tinygrad 0.8.0__py3-none-any.whl → 0.9.1__py3-none-any.whl

tinygrad/runtime/ops_python.py

@@ -0,0 +1,208 @@
+# pylint: disable=cell-var-from-loop
+# a python uops emulator
+# works to test the tensor cores, and all the uops in general
+# this is the (living) definition of uops
+from typing import Tuple, List, Optional, Any, Dict
+import pickle, base64, itertools, time, struct
+from tinygrad.dtype import DType, dtypes, ImageDType
+from tinygrad.helpers import all_same, getenv, flatten
+from tinygrad.device import Compiled, Compiler, Allocator
+from tinygrad.codegen.uops import UOpGraph, UOps
+from tinygrad.ops import BinaryOps, TernaryOps, exec_alu, truncate
+from tinygrad.renderer import Renderer
+from tinygrad.renderer.cstyle import CUDARenderer, MetalRenderer, AMDRenderer
+
+def _load(m, i):
+  if i < 0 or i >= len(m): raise IndexError(f"load out of bounds, size is {len(m)} and access is {i}")
+  return m[i]
+
+def load(inp, j=0):
+  if len(inp) == 4: return [_load(m, x+j) if gate else default for m,x,gate,default in zip(*inp)]
+  return [_load(m, x+j) for m,x in zip(inp[0], inp[1])]
+
+def _store(m, i, v):
+  if i < 0 or i >= len(m): raise IndexError(f"store out of bounds, size is {len(m)}, access is {i}, value is {v}")
+  m[i] = v
+
+class PythonProgram:
+  def __init__(self, name:str, lib:bytes):
+    self.uops: List[Tuple[UOps, Optional[DType], List[int], Any]] = pickle.loads(lib)
+  def __call__(self, *bufs, global_size:Tuple[int,int,int]=(1,1,1), local_size:Tuple[int,int,int]=(1,1,1), vals:Tuple[int, ...]=(), wait=False):
+    st = time.perf_counter()
+    warp = list(itertools.product(*[range(x) for x in local_size[::-1]]))
+    warp_size = len(warp)
+    for idxs in itertools.product(*[range(x) for x in global_size[::-1]]):
+      ul: Dict[int, Any] = {}
+      dl: Dict[int, DType] = {}
+      pbufs: List[memoryview] = list(bufs)
+      pvals: List[int] = list(vals)
+      i = 0
+      loop_ends: Dict[int, int] = {}
+      while i < len(self.uops):
+        uop, dtype, idp, arg = self.uops[i]
+        void_ops = {UOps.STORE, UOps.ENDRANGE, UOps.BARRIER, UOps.IF, UOps.ENDIF}
+        if uop is UOps.DEFINE_ACC: idp = [idp[0]]
+        inp = [ul[v] for v in idp if self.uops[v][0] not in void_ops]
+        dtp = [dl[v] for v in idp if self.uops[v][0] not in void_ops]
+        if getenv("TRACE"): print(i, uop, dtype, arg, inp, dtp)
+        if uop is UOps.STORE:
+          if len(inp) == 3: inp.append([True] * len(inp[0]))  # set the gate to True
+          if isinstance(dtp[0], ImageDType):
+            # image store
+            assert dtp[2].count == 4
+            for j,val in enumerate(inp[2]):
+              for m,ox,oy,v,g in zip(inp[0], inp[1][0], inp[1][1], val, inp[3]):
+                assert ox >= 0 and ox < dtp[0].shape[1] and oy >= 0 and oy < dtp[0].shape[0]
+                if g: _store(m, ox*4 + oy*dtp[0].shape[1]*4 + j, v)
+          elif dtp[2].count > 1:
+            for j,val in enumerate(inp[2]):
+              for m,o,v,g in zip(inp[0], inp[1], val, inp[3]):
+                if g: _store(m, o+j, v)
+          else:
+            for m,o,v,g in zip(*inp):
+              if g: _store(m, o, v)
+          i += 1
+          continue
+        if uop is UOps.ENDRANGE:
+          loop_ends[idp[0]] = i
+          i = idp[0]
+          continue
+        if uop in (UOps.BARRIER, UOps.IF, UOps.ENDIF):
+          # in the python emulator, the warp is always in sync
+          i += 1
+          continue
+        assert dtype is not None, f"{uop} is missing a dtype"
+        dl[i] = dtype
+        if uop is UOps.DEFINE_GLOBAL:
+          assert dtype.fmt is not None
+          ul[i] = [pbufs.pop(0).cast(dtype.fmt)] * warp_size
+        elif uop is UOps.DEFINE_LOCAL:
+          assert dtype.fmt is not None
