Merge branch 'prv-calibration'

paddle/fluid/operators/conv_op.cc

@@ -131,21 +131,14 @@ void Conv2DOpMaker::Make() {
            "The format of output tensor is X (one-dimensional) of size equal"
            "to the number of output channels. Only used with MKL-DNN.")
       .AsDispensable();
-  AddInput("Scale_in",
-           "(Tensor) Scale_in to be used for int8 input data."
-           "Only used with INT8.")
-      .AsDispensable();
-  AddInput("Scale_in_eltwise",
-           "(Tensor) Scale_in_eltwise to be used for int8 eltwise input data."
-           "Only used with MKL-DNN.")
-      .AsDispensable();
-  AddInput("Scale_weights",
-           "(Tensor) Scale_weights to be used for int8 weights data."
-           "Only used with MKL-DNN.")
-      .AsDispensable();
-  AddInput("Scale_out",
-           "(Tensor) Scale_out to be used for int8 output data."
-           "Only used with MKL-DNN.")
+  AddOutput("Output",
+            "(Tensor) The output tensor of convolution operator. "
+            "The format of output tensor is also NCHW.");
+
+  AddInput("ResidualData",
+           "(Tensor) Tensor with residual data "
+           "to which convolution output will be added."
+           "Used with fuse_residual_connection fusion.")
       .AsDispensable();
   AddOutput("Output",
             "(Tensor) The output tensor of convolution operator. "
@@ -193,6 +186,22 @@ void Conv2DOpMaker::Make() {
                 "whenever convolution output is as an input to residual "
                 "connection.")
       .SetDefault(false);
+  AddAttr<float>("Scale_in",
+           "Scale_in to be used for int8 input data."
+           "Only used with INT8.")
+      .SetDefault(1.0f);
+  AddAttr<float>("Scale_out",
+           "Scale_out to be used for int8 output data."
+           "Only used with MKL-DNN.")
+      .SetDefault(1.0f);
+  AddAttr<float>("Scale_in_eltwise",
+           "Scale_in_eltwise to be used for int8 eltwise input data."
+           "Only used with MKL-DNN.")
+      .SetDefault(1.0f);
+  AddAttr<std::vector<float>>("Scale_weights",
+           "Scale_weights to be used for int8 weights data."
+           "Only used with MKL-DNN.")
+      .SetDefault({1.0f});
   AddAttr<bool>("force_fp32_output", "(bool, default false) Force INT8 kernel output FP32, only used in mkldnn kernel")
       .SetDefault(false);
   AddAttr<std::string>(

paddle/fluid/operators/dequantize_op.cc

@@ -37,7 +37,7 @@ class DeQuantOpKernel : public framework::OpKernel<T> {
 
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* input = ctx.Input<Tensor>("Input");
-    auto* scale = ctx.Input<Tensor>("Scale");
+    auto scale_data = ctx.Attr<float>("Scale");
     auto* output = ctx.Output<Tensor>("Output");
     auto& dev_ctx =
         ctx.template device_context<platform::MKLDNNDeviceContext>();
@@ -45,8 +45,7 @@ class DeQuantOpKernel : public framework::OpKernel<T> {
  
     const T* input_data = input->data<T>();
     float* output_data = output->mutable_data<float>(ctx.GetPlace());
-    std::vector<float> scale_data = {*(scale->data<float>())};
-    std::vector<float> reorder_scale = {1.0f / scale_data[0]};
+    std::vector<float> reorder_scale = {1.0f / scale_data};
 
     std::vector<primitive> pipeline;
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
@@ -99,8 +98,8 @@ framework::OpKernelType DeQuantOp::GetExpectedKernelType(const framework::Execut
 
 void DeQuantOpMaker::Make() {
   AddInput("Input","input data");
-  AddInput("Scale","scale data");
   AddOutput("Output","output data");
+  AddAttr<float>("Scale","scale data").SetDefault({1.0f});
   AddComment(R"DOC(This op will quantize data from INT8 to FP32)DOC");
 }
 

paddle/fluid/operators/quantize_op.cc

@@ -35,7 +35,7 @@ class QuantOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* input = ctx.Input<Tensor>("Input");
-    auto* scale = ctx.Input<Tensor>("Scale");
+    auto scale_data = ctx.Attr<float>("Scale");
     auto* output = ctx.Output<Tensor>("Output");
     auto& dev_ctx =
         ctx.template device_context<platform::MKLDNNDeviceContext>();
@@ -47,11 +47,9 @@ class QuantOpKernel : public framework::OpKernel<T> {
 
     const T* input_data = input->data<T>();
 
-    std::vector<T> scale_data = {*(scale->data<T>())};
-
     mkldnn::primitive_attr attri;
     int mask = 0;
-    attri.set_output_scales(mask, scale_data);
+    attri.set_output_scales(mask, {scale_data});
 
     auto src_md = platform::MKLDNNMemDesc(
             {src_tz}, memory::data_type::f32, input->format());
@@ -108,11 +106,12 @@ framework::OpKernelType QuantOp::GetExpectedKernelType(const framework::Executio
 
 void QuantOpMaker::Make() {
   AddInput("Input","input data");
-  AddInput("Scale","scale data");
   AddOutput("Output","output data");
   AddAttr<bool>("is_negative_input",
                 "(bool, default false) Only used in mkldnn INT8 kernel")
       .SetDefault(false);
+  AddAttr<float>("Scale","scale data")
+      .SetDefault({1.0f});
   AddComment(R"DOC(This op will quantize data from FP32 to INT8)DOC");
 }