Cherry-pick from 1662, 16797.. : add anakin int8 support

paddle/fluid/framework/ir/fc_fuse_pass.cc

@@ -48,8 +48,9 @@ void FCFusePass::ApplyImpl(ir::Graph* graph) const {
     GET_IR_NODE_FROM_SUBGRAPH(elementwise_add, elementwise_add, fc_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(mul_out, mul_out, fc_pattern);
 
+    auto base_op_desc = *mul->Op()->Proto();
     // Create an FC Node.
-    OpDesc desc;
+    OpDesc desc(base_op_desc, nullptr);
     std::string fc_x_in = subgraph.at(x)->Name();
     std::string fc_Y_in = w->Name();
     std::string fc_bias_in = fc_bias->Name();

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -1640,7 +1640,8 @@ PDNode *patterns::FillConstantElementWiseMulFuse::operator()(
 void patterns::QuantDequantOpFuse::operator()(PDNode *quant_op_input,
                                               const std::string &op_type,
                                               const std::string &weight_name,
-                                              int times) {
+                                              int times,
+                                              const std::string &quant_type) {
   const int kNumFields = 5;
   const int kQuantizedWeightOffset = 0;
   const int kQuantizedOpOffset = 1;
@@ -1648,22 +1649,20 @@ void patterns::QuantDequantOpFuse::operator()(PDNode *quant_op_input,
   const int kDequantOpOffset = 3;
   const int kDequantOpOutOffset = 4;
   // the quant op always be one.
-  auto quant_op_in_scale =
-      pattern->NewNode(GetNodeName("quant_op_in_scale"))
-          ->assert_is_op_input("fake_quantize_range_abs_max", "InScale")
+  auto quant_op_in_scale = pattern->NewNode(GetNodeName("quant_op_in_scale"))
+                               ->assert_is_op_input(quant_type, "InScale")
                                ->AsInput();
-  auto quant_op = pattern->NewNode(GetNodeName("quant_op"))
-                      ->assert_is_op("fake_quantize_range_abs_max");
+  auto quant_op =
+      pattern->NewNode(GetNodeName("quant_op"))->assert_is_op(quant_type);
 
   auto quant_op_out_scale =
       pattern->NewNode(GetNodeName("quant_op_out_scale"))
-          ->assert_is_op_output("fake_quantize_range_abs_max", "OutScale")
+          ->assert_is_op_output(quant_type, "OutScale")
           ->assert_is_op_input("fake_dequantize_max_abs", "Scale")
           ->AsIntermediate();
 
-  auto quant_op_out =
-      pattern->NewNode(GetNodeName("quant_op_out"))
-          ->assert_is_op_output("fake_quantize_range_abs_max", "Out")
+  auto quant_op_out = pattern->NewNode(GetNodeName("quant_op_out"))
+                          ->assert_is_op_output(quant_type, "Out")
                           ->assert_is_op_input(op_type)
                           ->AsIntermediate();
 

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -880,7 +880,8 @@ struct QuantDequantOpFuse : public PatternBase {
       : PatternBase(pattern, name_scope, "quant_dequant_fuse") {}
 
   void operator()(PDNode* quant_op_input, const std::string& op_name,
-                  const std::string& weight_name, int times = 1);
+                  const std::string& weight_name, int times,
+                  const std::string& quant_type);
 
   std::string GetNodeName(const std::string& op_type) {
     return PDNodeName(name_scope_, repr_, id_, op_type);

paddle/fluid/framework/ir/quant_conv2d_dequant_fuse_pass.cc

@@ -25,7 +25,8 @@ namespace framework {
 namespace ir {
 
 void RunQuantDequant(ir::Graph* graph, Scope* scope, int times,
-                     std::string op_type) {
+                     const std::string& op_type,
+                     const std::string& quant_type) {
   const std::string pattern_name = "quant_dequant_fuse";
   //  FusePassBase::Init(pattern_name, graph);
   const int kNumFields = 5;
@@ -38,7 +39,7 @@ void RunQuantDequant(ir::Graph* graph, Scope* scope, int times,
   GraphPatternDetector gpd;
   auto* x = gpd.mutable_pattern()
                 ->NewNode("x")
-                ->assert_is_op_input("fake_quantize_range_abs_max", "X")
+                ->assert_is_op_input(quant_type, "X")
                 ->AsInput();
 
   std::string quantized_op_type = "";
@@ -46,6 +47,9 @@ void RunQuantDequant(ir::Graph* graph, Scope* scope, int times,
   if (op_type == "conv2d") {
     quantized_op_type = "conv2d";
     weight_name = "Filter";
+  } else if (op_type == "depthwise_conv2d") {
+    quantized_op_type = "depthwise_conv2d";
+    weight_name = "Filter";
   } else if (op_type == "conv2d_fusion") {
     quantized_op_type = "conv2d_fusion";
     weight_name = "Filter";
@@ -62,7 +66,7 @@ void RunQuantDequant(ir::Graph* graph, Scope* scope, int times,
   }
 
   patterns::QuantDequantOpFuse pattern(gpd.mutable_pattern(), pattern_name);
-  pattern(x, quantized_op_type, weight_name, times);
+  pattern(x, quantized_op_type, weight_name, times, quant_type);
 
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
@@ -103,7 +107,6 @@ void RunQuantDequant(ir::Graph* graph, Scope* scope, int times,
     std::unordered_set<const Node*> delete_nodes;
 
     for (int i = 0; i < times; i++) {
-      // max_range = (range * range) / weight_scale
       float max_range = boost::get<float>(
           nodes[i * kNumFields + kDequantOpOffset]->Op()->GetAttr("max_range"));
       float weight_scale = (range * range) / max_range;
@@ -118,7 +121,8 @@ void RunQuantDequant(ir::Graph* graph, Scope* scope, int times,
       new_op_desc.SetType(quantized_op_type);
 
       if (quantized_op_type == "conv2d" ||
-          quantized_op_type == "conv2d_fusion") {
+          quantized_op_type == "conv2d_fusion" ||
+          quantized_op_type == "depthwise_conv2d") {
         new_op_desc.SetInput("Input", {new_input});
         new_op_desc.SetOutput("Output", {new_output});
       } else if (quantized_op_type == "fc") {
@@ -156,11 +160,17 @@ void QuantDequantFusePass::ApplyImpl(ir::Graph* graph) const {
   const std::string pattern_name = "quant_dequant_fuse";
   FusePassBase::Init(pattern_name, graph);
 
-  std::unordered_set<std::string> quantized_op_types = {"conv2d", "mul"};
+  std::unordered_set<std::string> quant_types = {
+      "fake_quantize_range_abs_max", "fake_quantize_moving_average_abs_max"};
+
+  std::unordered_set<std::string> quantized_op_types = {"conv2d", "mul",
+                                                        "depthwise_conv2d"};
   auto* scope = param_scope();
+  for (auto& quant_type : quant_types) {
     for (auto& op_type : quantized_op_types) {
-    for (int i = 1; i <= 6; i++) {
-      RunQuantDequant(graph, scope, i, op_type);
+      for (int i = 6; i >= 1; i--) {
+        RunQuantDequant(graph, scope, i, op_type, quant_type);
+      }
     }
   }
 }

paddle/fluid/inference/anakin/convert/CMakeLists.txt

-cc_library(anakin_op_converter SRCS fc.cc conv2d.cc conv2d_fusion.cc elementwise.cc activation.cc pool2d.cc concat.cc split.cc relu.cc softmax.cc batch_norm.cc reshape.cc flatten.cc transpose.cc density_prior_box.cc detection_out.cc scale.cc dropout.cc im2sequence.cc sum.cc affine_channel.cc roi_align.cc DEPS anakin_engine framework_proto scope op_registry)
+cc_library(anakin_op_converter SRCS fc.cc conv2d.cc conv2d_fusion.cc
+elementwise.cc activation.cc pool2d.cc concat.cc split.cc relu.cc softmax.cc
+batch_norm.cc reshape.cc flatten.cc transpose.cc density_prior_box.cc
+detection_out.cc scale.cc dropout.cc im2sequence.cc sum.cc affine_channel.cc
+roi_align.cc helper.cc DEPS anakin_engine framework_proto scope op_registry
+gtest)
 
 cc_test(test_anakin_fc SRCS test_fc_op.cc DEPS anakin_op_converter mul_op SERIAL)
 cc_test(test_anakin_conv2d SRCS test_conv2d_op.cc DEPS anakin_op_converter conv_op im2col vol2col depthwise_conv SERIAL)

paddle/fluid/inference/anakin/convert/activation.cc

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-ActivationOpConverter<TargetT>::ActivationOpConverter(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+ActivationOpConverter<TargetT, PrecisionT>::ActivationOpConverter(
     const std::string &op_type)
     : op_type_(op_type) {
   auto it = anakin_op_types_.find(op_type_);
@@ -30,8 +30,8 @@ ActivationOpConverter<TargetT>::ActivationOpConverter(
   anakin_op_type_ = it->second;
 }
 
-template <typename TargetT>
-void ActivationOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void ActivationOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -50,11 +50,40 @@ void ActivationOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(sigmoid,
-                                  SigmoidOpConverter<::anakin::saber::NV>);
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(tanh, TanhOpConverter<::anakin::saber::NV>);
+using sigmoid_nv_fp32 =
+    ::paddle::inference::anakin::SigmoidOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::FP32>;
+using sigmoid_nv_int8 =
+    ::paddle::inference::anakin::SigmoidOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::INT8>;
+using tanh_nv_fp32 =
+    ::paddle::inference::anakin::TanhOpConverter<::anakin::saber::NV,
+                                                 ::anakin::Precision::FP32>;
+using tanh_nv_int8 =
+    ::paddle::inference::anakin::TanhOpConverter<::anakin::saber::NV,
+                                                 ::anakin::Precision::INT8>;
+
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(sigmoid, sigmoid_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(sigmoid, sigmoid_nv_int8);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(tanh, tanh_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(tanh, tanh_nv_int8);
 #endif
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(sigmoid,
-                                 SigmoidOpConverter<::anakin::saber::X86>);
-REGISTER_CPU_ANAKIN_OP_CONVERTER(tanh, TanhOpConverter<::anakin::saber::X86>);
+using sigmoid_cpu_fp32 =
+    ::paddle::inference::anakin::SigmoidOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::FP32>;
+using sigmoid_cpu_int8 =
+    ::paddle::inference::anakin::SigmoidOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::INT8>;
+using tanh_cpu_fp32 =
+    ::paddle::inference::anakin::TanhOpConverter<::anakin::saber::X86,
+                                                 ::anakin::Precision::FP32>;
+using tanh_cpu_int8 =
+    ::paddle::inference::anakin::TanhOpConverter<::anakin::saber::X86,
+                                                 ::anakin::Precision::INT8>;
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(sigmoid, sigmoid_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(sigmoid, sigmoid_cpu_int8);
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(tanh, tanh_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(tanh, tanh_cpu_int8);

paddle/fluid/inference/anakin/convert/activation.h

@@ -22,8 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class ActivationOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class ActivationOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   explicit ActivationOpConverter(const std::string &op_type);
 
@@ -40,16 +40,17 @@ class ActivationOpConverter : public AnakinOpConverter<TargetT> {
                                                       {"sigmoid", "Sigmoid"}};
 };
 
-template <typename TargetT>
-class TanhOpConverter : public ActivationOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class TanhOpConverter : public ActivationOpConverter<TargetT, PrecisionT> {
  public:
-  TanhOpConverter() : ActivationOpConverter<TargetT>("tanh") {}
+  TanhOpConverter() : ActivationOpConverter<TargetT, PrecisionT>("tanh") {}
 };
 
-template <typename TargetT>
-class SigmoidOpConverter : public ActivationOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class SigmoidOpConverter : public ActivationOpConverter<TargetT, PrecisionT> {
  public:
-  SigmoidOpConverter() : ActivationOpConverter<TargetT>("sigmoid") {}
+  SigmoidOpConverter()
+      : ActivationOpConverter<TargetT, PrecisionT>("sigmoid") {}
 };
 }  // namespace anakin
 }  // namespace inference

paddle/fluid/inference/anakin/convert/affine_channel.cc

@@ -16,18 +16,14 @@
 #include <algorithm>
 #include <string>
 #include <vector>
-
-using anakin::graph::GraphGlobalMem;
-using anakin::PTuple;
-using anakin::AK_FLOAT;
-using anakin::saber::Shape;
+#include "paddle/fluid/inference/anakin/convert/helper.h"
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void AffineChannelOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void AffineChannelOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -35,60 +31,20 @@ void AffineChannelOpConverter<TargetT>::operator()(
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
 
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
-
   auto input_name = op_desc.Input("X").front();
   auto output_name = op_desc.Output("Out").front();
+  this->engine_->AddOp(op_name, "AffineChannel", {input_name}, {output_name});
 
   // Copy the Scale to CPUPlace and get the pointer.
   auto *scale_v = scope.FindVar(op_desc.Input("Scale").front());
   PADDLE_ENFORCE_NOT_NULL(scale_v);
-  auto *scale_t = scale_v->GetMutable<framework::LoDTensor>();
-  std::unique_ptr<framework::LoDTensor> scale_tensor(
-      new framework::LoDTensor());
-  scale_tensor->Resize(scale_t->dims());
-  TensorCopySync((*scale_t), platform::CPUPlace(), scale_tensor.get());
+  auto weight1 = pblock_from_var<TargetT>(*scale_v);
+  this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
 