+          lbuf = memoryview(bytearray(arg[1]*dtype.itemsize))
+          ul[i] = [lbuf.cast(dtype.fmt)] * warp_size
+        elif uop is UOps.DEFINE_VAR:
+          ul[i] = [pvals.pop(0)] * warp_size
+        elif uop is UOps.SPECIAL:
+          if arg[1][0] == 'g':
+            ul[i] = [idxs[2-arg[0]]] * warp_size
+          elif arg[1][0] == 'l':
+            ul[i] = [x[2-arg[0]] for x in warp]
+        elif uop is UOps.CONST:
+          ul[i] = [[arg] * warp_size for _ in range(dtype.count)] if dtype.count > 1 else [arg] * warp_size
+        elif uop is UOps.DEFINE_ACC:
+          ul[i] = [[inp[0][0]] * warp_size for _ in range(dtype.count)] if dtype.count > 1 else [inp[0][0]] * warp_size
+        elif uop is UOps.RANGE:
+          if i not in ul: ul[i] = [inp[0][0]] * warp_size
+          else:
+            for j in range(len(ul[i])):
+              ul[i][j] += 1
+            if ul[i][0] == inp[1][0]:
+              del ul[i]
+              i = loop_ends[i] + 1
+              continue
+        elif uop in (UOps.CAST, UOps.BITCAST):
+          if dtype.count > 1: ul[i] = inp
+          else:
+            assert dtp[0].fmt and dtype.fmt
+            pack_format, unpack_format = str(warp_size) + dtp[0].fmt, str(warp_size) + dtype.fmt
+            if uop is UOps.BITCAST: ul[i] = list(struct.unpack(unpack_format, struct.pack(pack_format, *inp[0])))
+            else:
+              casted = [dtypes.as_const(x, dtype) for x in inp[0]]
+              if dtypes.is_int(dtype):
+                overflow_adjust = 2**(dtype.itemsize*8 - 1) if not dtypes.is_unsigned(dtype) else 0
+                casted = [((x + overflow_adjust) % 2**(dtype.itemsize*8) - overflow_adjust) for x in casted]
+              elif dtypes.is_float(dtype):
+                casted = [truncate.get(dtype, lambda dt: dt)(x) for x in casted]
+              ul[i] = list(struct.unpack(unpack_format, struct.pack(unpack_format, *casted)))
+        elif uop is UOps.LOAD:
+          if isinstance(dtp[0], ImageDType):
+            assert dtype.count == 4
+            ul[i] = []
+            for j in range(dtype.count):
+              ret = []
+              for m,ox,oy in zip(inp[0], inp[1][0], inp[1][1]):
+                if ox < 0 or ox >= dtp[0].shape[1] or oy < 0 or oy >= dtp[0].shape[0]: ret.append(0)
+                else: ret.append(_load(m, ox*4 + oy*dtp[0].shape[1]*4 + j))
+              ul[i].append(ret)
+          elif dtype.count > 1:
+            ul[i] = [load([inp[i][j] if dtp[i].count > 1 else inp[i] for i in range(len(inp))], j) for j in range(dtype.count)]
+          else:
+            ul[i] = load(inp)
+        elif uop is UOps.PHI:
+          for j in range(len(inp[0])): inp[0][j] = inp[1][j]
+          ul[i] = inp[0]
+        elif uop is UOps.GEP:
+          ul[i] = inp[0][arg]
+        elif uop is UOps.WMMA:
+          # here are the models for the WMMA instruction on the different hardware
+          def wmma_helper(WARP_THREADS, K, NUM_A, NUM_B, NUM_C, a_elem, b_elem, c_map):
+            assert len(inp[0]) == NUM_A, f"A must have {NUM_A} elements per thread"
+            assert len(inp[1]) == NUM_B, f"B must have {NUM_B} elements per thread"
+            assert len(inp[2]) == NUM_C, f"C must have {NUM_C} elements per thread"
+            assert len(flatten(inp[0])) == NUM_A * warp_size, f"WMMA must have {NUM_A * warp_size} total elements for A in WMMA"
+            assert len(flatten(inp[1])) == NUM_B * warp_size, f"WMMA must have {NUM_B * warp_size} total elements for B in WMMA"
+            assert len(flatten(inp[2])) == NUM_C * warp_size, f"WMMA must have {NUM_C * warp_size} total elements for C in WMMA"
+            assert warp_size > 0 and warp_size % WARP_THREADS == 0, f"must have multiples of {WARP_THREADS} warp threads"
