tinygrad 0.10.0__py3-none-any.whl → 0.10.1__py3-none-any.whl

tinygrad/codegen/kernel.py

@@ -1,25 +1,25 @@
 from __future__ import annotations
-import itertools, functools
+import itertools, functools, math
 from dataclasses import dataclass
 from collections import defaultdict
-from typing import Optional, List, Tuple, cast, Dict, Final, DefaultDict, Callable, Sequence
+from typing import Optional, cast, Final, Callable, Sequence
 from enum import Enum, auto
 
-from tinygrad.ops import GroupOp, KernelInfo, UOp, Ops, PatternMatcher, can_pad, print_uops, type_verify, resolve, Variable, sint, \
-    graph_rewrite, track_rewrites, UPat
+from tinygrad.ops import GroupOp, KernelInfo, UOp, Ops, can_pad, resolve, Variable, sint, graph_rewrite, track_rewrites, view_left, print_uops
+from tinygrad.spec import type_verify, shape_spec
 from tinygrad.device import Device
-from tinygrad.renderer import Renderer, TensorCore, Program
+from tinygrad.renderer import Renderer, TensorCore, ProgramSpec
 from tinygrad.dtype import ImageDType
-from tinygrad.helpers import all_same, colored, ansilen, dedup, getenv, prod, round_up, all_int, to_function_name, diskcache_put, unwrap
-from tinygrad.helpers import DEBUG, TC_OPT, USE_TC, AMX
+from tinygrad.helpers import all_same, colored, ansilen, dedup, getenv, prod, round_up, all_int, to_function_name, diskcache_put, unwrap, ContextVar
+from tinygrad.helpers import DEBUG, TC_SELECT, TC_OPT, USE_TC, AMX, CAPTURE_PROCESS_REPLAY
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import strides_for_shape
 from tinygrad.codegen.linearize import linearize_uop
-from tinygrad.codegen.uopgraph import full_graph_rewrite
+from tinygrad.codegen.rewriter import full_graph_rewrite
 from tinygrad.codegen.lowerer import rewrite_shapetracker_with_index, get_contraction
 
 class OptOps(Enum):
-  TC = auto(); UPCAST = auto(); UPCASTMID = auto(); UNROLL = auto(); LOCAL = auto() # noqa: E702
+  TC = auto(); UPCAST = auto(); UNROLL = auto(); LOCAL = auto() # noqa: E702
   GROUP = auto(); GROUPTOP = auto(); NOLOCALS = auto(); PADTO = auto(); SWAP = auto() # noqa: E702
   def __lt__(self, x:OptOps): return self.value < x.value
 
@@ -32,8 +32,8 @@ def check(cond:bool, msg:str=""):
 class Opt:
   op: OptOps
   axis: Optional[int] = None
-  amt: Optional[int] = None
-  def __repr__(self): return f"Opt(op={self.op}, axis={self.axis}, amt={self.amt})"
+  arg: Optional[int | tuple] = None
+  def __repr__(self): return f"Opt(op={self.op}, axis={self.axis}, arg={self.arg})"
   def real_axis(self, k:Kernel):
     if self.axis is None: return -1
     if self.op is OptOps.UNROLL: return k.first_reduce+self.axis
@@ -42,10 +42,10 @@ class Opt:
 
 @dataclass
 class TensorCoreOptions:
-  axes: Tuple[int, ...] # the location of the original N and M axes if still in the shape
-  axes_exist: Tuple[bool, ...] # true if the original N and M axes are still in the shape
-  axis_pads: Tuple[Tuple[int, int], ...]
-  def fix_axes(self, removed_axis:int): # adjust the TC axes if necesssary when a dimension is removed
+  axes: tuple[int, ...] # the location of the original N and M axes if still in the shape
+  axes_exist: tuple[bool, ...] # true if the original N and M axes are still in the shape
+  axis_pads: tuple[tuple[int, int], ...]
+  def fix_axes(self, removed_axis:int): # adjust the TC axes if necessary when a dimension is removed
     axes, axes_exist = list(self.axes), list(self.axes_exist)
     for tc_dim in [i for i in range(2) if axes_exist[i]]:
       if removed_axis < axes[tc_dim]: axes[tc_dim] -= 1
@@ -57,32 +57,28 @@ class Kernel:
     if ast.op is Ops.SINK: self.ast = ast
 
     self.opts = opts if opts is not None else Device[Device.DEFAULT].renderer
-    try: uop_sts_map = verify_ast(self.ast)
-    except AssertionError as e:
-      print("INVALID AST")
-      print(self.ast)
-      raise e
+    # verify AST matches the spec
+    if __debug__: type_verify(list(self.ast.toposort), shape_spec)
 
-    @functools.lru_cache(None)
-    def ordered_parents(op:UOp) -> List[UOp]: return dedup([item for x in op.src for item in ordered_parents(x)] + [op])
-    self.reduceops = dedup([x for x in ordered_parents(self.ast) if x.op is Ops.REDUCE_AXIS])
+    self.reduceops = [x for x in self.ast.toposort if x.op is Ops.REDUCE_AXIS]
 
-    self.vars: List[Variable] = self.ast.variables()
-    self.bufs: List[UOp] = [x for x in self.ast.parents if x.op in GroupOp.Buffer]
+    self.vars: list[Variable] = self.ast.variables()
+    # NOTE: this requires a specific order with the [::-1], this is likely a bug
+    self.bufs: list[UOp] = [x for x in self.ast.toposort if x.op in GroupOp.Buffer][::-1]
 
