r.src[:off]+tuple(UOp.sink(*rngs).ranges))
 
 pm_flatten_range = PatternMatcher([
   # real ranges only
@@ -82,9 +81,9 @@ def reduce_unparented(red:UOp) -> UOp|None:
   if len(reduce_unparented) == 0: return None
   ret = red.replace(src=(red.src[0],)+tuple(reduce_parented)) if len(reduce_parented) or red.dtype != red.src[0].dtype else red.src[0]
   if red.arg[0] is Ops.ADD:
-    for r in reduce_unparented: ret = ret * r.src[0].cast(ret.dtype.scalar()).broadcast(ret.dtype.count)
+    for r in reduce_unparented: ret = ret * r.src[0].cast(ret.dtype)
   if red.arg[0] is Ops.MUL:
-    for r in reduce_unparented: ret = ret ** r.src[0].cast(ret.dtype.scalar()).broadcast(ret.dtype.count)
+    for r in reduce_unparented: ret = ret ** r.src[0].cast(ret.dtype)
   return ret
 
 pm_reduce_unparented = PatternMatcher([
@@ -150,5 +149,5 @@ def no_load(u:UOp) -> bool: return not any(x.op is Ops.INDEX for x in u.backward
 pm_load_collapse = PatternMatcher([
   (UPat(Ops.REDUCE, arg=(Ops.ADD, 0), src=(UPat.var("u"), UPat()), name="red"), reduce_load_collapse),
   # we want to make sure we dont do math on a loaded index since that can cause overflow, this undoes the rule in pm_reduce_load_collapse
-  ((UPat.var("x", dtypes.weakint)+UPat.var("y"))<UPat.var("c"), lambda x,y,c: x < c-y if no_load(y) and no_load(c) and not no_load(x) else None),
+  ((UPat.var("x", dtypes.index)+UPat.var("y"))<UPat.var("c"), lambda x,y,c: x < c-y if no_load(y) and no_load(c) and not no_load(x) else None),
 ])

tinygrad/device.py

@@ -6,7 +6,7 @@ import importlib, inspect, functools, pathlib, os, contextlib, re, atexit, pickl
 from tinygrad.helpers import LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored
 from tinygrad.helpers import Context, CCACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, suppress_finalizing
 from tinygrad.helpers import select_by_name, select_first_inited, DEV, TracingKey, size_to_str, pluralize
-from tinygrad.dtype import DType, PtrDType, _to_np_dtype
+from tinygrad.dtype import DType, _to_np_dtype
 if TYPE_CHECKING: from tinygrad.renderer import Renderer
 
 # **************** Device ****************
@@ -102,7 +102,7 @@ class Buffer:
   profile_events:list[ProfileEvent] = []
   def __init__(self, device:str, size:int, dtype:DType, opaque:Any=None, options:BufferSpec|None=None, initial_value:bytes|None=None,
                uop_refcount=0, base:Buffer|None=None, offset:int=0, preallocate=False):
-    assert isinstance(dtype, DType) and not isinstance(dtype, PtrDType)
+    assert isinstance(dtype, DType)
     self.device, self.size, self.dtype, self.options, self.offset, self.allocated_views = device, size, dtype, options, offset, 0
     self._bufs: dict[str, Any] = {}
     if base is None:
@@ -202,8 +202,8 @@ class Buffer:
     return self.copyout(memoryview(bytearray(self.nbytes)))
   def numpy(self) -> 'np.ndarray': # type: ignore [name-defined] # noqa: F821
     import numpy as np
-    assert _to_np_dtype(self.dtype.base) is not None, f"no np dtype for {self.dtype.base}"
-    return np.frombuffer(self.as_memoryview(), dtype=_to_np_dtype(self.dtype.base))
+    assert _to_np_dtype(self.dtype) is not None, f"no np dtype for {self.dtype}"
+    return np.frombuffer(self.as_memoryview(), dtype=_to_np_dtype(self.dtype))
   def copyin(self, mv:memoryview):
     mv = flat_mv(mv)
     assert len(mv) == self.nbytes, f"size mismatch, {len(mv)=} != {self.dtype=} {self.size=}"

tinygrad/dtype.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 from typing import Final, ClassVar, Callable, Literal
 import math, struct, ctypes, functools
 from dataclasses import dataclass, fields
-from tinygrad.helpers import getenv, prod, round_up, OSX
+from tinygrad.helpers import getenv
 from enum import IntEnum, auto
 
 class ConstFloat(float):
@@ -69,19 +69,12 @@ class DType(metaclass=DTypeMetaClass):
   def __reduce__(self): return type(self), tuple(getattr(self, f.name) for f in fields(self))
   def __repr__(self): return f"dtypes.{INVERSE_DTYPES_DICT[self.scalar().name]}"+(f".vec({self.count})" if self.count != 1 else "")
   def __lt__(self, o:DType): return (self.priority, self.bitsize, self.name, self.fmt, self.count) < (o.priority, o.bitsize, o.name, o.fmt, o.count)
-  @property
-  def base(self): return self
-  @property
-  def vcount(self): return self.count
   @functools.cache  # pylint: disable=method-cache-max-size-none
   def vec(self, sz:int) -> DType:
     assert self.count == 1, f"can't vectorize {self} with size {sz}"
     if sz == 1 or self == dtypes.void: return self  # void doesn't vectorize, and sz=1 is scalar
     return DType(self.priority, self.bitsize*sz, f"{INVERSE_DTYPES_DICT[self.name]}{sz}", None, sz, self)
-  def ptr(self, size=-1, addrspace=AddrSpace.GLOBAL) -> PtrDType:
-    return PtrDType(self.priority, self.bitsize, self.name, self.fmt, self.count, None, self, addrspace, 1, size)
   def scalar(self) -> DType: return self._scalar if self._scalar is not None else self
-  def nbytes(self) -> int: raise RuntimeError("only ptr types have nbytes")
   @functools.cached_property
   def min(self):
     if dtypes.is_int(self): return 0 if dtypes.is_unsigned(self) else -2**(self.scalar().bitsize-1)
@@ -100,51 +93,14 @@ class DType(metaclass=DTypeMetaClass):
     # int is the default. wrap floats in ConstFloat to distinguish -0.0 from 0.0 in cache
     return ConstFloat(float(val)) if dtypes.is_float(self) else bool(val) if dtypes.is_bool(self) else int(val)
 
-@dataclass(frozen=True, eq=False)
-class PtrDType(DType):
-  _base: DType
-  addrspace: AddrSpace
-  v: int
-  size: int = -1  # -1 is unlimited size
-  @property
-  def base(self): return self._base
-  @functools.cache  # pylint: disable=method-cache-max-size-none
-  def vec(self, sz:int) -> DType:
-    assert self.v == 1, f"can't vectorize ptr {self} with size {sz}"
-    if sz == 1: return self  # sz=1 is a scalar
-    if isinstance(self, ImageDType):
-      return ImageDType(self.priority, self.bitsize, self.name, self.fmt, self.count, self, self._base, self.addrspace, sz, self.size, self.shape)
-    return type(self)(self.priority, self.bitsize, self.name, self.fmt, self.count, self, self._base, self.addrspace, sz, self.size)
-  def ptr(self, size=-1, addrspace=AddrSpace.GLOBAL) -> PtrDType: raise RuntimeError("can't make a pointer from a pointer")
-  def nbytes(self) -> int:
-    if self.size == -1: raise RuntimeError("can't get nbytes of a pointer with unlimited size")
-    return self.size*self.itemsize
-  @property
-  def vcount(self): return self.v
-  def __repr__(self):
-    return f"{self.base.__repr__()}.ptr({self.size}{', '+str(self.addrspace) if self.addrspace != AddrSpace.GLOBAL else ''})" + \
-      (f'.vec({self.v})' if self.v != 1 else '')
-
-@dataclass(frozen=True, eq=False)
-class ImageDType(PtrDType):
-  shape: tuple[int, ...] = ()   # shape of the Image
-  def ptr(self, size=-1, addrspace=AddrSpace.GLOBAL) -> PtrDType:
-    assert addrspace == AddrSpace.GLOBAL, "images can't be local"
-    return self
-  def __repr__(self): return f"dtypes.{self.name}({self.shape})" + (f'.vec({self.v})' if self.v != 1 else '')
-
-  # for 1d images on macos, we need to round pitch up to 256 pixels to make CL happy
-  @property
-  def pitch(self): return (round_up(self.shape[1], 256) if OSX else self.shape[1]) * 4 * self.itemsize
-
 
 class dtypes:
   @staticmethod
   @functools.cache
-  def is_float(x: DType) -> bool: return x.scalar() in dtypes.floats or isinstance(x, ImageDType)
+  def is_float(x: DType) -> bool: return x.scalar() in (dtypes.floats + (dtypes.weakfloat,))
   @staticmethod # static methods on top, or bool in the type info will refer to dtypes.bool
   @functools.cache
-  def is_int(x: DType) -> bool: return x.scalar() in (dtypes.ints + (dtypes.weakint,))
+  def is_int(x: DType) -> bool: return x.scalar() in (dtypes.ints + (dtypes.weakint, dtypes.index))
   @staticmethod
   @functools.cache
   def is_unsigned(x: DType) -> bool: return x.scalar() in dtypes.uints
@@ -167,6 +123,7 @@ class dtypes:
             dtypes.fp8e4m3: (4, 3), dtypes.fp8e5m2: (5, 2), dtypes.fp8e4m3fnuz: (4, 3), dtypes.fp8e5m2fnuz: (5, 2)}[dtype]
   void: Final[DType] = DType.new(-1, 0, "void", None)
   weakint: Final[DType] = DType.new(0, 800, "weakint", None)
+  index: Final[DType] = DType.new(0, 800, "index", None)  # NOTE: not in the promo lattice: index math never mixes dtypes
   bool: Final[DType] = DType.new(0, 1, "bool", '?')
   int8: Final[DType] = DType.new(1, 8, "signed char", 'b')
   uint8: Final[DType] = DType.new(2, 8, "unsigned char", 'B')
@@ -178,27 +135,22 @@ class dtypes:
   uint64: Final[DType] = DType.new(8, 64, "unsigned long", 'Q')
   _uint128: Final[DType] = DType.new(8, 128, "uint128", None)
   _uint256: Final[DType] = DType.new(8, 256, "uint256", None)
-  fp8e4m3: Final[DType] = DType.new(9, 8, "float8_e4m3", None)
-  fp8e5m2: Final[DType] = DType.new(10, 8, "float8_e5m2", None)
-  fp8e4m3fnuz: Final[DType] = DType.new(9, 8, "float8_e4m3fnuz", None)
-  fp8e5m2fnuz: Final[DType] = DType.new(10, 8, "float8_e5m2fnuz", None)
-  float16: Final[DType] = DType.new(11, 16, "half", 'e')
+  weakfloat: Final[DType] = DType.new(9, 800, "weakfloat", None)
+  fp8e4m3: Final[DType] = DType.new(10, 8, "float8_e4m3", None)
+  fp8e5m2: Final[DType] = DType.new(11, 8, "float8_e5m2", None)
+  fp8e4m3fnuz: Final[DType] = DType.new(10, 8, "float8_e4m3fnuz", None)
+  fp8e5m2fnuz: Final[DType] = DType.new(11, 8, "float8_e5m2fnuz", None)
+  float16: Final[DType] = DType.new(12, 16, "half", 'e')
   # bfloat16 has higher priority than float16, so least_upper_dtype(dtypes.int64, dtypes.uint64) = dtypes.float16
-  bfloat16: Final[DType] = DType.new(12, 16, "__bf16", None)
-  float32: Final[DType] = DType.new(13, 32, "float", 'f')
-  float64: Final[DType] = DType.new(14, 64, "double", 'd')
+  bfloat16: Final[DType] = DType.new(13, 16, "__bf16", None)
+  float32: Final[DType] = DType.new(14, 32, "float", 'f')
+  float64: Final[DType] = DType.new(15, 64, "double", 'd')
 
   # dtype aliases
   half = float16; float = float32; double = float64 # noqa: E702
   uchar = uint8; ushort = uint16; uint = uint32; ulong = uint64 # noqa: E702
   char = int8; short = int16; int = int32; long = int64 # noqa: E702
 
-  # NOTE: these are image dtypes
-  @staticmethod
-  def imageh(shp): return ImageDType(100, 16, "imageh", 'e', 1, None, dtypes.float32, AddrSpace.GLOBAL, 1, prod(shp), shp)
-  @staticmethod
-  def imagef(shp): return ImageDType(100, 32, "imagef", 'f', 1, None, dtypes.float32, AddrSpace.GLOBAL, 1, prod(shp), shp)
-
   default_float: ClassVar[DType] = float32
   default_int: ClassVar[DType] = int32
 
@@ -213,7 +165,7 @@ class dtypes:
   uints = (uint8, uint16, uint32, uint64)
   sints = (int8, int16, int32, int64)
   ints = uints + sints
-  all = floats + ints + (bool, weakint) # noqa: A003
+  all = floats + ints + (bool,) # noqa: A003
 
 if (env_default_float := getenv("DEFAULT_FLOAT", "")):
   dtypes.default_float = getattr(dtypes, env_default_float.lower())
@@ -227,7 +179,8 @@ def to_dtype(dtype:DTypeLike) -> DType: return dtype if isinstance(dtype, DType)
 promo_lattice = { dtypes.bool: [dtypes.weakint], dtypes.weakint: [dtypes.int8, dtypes.uint8],
   dtypes.int8: [dtypes.int16], dtypes.int16: [dtypes.int32], dtypes.int32: [dtypes.int64],
   dtypes.int64: [dtypes.uint64], dtypes.uint8: [dtypes.int16, dtypes.uint16], dtypes.uint16: [dtypes.int32, dtypes.uint32],
-  dtypes.uint32: [dtypes.int64, dtypes.uint64], dtypes.uint64: [dtypes.fp8e4m3, dtypes.fp8e5m2, dtypes.fp8e4m3fnuz, dtypes.fp8e5m2fnuz],
+  dtypes.uint32: [dtypes.int64, dtypes.uint64], dtypes.uint64: [dtypes.weakfloat],
+  dtypes.weakfloat: [dtypes.fp8e4m3, dtypes.fp8e5m2, dtypes.fp8e4m3fnuz, dtypes.fp8e5m2fnuz],
   dtypes.fp8e4m3: [dtypes.float16, dtypes.bfloat16], dtypes.fp8e5m2: [dtypes.float16, dtypes.bfloat16],
   dtypes.fp8e4m3fnuz: [dtypes.float16, dtypes.bfloat16], dtypes.fp8e5m2fnuz: [dtypes.float16, dtypes.bfloat16],
   dtypes.float16: [dtypes.float32], dtypes.bfloat16: [dtypes.float32], dtypes.float32: [dtypes.float64], }
@@ -237,12 +190,11 @@ def _get_recursive_parents(dtype:DType) -> set[DType]:
   return set.union(*[_get_recursive_parents(d) for d in promo_lattice[dtype]], {dtype}) if dtype != dtypes.float64 else {dtypes.float64}
 @functools.cache
 def least_upper_dtype(*ds:DType) -> DType:
-  return min(set.intersection(*[_get_recursive_parents(d.scalar()) for d in ds])) \
-      if not (images:=[d for d in ds if isinstance(d, ImageDType)]) else images[0]
+  return min(set.intersection(*[_get_recursive_parents(d.scalar()) for d in ds]))
 def least_upper_float(dt:DType) -> DType: return dt if dtypes.is_float(dt) else least_upper_dtype(dt, dtypes.default_float)
 
-DTYPES_DICT = {k: v for k, v in dtypes.__dict__.items() if isinstance(v, DType) and not k.startswith(("default", "void", "weakint", "_"))}
-INVERSE_DTYPES_DICT = {**{v.name:k for k,v in DTYPES_DICT.items()}, "void": "void", "weakint":"weakint"}
+DTYPES_DICT = {k: v for k, v in dtypes.__dict__.items() if isinstance(v, DType) and not k.startswith(("default", "void", "weak", "index", "_"))}
+INVERSE_DTYPES_DICT = {**{v.name:k for k,v in DTYPES_DICT.items()}, "void": "void", "weakint":"weakint", "index":"index", "weakfloat":"weakfloat"}
 