+            out = [inp[2][elem_idx][:] for elem_idx in range(NUM_C)]
+            for goff in range(0, warp_size, WARP_THREADS):
+              for lane_id in range(WARP_THREADS):
+                for elem_idx in range(NUM_C): # calculate new muls and add to acc
+                  (c_i, c_j) = c_map(lane_id, elem_idx)
+                  out[elem_idx][goff+lane_id] += sum(a_elem(inp[0], _k, c_j, goff) * b_elem(inp[1], c_i, _k, goff) for _k in range(K))
+            return out
+
+          # TODO: refactor these to a shared TensorCoreLayout in kernel.py
+          if arg[5] == "METAL":
+            # A (2 elements on 32 threads): row major
+            def a_b_elem(x, i, j, goff): return x[(i%2)][goff+(i//2)%2+(j%4)*2+(i//4)*8+(j//4)*16]
+            # (i, j), C, D (2 elements on 32 threads): row major same as A/B
+            def c_map(lane, elem): return (elem + ((lane%2)*2) + ((lane//8)%2)*4, ((lane//2)%4) + (lane//16)*4)
+            ul[i] = wmma_helper(32, 8, 2, 2, 2, a_b_elem, a_b_elem, c_map)
+          elif arg[5] == "AMD":
+            # A (16 elements on 32 threads): col major, lane 16-32 == lane 0-15
+            def a_elem(x, i, j, goff):
+              assert x[i][goff+j] == x[i][goff+j+16], "warp elements not duplicated properly across lanes"
+              return x[i][goff+j]
+            # B (16 elements on 32 threads): row major, lane 16-32 == lane 0-15
+            def b_elem(x, i, j, goff): return a_elem(x, j, i, goff)  # pylint: disable=arguments-out-of-order
+            def c_map(lane, elem): return (lane%16, lane//16+elem*2) # (i, j), C, D (8 elements on 32 threads): row major
+            ul[i] = wmma_helper(32, 16, 16, 16, 8, a_elem, b_elem, c_map)
+          elif arg[5] == "CUDA":
+            # A (8 elements on 32 threads)
+            def a_elem(x, i, j, goff): return x[(i%2)+(j//8)*2+(i//8)*4][goff+((i//2)%4)+(j%8)*4]
+            # B (4 elements on 32 threads)
+            def b_elem(x, i, j, goff): return x[(j%2)+(j//8)*2][goff+(j//2)%4+(i)*4]
+            # (i, j), C, D (4 elements on 32 threads)
+            def c_map(lane, elem): return ((elem%2)+(lane%4)*2, (lane//4)+(elem//2)*8)
+            ul[i] = wmma_helper(32, 16, 8, 4, 4, a_elem, b_elem, c_map)
+          else: raise NotImplementedError(f"unimplemented tensor core {arg}")
+        elif uop is UOps.ALU:
+          assert all_same([len(x) for x in inp]), f"{[len(x) for x in inp]} doesn't match on {arg}"
+          assert all_same([dtype] + dtp) or arg in {BinaryOps.CMPNE, BinaryOps.CMPLT, TernaryOps.WHERE}, f"dtype mismatch on {arg}"
+          ul[i] = [exec_alu(arg, dtype, p) for p in zip(*inp)]
+        assert i in ul, (uop, dtype, idp, arg)
+        i += 1
+    return time.perf_counter() - st
+
+class PythonRenderer(Renderer):
+  device = "PYTHON"
+  def __init__(self):
+    if getenv("EMULATE_METAL"): self.device, self.tensor_cores = "METAL", MetalRenderer.tensor_cores
+    if getenv("EMULATE_AMD"): self.device, self.tensor_cores = "AMD", AMDRenderer.tensor_cores
+    if getenv("EMULATE_CUDA"): self.device, self.tensor_cores = "CUDA", CUDARenderer.tensor_cores
+
+  def render(self, name:str, uops:UOpGraph) -> str:
+    lops = [(u.op, u.dtype, [uops.uops.index(v) for v in u.src], u.arg) for u in uops]
+    return base64.b64encode(pickle.dumps(lops)).decode()
+
+class PythonCompiler(Compiler):
+  def compile(self, src:str) -> bytes: return base64.b64decode(src)
+
+class PythonAllocator(Allocator):
+  def _alloc(self, size, options): return memoryview(bytearray(size))
+  def copyin(self, dest, src:memoryview): dest[:] = src
+  def copyout(self, dest:memoryview, src): dest[:] = src
+
+class PythonDevice(Compiled):
+  def __init__(self, device:str):
+    super().__init__(device, PythonAllocator(), PythonRenderer(), PythonCompiler(), PythonProgram)

