Add gpu support for unittest

paddle/fluid/operators/math/sequence_pooling_test.cc

@@ -19,17 +19,18 @@ limitations under the License. */
 template <typename DeviceContext, typename Place, typename T>
 void TestSequencePoolingSum(const paddle::framework::LoD& lod) {
   paddle::framework::LoDTensor cpu_out_grad;
+  paddle::framework::LoDTensor cpu_in_grad;
   paddle::framework::LoDTensor out_grad;
   paddle::framework::LoDTensor in_grad;
   const size_t second_dim = 128u;
 
   // construct out_grad's tensor in cpu
-  const size_t out_first_dim = lod.size() - 1;
+  const size_t out_first_dim = lod[0].size() - 1;
   auto out_dims = paddle::framework::make_ddim(
       {static_cast<int64_t>(out_first_dim), static_cast<int64_t>(second_dim)});
 
   cpu_out_grad.mutable_data<T>(out_dims, paddle::platform::CPUPlace());
-  for (int64_t i = 0; i < out_grad.numel(); ++i) {
+  for (int64_t i = 0; i < cpu_out_grad.numel(); ++i) {
     cpu_out_grad.data<T>()[i] = static_cast<T>(i);
   }
 
@@ -58,9 +59,18 @@ void TestSequencePoolingSum(const paddle::framework::LoD& lod) {
   paddle::operators::math::SequencePoolGradFunctor<DeviceContext, T>()(
       *context, "SUM", out_grad, &in_grad);
 
+  if (paddle::platform::is_cpu_place(*place)) {
+    cpu_in_grad = in_grad;
+  } else {
+    TensorCopySync(in_grad, paddle::platform::CPUPlace(), &cpu_in_grad);
+    cpu_in_grad.set_lod(in_grad.lod());
+  }
+
   EXPECT_EQ(in_grad.numel(), lod[0].back() * second_dim);
   EXPECT_EQ(in_grad.lod(), lod);
-  for (int64_t i = 0; i < in_grad.lod()[0].size() - 1; ++i) {
+
+  if (paddle::platform::is_cpu_place(*place)) {
+    for (int64_t i = 0; i < cpu_in_grad.lod()[0].size() - 1; ++i) {
       int64_t begin = in_grad.lod()[0][i];
       int64_t end = in_grad.lod()[0][i + 1];
       paddle::framework::Tensor tmp = in_grad.Slice(begin, end);
@@ -71,6 +81,19 @@ void TestSequencePoolingSum(const paddle::framework::LoD& lod) {
         }
       }
     }
+  } else {
+    for (int64_t i = 0; i < cpu_in_grad.lod()[0].size() - 1; ++i) {
+      int64_t begin = cpu_in_grad.lod()[0][i];
+      int64_t end = cpu_in_grad.lod()[0][i + 1];
+      paddle::framework::Tensor tmp = cpu_in_grad.Slice(begin, end);
+      for (int64_t j = 0; j != tmp.numel() / second_dim; ++j) {
+        for (int64_t m = 0; m != second_dim; ++m) {
+          EXPECT_EQ(tmp.data<T>()[m + j * second_dim],
+                    cpu_out_grad.data<T>()[m + i * second_dim]);
+        }
+      }
+    }
+  }
 
   delete place;
   delete context;