move that

.github/workflows/test.yml

@@ -870,29 +870,29 @@ jobs:
     - name: Test ONNX Runner (WEBGPU)
       run: WEBGPU=1 PYTHONPATH=. python3 test/external/external_test_onnx_runner.py
 
-  #osxremote:
-  #  name: MacOS (remote metal)
-  #  runs-on: macos-15
-  #  timeout-minutes: 10
-  #  env:
-  #    REMOTE: 1
-  #    REMOTEDEV: METAL
-  #  steps:
-  #    - name: Checkout Code
-  #      uses: actions/checkout@v4
-  #    - name: Setup Environment
-  #      uses: ./.github/actions/setup-tinygrad
-  #      with:
-  #        key: macos-remote
-  #        deps: testing_minimal
-  #    - name: Check Device.DEFAULT and print some source
-  #      run: |
-  #        python -c "from tinygrad import Device; assert Device.DEFAULT == 'REMOTE', Device.DEFAULT"
-  #        python -c "from tinygrad import Device; assert Device.default.properties.real_device == 'METAL', Device.default.properties.real_device"
-  #        DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
-  #    - name: Run REMOTE=1 Test
-  #      run: |
-  #        python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_tensor_variable.py
+  osxremote:
+    name: MacOS (remote metal)
+    runs-on: macos-15
+    timeout-minutes: 10
+    env:
+      REMOTE: 1
+      REMOTEDEV: METAL
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v4
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: macos-remote
+          deps: testing_minimal
+      - name: Check Device.DEFAULT and print some source
+        run: |
+          python -c "from tinygrad import Device; assert Device.DEFAULT == 'REMOTE', Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.default.properties.real_device == 'METAL', Device.default.properties.real_device"
+          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
+      - name: Run REMOTE=1 Test
+        run: |
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_tensor_variable.py
 
   amdremote:
     name: Linux (remote)

test/test_linearizer.py

@@ -3,6 +3,7 @@ import unittest
 from dataclasses import replace
 
 from tinygrad.opt.kernel import Opt, OptOps, KernelOptError, Kernel, AxisType
+from tinygrad.codegen import rewrites_for_views, apply_rewrites
 from tinygrad.codegen.gpudims import get_grouped_dims
 from tinygrad.uop.ops import UOp, Ops, GroupOp, KernelInfo
 from tinygrad.device import Device, Buffer, is_dtype_supported
@@ -22,7 +23,7 @@ def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
   # now all input buffers in s[-1] should be realized
   # create fresh buffers for the outputs
   bufs = [Buffer((x).device, x.size, x.dtype).allocate() if i < len(s[-1].ast.src) else x for i,x in enumerate(s[-1].bufs)]
-  return s[-1].ast, bufs
+  return apply_rewrites(s[-1].ast, rewrites_for_views), bufs
 
 def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0, use_tensor_cores:int=1):
   a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)

tinygrad/codegen/__init__.py

@@ -16,8 +16,8 @@ from tinygrad.codegen.devectorizer import load_store_folding, load_store_indexin
   ReduceContext, correct_load_store, pm_render
 from tinygrad.codegen.optional import get_late_rewrite_patterns
 from tinygrad.codegen.linearize import block_create, pm_blockend_merge, block_merge, pm_finalize, BlockContext
+from tinygrad.opt.swizzler import view_left, view_right, cleanup_pm
 from tinygrad.opt import pm_optimize
-from tinygrad.opt.swizzler import view_left, view_right, fix_kernel_ops
 
 @dataclass
 class RewriteStep:
@@ -30,6 +30,12 @@ class RewriteStep:
 
 def apply_rewrites(sink:UOp, rewrites:list[RewriteStep]): return functools.reduce(lambda x,f: f(x), rewrites, sink)
 
+rewrites_for_views = [
+  RewriteStep(view_left, name="view left"),
+  RewriteStep(view_right, name="view right"),
+  RewriteStep(cleanup_pm, name="cleanup view"),
+]
+
 rewrites_for_linearizer = [
   RewriteStep(block_create, ctx=BlockContext.from_sink, name="Linearizer: Create Blocks", bottom_up=True),
   RewriteStep(pm_blockend_merge, name="Linearizer: Merge Blockends"),
@@ -45,10 +51,8 @@ def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVEC
   # ** lowerer (rewrite_shapetracker_with_index) **
   ret: list[RewriteStep] = []
 
-  # TODO: move these to codegen
-  ret.append(RewriteStep(view_left, name="Main View Left"))
-  ret.append(RewriteStep(view_right, name="Main View Right"))
-  ret.append(RewriteStep(view_left+fix_kernel_ops, bottom_up=True, name="replace buffer"))
+  # this used to be in schedule
+  ret.extend(rewrites_for_views)
 
