fix et-view (pytorch#2843)

Summary:

et-view should always copy the data pointer.

Reviewed By: JacobSzwejbka

Differential Revision: D55715318

Author: metascroy

kernels/prim_ops/et_view.cpp

@@ -87,18 +87,7 @@ void et_view(RuntimeContext& context, EValue** stack) {
   // Do some checks
   ET_CHECK(self.numel() == out.numel());
 
-  // If out has a data_ptr, it must match self
-  // We hit this path for memory-planned tensors
-  if (out.const_data_ptr() != nullptr) {
-    ET_CHECK_MSG(
-        self.const_data_ptr() == out.const_data_ptr(),
-        "out has a non-null data_ptr, but it does not equal self's data_ptr.");
-
-    // nothing else to do
-    return;
-  }
-
-  // out.const_data_ptr() == nullptr now
+  // Update data ptr
   ET_CHECK_MSG(
       internal::set_tensor_data(
           out,

kernels/prim_ops/test/prim_ops_test.cpp

@@ -331,14 +331,13 @@ TEST_F(RegisterPrimOpsTest, TestETView) {
       EValue(good_outs[0]), EValue(good_outs[1])};
 
   // bad outs expect death
-  constexpr int N_BAD_OUTS = 3;
+  constexpr int N_BAD_OUTS = 2;
   Tensor bad_outs[N_BAD_OUTS] = {
       tf.ones({1, 3, 2, 1}), // wrong rank
-      tf.ones({1, 3, 3}), // wrong size
-      tf.ones({1, 3, 2}) // occupied data_ptr
+      tf.ones({1, 3, 3}) // wrong size
   };
   EValue bad_out_evalues[N_BAD_OUTS] = {
-      EValue(bad_outs[0]), EValue(bad_outs[1]), EValue(bad_outs[2])};
+      EValue(bad_outs[0]), EValue(bad_outs[1])};
 
   // ***************************************************************************
   // Run tests
@@ -349,7 +348,6 @@ TEST_F(RegisterPrimOpsTest, TestETView) {
       // Bad out stacks
       {&self_evalue, &size_int_list_evalue, &bad_out_evalues[0]},
       {&self_evalue, &size_int_list_evalue, &bad_out_evalues[1]},
-      {&self_evalue, &size_int_list_evalue, &bad_out_evalues[2]},
       // Bad size stacks
       {&self_evalue, &bad_size_int_list_evalue1, &good_out_evalues[0]},
       {&self_evalue, &bad_size_int_list_evalue2, &good_out_evalues[0]}};