tinygrad/shape/shapetracker.py

@@ -1,70 +1,35 @@
 # ShapeTracker allows movement operations to a buffer that don't require a copy to be made.
 from __future__ import annotations
-import functools, itertools, operator
 from dataclasses import dataclass
-from typing import Tuple, List, Optional, Dict, Set, cast, Union, Iterable
-from tinygrad.ops import MovementOps
-from tinygrad.helpers import prod, DEBUG, merge_dicts, getenv
-from tinygrad.shape.symbolic import Variable, MulNode, Node, SumNode, NumNode, sint
-from tinygrad.shape.view import View, _merge_dims
-
-def expr_node_mask(view:View, idx:Node, valid:Optional[Node]=None) -> Node:
-  expr = [valid] if valid is not None else []
-  if view.mask is not None:
-    acc = 1
-    for d,(x,y) in zip(reversed(view.shape), reversed(view.mask)):
-      if (x,y) != (0,d):
-        base = ((idx//acc)%d)
-        expr += [base >= x, base < y]
-      acc *= d
-  return Node.ands(expr)
-
-# generate an expression if you have a single idx variable
-def expr_node(view:View, idx:Optional[Node]=None) -> Node:
-  if idx is None: idx = Variable('idx', 0, prod(view.shape)-1)
-  ret: List[Node] = [NumNode(view.offset) if isinstance(view.offset, int) else view.offset] if view.offset else []
-  acc = 1
-  for d,s,_ in reversed(_merge_dims(view.shape, view.strides)):
-    ret.append(((idx//acc)%d)*s)
-    acc *= d
-  return Node.sum(ret)
-
-# generate an expression if you have a variable or expression for each index
-def expr_idxs(view:View, idxs:Tuple[Node, ...]) -> Node:
+from typing import Tuple, List, Optional, Dict, Set, Iterable, cast
+from tinygrad.helpers import merge_dicts, getenv
+from tinygrad.shape.symbolic import Variable, MulNode, Node, SumNode, NumNode, create_lt_node, create_ge_node, sint
+from tinygrad.shape.view import View, strides_for_shape
+
+def _expr_view(view:View, idxs:List[Node], valid:Optional[Node]=None) -> Tuple[Node, Node]:
   assert len(idxs) == len(view.shape), f"need an idx for all dimensions {idxs} vs {view.shape}"
-  return Node.sum([NumNode(view.offset) if isinstance(view.offset, int) else view.offset] + [idx*st for idx,sh,st in zip(idxs, view.shape, view.strides) if sh != 1 and st != 0])  # noqa: E501
-
-@functools.lru_cache(maxsize=None)
-def merge_views(vm2:View, vm1:View) -> Optional[View]:
-  if vm1.contiguous and vm1.shape == vm2.shape: return vm2
-  if vm2.contiguous: return vm1
-  if vm2.mask or vm1.offset != 0: return None  # this isn't supported yet
-  if None in (strides := ShapeTracker((vm2, vm1)).real_strides()): return None
-  return View.create(vm1.shape, cast(Tuple[sint, ...], strides), vm2.offset, vm1.mask)
-
-@functools.lru_cache(maxsize=None)
-def idxs_to_idx(shape:Tuple[int, ...], idxs:Tuple[Node, ...]) -> Node:
-  assert len(idxs) == len(shape), "need an idx for all dimensions"
-  acc, ret = 1, []
-  for tidx,d in zip(reversed(idxs), reversed(shape)):
-    ret.append(tidx * acc)
-    acc *= d
-  return Node.sum(ret)
+  iexpr: List[Node] = [NumNode(view.offset) if isinstance(view.offset, int) else view.offset]
+  vexpr: List[Node] = [valid] if valid is not None else []
+  for idx,sh,st,m in zip(idxs, view.shape, view.strides, view.mask if view.mask is not None else [None]*len(view.shape)):
+    if sh != 1 and st != 0: iexpr.append(idx*st)
+    if m is not None: vexpr += [create_ge_node(idx, m[0]), create_lt_node(idx, m[1])]  # idx >= m[0], idx < m[1]
+  return Node.sum(iexpr), Node.ands(vexpr)
 
