From df000116ea6d2da1a8dd840a0cd951553922a889 Mon Sep 17 00:00:00 2001
From: George Hotz <geohot@gmail.com>
Date: Tue, 2 Jun 2026 13:02:28 -0700
Subject: [PATCH] more renderer cleanups

---
 tinygrad/dtype.py             |  4 ++--
 tinygrad/renderer/__init__.py | 12 ++++++------
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/tinygrad/dtype.py b/tinygrad/dtype.py
index a76c8b6da7..73925ed683 100644
--- a/tinygrad/dtype.py
+++ b/tinygrad/dtype.py
@@ -56,7 +56,7 @@ class AddrSpace(Enum):
 @dataclass(frozen=True, eq=False)
 class DType(metaclass=DTypeMetaClass):
   priority: int  # this determines when things get upcasted
-  bitsize: int
+  bitsize: int   # this is the bitsize of the base dtype
   name: str
   fmt: FmtStr|None
   count: int
@@ -76,7 +76,7 @@ class DType(metaclass=DTypeMetaClass):
   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)
+    return DType(self.priority, self.bitsize, 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
diff --git a/tinygrad/renderer/__init__.py b/tinygrad/renderer/__init__.py
index 9552b27700..f8f15c146c 100644
--- a/tinygrad/renderer/__init__.py
+++ b/tinygrad/renderer/__init__.py
@@ -3,7 +3,7 @@ from typing import Callable, cast
 from dataclasses import dataclass
 from tinygrad.helpers import prod, Target, EMULATED_DTYPES
 from tinygrad.uop.ops import Ops, UOp, sint, ssimplify, smin, GroupOp, PatternMatcher
-from tinygrad.dtype import AddrSpace, PtrDType, DType, dtypes
+from tinygrad.dtype import AddrSpace, DType, dtypes
 from tinygrad.codegen.opt.tc import TensorCore
 from tinygrad.device import Compiler
 
@@ -41,7 +41,7 @@ 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].dtype.base.itemsize * mults
+          accessed = mem.get((buf, u.op), 0) + u.max_numel() * u.src[0].dtype.itemsize * mults
           mem[(buf, u.op)] = smin(accessed, buf.max_numel() * buf.dtype.itemsize)
       if u.op is Ops.RANGE:
         mult_stack.append(mults)
@@ -51,10 +51,10 @@ class Estimates:
       elif u.op is Ops.END: mults = mult_stack.pop(-1)
       elif u.op is Ops.SPECIAL: mults *= cast(sint, u.src[0].ssimplify()) # NOTE: we don't push to the mult_stack here, you can't end these
       elif u.op is Ops.DEFINE_VAR and u.arg[0] == 'core_id': mults *= u.arg[2] + 1
-      elif u.op is Ops.LOAD and (not isinstance(u.src[0].dtype, PtrDType) or u.src[0].dtype.addrspace != AddrSpace.REG):
-        lds += u.dtype.itemsize * mults
-      elif u.op is Ops.STORE and (not isinstance(u.src[0].dtype, PtrDType) or u.src[0].dtype.addrspace != AddrSpace.REG):
-        lds += u.src[1].dtype.itemsize * mults
+      elif u.op is Ops.LOAD and u.src[0].addrspace != AddrSpace.REG:
+        lds += u.max_numel() * u.dtype.itemsize * mults
+      elif u.op is Ops.STORE and u.src[0].addrspace != AddrSpace.REG:
+        lds += u.max_numel() * u.src[1].dtype.itemsize * mults
       elif u.op in GroupOp.ALU and u not in dont_count: flops += (mults * (2 if u.op is Ops.MULACC else 1)) * u.dtype.count
       elif u.op is Ops.WMMA and u not in dont_count: flops += 2 * prod(u.arg[1]) // u.arg[5] * mults
     return Estimates(flops, lds, sum(mem.values()))