fix strided_slice_op's GetExpectedKernelType (#28192)

* fix strided_slice_op's GetExpectedKernelType when input tensor is at CUDAPinnedPlace

* add unittest for tensors in cuda pinned place

* skip test for cuda pinned place on cpu machines

paddle/fluid/operators/strided_slice_op.cc

@@ -154,9 +154,14 @@ class StridedSliceOp : public framework::OperatorWithKernel {
  protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
+    // NOTE: cuda pinned tensor need to copy its data to target place
+    auto in_tensor = ctx.Input<Tensor>("Input");
+    if (platform::is_cuda_pinned_place(in_tensor->place())) {
+      return framework::OpKernelType(in_tensor->type(), ctx.device_context());
+    }
     return framework::OpKernelType(
         OperatorWithKernel::IndicateVarDataType(ctx, "Input"),
-        ctx.Input<Tensor>("Input")->place());
+        in_tensor->place());
   }
   framework::OpKernelType GetKernelTypeForVar(
       const std::string &var_name, const Tensor &tensor,

python/paddle/fluid/tests/unittests/test_strided_slice_op.py

@@ -511,6 +511,17 @@ class TestStridedSliceAPI(unittest.TestCase):
             x, axes=axes, starts=starts, ends=ends, strides=strides_1)
         assert sliced_1.shape == (3, 2, 2, 2)
 
+    @unittest.skipIf(not paddle.is_compiled_with_cuda(),
+                     "Cannot use CUDAPinnedPlace in CPU only version")
+    def test_cuda_pinned_place(self):
+        with paddle.fluid.dygraph.guard():
+            x = paddle.to_tensor(
+                np.random.randn(2, 10), place=paddle.CUDAPinnedPlace())
+            self.assertTrue(x.place.is_cuda_pinned_place())
+            y = x[:, ::2]
+            self.assertFalse(x.place.is_cuda_pinned_place())
+            self.assertFalse(y.place.is_cuda_pinned_place())
+
 
 if __name__ == "__main__":
     unittest.main()