 @functools.cache
 def can_lossless_cast(dt0:DType, dt1:DType) -> bool:
@@ -250,7 +202,7 @@ def can_lossless_cast(dt0:DType, dt1:DType) -> bool:
   # similar to https://numpy.org/doc/stable/reference/generated/numpy.can_cast.html
   if dt0 == dt1 or dt0 == dtypes.bool: return True
   match dt1:
-    case dtypes.weakint: return dt0 in dtypes.ints
+    case dtypes.weakint | dtypes.index: return dt0 in dtypes.ints
     case dtypes.double: return dt0 in (dtypes.float, dtypes.half, dtypes.bfloat16, *dtypes.fp8s,
       dtypes.uint32, dtypes.uint16, dtypes.uint8, dtypes.int32, dtypes.int16, dtypes.int8)
     case dtypes.float: return dt0 in (dtypes.half, dtypes.bfloat16, *dtypes.fp8s, dtypes.uint16, dtypes.uint8, dtypes.int16, dtypes.int8)

tinygrad/engine/jit.py

@@ -1,16 +1,16 @@
 from typing import TypeVar, Generic, Callable, Any
 import functools, collections
 from tinygrad.tensor import Tensor, all_tensors
-from tinygrad.helpers import flatten, merge_dicts, DEBUG, Context, BEAM, getenv, JIT, JIT_BATCH_SIZE, dedup, pluralize, VIZ
+from tinygrad.helpers import flatten, merge_dicts, DEBUG, Context, BEAM, getenv, JIT, JIT_BATCH_SIZE, dedup, pluralize, VIZ, disable_gc
 from tinygrad.device import Buffer, Compiled, Device, MultiBuffer
-from tinygrad.dtype import DType, dtypes
+from tinygrad.dtype import DType
 from tinygrad.uop.ops import UOp, PatternMatcher, Variable, sym_infer, Ops, buffers, track_rewrites, graph_rewrite
 from tinygrad.renderer import Estimates
 from tinygrad.engine.realize import capturing, compile_linear, link_linear, run_linear, graph_cache, estimate_uop, get_runtime
 from tinygrad.engine.realize import unwrap_multi, resolve_params, get_call_arg_uops, get_call_outs_ins
 from tinygrad.schedule.memory import memory_plan_rewrite, _collect_bufs
 from tinygrad.nn.state import get_parameters
-from tinygrad.schedule.rangeify import mop_cleanup
+from tinygrad.uop.movement import mop_cleanup
 from dataclasses import dataclass
 
 def prune_linear(linear:UOp, needed:set[UOp]) -> tuple[UOp, UOp]:
@@ -26,8 +26,8 @@ def prune_linear(linear:UOp, needed:set[UOp]) -> tuple[UOp, UOp]:
 def create_graph_call(batch:list[UOp]) -> UOp:
   # all external inputs are PARAMs
   input_list = dedup(u for si in batch for b in si.src[1:] for u in b.toposort() if u.op is Ops.PARAM)
-  cf = UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(UOp(Ops.LINEAR, src=tuple(batch)),), arg="graph")
-  return cf.call(*input_list, metadata=tuple(m for si in batch for m in si.arg.metadata))
+  cf = UOp(Ops.CUSTOM_FUNCTION, src=(UOp(Ops.LINEAR, src=tuple(batch)),), arg="graph")
+  return cf.call(*input_list)
 
 def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
   new_src: list[UOp] = []
@@ -61,7 +61,7 @@ def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
   return linear.replace(src=tuple(new_src))
 
 def _copy_input(u:UOp) -> UOp:
-  run_linear(UOp(Ops.LINEAR, src=(u.copy_to_device(u.device).call(new:=UOp.new_buffer(u.device, u.max_numel(), u.dtype), u, metadata=()),)))
+  run_linear(UOp(Ops.LINEAR, src=(u.copy_to_device(u.device).call(new:=UOp.new_buffer(u.device, u.max_numel(), u.dtype), u),)))
   return new
 
 @track_rewrites(lambda linear,held_bufs,input_uops,ret=(): f"JIT {pluralize('call', len(linear.src))}")
@@ -71,7 +71,7 @@ def jit_lower(linear:UOp, held_bufs:set[UOp], input_uops:list[UOp]) -> UOp:
   # parametrize input buffers: map each input buffer UOp to a PARAM with the correct slot index
   linear = linear.substitute({u: UOp.param(i, u.dtype, u.shape, u.device) for i,u in enumerate(input_uops)}, walk=True)
   linear = memory_plan_rewrite(linear, held_bufs)
-  linear = compile_linear(linear, beam=getenv("JITBEAM", BEAM.value))
+  linear = compile_linear(linear, beam=getenv("JITBEAM", BEAM.value), jit=True)
   if JIT < 2: linear = graph_split_rewrite(linear, max_batch_size=JIT_BATCH_SIZE.value)
   if VIZ: graph_rewrite(linear, PatternMatcher([]), name="View graphed linear")
   return linear
@@ -197,7 +197,7 @@ class CapturedJit(Generic[ReturnType]):
   expected_input_info: list[tuple[UOp, tuple[Variable, ...], DType, str]]  # (view, variables, dtype, device) per input
 
   @functools.cached_property
-  def linear(self) -> UOp: return link_linear(self._linear)
+  def linear(self) -> UOp: return link_linear(self._linear, jit=True)
 
   def __reduce__(self): return self.__class__, (self.ret, self._linear, self.expected_names, self.expected_input_info)
 
@@ -241,7 +241,7 @@ def _prepare_jit_inputs(args, kwargs):
   # collect buffer UOps (including MultiBuffer)
   input_buf_uops: list[UOp] = [u.base for u in input_uops if u.base.realized is not None]
   if len(set(input_buf_uops)) != len(input_buf_uops): raise JitError("duplicate inputs to JIT")
-  inputs = [(*(u.substitute({u.base:UOp(Ops.NOOP)}, extra_pm=mop_cleanup).unbind_all()), u.dtype, u.device) for u in input_uops]
+  inputs = [(*(u.substitute({u.base:UOp(Ops.NOOP, u.base.dtype)}, extra_pm=mop_cleanup).unbind_all()), u.dtype, u.device) for u in input_uops]
   _var_vals = merge_dicts([x[1] for x in inputs] + [dict(v.unbind() for v in (args + tuple(kwargs.values())) if isinstance(v, UOp))])
   var_vals = {k.expr:v for k,v in _var_vals.items()}
   expected_input_info = [(x[0], tuple(sorted(x[1].keys(), key=lambda v: v.expr)), x[2], x[3]) for x in inputs]
@@ -268,6 +268,7 @@ class TinyJit(Generic[ReturnType]):
 
   def __get__(self, obj, objtype): return functools.partial(self.__call__, obj) # add support for instance methods
 
+  @disable_gc()
   def __call__(self, *args, **kwargs) -> ReturnType:
     input_buf_uops, var_vals, names, expected_input_info = _prepare_jit_inputs(args, kwargs)
     if not JIT or self.cnt == 0:

tinygrad/engine/realize.py

@@ -2,9 +2,8 @@ from __future__ import annotations
 from typing import cast, Iterator, Any, Sequence
 import time, random, itertools, math, contextlib, weakref, array
 from dataclasses import dataclass, replace, field
-from tinygrad.helpers import colored, DEBUG, GlobalCounters, ansilen, all_int, TRACEMETA, prod, flatten, Context, getenv, to_tuple
+from tinygrad.helpers import colored, DEBUG, GlobalCounters, ansilen, all_int, prod, flatten, Context, getenv, to_tuple
 from tinygrad.helpers import BEAM, size_to_str, time_to_str, VALIDATE_WITH_CPU, PROFILE, ProfilePointEvent, cpu_events
-from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, sym_infer, buffers, graph_rewrite, ProgramInfo
 from tinygrad.device import Device, Buffer, MultiBuffer
 from tinygrad.renderer import Estimates
@@ -54,7 +53,7 @@ first_run_cache:set[bytes] = set()
 def track_stats(ctx:ExecContext, call:UOp, device:str, bufs:list[Buffer], var_vals:dict[str, int]):
   if PROFILE:
     outputs, inputs = get_call_outs_ins(call)
-    cpu_events.append(ProfilePointEvent(device, "exec", len(cpu_events), {"metadata": call.arg.metadata, "var_vals": var_vals,
+    cpu_events.append(ProfilePointEvent(device, "exec", len(cpu_events), {"var_vals": var_vals,
       "bufs": [b.trace_num for b in bufs], "name": get_call_name(call, bufs, var_vals), "outputs": outputs, "inputs": inputs}))
   et: list[float|None] = [None]
   if DEBUG >= 2: st = time.perf_counter()
@@ -81,8 +80,7 @@ def track_stats(ctx:ExecContext, call:UOp, device:str, bufs:list[Buffer], var_va
       colored(f"{membw*1e-12:4.0f}|{ldsbw*1e-12:<6.0f} TB/s", 'green')
     print(f"{colored(f'*** {device[:7]:7s} {GlobalCounters.kernel_count:4d}', header_color)}"+
       f" {display_name+' '*(46-ansilen(display_name))} arg {len(bufs):2d} mem {GlobalCounters.mem_used/1e9:6.2f} GB"+
-      ("" if et[0] is None else f" tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({flops_str} {mem_str})")+
-      f" {[repr(m) if TRACEMETA >= 2 else str(m) for m in call.arg.metadata] if call.arg.metadata else ''}")
+      ("" if et[0] is None else f" tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({flops_str} {mem_str})"))
     first_run_cache.add(call.src[0].key)
 
 local_size_cache: dict[bytes, tuple[int, ...]] = {}
@@ -216,7 +214,7 @@ def exec_hcq(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
       buf = b.bufs[j] if isinstance(b:=call.src[1+call.arg.aux.inputs].buffer, MultiBuffer) else b
       buf.ensure_allocated()._buf.cpu_view().view(fmt='Q')[:len(addrs)] = array.array('Q', addrs)
 
-  pm_exec.rewrite(call.replace(src=(ast,) + call.src[1:]), replace(ctx, update_stats=False, wait=True))
+  pm_exec.rewrite(call.replace(src=(ast,) + call.src[1:]), replace(ctx, update_stats=False))
 
   for d in call.arg.aux.device:
     with track_stats(ctx, call, d, [], ctx.var_vals):
@@ -231,9 +229,9 @@ pm_flatten_linear = PatternMatcher([
 
 def _validate(call:UOp, sink:UOp) -> UOp:
   params = get_call_arg_uops(call)
-  shadows = tuple(UOp.new_buffer(("CPU",)*len(p.device) if isinstance(p.device, tuple) else "CPU", prod(p.max_shape), p.dtype.base) for p in params)
+  shadows = tuple(UOp.new_buffer(("CPU",)*len(p.device) if isinstance(p.device, tuple) else "CPU", prod(p.max_shape), p.dtype) for p in params)
   copies = tuple(p.copy_to_device(s.device).call(s, p) for s, p in zip(shadows, params))
-  return UOp(Ops.LINEAR, src=copies + (call, UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(sink,), arg="validate").call(*shadows, *params)))
+  return UOp(Ops.LINEAR, src=copies + (call, UOp(Ops.CUSTOM_FUNCTION, src=(sink,), arg="validate").call(*shadows, *params)))
 pm_validate = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.SINK, name="sink"),), name="call", allow_any_len=True), _validate)]) + pm_flatten_linear
 
 # ctx is beam value
@@ -261,24 +259,24 @@ pm_exec = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="validate", name="ast"),), name="call", allow_any_len=True), exec_validate),
 ])
 
-def compile_linear(linear:UOp, beam:int|None=None, validate=False, input_uops:list[UOp]|None=None) -> UOp:
+def compile_linear(linear:UOp, beam:int|None=None, validate=False, input_uops:list[UOp]|None=None, jit=False) -> UOp:
   if validate: linear = graph_rewrite(linear, pm_validate, name="validate", walk=True)
   if (beam_val:=BEAM.value if beam is None else beam) >= 1: linear = graph_rewrite(linear, pm_beam, ctx=beam_val, walk=True)
   linear = graph_rewrite(linear, pm_compile, name="precompile kernels", walk=True)
   if getenv("HCQ2"):
     from extra.hcq2.hcq2 import hcq_compile
-    linear = hcq_compile(linear, input_uops)
+    linear = hcq_compile(linear, input_uops, jit=jit)
   return graph_rewrite(linear, pm_optimize_local_size, name="optimize local size", walk=True)
 
-def link_linear(linear:UOp) -> UOp:
+def link_linear(linear:UOp, jit=False) -> UOp:
   if getenv("HCQ2"):
     from extra.hcq2.hcq2 import hcq_link
-    linear = hcq_link(linear)
+    linear = hcq_link(linear, jit=jit)
   return linear
 
 def run_linear(linear:UOp, var_vals:dict[str, int]|None=None, input_uops:Sequence[UOp]=(), update_stats=True, jit=False, wait=False):
   inputs = list(input_uops)
-  if not jit: linear = link_linear(compile_linear(linear, validate=VALIDATE_WITH_CPU, input_uops=inputs))
+  if not jit: linear = link_linear(compile_linear(linear, validate=VALIDATE_WITH_CPU, input_uops=inputs, jit=False))
   ctx = ExecContext(var_vals or {}, tuple(inputs), update_stats, jit, wait or DEBUG>=2)
   for call in linear.src: pm_exec.rewrite(call, ctx)
 

tinygrad/helpers.py

@@ -34,6 +34,7 @@ def get_shape(x) -> tuple[int, ...]:
   if not hasattr(x, "__len__") or isinstance(x, str) or getattr(x, "shape", None) == (): return ()
   if not all_same(subs:=[get_shape(xi) for xi in x]): raise ValueError(f"inhomogeneous shape from {x}")
   return (len(subs),) + (subs[0] if subs else ())
+def is_image_shape(shape): return shape is not None and len(shape) == 3 and shape[-1] == 4
 def all_int(t: Sequence[Any]) -> TypeGuard[tuple[int, ...]]: return all(isinstance(s, int) for s in t)
 def colored(st, color:str|None, background=False): # replace the termcolor library
   if NO_COLOR: return st
@@ -392,6 +393,7 @@ def db_connection():
     # another connection has set it already or is in the process of setting it
     # that connection will lock the database
     with contextlib.suppress(sqlite3.OperationalError): _db_connection.execute("PRAGMA journal_mode=WAL").fetchone()
+    _db_connection.execute("PRAGMA synchronous=NORMAL")
     if DEBUG >= 8: _db_connection.set_trace_callback(print)
   return _db_connection
 

tinygrad/mixin/dtype.py

@@ -31,7 +31,20 @@ class DTypeMixin:
     """
     return self if self.dtype == (dt:=to_dtype(dtype)) else self._wrap_uop(self._uop.cast(dt))
 
-  def bitcast(self, dtype:DTypeLike) -> Self: raise NotImplementedError
+  def bitcast(self, dtype:DTypeLike) -> Self:
+    """
+    Bitcasts `self` to the given `dtype`. If the itemsize differs, the last axis is rescaled.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor([-1, 2, 3], dtype=dtypes.int32)
+    print(t.dtype, t.numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = t.bitcast(dtypes.uint32)
+    print(t.dtype, t.numpy())
+    ```
+    """
+    return self if self.dtype == (dt:=to_dtype(dtype)) else self._wrap_uop(self._uop.bitcast(dt))
 
   def element_size(self) -> int:
     """
@@ -54,7 +67,7 @@ class DTypeMixin:
     print(t.is_floating_point())
     ```
     """
-    return dtypes.is_float(self.dtype.base)
+    return dtypes.is_float(self.dtype)
 
   def float(self) -> Self:
     """

tinygrad/mixin/elementwise.py

@@ -65,7 +65,7 @@ class ElementwiseMixin(CreationMixin):
     print(Tensor([-3., -2., -1., 0., 1., 2., 3.]).neg().numpy())
     ```
     """
-    return self.logical_not() if self.dtype.scalar() == dtypes.bool else self * (-1)
+    return self.logical_not() if self.dtype == dtypes.bool else self * (-1)
 
   def _check_dtype(self) -> None:
     if not (dtypes.is_bool(self.dtype) or dtypes.is_int(self.dtype)):

tinygrad/mixin/gradient.py

@@ -5,14 +5,7 @@ from tinygrad.helpers import argsort
 from tinygrad.dtype import sum_acc_dtype
 
 def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):
-  def broadcast_to_input(x):
-    shape, j = [], 0
-    for i in range(len(ret.src[0].shape)):
-      if i < ret.arg[1]: shape.append(1)
-      else:
-        shape.append(x.shape[j])
-        j += 1
-    return x.reshape(tuple(shape)).expand(ret.src[0].shape)
+  def broadcast_to_input(x:UOp) -> UOp: return x._broadcast_to(ret.src[0].shape)
   if op == Ops.ADD: return (broadcast_to_input(ctx),)
   if op == Ops.MAX:
     assert ret.op is Ops.REDUCE, "only works on REDUCE"
@@ -81,6 +74,7 @@ pm_gradient = PatternMatcher([
   (UPat(Ops.SHRINK, name="ret"), lambda ctx, ret: (ctx.pad(tuple([(p[0], s-p[0]-p[1]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
   (UPat(Ops.PERMUTE, name="ret"), lambda ctx, ret: (ctx.permute(argsort(ret.marg)),)),
   (UPat(Ops.FLIP, name="ret"), lambda ctx, ret: (ctx.flip([i for i,x in enumerate(ret.marg) if x]),)),
+  (UPat(Ops.STACK, name="ret"), lambda ctx, ret: tuple(ctx[i] for i in range(len(ret.src)))),
   (UPat(Ops.COPY, name="ret"), lambda ctx, ret: (ctx.copy_to_device(ret.src[0].device),)),
   (UPat(Ops.MULTI, name="ret"), lambda ctx, ret: ctx.shard(ret.device, ret.axis).src),
   (UPat(Ops.TUPLE), lambda ctx: ctx.src),

tinygrad/mixin/movement.py

@@ -241,6 +241,24 @@ class MovementMixin:
     flip_arg = tuple([i in axis_arg for i in range(len(self.shape))])
     return self._mop(Ops.FLIP, arg=flip_arg) if any(flip_arg) else self
 
+  def stack(self, *args: Self, dim: int = 0) -> Self:
+    """
+    Concatenates self with other tensors in `args` along a new dimension specified by `dim`.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t0, t1, t2 = Tensor([1, 2]), Tensor([3, 4]), Tensor([5, 6])
+    print(t0.stack(t1, t2, dim=0).numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t0.stack(t1, t2, dim=1).numpy())
+    ```
+    """
+    tensors = argfix(self, *args)
+    dim = tensors[0]._resolve_dim(dim, extra=True)
+    assert all(t.shape == tensors[0].shape for t in tensors), f"all shapes must match for stack, got {[t.shape for t in tensors]}"
+    ret = tensors[0]._mop(Ops.STACK, arg=tuple(t._uop for t in tensors[1:]))
+    return ret if dim == 0 else ret.permute(tuple(range(1, dim+1)) + (0,) + tuple(range(dim+1, ret.ndim)))
+
   # **** high level ****
 
   def shrink_to(self, shape, *args) -> Self:

tinygrad/mixin/rand.py

@@ -3,7 +3,7 @@ import math
 from typing import Self, cast
 from tinygrad.dtype import DType, DTypeLike, dtypes, least_upper_dtype, to_dtype
 from tinygrad.helpers import all_int, argfix, ceildiv, prod, TRAINING
-from tinygrad.mixin import OpMixin
+from tinygrad.mixin.op import OpMixin
 from tinygrad.device import canonicalize_device
 
