Move custom GPU atomics into separate namespace. Add a script to print the include path.

• Move the custom atomicMin/Max/Mul functions to namespace pystencils::runtime::gpu, such that no name conflicts may ever arise with native CUDA/HIP functions
  • Other than CUDA, HIP defines atomicMin, atomicMax for floating-point types, so our definitions were raising duplicate-definition errors
• Use C++17 in cupy JIT compiler (for qualified namespace names)
• Make the pystencils.include module executable as a script to print the path to the pystencils header directory to stdout

src/pystencils/backend/platforms/cuda.py

@@ -109,8 +109,14 @@ class CudaPlatform(GenericGpu):
             cond = is_valid_thread
 
         # use atomic operation
+        match actual_reduction_op:
+            case ReductionOp.Min | ReductionOp.Max | ReductionOp.Mul:
+                impl_namespace = "pystencils::runtime::gpu::"
+            case _:
+                impl_namespace = ""
+
         func = CFunction(
-            f"atomic{actual_reduction_op.name}", [ptrtype, stype], PsCustomType("void")
+            f"{impl_namespace}atomic{actual_reduction_op.name}", [ptrtype, stype], PsCustomType("void")
         )
         func_args = (ptr_expr, symbol_expr)