     # get earlybufs, before any reduceops
-    earlybufs: List[UOp] = [x for reduceop in self.reduceops for x in reduceop.parents if x.op in GroupOp.Buffer]
+    earlybufs: list[UOp] = [x for reduceop in self.reduceops for x in reduceop.src[0].toposort if x.op in GroupOp.Buffer]
     self.full_buf_index: int = self.bufs.index(earlybufs[0]) if earlybufs else 0
     # NOTE: full_shape can be wrong if there's a tree of reduces
 
     # create new shapetrackers inside this kernel, we will permute them
-    self.sts: List[ShapeTracker] = [x.st_arg for x in self.bufs]
+    self.sts: list[ShapeTracker] = [x.st_arg for x in self.bufs]
 
     # add the shapetrackers for each reduce
     # we use this to track which axes are reduced in each reduce
     for x in self.reduceops:
-      self.sts.append(uop_sts_map[x])
-      self.sts.append(uop_sts_map[x.src[0]])
+      self.sts.append(unwrap(x.st))
+      self.sts.append(unwrap(x.src[0].st))
 
     # move all reduce axes to the end
     reduce = list(enumerate(zip(self.full_shape, self.output_shape)))
@@ -90,15 +86,13 @@ class Kernel:
     self.reshape_and_permute(None, permute)
 
     # parameters for optimization
-    self.applied_opts: List[Opt] = []
+    self.applied_opts: list[Opt] = []
     self.group_for_reduces: int = 0
     self.upcasted: int = 0
     self.local_dims: int = 0
     self.tensor_core: Optional[TensorCore] = None
     self.tensor_core_opts: Optional[TensorCoreOptions] = None
     self.use_tensor_cores: int = 0
-    # the local aliased buffers for A and B
-    self.bufs_for_tensor_core: Dict[UOp, Tuple[int, int]] = {}
     self.dont_use_locals: bool = False
 
     # group simplifies
@@ -112,25 +106,23 @@ class Kernel:
     ret.opts, ret.ast = self.opts, self.ast
 
     # things downstream of the AST
-    ret.reduceops, ret.vars, ret.bufs, ret.full_buf_index = \
-      self.reduceops, self.vars, self.bufs, self.full_buf_index
+    ret.reduceops, ret.vars, ret.bufs, ret.full_buf_index = self.reduceops, self.vars, self.bufs, self.full_buf_index
     ret.sts = self.sts[:len(ret.bufs)+len(ret.reduceops)*2] # NOTE: must redo the local buffers with TC in beam
 
     # parameters for optimizations
     ret.applied_opts, ret.group_for_reduces, ret.upcasted, ret.local_dims, ret.dont_use_locals = \
       self.applied_opts[:], self.group_for_reduces, self.upcasted, self.local_dims, self.dont_use_locals
-    ret.tensor_core, ret.tensor_core_opts, ret.bufs_for_tensor_core, ret.use_tensor_cores = \
-      self.tensor_core, self.tensor_core_opts, self.bufs_for_tensor_core, self.use_tensor_cores
+    ret.tensor_core, ret.tensor_core_opts, ret.use_tensor_cores = self.tensor_core, self.tensor_core_opts, self.use_tensor_cores
 
     return ret
 
   @property
-  def membufs(self) -> List[UOp]: return dedup([x.src[0] for x in self.bufs if x.op in {Ops.LOAD, Ops.STORE}])
+  def membufs(self) -> list[UOp]: return dedup([x.src[0] for x in self.bufs if x.op in {Ops.LOAD, Ops.STORE}])
 
   # TODO: these need more tests or it might silently be no-op
   def float4_axis(self, i:int): return [x-self.first_upcast for x in self.sts[i].unit_stride_axes() if x >= self.first_upcast and self.sts[i].shape[x]%4 == 0]  # noqa: E501
 
-  def upcasted_axis(self, i:int) -> List[Tuple[int, Optional[sint], bool]]:
+  def upcasted_axis(self, i:int) -> list[tuple[int, Optional[sint], bool]]:
     upcasted_shape, upcasted_stride = self.sts[i].shape[self.first_upcast:], self.sts[i].real_strides()[self.first_upcast:]
     assert all_int(upcasted_shape), f"cannot upcast a symbolic amount {upcasted_shape=}"
     return list(zip(upcasted_shape, upcasted_stride,
@@ -144,24 +136,20 @@ class Kernel:
   def first_upcast(self) -> int: return self.shape_len-self.upcasted
 
   @property
-  def reduceop(self) -> Optional[UOp]: return self.reduceops[0] if len(self.reduceops) > 0 else None
+  def reduceop(self) -> UOp|None: return self.reduceops[0] if len(self.reduceops) > 0 else None
 
   @property
-  def output_shape(self) -> Tuple[sint, ...]: return self.sts[0].shape
+  def output_shape(self) -> tuple[sint, ...]: return self.sts[0].shape
 
   @property
-  def full_shape(self) -> Tuple[sint, ...]: return self.sts[self.full_buf_index].shape
+  def full_shape(self) -> tuple[sint, ...]: return self.sts[self.full_buf_index].shape
 
   @property
-  def full_unupcasted_shape(self) -> Tuple[sint, ...]: return self.full_shape[:self.first_upcast]
+  def full_unupcasted_shape(self) -> tuple[sint, ...]: return self.full_shape[:self.first_upcast]
 
   @property
   def shape_len(self) -> int: return len(self.sts[0].shape)
 
-  @property
-  def upcast_in_mid_reduce_axes(self) -> List[int]:
-    return [j for j in range(self.first_reduce, self.first_reduce+self.group_for_reduces) if self.full_shape[j] == self.sts[0].shape[j]]
-
   @property
   def global_dims(self) -> int: return self.first_reduce-self.local_dims
 