   // Copy the Bias to CPUPlace and get the pointer.
   auto *bias_v = scope.FindVar(op_desc.Input("Bias").front());
   PADDLE_ENFORCE_NOT_NULL(bias_v);
-  auto *bias_t = bias_v->GetMutable<framework::LoDTensor>();
-  std::unique_ptr<framework::LoDTensor> bias_tensor(new framework::LoDTensor());
-  bias_tensor->Resize(bias_t->dims());
-  TensorCopySync((*bias_t), platform::CPUPlace(), bias_tensor.get());
-
-  this->engine_->AddOp(op_name, "AffineChannel", {input_name}, {output_name});
-
-  // Generate the Scale parameter of Anakin.
-  auto scale_shape = framework::vectorize2int(scale_t->dims());
-  while (scale_shape.size() < 4) {
-    scale_shape.insert(scale_shape.begin(), 1);
-  }
-  Shape anakin_scale_shape(scale_shape);
-  auto *weight1 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-          anakin_scale_shape);
-  float *scale_cpu_data =
-      static_cast<float *>(weight1->h_tensor().mutable_data());
-  std::copy_n(scale_tensor->data<float>(), scale_tensor->numel(),
-              scale_cpu_data);
-  weight1->d_tensor().set_shape(anakin_scale_shape);
-  weight1->d_tensor().copy_from(weight1->h_tensor());
-  this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
-
-  // Generate the Bias parameter of Anakin.
-  auto bias_shape = framework::vectorize2int(bias_t->dims());
-  while (bias_shape.size() < 4) {
-    bias_shape.insert(bias_shape.begin(), 1);
-  }
-  Shape anakin_bias_shape(bias_shape);
-  auto *weight2 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-          anakin_bias_shape);
-  float *bias_cpu_data =
-      static_cast<float *>(weight2->h_tensor().mutable_data());
-  std::copy_n(bias_tensor->data<float>(), bias_tensor->numel(), bias_cpu_data);
-  weight2->d_tensor().set_shape(anakin_bias_shape);
-  weight2->d_tensor().copy_from(weight2->h_tensor());
+  auto weight2 = pblock_from_var<TargetT>(*bias_v);
   this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
 }
 
@@ -97,8 +53,21 @@ void AffineChannelOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(
-    affine_channel, AffineChannelOpConverter<::anakin::saber::NV>);
+using affine_channel_nv_fp32 =
+    ::paddle::inference::anakin::AffineChannelOpConverter<
+        ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using affine_channel_nv_int8 =
+    ::paddle::inference::anakin::AffineChannelOpConverter<
+        ::anakin::saber::NV, ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(affine_channel, affine_channel_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(affine_channel, affine_channel_nv_int8);
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(
-    affine_channel, AffineChannelOpConverter<::anakin::saber::X86>);
+
+using affine_channel_cpu_fp32 =
+    ::paddle::inference::anakin::AffineChannelOpConverter<
+        ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using affine_channel_cpu_int8 =
+    ::paddle::inference::anakin::AffineChannelOpConverter<
+        ::anakin::saber::X86, ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(affine_channel, affine_channel_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(affine_channel, affine_channel_cpu_int8);

paddle/fluid/inference/anakin/convert/affine_channel.h

@@ -21,8 +21,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class AffineChannelOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class AffineChannelOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   AffineChannelOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/batch_norm.cc

@@ -18,17 +18,14 @@
 #include <map>
 #include <string>
 #include <vector>
-
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::Shape;
+#include "paddle/fluid/inference/anakin/convert/helper.h"
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void BatchNormOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void BatchNormOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -36,87 +33,46 @@ void BatchNormOpConverter<TargetT>::operator()(
   std::map<std::string, std::string> inputs;
   for (auto k : {"X", "Scale", "Bias", "Mean", "Variance"}) {
     PADDLE_ENFORCE_EQ(op_desc.Input(k).size(), 1UL);
-    auto v = op_desc.Input(k).front();
-    inputs.insert({k, v});
   }
 
+  auto input = op_desc.Input("X").front();
   auto output = op_desc.Output("Y").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Y").front();
   auto epsilon = boost::get<float>(op_desc.GetAttr("epsilon"));
-  // auto momentum = boost::get<float>(op_desc.GetAttr("momentum"));
 
   auto bn_op_name = op_name + ":bn";
   auto bn_output = bn_op_name + "_output";
-  this->engine_->AddOp(bn_op_name, "BatchNorm", {inputs["X"]}, {bn_output});
+  this->engine_->AddOp(bn_op_name, "BatchNorm", {input}, {bn_output});
   this->engine_->AddOpAttr(bn_op_name, "epsilon", epsilon);
   this->engine_->AddOpAttr(bn_op_name, "momentum", static_cast<float>(1.0));
 
   auto scale_op_name = op_name + ":scale";
-  auto get_lod_tensor = [this, &scope, &op_name](const std::string &var_name,
-                                                 framework::LoDTensor *tensor) {
-    auto *v = scope.FindVar(var_name);
-    PADDLE_ENFORCE_NOT_NULL(v);
-    auto *t = v->GetMutable<framework::LoDTensor>();
-    tensor->Resize(t->dims());
-    TensorCopySync(*t, platform::CPUPlace(), tensor);
-  };
-
-  framework::LoDTensor bias_t;
-  framework::LoDTensor mean_t;
-  framework::LoDTensor scale_t;
-  framework::LoDTensor variance_t;
-  get_lod_tensor(inputs["Bias"], &bias_t);
-  get_lod_tensor(inputs["Mean"], &mean_t);
-  get_lod_tensor(inputs["Scale"], &scale_t);
-  get_lod_tensor(inputs["Variance"], &variance_t);
+  this->engine_->AddOp(scale_op_name, "Scale", {bn_output}, {output});
+  this->engine_->AddOpAttr(scale_op_name, "axis", 1);
+  this->engine_->AddOpAttr(scale_op_name, "num_axes", 1);
+  this->engine_->AddOpAttr(scale_op_name, "bias_term", true);
 
-  auto fill_shape = [](size_t n, std::vector<int> shape) {
-    shape.insert(shape.begin(), 1);
-    if (shape.size() < n) {
-      shape.insert(shape.end(), n - shape.size(), 1);
-    }
-    return shape;
-  };
-  Shape shape1(fill_shape(4, framework::vectorize2int(mean_t.dims())));
-  Shape shape2(fill_shape(4, framework::vectorize2int(variance_t.dims())));
-  auto *weight1 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(shape1);
-  auto *mean_data = static_cast<float *>(weight1->h_tensor().mutable_data());
-  std::copy_n(mean_t.data<float>(), mean_t.numel(), mean_data);
+  auto *mean_v = scope.FindVar(op_desc.Input("Mean").front());
+  PADDLE_ENFORCE_NOT_NULL(mean_v);
+  auto weight1 = pblock_from_var<TargetT>(*mean_v);
   this->engine_->AddOpAttr(bn_op_name, "weight_1", *weight1);
 
-  auto *weight2 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(shape2);
-  auto *variance_data =
-      static_cast<float *>(weight2->h_tensor().mutable_data());
-  std::copy_n(variance_t.data<float>(), variance_t.numel(), variance_data);
+  auto *variance_v = scope.FindVar(op_desc.Input("Variance").front());
+  PADDLE_ENFORCE_NOT_NULL(variance_v);
+  auto weight2 = pblock_from_var<TargetT>(*variance_v);
   this->engine_->AddOpAttr(bn_op_name, "weight_2", *weight2);
 
-  Shape shape3(std::vector<int>({1, 1, 1, 1}));
-  auto *weight3 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(shape3);
-  auto *alpha_data = static_cast<float *>(weight3->h_tensor().mutable_data());
-  float weight3_data[] = {1};
-  std::copy(std::begin(weight3_data), std::end(weight3_data), alpha_data);
+  auto *weight3 = pblock_from_vector<TargetT>(std::vector<float>({1}));
   this->engine_->AddOpAttr(bn_op_name, "weight_3", *weight3);
 
-  Shape scale_shape(fill_shape(4, framework::vectorize2int(scale_t.dims())));
-  auto *scale = GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-      scale_shape);
-  auto *scale_data = static_cast<float *>(scale->h_tensor().mutable_data());
-  std::copy_n(scale_t.data<float>(), scale_t.numel(), scale_data);
-
-  Shape bias_shape(fill_shape(4, framework::vectorize2int(bias_t.dims())));
-  auto *bias = GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-      bias_shape);
-  auto *bias_data = static_cast<float *>(bias->h_tensor().mutable_data());
-  std::copy_n(bias_t.data<float>(), bias_t.numel(), bias_data);
-
-  this->engine_->AddOp(scale_op_name, "Scale", {bn_output}, {output});
-  this->engine_->AddOpAttr(scale_op_name, "axis", 1);
-  this->engine_->AddOpAttr(scale_op_name, "num_axes", 1);
-  this->engine_->AddOpAttr(scale_op_name, "bias_term", true);
+  auto *scale_v = scope.FindVar(op_desc.Input("Scale").front());
+  PADDLE_ENFORCE_NOT_NULL(scale_v);
+  auto scale = pblock_from_var<TargetT>(*scale_v);
   this->engine_->AddOpAttr(scale_op_name, "weight_1", *scale);
+
+  auto *bias_v = scope.FindVar(op_desc.Input("Bias").front());
+  PADDLE_ENFORCE_NOT_NULL(bias_v);
+  auto bias = pblock_from_var<TargetT>(*bias_v);
   this->engine_->AddOpAttr(scale_op_name, "weight_2", *bias);
 }
 
@@ -125,9 +81,17 @@ void BatchNormOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(batch_norm,
-                                  BatchNormOpConverter<::anakin::saber::NV>);
+using bn_nv_fp32 = ::paddle::inference::anakin::BatchNormOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using bn_nv_int8 = ::paddle::inference::anakin::BatchNormOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(batch_norm, bn_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(batch_norm, bn_nv_int8);
 #endif
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(batch_norm,
-                                 BatchNormOpConverter<::anakin::saber::X86>);
+using bn_cpu_fp32 = ::paddle::inference::anakin::BatchNormOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using bn_cpu_int8 = ::paddle::inference::anakin::BatchNormOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(batch_norm, bn_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(batch_norm, bn_cpu_int8);

paddle/fluid/inference/anakin/convert/batch_norm.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class BatchNormOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class BatchNormOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   BatchNormOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/concat.cc

@@ -19,8 +19,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void ConcatOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void ConcatOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -39,8 +39,21 @@ void ConcatOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(concat,
-                                  ConcatOpConverter<::anakin::saber::NV>);
+using concat_nv_fp32 =
+    ::paddle::inference::anakin::ConcatOpConverter<::anakin::saber::NV,
+                                                   ::anakin::Precision::FP32>;
+using concat_nv_int8 =
+    ::paddle::inference::anakin::ConcatOpConverter<::anakin::saber::NV,
+                                                   ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(concat, concat_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(concat, concat_nv_int8);
+
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(concat,
-                                 ConcatOpConverter<::anakin::saber::X86>);
+using concat_cpu_fp32 =
+    ::paddle::inference::anakin::ConcatOpConverter<::anakin::saber::X86,
+                                                   ::anakin::Precision::FP32>;
+using concat_cpu_int8 =
+    ::paddle::inference::anakin::ConcatOpConverter<::anakin::saber::X86,
+                                                   ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(concat, concat_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(concat, concat_cpu_int8);

paddle/fluid/inference/anakin/convert/concat.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class ConcatOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class ConcatOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   ConcatOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/conv2d.cc

@@ -16,18 +16,16 @@
 #include <algorithm>
 #include <memory>
 #include <vector>
+#include "paddle/fluid/inference/anakin/convert/helper.h"
 
-using anakin::graph::GraphGlobalMem;
 using anakin::PTuple;
-using anakin::AK_FLOAT;
-using anakin::saber::Shape;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void Conv2dOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void Conv2dOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -42,11 +40,8 @@ void Conv2dOpConverter<TargetT>::operator()(
 
   auto *filter_v = scope.FindVar(op_desc.Input("Filter").front());
   PADDLE_ENFORCE_NOT_NULL(filter_v);
-  auto *filter_t = filter_v->GetMutable<framework::LoDTensor>();
-  std::unique_ptr<framework::LoDTensor> weight_tensor(
-      new framework::LoDTensor());
-  weight_tensor->Resize(filter_t->dims());
-  TensorCopySync((*filter_t), platform::CPUPlace(), weight_tensor.get());
+  auto weight_tensor = tensor_from_var(*filter_v, platform::CPUPlace());
+  auto weight_shape = framework::vectorize2int(weight_tensor->dims());
 
   PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL);
 
@@ -69,25 +64,61 @@ void Conv2dOpConverter<TargetT>::operator()(
   this->engine_->AddOpAttr(op_name, "axis", 1);
   this->engine_->AddOpAttr(op_name, "bias_term", false);
 
