diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc
index 64be8a4b16314d58fb84b734088b03c61d1e79d0..6a6ef818927b3646fdb4fa6b345d5ed420376d94 100644
--- a/paddle/fluid/operators/conv_mkldnn_op.cc
+++ b/paddle/fluid/operators/conv_mkldnn_op.cc
@@ -350,27 +350,59 @@ std::cout<<"fuse_relu = "<<fuse_relu<<"  fuse_residual_conn = "<<fuse_residual_c
 
     std::vector<float> output_shift_scale;
     float sum_scale = 1.0f;
+
+for(int i=0; i<50; i++){
+    printf("%f ", (float)*(input_data+i));
+}
+printf("\n");fflush(stdout);
+
     if(is_INT8){
 std::cout<<"this is conv int8 op .............."<<std::endl;
+
+unsigned char* a = (unsigned char*)(input_data);
+for(int i=0; i<50; i++){
+    printf("%d ", a[i]);
+}
+printf("\n");
+
         int count = is_multi_channel? (g>1? weights_tz[1]*weights_tz[0] : weights_tz[0]) : 1; 
         float scale_in_data = *(scale_in->data<float>());
+
+for(int i=0; i<50; i++){
+    printf("%f ", (float)(input_data[i])/scale_in_data);
+}
+printf("\n");fflush(stdout);
+
+std::cout<<"scale_in_data = "<<scale_in_data<<std::endl;
+
         std::vector<float> scale_weights_data(count);
+std::cout<<"scale_weights_data = ";
         for(int i=0; i<count; i++){
             scale_weights_data[i] =*(scale_weights->data<float>() + i);
+std::cout<<scale_weights_data[i]<<" ";
         }
+std::cout<<std::endl;
+
         float scale_out_data = *(scale_out->data<float>());
 
+std::cout<<"scale_out_data = "<<scale_out_data<<std::endl;
+
         output_shift_scale.resize(count);
+std::cout<<"output_shift_scale = ";
         for(int i=0; i<count; i++){
             if(scale_weights_data[i] == 0.0)
                 output_shift_scale[i] = scale_out_data;
             else 
                 output_shift_scale[i] = scale_out_data / (scale_in_data * scale_weights_data[i]);
+std::cout<<output_shift_scale[i]<<" ";
         }
+std::cout<<std::endl;
         if(fuse_residual_conn){
             float scale_in_eltwise_data = *(scale_in_eltwise->data<float>());
+std::cout<<"scale_in_eltwise_data = "<<scale_in_eltwise_data<<std::endl;
             sum_scale = scale_out_data / scale_in_eltwise_data;
         }
+std::cout<<"  sum_scale = "<<sum_scale<<std::endl;
     }
 
     // Get unique name for storing MKLDNN primitives
@@ -379,7 +411,7 @@ std::cout<<"this is conv int8 op .............."<<std::endl;
         ctx.op().Output("Output"));
     const std::string key_conv_pd = key + "@conv_pd";
 
-std::cout<<key_conv_pd<<std::endl;
+std::cout<<"current op is = "<<key_conv_pd<<std::endl;
 
     std::vector<primitive> pipeline;
     auto user_src_md = platform::MKLDNNMemDesc(
@@ -412,7 +444,7 @@ std::cout<<key_conv_pd<<std::endl;
             weights_tz, memory::data_type::s8,
             (g == 1) ? chosen_memory_format : mkldnn::memory::format::goihw);
         dst_md = platform::MKLDNNMemDesc(
-            dst_tz,
+            dst_tz,// memory::data_type::f32, chosen_memory_format);
             fuse_relu? paddle::framework::ToMKLDNNDataType(std::type_index(typeid(unsigned char))) : 
             paddle::framework::ToMKLDNNDataType(std::type_index(typeid(char))),
             chosen_memory_format);
@@ -467,7 +499,7 @@ std::cout<<key_conv_pd<<std::endl;
     std::shared_ptr<mkldnn::memory> weights_memory_p;// = handler.AcquireWeightsMemoryFromPrimitive(
         //user_weights_memory_p, pipeline, is_test);
     if(is_INT8){
-        int mask_reorder = is_multi_channel? 0 : ((g!= 1) ? (1<<1)+(1<<0) : 1<<0);
+        int mask_reorder = is_multi_channel? ((g!= 1) ? (1<<1)+(1<<0) : 1<<0) : 0;
         int count = is_multi_channel? (g>1? weights_tz[1]*weights_tz[0] : weights_tz[0]) : 1;
         std::vector<float> scale_weights_data(count);
         for(int i=0; i<count; i++){
@@ -479,7 +511,6 @@ std::cout<<key_conv_pd<<std::endl;
         weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive(
             user_weights_memory_p, pipeline, is_test);
     }
- 
 