@@ -170,18 +158,17 @@ class Kernel:
   # cyan   -- local dims (warp ones first)
   #  *** self.first_reduce
   # green  -- reduce-local dims
-  # white  -- reduce-late upcasted dim (self.upcast_in_mid_reduce_axes)
   # red    -- reduce loops
   #  *** self.upcasted
   # purple -- reduce upcasted
   # yellow -- normal upcasted dimensions
-  def colors(self) -> List[str]:
+  def colors(self) -> list[str]:
     # first non local non reduce dims are global (blue)
     colors = ["blue"] * self.global_dims if not self.dont_use_locals else ["BLUE"] * self.global_dims
     # after global are local_dims; warp ones used in tensor cores must be closest to first_reduce (cyan)
     colors += ["cyan"] * self.local_dims
-    # between first_reduce and first_reduce + group_for_reduces, they are either upcast mid reduce (white), or late upcasted (green)
-    colors += ["white" if i in self.upcast_in_mid_reduce_axes else "green" for i in range(self.first_reduce, self.first_reduce + self.group_for_reduces)]  # noqa: E501
+    # between first_reduce and first_reduce + group_for_reduces, they are late upcasted (green)
+    colors += ["green"] * self.group_for_reduces
     # between first_reduce + group_for_reduces and upcasted, they are reduce (red)
     colors += ["red"] * (self.first_upcast - (self.first_reduce + self.group_for_reduces))
     # upcasted dimensions are reduce (magenta) or normal (yellow)
@@ -198,7 +185,7 @@ class Kernel:
   # ******************** base simplifiers ********************
 
   # apply reshape and permute to all shapetrackers
-  def reshape_and_permute(self, new_shape_fxn:Optional[Callable[[Tuple[sint, ...]], Sequence[sint]]], axis:Optional[Sequence[int]]):
+  def reshape_and_permute(self, new_shape_fxn:Optional[Callable[[tuple[sint, ...]], Sequence[sint]]], axis:Optional[Sequence[int]]):
     def reshape(st:ShapeTracker): return st.reshape(tuple(new_shape_fxn(st.shape))) if new_shape_fxn is not None else st
     def permute(st:ShapeTracker): return st.permute(tuple(axis)) if axis is not None else st
     self.sts = [permute(reshape(st)) for st in self.sts]
@@ -240,7 +227,7 @@ class Kernel:
     if isinstance(self.membufs[0].dtype, ImageDType):
       base_shape = self.membufs[0].dtype.shape
       if shape_idx_groups := get_contraction(self.output_shape, base_shape):
-        special_strides: Tuple[sint, ...] = tuple()
+        special_strides: tuple[sint, ...] = tuple()
         for i,g in enumerate(shape_idx_groups):
           shape_piece = tuple(self.output_shape[x] for x in g)
           assert prod(shape_piece) == base_shape[i], f"get_contraction was wrong? {shape_piece} != {base_shape[i]}"
@@ -298,37 +285,34 @@ class Kernel:
     s0, s1, s2 = axis_choices[-(axis+1)][0][0], axis_choices[-(axis+1)][1][0], axis_choices[-(axis+1)][2]  # s0 is n, s1 is m, s2 is k
     axis_pads = tuple((x, tc.dims[i]) for i, x in enumerate([s0, s1, s2]) if resolve(self.full_shape[x]%tc.dims[i] != 0))
     if axis_pads and (opt_level < 2): return None
-    self.bufs_for_tensor_core[reduceop] = (buf0, buf1)
     if DEBUG >= 3: print("TENSOR CORES", axis_buf0, axis_buf1, tc)
     return TensorCoreOptions(axes=(s0, s1, s2), axes_exist=(True, True), axis_pads=axis_pads)
 
-  def _apply_tc_opt(self, use_tensor_cores:int, axis:int, opt_level:int) -> bool:
+  def _apply_tc_opt(self, use_tensor_cores:int, axis:int, tc_select:int, opt_level:int) -> bool:
     if use_tensor_cores and self.reduceop is not None and self.reduceop.arg[0] is Ops.ADD:
-      for tc in self.opts.tensor_cores:
+      tensor_cores = self.opts.tensor_cores if tc_select == -1 else [self.opts.tensor_cores[tc_select]]
+      for tc in tensor_cores:
         tensor_core_opts = [self._create_tc_opts(reduceop, tc, axis, opt_level) for reduceop in self.reduceops]
         # can only fuse reduces with the same tc options
         assert all_same(tensor_core_opts)
         if tensor_core_opts[0] is None: continue
-        # tensor core -- unroll the reduce dim, upcast input and local the correct thread pattern
         self.tensor_core_opts = tc_opts = tensor_core_opts[0]
 
         # attempt to pad the tensor axes that require it
         try:
           for axis, dim in tc_opts.axis_pads: self.apply_opt(Opt(OptOps.PADTO, axis, dim), append_opt=False) # PADTO might fail
         except KernelOptError: continue
-        for tc_dim, amt in tc.reduce_axes: self.apply_opt(Opt(OptOps.UNROLL,tc_opts.axes[2]-self.first_reduce,amt), append_opt=False)
-        for opt in tc.opts_seq:
-          if opt == "UP":
-            for tc_dim, amt in tc.early_upcast_axes: self.apply_opt(Opt(OptOps.UPCAST,tc_opts.axes[tc_dim],amt), append_opt=False)
-          elif opt == "LC":
-            for tc_dim, amt in tc.threads: self.apply_opt(Opt(OptOps.LOCAL,tc_opts.axes[tc_dim],amt), append_opt=False)
+        # tensor core -- unroll the reduce dim (K), upcast and local the inner and outer dims (N, M)
+        for dim, amt in tc.get_reduce_axes(): self.apply_opt(Opt(OptOps.UNROLL, tc_opts.axes[2]-self.first_reduce, amt), append_opt=False)
+        for opt in tc.opts: self.apply_opt(Opt({"u":OptOps.UPCAST, "l":OptOps.LOCAL}[opt[0]], tc_opts.axes[int(opt[1])], 2), append_opt=False)
         self.tensor_core = tc
         self.use_tensor_cores = use_tensor_cores  # TC=2 will do the shape ops without the WMMA
         return True
     return False
 