-  auto weight_shape = framework::vectorize2int(filter_t->dims());
-  Shape anakin_shape(weight_shape);
-  auto *weight1 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-          anakin_shape);
-  float *cpu_data = static_cast<float *>(weight1->h_tensor().mutable_data());
-  std::copy_n(weight_tensor->data<float>(), weight_tensor->numel(), cpu_data);
+  ::anakin::saber::Shape anakin_shape(weight_shape);
+  bool enable_int8 = boost::get<bool>(op_desc.HasAttr("enable_int8"));
+
+  if (enable_int8) {
+    const float int8_range = 127.;
+    float in_scale = boost::get<float>(op_desc.GetAttr("input_scale"));
+    float weight_scale = boost::get<float>(op_desc.GetAttr("weight_scale"));
+    auto *weight1 = ::anakin::graph::GraphGlobalMem<TargetT>::Global()
+                        .template new_block<::anakin::AK_INT8>(anakin_shape);
+    float *weight_data = weight_tensor->data<float>();
+    std::vector<char> weight_int8;
+    int weight_num = weight_tensor->numel();
+    for (int i = 0; i < weight_tensor->numel(); i++) {
+      bool is_valid_int8 =
+          ((weight_data[i] >= -128) && (weight_data[i] <= 127));
+      PADDLE_ENFORCE(is_valid_int8,
+                     "We are in anakin subgraph int8 mode, the weight of conv "
+                     "should be in range [-128, 127]");
+      weight_int8.push_back(static_cast<char>(weight_data[i]));
+    }
+    memcpy(static_cast<void *>(weight1->h_tensor().mutable_data()),
+           static_cast<void *>(weight_int8.data()), sizeof(char) * weight_num);
     weight1->d_tensor().set_shape(anakin_shape);
     weight1->d_tensor().copy_from(weight1->h_tensor());
     this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
+    this->engine_->Graph()->SetOpPrec(op_name, ::anakin::AK_INT8);
+    this->engine_->Graph()->SetWeightsScale(op_name,
+                                            {weight_scale / int8_range}, false);
+    this->engine_->AddTensorScale(input_name, in_scale / int8_range);
+  } else {
+    auto *weight1 = pblock_from_tensor<TargetT>(*weight_tensor, weight_shape);
+    this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
+  }
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(conv2d,
-                                 Conv2dOpConverter<::anakin::saber::X86>);
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(conv2d,
-                                  Conv2dOpConverter<::anakin::saber::NV>);
+using conv2d_nv_fp32 =
+    ::paddle::inference::anakin::Conv2dOpConverter<::anakin::saber::NV,
+                                                   ::anakin::Precision::FP32>;
+using conv2d_nv_int8 =
+    ::paddle::inference::anakin::Conv2dOpConverter<::anakin::saber::NV,
+                                                   ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(conv2d, conv2d_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(conv2d, conv2d_nv_int8);
 #endif
+
+using conv2d_cpu_fp32 =
+    ::paddle::inference::anakin::Conv2dOpConverter<::anakin::saber::X86,
+                                                   ::anakin::Precision::FP32>;
+using conv2d_cpu_int8 =
+    ::paddle::inference::anakin::Conv2dOpConverter<::anakin::saber::X86,
+                                                   ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(conv2d, conv2d_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(conv2d, conv2d_cpu_int8);

paddle/fluid/inference/anakin/convert/conv2d.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class Conv2dOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class Conv2dOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   Conv2dOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/conv2d_fusion.cc

@@ -16,18 +16,16 @@
 #include <algorithm>
 #include <memory>
 #include <vector>
+#include "paddle/fluid/inference/anakin/convert/helper.h"
 
-using anakin::graph::GraphGlobalMem;
 using anakin::PTuple;
-using anakin::AK_FLOAT;
-using anakin::saber::Shape;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void Conv2dFusionOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void Conv2dFusionOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -43,24 +41,16 @@ void Conv2dFusionOpConverter<TargetT>::operator()(
 
   auto *filter_v = scope.FindVar(op_desc.Input("Filter").front());
   PADDLE_ENFORCE_NOT_NULL(filter_v);
-  auto *filter_t = filter_v->GetMutable<framework::LoDTensor>();
+
+  auto weight_tensor = tensor_from_var(*filter_v, platform::CPUPlace());
+  auto weight_shape = framework::vectorize2int(weight_tensor->dims());
 
   auto *b_v = scope.FindVar(op_desc.Input("Bias").front());
   PADDLE_ENFORCE_NOT_NULL(b_v);
-  auto *b_t = b_v->GetMutable<framework::LoDTensor>();
-
-  std::unique_ptr<framework::LoDTensor> weight_tensor(
-      new framework::LoDTensor());
-  weight_tensor->Resize(filter_t->dims());
-  TensorCopySync((*filter_t), platform::CPUPlace(), weight_tensor.get());
 
   PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL);
-
-  // const int n_output = weight_tensor->dims()[0];
-  // const int n_input = weight_tensor->dims()[1];
   const int filter_h = weight_tensor->dims()[2];
   const int filter_w = weight_tensor->dims()[3];
-  // auto filter_num = n_input * filter_h * filter_w ;
   auto filter_num = weight_tensor->dims()[0];
   this->engine_->template AddOpAttr<int>(op_name, "filter_num", filter_num);
   this->engine_->template AddOpAttr<PTuple<int>>(op_name, "kernel_size",
@@ -77,37 +67,42 @@ void Conv2dFusionOpConverter<TargetT>::operator()(
   this->engine_->AddOpAttr(op_name, "axis", 1);
   this->engine_->AddOpAttr(op_name, "bias_term", true);
 
-  auto weight_shape = framework::vectorize2int(filter_t->dims());
-  Shape anakin_shape(weight_shape);
-  auto *weight1 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-          anakin_shape);
-  float *cpu_data = static_cast<float *>(weight1->h_tensor().mutable_data());
-  std::copy_n(weight_tensor->data<float>(), weight_tensor->numel(), cpu_data);
+  ::anakin::saber::Shape anakin_shape(weight_shape);
+  bool enable_int8 = boost::get<bool>(op_desc.HasAttr("enable_int8"));
+  if (enable_int8) {
+    const float int8_range = 127.;
+    float in_scale = boost::get<float>(op_desc.GetAttr("input_scale"));
+    float weight_scale = boost::get<float>(op_desc.GetAttr("weight_scale"));
+    auto *weight1 = ::anakin::graph::GraphGlobalMem<TargetT>::Global()
+                        .template new_block<::anakin::AK_INT8>(anakin_shape);
+    float *weight_data = weight_tensor->data<float>();
+    std::vector<char> weight_int8;
+    int weight_num = weight_tensor->numel();
+    for (int i = 0; i < weight_tensor->numel(); i++) {
+      bool is_valid_int8 =
+          ((weight_data[i] >= -128) && (weight_data[i] <= 127));
+      PADDLE_ENFORCE(is_valid_int8,
+                     "We are in anakin subgraph int8 mode, the weight of conv "
+                     "should be in range [-128, 127]");
+      weight_int8.push_back(static_cast<char>(weight_data[i]));
+    }
+    memcpy(static_cast<void *>(weight1->h_tensor().mutable_data()),
+           static_cast<void *>(weight_int8.data()), sizeof(char) * weight_num);
     weight1->d_tensor().set_shape(anakin_shape);
     weight1->d_tensor().copy_from(weight1->h_tensor());
     this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
-
-  auto bias_shape = framework::vectorize2int(b_t->dims());
-  framework::LoDTensor bias_tensor;
-  bias_tensor.Resize(b_t->dims());
-  TensorCopySync((*b_t), platform::CPUPlace(), &bias_tensor);
-  auto *bias_data = bias_tensor.data<float>();
-  bias_shape.insert(bias_shape.begin(), 1);
-  bias_shape.insert(bias_shape.begin(), 1);
-  bias_shape.insert(bias_shape.begin(), 1);
-  // bias_shape.push_back(1);
-  // bias_shape.push_back(1);
-  Shape anakin_bias_shape(bias_shape);
-
-  auto *weight2 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-          anakin_bias_shape);
-  float *cpu_data2 = static_cast<float *>(weight2->h_tensor().mutable_data());
-  std::copy_n(bias_data, bias_tensor.numel(), cpu_data2);
-  weight2->d_tensor().set_shape(anakin_bias_shape);
-  weight2->d_tensor().copy_from(weight2->h_tensor());
+    this->engine_->Graph()->SetOpPrec(op_name, ::anakin::AK_INT8);
+    this->engine_->Graph()->SetWeightsScale(op_name,
+                                            {weight_scale / int8_range}, false);
+    this->engine_->AddTensorScale(input_name, in_scale / int8_range);
+  } else {
+    auto weight_tensor = tensor_from_var(*filter_v, platform::CPUPlace());
+    auto weight_shape = framework::vectorize2int(weight_tensor->dims());
+    auto *weight1 = pblock_from_tensor<TargetT>(*weight_tensor, weight_shape);
+    this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
+    auto weight2 = pblock_from_var<TargetT>(*b_v);
     this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
+  }
 }
 
 }  // namespace anakin
@@ -115,9 +110,21 @@ void Conv2dFusionOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(conv2d_fusion,
-                                  Conv2dFusionOpConverter<::anakin::saber::NV>);
+using conv2d_fusion_nv_fp32 =
+    ::paddle::inference::anakin::Conv2dFusionOpConverter<
+        ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using conv2d_fusion_nv_int8 =
+    ::paddle::inference::anakin::Conv2dFusionOpConverter<
+        ::anakin::saber::NV, ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(conv2d_fusion, conv2d_fusion_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(conv2d_fusion, conv2d_fusion_nv_int8);
 #endif
-
-REGISTER_CPU_ANAKIN_OP_CONVERTER(conv2d_fusion,
-                                 Conv2dFusionOpConverter<::anakin::saber::X86>);
+using conv2d_fusion_cpu_fp32 =
+    ::paddle::inference::anakin::Conv2dFusionOpConverter<
+        ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using conv2d_fusion_cpu_int8 =
+    ::paddle::inference::anakin::Conv2dFusionOpConverter<
+        ::anakin::saber::X86, ::anakin::Precision::INT8>;
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(conv2d_fusion, conv2d_fusion_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(conv2d_fusion, conv2d_fusion_cpu_int8);

paddle/fluid/inference/anakin/convert/conv2d_fusion.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class Conv2dFusionOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class Conv2dFusionOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   Conv2dFusionOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/density_prior_box.cc

@@ -23,8 +23,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void DensityPriorBoxOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void DensityPriorBoxOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc& op, const framework::BlockDesc& block_desc,
     const framework::Scope& scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -109,13 +109,24 @@ void DensityPriorBoxOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(
-    density_prior_box, DensityPriorBoxOpConverter<::anakin::saber::NV>);
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(
-    prior_box, DensityPriorBoxOpConverter<::anakin::saber::NV>);
+using ds_pr_nv_fp32 = ::paddle::inference::anakin::DensityPriorBoxOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using ds_pr_nv_int8 = ::paddle::inference::anakin::DensityPriorBoxOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::INT8>;
+
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(density_prior_box, ds_pr_nv_fp32);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(prior_box, ds_pr_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(density_prior_box, ds_pr_nv_int8);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(prior_box, ds_pr_nv_int8);
 #endif
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(
-    density_prior_box, DensityPriorBoxOpConverter<::anakin::saber::X86>);
-REGISTER_CPU_ANAKIN_OP_CONVERTER(
-    prior_box, DensityPriorBoxOpConverter<::anakin::saber::X86>);
+using ds_pr_cpu_fp32 = ::paddle::inference::anakin::DensityPriorBoxOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using ds_pr_cpu_int8 = ::paddle::inference::anakin::DensityPriorBoxOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::INT8>;
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(density_prior_box, ds_pr_cpu_fp32);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(prior_box, ds_pr_cpu_fp32);
+
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(density_prior_box, ds_pr_cpu_int8);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(prior_box, ds_pr_cpu_int8);

paddle/fluid/inference/anakin/convert/density_prior_box.h

@@ -22,8 +22,9 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class DensityPriorBoxOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class DensityPriorBoxOpConverter
+    : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   DensityPriorBoxOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/detection_out.cc

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void DetectionOutOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void DetectionOutOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -67,8 +67,21 @@ void DetectionOutOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(detection_out,
-                                  DetectionOutOpConverter<::anakin::saber::NV>);
+using detection_out_nv_fp32 =
+    ::paddle::inference::anakin::DetectionOutOpConverter<
+        ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using detection_out_nv_int8 =
+    ::paddle::inference::anakin::DetectionOutOpConverter<
+        ::anakin::saber::NV, ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(detection_out, detection_out_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(detection_out, detection_out_nv_int8);
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(detection_out,
-                                 DetectionOutOpConverter<::anakin::saber::X86>);
+
+using detection_out_cpu_fp32 =
+    ::paddle::inference::anakin::DetectionOutOpConverter<
+        ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using detection_out_cpu_int8 =
+    ::paddle::inference::anakin::DetectionOutOpConverter<
+        ::anakin::saber::X86, ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(detection_out, detection_out_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(detection_out, detection_out_cpu_int8);

paddle/fluid/inference/anakin/convert/detection_out.h

@@ -22,8 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class DetectionOutOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class DetectionOutOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   DetectionOutOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/dropout.cc

