fea/lightly support lod (#12451)

Author: Superjomn

paddle/fluid/inference/api/api_anakin_engine_tester.cc

@@ -37,19 +37,21 @@ TEST(inference, anakin) {
 
   float data[1 * 3 * 224 * 224] = {1.0f};
 
-  PaddleTensor tensor{.name = "input_0",
-                      .shape = std::vector<int>({1, 3, 224, 224}),
-                      .data = PaddleBuf(data, sizeof(data)),
-                      .dtype = PaddleDType::FLOAT32};
+  PaddleTensor tensor;
+  tensor.name = "input_0";
+  tensor.shape = std::vector<int>({1, 3, 224, 224});
+  tensor.data = PaddleBuf(data, sizeof(data));
+  tensor.dtype = PaddleDType::FLOAT32;
 
   // For simplicity, we set all the slots with the same data.
   std::vector<PaddleTensor> paddle_tensor_feeds;
   paddle_tensor_feeds.emplace_back(std::move(tensor));
 
-  PaddleTensor tensor_out{.name = "prob_out",
-                          .shape = std::vector<int>({1000, 1}),
-                          .data = PaddleBuf(),
-                          .dtype = PaddleDType::FLOAT32};
+  PaddleTensor tensor_out;
+  tensor_out.name = "prob_out";
+  tensor_out.shape = std::vector<int>({1000, 1});
+  tensor_out.data = PaddleBuf();
+  tensor_out.dtype = PaddleDType::FLOAT32;
 
   std::vector<PaddleTensor> outputs;
   outputs.emplace_back(std::move(tensor_out));

paddle/fluid/inference/api/api_impl.cc

@@ -183,6 +183,13 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
     // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
     std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
                 inputs[i].data.length());
+    // TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
+    framework::LoD lod;
+    for (auto &level : inputs[i].lod) {
+      lod.emplace_back(level);
+    }
+    input.set_lod(lod);
+
     feeds->push_back(input);
   }
   return true;
@@ -248,6 +255,10 @@ bool NativePaddlePredictor::GetFetch(
       buffer.Resize(sizeof(float) * data.size());
     }
     std::memcpy(buffer.data(), data.data(), buffer.length());
+    // copy LoD
+    for (const auto &level : fetchs[i].lod()) {
+      outputs->at(i).lod.emplace_back(level);
+    }
     outputs->at(i).dtype = PaddleDType::FLOAT32;
     // TODO(panyx0718): support other types? fill tensor name? avoid a copy.
   }

paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc

@@ -49,11 +49,10 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
     std::vector<int64_t> data(20);
     for (int i = 0; i < 20; i++) data[i] = i;
 
-    PaddleTensor tensor{
-        .name = "",
-        .shape = std::vector<int>({10, 1}),
-        .data = PaddleBuf(data.data(), data.size() * sizeof(int64_t)),
-        .dtype = PaddleDType::INT64};
+    PaddleTensor tensor;
+    tensor.shape = std::vector<int>({10, 1});
+    tensor.data = PaddleBuf(data.data(), data.size() * sizeof(int64_t));
+    tensor.dtype = PaddleDType::INT64;
 
     // For simplicity, we set all the slots with the same data.
     std::vector<PaddleTensor> slots(4, tensor);

paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc

@@ -47,10 +47,10 @@ void Main(bool use_gpu) {
     //# 2. Prepare input.
     int64_t data[4] = {1, 2, 3, 4};
 
-    PaddleTensor tensor{.name = "",
-                        .shape = std::vector<int>({4, 1}),
-                        .data = PaddleBuf(data, sizeof(data)),
-                        .dtype = PaddleDType::INT64};
+    PaddleTensor tensor;
+    tensor.shape = std::vector<int>({4, 1});
+    tensor.data = PaddleBuf(data, sizeof(data));
+    tensor.dtype = PaddleDType::INT64;
 
     // For simplicity, we set all the slots with the same data.
     std::vector<PaddleTensor> slots(4, tensor);
@@ -94,10 +94,11 @@ void MainThreads(int num_threads, bool use_gpu) {
       for (int batch_id = 0; batch_id < num_batches; ++batch_id) {
         // 2. Dummy Input Data
         int64_t data[4] = {1, 2, 3, 4};
-        PaddleTensor tensor{.name = "",
-                            .shape = std::vector<int>({4, 1}),
-                            .data = PaddleBuf(data, sizeof(data)),
-                            .dtype = PaddleDType::INT64};
+        PaddleTensor tensor;
+        tensor.shape = std::vector<int>({4, 1});
+        tensor.data = PaddleBuf(data, sizeof(data));
+        tensor.dtype = PaddleDType::INT64;
+
         std::vector<PaddleTensor> inputs(4, tensor);
         std::vector<PaddleTensor> outputs;
         // 3. Run

paddle/fluid/inference/api/demo_ci/vis_demo.cc

@@ -123,11 +123,11 @@ void Main(bool use_gpu) {
   file.close();
 
   // Inference.
-  PaddleTensor input{
-      .name = "xx",
-      .shape = record.shape,
-      .data = PaddleBuf(record.data.data(), record.data.size() * sizeof(float)),
-      .dtype = PaddleDType::FLOAT32};
+  PaddleTensor input;
+  input.shape = record.shape;
+  input.data =
+      PaddleBuf(record.data.data(), record.data.size() * sizeof(float));
+  input.dtype = PaddleDType::FLOAT32;
 
   VLOG(3) << "run executor";
   std::vector<PaddleTensor> output;

paddle/fluid/inference/api/paddle_inference_api.h

@@ -67,9 +67,9 @@ struct PaddleTensor {
   PaddleTensor() = default;
   std::string name;  // variable name.
   std::vector<int> shape;
-  // TODO(Superjomn) for LoD support, add a vector<vector<int>> field if needed.
   PaddleBuf data;  // blob of data.
   PaddleDType dtype;
+  std::vector<std::vector<uint64_t>> lod;  // lod data
 };
 
 enum class PaddleEngineKind {