-  def apply_tensor_cores(self, use_tensor_cores=1, extra_opts:Optional[List[Opt]]=None, axis:int=0, tc_opt:Optional[int]=None) -> bool:
-    """ Attempts to apply a tensor core optimization to the kernel.  If one exists and applies properly, return true, otherwise return false.
+  def apply_tensor_cores(self, use_tensor_cores=1, extra_opts:Optional[list[Opt]]=None, axis:int=0, tc_select:Optional[int]=None,
+                         tc_opt:Optional[int]=None) -> bool:
+    """ Attempts to apply a tensor core optimization to the kernel. If one exists and applies properly, return true, otherwise return false.
     Tensor cores are optimized instructions that matrix multiply-accumulate across a wave of threads: D(M, N) = A(M, K) * B(K, N) + C(M, N).
 
     Keyword arguments:
@@ -337,15 +321,19 @@ class Kernel:
       1: enable tensor cores
       2: apply tensor core shape but don't use UOp.WMMA
     extra_opts -- additional Opt's to apply after the tensor core instead of the hand-coded additional Opt's (default None)
+    tc_select -- specifies which tensor core(s) to use for optimization (default -1)
+      -1: iterates through all available tensor cores in order and uses the first one that matches the requirements (dims and dtypes)
+      [0-N]: uses only the n'th tensor core available; useful for search
     tc_opt -- controls which kinds of kernels may be eligible for tensor cores application (default 2 during BEAM, 0 otherwise)
       0: applies to only kernels with a single reduce axis and direct UOps.LOAD into Ops.MUL
       1: allows kernels with multiple reduce axes and also multiplication of UOps.CAST'd buffers
       2: allows kernels with M, N, K axes that are not multiples of the tensor core dimensions by applying padding those axes as needed
     """
+    if tc_select is None: tc_select = TC_SELECT.value
     if tc_opt is None: tc_opt = TC_OPT.value
     if not self.opts.tensor_cores and use_tensor_cores != 2: return False
     try: # check TC first and apply hand-coded opts if successful
-      self.apply_opt(Opt(OptOps.TC, axis, tc_opt))
+      self.apply_opt(Opt(OptOps.TC, axis, (tc_select, tc_opt)))
 
       if (tc_opts:=self.tensor_core_opts) is not None:
         if extra_opts is not None:
@@ -364,24 +352,28 @@ class Kernel:
       return False
 
   def apply_opt(self, opt:Opt, append_opt:bool=True):
-    if self.dont_use_locals: check(opt.op not in {OptOps.LOCAL, OptOps.GROUP, OptOps.GROUPTOP, OptOps.UPCASTMID}, "not using locals")
+    if self.dont_use_locals: check(opt.op not in {OptOps.LOCAL, OptOps.GROUP, OptOps.GROUPTOP}, "not using locals")
 
     if opt.op is OptOps.TC:
       check(len(self.applied_opts) == 0, "tensor core opts must be first") # TODO: things like PADTO might be fine
-      check(opt.axis is not None and opt.amt is not None, "tensor core opts must have an axis and amt")
       check((use_tensor_cores:=USE_TC.value) == 2 or len(self.opts.tensor_cores) > 0, "must have tensor cores or TC=2")
-      check(self._apply_tc_opt(use_tensor_cores, cast(int, opt.axis), cast(int, opt.amt)), "no tensor core available")
+      check(opt.axis is not None, "tensor core opts must have an axis")
+      check(opt.arg is not None and isinstance(opt.arg, tuple) and len(opt.arg) == 2, "tensor core opts must have tc_select and tc_opt")
+      check(-1 <= (tc_select:=cast(tuple, opt.arg)[0]) < len(self.opts.tensor_cores), "tensor core opts must have valid tc_select")
+      check(0 <= (tc_opt:=cast(tuple, opt.arg)[1]) <= 2, "tensor core opts must have valid tc_opt")
+      check(self._apply_tc_opt(use_tensor_cores, cast(int, opt.axis), tc_select, tc_opt), "no tensor core available")
       self.applied_opts.append(opt)
       return
 
     axis = opt.real_axis(self)
     check(axis < len(self.full_shape), "invalid axis")
 
-    if opt.op is OptOps.SWAP: amt = cast(int, opt.amt)  # amt is an axis in the SWAPs
-    elif opt.amt is not None:
-      amt = opt.amt if opt.amt != 0 else self.full_shape[axis]
-      check(isinstance(amt, int) and amt != 1, "shift/padto of amt 1 or Node is meaningless")
-      if opt.op is not OptOps.PADTO: check(self.full_shape[axis] % amt == 0, "no longer valid shift")
+    if opt.op is OptOps.SWAP: amt = cast(int, opt.arg)  # arg is an axis in the SWAPs
+    elif opt.arg is not None:
+      check(isinstance(opt.arg, int), "arg should be int")
+      amt = arg if (arg:=cast(int, opt.arg)) != 0 else self.full_shape[axis]
+      check(isinstance(amt, int) and amt != 1, f"shift/padto of {amt=}, 1 or symbolic amount is meaningless")
+      if opt.op is not OptOps.PADTO: check(self.full_shape[axis] % amt == 0, f"no longer valid shift {self.full_shape[axis]=}, {amt=}")
     else: amt = -1
 