   # this is kernel.py
   ret.append(RewriteStep(pm_optimize, ctx=lambda _: opts, name="optimize ast"))

tinygrad/opt/kernel.py

@@ -73,7 +73,7 @@ class Kernel:
       self.sts.append(unwrap(x.src[0].st))
 
     # add a shapetracker to the end to track the full shape, with 0 strides so it can merge
-    full_shape = ast.full_shape
+    full_shape = self.ast.full_shape
     self.sts.append(ShapeTracker.from_shape(full_shape, (0,)*len(full_shape)))
 
     # parameters for optimization

tinygrad/opt/swizzler.py

@@ -1,9 +1,9 @@
 from tinygrad.uop.ops import UOp, Ops, GroupOp, PatternMatcher, UPat, graph_rewrite, resolve, sint
 from tinygrad.helpers import all_same, prod, unwrap, colored
+from tinygrad.dtype import dtypes
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View, strides_for_shape, get_contraction_with_reduce
 from tinygrad.schedule.grouper import ALWAYS_CONTIGUOUS
-from tinygrad.dtype import ImageDType, dtypes
 
 merge_views = PatternMatcher([
   # merge adjacent views
@@ -96,7 +96,7 @@ view_right = merge_views+PatternMatcher([
   # apply view after reduceops
   (UPat(Ops.REDUCE_AXIS, src=(UPat(Ops.VIEW, src=(UPat(GroupOp.All-ALWAYS_CONTIGUOUS, name="src"),), name="v"),), name="r"), reduceop_view_right),
   # apply view after elementwise ops
-  (UPat(GroupOp.All-{Ops.SINK, Ops.REDUCE_AXIS, Ops.LOAD, Ops.STORE}, name="root"), elementwise_view_right),
+  (UPat(GroupOp.All-{Ops.SINK, Ops.REDUCE_AXIS}, name="root"), elementwise_view_right),
   # merge axes for double reduce (invert of SPLIT_REDUCEOP=1)
   (UPat(Ops.REDUCE_AXIS, src=(UPat(Ops.REDUCE_AXIS, name="r1"),), name="r2"),
    lambda r1,r2: r1.replace(arg=(r1.arg[0], r2.arg[1]+r1.arg[1])) if r1.arg[0] is r2.arg[0] else None),
@@ -112,22 +112,14 @@ def check_load_st(glbl:UOp, view:UOp):
   raise RuntimeError("self operand of augmented assign must be contiguous.\nhelp: consider using .contiguous():\n"
                      +colored("   - a += a.T\n", "red")+colored("   + a += a.T.contiguous()", "green"))
 
-fix_kernel_ops = PatternMatcher([
-  # add the LOAD
-  (UPat(Ops.DEFINE_GLOBAL, name="x"), lambda x: x.replace(tag=None).view(x.st).load() if x.tag is not None else None),
-  # STORE (except for meta ops)
-  (UPat(Ops.SINK, src=UPat(GroupOp.All-{Ops.STORE}), name="sink"), lambda sink:
-   UOp.sink(*[UOp.store(UOp(Ops.DEFINE_GLOBAL, (s:=x.base).dtype.ptr(s.st.real_size()), (), i).view(s.st), s) for i,x in enumerate(sink.src)])),
-  # passthrough ASSIGN
-  (UPat(Ops.ASSIGN, name="x"), lambda x: x.src[1]),
+cleanup_pm = PatternMatcher([
+  # add VIEW to any DEFINE_GLOBAL that somehow lost its view
+  (UPat(Ops.STORE, src=(UPat(Ops.DEFINE_GLOBAL, name="d"),), name="x", allow_any_len=True), lambda d,x: x.replace(src=(d.view(d.st),)+x.src[1:])),
   # VALID
   (UPat(Ops.VIEW, src=(UPat.cvar(),), name="self"),
    lambda self: UOp.where(UOp(Ops.VALID, dtypes.bool, (UOp(Ops.VIEW, arg=self.st),)), self.const_like(self.base.arg), 0)),
-  # remove CONTIGUOUS/DEVICE from kernel AST
-  (UPat((Ops.CONTIGUOUS, Ops.MSELECT), src=(UPat.var("x"),)), lambda x: x),
-  (UPat(Ops.VIEW, src=(UPat(Ops.DEVICE),), name="view"), lambda view: view.replace(src=())),
-  # no ImageDType after index
-  (UPat(GroupOp.All-{Ops.DEFINE_GLOBAL, Ops.VIEW}, name="x"), lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
+  # VIEW on SINK is SINK
+  (UPat(Ops.VIEW, name="v").sink(), lambda v: v.src[0].sink()),
   # if this kernel also assigns to the loaded buffer, ensure we can index it correctly
   (UPat(Ops.LOAD, src=(UPat.var("glbl").view(name="view"),)), check_load_st),
 ])