 

tinygrad/nn/__init__.py

@@ -348,18 +348,18 @@ def _embedding_bwd(grad_emb:UOp, call:UOp) -> tuple:
       # each device owns [offset, offset+local_vocab_size) of the global vocabulary
       dnum = UOp.variable("_device_num", 0, ndev-1)
       offset = dnum * local_vocab_size
-      global_token_id = idx_flat[i].cast(dtypes.weakint)
+      global_token_id = idx_flat[i].cast(dtypes.index)
       local_token_id = (global_token_id - offset).clip(0, grad_weight.shape[0]-1)
       in_range = (global_token_id >= offset) & (global_token_id < (offset + local_vocab_size)) & j_ok
       grad_val = in_range.where(grad_emb_flat[i, j_idx].load().cast(dtypes.float), 0.0)
     else:
-      local_token_id = idx_flat[i].clip(0, grad_weight.shape[0]-1).cast(dtypes.weakint)
+      local_token_id = idx_flat[i].clip(0, grad_weight.shape[0]-1).cast(dtypes.index)
       grad_val = j_ok.where(grad_emb_flat[i, j_idx].load().cast(dtypes.float), 0.0)
     # atomic scatter-add: grad_weight[token_id, j] += grad_emb_flat[i, j]
     if device in ("CPU", "NULL"): atomic_arg = "__atomic_fetch_add({0}, {1}, __ATOMIC_RELAXED);"
     elif device == "AMD": atomic_arg = "__hip_atomic_fetch_add({0}, {1}, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);"
     else: raise NotImplementedError(f"no atomics for device {device}")
-    atomic = UOp(Ops.CUSTOM, dtypes.void, (grad_weight.index(local_token_id, j_idx), grad_val), arg = atomic_arg)
+    atomic = UOp(Ops.CUSTOM, src=(grad_weight.index(local_token_id, j_idx), grad_val), arg = atomic_arg)
     return atomic.end(i, j_outer, j_inner).sink(arg=KernelInfo(name="embedding_bwd", opts_to_apply=()))
 
   grad_weight_uop = grad_weight_uop.custom_kernel(grad_emb, idx, fxn=_embedding_bwd_kernel)[0]

tinygrad/nn/datasets.py

@@ -8,7 +8,7 @@ def mnist(device=None, fashion=False):
          _mnist("t10k-images-idx3-ubyte.gz")[0x10:].reshape(-1,1,28,28).to(device), _mnist("t10k-labels-idx1-ubyte.gz")[8:].to(device)
 
 def cifar(device=None):
-  tt = tar_extract(Tensor.from_url('https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz', gunzip=True))
+  tt = tar_extract(Tensor.from_url('https://data.brainchip.com/dataset-mirror/cifar10/cifar-10-binary.tar.gz', gunzip=True))
   train = Tensor.cat(*[tt[f"cifar-10-batches-bin/data_batch_{i}.bin"].reshape(-1, 3073).to(device) for i in range(1,6)])
   test = tt["cifar-10-batches-bin/test_batch.bin"].reshape(-1, 3073).to(device)
   return train[:, 1:].reshape(-1,3,32,32), train[:, 0], test[:, 1:].reshape(-1,3,32,32), test[:, 0]

tinygrad/nn/onnx.py

@@ -3,7 +3,7 @@ from typing import Any, Sequence, cast, Literal, NamedTuple, Generator
 import dataclasses, functools, io, math, types, warnings, pathlib, sys, os, struct, enum
 from tinygrad.nn.state import TensorIO
 from tinygrad.tensor import Tensor
-from tinygrad.mixin import ReductionStr
+from tinygrad.mixin.op import ReductionStr
 from tinygrad.helpers import getenv, all_same, prod, flatten, make_tuple, argsort, is_numpy_ndarray, get_single_element, polyN, Context
 from tinygrad.dtype import DType, ConstType, dtypes, _from_np_dtype, truncate, least_upper_dtype, DTYPES_DICT
 from tinygrad.device import Device

tinygrad/nn/state.py

@@ -102,7 +102,8 @@ def fs_store(t:Tensor) -> Tensor:
 
   level_chunks = base_chunks
   for _ in range(tree_depth + 1):
-    data = data.to("tinyfs:store")[:level_chunks * 16].contiguous().to(to_device)
+    # assign data into tinyfs:store and read back hashes
+    data = Tensor.empty(data.shape[0], dtype=dtypes.uint8, device="tinyfs:store").assign(data)[:level_chunks * 16].to(to_device)
     if (tsize := data.shape[0]) % CHUNK_SIZE != 0: data = data.pad((0, CHUNK_SIZE - tsize % CHUNK_SIZE))
     level_chunks = math.ceil(data.shape[0] / CHUNK_SIZE)
 
@@ -123,18 +124,16 @@ def fs_load(t:Tensor, size:int) -> Tensor:
   tree_depth = math.ceil(math.log(base_chunks, CHUNK_SIZE // 16))
   data, level_chunks = h, 0
   for i in reversed(range(tree_depth + 1)):
-    data = data.to("tinyfs:load")
-
     # if not last level, its still hashes
     if i > 0 or tree_depth == 0:
       level_chunks = max(1, math.ceil(base_chunks / (CHUNK_SIZE // 16)**(i-1)))
-      pad_amt = 16 * level_chunks
-    else: pad_amt = CHUNK_SIZE * level_chunks
-    if (tsize := data.shape[0]) < pad_amt: data = data.pad((0, pad_amt - tsize))
-    data = data[:pad_amt].contiguous()
-    if i != 0: data = data.to(t.device)
+      out_sz = 16 * level_chunks
+    else: out_sz = CHUNK_SIZE * level_chunks
+    # assign hash into tinyfs:load and read back data
+    (load:=Tensor.empty(out_sz, dtype=dtypes.uint8, device="tinyfs:load"))[:data.shape[0]].assign(data)
+    data = load
 
-  return data[:size]
+  return data.to(t.device)[:size]
 
 # state dict
 

tinygrad/renderer/__init__.py

@@ -35,7 +35,7 @@ class Estimates:
         while len(buf.src) and buf.op is not Ops.PARAM: buf = buf.src[0]
         if buf.op is Ops.PARAM:
           # u.src[0] is INDEX, cap at buffer size for re-reads (e.g. matmul)
-          accessed = mem.get((buf, u.op), 0) + u.src[0].max_numel() * u.src[0].dtype.base.scalar().itemsize * mults
+          accessed = mem.get((buf, u.op), 0) + u.src[0].max_numel() * u.src[0].dtype.scalar().itemsize * mults
           mem[(buf, u.op)] = smin(accessed, buf.max_numel() * buf.dtype.scalar().itemsize)
       if u.op is Ops.RANGE:
         mult_stack.append(mults)
@@ -70,7 +70,6 @@ class Renderer:
   global_prod_max: tuple[int, ...]|None = None
   shared_max: int = 32768
   tensor_cores: list[TensorCore] = []
-  pre_matcher: PatternMatcher|None = None
   extra_matcher: PatternMatcher|None = None
   code_for_op: dict[Ops, Callable] = {}
 
@@ -83,4 +82,4 @@ class Renderer:
   def aux(self, uops:list[UOp]) -> dict: raise NotImplementedError("needs aux")
   def supported_dtypes(self) -> set[DType]:
     # double can't be bitcast to anything without long support
-    return set(dtypes.all) - {dtypes.weakint} - ({dtypes.double} if dtypes.long in EMULATED_DTYPES.tolist(dtypes) else set())
+    return set(dtypes.all) - ({dtypes.double} if dtypes.long in EMULATED_DTYPES.tolist(dtypes) else set())

tinygrad/renderer/cstyle.py

@@ -3,8 +3,8 @@ import math, sys, struct
 from collections import defaultdict, Counter
 from tinygrad.codegen.opt import tc
 from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat, range_str, axis_letters
-from tinygrad.helpers import strip_parens, getenv, prod, dedup, Target, CPU_COUNT, IMAGE, FLOAT16
-from tinygrad.dtype import ImageDType, dtypes, DType, PtrDType, AddrSpace, truncate, float_to_bf16
+from tinygrad.helpers import strip_parens, getenv, prod, dedup, Target, CPU_COUNT, IMAGE, FLOAT16, is_image_shape
+from tinygrad.dtype import dtypes, DType, AddrSpace, truncate, float_to_bf16
 from tinygrad.renderer import Renderer
 
 
@@ -70,11 +70,11 @@ base_rewrite = PatternMatcher([
 def create_non_native_float_pats(dts:tuple[DType, ...], casting:bool=True):
   patterns = PatternMatcher([
     (UPat(Ops.WHERE, src=(UPat.var("b"), UPat.var("x", dtype=dts), UPat.var("y", dtype=dts))),
-     lambda b,x,y: UOp(Ops.WHERE, dtype=dtypes.float, src=(b,x.cast(dtypes.float),y.cast(dtypes.float))).cast(x.dtype)),
+     lambda b,x,y: UOp(Ops.WHERE, src=(b,x.cast(dtypes.float),y.cast(dtypes.float))).cast(x.dtype)),
     (UPat(GroupOp.ALU, dtype=dts, name="x"),
-     lambda x: UOp(x.op, dtypes.float, tuple(vv.cast(dtypes.float) for vv in x.src), x.arg).cast(x.dtype)),
+     lambda x: UOp(x.op, src=tuple(vv.cast(dtypes.float) for vv in x.src), arg=x.arg).cast(x.dtype)),
     (UPat(GroupOp.ALU, dtypes.bool, name="alu", src=(UPat.var("x", dtype=dts), UPat.var("y", dtype=dts))),
-     lambda alu,x,y: UOp(alu.op, dtypes.bool, (x.cast(dtypes.float), y.cast(dtypes.float)), alu.arg))])
+     lambda alu,x,y: UOp(alu.op, src=(x.cast(dtypes.float), y.cast(dtypes.float)), arg=alu.arg))])
   if casting:
     # add float intermediate casting
     patterns += PatternMatcher([
@@ -95,15 +95,8 @@ pm_manual_bf16_cast = PatternMatcher([
   (UPat(Ops.CAST, dtype=dtypes.bfloat16, src=(UPat.var("x", dtype=dtypes.float),)), cast_float_to_bf16),
 ])
 
-def uops_to_dtypes(uops:list[UOp]) -> list[DType]:
-  ret = []
-  seen = set()
-  for u in uops:
-    if u.addrspace in (AddrSpace.ALU, None) and u.dtype != dtypes.void and u._shape is not None and (key:=(u.dtype, u.max_numel())) not in seen:
-      # TODO: this eventually needs to be removed
-      ret.append(u.dtype.vec(u.max_numel()))
-      seen.add(key)
-  return ret
+def uops_to_dtypes(uops:list[UOp]) -> list[tuple[DType, int]]:
+  return dedup((u.dtype, u.max_numel()) for u in uops if u.addrspace in (AddrSpace.ALU, None) and u.dtype != dtypes.void and u._shape is not None)
 
 # (name, dims, dtype_in, dtype_out, device, threads, upcast_axes, reduce_axes)
 def wmma_args(uops:list[UOp]):
@@ -140,9 +133,9 @@ class CStyleLanguage(Renderer):
 
   def render_kernel(self, function_name:str, kernel:list[str], bufs:list[tuple[str,tuple[UOp,bool]]], uops:list[UOp], prefix=None) -> str:
     tmp = ""
-    if any(isinstance(u.dtype, ImageDType) for _,(u,_) in bufs):
+    if any(is_image_shape(u._shape) for _,(u,_) in bufs):
       tmp = "const sampler_t smp = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;\n"
-    buftypes = [(name, self._render_dtype(u.dtype, sz=1, addrspace=u.addrspace, mutable=mutable)+self.buffer_suffix \
+    buftypes = [(name, self._render_dtype(u.dtype, sz=1, addrspace=u.addrspace, mutable=mutable, shape=u._shape)+self.buffer_suffix \
                  if u.addrspace == AddrSpace.GLOBAL else self.arg_int_prefix if u.dtype == dtypes.int else None) for name,(u,mutable) in bufs]
     local_dims = [u.src[0] for u in uops if u.op is Ops.SPECIAL and u.arg[0] == "l"]
     launch_bounds = prod([d.vmax for d in local_dims])
@@ -164,8 +157,8 @@ class CStyleLanguage(Renderer):
     suffix = f"[{x.max_numel()}]"
     return f"{prefix}{self._render_dtype(x.dtype, sz=lanes)} {self[x]}{suffix};"
 
-  def _render_dtype(self, dtype:DType, sz:int=1, addrspace=AddrSpace.ALU, mutable=True, override_ptr=False):
-    if isinstance(dtype, ImageDType): return f"{'write_only' if mutable else 'read_only'} image2d_t"
+  def _render_dtype(self, dtype:DType, sz:int=1, addrspace=AddrSpace.ALU, mutable=True, override_ptr=False, shape=None):
+    if is_image_shape(shape): return f"{'write_only' if mutable else 'read_only'} image2d_t"
     prefix, suffix = "", ""
     if addrspace in (AddrSpace.LOCAL, AddrSpace.GLOBAL):
       if addrspace == AddrSpace.LOCAL and self.smem_prefix_for_cast: prefix = self.smem_prefix
@@ -176,16 +169,16 @@ class CStyleLanguage(Renderer):
       return prefix + self.type_map.get(scalar:=dtype.scalar(), scalar.name).replace(" ", "_") + str(sz) + suffix
     return prefix + self.type_map.get(scalar:=dtype.scalar(), scalar.name) + suffix
 
-  def render_type(self, u:UOp): return self._render_dtype(u.dtype, u.max_numel(), u.addrspace)
+  def render_type(self, u:UOp): return self._render_dtype(u.dtype, u.max_numel(), u.addrspace, shape=u._shape)
   def render_access(self, u:UOp):
     if u.max_numel() > 1 or u.dtype != u.src[0].dtype:
-      return f"*(({self._render_dtype(u.dtype, u.max_numel(), u.addrspace, override_ptr=True)})({self[u]}))"
+      return f"*(({self._render_dtype(u.dtype, u.max_numel(), u.addrspace, override_ptr=True, shape=u._shape)})({self[u]}))"
     else: return f"*{self[u]}"
   def render_cast(self, u:UOp, val:str) -> str: return f"({self.render_type(u)})({val})"
 
