From d42cc8d10d65921b6412e4f7b2fe38caa57de8db Mon Sep 17 00:00:00 2001
From: zhupengyang <zhu_py@qq.com>
Date: Thu, 7 May 2020 13:53:45 +0800
Subject: [PATCH] [NPU] init device tensor when needed (#3552)

---
 lite/kernels/npu/bridges/engine.cc   |  3 ++
 lite/kernels/npu/bridges/engine.h    |  1 +
 lite/kernels/npu/subgraph_compute.cc | 44 +++++++++++++++++++---------
 lite/kernels/npu/subgraph_compute.h  |  4 +++
 lite/tools/ci_build.sh               |  2 +-
 5 files changed, 39 insertions(+), 15 deletions(-)

diff --git a/lite/kernels/npu/bridges/engine.cc b/lite/kernels/npu/bridges/engine.cc
index 6e639a37ba..8ca8357710 100644
--- a/lite/kernels/npu/bridges/engine.cc
+++ b/lite/kernels/npu/bridges/engine.cc
@@ -95,6 +95,8 @@ int Engine::Build() {
   return build_device_program_status_;
 }
 
+void Engine::InitDeviceTensor() { return; }
+
 bool Engine::InputShapeChanged() {
   for (size_t i = 0; i < origin_itensors_.size(); i++) {
     if (origin_itensors_[i]->dims() != origin_idims_[i]) {
@@ -110,6 +112,7 @@ int Engine::Launch() {
       CHECK_REBUILD_WHEN_SHAPE_CHANGED(build_device_program_status_) &&
       InputShapeChanged()) {
     Build();
+    InitDeviceTensor();
   }
   if (CHECK_FAILED(build_device_program_status_)) {
     LaunchOriginProgram();
diff --git a/lite/kernels/npu/bridges/engine.h b/lite/kernels/npu/bridges/engine.h
index 61a4e12cf3..34ec923889 100644
--- a/lite/kernels/npu/bridges/engine.h
+++ b/lite/kernels/npu/bridges/engine.h
@@ -55,6 +55,7 @@ class Engine {
   virtual int BuildOriginProgram();
   virtual int LaunchOriginProgram();
 
+  virtual void InitDeviceTensor();
   virtual bool InputShapeChanged();
 
   KernelContext *ctx_{nullptr};
diff --git a/lite/kernels/npu/subgraph_compute.cc b/lite/kernels/npu/subgraph_compute.cc
index 1baa5a0de4..1a991bfc74 100644
--- a/lite/kernels/npu/subgraph_compute.cc
+++ b/lite/kernels/npu/subgraph_compute.cc
@@ -195,18 +195,6 @@ int SubgraphEngine::LaunchDeviceProgram() {
   // Copy the data of origin input tensors to the buffer of input HiAI tensors
   // init device_itensors_, device_otensors_, origin_otensors_
   auto device_program = device_program_map_[inputs_shape_];
-  for (size_t i = 0; i < device_itensors_.size(); i++) {
-    device_itensors_[i]->Init(&(device_program->device_idims[i]));
-    std::memcpy(device_itensors_[i]->GetBuffer(),
-                origin_itensors_[i]->raw_data(),
-                origin_itensors_[i]->memory_size());
-  }
-  for (size_t i = 0; i < device_otensors_.size(); i++) {
-    device_otensors_[i]->Init(&(device_program->device_odims[i]));
-  }
-  for (size_t i = 0; i < origin_otensors_.size(); i++) {
-    origin_otensors_[i]->Resize(device_program->origin_odims[i]);
-  }
 
   // Run the HiAI model by name
   std::string key = "model_name";  // Note: key seems must be model_name
@@ -233,15 +221,43 @@ int SubgraphEngine::LaunchDeviceProgram() {
   return 0;
 }
 
+int SubgraphEngine::Build() {
+  if (device_program_map_.count(inputs_shape_) > 0) {
+    return subgraph::SUCCESS;
+  }
+  // In order to attach all of the ops of the block desc, we need to build the
+  // original program firstly.
+  BuildOriginProgram();
+  // Run InferShape() of all of ops, and convert Paddle ops to NPU/XPU IR graph
+  build_device_program_status_ = BuildDeviceProgram();
+  return build_device_program_status_;
+}
+
+void SubgraphEngine::InitDeviceTensor() {
+  auto device_program = device_program_map_[inputs_shape_];
+  for (size_t i = 0; i < device_itensors_.size(); i++) {
+    device_itensors_[i]->Init(&(device_program->device_idims[i]));
+    std::memcpy(device_itensors_[i]->GetBuffer(),
+                origin_itensors_[i]->raw_data(),
+                origin_itensors_[i]->memory_size());
+  }
+  for (size_t i = 0; i < device_otensors_.size(); i++) {
+    device_otensors_[i]->Init(&(device_program->device_odims[i]));
+  }
+  for (size_t i = 0; i < origin_otensors_.size(); i++) {
+    origin_otensors_[i]->Resize(device_program->origin_odims[i]);
+  }
+}
+
 bool SubgraphEngine::InputShapeChanged() {
   std::vector<std::vector<int64_t>> new_shape;
   for (auto origin_itensor : origin_itensors_) {
     new_shape.push_back(origin_itensor->dims().Vectorize());
   }
-  inputs_shape_ = new_shape;
-  if (device_program_map_.count(inputs_shape_) > 0) {
+  if (inputs_shape_ == new_shape) {
     return false;
   }
+  inputs_shape_ = new_shape;
   return true;
 }
 
diff --git a/lite/kernels/npu/subgraph_compute.h b/lite/kernels/npu/subgraph_compute.h
index 801f61b036..db84fc1883 100644
--- a/lite/kernels/npu/subgraph_compute.h
+++ b/lite/kernels/npu/subgraph_compute.h
@@ -49,9 +49,13 @@ class SubgraphEngine : public subgraph::Engine {
     std::vector<hiai::TensorDimension> device_odims{};
   };
 
+  int Build() override;
+
  protected:
   int BuildDeviceProgram() override;
   int LaunchDeviceProgram() override;
+
+  void InitDeviceTensor() override;
   bool InputShapeChanged() override;
 
   std::string model_name_{"model.om"};
diff --git a/lite/tools/ci_build.sh b/lite/tools/ci_build.sh
index 270c3cf79c..cda8bbd4e0 100755
--- a/lite/tools/ci_build.sh
+++ b/lite/tools/ci_build.sh
@@ -396,7 +396,7 @@ function test_arm_android {
 
     adb -s ${device} push ${testpath} ${adb_work_dir}
     adb -s ${device} shell "cd ${adb_work_dir} && ./${test_name}"
-    adb -s ${device} shell "rm ${adb_work_dir}/${test_name}"
+    adb -s ${device} shell "rm -f ${adb_work_dir}/${test_name}"
 }
 
 # test_npu <some_test_name> <adb_port_number>
-- 
GitLab