     std::shared_ptr<mkldnn::memory> dst_memory_p;
     if(is_INT8){
@@ -510,7 +541,7 @@ std::cout<<key_conv_pd<<std::endl;
                   handler.AcquireDstMemoryFromPrimitive(to_void_cast<int8_t>(output_data));
           }
         }
-std::cout<<"input fmt = "<<input->format()<<"  output fmt = "<<output->format()<<"   dst fmt = "<<dst_memory_p->get_primitive_desc().desc().data.format<<std::endl;
+std::cout<<"input fmt = "<<input->format()<<"   input dt = "<<paddle::framework::ToMKLDNNDataType(input->type())<<"  output fmt = "<<output->format()<<"   output dt = "<<paddle::framework::ToMKLDNNDataType(output->type())<<"   dst fmt = "<<dst_memory_p->get_primitive_desc().desc().data.format<<"  dst dt = "<<dst_memory_p->get_primitive_desc().desc().data.data_type<<std::endl;
     } else{
         T* output_data = nullptr;
         if (fuse_residual_conn) {
@@ -545,7 +576,7 @@ std::cout<<"input fmt = "<<input->format()<<"  output fmt = "<<output->format()<
       std::shared_ptr<mkldnn::memory>  bias_memory_p;// =
           //handler.AcquireBiasMemoryFromPrimitive(user_bias_memory_p, pipeline);
       if(is_INT8){
-          int mask_reorder = is_multi_channel? 0 : 1<<0;
+          int mask_reorder = is_multi_channel? 1<<0 : 1;
           int count = is_multi_channel? (g>1? weights_tz[1]*weights_tz[0] : weights_tz[0]) : 1;
           std::vector<float> scale_bias_data(count);
           for(int i=0; i<count; i++){
@@ -594,7 +625,7 @@ std::cout<<"input fmt = "<<input->format()<<"  output fmt = "<<output->format()<
       if (fuse_relu) {
         constexpr float scale = 1.0f;
         constexpr float negative_slope = 0.0f;
-        constexpr float placeholder = 0.0f; //beta
+        constexpr float placeholder = 1.0f; //beta
         post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
                                        negative_slope, placeholder);
       }
diff --git a/paddle/fluid/operators/dequantize_op.cc b/paddle/fluid/operators/dequantize_op.cc
index 5fe6010dba8a10846f9e5b2f6a8ba239b77839c1..d7504972bce87753e1916769428a3305b5fbec67 100644
--- a/paddle/fluid/operators/dequantize_op.cc
+++ b/paddle/fluid/operators/dequantize_op.cc
@@ -72,7 +72,7 @@ std::cout<<"this is dequant op ***********"<<std::endl;
     auto dst_memory = mkldnn::memory(dst_pd, to_void_cast<float>(output_data));
     
     auto reorder_pd = std::shared_ptr<reorder::primitive_desc>(
-        new reorder::primitive_desc(dst_pd, src_pd, attri));    
+        new reorder::primitive_desc(src_pd, dst_pd, attri));    
     auto reorder_p= std::shared_ptr<reorder>(new reorder(*reorder_pd, *src_memory_p, dst_memory));
     pipeline.push_back(*reorder_p);
 
diff --git a/paddle/fluid/operators/quantize_op.cc b/paddle/fluid/operators/quantize_op.cc
index cb5a9e4c1c8d6f94e8dce9eb753fb3763625baa4..1748e645d76a79a1305dcd0c13b6dd8eb91eb0ec 100644
--- a/paddle/fluid/operators/quantize_op.cc
+++ b/paddle/fluid/operators/quantize_op.cc
@@ -47,9 +47,35 @@ std::cout<<"this is quantize op!!!!!!!!!!!!!!"<<std::endl;
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
 
     const T* input_data = input->data<T>();
-    T* output_data = output->mutable_data<T>(ctx.GetPlace());
+    uint8_t* output_data = output->mutable_data<uint8_t>(ctx.GetPlace());
     std::vector<T> scale_data = {*(scale->data<T>())};
 
+FILE *fp = fopen("quant_input.txt","w");
+printf("quantize check!!!!\n");
+std::vector<int> tz = paddle::framework::vectorize2int(input->dims());
+int count=1;
+for(int i=0; i<tz.size(); i++){
+    count*=tz[i];
+    printf("%d ",tz[i]);
+}
+printf("\n");
+int num=0;
+for(int i=0; i<count; i++){
+    if(num==32){ fprintf(fp,"\n"); num=0;}
+    fprintf(fp,"%f ", *(input_data+i));
+    num ++;
+}
+fprintf(fp,"\n");
+fclose(fp);
+for(int i=0; i<scale_data.size(); i++){
+    printf("%f", scale_data[i]);
+}
+printf("\n");
+for(int i=0; i<50; i++){
+    printf("%f ", (*(input_data+i))*scale_data[0]);
+}
+printf("\n");
+fflush(stdout);
     mkldnn::primitive_attr attri;
     int mask = 0;
     attri.set_output_scales(mask, scale_data);
@@ -63,10 +89,10 @@ std::cout<<"this is quantize op!!!!!!!!!!!!!!"<<std::endl;
     auto dst_md = platform::MKLDNNMemDesc(
             {dst_tz}, memory::data_type::u8, memory::format::nhwc);
     auto dst_pd = mkldnn::memory::primitive_desc(dst_md, engine);
-    auto dst_memory = mkldnn::memory(dst_pd, to_void_cast<T>(output_data));
+    auto dst_memory = mkldnn::memory(dst_pd, to_void_cast<uint8_t>(output_data));
     
     auto reorder_pd = std::shared_ptr<reorder::primitive_desc>(
-        new reorder::primitive_desc(dst_pd, src_pd, attri));    
+        new reorder::primitive_desc(src_pd, dst_pd, attri));    
     auto reorder_p= std::shared_ptr<reorder>(new reorder(*reorder_pd, *src_memory_p, dst_memory));
 
     pipeline.push_back(*reorder_p);