 @dataclass(frozen=True)
 class ShapeTracker:
   views: Tuple[View, ...]
-  def __post_init__(self):
-    assert isinstance(self.views, tuple) and all(isinstance(v, View) for v in self.views), "ShapeTracker must be created with a tuple of Views"
 
   def __add__(self, st:ShapeTracker) -> ShapeTracker:
-    base = ShapeTracker(self.views)
-    for v in st.views: base = ShapeTracker(base.views + (v,)).simplify() # one view at a time = better simplification
-    return base
+    ret = self
+    for v in st.views: ret = ShapeTracker(ret.views + (v,)).simplify() # one view at a time = better simplification
+    return ret
 
   def invert(self, out_shape:Tuple[sint, ...]) -> Optional[ShapeTracker]:
-    ret = tuple(v.invert(s) for v,s in zip(self.views[::-1], [x.shape for x in self.views[::-1][1:]]+[out_shape]))
-    return ShapeTracker(cast(Tuple[View, ...], ret)).reshape(out_shape) if all(x is not None for x in ret) else None
+    inverted_views:List[View] = []
+    for v,s in zip(self.views[::-1], [x.shape for x in self.views[::-1][1:]]+[out_shape]):
+      if (inverted:= v.invert(s)) is None: return None
+      inverted_views.append(inverted)
+    return ShapeTracker(tuple(inverted_views)).reshape(out_shape)
 
   @staticmethod
   def from_shape(shape:Tuple[sint, ...]): return ShapeTracker((View.create(shape),))
@@ -72,16 +37,22 @@ class ShapeTracker:
   @property
   def contiguous(self) -> bool: return len(self.views) == 1 and self.views[0].contiguous
 
+  @property
+  def consecutive(self) -> bool: return len(self.views) == 1 and (v:=self.views[0]).mask is None and v.strides == strides_for_shape(v.shape)
+
   @property
   def shape(self) -> Tuple[sint, ...]: return self.views[-1].shape
 