@@ -16,17 +16,14 @@
 #include <algorithm>
 #include <string>
 #include <vector>
-
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::Shape;
+#include "paddle/fluid/inference/anakin/convert/helper.h"
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void DropoutOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void DropoutOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -42,12 +39,7 @@ void DropoutOpConverter<TargetT>::operator()(
 
   auto dropout_prob = boost::get<float>(op_desc.GetAttr("dropout_prob"));
   auto factor = 1 - dropout_prob;
-  Shape shape1(std::vector<int>({1, 1, 1, 1}));
-  auto *weight1 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(shape1);
-  auto *factor_data = static_cast<float *>(weight1->h_tensor().mutable_data());
-  float weight1_data[] = {factor};
-  std::copy(std::begin(weight1_data), std::end(weight1_data), factor_data);
+  auto *weight1 = pblock_from_vector<TargetT>(std::vector<float>({factor}));
 
   this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
   this->engine_->AddOpAttr(op_name, "axis", 0);
@@ -60,8 +52,21 @@ void DropoutOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(dropout,
-                                  DropoutOpConverter<::anakin::saber::NV>);
+using dropout_nv_fp32 =
+    ::paddle::inference::anakin::DropoutOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::FP32>;
+using dropout_nv_int8 =
+    ::paddle::inference::anakin::DropoutOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(dropout, dropout_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(dropout, dropout_nv_int8);
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(dropout,
-                                 DropoutOpConverter<::anakin::saber::X86>);
+
+using dropout_cpu_fp32 =
+    ::paddle::inference::anakin::DropoutOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::FP32>;
+using dropout_cpu_int8 =
+    ::paddle::inference::anakin::DropoutOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(dropout, dropout_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(dropout, dropout_cpu_int8);

paddle/fluid/inference/anakin/convert/dropout.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class DropoutOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class DropoutOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   DropoutOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/elementwise.cc

@@ -17,17 +17,14 @@
 #include <string>
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::Shape;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void ElementwiseAddOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void ElementwiseAddOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -48,8 +45,8 @@ void ElementwiseAddOpConverter<TargetT>::operator()(
   this->engine_->template AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
 }
 
-template <typename TargetT>
-void ElementwiseMulOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void ElementwiseMulOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -75,12 +72,31 @@ void ElementwiseMulOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(
-    elementwise_add, ElementwiseAddOpConverter<::anakin::saber::NV>);
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(
-    elementwise_mul, ElementwiseMulOpConverter<::anakin::saber::NV>);
+using elet_nv_fp32 = ::paddle::inference::anakin::ElementwiseAddOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using elet_nv_int8 = ::paddle::inference::anakin::ElementwiseAddOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::INT8>;
+using eletmul_nv_fp32 = ::paddle::inference::anakin::ElementwiseMulOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using eletmul_nv_int8 = ::paddle::inference::anakin::ElementwiseMulOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::INT8>;
+
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(elementwise_add, elet_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(elementwise_add, elet_nv_int8);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(elementwise_mul, eletmul_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(elementwise_mul, eletmul_nv_int8);
+
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(
-    elementwise_add, ElementwiseAddOpConverter<::anakin::saber::X86>);
-REGISTER_CPU_ANAKIN_OP_CONVERTER(
-    elementwise_mul, ElementwiseMulOpConverter<::anakin::saber::X86>);
+using elet_cpu_fp32 = ::paddle::inference::anakin::ElementwiseAddOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using elet_cpu_int8 = ::paddle::inference::anakin::ElementwiseAddOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::INT8>;
+using eletmul_cpu_fp32 = ::paddle::inference::anakin::ElementwiseMulOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using eletmul_cpu_int8 = ::paddle::inference::anakin::ElementwiseMulOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::INT8>;
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(elementwise_add, elet_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(elementwise_add, elet_cpu_int8);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(elementwise_mul, eletmul_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(elementwise_mul, eletmul_cpu_int8);

paddle/fluid/inference/anakin/convert/elementwise.h

@@ -20,8 +20,9 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class ElementwiseAddOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class ElementwiseAddOpConverter
+    : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   ElementwiseAddOpConverter() = default;
 
@@ -34,8 +35,9 @@ class ElementwiseAddOpConverter : public AnakinOpConverter<TargetT> {
  private:
 };
 
-template <typename TargetT>
-class ElementwiseMulOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class ElementwiseMulOpConverter
+    : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   ElementwiseMulOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/fc.cc

@@ -16,22 +16,19 @@
 #include <algorithm>
 #include <string>
 #include <vector>
-
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::Shape;
+#include "paddle/fluid/inference/anakin/convert/helper.h"
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void FcBaseOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void FcBaseOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   auto input_names = op_desc.InputNames();
-  bool with_bias = input_names.size() == 3;
+  bool with_bias = input_names.size() >= 3;
 
   std::string w_name = "Y";
   std::string i_name = "X";
@@ -45,7 +42,12 @@ void FcBaseOpConverter<TargetT>::operator()(
   // get weights
   auto *y_v = scope.FindVar(op_desc.Input(w_name).front());
   PADDLE_ENFORCE_NOT_NULL(y_v);
-  auto *y_t = y_v->GetMutable<framework::LoDTensor>();
+  auto weight_tensor = tensor_from_var(*y_v, platform::CPUPlace());
+  auto weight_shape = framework::vectorize2int(weight_tensor->dims());
+
+  int out_dim = weight_shape[1];
+  const int w_m = weight_shape[0];
+  const int w_k = weight_shape[1];
 
   auto input_name = op_desc.Input(i_name).front();
   auto output_name = op_desc.Output("Out").front();
@@ -53,64 +55,58 @@ void FcBaseOpConverter<TargetT>::operator()(
   this->engine_->AddOp(op_name, "Dense", {input_name}, {output_name});
   this->engine_->AddOpAttr(op_name, "bias_term", with_bias);
   this->engine_->AddOpAttr(op_name, "axis", 1);
-
-  auto weight_shape = framework::vectorize2int(y_t->dims());
-  int out_dim = weight_shape[1];
   this->engine_->AddOpAttr(op_name, "out_dim", out_dim);
-  const int w_m = weight_shape[0];
-  const int w_k = weight_shape[1];
-
-  if (weight_shape.size() < 4UL) {
-    weight_shape.insert(weight_shape.begin(), 4UL - weight_shape.size(), 1);
-  }
-  Shape anakin_shape(weight_shape);
 
-  framework::LoDTensor weight_tensor;
-  weight_tensor.Resize(y_t->dims());
-  TensorCopySync((*y_t), platform::CPUPlace(), &weight_tensor);
-  auto *weight_data = weight_tensor.data<float>();
-  PADDLE_ENFORCE(w_m * w_k == weight_tensor.numel());
+  auto *weight_data = weight_tensor->data<float>();
+  PADDLE_ENFORCE(w_m * w_k == weight_tensor->numel());
 
-  std::vector<float> trans_weight_data(weight_tensor.numel());
+  std::vector<float> trans_weight_data(weight_tensor->numel());
   for (int i = 0; i < w_m; i++) {
     for (int j = 0; j < w_k; j++) {
       trans_weight_data[i + j * w_m] = weight_data[i * w_k + j];
     }
   }
-  auto *weight1 =
-      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-          anakin_shape);
-  float *cpu_data = static_cast<float *>(weight1->h_tensor().mutable_data());
-  std::copy_n(trans_weight_data.data(), weight_tensor.numel(), cpu_data);
+
+  int weight_num = weight_tensor->numel();
+  bool enable_int8 = boost::get<bool>(op_desc.HasAttr("enable_int8"));
+  if (enable_int8) {
+    if (weight_shape.size() < 4UL) {
+      weight_shape.insert(weight_shape.begin(), 4UL - weight_shape.size(), 1);
+    }
+    ::anakin::saber::Shape anakin_shape(weight_shape);
+    const float int8_range = 127.;
+    float in_scale = boost::get<float>(op_desc.GetAttr("input_scale"));
+    float weight_scale = boost::get<float>(op_desc.GetAttr("weight_scale"));
+    auto *weight1 = ::anakin::graph::GraphGlobalMem<TargetT>::Global()
+                        .template new_block<::anakin::AK_INT8>(anakin_shape);
+    std::vector<char> weight_int8;
+    for (int i = 0; i < weight_num; i++) {
+      bool is_valid_int8 =
+          ((trans_weight_data[i] >= -128) && (trans_weight_data[i] <= 127));
+      PADDLE_ENFORCE(is_valid_int8,
+                     "We are in anakin subgraph int8 mode, the weight of fc "
+                     "should be in range [-128, 127]");
+      weight_int8.push_back(static_cast<char>(trans_weight_data[i]));
+    }
+    memcpy(static_cast<void *>(weight1->h_tensor().mutable_data()),
+           static_cast<void *>(weight_int8.data()), sizeof(char) * weight_num);
     weight1->d_tensor().set_shape(anakin_shape);
     weight1->d_tensor().copy_from(weight1->h_tensor());
     this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
+    this->engine_->Graph()->SetOpPrec(op_name, ::anakin::AK_INT8);
+    this->engine_->Graph()->SetWeightsScale(op_name,
+                                            {weight_scale / int8_range}, false);
+    this->engine_->AddTensorScale(input_name, in_scale / int8_range);
+  } else {
+    auto *weight1 = pblock_from_vector<TargetT>(trans_weight_data);
+    this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
+  }
 
   // get bias
   if (with_bias) {
     auto *b_v = scope.FindVar(op_desc.Input("Bias").front());
     PADDLE_ENFORCE_NOT_NULL(b_v);
-    auto *b_t = b_v->GetMutable<framework::LoDTensor>();
-
-    auto bias_shape = framework::vectorize2int(b_t->dims());
-    framework::LoDTensor bias_tensor;
-    bias_tensor.Resize(b_t->dims());
-    TensorCopySync((*b_t), platform::CPUPlace(), &bias_tensor);
-    auto *bias_data = bias_tensor.data<float>();
-    bias_shape.insert(bias_shape.begin(), 1);
-    bias_shape.insert(bias_shape.begin(), 1);
-    bias_shape.insert(bias_shape.begin(), 1);
-    // bias_shape.push_back(1);
-    // bias_shape.push_back(1);
-    Shape anakin_bias_shape(bias_shape);
-
-    auto *weight2 =
-        GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
-            anakin_bias_shape);
-    float *cpu_data2 = static_cast<float *>(weight2->h_tensor().mutable_data());
-    std::copy_n(bias_data, bias_tensor.numel(), cpu_data2);
-    weight2->d_tensor().set_shape(anakin_bias_shape);
-    weight2->d_tensor().copy_from(weight2->h_tensor());
+    auto weight2 = pblock_from_var<TargetT>(*b_v);
     this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
   }
 }
@@ -120,9 +116,39 @@ void FcBaseOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(mul, MulOpConverter<::anakin::saber::NV>);
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(fc, FcOpConverter<::anakin::saber::NV>);
+using mul_nv_fp32 =
+    ::paddle::inference::anakin::MulOpConverter<::anakin::saber::NV,
+                                                ::anakin::Precision::FP32>;
+using fc_nv_fp32 =
+    ::paddle::inference::anakin::FcOpConverter<::anakin::saber::NV,
+                                               ::anakin::Precision::FP32>;
+using mul_nv_int8 =
+    ::paddle::inference::anakin::MulOpConverter<::anakin::saber::NV,
+                                                ::anakin::Precision::INT8>;
+using fc_nv_int8 =
+    ::paddle::inference::anakin::FcOpConverter<::anakin::saber::NV,
+                                               ::anakin::Precision::INT8>;
+
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(mul, mul_nv_fp32);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(fc, fc_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(mul, mul_nv_int8);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(fc, fc_nv_int8);
 #endif
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(mul, MulOpConverter<::anakin::saber::X86>);
-REGISTER_CPU_ANAKIN_OP_CONVERTER(fc, FcOpConverter<::anakin::saber::X86>);
+using mul_cpu_fp32 =
+    ::paddle::inference::anakin::MulOpConverter<::anakin::saber::X86,
+                                                ::anakin::Precision::FP32>;
+using fc_cpu_fp32 =
+    ::paddle::inference::anakin::FcOpConverter<::anakin::saber::X86,
+                                               ::anakin::Precision::FP32>;
+using mul_cpu_int8 =
+    ::paddle::inference::anakin::MulOpConverter<::anakin::saber::X86,
+                                                ::anakin::Precision::INT8>;
+using fc_cpu_int8 =
+    ::paddle::inference::anakin::FcOpConverter<::anakin::saber::X86,
+                                               ::anakin::Precision::INT8>;
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(mul, mul_cpu_fp32);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(fc, fc_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(mul, mul_cpu_int8);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(fc, fc_cpu_int8);

paddle/fluid/inference/anakin/convert/fc.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class FcBaseOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class FcBaseOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   FcBaseOpConverter() = default;
 
@@ -33,15 +33,15 @@ class FcBaseOpConverter : public AnakinOpConverter<TargetT> {
 };
 
 // with bias
-template <typename TargetT>
-class FcOpConverter : public FcBaseOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class FcOpConverter : public FcBaseOpConverter<TargetT, PrecisionT> {
  public:
   FcOpConverter() = default;
 };
 
