[MPSInductor] Implement bitcasts (pytorch#144638)

That will be used to compile something like `torch.rand(32, device='mps').view(dtype=torch.int32)`

Pull Request resolved: pytorch#144638
Approved by: https://github.com/dcci

test/inductor/test_mps_basic.py

@@ -61,6 +61,7 @@ class MPSBasicTests(TestCase):
     test_tanh = CommonTemplate.test_tanh
     test_view_as_complex = CommonTemplate.test_view_as_complex
     test_views6 = CommonTemplate.test_views6
+    test_views7 = CommonTemplate.test_views7
     test_zero_dim_reductions = CommonTemplate.test_zero_dim_reductions
 
     @parametrize("dtype", MPS_DTYPES)

torch/_inductor/codegen/mps.py

@@ -74,6 +74,12 @@ def to_dtype(
     ) -> str:
         return f"static_cast<{DTYPE_TO_METAL[dtype]}>({x})"
 
+    @staticmethod
+    def to_dtype_bitcast(
+        x: CSEVariable, dtype: torch.dtype, src_dtype: torch.dtype
+    ) -> str:
+        return f"*reinterpret_cast<thread {DTYPE_TO_METAL[dtype]}*>(&{x})"
+
     @staticmethod
     def constant(val: CSEVariable, dtype: torch.dtype) -> str:
         return value_to_metal(val)