Remove/edit some comments.

lite/api/CMakeLists.txt

@@ -87,7 +87,7 @@ else()
             target_link_libraries(paddle_light_api_shared ${rknpu_builder_libs} ${rknpu_runtime_libs})
         endif()
         if (LITE_WITH_IMAGINATION_NNA)
-            # Need to add IMG IMAGINATION_NNA runtime libs (libhiai.so) dependency
+            # Need to add IMG IMAGINATION_NNA runtime libs (libimgdnn.so, libnnasession.so) dependency
             #target_link_libraries(paddle_light_api_shared ${nna_builder_libs} ${nna_runtime_libs})
         endif()
     endif()
@@ -123,8 +123,8 @@ if(LITE_WITH_RKNPU)
 endif()
 
 if(LITE_WITH_IMAGINATION_NNA)
-    set(light_api_deps ${light_api_deps} ${nna_deps})
-    set(cxx_api_deps ${cxx_api_deps} ${nna_deps})
+    set(light_api_deps ${light_api_deps} ${imagination_nna_deps})
+    set(cxx_api_deps ${cxx_api_deps} ${imagination_nna_deps})
 endif()
 
 if(LITE_WITH_HUAWEI_ASCEND_NPU)

lite/backends/imagination_nna/imgdnn_manager.h

@@ -222,6 +222,13 @@ class ImgdnnManager {
     return desc;
   }
 
+  imgdnn_tensor_descriptor getTensorDescriptor(imgdnn_tensor tensor) {
+    imgdnn_tensor_descriptor desc;
+    err_ = imgdnnGetTensorDescriptor(tensor, &desc);
+    ASSERT(err_ != IMGDNN_SUCCESS, "GetTensorDescriptors failed!");
+    return desc;
+  }
+
   size_t getDescriptorSize(const imgdnn_tensor_descriptor *const descriptor) {
     size_t size = imgdnnGetDescriptorSize(descriptor, &err_);
     ASSERT(err_ != IMGDNN_SUCCESS, "GetDescriptorSize failed!");

lite/kernels/imagination_nna/bridges/CMakeLists.txt

@@ -9,37 +9,8 @@ set(imagination_nna_subgraph_bridge_deps subgraph_bridge_registry subgraph_bridg
 
 lite_cc_library(subgraph_bridge_fc_op_imagination_nna SRCS fc_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
 lite_cc_library(subgraph_bridge_conv_op_imagination_nna SRCS conv_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_matmul_op_imagination_nna SRCS matmul_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_mul_op_imagination_nna SRCS mul_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
 lite_cc_library(subgraph_bridge_act_op_imagination_nna SRCS act_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_scale_op_imagination_nna SRCS scale_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_softmax_op_imagination_nna SRCS softmax_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
 lite_cc_library(subgraph_bridge_pool_op_imagination_nna SRCS pool_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_batch_norm_op_imagination_nna SRCS batch_norm_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_elementwise_ops_imagination_nna SRCS elementwise_ops.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_reshape_op_imagination_nna SRCS reshape_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_conv_transpose_op_imagination_nna SRCS conv_transpose_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_interpolate_op_imagination_nna SRCS interpolate_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_transpose_op_imagination_nna SRCS transpose_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_split_op_imagination_nna SRCS split_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_concat_op_imagination_nna SRCS concat_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_shuffle_channel_op_imagination_nna SRCS shuffle_channel_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_pad2d_op_imagination_nna SRCS pad2d_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_reduce_mean_op_imagination_nna SRCS reduce_mean_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_unsqueeze_op_imagination_nna SRCS unsqueeze_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_gather_op_imagination_nna SRCS gather_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_lookup_table_op_imagination_nna SRCS lookup_table_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_argmax_op_imagination_nna SRCS argmax_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_instance_norm_op_imagination_nna SRCS instance_norm_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_dropout_op_imagination_nna SRCS dropout_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_topk_op_imagination_nna SRCS topk_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_layer_norm_op_imagination_nna SRCS layer_norm_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_fill_constant_op_imagination_nna SRCS fill_constant_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_fill_constant_batch_size_like_op_imagination_nna SRCS fill_constant_batch_size_like_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_increment_op_imagination_nna SRCS increment_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_compare_op_imagination_nna SRCS compare_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-#lite_cc_library(subgraph_bridge_shape_op_imagination_nna SRCS shape_op.cc DEPS ${imagination_nna_subgraph_bridge_deps})
-
 