 // without bias
-template <typename TargetT>
-class MulOpConverter : public FcBaseOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class MulOpConverter : public FcBaseOpConverter<TargetT, PrecisionT> {
  public:
   MulOpConverter() = default;
 };

paddle/fluid/inference/anakin/convert/flatten.cc

@@ -21,8 +21,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void FlattenOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void FlattenOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -46,8 +46,21 @@ void FlattenOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(flatten,
-                                  FlattenOpConverter<::anakin::saber::NV>);
+using flatten_nv_fp32 =
+    ::paddle::inference::anakin::FlattenOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::FP32>;
+using flatten_nv_int8 =
+    ::paddle::inference::anakin::FlattenOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::INT8>;
+
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(flatten, flatten_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(flatten, flatten_nv_int8);
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(flatten,
-                                 FlattenOpConverter<::anakin::saber::X86>);
+using flatten_cpu_fp32 =
+    ::paddle::inference::anakin::FlattenOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::FP32>;
+using flatten_cpu_int8 =
+    ::paddle::inference::anakin::FlattenOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(flatten, flatten_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(flatten, flatten_cpu_int8);

paddle/fluid/inference/anakin/convert/flatten.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class FlattenOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class FlattenOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   FlattenOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/helper.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/anakin/convert/helper.h"
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+std::unique_ptr<framework::LoDTensor> tensor_from_var(
+    const framework::Variable& var, const platform::Place& place) {
+  auto& src = var.Get<framework::LoDTensor>();
+  std::unique_ptr<framework::LoDTensor> dst(new framework::LoDTensor());
+  dst->Resize(src.dims());
+  TensorCopySync((src), place, dst.get());
+  return dst;
+}
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/anakin/convert/helper.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+#include <algorithm>
+#include <map>
+#include <memory>
+#include <vector>
+
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/variable.h"
+
+#include "framework/core/net/net.h"
+#include "framework/core/types.h"
+#include "framework/graph/graph.h"
+#include "framework/graph/graph_global_mem.h"
+#include "saber/saber_types.h"
+
+using anakin::saber::Shape;
+using anakin::AK_FLOAT;
+using anakin::PBlock;
+using anakin::graph::GraphGlobalMem;
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+std::unique_ptr<framework::LoDTensor> tensor_from_var(
+    const framework::Variable& var, const platform::Place& place);
+template <typename T>
+PBlock<T>* pblock_from_tensor(const framework::LoDTensor& tensor,
+                              std::vector<int> shape) {
+  while (shape.size() < 4) {
+    shape.insert(shape.begin(), 1);
+  }
+  Shape anakin_shape(shape);
+  auto* weight =
+      GraphGlobalMem<T>::Global().template new_block<AK_FLOAT>(anakin_shape);
+  float* cpu_data = static_cast<float*>(weight->h_tensor().mutable_data());
+  std::copy_n(tensor.data<float>(), tensor.numel(), cpu_data);
+  weight->d_tensor().set_shape(anakin_shape);
+  weight->d_tensor().copy_from(weight->h_tensor());
+  return weight;
+}
+
+template <typename T>
+PBlock<T>* pblock_from_vector(const std::vector<float>& vec,
+                              std::vector<int> shape_vec) {
+  while (shape_vec.size() < 4) {
+    shape_vec.insert(shape_vec.begin(), 1);
+  }
+  Shape shape(shape_vec);
+  auto* weight =
+      GraphGlobalMem<T>::Global().template new_block<AK_FLOAT>(shape);
+  auto* weight_data = static_cast<float*>(weight->h_tensor().mutable_data());
+  std::copy(std::begin(vec), std::end(vec), weight_data);
+  weight->d_tensor().set_shape(shape);
+  weight->d_tensor().copy_from(weight->h_tensor());
+  return weight;
+}
+
+template <typename T>
+PBlock<T>* pblock_from_vector(const std::vector<float>& vec) {
+  int size = vec.size();
+  return pblock_from_vector<T>(vec, std::vector<int>({1, 1, 1, size}));
+}
+
+template <typename T>
+PBlock<T>* pblock_from_var(const framework::Variable& var) {
+  auto tensor = tensor_from_var(var, platform::CPUPlace());
+  auto shape = framework::vectorize2int(tensor->dims());
+  return pblock_from_tensor<T>(*tensor, shape);
+}
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/anakin/convert/im2sequence.cc

@@ -23,8 +23,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void Im2SequenceConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void Im2SequenceConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -55,5 +55,18 @@ void Im2SequenceConverter<TargetT>::operator()(
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(im2sequence,
-                                  Im2SequenceConverter<::anakin::saber::NV>);
+#ifdef PADDLE_WITH_CUDA
+using im2sequence_nv_fp32 = ::paddle::inference::anakin::Im2SequenceConverter<
+    ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using im2sequence_nv_int8 = ::paddle::inference::anakin::Im2SequenceConverter<
+    ::anakin::saber::NV, ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(im2sequence, im2sequence_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(im2sequence, im2sequence_nv_int8);
+#endif
+
+using im2sequence_cpu_fp32 = ::paddle::inference::anakin::Im2SequenceConverter<
+    ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using im2sequence_cpu_int8 = ::paddle::inference::anakin::Im2SequenceConverter<
+    ::anakin::saber::X86, ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(im2sequence, im2sequence_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(im2sequence, im2sequence_cpu_int8);

paddle/fluid/inference/anakin/convert/im2sequence.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class Im2SequenceConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class Im2SequenceConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   Im2SequenceConverter() = default;
 

paddle/fluid/inference/anakin/convert/op_converter.h

@@ -32,9 +32,9 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
+template <typename TargetT, ::anakin::Precision PrecisionT>
 class AnakinOpConverter {
-  using AnakinEngineT = AnakinEngine<TargetT, ::anakin::Precision::FP32>;
+  using AnakinEngineT = AnakinEngine<TargetT, PrecisionT>;
 
  public:
   AnakinOpConverter() = default;
@@ -96,6 +96,13 @@ class AnakinOpConverter {
       engine->Graph()->RegistVar(output);
     }
     engine->Freeze();
+    // Add scale for tensor in int8 mode.
+    auto tensor_scales = engine->GetTensorScales();
+
+    for (auto &item : tensor_scales) {
+      engine->Graph()->SetVarScale(item.first, item.second);
+    }
+
     for (auto &input : inputs) {
       if (parameters.count(input)) continue;
       std::vector<int> input_shape;
@@ -136,52 +143,78 @@ class AnakinOpConverter {
   AnakinEngineT *engine_{nullptr};
 
  private:
-  std::unordered_map<std::string, AnakinOpConverter<TargetT> *> converters_;
+  std::unordered_map<std::string, AnakinOpConverter<TargetT, PrecisionT> *>
+      converters_;
   framework::Scope *scope_{nullptr};
   std::mutex mutex_;
 };
 
-template class AnakinOpConverter<::anakin::saber::NV>;
-template class AnakinOpConverter<::anakin::saber::X86>;
+template class AnakinOpConverter<::anakin::saber::NV,
+                                 ::anakin::Precision::FP32>;
+template class AnakinOpConverter<::anakin::saber::NV,
+                                 ::anakin::Precision::INT8>;
+
+template class AnakinOpConverter<::anakin::saber::X86,
+                                 ::anakin::Precision::FP32>;
+template class AnakinOpConverter<::anakin::saber::X86,
+                                 ::anakin::Precision::INT8>;
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 #define REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__,              \
-                                          place_type__, place_class__)         \
-  struct anakin_##op_type__##_##place_type__##_converter                       \
+                                          place_type__, place_class__,         \
+                                          precision_type__, precision_class__) \
+  struct anakin_##op_type__##_##place_type__##_##precision_type__##_converter  \
       : public ::paddle::framework::Registrar {                                \
-    anakin_##op_type__##_##place_type__##_converter() {                        \
+    anakin_##op_type__##_##place_type__##_##precision_type__##_converter() {   \
       LOG(INFO) << "register convert " << #op_type__ << " ";                   \
       ::paddle::inference::Registry<                                           \
-          ::paddle::inference::anakin::AnakinOpConverter<place_class__>>::     \
-          Global()                                                             \
-              .Register<::paddle::inference::anakin::Converter__>(#op_type__); \
+          ::paddle::inference::anakin::AnakinOpConverter<                      \
+              place_class__, precision_class__>>::Global()                     \
+          .Register<Converter__>(#op_type__);                                  \
     }                                                                          \
   };                                                                           \
-  anakin_##op_type__##_##place_type__##_converter                              \
-      anakin_##op_type__##_##place_type__##_converter__;                       \
-  int TouchConverterRegister_anakin_##op_type__##_##place_type__() {           \
-    anakin_##op_type__##_##place_type__##_converter__.Touch();                 \
+  anakin_##op_type__##_##place_type__##_##precision_type__##_converter         \
+      anakin_##op_type__##_##place_type__##_##precision_type__##_converter__;  \
+  int Touch_anakin_##op_type__##_##place_type__##_##precision_type__() {       \
+    anakin_##op_type__##_##place_type__##_##precision_type__##_converter__     \
+        .Touch();                                                              \
     return 0;                                                                  \
   }
 
 #define REGISTER_CUDA_ANAKIN_OP_CONVERTER(op_type__, Converter__) \
   REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__, CUDA, \
-                                    ::anakin::saber::NV)
+                                    ::anakin::saber::NV, FP32,    \
+                                    ::anakin::Precision::FP32)
+
+#define REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(op_type__, Converter__) \
+  REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__, CUDA,      \
+                                    ::anakin::saber::NV, INT8,         \
+                                    ::anakin::Precision::INT8)
 
 #define REGISTER_CPU_ANAKIN_OP_CONVERTER(op_type__, Converter__) \
   REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__, CPU, \
-                                    ::anakin::saber::X86)
+                                    ::anakin::saber::X86, FP32,  \
+                                    ::anakin::Precision::FP32)
+
+#define REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(op_type__, Converter__) \
+  REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__, CPU,      \
+                                    ::anakin::saber::X86, INT8,       \
+                                    ::anakin::Precision::INT8)
 
-#define USE_ANAKIN_CONVERTER_BASE(op_type__, place_type__)                 \
-  extern int TouchConverterRegister_anakin_##op_type__##_##place_type__(); \
-  int use_op_converter_anakin_##op_type__##_##place_type__                 \
+#define USE_ANAKIN_CONVERTER_BASE(op_type__, place_type__, precision_type__)   \
+  extern int Touch_anakin_##op_type__##_##place_type__##_##precision_type__(); \
+  int use_converter_anakin_##op_type__##_##place_type__##_##precision_type__   \
       __attribute__((unused)) =                                                \
-          TouchConverterRegister_anakin_##op_type__##_##place_type__();
+          Touch_anakin_##op_type__##_##place_type__##_##precision_type__();
 
 #define USE_ANAKIN_CONVERTER(op_type__) \
-  USE_ANAKIN_CONVERTER_BASE(op_type__, CUDA)
+  USE_ANAKIN_CONVERTER_BASE(op_type__, CUDA, FP32)
+#define USE_INT8_ANAKIN_CONVERTER(op_type__) \
+  USE_ANAKIN_CONVERTER_BASE(op_type__, CUDA, INT8)
 
 #define USE_CPU_ANAKIN_CONVERTER(op_type__) \
-  USE_ANAKIN_CONVERTER_BASE(op_type__, CPU)
+  USE_ANAKIN_CONVERTER_BASE(op_type__, CPU, FP32)
+#define USE_CPU_INT8_ANAKIN_CONVERTER(op_type__) \
+  USE_ANAKIN_CONVERTER_BASE(op_type__, CPU, INT8)

paddle/fluid/inference/anakin/convert/pool2d.cc

@@ -23,8 +23,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void Pool2dOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void Pool2dOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -72,8 +72,21 @@ void Pool2dOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(pool2d,
-                                  Pool2dOpConverter<::anakin::saber::NV>);
+using pool2d_nv_float32 =
+    ::paddle::inference::anakin::Pool2dOpConverter<::anakin::saber::NV,
+                                                   ::anakin::Precision::FP32>;
+using pool2d_nv_int8 =
+    ::paddle::inference::anakin::Pool2dOpConverter<::anakin::saber::NV,
+                                                   ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(pool2d, pool2d_nv_float32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(pool2d, pool2d_nv_int8);
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(pool2d,
-                                 Pool2dOpConverter<::anakin::saber::X86>);
+
+using pool2d_cpu_float32 =
+    ::paddle::inference::anakin::Pool2dOpConverter<::anakin::saber::X86,
+                                                   ::anakin::Precision::FP32>;
+using pool2d_cpu_int8 =
+    ::paddle::inference::anakin::Pool2dOpConverter<::anakin::saber::X86,
+                                                   ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(pool2d, pool2d_cpu_float32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(pool2d, pool2d_cpu_int8);

paddle/fluid/inference/anakin/convert/pool2d.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class Pool2dOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class Pool2dOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   Pool2dOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/relu.cc

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void ReluOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void ReluOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -36,8 +36,8 @@ void ReluOpConverter<TargetT>::operator()(
   this->engine_->AddOpAttr(op_name, "alpha", 0);
 }
 
-template <typename TargetT>
-void LeakyReluOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void LeakyReluOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -58,10 +58,35 @@ void LeakyReluOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(relu, ReluOpConverter<::anakin::saber::NV>);
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(leaky_relu,
-                                  LeakyReluOpConverter<::anakin::saber::NV>);
+using relu_nv_fp32 =
+    ::paddle::inference::anakin::ReluOpConverter<::anakin::saber::NV,
+                                                 ::anakin::Precision::FP32>;
+using leaky_nv_fp32 = ::paddle::inference::anakin::LeakyReluOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using relu_nv_int8 =
+    ::paddle::inference::anakin::ReluOpConverter<::anakin::saber::NV,
+                                                 ::anakin::Precision::INT8>;
+using leaky_nv_int8 = ::paddle::inference::anakin::LeakyReluOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::INT8>;
+
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(relu, relu_nv_fp32);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(leaky_relu, leaky_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(relu, relu_nv_int8);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(leaky_relu, leaky_nv_int8);
+
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(relu, ReluOpConverter<::anakin::saber::X86>);
-REGISTER_CPU_ANAKIN_OP_CONVERTER(leaky_relu,
-                                 LeakyReluOpConverter<::anakin::saber::X86>);
+
+using relu_cpu_fp32 =
+    ::paddle::inference::anakin::ReluOpConverter<::anakin::saber::X86,
+                                                 ::anakin::Precision::FP32>;
+using leaky_cpu_fp32 = ::paddle::inference::anakin::LeakyReluOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using relu_cpu_int8 =
+    ::paddle::inference::anakin::ReluOpConverter<::anakin::saber::X86,
+                                                 ::anakin::Precision::INT8>;
+using leaky_cpu_int8 = ::paddle::inference::anakin::LeakyReluOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(relu, relu_cpu_fp32);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(leaky_relu, leaky_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(relu, relu_cpu_int8);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(leaky_relu, leaky_cpu_int8);

paddle/fluid/inference/anakin/convert/relu.h

@@ -22,8 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class ReluOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class ReluOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   ReluOpConverter() = default;
 