tinygrad/schedule/kernelize.py

@@ -149,11 +149,23 @@ create_kernels = PatternMatcher([
 
 # **** fix kernel AST
 
-early_buffer_ops = PatternMatcher([
+add_buffer_ops = PatternMatcher([
   # LOAD
-  (UPat(Ops.BUFFER, name="x"), lambda ctx,x: UOp(Ops.DEFINE_GLOBAL, x.dtype.ptr(x.size), (), ctx.index(x), tag=1)),
-  # no SINK for meta ops
+  (UPat(Ops.BUFFER, name="x"), lambda ctx,x: UOp.load(UOp(Ops.DEFINE_GLOBAL, x.dtype.ptr(x.size), (), ctx.index(x)).view(x.st),)),
+  # STORE (except for meta ops)
   (UPat(Ops.SINK, src=(UPat(Ops.CONTIGUOUS, src=(UPat(GroupOp.Meta, name="x"),),))), lambda x:x),
+  (UPat(Ops.SINK, src=UPat(GroupOp.All-{Ops.STORE}), name="sink"), lambda ctx,sink:
+   UOp.sink(*[UOp.store(UOp(Ops.DEFINE_GLOBAL, (s:=x.base).dtype.ptr(ctx[i].size), (), i).view(s.st), s) for i,x in enumerate(sink.src)])),
+  # passthrough ASSIGN
+  (UPat(Ops.ASSIGN, name="x"), lambda x: x.src[1]),
+])
+
+fix_kernel_ops = PatternMatcher([
+  # remove CONTIGUOUS/DEVICE from kernel AST
+  (UPat((Ops.CONTIGUOUS, Ops.MSELECT), src=(UPat.var("x"),)), lambda x: x),
+  (UPat(Ops.VIEW, src=(UPat(Ops.DEVICE),), name="view"), lambda view: view.replace(src=())),
+  # no ImageDType after index
+  (UPat(GroupOp.All-{Ops.DEFINE_GLOBAL, Ops.VIEW}, name="x"), lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
 ])
 
 replace_globals = PatternMatcher([
@@ -165,6 +177,8 @@ replace_globals = PatternMatcher([
 
 def fix_kernel_ast(k:UOp) -> UOp|None:
   if k.arg.ast.op in GroupOp.Meta or all(s.op is Ops.STORE for s in k.arg.ast.src): return None
+  # replace global memory ops with the BUFFER they write to
+  ast = graph_rewrite(k.arg.ast, replace_globals, bottom_up=True, name="replace globals")
   # replace buffer with define_global + add load/store last
   bufs = []
   for s in k.src:
@@ -172,9 +186,7 @@ def fix_kernel_ast(k:UOp) -> UOp|None:
     # traverse back through MSELECT and MSTACK. HACK: 0 branch of MSTACK only
     while s.op in {Ops.MSELECT, Ops.MSTACK}: s = s.src[0]
     bufs.append(s)
-  # replace global memory ops with the BUFFER they write to
-  ast = graph_rewrite(k.arg.ast, replace_globals, bottom_up=True, name="replace globals")
-  ast = graph_rewrite(ast, early_buffer_ops, bufs, bottom_up=True, name="replace buffer early")
+  ast = graph_rewrite(ast, merge_views+add_buffer_ops+fix_kernel_ops, bufs, bottom_up=True, name="replace buffer")
   if ast.op is Ops.SINK and not all_same([x.device for x in k.src]):
     raise RuntimeError(f"all buffers must be on the same device: {tuple(b.buf_uop.buffer for b in k.src)}")
   return k.replace(arg=Kernel(ast, k.arg.metadata))

tinygrad/uop/ops.py

@@ -154,8 +154,8 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
       sz = cast(PtrDType, self.dtype).size
       return ShapeTracker.from_shape((sz,)) if sz > 0 else None
 
-    # hack for PTX, CASTing the ptr loses the shape. even worse hack with tag
-    if self.op is Ops.CAST and self.src[0].op is Ops.DEFINE_GLOBAL and self.src[0].tag is None: return None
+    # hack for PTX, CASTing the ptr loses the shape
+    if self.op is Ops.CAST and self.src[0].op is Ops.DEFINE_GLOBAL: return None
 
     # otherwise we get the shape from sources
     if not (src_sts := [x.st for x in self.src if x.st is not None]): return None
@@ -440,7 +440,7 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     all_vars = set([x for x in self.toposort() if x.op is Ops.DEFINE_VAR])
     return bound_vars.union(set([x for x in all_vars if x not in bound_var_base]))
   def variables(self) -> list[Variable]:
-    st_vars: list[set[Variable]] = [x.arg.vars() for x in self.toposort() if x.op is Ops.VIEW]
+    st_vars: list[set[Variable]] = [x.st_arg.vars() for x in self.toposort() if x.op in GroupOp.Buffer]
     return sorted(set.union(*st_vars, set([x.unbind()[0] if x.op is not Ops.DEFINE_VAR else x for x in self.vars()])), key=lambda v: v.arg)
 
   # *** uop symbolic stuff ***