   # LEGACY
   def render_dtype(self, dt:DType, mutable=True) -> str:
-    return self._render_dtype(dt, dt.count, dt.addrspace if isinstance(dt, PtrDType) else AddrSpace.REG)
+    return self._render_dtype(dt, 1, AddrSpace.REG)
 
   def __getitem__(self, key): return self.r[key]  # hacky helper
   def _render(self, uops:list[UOp]) -> tuple[str, list[str], list[tuple[str,tuple[UOp,bool]]]]:
@@ -227,7 +220,7 @@ class CStyleLanguage(Renderer):
       assert l is not None, f"failed to render {u.op} {u.dtype} {[(x.op,x.dtype) for x in u.src]} {u.arg}"
 
       if u.op in {Ops.ENDIF, Ops.END}: depth -= 1
-      if (u.op is not Ops.CAST or u.dtype.vcount == 1) and (u.op in {Ops.CONST, Ops.INDEX, Ops.SHRINK, Ops.CUSTOMI} or \
+      if (u.op is not Ops.CAST or u.max_numel() == 1) and (u.op in {Ops.CONST, Ops.INDEX, Ops.SHRINK, Ops.CUSTOMI} or \
         (u.op is Ops.LOAD and u.src[0].addrspace == AddrSpace.REG) or \
         (u.op is Ops.CAST and u.addrspace in (AddrSpace.GLOBAL, AddrSpace.LOCAL)) or \
         (u.op in {Ops.STACK, *(GroupOp.ALU-{Ops.WHERE}), Ops.CAST, Ops.BITCAST} and child_count[u] == 1 and not getenv("EXPAND_SSA"))):
@@ -271,12 +264,13 @@ class ClangRenderer(CStyleLanguage):
 
   if sys.platform == 'win32':
     kernel_typedef = "__attribute__((ms_abi)) void"
-  def render_vector_prefix(self, dt:DType) -> str:
+  def render_vector_prefix(self, dt:DType, count:int) -> str:
     # round (down) to power of two (this is actually the default clang behavior)
-    alignment = 2**int(math.log2(dt.itemsize)) if getenv("ALIGNED", 1) and not dtypes.is_bool(dt) else 1
-    return f"typedef {self.render_dtype(dt.scalar())} {self.render_dtype(dt)} __attribute__((aligned({alignment}),ext_vector_type({dt.count})));"
+    alignment = 2**int(math.log2(dt.itemsize * count)) if getenv("ALIGNED", 1) and not dtypes.is_bool(dt) else 1
+    vec = self._render_dtype(dt, count, AddrSpace.REG)
+    return f"typedef {self.render_dtype(dt)} {vec} __attribute__((aligned({alignment}),ext_vector_type({count})));"
 
-  def _render_defines(self, uops) -> list[str]: return [self.render_vector_prefix(dt) for dt in uops_to_dtypes(uops) if dt.count > 1]
+  def _render_defines(self, uops) -> list[str]: return [self.render_vector_prefix(dt, count) for dt, count in uops_to_dtypes(uops) if count > 1]
   def _render_body(self, function_name, kernel, bufs, uops, pref=None) -> str: return super().render_kernel(function_name, kernel, bufs, uops, pref)
   def _render_entry(self, function_name:str, bufs:list[tuple[str,tuple[UOp,bool]]]) -> str: return ""
 
@@ -323,14 +317,14 @@ class OpenCLRenderer(CStyleLanguage):
   ]) + base_rewrite
 
   def render_kernel(self, function_name, kernel, bufs, uops, prefix=None) -> str:
-    if any(uop.dtype.base == dtypes.half for uop in uops): prefix = (["#pragma OPENCL EXTENSION cl_khr_fp16 : enable"] + (prefix or []))
+    if any(uop.dtype == dtypes.half for uop in uops): prefix = (["#pragma OPENCL EXTENSION cl_khr_fp16 : enable"] + (prefix or []))
     return super().render_kernel(function_name, kernel, bufs, uops, prefix)
 
   def aux(self, uops:list[UOp]):
-    arg_dtypes:list[list[tuple[int, DType]]] = []
+    arg_dtypes:list[list[tuple[int, DType, tuple|None]]] = []
     for i,u in enumerate(u for u in uops if u.op is Ops.PARAM):
       while len(arg_dtypes) <= u.arg.slot: arg_dtypes.append([])
-      arg_dtypes[u.arg.slot].append((i, u.dtype))
+      arg_dtypes[u.arg.slot].append((i, u.dtype, u._shape))
     return tuple(tuple(a) for a in arg_dtypes),
 
   def supported_dtypes(self): return {d for d in super().supported_dtypes()
@@ -363,7 +357,7 @@ class MetalRenderer(CStyleLanguage):
   extra_matcher = PatternMatcher([
     # NOTE: this is copied from PTX
     (UPat((Ops.SQRT, Ops.EXP2, Ops.LOG2, Ops.SIN), dtype=dtypes.bfloat16, name="x"),
-      lambda x: (UOp(x.op, dtypes.float, tuple(vv.cast(dtypes.float) for vv in x.src), x.arg).cast(dtypes.bfloat16))),
+      lambda x: (UOp(x.op, src=tuple(vv.cast(dtypes.float) for vv in x.src), arg=x.arg).cast(dtypes.bfloat16))),
   ])
 
   string_rewrite = PatternMatcher([
@@ -373,8 +367,10 @@ class MetalRenderer(CStyleLanguage):
   def render_kernel(self, function_name, kernel, bufs, uops, prefix=None):
     prefix = ["#include <metal_stdlib>","using namespace metal;"]
     deduped_wmma_args = dedup([(name, dtype_in, dtype_out) for name, _, dtype_in, dtype_out, _, _, _, _ in wmma_args(uops)])
-    for name, dtype_in, dtype_out in deduped_wmma_args: prefix.append(
-  f"""{(dstr_out:=self.render_dtype(dtype_out.vec(2)))} __{name}({(dstr_in:=self.render_dtype(dtype_in.vec(2)))} a, {dstr_in} b, {dstr_out} c){{
+    for name, dtype_in, dtype_out in deduped_wmma_args:
+      dstr_out, dstr_in = self._render_dtype(dtype_out, 2, AddrSpace.REG), self._render_dtype(dtype_in, 2, AddrSpace.REG)
+      prefix.append(
+f"""{dstr_out} __{name}({dstr_in} a, {dstr_in} b, {dstr_out} c){{
   simdgroup_{self.render_dtype(dtype_in)}8x8 mat_a, mat_b; simdgroup_{self.render_dtype(dtype_out)}8x8 mat_c;
   mat_a.thread_elements()[0] = a[0]; mat_b.thread_elements()[0] = b[0]; mat_c.thread_elements()[0] = c[0];
   mat_a.thread_elements()[1] = a[1]; mat_b.thread_elements()[1] = b[1]; mat_c.thread_elements()[1] = c[1];
@@ -424,26 +420,27 @@ class CUDARenderer(CStyleLanguage):
     (UPat(Ops.BITCAST, name="x"), lambda ctx,x: f"tg_bitcast<{ctx.render_dtype(x.dtype)}>(({ctx.render_dtype(x.src[0].dtype)})({ctx[x.src[0]]}))"),
   ]) + base_rewrite
 
-  def render_vector_prefix(self, dt:DType) -> str:
-    vec, scal = self.render_dtype(dt), self.render_dtype(dt.scalar()),
-    elems, header = ', '.join(_nms[:dt.count]), ', '.join([f"{scal} {x}" for x in _nms[:dt.count]])
-    return f"struct __align__({dt.itemsize}) {vec} {{ {scal} {elems}; }}; __device__ {vec} make_{vec}({header}) {{ {vec} r={{{elems}}}; return r; }}"
+  def render_vector_prefix(self, dt:DType, count:int) -> str:
+    vec, scal = self._render_dtype(dt, count, AddrSpace.REG), self.render_dtype(dt)
+    elems, header = ', '.join(_nms[:count]), ', '.join([f"{scal} {x}" for x in _nms[:count]])
+    return f"struct __align__({dt.itemsize * count}) {vec} {{ {scal} {elems}; }}; " \
+           f"__device__ {vec} make_{vec}({header}) {{ {vec} r={{{elems}}}; return r; }}"
 
   def render_kernel(self, function_name, kernel, bufs, uops, prefix=None):
     # TODO: why is dtypes.bfloat16.name == "__bf16"? would be easier not override dtypes.name
     prefix = ["#define INFINITY (__int_as_float(0x7f800000))", "#define NAN (__int_as_float(0x7fffffff))",
               "template <class T, class F> __device__ __forceinline__ T tg_bitcast(F v) { union U { F f; T t; }; U u; u.f = v; return u.t; }"]
     used_dtypes = uops_to_dtypes(uops)
-    if any(dt.scalar() in dtypes.fp8s for dt in used_dtypes): prefix.append("#include <cuda_fp8.h>")
-    if any(dt.scalar() == dtypes.half for dt in used_dtypes): prefix.append("#include <cuda_fp16.h>")
-    if any(dt.scalar() == dtypes.bfloat16 for dt in used_dtypes): prefix.append("#include <cuda_bf16.h>")
-    prefix += [self.render_vector_prefix(dt) for dt in used_dtypes if (dt.count in (4,8) and dt.scalar() in {dtypes.half, dtypes.bfloat16})
-      or (dt.count in (2,4,8,16) and dt.scalar() in dtypes.fp8s)]
+    if any(dt in dtypes.fp8s for dt, _ in used_dtypes): prefix.append("#include <cuda_fp8.h>")
+    if any(dt == dtypes.half for dt, _ in used_dtypes): prefix.append("#include <cuda_fp16.h>")
+    if any(dt == dtypes.bfloat16 for dt, _ in used_dtypes): prefix.append("#include <cuda_bf16.h>")
+    prefix += [self.render_vector_prefix(dt, count) for dt, count in used_dtypes if (count in (4,8) and dt in {dtypes.half, dtypes.bfloat16})
+      or (count in (2,4,8,16) and dt in dtypes.fp8s)]
     dt_map_in = { dtypes.float: "tf32", dtypes.half: "f16", dtypes.bfloat16: "bf16", dtypes.fp8e4m3: "e4m3", dtypes.fp8e5m2: "e5m2" }
     dt_map_out = { dtypes.float: "f32", dtypes.half: "f16" }
     for name, (N, M, K), dtype_in, dtype_out, _, _, upcast_axes, _ in wmma_args(uops):
       upcast_sizes = [prod(size for _, size in upcast) for upcast in upcast_axes]
-      wmma_dtypes = [self.render_dtype(dtype.vec(size)) for dtype, size in zip([dtype_in, dtype_in, dtype_out], upcast_sizes)]
+      wmma_dtypes = [self._render_dtype(dtype, size, AddrSpace.REG) for dtype, size in zip([dtype_in, dtype_in, dtype_out], upcast_sizes)]
       n_operands = [size*dtype.itemsize//4 for dtype, size in zip([dtype_in, dtype_in, dtype_out], upcast_sizes)] # 4 => CUDA reg size in bytes
       operands = [f"%{i}" for i in range(sum(n_operands))]
 
@@ -514,9 +511,9 @@ class HIPRenderer(CStyleLanguage):
   float4 = "make_float4"
   type_map = {dtypes.bfloat16: "hip_bfloat16", dtypes.fp8e4m3: "hip_fp8", dtypes.fp8e5m2: "hip_bf8"}
   extra_matcher = create_non_native_float_pats((dtypes.bfloat16, *dtypes.fp8s)) + PatternMatcher([
-    (UPat(Ops.WMMA, name="x", dtype=dtypes.float.vec(4)),
-      lambda x: UOp(Ops.WMMA, x.dtype, (x.src[0].bitcast(dtypes.uint64), x.src[1].bitcast(dtypes.uint64),
-        x.src[2]), (*x.arg,)) if x.src[0].dtype in (dtypes.fp8e4m3.vec(8), dtypes.fp8e5m2.vec(8)) else None),
+    (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+      lambda x: UOp(Ops.WMMA, src=(x.src[0].bitcast(dtypes.uint64), x.src[1].bitcast(dtypes.uint64),
+        x.src[2]), arg=(*x.arg,)) if x.src[0].max_numel() == 8 and x.src[0].dtype in dtypes.fp8_ocp else None),
     # bfloat16 constant casting
     (UPat.cvar('x', dtypes.bfloat16), lambda x: cast_float_to_bf16(UOp.const(dtypes.float, x.arg))),
   ])
@@ -525,10 +522,10 @@ class HIPRenderer(CStyleLanguage):
     from tinygrad.renderer.amd.elf import assemble_linear
     return assemble_linear(prg, lin, self.target.arch)
 
-  def render_vector_prefix(self, dtype:DType) -> str:
-    vec, scal = self.render_dtype(dtype), self.render_dtype(dtype.scalar())
-    return f"typedef {scal} {vec} __attribute__((ext_vector_type({dtype.count})));\nstatic inline __attribute__((device)) "+ \
-           f"{vec} make_{vec}({', '.join([f'{scal} {x}' for x in _nms[:dtype.count]])}) {{ return {{ {', '.join(_nms[:dtype.count])} }}; }}"
+  def render_vector_prefix(self, dtype:DType, count:int) -> str:
+    vec, scal = self._render_dtype(dtype, count, AddrSpace.REG), self.render_dtype(dtype)
+    return f"typedef {scal} {vec} __attribute__((ext_vector_type({count})));\nstatic inline __attribute__((device)) "+ \
+           f"{vec} make_{vec}({', '.join([f'{scal} {x}' for x in _nms[:count]])}) {{ return {{ {', '.join(_nms[:count])} }}; }}"
 
   def render_kernel(self, function_name, kernel, bufs, uops, prefix=None) -> str:
     prefix, ockl = [], []
@@ -542,10 +539,10 @@ class HIPRenderer(CStyleLanguage):
     ocml_ops = {Ops.EXP2: ("exp2", "pure"), Ops.LOG2: ("log2", "pure"), Ops.SQRT: ("sqrt", "const"), Ops.SIN: ("sin", ""), Ops.TRUNC: ("trunc", "")}
     ocml = [(f"__ocml_{ocml_ops[op][0]}_f{dt.bitsize}", dt.name, dt.name, ocml_ops[op][1])
       for op, dt in dedup((u.op, u.dtype.scalar()) for u in uops) if op in ocml_ops and dt in (dtypes.half, dtypes.float, dtypes.double)]
-    if any(dt.scalar() == dtypes.bfloat16 for dt in used_dtypes):
+    if any(dt == dtypes.bfloat16 for dt, _ in used_dtypes):
       prefix.append(f"typedef {'__bf16' if self.is_cdna4(self.target.arch) else 'unsigned short'} hip_bfloat16;")
-    if any(dt.scalar() == dtypes.half for dt in used_dtypes): prefix.append("#define half _Float16")
-    if any(dt.scalar() in dtypes.fp8s for dt in used_dtypes):
+    if any(dt == dtypes.half for dt, _ in used_dtypes): prefix.append("#define half _Float16")
+    if any(dt in dtypes.fp8s for dt, _ in used_dtypes):
       prefix += ["typedef unsigned char hip_bf8;", "typedef unsigned char hip_fp8;"]
     if any((u.op is Ops.CAST and u.dtype in dtypes.fp8s and u.src[0].dtype == dtypes.float) or
            (u.op is Ops.CONST and u.dtype in dtypes.fp8s) for u in uops):
@@ -553,7 +550,7 @@ class HIPRenderer(CStyleLanguage):
   v = (((*(unsigned*)&v)&0x7F800000)!=0x7F800000)?__builtin_amdgcn_fmed3f(v,is_bf8?57344.0f:448.0f,is_bf8?-57344.0f:-448.0f) : v;
   return (unsigned char)(is_bf8?__builtin_amdgcn_cvt_pk_bf8_f32(v,v,0,false):__builtin_amdgcn_cvt_pk_fp8_f32(v,v,0,false));\n}""")
     prefix += [f'extern "C" __attribute__((device{f", {atr}" if atr else ""})) {dto} {meth}({dti});' for meth,dti,dto,atr in ockl+ocml]
-    prefix += [self.render_vector_prefix(dt) for dt in used_dtypes if dt.count > 1]
+    prefix += [self.render_vector_prefix(dt, count) for dt, count in used_dtypes if count > 1]
 
     for name, (N, M, K), dtype_in, dtype_out, _, _, _, _ in wmma_args(uops): # TODO: handle TCs f32_bf16 and bf16_bf16 w/ wrapper
       if self.is_cdna(self.target.arch):

tinygrad/renderer/isa/__init__.py

@@ -9,6 +9,7 @@ class Register:
   name: str
   index: int
   _cons: tuple[Register, ...] = field(default_factory=tuple)
+  size: int = 8
   @property
   def cons(self): return self._cons or (self,)
   def __repr__(self): return self.name

tinygrad/renderer/isa/x86.py

@@ -44,11 +44,9 @@ class X86Ops(FastEnum):
   # jumps
   JNE = auto(); JE = auto(); JL = auto(); JB = auto(); JGE = auto(); JMP = auto()
   # vectorize / gep
-  VSHUFPS = auto(); VSHUFPD = auto(); VINSERTPS = auto(); VPSRLDQ = auto()
+  VINSERTPS = auto(); VPSRLDQ = auto()
   VPEXTRB = auto(); VPEXTRW = auto(); VPEXTRD = auto(); VPEXTRQ = auto()
   VPINSRB = auto(); VPINSRW = auto(); VPINSRD = auto(); VPINSRQ = auto()
-  VPBROADCASTB = auto(); VPBROADCASTW = auto(); VPBROADCASTD = auto(); VPBROADCASTQ = auto()
-  VBROADCASTSS = auto()
   # int binary
   IDIV = auto(); DIV = auto()
   ADD = auto(); ADDi = auto(); SUB = auto(); SUBi = auto(); IMUL = auto(); IMULi = auto()
@@ -62,8 +60,6 @@ class X86Ops(FastEnum):
   VSUBSS = auto(); VSUBSD = auto(); VSUBPS = auto(); VSUBPD = auto()
   VMULSS = auto(); VMULSD = auto(); VMULPS = auto(); VMULPD = auto()
   VDIVSS = auto(); VDIVSD = auto(); VDIVPS = auto(); VDIVPD = auto()
-  VMAXSS = auto(); VMAXSD = auto(); VMAXPS = auto(); VMAXPD = auto()
-  VMINSS = auto(); VMINSD = auto(); VMINPS = auto(); VMINPD = auto()
   # int vector binary
   VPADDB = auto(); VPADDW = auto(); VPADDD = auto(); VPADDQ = auto()
   VPSUBB = auto(); VPSUBW = auto(); VPSUBD = auto(); VPSUBQ = auto()
@@ -72,8 +68,6 @@ class X86Ops(FastEnum):
   VPAND = auto(); VPOR = auto(); VPXOR = auto()
   # packed variable shifts
   VPSLLVD = auto(); VPSLLVQ = auto(); VPSRLVD = auto(); VPSRLVQ = auto(); VPSRAVD = auto()
-  # fused multiply add
-  VFMADD213SS = auto(); VFMADD213SD = auto(); VFMADD213PS = auto(); VFMADD213PD = auto()
   # return
   RET = auto()
 