@@ -34,8 +34,8 @@ class ReluOpConverter : public AnakinOpConverter<TargetT> {
   virtual ~ReluOpConverter() {}
 };
 
-template <typename TargetT>
-class LeakyReluOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class LeakyReluOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   LeakyReluOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/reshape.cc

@@ -21,8 +21,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void ReshapeOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void ReshapeOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -47,9 +47,21 @@ void ReshapeOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(reshape,
-                                  ReshapeOpConverter<::anakin::saber::NV>);
+using reshape_nv_fp32 =
+    ::paddle::inference::anakin::ReshapeOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::FP32>;
+using reshape_nv_int8 =
+    ::paddle::inference::anakin::ReshapeOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(reshape, reshape_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(reshape, reshape_nv_int8);
 #endif
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(reshape,
-                                 ReshapeOpConverter<::anakin::saber::X86>);
+using reshape_cpu_fp32 =
+    ::paddle::inference::anakin::ReshapeOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::FP32>;
+using reshape_cpu_int8 =
+    ::paddle::inference::anakin::ReshapeOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(reshape, reshape_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(reshape, reshape_cpu_int8);

paddle/fluid/inference/anakin/convert/reshape.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class ReshapeOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class ReshapeOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   ReshapeOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/roi_align.cc

@@ -16,17 +16,12 @@
 #include <algorithm>
 #include <map>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
-
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void RoiAlignOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void RoiAlignOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -57,8 +52,21 @@ void RoiAlignOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(roi_align,
-                                  RoiAlignOpConverter<::anakin::saber::NV>);
+using roi_align_nv_fp32 =
+    ::paddle::inference::anakin::RoiAlignOpConverter<::anakin::saber::NV,
+                                                     ::anakin::Precision::FP32>;
+using roi_align_nv_int8 =
+    ::paddle::inference::anakin::RoiAlignOpConverter<::anakin::saber::NV,
+                                                     ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(roi_align, roi_align_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(roi_align, roi_align_nv_int8);
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(roi_align,
-                                 RoiAlignOpConverter<::anakin::saber::X86>);
+
+using roi_align_cpu_fp32 =
+    ::paddle::inference::anakin::RoiAlignOpConverter<::anakin::saber::X86,
+                                                     ::anakin::Precision::FP32>;
+using roi_align_cpu_int8 =
+    ::paddle::inference::anakin::RoiAlignOpConverter<::anakin::saber::X86,
+                                                     ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(roi_align, roi_align_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(roi_align, roi_align_cpu_int8);

paddle/fluid/inference/anakin/convert/roi_align.h

@@ -22,8 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class RoiAlignOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class RoiAlignOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   RoiAlignOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/scale.cc

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void ScaleOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void ScaleOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -49,4 +49,22 @@ void ScaleOpConverter<TargetT>::operator()(
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(scale, ScaleOpConverter<::anakin::saber::NV>);
+#ifdef PADDLE_WITH_CUDA
+using scale_nv_fp32 =
+    ::paddle::inference::anakin::ScaleOpConverter<::anakin::saber::NV,
+                                                  ::anakin::Precision::FP32>;
+using scale_nv_int8 =
+    ::paddle::inference::anakin::ScaleOpConverter<::anakin::saber::NV,
+                                                  ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(scale, scale_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(scale, scale_nv_int8);
+#endif
+
+using scale_cpu_fp32 =
+    ::paddle::inference::anakin::ScaleOpConverter<::anakin::saber::X86,
+                                                  ::anakin::Precision::FP32>;
+using scale_cpu_int8 =
+    ::paddle::inference::anakin::ScaleOpConverter<::anakin::saber::X86,
+                                                  ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(scale, scale_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(scale, scale_cpu_int8);

paddle/fluid/inference/anakin/convert/scale.h

@@ -22,8 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class ScaleOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class ScaleOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   ScaleOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/softmax.cc

@@ -18,8 +18,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void SoftMaxOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void SoftMaxOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -45,9 +45,22 @@ void SoftMaxOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(softmax,
-                                  SoftMaxOpConverter<::anakin::saber::NV>);
+using sm_nv_fp32 =
+    ::paddle::inference::anakin::SoftMaxOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::FP32>;
+using sm_nv_int8 =
+    ::paddle::inference::anakin::SoftMaxOpConverter<::anakin::saber::NV,
+                                                    ::anakin::Precision::INT8>;
+
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(softmax, sm_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(softmax, sm_nv_int8);
 #endif
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(softmax,
-                                 SoftMaxOpConverter<::anakin::saber::X86>);
+using sm_cpu_fp32 =
+    ::paddle::inference::anakin::SoftMaxOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::FP32>;
+using sm_cpu_int8 =
+    ::paddle::inference::anakin::SoftMaxOpConverter<::anakin::saber::X86,
+                                                    ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(softmax, sm_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(softmax, sm_cpu_int8);

paddle/fluid/inference/anakin/convert/softmax.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class SoftMaxOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class SoftMaxOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   SoftMaxOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/split.cc

@@ -22,8 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void SplitOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void SplitOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -56,7 +56,22 @@ void SplitOpConverter<TargetT>::operator()(
 }  // namespace inference
 }  // namespace paddle
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(split, SplitOpConverter<::anakin::saber::NV>);
+using split_nv_fp32 =
+    ::paddle::inference::anakin::SplitOpConverter<::anakin::saber::NV,
+                                                  ::anakin::Precision::FP32>;
+using split_nv_int8 =
+    ::paddle::inference::anakin::SplitOpConverter<::anakin::saber::NV,
+                                                  ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(split, split_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(split, split_nv_int8);
 #endif
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(split, SplitOpConverter<::anakin::saber::X86>);
+using split_cpu_fp32 =
+    ::paddle::inference::anakin::SplitOpConverter<::anakin::saber::X86,
+                                                  ::anakin::Precision::FP32>;
+using split_cpu_int8 =
+    ::paddle::inference::anakin::SplitOpConverter<::anakin::saber::X86,
+                                                  ::anakin::Precision::INT8>;
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(split, split_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(split, split_cpu_int8);

paddle/fluid/inference/anakin/convert/split.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class SplitOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class SplitOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   SplitOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/sum.cc

@@ -23,11 +23,10 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void SumOpConverter<TargetT>::operator()(const framework::proto::OpDesc &op,
-                                         const framework::BlockDesc &block_desc,
-                                         const framework::Scope &scope,
-                                         bool test_mode) {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void SumOpConverter<TargetT, PrecisionT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 2);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
@@ -49,6 +48,21 @@ void SumOpConverter<TargetT>::operator()(const framework::proto::OpDesc &op,
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(sum, SumOpConverter<::anakin::saber::NV>);
+using sum_nv_fp32 =
+    ::paddle::inference::anakin::SumOpConverter<::anakin::saber::NV,
+                                                ::anakin::Precision::FP32>;
+using sum_nv_int8 =
+    ::paddle::inference::anakin::SumOpConverter<::anakin::saber::NV,
+                                                ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(sum, sum_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(sum, sum_nv_int8);
 #endif
-REGISTER_CPU_ANAKIN_OP_CONVERTER(sum, SumOpConverter<::anakin::saber::X86>);
+
+using sum_cpu_fp32 =
+    ::paddle::inference::anakin::SumOpConverter<::anakin::saber::X86,
+                                                ::anakin::Precision::FP32>;
+using sum_cpu_int8 =
+    ::paddle::inference::anakin::SumOpConverter<::anakin::saber::X86,
+                                                ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(sum, sum_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(sum, sum_cpu_int8);

paddle/fluid/inference/anakin/convert/sum.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class SumOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class SumOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   SumOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/test_activation_op.cc

@@ -27,8 +27,8 @@ static void test_activation_op(const std::string& op_type,
                                bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("act-X", {10, 6, 1, 1});
   validator.DeclOutputVar("act-Out", {10, 6, 1, 1});
   framework::OpDesc desc;
@@ -57,6 +57,7 @@ TEST(tanh_op, gpu) {
 }
 #endif
 
+/*
 TEST(sigm_op, cpu) {
   platform::CPUPlace cpu_place;
   platform::CPUDeviceContext ctx(cpu_place);
@@ -68,6 +69,7 @@ TEST(tanh_op, cpu) {
   platform::CPUDeviceContext ctx(cpu_place);
   test_activation_op<::anakin::saber::X86>("tanh", ctx, false);
 }
+*/
 
 }  // namespace anakin
 }  // namespace inference

paddle/fluid/inference/anakin/convert/test_affine_channel_op.cc

@@ -28,8 +28,8 @@ void test_affine_channel_op(const platform::DeviceContext& context,
   std::unordered_set<std::string> parameters({"scale", "bias"});
 
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("x", {1, 3, 5, 2});
   validator.DeclOutputVar("out", {1, 3, 5, 2});
   validator.DeclParamVar("scale", {3});

paddle/fluid/inference/anakin/convert/test_batch_norm_op.cc

@@ -25,8 +25,8 @@ void test_batchnorm_op(const platform::DeviceContext& context, bool use_gpu) {
       {"batch_norm_scale", "batch_norm_bias", "batch_norm_mean",
        "batch_norm_variance"});
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   std::vector<int> param_shape{2};
 
   validator.DeclInputVar("batch_norm_X", {1, 2, 5, 5});

paddle/fluid/inference/anakin/convert/test_concat_op.cc

@@ -25,8 +25,8 @@ template <typename TargetT>
 void test_concat_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters({""});
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("concat_x1", {1, 2, 1, 1});
   validator.DeclInputVar("concat_x2", {1, 3, 1, 1});
   validator.DeclInputVar("concat_x3", {1, 1, 1, 1});

paddle/fluid/inference/anakin/convert/test_conv2d_op.cc

@@ -25,8 +25,8 @@ template <typename TargetT>
 void test_conv2d_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters({"conv2d-Y"});
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("conv2d-X", {1, 3, 3, 3});
   validator.DeclParamVar("conv2d-Y", {4, 3, 1, 1});
   validator.DeclOutputVar("conv2d-Out", {1, 4, 3, 3});

paddle/fluid/inference/anakin/convert/test_dropout_op.cc

@@ -25,8 +25,8 @@ template <typename TargetT>
 void test_dropout_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("x", {1, 1, 2, 2});
   validator.DeclOutputVar("out", {1, 1, 2, 2});
   validator.DeclOutputVar("mask", {1, 1, 2, 2});

paddle/fluid/inference/anakin/convert/test_elementwise_op.cc

@@ -27,8 +27,8 @@ static void test_elementwise_op(const std::string& op_type,
                                 bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("x", {1, 1, 2, 2});
   validator.DeclInputVar("y", {1, 1, 2, 2});
   validator.DeclOutputVar("out", {1, 1, 2, 2});

paddle/fluid/inference/anakin/convert/test_fc_op.cc

@@ -25,8 +25,8 @@ void test_mul_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters({"mul_y"});
   framework::Scope scope;
 
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("mul_x", {1, 1, 2, 2});
   validator.DeclParamVar("mul_y", {4, 2});
   validator.DeclOutputVar("mul_out", {1, 2});

paddle/fluid/inference/anakin/convert/test_flatten_op.cc

@@ -24,8 +24,8 @@ template <typename TargetT>
 void test_flatten_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("flatten-X", {3, 10, 10, 4});
   validator.DeclOutputVar("flatten-Out", {3, 400, 1, 1});
   framework::OpDesc desc;

paddle/fluid/inference/anakin/convert/test_pool2d_op.cc

@@ -25,8 +25,8 @@ void test_pool2d(const platform::DeviceContext& context, bool use_gpu,
                  std::string pool_type = "max") {
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
 
   // The ITensor's Dims should not contain the batch size.
   // So, the ITensor's Dims of input and output should be C * H * W.

paddle/fluid/inference/anakin/convert/test_relu_op.cc

@@ -27,8 +27,8 @@ static void test_activation_op(const std::string& op_type,
                                bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("act-X", {10, 6, 1, 1});
   validator.DeclOutputVar("act-Out", {10, 6, 1, 1});
   framework::OpDesc desc;
@@ -60,20 +60,6 @@ TEST(leaky_relu_op, gpu) {
 }
 #endif
 
-/* seems bug here
-TEST(relu_op, cpu) {
-  platform::CPUPlace cpu_place;
-  platform::CPUDeviceContext ctx(cpu_place);
-  test_activation_op<::anakin::saber::X86>("relu", ctx, false);
-}
-
-TEST(leaky_relu_op, cpu) {
-  platform::CPUPlace cpu_place;
-  platform::CPUDeviceContext ctx(cpu_place);
-  test_activation_op<::anakin::saber::X86>("leaky_relu", ctx, false);
-}
-*/
-
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/anakin/convert/test_reshape_op.cc