     if self.reduceop is not None and (opt.op in {OptOps.GROUP, OptOps.GROUPTOP} or \
@@ -416,18 +408,10 @@ class Kernel:
       self.upcast()
     elif opt.op is OptOps.UPCAST:                     # yellow
       check(axis < self.first_reduce, "upcast is for non-reduce")
-      check(not (self.tensor_core and self.global_dims <= axis < self.global_dims+len(self.tensor_core.threads)), "can't upcast TC locals")
+      check(not (self.tensor_core and self.global_dims <= axis < self.global_dims+len(self.tensor_core.get_local_axes())), "can't upcast TC locals")
       check(amt <= 16, "don't upcast more than 16")
       self.shift_to(axis, amt, insert_before=None)
       self.upcast()
-    elif opt.op is OptOps.UPCASTMID:                  # white
-      check(self.bufs[0].src[0].dtype.name.startswith('image') and not self.float4_axis(0) and self.group_for_reduces != 0 and self.first_reduce <= 2 and prod(self.sts[0].shape) > 1, "invalid upcast mid reduce")  # noqa: E501
-      axes = self.sts[0].unit_stride_axes()
-      check(len(axes) == 1, f"wrong number of stride 1 axis : {axes}")
-      check(axes[0] == axis, "wrong axis")
-      check(amt == 4, "don't upcast mid anything but 4")
-      self.shift_to(axis, amt, insert_before=self.first_reduce + self.group_for_reduces)
-      self.group_for_reduces += 1
     elif opt.op is OptOps.NOLOCALS:
       check(self.opts.has_local and not self.dont_use_locals, "NOLOCALS is meaningless if target does not support local or already not using locals")
       check(self.local_dims == 0 and self.group_for_reduces == 0, "can't have no locals with locals")
@@ -441,7 +425,7 @@ class Kernel:
       check(not self.vars, "does not work with symbolic shape")
       check(axis < self.first_upcast, "cannot pad upcasted")
       # ok to pad SUM if all parent ALU ops have f(0) = 0
-      if (r:=self.reduceop) is not None and self.first_reduce <= axis: check(r.arg[0] is Ops.ADD and can_pad(r), f"cannot pad {r}")
+      if (r:=self.reduceop) is not None and self.first_reduce <= axis: check(r.arg[0] is Ops.ADD and can_pad(r, {}, {}), f"cannot pad {r}")
       padded = False
       for i,st in enumerate(self.sts):
         if (s:=st.shape[axis]) == 1: continue  # reduced
@@ -460,8 +444,7 @@ class Kernel:
     if isinstance(self.membufs[0].dtype, ImageDType):
       unit_stride_axes_mul_4 = [i for i in self.sts[0].unit_stride_axes(ignore_valid=True) if self.sts[0].shape[i]%4 == 0]
       assert unit_stride_axes_mul_4, f"needs a unit stride axis in {self.bufs[0]}"
-      if all(x < self.first_upcast for x in unit_stride_axes_mul_4) and unit_stride_axes_mul_4[0] not in self.upcast_in_mid_reduce_axes:
-        self.apply_opt(Opt(OptOps.UPCAST, unit_stride_axes_mul_4[0], 4))
+      if all(x < self.first_upcast for x in unit_stride_axes_mul_4): self.apply_opt(Opt(OptOps.UPCAST, unit_stride_axes_mul_4[0], 4))
     return self
 
   def hand_coded_optimizations(self) -> Kernel:
@@ -496,19 +479,12 @@ class Kernel:
               break
             except KernelOptError: pass
 
-      # are we upcasting in mid reduce? (only for images)
-      if self.bufs[0].src[0].dtype.name.startswith('image') and not self.float4_axis(0) and self.group_for_reduces and self.first_reduce <= 2 and prod(self.sts[0].shape) > 1:  # noqa: E501
-        axes = self.sts[0].unit_stride_axes()
-        assert len(axes) == 1, f"wrong number of stride 1 axis : {axes}"
-        if self.sts[0].shape[axes[0]]%4 == 0:
-          self.apply_opt(Opt(OptOps.UPCASTMID, axes[0], 4))
-
     # upcast float4 images
     for buf_index,buf in enumerate(self.bufs):
       unit_stride_axes_mul_4 = [i for i in self.sts[buf_index].unit_stride_axes(ignore_valid=True) if self.sts[buf_index].shape[i]%4 == 0]
       if buf.src[0].dtype.__class__ is ImageDType:
         #assert len(unit_stride_axes_mul_4) >= 1, f"needs a unit stride axis in {self.bufs[buf_index]}"
-        if len(unit_stride_axes_mul_4) and all(x < self.first_upcast for x in unit_stride_axes_mul_4) and unit_stride_axes_mul_4[0] not in self.upcast_in_mid_reduce_axes:  # noqa: E501
+        if len(unit_stride_axes_mul_4) and all(x < self.first_upcast for x in unit_stride_axes_mul_4):
           if unit_stride_axes_mul_4[0] < self.first_reduce:
             self.apply_opt(Opt(OptOps.UPCAST, unit_stride_axes_mul_4[0], 4))
           else:
@@ -524,7 +500,7 @@ class Kernel:
     # expression and run test/test_ops.py with IMAGE=2
     # if there are small dims with lots of valid masks, upcast them (they might be from Tensor.stack)
     # this can be made much smarter
-    to_upcast: List[int] = []
+    to_upcast: list[int] = []
     # upcast leading axes first (hack-ish for winograd; we actually want to upcast masked axes with low stride first)
     for axis in range(self.first_reduce):
       # we might want to be able to split axes that are masked, or refuse to merge them in simplify_merge_adjacent
@@ -577,7 +553,7 @@ class Kernel:
       else:
         # prioritize making expand axes local
         local_axis_ranking = [(any(self.sts[buf_index].views[-1].strides[axis] == 0 for buf_index in range(len(self.sts))), axis) for axis in range(len(self.full_shape[:self.first_reduce]))]  # noqa: E501
-        to_local: List[Tuple[int, int]] = []
+        to_local: list[tuple[int, int]] = []
         for _, axis in sorted(local_axis_ranking, key=lambda x: (-x[0], -x[1])):
           local_size = prod(sz for _, sz in to_local)
           local_sz: Optional[int] = next((x for x in ([32] * (axis == 0) + [16, 8, 4, 3, 2]) if self.full_shape[axis] % x == 0 and local_size * x <= 128), None)  # noqa: E501
@@ -593,11 +569,11 @@ class Kernel:
 