   @property
-  def size(self) -> int: return prod([x.max if isinstance(x, Node) else x for x in self.views[-1].shape])
+  def size(self) -> int: return self.views[-1].size()
 
   def real_size(self) -> int:
     if 0 in self.shape: return 0
-    ret = self.expr_idxs()[0].max
-    while not isinstance(ret, int): ret = ret.max    # TODO: this is a while loop?!? it should be more clear what max does
+    idx, valid = self.expr_idxs()
+    if not valid: return 0
+    # TODO: it's possible that the real_size is smaller condition on valid being true
+    ret = idx.max
+    if not isinstance(ret, int): ret = ret.max  # might be represent by symbolic shape, take one more max for int max
     assert isinstance(ret, int), f"ret must be integer, {ret=} isn't"
     return ret+1
 
@@ -90,30 +61,9 @@ class ShapeTracker:
   @property
   def var_vals(self) -> Dict[Variable, int]: return merge_dicts([dict([v.unbind()]) for v in self.vars()])
 
-  def unbind(self) -> ShapeTracker: return ShapeTracker(tuple(v.unbind() for v in self.views))
-
-  def to_movement_ops(self) -> List[Tuple[MovementOps, Tuple]]:
-    to_apply:List[Tuple[MovementOps, Tuple]] = []
-    for v in self.views:
-      real_shape = tuple(y-x for x,y in v.mask) if v.mask else v.shape
-      real_offset = 0 if 0 in real_shape else (v.offset + (sum(x*st for (x,_),st in zip(v.mask, v.strides)) if v.mask else 0))
-      # first, we apply the offset
-      # then, we make it the correct shape
-      # then, we apply permutations
-      to_apply.append((MovementOps.AS_STRIDED, (tuple([s if st != 0 else 1 for s,st in zip(real_shape, v.strides)]), v.strides, real_offset)))
-      # then, we apply pre expand pads
-      if v.mask is not None:
-        pre_expand_pads = tuple((x,s-y) if st != 0 else (0,0) for (x,y),s,st in zip(v.mask, v.shape, v.strides))
-        post_expand_pads = tuple((x,s-y) if st == 0 else (0,0) for (x,y),s,st in zip(v.mask, v.shape, v.strides))
-        if any(x != (0,0) for x in pre_expand_pads):
-          to_apply.append((MovementOps.PAD, pre_expand_pads))
-          real_shape = tuple(x+s[0]+s[1] for x,s in zip(real_shape, pre_expand_pads))
-      # then, we do any expands
-      # NOTE: this is a good idea even without masks, since torch doesn't support negative strides and has to make a copy
-      if any(s != 1 and st == 0 for s,st in zip(real_shape, v.strides)): to_apply.append((MovementOps.EXPAND, real_shape))
-      # lastly, we apply post expand pads
-      if v.mask is not None and any(x != (0,0) for x in post_expand_pads): to_apply.append((MovementOps.PAD, post_expand_pads))
-    return to_apply
+  def unbind(self) -> Tuple[ShapeTracker, Dict[Variable, int]]:
+    unbound_views, var_vals = zip(*[v.unbind() for v in self.views])
+    return ShapeTracker(tuple(unbound_views)), merge_dicts(var_vals)
 
   # NOTE: if a stride is not always valid, it will be None
   def real_strides(self, ignore_valid=False) -> Tuple[Optional[sint], ...]:
@@ -124,7 +74,7 @@ class ShapeTracker:
     bad_idx_vars: Set[Variable] = set()
     for this_dim in (idx.nodes if isinstance(idx, SumNode) else [idx]):
       idx_maybe, stride_maybe = (this_dim.a, this_dim.b) if isinstance(this_dim, MulNode) else (this_dim, 1)
-      try: ret[idxs.index(idx_maybe)] = stride_maybe
+      try: ret[idxs.index(idx_maybe)] = cast(sint, stride_maybe)
       except ValueError: bad_idx_vars = bad_idx_vars.union(idx_maybe.vars())
     idx_vars, valid_vars = idx.vars(), valid.vars()
     for i,tidx in enumerate(idxs):
@@ -134,30 +84,27 @@ class ShapeTracker:
 
   def unit_stride_axes(self, ignore_valid=False) -> List[int]: return [i for i,st in enumerate(self.real_strides(ignore_valid)) if st == 1]
 
