rewrite name

.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_jit.py

@@ -792,6 +792,7 @@ class TestJitGraphSplit(unittest.TestCase):
       multigraph=[self.ji_graph(5)],
       hcqgraph=[self.ji_graph(5)])
 
+  @unittest.skip("flaky")
   def test_jit_multidev_xfer(self):
     if Device.DEFAULT in {"CPU", "LLVM"}: raise unittest.SkipTest("CPU/LLVM is not a valid default device for this test (zero-copies)")
 

tinygrad/codegen/__init__.py

@@ -17,7 +17,6 @@ from tinygrad.codegen.devectorizer import load_store_folding, load_store_indexin
 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 import pm_optimize
-from tinygrad.opt.swizzler import view_left, view_right, fix_kernel_ops
 
 @dataclass
 class RewriteStep:
@@ -45,11 +44,6 @@ 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 is kernel.py
   ret.append(RewriteStep(pm_optimize, ctx=lambda _: opts, name="optimize ast"))
 

tinygrad/opt/kernel.py

@@ -14,7 +14,7 @@ from tinygrad.dtype import ImageDType, AddrSpace
 from tinygrad.helpers import all_same, colored, ansilen, dedup, prod, round_up, to_function_name, unwrap, argfix, DEBUG, TC_SELECT, TC_OPT, AMX
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import strides_for_shape, get_contraction
-from tinygrad.opt.swizzler import view_left
+from tinygrad.opt.swizzler import view_left, view_left_through_load
 
 class OptOps(Enum):
   TC = auto(); UPCAST = auto(); UNROLL = auto(); LOCAL = auto() # noqa: E702
@@ -503,4 +503,4 @@ class Kernel:
     self.finalized = True
     fixed_ast = fixup_ast(self.ast)
     del fixup_ast
-    return graph_rewrite(fixed_ast, view_left, name="fixup optimized AST")
+    return graph_rewrite(fixed_ast, view_left+view_left_through_load, name="fixup optimized AST")

tinygrad/opt/swizzler.py

@@ -44,12 +44,18 @@ def reduce_push_add_ones(src:UOp, r:UOp, view:UOp):
 
 view_left = merge_views+PatternMatcher([
   # view before elementwise and buffer ops
-  (UPat(Ops.VIEW, src=(UPat({*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.BIND, Ops.LOAD, Ops.STORE, Ops.VALID, Ops.SINK}, name="e"),), name="view"),
+  (UPat(Ops.VIEW, src=(UPat({*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.BIND, Ops.STORE, Ops.VALID, Ops.SINK}, name="e"),), name="view"),
    lambda e,view: e.replace(src=tuple(s.view(view.st) for s in e.src))),
   # if there's ones added after reduce, put this before the reduce
   (UPat(Ops.VIEW, src=(UPat(Ops.REDUCE_AXIS, src=(UPat.var("src"),), name="r"),), name="view"), reduce_push_add_ones),
 ])
 
+view_left_through_load = PatternMatcher([
+  # view before load
+  (UPat(Ops.VIEW, src=(UPat(Ops.LOAD, name="e"),), name="view"),
+   lambda e,view: e.replace(src=tuple(s.view(view.st) for s in e.src))),
+])
+
 def apply_swizzle(u:UOp) -> UOp: return graph_rewrite(u, view_left, name="Sub View Left")
 
 # change reduceop axes and input ShapeTrackers, view gets replaced with a reshape.
@@ -96,7 +102,7 @@ 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, Ops.STORE}, 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),
@@ -112,9 +118,9 @@ 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),
+fix_kernel_ops = view_left_through_load+PatternMatcher([
+  # add view to LOAD
+  (UPat(Ops.DEFINE_GLOBAL, name="g").load(), lambda g: g.view(g.st).load()),
   # 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)])),

tinygrad/schedule/grouper.py

@@ -3,7 +3,7 @@ from tinygrad.helpers import all_int, prod, unwrap, dedup, DONT_REALIZE_EXPAND,
 from tinygrad.shape.shapetracker import ShapeTracker
 
 ALWAYS_CONTIGUOUS = {Ops.CONTIGUOUS, Ops.ASSIGN, Ops.COPY, Ops.BUFFER, Ops.BUFFER_VIEW,
-                     Ops.CONST, Ops.BIND, Ops.DEVICE, Ops.MSELECT, Ops.MSTACK, Ops.DEFINE_GLOBAL}
+                     Ops.CONST, Ops.BIND, Ops.DEVICE, Ops.MSELECT, Ops.MSTACK, Ops.DEFINE_GLOBAL, Ops.LOAD}
 
 # **** Grouper decides which of the UOps realize
 

tinygrad/schedule/kernelize.py

@@ -7,7 +7,7 @@ from tinygrad.helpers import Metadata, all_int, all_same, prod, dedup, unwrap, g
 from tinygrad.dtype import ImageDType
 from tinygrad.schedule.multi import multi_pm
 from tinygrad.schedule.grouper import group_realizes, ALWAYS_CONTIGUOUS
-from tinygrad.opt.swizzler import merge_views, apply_swizzle, swizzle_reduceop
+from tinygrad.opt.swizzler import merge_views, view_left, view_right, fix_kernel_ops, apply_swizzle, swizzle_reduceop
 
 # creation can recurse a lot
 import sys
@@ -149,18 +149,15 @@ create_kernels = PatternMatcher([
 
 # **** fix kernel AST
 
-early_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.SINK, src=(UPat(Ops.CONTIGUOUS, src=(UPat(GroupOp.Meta, name="x"),),))), lambda x:x),
-])
-
-replace_globals = PatternMatcher([
+replace_buffers = PatternMatcher([
   # replace ASSIGN with the target BUFFER
-  (UPat(Ops.ASSIGN, src=(UPat(Ops.BUFFER), UPat(Ops.KERNEL)), name="assign", allow_any_len=True), lambda assign: assign.src[0]),
+  (UPat(Ops.ASSIGN, src=(UPat((Ops.BUFFER, Ops.LOAD)), UPat(Ops.KERNEL)), name="assign", allow_any_len=True), lambda assign: assign.src[0]),
   # HACK: select the 0 branch of MSTACK (the device is wrong after this, is that okay?)
   (UPat(Ops.MSTACK, name="x"), lambda x: x.src[0]),
+  # LOAD
+  (UPat(Ops.BUFFER, name="x"), lambda ctx,x: UOp(Ops.DEFINE_GLOBAL, x.dtype.ptr(x.size), (), ctx.index(x)).load()),
+  # no SINK for meta ops
+  (UPat(Ops.SINK, src=(UPat(Ops.CONTIGUOUS, src=(UPat(GroupOp.Meta, name="x"),),))), lambda x:x),
 ])
 
 def fix_kernel_ast(k:UOp) -> UOp|None:
@@ -173,10 +170,13 @@ def fix_kernel_ast(k:UOp) -> UOp|None:
     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(k.arg.ast, replace_buffers, bufs, bottom_up=True, name="replace buffers")
   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)}")
+  # TODO: move these to codegen
+  ast = graph_rewrite(ast, view_left, name="Main View Left")
+  ast = graph_rewrite(ast, view_right, name="Main View Right")
+  ast = graph_rewrite(ast, view_left+fix_kernel_ops, bottom_up=True, name="Finalize Kernel")
   return k.replace(arg=Kernel(ast, k.arg.metadata))
 
 create_ast = PatternMatcher([(UPat(Ops.KERNEL, name="k"), fix_kernel_ast),])

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