 set(imagination_nna_subgraph_bridges
         subgraph_bridge_registry
@@ -47,35 +18,8 @@ set(imagination_nna_subgraph_bridges
         subgraph_bridge_graph_imagination_nna
         subgraph_bridge_fc_op_imagination_nna
         subgraph_bridge_conv_op_imagination_nna
-        #subgraph_bridge_matmul_op_imagination_nna
-        #subgraph_bridge_mul_op_imagination_nna
         subgraph_bridge_act_op_imagination_nna
-        #subgraph_bridge_scale_op_imagination_nna
-        #subgraph_bridge_softmax_op_imagination_nna
         subgraph_bridge_pool_op_imagination_nna
-        #subgraph_bridge_batch_norm_op_imagination_nna
-        #subgraph_bridge_elementwise_ops_imagination_nna
-        #subgraph_bridge_reshape_op_imagination_nna
-        #subgraph_bridge_conv_transpose_op_imagination_nna
-        #subgraph_bridge_interpolate_op_imagination_nna
-        #subgraph_bridge_transpose_op_imagination_nna
-        #subgraph_bridge_split_op_imagination_nna
-        #subgraph_bridge_concat_op_imagination_nna
-        #subgraph_bridge_shuffle_channel_op_imagination_nna
-        #subgraph_bridge_pad2d_op_imagination_nna
-        #subgraph_bridge_reduce_mean_op_imagination_nna
-        #subgraph_bridge_unsqueeze_op_imagination_nna
-        #subgraph_bridge_gather_op_imagination_nna
-        #subgraph_bridge_lookup_table_op_imagination_nna
-        #subgraph_bridge_argmax_op_imagination_nna
-        #subgraph_bridge_instance_norm_op_imagination_nna
-        #subgraph_bridge_dropout_op_imagination_nna
-        #subgraph_bridge_topk_op_imagination_nna
-        #subgraph_bridge_layer_norm_op_imagination_nna
-        #subgraph_bridge_fill_constant_op_imagination_nna
-        #subgraph_bridge_fill_constant_batch_size_like_op_imagination_nna
-        #subgraph_bridge_increment_op_imagination_nna
-        #subgraph_bridge_compare_op_imagination_nna
         CACHE INTERNAL "imagination_nna_subgraph_bridges")
 
 message(STATUS "+++++ imagination_nna_subgraph_bridges: ${imagination_nna_subgraph_bridges}")

lite/kernels/imagination_nna/bridges/batch_norm_op.cc

@@ -30,7 +30,7 @@ int BatchNormConverter(void* ctx, OpLite* op, KernelBase* kernel) {
   auto scope = op->scope();
   VLOG(3) << "[NNA] Converting " + op_type + "...";
 
-  // Get iimagination_nnat and output vars and op attributes
+  // Get input and output vars and op attributes
   auto x_name = op_info->Input("X").front();
   auto x = scope->FindMutableTensor(x_name);
   auto x_dims = x->dims();
@@ -43,24 +43,14 @@ int BatchNormConverter(void* ctx, OpLite* op, KernelBase* kernel) {
   auto variance_name = op_info->Input("Variance").front();
   auto variance = scope->FindMutableTensor(variance_name);
   auto y_name = op_info->Output("Y").front();
-  // float momentum = op_info->GetAttr<float>("momentum");
   float epsilon = op_info->GetAttr<float>("epsilon");
-  // int mode = 1;  // bnScale, bnBias tensor dims are 1xCx1x1
-  /*
-  bool use_global_stats = !op_info->HasAttr("use_global_stats") ||
-                          op_info->GetAttr<bool>("use_global_stats");
-  if (!use_global_stats) {
-    LOG(WARNING) << "[NNA] Only use_global_stats=true is supported by DDK";
-  }
-  */
 
   // X node
   std::shared_ptr<Node> x_node = nullptr;
   if (graph->Has(x_name)) {
     x_node = graph->Get(x_name);
   } else {
-    // x_node = graph->Add(x_name, *x);
-    LOG(WARNING) << "BatchNormConverter:x_node not in graph";
+    LOG(WARNING) << "[Imagination NNA] BatchNormConverter: x_node not in graph";
   }
 
   ConvNetBuilder& builder = graph->GetBuilder();
@@ -71,7 +61,6 @@ int BatchNormConverter(void* ctx, OpLite* op, KernelBase* kernel) {
   bn_out = builder.createScaleLayer(
       bn_out, true, scale->mutable_data<float>(), bias->mutable_data<float>());
 