-  def _expr_idx(self, idx:Node, valid:Node) -> Tuple[Node, Node]:
-    for v in reversed(self.views[0:-1]):
+  def expr_idxs(self, idxs:Optional[Iterable[Node]]=None) -> Tuple[Node, Node]:
+    idxs = [Variable(f"idx{i}", 0, s-1) for i,s in enumerate(self.shape)] if idxs is None else list(idxs)
+    idx, valid = _expr_view(self.views[-1], idxs)
+    for view in reversed(self.views[0:-1]):
       if valid.max == 0: return NumNode(-1), valid
-      valid = expr_node_mask(v, idx, valid)
-      idx = expr_node(v, idx)
+      view = view.minify()
+      acc, idxs = 1, []
+      for d in reversed(view.shape):
+        idxs.append((idx//acc)%d)
+        acc *= d
+      idx, valid = _expr_view(view, idxs[::-1], valid)
+    assert not isinstance(idx.min, int) or idx.min >= -2**31, f"idx.min too small. {idx=}, {idx.min=}"
+    assert not isinstance(idx.max, int) or idx.max < 2**31, f"idx.max too big. {idx=}, {idx.max=}"
     return idx, valid
 
-  def expr_idxs(self, idxs:Optional[Iterable[Node]]=None):
-    if idxs is None: idxs = [Variable(f"idx{i}", 0, s-1) for i,s in enumerate(self.shape)]
-    idx = expr_idxs(self.views[-1], tuple(idxs))
-    valid = expr_node_mask(self.views[-1], idxs_to_idx(self.views[-1].shape, tuple(idxs)))
-    return self._expr_idx(idx, valid)
-
-  def expr_node(self, idx:Union[Node,str]='idx'):
-    if isinstance(idx, str): idx = Variable(idx, 0, prod(self.shape)-1)
-    return self._expr_idx(expr_node(self.views[-1], idx), expr_node_mask(self.views[-1], idx))
-
   def axis_is_masked(self, axis:int) -> bool:
     _, valid = self.expr_idxs()
     return f'idx{axis}' in [v.expr for v in valid.vars()]
 
   def simplify(self) -> ShapeTracker:
-    if len(self.views) >= 2 and (new_view := merge_views(self.views[-2], self.views[-1])) is not None:
-      if DEBUG >= 5: print(f"st simplify : {self.views[-2]} + {self.views[-1]} = {new_view}")
+    if len(self.views) >= 2 and (new_view := self.views[-2] + self.views[-1]) is not None:
       return ShapeTracker(self.views[:-2] + (new_view,)).simplify()
     return self
 
@@ -172,11 +119,3 @@ class ShapeTracker:
   def reshape(self, new_shape: Tuple[sint, ...]) -> ShapeTracker:
     if getenv("MERGE_VIEW", 1) and (new_view := self.views[-1].reshape(new_shape)) is not None: return ShapeTracker(self.views[0:-1] + (new_view,))
     return ShapeTracker(self.views + (View.create(new_shape), ))
-
-# returns the axes to create new_shape if new_shape can be created by combining axis from old_shape
-# TODO: if we remove movementops from lazy.py we can delete this
-def get_contraction(old_shape:Tuple[sint, ...], new_shape:Tuple[sint, ...]) -> Optional[List[List[int]]]:
-  acc_old, acc_new = list(itertools.accumulate(old_shape, operator.mul)), list(itertools.accumulate(new_shape, operator.mul))
-  try: split = [acc_old.index(acc)+1 if acc != 1 else 0 for acc in acc_new]
-  except ValueError: return None
-  return [list(range(st,ed)) for st,ed in zip([0]+split[:-1], split[:-1]+[len(old_shape)])]