improve cast to unsign min_max [pr] (#17078)

test/null/test_uop_vmin_vmax.py

@@ -127,6 +127,13 @@ class TestVminVmaxProperties(unittest.TestCase):
     self.assertEqual(x.vmin, 0)
     self.assertEqual(x.vmax, 10 >> 2)
 
+  def test_vmin_vmax_cast_unsigned(self):
+    # a fitting source keeps exact bounds: no wrap can occur
+    self.assertEqual(UOp.variable('x', 5, 10).cast(dtypes.uint8)._min_max, (5, 10))
+    # a possibly-negative or too-large source can wrap: conservative
+    self.assertEqual(UOp.variable('x', -1, 10).cast(dtypes.uint8)._min_max, (0, 255))
+    self.assertEqual(UOp.variable('x', 250, 260).cast(dtypes.uint8)._min_max, (0, 255))
+
   def test_vmin_vmax_xor_neg1(self):
     x = UOp.variable('x', 3, 7)
     uop = x ^ -1

tinygrad/uop/ops.py

@@ -1029,9 +1029,12 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
     if self.op is Ops.STACK: return min(x.vmin for x in self.src), max(x.vmax for x in self.src)
     if self.op is Ops.CONST and self.arg is not Invalid: return self.arg, self.arg
     if self.op is Ops.INDEX: return self.src[0]._min_max
-    # TODO: CAST to bool/unsigned is not monotone, still some case can be simplified
-    if self.op is Ops.CAST and self.dtype in dtypes.floats+dtypes.sints+(dtypes.index,):
-      return max(self.dtype.min, self.src[0].vmin), min(self.src[0].vmax, self.dtype.max)
+    if self.op is Ops.CAST:
+      # a cast to unsigned keeps exact bounds when the source fits
+      # TODO: can do more based on new dtype window
+      if dtypes.is_unsigned(self.dtype) and 0 <= self.src[0].vmin and self.src[0].vmax <= self.dtype.max: return self.src[0]._min_max
+      if self.dtype in dtypes.floats+dtypes.sints+(dtypes.index,):
+        return max(self.dtype.min, self.src[0].vmin), min(self.src[0].vmax, self.dtype.max)
     return self.dtype.min, self.dtype.max
 
   @functools.cached_property