@@ -24,8 +24,8 @@ template <typename TargetT>
 void test_reshape1_op(const platform::DeviceContext& context, bool use_gpu) {
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
 
   // validator.DeclInputVar("reshape-X", {2, 3, 3, 1});
   // validator.DeclOutputVar("reshape-Out", {3, 2, 1, 3});
@@ -49,8 +49,8 @@ template <typename TargetT>
 void test_reshape2_op(const platform::DeviceContext& context, bool use_gpu) {
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
 
   validator.DeclInputVar("reshape-X", {1, 2, 4});
   validator.DeclOutputVar("reshape-Out", {1, 4, 2});

paddle/fluid/inference/anakin/convert/test_softmax_op.cc

@@ -24,8 +24,8 @@ template <typename TargetT>
 void test_softmax_op(const platform::DeviceContext& context, bool use_gpu) {
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
 
   validator.DeclInputVar("softmax-X", {1, 10, 2});
   validator.DeclOutputVar("softmax-Out", {1, 10, 2});

paddle/fluid/inference/anakin/convert/test_split_op.cc

@@ -27,8 +27,8 @@ void AnakinSliceTest(const platform::DeviceContext &context, bool use_gpu,
                      const std::vector<int> &sections) {
   std::unordered_set<std::string> parameters({""});
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
 
   validator.DeclInputVar("split_input", in_shape);
   std::vector<std::string> output_vars;

paddle/fluid/inference/anakin/convert/test_sum_op.cc

@@ -26,8 +26,8 @@ template <typename TargetT>
 static void test_sum_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("sum_x1", {1, 2, 1, 2});
   validator.DeclInputVar("sum_x2", {1, 2, 1, 2});
   validator.DeclOutputVar("sum_out", {1, 2, 1, 2});

paddle/fluid/inference/anakin/convert/test_transpose_op.cc

@@ -24,8 +24,8 @@ template <typename TargetT>
 void test_transpose1_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("transpose-X", {2, 3, 4, 5});
   validator.DeclOutputVar("transpose-Out", {4, 2, 5, 3});
 
@@ -47,8 +47,8 @@ template <typename TargetT>
 void test_transpose2_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
-                                             use_gpu);
+  AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
+      parameters, &scope, context, use_gpu);
   validator.DeclInputVar("transpose-X", {3, 4, 5});
   validator.DeclOutputVar("transpose-Out", {3, 5, 4});
 

paddle/fluid/inference/anakin/convert/transpose.cc

@@ -23,8 +23,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-void TransposeOpConverter<TargetT>::operator()(
+template <typename TargetT, ::anakin::Precision PrecisionT>
+void TransposeOpConverter<TargetT, PrecisionT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -50,9 +50,17 @@ void TransposeOpConverter<TargetT>::operator()(
 }  // namespace paddle
 
 #ifdef PADDLE_WITH_CUDA
-REGISTER_CUDA_ANAKIN_OP_CONVERTER(transpose,
-                                  TransposeOpConverter<::anakin::saber::NV>);
+using transpose_nv_fp32 = ::paddle::inference::anakin::TransposeOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::FP32>;
+using transpose_nv_int8 = ::paddle::inference::anakin::TransposeOpConverter<
+    ::anakin::saber::NV, ::anakin::Precision::INT8>;
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(transpose, transpose_nv_fp32);
+REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(transpose, transpose_nv_int8);
 #endif
 
-REGISTER_CPU_ANAKIN_OP_CONVERTER(transpose,
-                                 TransposeOpConverter<::anakin::saber::X86>);
+using transpose_cpu_fp32 = ::paddle::inference::anakin::TransposeOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::FP32>;
+using transpose_cpu_int8 = ::paddle::inference::anakin::TransposeOpConverter<
+    ::anakin::saber::X86, ::anakin::Precision::INT8>;
+REGISTER_CPU_ANAKIN_OP_CONVERTER(transpose, transpose_cpu_fp32);
+REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(transpose, transpose_cpu_int8);

paddle/fluid/inference/anakin/convert/transpose.h

@@ -20,8 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <typename TargetT>
-class TransposeOpConverter : public AnakinOpConverter<TargetT> {
+template <typename TargetT, ::anakin::Precision PrecisionT>
+class TransposeOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
  public:
   TransposeOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/ut_helper.h

@@ -61,7 +61,7 @@ void RandomizeTensor(framework::LoDTensor* tensor,
   auto* temp_data = temp_tensor.mutable_data<float>(cpu_place);
 
   for (size_t i = 0; i < num_elements; i++) {
-    *(temp_data + i) = random(-128., 128.);
+    *(temp_data + i) = random(0., 1.);
   }
 
   TensorCopySync(temp_tensor, place, tensor);
@@ -72,9 +72,9 @@ void RandomizeTensor(framework::LoDTensor* tensor,
  * anakin
  * layer.
  */
-template <typename TargetT>
+template <typename TargetT, ::anakin::Precision PrecisionT>
 class AnakinConvertValidation {
-  using AnakinNvEngineT = AnakinEngine<TargetT, Precision::FP32>;
+  using AnakinNvEngineT = AnakinEngine<TargetT, PrecisionT>;
 
  public:
   AnakinConvertValidation() = delete;
@@ -84,7 +84,7 @@ class AnakinConvertValidation {
                           const platform::DeviceContext& ctx,
                           bool use_gpu = true)
       : parameters_(parameters), scope_(scope), ctx_(ctx), use_gpu_(use_gpu) {
-    engine_.reset(new AnakinEngine<TargetT, Precision::FP32>(true));
+    engine_.reset(new AnakinEngine<TargetT, PrecisionT>(true));
   }
 
   // Declare a Variable as input with random initialization.
@@ -127,7 +127,7 @@ class AnakinConvertValidation {
     // should init anakin engine here.
 
     auto& block_desc = program_desc_.Block(framework::kRootBlockIndex);
-    Singleton<AnakinOpConverter<TargetT>>::Global().ConvertOp(
+    Singleton<AnakinOpConverter<TargetT, PrecisionT>>::Global().ConvertOp(
         desc, block_desc, parameters_, *scope_, engine_.get(),
         true /*test_mode*/);
     engine_->Freeze();
@@ -213,8 +213,15 @@ class AnakinConvertValidation {
   bool use_gpu_{true};
 };
 
-template class AnakinConvertValidation<::anakin::saber::NV>;
-template class AnakinConvertValidation<::anakin::saber::X86>;
+template class AnakinConvertValidation<::anakin::saber::NV,
+                                       ::anakin::Precision::FP32>;
+template class AnakinConvertValidation<::anakin::saber::X86,
+                                       ::anakin::Precision::FP32>;
+
+template class AnakinConvertValidation<::anakin::saber::NV,
+                                       ::anakin::Precision::INT8>;
+template class AnakinConvertValidation<::anakin::saber::X86,
+                                       ::anakin::Precision::INT8>;
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/anakin/engine.cc

@@ -172,11 +172,20 @@ AnakinEngine<TargetT, PrecisionType, RunType>::Clone() {
 
 #ifdef PADDLE_WITH_CUDA
 template class AnakinEngine<::anakin::saber::NV, ::anakin::Precision::FP32>;
-template class AnakinEngineManager<::anakin::saber::NV>;
+template class AnakinEngineManager<::anakin::saber::NV,
+                                   ::anakin::Precision::FP32>;
+
+template class AnakinEngine<::anakin::saber::NV, ::anakin::Precision::INT8>;
+template class AnakinEngineManager<::anakin::saber::NV,
+                                   ::anakin::Precision::INT8>;
 #endif
 
 template class AnakinEngine<::anakin::saber::X86, ::anakin::Precision::FP32>;
-template class AnakinEngineManager<::anakin::saber::X86>;
+template class AnakinEngineManager<::anakin::saber::X86,
+                                   ::anakin::Precision::FP32>;
+template class AnakinEngine<::anakin::saber::X86, ::anakin::Precision::INT8>;
+template class AnakinEngineManager<::anakin::saber::X86,
+                                   ::anakin::Precision::INT8>;
 
 // template class AnakinEngine<::anakin::saber::X86, ::anakin::Precision::FP32>;
 }  // namespace anakin

paddle/fluid/inference/anakin/engine.h

@@ -93,6 +93,12 @@ class AnakinEngine {
   void Save(std::string path) { graph_->save(path); }
   bool IsInit() { return initialized_; }
   int GetDevice() { return device_; }
+  void AddTensorScale(const std::string &tensor_name, float scale) {
+    tensor_scales_[tensor_name] = scale;
+  }
+  std::unordered_map<std::string, float> GetTensorScales() {
+    return tensor_scales_;
+  }
   void Execute(const std::map<std::string, framework::LoDTensor *> &inputs,
                const std::map<std::string, framework::LoDTensor *> &outputs);
 #ifdef PADDLE_WITH_CUDA
@@ -112,11 +118,12 @@ class AnakinEngine {
   std::unique_ptr<GraphT> graph_;
   std::unique_ptr<NetT> net_;
   std::vector<std::string> program_inputs_;
+  std::unordered_map<std::string, float> tensor_scales_;
 };
 
-template <typename TargetT>
+template <typename TargetT, ::anakin::Precision PrecisionType>
 class AnakinEngineManager {
-  using AnakinEngineT = AnakinEngine<TargetT, Precision::FP32>;
+  using AnakinEngineT = AnakinEngine<TargetT, PrecisionType>;
 
  public:
   bool HasEngine(const std::string &name) const {
@@ -132,7 +139,7 @@ class AnakinEngineManager {
                         std::vector<std::string> program_inputs,
                         std::string engine_name) {
     std::unique_lock<std::mutex> lk(mut_);
-    auto *p = new AnakinEngine<TargetT, Precision::FP32>(
+    auto *p = new AnakinEngine<TargetT, PrecisionType>(
         need_summary, device, max_batch_size, max_input_shape, program_inputs);
     engines_[engine_name].reset(p);
     return p;

paddle/fluid/inference/analysis/argument.h

@@ -169,7 +169,13 @@ struct Argument {
                       anakin_max_shape_t);
   DECL_ARGUMENT_FIELD(anakin_max_batch_size, AnakinMaxBatchSize, int);
   DECL_ARGUMENT_FIELD(anakin_min_subgraph_size, AnakinMinSubgraphSize, int);
+  DECL_ARGUMENT_FIELD(anakin_precision_mode, AnakinPrecisionMode,
+                      AnalysisConfig::Precision);
   DECL_ARGUMENT_FIELD(use_anakin, UseAnakin, bool);
+  DECL_ARGUMENT_FIELD(anakin_passes_filter, AnakinPassesFilter,
+                      std::vector<std::string>);
+  DECL_ARGUMENT_FIELD(anakin_ops_filter, AnakinOpsFilter,
+                      std::vector<std::string>);
 
   // Memory optimized related.
   DECL_ARGUMENT_FIELD(enable_memory_optim, EnableMemoryOptim, bool);

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -123,6 +123,11 @@ void IRPassManager::CreatePasses(Argument *argument,
       pass->Set("max_input_shape", new std::map<std::string, std::vector<int>>(
                                        argument->anakin_max_input_shape()));
       pass->Set("max_batch_size", new int(argument->anakin_max_batch_size()));
+      bool enable_int8 =
+          argument->anakin_precision_mode() == AnalysisConfig::Precision::kInt8;
+      pass->Set("enable_int8", new bool(enable_int8));
+      pass->Set("anakin_ops_filter",
+                new std::vector<std::string>(argument->anakin_ops_filter()));
     }
 
     pre_pass = pass_name;

paddle/fluid/inference/analysis/ir_passes/anakin_subgraph_pass.cc

@@ -39,8 +39,14 @@ void analysis::AnakinSubgraphPass::ApplyImpl(
     framework::ir::Graph *graph) const {
   framework::ir::FusePassBase::Init("anakin_subgraph_pass", graph);
 
-  auto teller = [](const framework::ir::Node *node) {
-    if (!node->IsOp() || !node->Op()) return false;
+  auto &anakin_ops_filter = Get<std::vector<std::string>>("anakin_ops_filter");
+
+  auto teller = [&anakin_ops_filter](const framework::ir::Node *node) {
+    if (!node->IsOp() || !node->Op())
+      return false;
+    else if (std::find(anakin_ops_filter.begin(), anakin_ops_filter.end(),
+                       node->Op()->Type()) != anakin_ops_filter.end())
+      return false;
     return anakin::OpTeller::Global().Tell(node->Op()->Type(), *node->Op());
   };
 
@@ -191,47 +197,71 @@ void AnakinSubgraphPass::CreateAnakinOp(
   SetAttr(op_desc->Proto(), "engine_key", engine_key);
   auto max_input_shape =
       Get<std::map<std::string, std::vector<int>>>("max_input_shape");
-  auto max_batch_size = Get<int>("max_batch_size");
   auto program_inputs = program_desc->GetFeedTargetNames();
 
   bool use_gpu = Get<bool>("use_gpu");
   SetAttr(op_desc->Proto(), "use_gpu", use_gpu);
+  bool enable_int8 = Get<bool>("enable_int8");
+  SetAttr(op_desc->Proto(), "enable_int8", enable_int8);
+  if (enable_int8) {
+    CreateAnakinEngine<::anakin::Precision::INT8>(&block_desc, params,
+                                                  input_names, output_mapping,
+                                                  program_inputs, engine_key);
+  } else {
+    CreateAnakinEngine<::anakin::Precision::FP32>(&block_desc, params,
+                                                  input_names, output_mapping,
+                                                  program_inputs, engine_key);
+  }
+}
 