@@ -81,8 +75,7 @@ class X86GroupOp:
   # X86Ops whose first src is also the destination
   TwoAddress = {X86Ops.ADD, X86Ops.ADDi, X86Ops.AND, X86Ops.ANDi, X86Ops.XOR, X86Ops.XORi, X86Ops.OR, X86Ops.ORi, X86Ops.IMUL,
                 X86Ops.SUB, X86Ops.SUBi, X86Ops.SHL, X86Ops.SHLi, X86Ops.SHR, X86Ops.SHRi, X86Ops.SAR, X86Ops.SARi,
-                X86Ops.IDIV, X86Ops.DIV, X86Ops.VFMADD213SS, X86Ops.VFMADD213SD, X86Ops.VFMADD213PS, X86Ops.VFMADD213PD,
-                X86Ops.CMOVNE, X86Ops.CMOVE, X86Ops.CMOVL, X86Ops.CMOVB}
+                X86Ops.IDIV, X86Ops.DIV, X86Ops.CMOVNE, X86Ops.CMOVE, X86Ops.CMOVL, X86Ops.CMOVB}
 
   # X86Ops whose first src can read from memory
   ReadMem1st = {X86Ops.MOV, X86Ops.VMOVSS, X86Ops.VMOVSD, X86Ops.VMOVUPS, X86Ops.MOVZX, X86Ops.MOVSX, X86Ops.MOVSXD, X86Ops.VMOVD, X86Ops.VMOVQ,
@@ -90,7 +83,6 @@ class X86GroupOp:
                 X86Ops.VPMOVSXBW, X86Ops.VPMOVSXBD, X86Ops.VPMOVSXBQ, X86Ops.VPMOVSXWD, X86Ops.VPMOVSXWQ, X86Ops.VPMOVSXDQ,
                 X86Ops.VCVTDQ2PS, X86Ops.VCVTDQ2PD, X86Ops.VCVTTPS2DQ, X86Ops.VCVTTPD2DQ, X86Ops.VCVTTSS2SI, X86Ops.VCVTTSD2SI,
                 X86Ops.VCVTPH2PS, X86Ops.VCVTPS2PD, X86Ops.VCVTPD2PS, X86Ops.VROUNDPS, X86Ops.VROUNDPD, X86Ops.VSQRTPS, X86Ops.VSQRTPD,
-                X86Ops.VPBROADCASTB, X86Ops.VPBROADCASTW, X86Ops.VPBROADCASTD, X86Ops.VPBROADCASTQ, X86Ops.VBROADCASTSS,
                 X86Ops.CMPi, X86Ops.IMULi, X86Ops.LEA}
 
   # X86Ops whose second src can read from memory NOTE: some of these are TwoAddress so the second src is actually the first
@@ -100,15 +92,10 @@ class X86GroupOp:
                 X86Ops.VPADDB, X86Ops.VPADDW, X86Ops.VPADDD, X86Ops.VPADDQ, X86Ops.VPSUBB, X86Ops.VPSUBW, X86Ops.VPSUBD, X86Ops.VPSUBQ,
                 X86Ops.VPCMPEQB, X86Ops.VPCMPEQW, X86Ops.VPCMPEQD, X86Ops.VPCMPEQQ, X86Ops.VPBLENDVB, X86Ops.VBLENDVPS, X86Ops.VBLENDVPD,
                 X86Ops.VPCMPGTB, X86Ops.VPCMPGTW, X86Ops.VPCMPGTD, X86Ops.VPCMPGTQ, X86Ops.VCMPSS, X86Ops.VCMPSD, X86Ops.VCMPPS, X86Ops.VCMPPD,
-                X86Ops.VPMULLW, X86Ops.VPMULLD, X86Ops.VROUNDSS, X86Ops.VROUNDSD, X86Ops.VSQRTSS, X86Ops.VSQRTSD, X86Ops.VSHUFPS, X86Ops.VINSERTPS,
+                X86Ops.VPMULLW, X86Ops.VPMULLD, X86Ops.VROUNDSS, X86Ops.VROUNDSD, X86Ops.VSQRTSS, X86Ops.VSQRTSD, X86Ops.VINSERTPS,
                 X86Ops.VPINSRB, X86Ops.VPINSRW, X86Ops.VPINSRD, X86Ops.VPINSRQ, X86Ops.VPAND, X86Ops.VPOR, X86Ops.VPXOR, X86Ops.VPSLLVD,
                 X86Ops.VPSLLVQ, X86Ops.VPSRLVD, X86Ops.VPSRLVQ, X86Ops.VPSRAVD, X86Ops.CMOVNE, X86Ops.CMOVE, X86Ops.CMOVL, X86Ops.CMOVB,
-                X86Ops.VMAXSS, X86Ops.VMAXSD, X86Ops.VMAXPS, X86Ops.VMAXPD, X86Ops.VMINSS, X86Ops.VMINSD, X86Ops.VMINPS, X86Ops.VMINPD,
-                X86Ops.VCVTSI2SS, X86Ops.VCVTSI2SD, X86Ops.VCVTSS2SD, X86Ops.VCVTSD2SS, X86Ops.VUCOMISS, X86Ops.VUCOMISD, X86Ops.IDIV, X86Ops.DIV,
-                X86Ops.VSHUFPD}
-
-  # X86Ops whose third src can read from memory NOTE: these are TwoAddress so the third src is actually the second
-  ReadMem3rd = {X86Ops.VFMADD213SS, X86Ops.VFMADD213SD, X86Ops.VFMADD213PS, X86Ops.VFMADD213PD}
+                X86Ops.VCVTSI2SS, X86Ops.VCVTSI2SD, X86Ops.VCVTSS2SD, X86Ops.VCVTSD2SS, X86Ops.VUCOMISS, X86Ops.VUCOMISD, X86Ops.IDIV, X86Ops.DIV}
 
   # X86Ops that can write to memory
   WriteMem = {X86Ops.MOVm, X86Ops.MOVi, X86Ops.VMOVSSm, X86Ops.VMOVSDm, X86Ops.VMOVUPSm, X86Ops.VMOVDm, X86Ops.VMOVQm,
@@ -128,7 +115,7 @@ class X86GroupOp:
   Rm1st = ReadMem1st | (ReadMem2nd & TwoAddress) | {X86Ops.VPSRLDQ}
 
   # X86Ops whose second src is the rm field
-  Rm2nd = ReadMem2nd | (ReadMem3rd & TwoAddress)
+  Rm2nd = ReadMem2nd
 
   All = set(X86Ops)
 
@@ -153,20 +140,20 @@ extra_matcher = PatternMatcher([
   # no int8 mul or cmove, cast to int16
   (UPat.var("a", dtypes.int8s) * UPat.var("b"), lambda a,b: (a.cast(dtypes.int16) * b.cast(dtypes.int16)).cast(a.dtype)),
   (UPat.var("m").where(UPat.var("a", (dtypes.bool,)+dtypes.int8s), UPat.var("b")),
-   lambda m,a,b: m.where(a.cast(dtypes.int16), b.cast(dtypes.int16)).cast(a.dtype) if a.dtype.count == 1 else None),
+   lambda m,a,b: m.where(a.cast(dtypes.int16), b.cast(dtypes.int16)).cast(a.dtype) if a.max_numel() == 1 else None),
   # float16 alus are done in float32
-  (UPat(GroupOp.ALU, dtypes.float16, name="x"), lambda x: UOp(x.op, dtypes.float.vec(x.dtype.count),
+  (UPat(GroupOp.ALU, dtypes.float16, name="x"), lambda x: UOp(x.op, dtypes.float,
    tuple(s.cast(dtypes.float) if s.dtype != dtypes.bool else s for s in x.src)).cast(x.dtype)),
   (UPat(GroupOp.Comparison, src=(UPat.var("a", dtypes.float16), UPat.var("b")), name="x"),
-   lambda x,a,b: UOp(x.op, x.dtype, (a.cast(dtypes.float32), b.cast(dtypes.float32))).cast(x.dtype)),
+   lambda x,a,b: UOp(x.op, src=(a.cast(dtypes.float32), b.cast(dtypes.float32))).cast(x.dtype)),
   # no cmpne for packed ints, y != x => !(y==x)
   (UPat(Ops.CMPNE, src=(UPat.var("y", dtypes.ints), UPat.var("x")), name="cmp"),
-   lambda y,x,cmp: UOp(Ops.CMPEQ, cmp.dtype, (y,x))^True if y.dtype.count > 1 else None),
+   lambda y,x,cmp: UOp(Ops.CMPEQ, src=(y,x))^True if y.max_numel() > 1 else None),
   # float where expects a mask
   (UPat.var("m", dtypes.bool).where(UPat.var("a", dtypes.floats), UPat.var("b")),
    lambda m,a,b: m.cast(a.dtype).ne(0).where(a, b) if m.src[0].dtype not in dtypes.floats else None),
   # rewrite -x -> 0 - x
-  (UPat(Ops.NEG, name="x"), lambda x: UOp(Ops.SUB, x.dtype, (x.const_like(0),) + x.src)),
+  (UPat(Ops.NEG, name="x"), lambda x: UOp(Ops.SUB, src=(x.const_like(0),) + x.src)),
   # TODO: add support for mod, requires support for accessing the 2nd+ reg of a multi output instruction
   (UPat(Ops.CMOD, src=(UPat.var("x"), UPat.var("y"))), lambda x,y: x - y * x.alu(Ops.CDIV, y)),
 ])
@@ -177,50 +164,36 @@ def scratch_buffer(elem_dt:DType, count:int, slot:int) -> UOp:
   return UOp.placeholder((count,), elem_dt, slot, AddrSpace.LOCAL)
 
 def gated_load(ctx, addr:UOp, alt:UOp, gate:UOp, x:UOp):
-  local = scratch_buffer(addr.src[0].dtype.scalar(), x.dtype.count, next(ctx))
+  local = scratch_buffer(addr.src[0].dtype.scalar(), x.max_numel(), next(ctx))
   local_idx = local.index(UOp.const(dtypes.int32, 0), dtype=dtypes.uint64)
   # the selected address is a 64bit value, the AFTER orders the load after the scratch store and carries the element dtype for the encoder
   sel = gate.where(addr.replace(dtype=dtypes.uint64), local_idx)
-  ptr = UOp(Ops.AFTER, addr.dtype, (sel, (local_idx if x.dtype.count == 1 else local).store(alt)))
+  ptr = UOp(Ops.AFTER, addr.dtype, (sel, (local_idx if x.max_numel() == 1 else local).store(alt)))
   return ptr.load(dtype=x.dtype)
 
 def gated_store(addr:UOp, gate:UOp, val:UOp):
-  local = scratch_buffer(addr.src[0].dtype.scalar(), val.dtype.count, -1)
+  local = scratch_buffer(addr.src[0].dtype.scalar(), val.max_numel(), -1)
   sel = gate.where(addr.replace(dtype=dtypes.uint64), local.index(UOp.const(dtypes.int32, 0), dtype=dtypes.uint64))
   return UOp(Ops.AFTER, addr.dtype, (sel,)).store(val)
 
 # legalize the new style graph for isel. NOTE: this runs after the spec is verified, some of these rewrites violate it
 pre_isel_matcher = PatternMatcher([
-  # x86 registers are typed by their width, materialize the structural width of the graph into vec dtypes (this is still valid new style)
-  (UPat(Ops.SHRINK, src=(UPat(), UPat(), UPat.cvar("c"))).load(allow_any_len=True, name="x"), lambda x,c:
-   x.replace(dtype=x.dtype.scalar().vec(c.arg)) if c.arg > x.dtype.count else None),
-  (UPat(Ops.STACK, name="x"), lambda x: x.replace(dtype=x.dtype.scalar().vec(len(x.src))) if 1 < len(x.src) != x.dtype.count else None),
-  (UPat(GroupOp.ALU.union({Ops.CAST, Ops.BITCAST}), name="x"), lambda x: x.replace(dtype=x.dtype.scalar().vec(c)) \
-    if (c:=max([s.dtype.count for s in x.src], default=1)) > x.dtype.count else None),
-  # zero extending scalar 32bit int is a noop
-  (UPat.var("y", dtypes.uint32).cast(dtypes.int64s, name="x"), lambda y,x: x.replace(op=Ops.NOOP) if y.dtype.count == 1 else None),
-  # cast between signed and unsigned int is a noop
+  # noop casts: zero extending scalar 32bit int, same-width signed/unsigned, narrowing scalar int
+  (UPat.var("y", dtypes.uint32).cast(dtypes.int64s, name="x"), lambda y,x: x.replace(op=Ops.NOOP, arg=None) if y.max_numel() == 1 else None),
   (UPat.var("y", dtypes.ints+(dtypes.bool,)).cast(dtypes.ints, name="x"),
-   lambda y,x: x.replace(op=Ops.NOOP) if x.dtype.itemsize == y.dtype.itemsize else None),
-  # cast to < scalar int is a noop
+   lambda y,x: x.replace(op=Ops.NOOP, arg=None) if x.dtype.itemsize == y.dtype.itemsize and y.max_numel() == 1 else None),
   (UPat.var("y", dtypes.ints).cast(dtypes.ints, name="x"),
-   lambda y,x: x.replace(op=Ops.NOOP) if x.dtype.itemsize < y.dtype.itemsize and y.dtype.count == 1 else None),
+   lambda y,x: x.replace(op=Ops.NOOP, arg=None) if x.dtype.itemsize < y.dtype.itemsize and y.max_numel() == 1 else None),
   # bitcasts between scalar floats and ints are real, rest are noops
   (UPat.var("y").bitcast().named("x"), lambda y,x: None if y.dtype in dtypes.floats and x.dtype in dtypes.ints or \
-   y.dtype in dtypes.ints and x.dtype in dtypes.floats else x.replace(op=Ops.NOOP)),
-  # noop of a noop is removed
-  (UPat(Ops.NOOP, src=(UPat(Ops.NOOP),), name="x"), lambda x: x.replace(src=x.src[0].src)),
-  # moving elements of a single register to another without shuffling is a noop
-  (UPat(Ops.STACK, src=(UPat.var("y").index(UPat()),), allow_any_len=True, name="x"),
-   lambda y,x: UOp(Ops.NOOP, x.dtype, (y,)) if all(s.op is Ops.INDEX and len(s.src) == 2 and s.src[0] is y \
-     and s.src[1].op is Ops.CONST and s.src[1].arg == i for i,s in enumerate(x.src)) else None),
+   y.dtype in dtypes.ints and x.dtype in dtypes.floats else x.replace(op=Ops.NOOP, arg=None)),
   # gated load/store become a conditional move on the address, the load/store are unconditional
   (UPat((Ops.INDEX, Ops.SHRINK), name="addr").load(UPat.var("alt"), UPat.var("gate"), name="x"), gated_load),
   (UPat((Ops.INDEX, Ops.SHRINK), name="addr").store(UPat.var("val"), UPat.var("gate")), gated_store),
   # TODO: remove this once we allow all flag producing ops in cmove
   # if gate in scalar int cmove is not a comparison need to add one to set the flag
   (UPat.var("m", dtypes.bool).where(UPat.var("a"), UPat.var("b")),
-   lambda m,a,b: m.ne(0).where(a,b) if m.op not in GroupOp.Comparison and a.dtype.count == 1 else None),
+   lambda m,a,b: m.ne(0).where(a,b) if m.op not in GroupOp.Comparison else None),
 ])
 
 # ***** X86 registers *****
@@ -234,7 +207,7 @@ RBP = Register("rbp", 5)
 RSI = Register("rsi", 6)
 RDI = Register("rdi", 7)
 GPR = (RAX, RCX, RDX, RBX, RSP, RBP, RSI, RDI) + tuple(Register(f"r{i}", i) for i in range(8, 16))
-XMM = tuple(Register(f"xmm{i}", i) for i in range(16))
+XMM = tuple(Register(f"xmm{i}", i, size=16) for i in range(16))
 # gprs you can write to
 WGPR = tuple(r for r in GPR if r != RSP)
 