   # **** kernel outputs ****
 
-  kernel_cnt: Final[DefaultDict[str, int]] = defaultdict(int)
+  kernel_cnt: Final[defaultdict[str, int]] = defaultdict(int)
   @functools.cached_property
   def name(self) -> str:
     # kernel name (before late upcast)
-    kernel_type = "r" if self.reduceop is not None else ("C" if all(x.op in GroupOp.Buffer for x in self.ast.parents) else "E")
+    kernel_type = "r" if self.reduceop is not None else ("C" if all(x.op is Ops.SINK or x.op in GroupOp.Buffer for x in self.ast.toposort) else "E")
     suffix = colored('_', 'BLACK').join([colored(x.render() if isinstance(x, UOp) else str(x), c) for x,c in zip(self.full_shape, self.colors())])
     name = kernel_type + (f"{len(self.ast.src)}" if len(self.ast.src) > 1 else "") + "_" + suffix
 
@@ -612,7 +588,10 @@ class Kernel:
       ret = op.replace(src=tuple(fixup_ast(x) for x in op.src))
       if op.op in GroupOp.Buffer and op in self.bufs:
         st_uop = self.sts[self.bufs.index(op)].to_uop()
-        return ret.replace(src=(st_uop,)) if op.op is Ops.VALID else ret.replace(src=(ret.src[0], st_uop, *ret.src[2:]))
+        # NOTE: if CONST got masked after applying opts, we create a new VALID
+        if op.op is Ops.CONST and any(v.mask is not None for v in unwrap(st_uop.st).views): return op.valid(unwrap(st_uop.st))
+        # otherwise we just replace the VIEW source
+        return ret.replace(src=(st_uop,)) if len(op.src) == 1 else ret.replace(src=(ret.src[0], st_uop, *ret.src[2:]))
       if op.op is Ops.SINK: return ret.replace(arg = KernelInfo(self.local_dims, self.upcasted, self.dont_use_locals))
       if op.op is Ops.REDUCE_AXIS:
         reduce_idx = len(self.bufs) + self.reduceops.index(op) * 2
@@ -623,47 +602,43 @@ class Kernel:
         grouped_axes = reduced_axes(self.first_reduce, self.first_reduce + self.group_for_reduces)
 
         if (tc := self.tensor_core) and (self.use_tensor_cores == 1 or self.use_tensor_cores == 3):
-          def fix_st(st: ShapeTracker, wd_pattern, tcd_pattern):
-            wd, warp_dims = self.global_dims,  tuple(sz for _, sz in tc.threads)
-            tcd, tcd_dims = self.first_upcast, tuple(sz for _, sz in tc.reduce_axes + tc.early_upcast_axes)
-
-            assert st.shape[wd:wd+len(warp_dims)] == warp_dims, f"warp dims wrong: {st.shape[wd:wd+len(warp_dims)]=} != {warp_dims=}"
-            assert st.shape[tcd:tcd+len(tcd_dims)] == tcd_dims, f"tcd dims wrong: {st.shape[tcd:tcd+len(tcd_dims)]=} != {tcd_dims=}"
-            assert tc.expanded_shape is not None
-
-            new_shape = st.shape[:tcd] + tc.expanded_shape + st.shape[tcd+len(tcd_dims):]  # expand the tcd
-            permaxis = list(range(wd)) + [y + (wd if x == 0 else tcd) for x,y in wd_pattern]  + list(range(wd+len(warp_dims),tcd)) + \
-                                         [y + (wd if x == 0 else tcd) for x,y in tcd_pattern] + list(range(tcd+len(tc.expanded_shape),len(new_shape)))
-            return st.reshape(new_shape).permute(tuple(permaxis)).reshape(st.shape).simplify()
+          wd, tcd = self.global_dims, self.first_upcast
+          def get_upcast_axes(buf): # upcast along non-zero dimensions of (tc_reduce + tc_upcast)
+            upcast_axes = int(math.log2(tc.elements_per_thread[buf]))
+            return tuple((tcd + len(tc.get_reduce_axes()) + len(tc.get_upcast_axes()) - (i+1), 2) for i in range(upcast_axes))
+          def get_tc_swizzle_st(shape, local_perm, upcast_perm):
+            offset = (tcd - (wd + len(local_perm)))
+            permaxis = list(range(wd)) \
+              + [wd + x + (offset if x >= len(local_perm) else 0) for x in local_perm]  + list(range(wd + len(local_perm), tcd)) \
+              + [wd + x + (offset if x >= len(local_perm) else 0) for x in upcast_perm] + list(range(tcd + len(upcast_perm), len(shape)))
+            return ShapeTracker.from_shape(shape).permute(tuple(permaxis))
 
           srcs = list((ret.src[0] if ret.src[0].op is not Ops.CAST else ret.src[0].src[0]).src)
-          for i, tc_pattern in enumerate([tc.st1_pattern, tc.st2_pattern]):
-            if tc_pattern: srcs[i] = srcs[i].view(fix_st(unwrap(srcs[i].st), *tc_pattern))
+          for i, (src, swizzle) in enumerate(zip(srcs, tc.swizzle)):
+            if swizzle: srcs[i] = src.view(get_tc_swizzle_st((src if src.op is Ops.LOAD else src.src[0]).st_arg.shape, *swizzle))
 