+template <::anakin::Precision PrecisionT>
+void AnakinSubgraphPass::CreateAnakinEngine(
+    framework::BlockDesc *block_desc, const std::vector<std::string> &params,
+    const std::set<std::string> &input_names,
+    const std::vector<std::string> &output_mapping,
+    const std::vector<std::string> &program_inputs,
+    const std::string &engine_key) const {
+  framework::BlockDesc block_desc_temp(nullptr, block_desc->Proto());
+  bool use_gpu = Get<bool>("use_gpu");
+  auto max_batch_size = Get<int>("max_batch_size");
+  auto max_input_shape =
+      Get<std::map<std::string, std::vector<int>>>("max_input_shape");
   if (use_gpu) {
 #ifdef PADDLE_WITH_CUDA
     inference::Singleton<
-        anakin::AnakinEngineManager<::anakin::saber::NV>>::Global()
+        anakin::AnakinEngineManager<::anakin::saber::NV, PrecisionT>>::Global()
         .Create(true, Get<int>("gpu_device_id"), max_batch_size,
                 max_input_shape, program_inputs, engine_key);
 #endif
   } else {
     inference::Singleton<
-        anakin::AnakinEngineManager<::anakin::saber::X86>>::Global()
+        anakin::AnakinEngineManager<::anakin::saber::X86, PrecisionT>>::Global()
         .Create(true, Get<int>("gpu_device_id"), max_batch_size,
                 max_input_shape, program_inputs, engine_key);
   }
 
   auto *scope = param_scope();
   std::unordered_set<std::string> param_set(params.begin(), params.end());
-  framework::BlockDesc block_desc_temp(nullptr, block_desc.Proto());
   if (use_gpu) {
+#ifdef PADDLE_WITH_CUDA
     auto *anakin_engine =
         inference::Singleton<inference::anakin::AnakinEngineManager<
-            ::anakin::saber::NV>>::Global()
+            ::anakin::saber::NV, PrecisionT>>::Global()
             .Get(engine_key);
-    inference::Singleton<
-        inference::anakin::AnakinOpConverter<::anakin::saber::NV>>::Global()
+    inference::Singleton<inference::anakin::AnakinOpConverter<
+        ::anakin::saber::NV, PrecisionT>>::Global()
         .ConvertBlockToAnakinEngine(
             &block_desc_temp, scope,
             std::vector<std::string>(input_names.begin(), input_names.end()),
             param_set, output_mapping, anakin_engine);
+#endif
   } else {
     auto *anakin_engine =
         inference::Singleton<inference::anakin::AnakinEngineManager<
-            ::anakin::saber::X86>>::Global()
+            ::anakin::saber::X86, PrecisionT>>::Global()
             .Get(engine_key);
-    inference::Singleton<
-        inference::anakin::AnakinOpConverter<::anakin::saber::X86>>::Global()
+    inference::Singleton<inference::anakin::AnakinOpConverter<
+        ::anakin::saber::X86, PrecisionT>>::Global()
         .ConvertBlockToAnakinEngine(
             &block_desc_temp, scope,
             std::vector<std::string>(input_names.begin(), input_names.end()),

paddle/fluid/inference/analysis/ir_passes/anakin_subgraph_pass.h

@@ -15,6 +15,7 @@
 #pragma once
 #include <paddle/fluid/framework/ir/fuse_pass_base.h>
 #include <memory>
+#include <set>
 #include <string>
 #include <vector>
 #include "paddle/fluid/framework/ir/pass.h"
@@ -36,6 +37,13 @@ class AnakinSubgraphPass : public framework::ir::FusePassBase {
                       const std::vector<std::string> &graph_params,
                       std::vector<std::string> *repetitive_params) const;
   void CleanIntermediateOutputs(framework::ir::Node *node);
+  template <::anakin::Precision PrecisionT>
+  void CreateAnakinEngine(framework::BlockDesc *block_desc,
+                          const std::vector<std::string> &params,
+                          const std::set<std::string> &input_names,
+                          const std::vector<std::string> &output_mapping,
+                          const std::vector<std::string> &program_inputs,
+                          const std::string &engine_key) const;
 };
 
 }  // namespace analysis

paddle/fluid/inference/api/analysis_config.cc

@@ -116,6 +116,9 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
   CP_MEMBER(anakin_max_batchsize_);
   CP_MEMBER(anakin_max_input_shape_);
   CP_MEMBER(anakin_min_subgraph_size_);
+  CP_MEMBER(anakin_precision_mode_);
+  CP_MEMBER(anakin_passes_filter_);
+  CP_MEMBER(anakin_ops_filter_);
 
   // Ir related.
   CP_MEMBER(enable_ir_optim_);
@@ -276,9 +279,12 @@ void AnalysisConfig::Update() {
 
     pass_builder()->ClearPasses();
     for (const auto &pass : kAnakinSubgraphPasses) {
+      if (std::find(anakin_passes_filter_.begin(), anakin_passes_filter_.end(),
+                    pass) == anakin_passes_filter_.end()) {
         pass_builder()->AppendPass(pass);
       }
     }
+  }
 
   if (ir_debug_) {
     pass_builder()->TurnOnDebug();
@@ -391,11 +397,16 @@ void AnalysisConfig::SwitchIrDebug(int x) {
 }
 void AnalysisConfig::EnableAnakinEngine(
     int max_batch_size, std::map<std::string, std::vector<int>> max_input_shape,
-    int min_subgraph_size) {
+    int min_subgraph_size, AnalysisConfig::Precision precision_mode,
+    std::vector<std::string> passes_filter,
+    std::vector<std::string> ops_filter) {
   anakin_max_batchsize_ = max_batch_size;
   anakin_max_input_shape_ = max_input_shape;
   anakin_min_subgraph_size_ = min_subgraph_size;
+  anakin_passes_filter_ = passes_filter;
+  anakin_ops_filter_ = ops_filter;
   use_anakin_ = true;
+  anakin_precision_mode_ = precision_mode;
   Update();
 }
 }  // namespace paddle

paddle/fluid/inference/api/analysis_predictor.cc

@@ -386,6 +386,9 @@ void AnalysisPredictor::PrepareArgument() {
     argument_.SetAnakinMaxBatchSize(config_.anakin_max_batchsize_);
     argument_.SetAnakinMaxInputShape(config_.anakin_max_input_shape_);
     argument_.SetAnakinMinSubgraphSize(config_.anakin_min_subgraph_size_);
+    argument_.SetAnakinPrecisionMode(config_.anakin_precision_mode_);
+    argument_.SetAnakinPassesFilter(config_.anakin_passes_filter_);
+    argument_.SetAnakinOpsFilter(config_.anakin_ops_filter_);
     LOG(INFO) << "Anakin subgraph engine is enabled";
   }
 

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -152,7 +152,9 @@ struct AnalysisConfig {
   void EnableAnakinEngine(
       int max_batch_size = 1,
       std::map<std::string, std::vector<int>> max_input_shape = {},
-      int min_subgraph_size = 6);
+      int min_subgraph_size = 6, Precision precision = Precision::kFloat32,
+      std::vector<std::string> passes_filter = {},
+      std::vector<std::string> ops_filter = {});
 
   /** A boolean state indicating whether the Anakin sub-graph engine is used.
   */
@@ -291,6 +293,9 @@ struct AnalysisConfig {
   int anakin_max_batchsize_;
   int anakin_min_subgraph_size_{6};
   std::map<std::string, std::vector<int>> anakin_max_input_shape_;
+  Precision anakin_precision_mode_;
+  std::vector<std::string> anakin_passes_filter_;
+  std::vector<std::string> anakin_ops_filter_;
   std::map<std::string, std::string> engine_opt_info_;
 
   bool use_mkldnn_quantizer_{false};

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -73,15 +73,21 @@ void PaddlePassBuilder::ClearPasses() { passes_.clear(); }
 // The following passes works for Anakin sub-graph engine.
 const std::vector<std::string> kAnakinSubgraphPasses({
     "infer_clean_graph_pass",                       //
+    "graph_viz_pass",                               //
+    "quant_conv2d_dequant_fuse_pass",               //
+    "graph_viz_pass",                               //
     "simplify_anakin_priorbox_detection_out_pass",  //
     "fillconstant_elementwisemul_fuse",             //
     "fc_fuse_pass",                                 //
     "conv_elementwise_add_fuse_pass",               //
-    "conv_bn_fuse_pass",                            //
-    "conv_elementwise_add_fuse_pass",               //
+    // "conv_bn_fuse_pass",                            //
+    // "conv_elementwise_add_fuse_pass",               //
     "fc_gru_fuse_pass",      //
-    "quant_conv2d_dequant_fuse_pass",               //
-    "anakin_subgraph_pass",
+    "graph_viz_pass",        //
+    "anakin_subgraph_pass",  //
+    "graph_viz_pass",        //
+    "fc_gru_fuse_pass",      //
+    "graph_viz_pass",        //
 });
 
 GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {

paddle/fluid/operators/anakin/anakin_engine_op.h

@@ -44,6 +44,7 @@ class AnakinEngineOp : public framework::OperatorBase {
   std::string engine_key_;
   std::string engine_serialized_data_;
   bool use_gpu_;
+  bool enable_int8_;
 
  public:
   AnakinEngineOp(const std::string &type,
@@ -55,6 +56,7 @@ class AnakinEngineOp : public framework::OperatorBase {
     engine_key_ = Attr<std::string>("engine_key");
     auto params = Attr<std::vector<std::string>>("parameters");
     use_gpu_ = Attr<bool>("use_gpu");
+    enable_int8_ = Attr<bool>("enable_int8");
     for (const auto &param : params) {
       param_names_.insert(param);
     }
@@ -68,11 +70,6 @@ class AnakinEngineOp : public framework::OperatorBase {
 
   void RunAnakin(const framework::Scope &scope,
                  const platform::Place &dev_place) const {
-    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
-    auto &dev_ctx = *pool.Get(dev_place);
-    auto stream =
-        reinterpret_cast<const platform::CUDADeviceContext &>(dev_ctx).stream();
-
     PADDLE_ENFORCE(!input_names_.empty(), "should pass more than one inputs");
 
     std::vector<std::string> output_maps =
@@ -96,18 +93,35 @@ class AnakinEngineOp : public framework::OperatorBase {
       outputs.insert({output_maps[output_index], fluid_t});
       output_index += 1;
     }
+    if (enable_int8_) {
+      Execute<::anakin::Precision::INT8>(inputs, outputs, dev_place);
+    } else {
+      Execute<::anakin::Precision::FP32>(inputs, outputs, dev_place);
+    }
+  }
+
+  template <::anakin::Precision PrecisionT>
+  void Execute(const std::map<std::string, framework::LoDTensor *> &inputs,
+               const std::map<std::string, framework::LoDTensor *> &outputs,
+               const platform::Place &dev_place) const {
     if (use_gpu_) {
 #ifdef PADDLE_WITH_CUDA
+      platform::DeviceContextPool &pool =
+          platform::DeviceContextPool::Instance();
+      auto &dev_ctx = *pool.Get(dev_place);
+      auto stream =
+          reinterpret_cast<const platform::CUDADeviceContext &>(dev_ctx)
+              .stream();
       auto *engine =
           inference::Singleton<inference::anakin::AnakinEngineManager<
-              ::anakin::saber::NV>>::Global()
+              ::anakin::saber::NV, PrecisionT>>::Global()
               .Get(engine_key_);
       engine->Execute(inputs, outputs, stream);
 #endif
     } else {
       auto *engine =
           inference::Singleton<inference::anakin::AnakinEngineManager<
-              ::anakin::saber::X86>>::Global()
+              ::anakin::saber::X86, PrecisionT>>::Global()
               .Get(engine_key_);
       engine->Execute(inputs, outputs);
     }

paddle/fluid/pybind/inference_api.cc

@@ -16,6 +16,7 @@
 #include <pybind11/stl.h>
 #include <cstring>
 #include <iostream>
+#include <map>
 #include <string>
 #include <vector>
 #include "paddle/fluid/inference/api/analysis_predictor.h"
@@ -230,8 +231,13 @@ void BindAnalysisConfig(py::module *m) {
            py::arg("precision_mode") = AnalysisConfig::Precision::kFloat32,
            py::arg("use_static") = true)
       .def("enable_anakin_engine", &AnalysisConfig::EnableAnakinEngine,
-           py::arg("max_batch_size") = 1, py::arg("max_input_shape") = {},
-           py::arg("min_subgraph_size") = 6)
+           py::arg("max_batch_size") = 1,
+           py::arg("max_input_shape") =
+               std::map<std::string, std::vector<int>>(),
+           py::arg("min_subgraph_size") = 6,
+           py::arg("precision_mode") = AnalysisConfig::Precision::kFloat32,
+           py::arg("passes_filter") = std::vector<std::string>(),
+           py::arg("ops_filter") = std::vector<std::string>())
       .def("tensorrt_engine_enabled", &AnalysisConfig::tensorrt_engine_enabled)
       .def("switch_ir_debug", &AnalysisConfig::SwitchIrDebug,
            py::arg("x") = true)