@@ -245,8 +218,6 @@ reg_strs = {"rax": {4:"eax", 2:"ax", 1:"al"}, "rcx": {4:"ecx", 2:"cx", 1:"cl"},
         **{f"r{i}": {4:f"r{i}d", 2:f"r{i}w", 1:f"r{i}b"} for i in range(8, 16)}, **{f"xmm{i}": {64:f"zmm{i}", 32:f"ymm{i}"} for i in range(16)}}
 
 # ***** X86 instruction selection *****
-# if s is used multiple times we don't fold
-def is_foldable(ctx:IselContext, x:UOp, s:UOp) -> bool: return len(ctx.uses[s]) == x.src.count(s) == 1
 def base(x:UOp, i:int) -> UOp: return s.src[0] if (s:=x.src[i]).op is Ops.INDEX else s
 def lane(x:UOp, i:int) -> int: return s.src[1].arg if (s:=x.src[i]).op is Ops.INDEX else 0
 def to_int(dt:DType): return {dtypes.float16: dtypes.int16, dtypes.float32: dtypes.int32, dtypes.float64: dtypes.int64}[dt]
@@ -264,27 +235,8 @@ def cmp(x:UOp) -> UOp:
   return x.ins(X86Ops.CMP, dtype=dtypes.void) if (i:=to_imm(x.src[1])) is None else x.ins(X86Ops.CMPi, dtype=dtypes.void, src=(x.src[0], i))
 def vcmp(x:UOp) -> UOp:
   v = imm(dtypes.uint8, {Ops.CMPLT: 1, Ops.CMPNE: 4, Ops.CMPEQ: 0}[x.op])
-  if x.dtype.scalar() is dtypes.float32: return x.ins(X86Ops.VCMPSS if x.dtype.count == 1 else X86Ops.VCMPPS, src=x.src + (v,))
-  return x.ins(X86Ops.VCMPSD if x.dtype.count == 1 else X86Ops.VCMPPD, src=x.src + (v,))
-
-# vshufps xmm2, xmm0, xmm1, imm
-# for 128 bit xmm2 selects its lower 2 32 bits from xmm0 and its upper 2 32 bits from xmm1 according to imm
-# for 256 bit ymm2 repeats the shuffle for its upper 128 bits selecting from the upper 128 bits of ymm0 and ymm1
-def vshufps(x:UOp) -> UOp|None:
-  a, b = base(x, 0), base(x, 2)
-  if not (a is base(x, 1) and b is base(x, 3)) or any(lane(x, i) > 3 for i in range(4)): return None
-  if len(x.src) == 8:
-    if not (a is base(x, 4) is base(x, 5) and b is base(x, 6) is base(x, 7)) or any(lane(x, i+4) != lane(x, i)+4 for i in range(4)): return None
-  return x.ins(X86Ops.VSHUFPS, src=(a, b, imm(dtypes.uint8, sum(lane(x, i) << 2*i for i in range(4)))))
-
-# vshufpd xmm2, xmm0, xmm1, imm
-# for 128 bit xmm2 selects its lower 64 bits from xmm0 and its upper 64 bits from xmm1 according to imm
-# for 256 bit ymm2 also selects its upper 128 bits from the upper 128 bits of ymm0 and ymm1 following the same constraint
-def vshufpd(x:UOp) -> UOp|None:
-  a, b = base(x, 0), base(x, 1)
-  if lane(x, 0) > 1 or lane(x, 1) > 1: return None
-  if len(x.src) == 4 and not (a is base(x, 2) and b is base(x, 3) and lane(x, 2) > 1 and lane(x, 3) > 1): return None
-  return x.ins(X86Ops.VSHUFPD, src=(a, b, imm(dtypes.uint8, sum(lane(x, i) << i for i in range(len(x.src))))))
+  if x.dtype.scalar() is dtypes.float32: return x.ins(X86Ops.VCMPSS if x.max_numel() == 1 else X86Ops.VCMPPS, src=x.src + (v,))
+  return x.ins(X86Ops.VCMPSD if x.max_numel() == 1 else X86Ops.VCMPPD, src=x.src + (v,))
 
 # vinsertps xmm2, xmm0, xmm1, imm
 # inserts any 32 bit element in xmm1 into any position in xmm0 according to immm, result is written to xmm2
@@ -292,8 +244,7 @@ def vshufpd(x:UOp) -> UOp|None:
 def vinsertps(x:UOp) -> UOp:
   def _insert(ret:UOp, i:int) -> UOp:
     s, v = base(x, i), lane(x, i)
-    # moving the 0th element into the 0th position does nothing
-    return s if i == v == 0 else x.ins(X86Ops.VINSERTPS, src=(ret, s, imm(dtypes.uint8, v << 6 | i << 4)))
+    return x.ins(X86Ops.VINSERTPS, src=(ret, s, imm(dtypes.uint8, v << 6 | i << 4)))
   return functools.reduce(_insert, range(len(x.src)), def_reg(x.dtype))
 
 # vpinsq xmm2, xmm0, rax, imm
@@ -302,16 +253,6 @@ def vpins(x:UOp) -> UOp:
   op = {1: X86Ops.VPINSRB, 2: X86Ops.VPINSRW, 4: X86Ops.VPINSRD, 8: X86Ops.VPINSRQ}[x.dtype.scalar().itemsize]
   return functools.reduce(lambda ret,i: x.ins(op, src=(ret, x.src[i], imm(dtypes.uint8, i))), range(len(x.src)), def_reg(x.dtype))
 
-# vpbroadcastd xmm1, xmm0
-# inserts scalar int in xmm0 into all lanes of xmm1
-def vpbroadcast(ctx:IselContext, x:UOp, y:UOp) -> UOp:
-  n = x.ins({1: X86Ops.VPBROADCASTB, 2: X86Ops.VPBROADCASTW, 4: X86Ops.VPBROADCASTD, 8: X86Ops.VPBROADCASTQ}[y.dtype.itemsize], src=(y,))
-  if y.op is Ops.LOAD and len(y.src) == 1 and is_foldable(ctx, n, y): return n
-  # if there isn't a load we can fold we need to move y from gpr to xmm
-  # this is hacky but required because int.vec(1) isn't supported
-  y = y if y.dtype.itemsize > 1 else y.cast(dtypes.int16)
-  return n.replace(src=(y.bitcast({2:dtypes.float16, 4:dtypes.float32, 8:dtypes.float64}[y.dtype.itemsize]),))
-
 # we don't call ctx.vreg on the srcs to avoid duplicates, a rewrite will assign the tuple of valid registers to a vreg
 def idiv(ctx:IselContext, x:UOp) -> UOp:
   op = X86Ops.DIV if x.dtype in dtypes.uints else X86Ops.IDIV
@@ -359,6 +300,25 @@ def abi(ctx:IselContext, x:UOp) -> UOp|None:
   # this move "cleanses" the abi register constraint
   return x.ins(X86Ops.MOV, dtype=dt, src=src)
 
+GPR_DEST_OPS = {X86Ops.VPEXTRB, X86Ops.VPEXTRW, X86Ops.VPEXTRD, X86Ops.VPEXTRQ, X86Ops.VCVTTSS2SI, X86Ops.VCVTTSD2SI,
+                 X86Ops.VMOVDm, X86Ops.VMOVQm}
+XMM_OPS = {op for op in X86Ops if op.name.startswith('V')} - GPR_DEST_OPS
+
+def _is_vec_xmm(y: UOp) -> bool:
+  return (y.op is Ops.INS and y.arg in XMM_OPS) or (y.op not in (Ops.BUFFER, Ops.PARAM, Ops.AFTER, Ops.INS) and y.max_numel() > 1)
+
+def _xmm_sz(x: UOp) -> X86Ops:
+  bits = x.max_numel() * x.dtype.itemsize
+  if bits >= 16: return X86Ops.VMOVUPS
+  if bits >= 8: return X86Ops.VMOVSD
+  return X86Ops.VMOVSS
+
+def _xmm_sz_m(x: UOp) -> X86Ops:
+  bits = x.max_numel() * x.dtype.itemsize
+  if bits >= 16: return X86Ops.VMOVUPSm
+  if bits >= 8: return X86Ops.VMOVSDm
+  return X86Ops.VMOVSSm
+
 def alloc_vregs(ctx:IselContext, x:UOp) -> UOp|None:
   # register placeholders with real registers
   if x.arg is X86Ops.DEFINE and x.tag is not None: return None
@@ -371,30 +331,19 @@ def alloc_vregs(ctx:IselContext, x:UOp) -> UOp|None:
   # allocate vreg definitions, the value of a BUFFER is its address so it lives in a gpr
   defs = []
   if isinstance(x.tag, tuple): defs = [ctx.vreg(x.tag)]
-  elif x.op is Ops.BUFFER or x.dtype in dtypes.ints+(dtypes.bool,): defs = [ctx.vreg(WGPR)]
-  elif x.dtype in dtypes.floats or x.dtype.count > 1: defs = [ctx.vreg(XMM)]
+  elif x.op is Ops.BUFFER: defs = [ctx.vreg(WGPR)]
+  elif x.dtype in dtypes.floats or (x.op is Ops.INS and x.arg in XMM_OPS) or x.dtype.count > 1: defs = [ctx.vreg(XMM)]
+  elif x.dtype in dtypes.ints+(dtypes.bool,): defs = [ctx.vreg(WGPR)]
   # TODO: add this once the scheduler can track register pressure
   # if x.arg in X86GroupOp.WriteFlags: defs.append(ctx.vreg(RFLAGS))
   # the size src of a BUFFER is not a value, tag it so it isn't materialized into a register
   if x.op is Ops.BUFFER: return x.replace(src=tuple(s.rtag() for s in x.src), tag=tuple(defs))
   return x.replace(tag=tuple(defs))
 