             if self.use_tensor_cores == 3:  # for TC=3, emulate the warp addressing with locals
               local_shape = tuple(1 if i >= self.first_reduce and i < self.first_upcast else s for i, s in enumerate(self.full_shape))
               st = store_st = ShapeTracker.from_shape(local_shape)
-              local_buffer = UOp(Ops.DEFINE_LOCAL, tc.dtype_in.ptr(local=True), (), (f"temp{i + 1}", st.real_size()))
-              if tc_pattern: store_st = fix_st(store_st, *tc_pattern)
+              local_buffer = UOp(Ops.DEFINE_LOCAL, tc.dtype_in.ptr(size=st.real_size(), local=True), (), f"temp{i}")
+              if swizzle: store_st = get_tc_swizzle_st(store_st.shape, *swizzle)
               local_store = UOp.store(local_buffer, store_st.to_uop(), srcs[i])
               srcs[i] = UOp(Ops.LOAD, tc.dtype_in, (local_buffer, st.to_uop(), local_store))
 
-          tc_reduce_axes = tuple(self.first_upcast + ax for ax, _ in tc.reduce_axes)
-          if self.use_tensor_cores == 1: # real WMMA, use CONTRACT/EXPAND to get the vectorization right
-            upcast_axes = tuple(tuple((self.first_upcast + ax, sz) for ax, sz in up) for up in tc.upcast_axes)
-            wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.opts.device, prod(sz for _, sz in tc.threads), upcast_axes, tc_reduce_axes)
-            wmma_sz = [prod(x[1] for x in l) for l in upcast_axes]
-            wmma = UOp(Ops.WMMA, dtype=tc.dtype_out.vec(wmma_sz[2]), src=(
-              UOp(Ops.CONTRACT, dtype=srcs[0].dtype.vec(wmma_sz[0]), src=(srcs[0],), arg=upcast_axes[0]),
-              UOp(Ops.CONTRACT, dtype=srcs[1].dtype.vec(wmma_sz[1]), src=(srcs[1],), arg=upcast_axes[1]),
-              UOp.const(tc.dtype_out.vec(wmma_sz[2]), 0.0)), arg=wmma_arg)
-            tc_uop = UOp(Ops.EXPAND, tc.dtype_out, (wmma,), arg=upcast_axes[2])
+          tc_reduce_axes = tuple(tcd + ax for ax, _ in tc.get_reduce_axes())
+          if self.use_tensor_cores == 1: # real WMMA, use CONTRACT/UNROLL to get the vectorization right
+            tc_upcast_axes = (get_upcast_axes(0), get_upcast_axes(1), get_upcast_axes(2))
+            wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.opts.device, tc.threads, tc_upcast_axes, tc_reduce_axes)
+            wmma = UOp(Ops.WMMA, dtype=tc.dtype_out.vec(tc.elements_per_thread[2]), src=(
+              UOp(Ops.CONTRACT, dtype=srcs[0].dtype.vec(tc.elements_per_thread[0]), src=(srcs[0],), arg=tc_upcast_axes[0]),
+              UOp(Ops.CONTRACT, dtype=srcs[1].dtype.vec(tc.elements_per_thread[1]), src=(srcs[1],), arg=tc_upcast_axes[1]),
+              UOp.const(tc.dtype_out.vec(tc.elements_per_thread[2]), 0.0)), arg=wmma_arg)
+            tc_uop = UOp(Ops.UNROLL, tc.dtype_out, (wmma,), arg=tc_upcast_axes[2])
 
           else: # for TC=3 MUL/SUM instead of WMMA
             tc_uop = UOp(Ops.REDUCE_AXIS, tc.dtype_out, ((srcs[0] * srcs[1]).cast(tc.dtype_out),), (Ops.ADD, tc_reduce_axes))
 
-          new_reduce_axes = tuple(i for i in axes if i not in tc_reduce_axes)
-          return ret.replace(src=(tc_uop,), arg=(Ops.ADD, new_reduce_axes)) if new_reduce_axes else tc_uop
+          return ret.replace(src=(tc_uop,), arg=(Ops.ADD, new_axes)) if (new_axes := tuple(i for i in axes if i not in tc_reduce_axes)) else tc_uop
 
         ret = ret.replace(arg = (op.arg[0], axes))
         if self.group_for_reduces and grouped_axes:
@@ -672,7 +647,8 @@ class Kernel:
               for i in range(self.first_reduce, self.first_reduce+self.group_for_reduces)]) + \
             (1,) * (self.shape_len - self.upcasted - self.group_for_reduces - self.first_reduce) + tuple([x[0] for x in self.upcasted_axis(0)])
           st_uop = ShapeTracker.from_shape(local_shape).to_uop()
-          local_buffer = UOp(Ops.DEFINE_LOCAL, op.dtype.ptr(local=True), (), (f"temp{self.reduceops.index(op)+1}", st_uop.arg.real_size()))
+          local_size = st_uop.arg.real_size()
+          local_buffer = UOp(Ops.DEFINE_LOCAL, op.dtype.ptr(local_size, local=True), (), f"temp{self.reduceops.index(op)}")
           local_load = UOp(Ops.LOAD, op.dtype, (local_buffer, st_uop, UOp.store(local_buffer, st_uop, ret)))
           grouped_reduce = UOp(Ops.REDUCE_AXIS, op.dtype, (local_load,), arg=(op.arg[0], grouped_axes))
           if op is self.reduceops[-1]: return grouped_reduce
@@ -681,9 +657,7 @@ class Kernel:
 
       return ret
 
-    return graph_rewrite(fixup_ast(self.ast), PatternMatcher([
-      (UPat({*GroupOp.ALU,Ops.CAST,Ops.BITCAST,Ops.ASSIGN}, name="e").view(name="v"), lambda e,v: e.replace(src=tuple(s.view(v.st) for s in e.src))),
-      (UPat(Ops.LOAD, name="b").view(name="v"), lambda b,v: b.replace(src=tuple((v.arg).to_uop() if s.op is Ops.VIEW else s for s in b.src)))]))
+    return graph_rewrite(fixup_ast(self.ast), view_left)
 