-  // PrecisionType precision = x->precision();
   imgdnn_tensor_descriptor desc;
   imgdnn_err_code err = imgdnnGetTensorDescriptor(bn_out, &desc);
   CHECK(err == IMGDNN_SUCCESS) << "fail get tensor description(BN)";

lite/kernels/imagination_nna/bridges/graph.cc

@@ -45,7 +45,7 @@ std::shared_ptr<Node> Graph::Add(const std::string& name,
                                  const void* const const_data,
                                  std::vector<int64_t> shape,
                                  const TensorInfo& qnt,
-                                 Node::Role role /* = Node::Role::kData*/) {
+                                 Node::Role role) {
   auto node = std::make_shared<Node>(qnt.type, qnt.layout, role);
   auto idx = Add(name, node);
   CHECK_GE(idx, 1);

lite/kernels/imagination_nna/bridges/graph.h

@@ -79,28 +79,16 @@ class Graph {
  public:
   explicit Graph(lite::imagination_nna::ImgdnnManager* pMgr) {
     pImgdnnMgr = pMgr;
-    std::cout << "graph construct" << std::endl;
   }
 
-  ~Graph() { std::cout << "Graph deconst" << std::endl; }
-
-  // Add 1
-  int Add(const std::string& name, std::shared_ptr<Node> node);
-
-  // Add 2, weights,bias
+  // Add constant tensor, such as weights,bias
   std::shared_ptr<Node> Add(const std::string& name,
                             const void* const const_data,
                             std::vector<int64_t> shape,
                             const TensorInfo& qnt,
-                            Node::Role role /* = Node::Role::kData*/);
-
-  // Add 3
-  std::shared_ptr<Node> Add(const std::string& name,
-                            const Tensor& tensor,
-                            std::vector<int64_t> shape,
-                            const TensorInfo& qnt,
                             Node::Role role);
-  // Add 4
+
+  // Add input tensor
   std::shared_ptr<Node> Add(const std::string& name,
                             const Tensor& tensor,
                             const TensorInfo& qnt,
@@ -108,15 +96,14 @@ class Graph {
     return Add(name, tensor, tensor.dims().Vectorize(), qnt, role);
   }
 
-  // Used to add intermediate tensor
-  // Add 5
+  // Add intermediate activation tensor
   int Add(const std::string& name,
-          imgdnn_tensor tensor,
+          imgdnn_tensor img_tensor,
           imgdnn_type type,
           DataLayoutType layout = DATALAYOUT(kNCHW)) {
     Node::Role role = Node::Role::kData;
     auto node = std::make_shared<Node>(type, layout, role);
-    node->set_data(tensor);
+    node->set_data(img_tensor);
     return Add(name, node);  // call Add 1
   }
 
@@ -135,11 +122,19 @@ class Graph {
   }
 
  private:
+  int Add(const std::string& name, std::shared_ptr<Node> node);
+
+  std::shared_ptr<Node> Add(const std::string& name,
+                            const Tensor& tensor,
+                            std::vector<int64_t> shape,
+                            const TensorInfo& qnt,
+                            Node::Role role);
+
   std::unordered_map<std::string, std::vector<std::shared_ptr<Node>>> nodes_;
   lite::imagination_nna::ImgdnnManager* pImgdnnMgr{nullptr};
 };
-}  // namespace imagination_nna
 
+}  // namespace imagination_nna
 }  // namespace subgraph
 }  // namespace lite
 }  // namespace paddle

lite/kernels/imagination_nna/bridges/paddle_use_bridges.h

@@ -19,4 +19,3 @@ USE_SUBGRAPH_BRIDGE(conv2d, kImaginationNNA);
 USE_SUBGRAPH_BRIDGE(depthwise_conv2d, kImaginationNNA);
 USE_SUBGRAPH_BRIDGE(fc, kImaginationNNA);
 USE_SUBGRAPH_BRIDGE(pool2d, kImaginationNNA);
-// USE_SUBGRAPH_BRIDGE(softmax, kImaginationNNA);

lite/kernels/imagination_nna/bridges/pool_op.cc

@@ -97,9 +97,9 @@ int PoolConverter(void* ctx, OpLite* op, KernelBase* kernel) {
                                  ksize);
 