-dts = dtypes.ints + (dtypes.bool, dtypes.float16, dtypes.float32, dtypes.float64)
-dt_16bit = tuple(dt.vec(l) for dt in dts for l in [2,1] if l*dt.itemsize == 2 and dt not in dtypes.int16s)
-dt_32bit = tuple(dt.vec(l) for dt in dts for l in [4,2,1] if l*dt.itemsize == 4 and dt not in dtypes.int32s)
-dt_64bit = tuple(dt.vec(l) for dt in dts for l in [8,4,2,1] if l*dt.itemsize == 8 and dt not in dtypes.int64s)
-dt_128bit = tuple(dt.vec(l) for dt in dts for l in [16,8,4,2,1] if l*dt.itemsize == 16)
-
 isel_matcher = PatternMatcher([
   # **** Op -> Op ****
-  # materialize the structural width of a STACK into a vec dtype
-  (UPat(Ops.STACK, name="x"), lambda x: x.replace(dtype=x.dtype.scalar().vec(len(x.src))) if 1 < len(x.src) != x.dtype.count else None),
   # cast of void is a noop
   (UPat.var("y").cast(name="x"), lambda y,x: y if y.dtype == dtypes.void else None),
-  # extracting the 0th float element is a noop as it just moves the 0th element from one xmm register to another
-  # this is done here to not interfere with shuffles
-  (UPat(dtype=dtypes.floats).index(UPat(Ops.CONST, arg=0), name="x"),
-   lambda x: x.replace(op=Ops.NOOP, src=x.src[:1]) if x.src[0].dtype.count > 1 else None),
   # range is lowered to acc, cmp, jmp after regalloc
   (UPat(Ops.RANGE, src=(UPat.cvar("c"),), allow_any_len=True, name="x"), lambda c,x: x.replace(src=(imm(c.dtype, c.arg),) + x.src[1:])),
   (UPat(Ops.RANGE, name="x"), lambda ctx,x: x.replace(tag=(ctx.vreg(WGPR),)) if not isinstance(x.tag, tuple) else None),
@@ -402,7 +351,7 @@ isel_matcher = PatternMatcher([
   # add callee saved registers to the RET, these will be scheduled at the top of the kernel and will be saved/restored if they are used in regalloc
   # so regalloc builds the prologue/epilogue naturally
   (UPat(Ops.SINK, name="x"), lambda x:
-   x.replace(src=(x.ins(X86Ops.RET, src=x.src + tuple(def_reg(dtypes.uint64 if r in GPR else dtypes.float64.vec(2), r) for r in CALLEE_SAVED)),)) \
+   x.replace(src=(x.ins(X86Ops.RET, src=x.src + tuple(def_reg(dtypes.uint64 if r in GPR else dtypes.float64, r) for r in CALLEE_SAVED)),)) \
     if not x.src or x.src[0].arg is not X86Ops.RET else None),
   # function abi constraints
   (UPat((Ops.PARAM, Ops.SPECIAL), name="x"), abi),
@@ -411,15 +360,6 @@ isel_matcher = PatternMatcher([
   (UPat.cvar("x", dtypes.ints+(dtypes.bool,)), lambda x: x.ins(X86Ops.MOVi, src=(imm(x.dtype, x.arg),)) if not x.tag else None),
   (UPat.cvar("x", dtypes.floats), lambda x:
    UOp.const(dt:=to_int(x.dtype), struct.unpack(dt.fmt, struct.pack(x.dtype.fmt, x.arg))[0]).bitcast(x.dtype) if not x.tag else None),
-  # TODO: these should use a.maximum(b) / a.minimum(b)
-  ((UPat.var("a") < UPat.var("b")).where(UPat.var("b", dtypes.float32), UPat.var("a")), lambda a,b:
-   a.ins(X86Ops.VMAXSS if a.dtype.count == 1 else X86Ops.VMAXPS, src=(a, b))),
-  ((UPat.var("a") < UPat.var("b")).where(UPat.var("b", dtypes.float64), UPat.var("a")), lambda a,b:
-   a.ins(X86Ops.VMAXSD if a.dtype.count == 1 else X86Ops.VMAXPD, src=(a, b))),
-  ((UPat.var("a") < UPat.var("b")).where(UPat.var("a", dtypes.float32), UPat.var("b")), lambda a,b:
-   a.ins(X86Ops.VMINSS if a.dtype.count == 1 else X86Ops.VMINPS, src=(a, b))),
-  ((UPat.var("a") < UPat.var("b")).where(UPat.var("a", dtypes.float64), UPat.var("b")), lambda a,b:
-   a.ins(X86Ops.VMINSD if a.dtype.count == 1 else X86Ops.VMINPD, src=(a, b))),
   # conditional moves that use masks NOTE: these currently assume a mask producing cmp exists
   (UPat.var("m").where(UPat.var("a", dtypes.ints), UPat.var("b")), lambda m,a,b:
    a.ins(X86Ops.VPBLENDVB, src=(b, a, m.replace(dtype=m.src[0].dtype))) if a.dtype.count > 1 else None),
@@ -429,7 +369,7 @@ isel_matcher = PatternMatcher([
    a.ins(X86Ops.VBLENDVPD, src=(b, a, m.replace(dtype=m.src[0].dtype)))),
   # in this case we have a mask producing comparison whose user expects a bool, so we convert to bool
   (UPat(GroupOp.Comparison, dtypes.bool, (UPat.var("y", (dtypes.float32, dtypes.float64)), UPat()), name="x"), lambda y,x:
-   UOp(Ops.AND, dt:=to_int(y.dtype), (x.replace(dtype=y.dtype).bitcast(dt), UOp.const(dt, 1))).f(Ops.NOOP, dtype=dtypes.bool)),
+   UOp(Ops.AND, src=(x.replace(dtype=y.dtype).bitcast(dt:=to_int(y.dtype)), UOp.const(dt, 1))).f(Ops.NOOP, dtype=dtypes.bool)),
   # conditional moves that use flags
   (UPat(Ops.CMPLT, src=(UPat(dtype=dtypes.sints), UPat()), name="m").where(UPat.var("a"), UPat.var("b")), lambda m,a,b:
    a.ins(X86Ops.CMOVL, src=(b, a, cmp(m)))),
@@ -457,58 +397,48 @@ isel_matcher = PatternMatcher([
   (UPat(Ops.CMPLT, src=(UPat.var("a", dtypes.int32s), UPat.var("b")), name="x"), lambda a,b,x: x.ins(X86Ops.VPCMPGTD, src=(b, a))),
   (UPat(Ops.CMPLT, src=(UPat.var("a", dtypes.int64s), UPat.var("b")), name="x"), lambda a,b,x: x.ins(X86Ops.VPCMPGTQ, src=(b, a))),
   # float unary
-  (UPat.var("y", dtypes.float32).sqrt().named("x"), lambda y,x: x.ins(X86Ops.VSQRTSS, src=(y, y)) if x.dtype.count == 1 else x.ins(X86Ops.VSQRTPS)),
-  (UPat.var("y", dtypes.float64).sqrt().named("x"), lambda y,x: x.ins(X86Ops.VSQRTSD, src=(y, y)) if x.dtype.count == 1 else x.ins(X86Ops.VSQRTPD)),
+  (UPat.var("y", dtypes.float32).sqrt().named("x"), lambda y,x: x.ins(X86Ops.VSQRTSS, src=(y, y)) if x.max_numel() == 1 else x.ins(X86Ops.VSQRTPS)),
+  (UPat.var("y", dtypes.float64).sqrt().named("x"), lambda y,x: x.ins(X86Ops.VSQRTSD, src=(y, y)) if x.max_numel() == 1 else x.ins(X86Ops.VSQRTPD)),
   (UPat.var("y", dtypes.float32).trunc().named("x"), lambda y,x:
-   x.ins(X86Ops.VROUNDSS, src=(y, y, imm(dtypes.uint8, 3))) if x.dtype.count == 1 else x.ins(X86Ops.VROUNDPS, src=(y, imm(dtypes.uint8, 3)))),
+   x.ins(X86Ops.VROUNDSS, src=(y, y, imm(dtypes.uint8, 3))) if x.max_numel() == 1 else x.ins(X86Ops.VROUNDPS, src=(y, imm(dtypes.uint8, 3)))),
   (UPat.var("y", dtypes.float64).trunc().named("x"), lambda y,x:
-   x.ins(X86Ops.VROUNDSD, src=(y, y, imm(dtypes.uint8, 3))) if x.dtype.count == 1 else x.ins(X86Ops.VROUNDPD, src=(y, imm(dtypes.uint8, 3)))),
-  # shufles
-  (UPat.var("y", dtypes.float32).broadcast(name="x"), lambda y,x: x.ins(X86Ops.VBROADCASTSS, src=(y,))),
-  # for float16 we route the srcs through gprs unless we can fold them, this is suboptimal for values in xmms, in that case we want vpunpcklwd
-  (UPat(Ops.STACK, dtypes.float16, name="x"), lambda ctx,x:
-   vpins(x.replace(src=tuple(s if s.op is Ops.LOAD and is_foldable(ctx, x, s) else s.bitcast(dtypes.int16) for s in x.src)))),
-  (UPat(Ops.STACK, (dtypes.float32.vec(4), dtypes.float32.vec(8)), name="x"), vshufps),
-  (UPat(Ops.STACK, (dtypes.float64.vec(2), dtypes.float64.vec(4)), name="x"), vshufpd),
+   x.ins(X86Ops.VROUNDSD, src=(y, y, imm(dtypes.uint8, 3))) if x.max_numel() == 1 else x.ins(X86Ops.VROUNDPD, src=(y, imm(dtypes.uint8, 3)))),
+  # for float16 we route the srcs through gprs, this is suboptimal for values in xmms, in that case we want vpunpcklwd
+  (UPat(Ops.STACK, dtypes.float16, name="x"), lambda x:
+   vpins(x.replace(src=tuple(s.bitcast(dtypes.int16) for s in x.src)))),
   (UPat(Ops.STACK, dtypes.float32, name="x"), vinsertps),
-  (UPat.var("y", dtypes.ints+(dtypes.bool,)).broadcast(name="x"), vpbroadcast),
   (UPat(Ops.STACK, dtypes.ints+(dtypes.bool,), name="x"), vpins),
   # INDEX on a vector register value extracts a single element
   (UPat.var("y", dtypes.int8s+(dtypes.bool,)).index(UPat.cvar("c"), name="x"),
-   lambda y,c,x: x.ins(X86Ops.VPEXTRB, src=(y, imm(dtypes.uint8, c.arg))) if y.dtype.count > 1 else None),
+   lambda y,c,x: x.ins(X86Ops.VPEXTRB, src=(y, imm(dtypes.uint8, c.arg))) if _is_vec_xmm(y) else None),
   (UPat.var("y", dtypes.int16s).index(UPat.cvar("c"), name="x"),
-   lambda y,c,x: x.ins(X86Ops.VPEXTRW, src=(y, imm(dtypes.uint8, c.arg))) if y.dtype.count > 1 else None),
+   lambda y,c,x: x.ins(X86Ops.VPEXTRW, src=(y, imm(dtypes.uint8, c.arg))) if _is_vec_xmm(y) else None),
   (UPat.var("y", dtypes.int32s).index(UPat.cvar("c"), name="x"),
-   lambda y,c,x: x.ins(X86Ops.VPEXTRD, src=(y, imm(dtypes.uint8, c.arg))) if y.dtype.count > 1 else None),
+   lambda y,c,x: x.ins(X86Ops.VPEXTRD, src=(y, imm(dtypes.uint8, c.arg))) if _is_vec_xmm(y) else None),
   (UPat.var("y", dtypes.int64s).index(UPat.cvar("c"), name="x"),
-   lambda y,c,x: x.ins(X86Ops.VPEXTRQ, src=(y, imm(dtypes.uint8, c.arg))) if y.dtype.count > 1 else None),
+   lambda y,c,x: x.ins(X86Ops.VPEXTRQ, src=(y, imm(dtypes.uint8, c.arg))) if _is_vec_xmm(y) else None),
   (UPat.var("y", dtypes.floats).index(UPat.cvar("c"), name="x"),
-   lambda y,c,x: x.ins(X86Ops.VPSRLDQ, src=(y, imm(dtypes.uint8, c.arg * x.dtype.itemsize))) if y.dtype.count > 1 else None),
-  # fused multiply add
-  ((UPat(Ops.MUL, dtypes.float32, name="a") + UPat.var("b")).named("c"), lambda ctx,a,b,c:
-   a.ins(X86Ops.VFMADD213SS if a.dtype.count == 1 else X86Ops.VFMADD213PS, src=(*a.src, b)) if is_foldable(ctx, c, a) else None),
-  ((UPat(Ops.MUL, dtypes.float64, name="a") + UPat.var("b")).named("c"), lambda ctx,a,b,c:
-   a.ins(X86Ops.VFMADD213SD if a.dtype.count == 1 else X86Ops.VFMADD213PD, src=(*a.src, b)) if is_foldable(ctx, c, a) else None),
+   lambda y,c,x: x.ins(X86Ops.VPSRLDQ, src=(y, imm(dtypes.uint8, c.arg * x.dtype.itemsize))) if _is_vec_xmm(y) else None),
   # packed bitwise
-  ((UPat() & UPat()).named("x"), lambda x: x.ins(X86Ops.VPAND) if x.dtype.count > 1 else None),
-  ((UPat() | UPat()).named("x"), lambda x: x.ins(X86Ops.VPOR) if x.dtype.count > 1 else None),
-  ((UPat() ^ UPat()).named("x"), lambda x: x.ins(X86Ops.VPXOR) if x.dtype.count > 1 else None),
+  ((UPat() & UPat()).named("x"), lambda x: x.ins(X86Ops.VPAND) if x.max_numel() > 1 else None),
+  ((UPat() | UPat()).named("x"), lambda x: x.ins(X86Ops.VPOR) if x.max_numel() > 1 else None),
+  ((UPat() ^ UPat()).named("x"), lambda x: x.ins(X86Ops.VPXOR) if x.max_numel() > 1 else None),
   # packed int binary
-  ((UPat(dtype=dtypes.int32s) << UPat()).named("x"), lambda x: x.ins(X86Ops.VPSLLVD) if x.dtype.count > 1 else None),
-  ((UPat(dtype=dtypes.int64s) << UPat()).named("x"), lambda x: x.ins(X86Ops.VPSLLVQ) if x.dtype.count > 1 else None),
-  ((UPat(dtype=dtypes.uint32) >> UPat()).named("x"), lambda x: x.ins(X86Ops.VPSRLVD) if x.dtype.count > 1 else None),
-  ((UPat(dtype=dtypes.uint64) >> UPat()).named("x"), lambda x: x.ins(X86Ops.VPSRLVQ) if x.dtype.count > 1 else None),
-  ((UPat(dtype=dtypes.int32) >> UPat()).named("x"), lambda x: x.ins(X86Ops.VPSRAVD) if x.dtype.count > 1 else None),
-  ((UPat(dtype=dtypes.int8s) + UPat()).named("x"), lambda x: x.ins(X86Ops.VPADDB) if x.dtype.count > 1 else None),
-  ((UPat(dtype=dtypes.int16s) + UPat()).named("x"), lambda x: x.ins(X86Ops.VPADDW) if x.dtype.count > 1 else None),
-  ((UPat(dtype=dtypes.int32s) + UPat()).named("x"), lambda x: x.ins(X86Ops.VPADDD) if x.dtype.count > 1 else None),
-  ((UPat(dtype=dtypes.int64s) + UPat()).named("x"), lambda x: x.ins(X86Ops.VPADDQ) if x.dtype.count > 1 else None),
-  (UPat(Ops.SUB, dtypes.int8s, name="x"), lambda x: x.ins(X86Ops.VPSUBB) if x.dtype.count > 1 else None),
-  (UPat(Ops.SUB, dtypes.int16s, name="x"), lambda x: x.ins(X86Ops.VPSUBW) if x.dtype.count > 1 else None),
-  (UPat(Ops.SUB, dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VPSUBD) if x.dtype.count > 1 else None),
-  (UPat(Ops.SUB, dtypes.int64s, name="x"), lambda x: x.ins(X86Ops.VPSUBQ) if x.dtype.count > 1 else None),
-  (UPat(Ops.MUL, dtypes.int16s, name="x"), lambda x: x.ins(X86Ops.VPMULLW) if x.dtype.count > 1 else None),
-  (UPat(Ops.MUL, dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VPMULLD) if x.dtype.count > 1 else None),
+  ((UPat(dtype=dtypes.int32s) << UPat()).named("x"), lambda x: x.ins(X86Ops.VPSLLVD) if x.max_numel() > 1 else None),
+  ((UPat(dtype=dtypes.int64s) << UPat()).named("x"), lambda x: x.ins(X86Ops.VPSLLVQ) if x.max_numel() > 1 else None),
+  ((UPat(dtype=dtypes.uint32) >> UPat()).named("x"), lambda x: x.ins(X86Ops.VPSRLVD) if x.max_numel() > 1 else None),
+  ((UPat(dtype=dtypes.uint64) >> UPat()).named("x"), lambda x: x.ins(X86Ops.VPSRLVQ) if x.max_numel() > 1 else None),
+  ((UPat(dtype=dtypes.int32) >> UPat()).named("x"), lambda x: x.ins(X86Ops.VPSRAVD) if x.max_numel() > 1 else None),
+  ((UPat(dtype=dtypes.int8s) + UPat()).named("x"), lambda x: x.ins(X86Ops.VPADDB) if x.max_numel() > 1 else None),
+  ((UPat(dtype=dtypes.int16s) + UPat()).named("x"), lambda x: x.ins(X86Ops.VPADDW) if x.max_numel() > 1 else None),
+  ((UPat(dtype=dtypes.int32s) + UPat()).named("x"), lambda x: x.ins(X86Ops.VPADDD) if x.max_numel() > 1 else None),
+  ((UPat(dtype=dtypes.int64s) + UPat()).named("x"), lambda x: x.ins(X86Ops.VPADDQ) if x.max_numel() > 1 else None),
+  (UPat(Ops.SUB, dtypes.int8s, name="x"), lambda x: x.ins(X86Ops.VPSUBB) if x.max_numel() > 1 else None),
+  (UPat(Ops.SUB, dtypes.int16s, name="x"), lambda x: x.ins(X86Ops.VPSUBW) if x.max_numel() > 1 else None),
+  (UPat(Ops.SUB, dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VPSUBD) if x.max_numel() > 1 else None),
+  (UPat(Ops.SUB, dtypes.int64s, name="x"), lambda x: x.ins(X86Ops.VPSUBQ) if x.max_numel() > 1 else None),
+  (UPat(Ops.MUL, dtypes.int16s, name="x"), lambda x: x.ins(X86Ops.VPMULLW) if x.max_numel() > 1 else None),
+  (UPat(Ops.MUL, dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VPMULLD) if x.max_numel() > 1 else None),
   # scalar int binary
   ((UPat(dtype=dtypes.ints).alu(Ops.CDIV, UPat())).named("x"), idiv),
   # scalar int binary with immediate
@@ -532,21 +462,21 @@ isel_matcher = PatternMatcher([
   (UPat.var("a", dtypes.ints+(dtypes.bool,)) ^ UPat.var("b"), lambda a,b: a.ins(X86Ops.XOR, src=(a, b))),
   (UPat(Ops.SUB, dtypes.ints, (UPat.var("a"), UPat.var("b"))), lambda a,b: a.ins(X86Ops.SUB, src=(a, b))),
   # float binary
-  ((UPat(dtype=dtypes.float32) + UPat()).named("x"), lambda x: x.ins(X86Ops.VADDSS if x.dtype.count == 1 else X86Ops.VADDPS)),
-  ((UPat(dtype=dtypes.float64) + UPat()).named("x"), lambda x: x.ins(X86Ops.VADDSD if x.dtype.count == 1 else X86Ops.VADDPD)),
-  ((UPat(dtype=dtypes.float32) * UPat()).named("x"), lambda x: x.ins(X86Ops.VMULSS if x.dtype.count == 1 else X86Ops.VMULPS)),
-  ((UPat(dtype=dtypes.float64) * UPat()).named("x"), lambda x: x.ins(X86Ops.VMULSD if x.dtype.count == 1 else X86Ops.VMULPD)),
-  (UPat(Ops.SUB, dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VSUBSS if x.dtype.count == 1 else X86Ops.VSUBPS)),
-  (UPat(Ops.SUB, dtypes.float64, name="x"), lambda x: x.ins(X86Ops.VSUBSD if x.dtype.count == 1 else X86Ops.VSUBPD)),
-  (UPat(Ops.FDIV, dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VDIVSS if x.dtype.count == 1 else X86Ops.VDIVPS)),
-  (UPat(Ops.FDIV, dtypes.float64, name="x"), lambda x: x.ins(X86Ops.VDIVSD if x.dtype.count == 1 else X86Ops.VDIVPD)),
+  ((UPat(dtype=dtypes.float32) + UPat()).named("x"), lambda x: x.ins(X86Ops.VADDSS if x.max_numel() == 1 else X86Ops.VADDPS)),
+  ((UPat(dtype=dtypes.float64) + UPat()).named("x"), lambda x: x.ins(X86Ops.VADDSD if x.max_numel() == 1 else X86Ops.VADDPD)),
+  ((UPat(dtype=dtypes.float32) * UPat()).named("x"), lambda x: x.ins(X86Ops.VMULSS if x.max_numel() == 1 else X86Ops.VMULPS)),
+  ((UPat(dtype=dtypes.float64) * UPat()).named("x"), lambda x: x.ins(X86Ops.VMULSD if x.max_numel() == 1 else X86Ops.VMULPD)),
+  (UPat(Ops.SUB, dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VSUBSS if x.max_numel() == 1 else X86Ops.VSUBPS)),
+  (UPat(Ops.SUB, dtypes.float64, name="x"), lambda x: x.ins(X86Ops.VSUBSD if x.max_numel() == 1 else X86Ops.VSUBPD)),
+  (UPat(Ops.FDIV, dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VDIVSS if x.max_numel() == 1 else X86Ops.VDIVPS)),
+  (UPat(Ops.FDIV, dtypes.float64, name="x"), lambda x: x.ins(X86Ops.VDIVSD if x.max_numel() == 1 else X86Ops.VDIVPD)),
   # casts
-  (UPat(dtype=dtypes.int32).cast(dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VCVTDQ2PS) if x.dtype.count > 1 else None),
-  (UPat(dtype=dtypes.int32).cast(dtypes.float64, name="x"), lambda x: x.ins(X86Ops.VCVTDQ2PD) if x.dtype.count > 1 else None),
-  (UPat(dtype=dtypes.float32).cast(dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VCVTTPS2DQ) if x.dtype.count > 1 else None),
-  (UPat(dtype=dtypes.float64).cast(dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VCVTTPD2DQ) if x.dtype.count > 1 else None),
-  (UPat(dtype=dtypes.float32).cast(dtypes.float64, name="x"), lambda x: x.ins(X86Ops.VCVTPS2PD) if x.dtype.count > 1 else None),
-  (UPat(dtype=dtypes.float64).cast(dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VCVTPD2PS) if x.dtype.count > 1 else None),
+  (UPat(dtype=dtypes.int32).cast(dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VCVTDQ2PS) if x.max_numel() > 1 else None),
+  (UPat(dtype=dtypes.int32).cast(dtypes.float64, name="x"), lambda x: x.ins(X86Ops.VCVTDQ2PD) if x.max_numel() > 1 else None),
+  (UPat(dtype=dtypes.float32).cast(dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VCVTTPS2DQ) if x.max_numel() > 1 else None),
+  (UPat(dtype=dtypes.float64).cast(dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VCVTTPD2DQ) if x.max_numel() > 1 else None),
+  (UPat(dtype=dtypes.float32).cast(dtypes.float64, name="x"), lambda x: x.ins(X86Ops.VCVTPS2PD) if x.max_numel() > 1 else None),
+  (UPat(dtype=dtypes.float64).cast(dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VCVTPD2PS) if x.max_numel() > 1 else None),
   (UPat(dtype=dtypes.float32).cast(dtypes.float16, name="x"), lambda x: x.ins(X86Ops.VCVTPS2PH, src=x.src + (imm(dtypes.uint8, 4),))),
   (UPat(dtype=dtypes.float16).cast(dtypes.float32, name="x"), lambda x: x.ins(X86Ops.VCVTPH2PS)),
   (UPat(dtype=dtypes.float32).cast(dtypes.int32s+dtypes.int64s, name="x"), lambda x: x.ins(X86Ops.VCVTTSS2SI)),
@@ -555,9 +485,11 @@ isel_matcher = PatternMatcher([
   (UPat.var("y", dtypes.float64).cast(dtypes.float32, name="x"), lambda y,x: x.ins(X86Ops.VCVTSD2SS, src=(y, y))),
   (UPat.var("y", (dtypes.int32, dtypes.int64)).cast(dtypes.float32, name="x"), lambda y,x: x.ins(X86Ops.VCVTSI2SS, src=(def_reg(x.dtype), y))),
   (UPat.var("y", (dtypes.int32, dtypes.int64)).cast(dtypes.float64, name="x"), lambda y,x: x.ins(X86Ops.VCVTSI2SD, src=(def_reg(x.dtype), y))),
-  (UPat(dtype=dtypes.uints+(dtypes.bool,)).cast(dtypes.ints, name="x"), lambda x: x.ins(X86Ops.MOVZX) if x.dtype.count == 1 else None),
-  (UPat(dtype=dtypes.int32).cast(dtypes.int64s, name="x"), lambda x: x.ins(X86Ops.MOVSXD) if x.dtype.count == 1 else None),
-  (UPat(dtype=dtypes.sints).cast(dtypes.ints, name="x"), lambda x: x.ins(X86Ops.MOVSX) if x.dtype.count == 1 else None),
+  (UPat(dtype=dtypes.uints+(dtypes.bool,)).cast(dtypes.ints, name="x"), lambda x:
+   x.ins(X86Ops.MOVZX) if x.max_numel() == 1 and x.src[0].dtype.itemsize < x.dtype.itemsize else None),
+  (UPat(dtype=dtypes.int32).cast(dtypes.int64s, name="x"), lambda x: x.ins(X86Ops.MOVSXD) if x.max_numel() == 1 else None),
+  (UPat(dtype=dtypes.sints).cast(dtypes.ints, name="x"), lambda x:
+   x.ins(X86Ops.MOVSX) if x.max_numel() == 1 and x.src[0].dtype.itemsize < x.dtype.itemsize else None),
   (UPat(dtype=(dtypes.uint8, dtypes.bool)).cast(dtypes.int16s, name="x"), lambda x: x.ins(X86Ops.VPMOVZXBW)),
   (UPat(dtype=(dtypes.uint8, dtypes.bool)).cast(dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VPMOVZXBD)),
   (UPat(dtype=(dtypes.uint8, dtypes.bool)).cast(dtypes.int64s, name="x"), lambda x: x.ins(X86Ops.VPMOVZXBQ)),
@@ -570,7 +502,7 @@ isel_matcher = PatternMatcher([
   (UPat(dtype=dtypes.int16).cast(dtypes.int32s, name="x"), lambda x: x.ins(X86Ops.VPMOVSXWD)),
   (UPat(dtype=dtypes.int16).cast(dtypes.int64s, name="x"), lambda x: x.ins(X86Ops.VPMOVSXWQ)),
   (UPat(dtype=dtypes.int32).cast(dtypes.int64s, name="x"), lambda x: x.ins(X86Ops.VPMOVSXDQ)),
-  # bitcasts
+  # bitcasts between scalar floats and ints
   (UPat.var("y", dtypes.float16).bitcast(dtypes.int16s).named("x"), lambda y,x: x.ins(X86Ops.VPEXTRW, src=(y, imm(dtypes.uint8, 0)))),
   (UPat(dtype=dtypes.int16s).bitcast(dtypes.float16).named("x"), vpins),
   (UPat(dtype=dtypes.int32s).bitcast(dtypes.float32).named("x"), lambda x: x.ins(X86Ops.VMOVD)),
@@ -579,34 +511,26 @@ isel_matcher = PatternMatcher([
   (UPat(dtype=dtypes.float64).bitcast(dtypes.int64s).named("x"), lambda x: x.ins(X86Ops.VMOVQm)),
   # index on a buffer (or the stack pointer) computes an address, addresses are 64bit values
   (UPat((Ops.INDEX, Ops.SHRINK), name="x"),
-   lambda x: x.ins(X86Ops.LEA, dtype=dtypes.uint64, src=fold_address(x)) if x.src[0].dtype.count == 1 else None),
+   lambda x: x.ins(X86Ops.LEA, dtype=dtypes.uint64, src=fold_address(x)) if not _is_vec_xmm(x.src[0]) else None),
   # TODO: fuse stores, very few cases -- store cmp becomes setcc, store gep int becomes vpextr, store bitcast to int becomes vmovd/q
   # copy, load, store
   # NOTE: copy here violates the spec, it only happens post register allocation when a reg to reg move needs to be inserted
-  (UPat(Ops.COPY, dt_128bit, name="x"), lambda x: x.ins(X86Ops.VMOVUPS)),
-  (UPat(Ops.COPY, dt_64bit, name="x"), lambda x: x.ins(X86Ops.VMOVSD)),
-  (UPat(Ops.COPY, dt_32bit+dt_16bit, name="x"), lambda x: x.ins(X86Ops.VMOVSS)),
-  (UPat(Ops.COPY, dtypes.ints+(dtypes.bool,), name="x"), lambda x: x.ins(X86Ops.MOV)),
-  (UPat(Ops.LOAD, dt_128bit, src=(UPat(name="a"),), name="x"), lambda x,a: x.ins(X86Ops.VMOVUPS, src=fold_address(a))),
-  (UPat(Ops.LOAD, dt_64bit, src=(UPat(name="a"),), name="x"), lambda x,a: x.ins(X86Ops.VMOVSD, src=fold_address(a))),
-  (UPat(Ops.LOAD, dt_32bit, src=(UPat(name="a"),), name="x"), lambda x,a: x.ins(X86Ops.VMOVSS, src=fold_address(a))),
-  (UPat(Ops.LOAD, dt_16bit, src=(UPat(name="a"),), name="x"), lambda x,a:
-   x.ins(X86Ops.VPINSRW, src=(def_reg(x.dtype, x.tag),) + fold_address(a) + (imm(dtypes.uint8, 0),))),
-  (UPat(Ops.LOAD, dtypes.ints+(dtypes.bool,), src=(UPat(name="a"),), name="x"), lambda x,a: x.ins(X86Ops.MOV, src=fold_address(a))),
-  (UPat.var("a").store(UPat.var("b", dt_128bit), name="x"), lambda a,b,x: x.ins(X86Ops.VMOVUPSm, src=fold_address(a) + (b,))),
-  (UPat.var("a").store(UPat.var("b", dt_64bit), name="x"), lambda a,b,x: x.ins(X86Ops.VMOVSDm, src=fold_address(a) + (b,))),
-  (UPat.var("a").store(UPat.var("b", dt_32bit), name="x"), lambda a,b,x: x.ins(X86Ops.VMOVSSm, src=fold_address(a) + (b,))),
-  (UPat.var("a").store(UPat.var("b", dt_16bit), name="x"), lambda a,b,x: x.ins(X86Ops.VPEXTRW, src=fold_address(a) + (b, imm(dtypes.uint8, 0)))),
+  (UPat(Ops.COPY, dtypes.floats, name="x"), lambda x: x.ins(_xmm_sz(x))),
+  (UPat(Ops.COPY, dtypes.ints+(dtypes.bool,), name="x"), lambda x: x.ins(X86Ops.MOV) if x.max_numel() == 1 else x.ins(_xmm_sz(x))),
+  (UPat(Ops.LOAD, dtypes.floats, src=(UPat(name="a"),), name="x"), lambda x,a:
+   x.ins(X86Ops.VPINSRW, src=(def_reg(x.dtype, x.tag),) + fold_address(a) + (imm(dtypes.uint8, 0),)) if x.max_numel() * x.dtype.itemsize == 2 else
+   x.ins(_xmm_sz(x), src=fold_address(a))),
+  (UPat(Ops.LOAD, dtypes.ints+(dtypes.bool,), src=(UPat(name="a"),), name="x"), lambda x,a:
+   x.ins(X86Ops.MOV, src=fold_address(a)) if x.max_numel() == 1 else
+   x.ins(X86Ops.VPINSRW, src=(def_reg(x.dtype, x.tag),) + fold_address(a) + (imm(dtypes.uint8, 0),)) if x.max_numel() * x.dtype.itemsize == 2 else
+   x.ins(_xmm_sz(x), src=fold_address(a))),
+  (UPat.var("a").store(UPat.var("b", dtypes.floats), name="x"), lambda a,b,x:
+   x.ins(X86Ops.VPEXTRW, src=fold_address(a) + (b, imm(dtypes.uint8, 0))) if b.max_numel() * b.dtype.itemsize == 2 else
+   x.ins(_xmm_sz_m(b), src=fold_address(a) + (b,))),
   (UPat.var("a").store(UPat.var("b", dtypes.ints+(dtypes.bool,)), name="x"), lambda a,b,x:
+   x.ins(X86Ops.VPEXTRW, src=fold_address(a) + (b, imm(dtypes.uint8, 0))) if b.max_numel() > 1 and b.max_numel() * b.dtype.itemsize == 2 else
+   x.ins(_xmm_sz_m(b), src=fold_address(a) + (b,)) if b.max_numel() > 1 else
    x.ins(X86Ops.MOVm, src=fold_address(a) + (b,)) if (i:=to_imm(b)) is None else x.ins(X86Ops.MOVi, src=fold_address(a) + (i,))),
-  # **** X86Op -> X86Op ****
-  # fold loads into X86Ops that allow it, if beneficial
-  (UPat(Ops.INS, src=(UPat(Ops.LOAD, src=(UPat(name="a"),), name="y"),), allow_any_len=True, name="x"), lambda ctx,y,a,x:
-   x.replace(src=fold_address(a) + x.src[1:]) if x.arg in X86GroupOp.ReadMem1st and is_foldable(ctx, x, y) else None),
-  (UPat(Ops.INS, src=(UPat(), UPat(Ops.LOAD, src=(UPat(name="a"),), name="y")), allow_any_len=True, name="x"), lambda ctx,y,a,x:
-   x.replace(src=x.src[:1] + fold_address(a) + x.src[2:]) if x.arg in X86GroupOp.ReadMem2nd and is_foldable(ctx, x, y) else None),
-  (UPat(Ops.INS, src=(UPat(), UPat(), UPat(Ops.LOAD, src=(UPat(name="a"),), name="y")), allow_any_len=True, name="x"), lambda ctx,y,a,x:
-   x.replace(src=x.src[:2] + fold_address(a) + x.src[3:]) if x.arg in X86GroupOp.ReadMem3rd and is_foldable(ctx, x, y) else None),
   # allocate virtual registers
   (UPat((Ops.INS, Ops.BUFFER), name="x"), alloc_vregs),
 ])
@@ -825,24 +749,15 @@ encodings = {
   X86Ops.VDIVSD: lambda x: encode(x, 0x5E, pp=3, sel=1), X86Ops.VDIVPD: lambda x: encode(x, 0x5E, pp=1, sel=1),
   X86Ops.VCMPSS: lambda x: encode(x, 0xC2, pp=2, sel=1), X86Ops.VCMPPS: lambda x: encode(x, 0xC2, pp=0, sel=1),
   X86Ops.VCMPSD: lambda x: encode(x, 0xC2, pp=3, sel=1), X86Ops.VCMPPD: lambda x: encode(x, 0xC2, pp=1, sel=1),
-  X86Ops.VMAXSS: lambda x: encode(x, 0x5F, pp=2, sel=1), X86Ops.VMAXPS: lambda x: encode(x, 0x5F, pp=0, sel=1),
-  X86Ops.VMAXSD: lambda x: encode(x, 0x5F, pp=3, sel=1), X86Ops.VMAXPD: lambda x: encode(x, 0x5F, pp=1, sel=1),
-  X86Ops.VMINSS: lambda x: encode(x, 0x5D, pp=2, sel=1), X86Ops.VMINPS: lambda x: encode(x, 0x5D, pp=0, sel=1),
-  X86Ops.VMINSD: lambda x: encode(x, 0x5D, pp=3, sel=1), X86Ops.VMINPD: lambda x: encode(x, 0x5D, pp=1, sel=1),
   # ternary
   X86Ops.CMOVB: lambda x: encode(x, 0x0F42), X86Ops.CMOVL: lambda x: encode(x, 0x0F4C),
   X86Ops.CMOVE: lambda x: encode(x, 0x0F44), X86Ops.CMOVNE: lambda x: encode(x, 0x0F45),
-  X86Ops.VFMADD213SS: lambda x: encode(x, 0xA9, pp=1, sel=2), X86Ops.VFMADD213SD: lambda x: encode(x, 0xA9, pp=1, sel=2, we=1),
-  X86Ops.VFMADD213PS: lambda x: encode(x, 0xA8, pp=1, sel=2), X86Ops.VFMADD213PD: lambda x: encode(x, 0xA8, pp=1, sel=2, we=1),
   X86Ops.VBLENDVPS: lambda x: encode(x, 0x4A, pp=1, sel=3), X86Ops.VBLENDVPD: lambda x: encode(x, 0x4B, pp=1, sel=3),
   X86Ops.VPBLENDVB: lambda x: encode(x, 0x4C, pp=1, sel=3),
   # shuffles
-  X86Ops.VPBROADCASTB: lambda x: encode(x, 0x78, pp=1, sel=2), X86Ops.VPBROADCASTW: lambda x: encode(x, 0x79, pp=1, sel=2),
-  X86Ops.VPBROADCASTD: lambda x: encode(x, 0x58, pp=1, sel=2), X86Ops.VPBROADCASTQ: lambda x: encode(x, 0x59, pp=1, sel=2),
-  X86Ops.VBROADCASTSS: lambda x: encode(x, 0x18, pp=1, sel=2), X86Ops.VPSRLDQ: lambda x: encode(x, 0x73, reg=3, pp=1, sel=1),
+  X86Ops.VPSRLDQ: lambda x: encode(x, 0x73, reg=3, pp=1, sel=1),
   X86Ops.VPINSRB: lambda x: encode(x, 0x20, pp=1, sel=3), X86Ops.VPINSRW: lambda x: encode(x, 0xC4, pp=1, sel=1),
   X86Ops.VPINSRD: lambda x: encode(x, 0x22, pp=1, sel=3), X86Ops.VPINSRQ: lambda x: encode(x, 0x22, pp=1, sel=3, we=1),
-  X86Ops.VSHUFPS: lambda x: encode(x, 0xC6, pp=0, sel=1), X86Ops.VSHUFPD: lambda x: encode(x, 0xC6, pp=1, sel=1),
   X86Ops.VINSERTPS: lambda x: encode(x, 0x21, pp=1, sel=3),
   # extract
   X86Ops.VPEXTRB: lambda x: encode(x, 0x14, pp=1, sel=3), X86Ops.VPEXTRW: lambda x: encode(x, 0x15, pp=1, sel=3),
@@ -876,20 +791,21 @@ class X86Renderer(ISARenderer):
   def stack_pointer(self) -> UOp: return def_reg(dtypes.uint64, RSP)
   # the value of a BUFFER is its address, it moves through registers and the stack as a 64bit int
   def copy(self, x:UOp, reg:Register):
-    ret = isel_matcher.rewrite(UOp(Ops.COPY, dtypes.uint64 if x.op is Ops.BUFFER else x.dtype, (x,), tag=reg))
-    assert ret is not None
+    dt = dtypes.uint64 if x.op is Ops.BUFFER else x.dtype
+    ret = isel_matcher.rewrite(UOp(Ops.COPY, dt, (x,), tag=reg))
+    assert ret is not None, f"failed to copy {x}"
     return ret
 