   # **** this is the lowerer ****
 
@@ -696,58 +670,26 @@ class Kernel:
       if getenv("RAWAST"): print(self.ast)
       print(modified_ast)
       print(self.applied_opts)
-    verify_ast(modified_ast)
+    # verify AST matches the spec after applying opts
+    if __debug__: type_verify(list(modified_ast.toposort))
+    # TODO: sadly modified_ast doesn't pass the shape spec because of how group_for_reduces constructs UOps, there's probably a way to fix this
+    #if __debug__: type_verify(list(modified_ast.toposort), shape_spec)
 
-    self.uops:List[UOp] = linearize_uop(full_graph_rewrite(rewrite_shapetracker_with_index(modified_ast, self.opts), self.opts))
+    self.uops:list[UOp] = linearize_uop(full_graph_rewrite(rewrite_shapetracker_with_index(modified_ast, self.opts), self.opts))
     if DEBUG >= 5: print_uops(self.uops)
     return self
 
-  def to_program(self, name_override:Optional[str]=None) -> Program:
+  def to_program(self, name_override:Optional[str]=None) -> ProgramSpec:
     self.linearize()
     src = self.opts.render(name:=to_function_name(ansiname:=(name_override if name_override is not None else self.name)), self.uops)
 
-    if getenv("RUN_PROCESS_REPLAY"):
-      from test.external.process_replay.helpers import get_process_replay_ctx
-      diskcache_put("kernel_process_replay", str(id(self)), (self.ast, self.opts, self.applied_opts, name, *get_process_replay_ctx(), src))
+    if CAPTURE_PROCESS_REPLAY:
+      diskcache_put("kernel_process_replay", str(id(self)), (self.ast, self.opts, self.applied_opts, name, ContextVar._cache, src))
 
     # group non-local bufs by the op type (LOAD or STORE) and the buffer arg. take the max access of that buffer in bytes
     # TODO: these max and min don't work on symbolic, and results are very wrong.
     mem_bytes = sum(max(x.src[0].dtype.itemsize * x.st_arg.real_size() for x in group)
-      for _, group in itertools.groupby([x for x in self.ast.parents if x.op in GroupOp.Buffer and x.src[0].op is Ops.DEFINE_GLOBAL],
+      for _, group in itertools.groupby([x for x in self.ast.toposort if x.op in GroupOp.Buffer and x.src[0].op is Ops.DEFINE_GLOBAL],
                         key=lambda x: (x.op, x.src[0].arg)))
-    return Program(ansiname, src, self.opts.device, self.uops, mem_estimate=mem_bytes,
+    return ProgramSpec(ansiname, src, self.opts.device, self.uops, mem_estimate=mem_bytes,
                    global_size=[1,1,1] if self.opts.has_local else None, local_size=[1,1,1] if self.opts.has_local else None)
-
-# the living definition of intermediate UOps
-
-def _assert_valid_uop(uop:UOp, st:ShapeTracker, sts:Dict[UOp, ShapeTracker]) -> None:
-  if not uop.has_st or uop in sts: return
-  # restore globals from the two stage reduce
-  if uop.op is Ops.LOAD and uop.src[0].op is Ops.DEFINE_LOCAL:
-    _assert_valid_uop(local_reduce:=uop.src[2].src[2], uop.st_arg, sts)
-    sts[uop] = sts[local_reduce]
-    return
-  for x in uop.src: _assert_valid_uop(x, st, sts)
-  # only reduceuop is allowed to change shape, limited to turning n to 1
-  if uop.op in {Ops.REDUCE_AXIS, Ops.WMMA}: st = ShapeTracker.from_shape(sts[uop.src[0]].reduce(uop.axis_arg))
-  # movementops are pushed to VIEW
-  elif uop.op is Ops.VIEW:
-    assert len(uop.src) == 0, f"can't swizzle in kernel yet {uop}"
-    st = uop.arg
-  # everything else inherits shape
-  else:
-    st = (src_sts:=[sts[x] for x in uop.src if x.has_st])[0]
-    if not all_same(shapes:=[x.shape for x in src_sts]):
-      if all_same(sizes:=[prod(x) for x in shapes]): raise AssertionError(f"found implicit reshape {shapes}")
-      raise AssertionError(f"found implicit expand {sizes} {shapes}")
-  sts[uop] = st
-
-def verify_ast(ast:UOp) -> Dict[UOp, ShapeTracker]:
-  assert ast.op is Ops.SINK and all(x.op is Ops.STORE for x in ast.src), "must be SINK"
-  assert all_same([x.st_arg.size for x in ast.src]), "outputs must be exactly the same size"
-  sts: Dict[UOp, ShapeTracker] = {}
-  for out in ast.src: _assert_valid_uop(out, out.st_arg, sts)
-  shape_dims = [sorted(dedup(dims)) for dims in zip(*[x.shape for x in sts.values()])]
-  assert all(len(x) == 1 or (len(x) == 2 and x[0] == 1) for x in shape_dims), f"shapes must have either 1 or n in each dimension, {shape_dims}"
-  type_verify(list(sts))
-  return sts