   // ceil mode
-  /* bool ceil_mode =
-      op_info->HasAttr("ceil_mode") && op_info->GetAttr<bool>("ceil_mode");
-  */
+  if (op_info->HasAttr("ceil_mode"))
+    LOG(WARNING) << "[NNA] imgdnn has no ceil_mode: "
+                 << op_info->GetAttr<bool>("ceil_mode");
 
   unsigned int img_ksize[2] = {(unsigned int)ksize[0], (unsigned int)ksize[1]};
   unsigned int img_stride[2] = {(unsigned int)strides[0],
@@ -126,11 +126,8 @@ int PoolConverter(void* ctx, OpLite* op, KernelBase* kernel) {
                                               pad_to_end,
                                               img_pool_type);
 
-  // LOG(INFO) << "pooling op output:" << static_cast<int>(pooling_out);
-
-  imgdnn_tensor_descriptor desc;
-  imgdnn_err_code err = imgdnnGetTensorDescriptor(pooling_out, &desc);
-  CHECK(err == IMGDNN_SUCCESS) << "fail get tensor description(POOL)";
+  imgdnn_tensor_descriptor desc =
+      graph->GetBuilder()->getTensorDescriptor(pooling_out);
 
   graph->Add(out_name, pooling_out, desc.type);
 

lite/kernels/imagination_nna/bridges/utility.cc

@@ -55,9 +55,7 @@ void TensorInfoReset(TensorInfo* qnt) {
   qnt->count = 0;
   qnt->axis = 0;
   qnt->scales.clear();
-  // qnt.scales.shrink_to_fit();
   qnt->zero_points.clear();
-  // qnt.zero_points.shrink_to_fit();
   qnt->layout = DATALAYOUT(kNCHW);
 }
 

lite/kernels/imagination_nna/subgraph_compute.cc

@@ -97,8 +97,6 @@ bool SubgraphEngine::BuildDeviceProgram() {
   imgdnn_mgr_.getNetworkObjectInputs(
       std::numeric_limits<unsigned int>::max(), nullptr, &num_inputs);
   CHECK_EQ(num_inputs, device_inames_.size());
-  // origin_idims_.resize(num_inputs);
-  // origin_itensors_.resize(num_inputs);
   device_itensors_.resize(num_inputs);
   imgdnn_mgr_.getNetworkObjectInputs(
       num_inputs, device_itensors_.data(), nullptr);
@@ -108,9 +106,6 @@ bool SubgraphEngine::BuildDeviceProgram() {
     auto node = graph.Get(device_inames_[i]);
     auto type = node->type();
     auto layout = node->layout();
-    // origin_itensors_[i] = scope_->FindMutableTensor(device_inames_[i]);
-    // CHECK(origin_itensors_[i]);
-    // origin_idims_[i] = origin_itensors_[i]->dims();
     VLOG(3) << "[NNA] Inputs[" << i << "] name: " << device_inames_[i]
             << " type: " << type << " layout: " << DataLayoutToStr(layout);
   }
@@ -119,8 +114,6 @@ bool SubgraphEngine::BuildDeviceProgram() {
   imgdnn_mgr_.getNetworkObjectOutputs(
       std::numeric_limits<unsigned int>::max(), nullptr, &num_outputs);
   CHECK_EQ(num_outputs, device_onames_.size());
-  // origin_odims_.resize(num_outputs);
-  // origin_otensors_.resize(num_outputs);
   device_otensors_.resize(num_outputs);
   imgdnn_mgr_.getNetworkObjectOutputs(
       num_outputs, device_otensors_.data(), nullptr);
@@ -129,9 +122,6 @@ bool SubgraphEngine::BuildDeviceProgram() {
     auto node = graph.Get(device_onames_[i]);
     auto type = node->type();
     auto layout = node->layout();
-    // origin_otensors_[i] = scope_->FindMutableTensor(device_onames_[i]);
-    // CHECK(origin_otensors_[i]);
-    // origin_odims_[i] = origin_otensors_[i]->dims();
     VLOG(3) << "[NNA] Outputs[" << i << "] name: " << device_onames_[i]
             << " type: " << type << " layout: " << DataLayoutToStr(layout);
     // Prepare the device output tensors
@@ -161,8 +151,10 @@ bool SubgraphEngine::BuildDeviceProgram() {
 }
 
 bool SubgraphEngine::LaunchDeviceProgram() {
-  if (!device_program_ready)  // build device program fail
+  if (!device_program_ready) {
+    LOG(WARNING) << "[NNA] Build device program fail, run origin program";
     LaunchOriginProgram();
+  }
 
   // Set input buffer
   for (size_t i = 0; i < origin_itensors_.size(); i++) {

tools/codestyle/clang_format.hook

 #!/bin/bash
 set -e
 
-readonly VERSION="6.0.0"
+readonly VERSION="3.8"
 
 version=$(clang-format -version)