From ca174d49104210cbc334a5ec0b9b2e66586e155f Mon Sep 17 00:00:00 2001
From: xuyongfei <xuyongfei.huawei.com>
Date: Sun, 27 Dec 2020 10:50:22 +0800
Subject: [PATCH] serving, avoid output memory copy, performance

---
 mindspore_serving/ccsrc/common/buffer_tensor.h  | 11 +++++++++++
 mindspore_serving/ccsrc/worker/work_executor.cc | 13 +++++++------
 third_party/mindspore                           |  2 +-
 3 files changed, 19 insertions(+), 7 deletions(-)

diff --git a/mindspore_serving/ccsrc/common/buffer_tensor.h b/mindspore_serving/ccsrc/common/buffer_tensor.h
index 9a6b4e3..874d440 100644
--- a/mindspore_serving/ccsrc/common/buffer_tensor.h
+++ b/mindspore_serving/ccsrc/common/buffer_tensor.h
@@ -57,6 +57,17 @@ class BufferTensor : public TensorBase {
   bool data_readonly_ = false;
 };
 
+class BufferTensorWithOwner : public BufferTensor {
+ public:
+  BufferTensorWithOwner(TensorBasePtr buffer_tensor_owner, DataType type, std::vector<int64_t> shape, uint8_t *data,
+                        size_t data_len, bool data_readonly)
+      : BufferTensor(type, shape, data, data_len, data_readonly), buffer_tensor_owner_(buffer_tensor_owner) {}
+  ~BufferTensorWithOwner() = default;
+
+ private:
+  TensorBasePtr buffer_tensor_owner_;
+};
+
 }  // namespace mindspore::serving
 
 #endif  // MINDSPORE_SERVING_BUFFER_TENSOR_H
diff --git a/mindspore_serving/ccsrc/worker/work_executor.cc b/mindspore_serving/ccsrc/worker/work_executor.cc
index 68bfe08..4d8eab1 100644
--- a/mindspore_serving/ccsrc/worker/work_executor.cc
+++ b/mindspore_serving/ccsrc/worker/work_executor.cc
@@ -288,7 +288,10 @@ bool WorkExecutor::ReplyRequest(const std::vector<Instance> &outputs) {
 bool WorkExecutor::ReplyRequest(const Instance &outputs) {
   Status status;
   Instance trans_outputs = CreateInputInstance(outputs, kPredictPhaseTag_Output);
-  outputs.context.user_context->worker_call_back(trans_outputs, Status(SUCCESS));
+  Instance real_outputs;
+  real_outputs.data = trans_outputs.data;
+  real_outputs.context = trans_outputs.context;
+  outputs.context.user_context->worker_call_back(real_outputs, Status(SUCCESS));
   outputs.context.promise->set_value();
   return true;
 }
@@ -394,12 +397,10 @@ Status WorkExecutor::PostPredict(const std::vector<Instance> &inputs, const std:
       }
       shape.erase(shape.begin());
     }
-    auto src_buffer = reinterpret_cast<const uint8_t *>(item->data());
+    auto src_buffer = const_cast<uint8_t *>(item->data());
     for (uint32_t k = 0; k < input_batch_size; k++) {
-      auto tensor = std::make_shared<Tensor>();
-      tensor->set_shape(shape);
-      tensor->set_data_type(item->data_type());
-      tensor->set_data(src_buffer + item_size * k, item_size);
+      auto tensor = std::make_shared<BufferTensorWithOwner>(item, item->data_type(), shape, src_buffer + item_size * k,
+                                                            item_size, true);
       results_data[k].data.push_back(tensor);
     }
   }
diff --git a/third_party/mindspore b/third_party/mindspore
index f79e3bd..7a3f342 160000
--- a/third_party/mindspore
+++ b/third_party/mindspore
@@ -1 +1 @@
-Subproject commit f79e3bdb39b4de3fca1dc973feb2fac9f3cb13bb
+Subproject commit 7a3f34247db5a6c7a00b00b16d2bfa2ff94e0b08