fix quantize bug and add debug log

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -350,10 +350,29 @@ 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 ", *(input_data+i)/scale_in_data);
+}
+printf("\n");fflush(stdout);
+
         std::vector<float> scale_weights_data(count);
         for(int i=0; i<count; i++){
             scale_weights_data[i] =*(scale_weights->data<float>() + i);
@@ -479,7 +498,6 @@ 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 +528,7 @@ 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) {

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);
 

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);

paddle/fluid/platform/mkldnn_helper.h

@@ -212,16 +212,18 @@ class MKLDNNHandler {
       std::shared_ptr<mkldnn::primitive> reorder_p;
       if (mpd != user_mpd) {
         target_memory_p = std::make_shared<mkldnn::memory>(mpd);
-        auto reorder_p =
-            std::make_shared<mkldnn::reorder>(*user_memory_p, *target_memory_p);
+        std::shared_ptr<mkldnn::reorder> reorder_p;// =
+            //std::make_shared<mkldnn::reorder>(*user_memory_p, *target_memory_p);
         if(is_INT8){
             mkldnn::primitive_attr attri;
             attri.set_output_scales(mask, scale_data);
 
             auto reorder_pd = std::shared_ptr<mkldnn::reorder::primitive_desc>(
-                    new mkldnn::reorder::primitive_desc(mpd, user_mpd, attri));
-            auto reorder_p =
+                    new mkldnn::reorder::primitive_desc(user_mpd, mpd, attri));
+            reorder_p =
                 std::shared_ptr<mkldnn::reorder>(new mkldnn::reorder(*reorder_pd, *user_memory_p, *target_memory_p));
+        } else{
+            reorder_p = std::make_shared<mkldnn::reorder>(*user_memory_p, *target_memory_p);
         }
         dev_ctx_.SetBlob(key_reorder_p, reorder_p);
         pipeline.push_back(*reorder_p);