   def spill(self, disp:UOp, x:UOp) -> UOp:
     if x.op is Ops.BUFFER: x = x.replace(dtype=dtypes.uint64)
-    ret = isel_matcher.rewrite(self.stack_pointer().index(disp).store(x))
-    assert ret is not None
-    return ret
+    is_xmm = isinstance(x.tag, tuple) and x.tag[0].cons[0].size == 16
+    op = X86Ops.VMOVUPSm if is_xmm else X86Ops.MOVm
+    return UOp(Ops.INS, dtypes.void, fold_address(self.stack_pointer().index(disp)) + (x,), op, x.tag)
 
   def fill(self, disp:UOp, x:UOp, reg:Register) -> UOp:
-    ret = isel_matcher.rewrite(self.stack_pointer().index(disp).load(dtype=dtypes.uint64 if x.op is Ops.BUFFER else x.dtype, tag=reg))
-    assert ret is not None
-    return ret
+    is_xmm = reg.cons[0].size == 16
+    dt = dtypes.uint64 if x.op is Ops.BUFFER else x.dtype
+    return UOp(Ops.INS, dt, fold_address(self.stack_pointer().index(disp)), X86Ops.VMOVUPS if is_xmm else X86Ops.MOV, (reg,))
 
   def asm_str(self, uops:list[UOp], function_name:str) -> str:
     def _format_op(x:UOp) -> str: return f"    {(o[7:-1] if (o:=str(x.arg))[-1] in ('i', 'm') else o[7:]).lower():7s}"