clean code for int8

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -275,7 +275,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
   void Compute(const paddle::framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                    "It must use CPUPlace.");
-std::cout<<"this is conv kernel op....................."<<std::endl;
     const bool is_test = ctx.Attr<bool>("is_test");
 
     auto& dev_ctx =
@@ -320,8 +319,6 @@ std::cout<<"this is conv kernel op....................."<<std::endl;
     bool fuse_residual_conn = ctx.Attr<bool>("fuse_residual_connection");
     int groups = ctx.Attr<int>("groups");
 
-std::cout<<"fuse_relu = "<<fuse_relu<<"  fuse_residual_conn = "<<fuse_residual_conn<<std::endl;
-
     // TODO(tpatejko): add support for dilation
     PADDLE_ENFORCE(
         dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1,
@@ -351,36 +348,15 @@ 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;
-        //const uint8_t* input_data_int8 = input->data<uint8_t>(); //FIX ME XIAOLI
-//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::cout<<"scale_in = "<<scale_in_data<<std::endl;
-
         std::vector<float> scale_weights_data(count);
         for(int i=0; i<count; i++){
             scale_weights_data[i] =*(scale_weights->data<float>() + i);
         }
         float scale_out_data = *(scale_out->data<float>());
-std::cout<<"scale_out = "<<scale_out_data<<std::endl;
         output_shift_scale.resize(count);
         for(int i=0; i<count; i++){
             if(scale_weights_data[i] == 0.0)
@@ -391,7 +367,6 @@ std::cout<<"scale_out = "<<scale_out_data<<std::endl;
         if(fuse_residual_conn){
             float scale_in_eltwise_data = *(scale_in_eltwise->data<float>());
             sum_scale = scale_out_data / scale_in_eltwise_data;
-std::cout<<"scale_in_eltwise_data = "<<scale_in_eltwise_data<<" scale_out_data = "<<scale_out_data<<" sum_scale = "<<sum_scale<<std::endl;
         }
     }
 
@@ -401,8 +376,6 @@ std::cout<<"scale_in_eltwise_data = "<<scale_in_eltwise_data<<" scale_out_data =
         ctx.op().Output("Output"));
     const std::string key_conv_pd = key + "@conv_pd";
 
-std::cout<<"current op is = "<<key_conv_pd<<std::endl;
-
     std::vector<primitive> pipeline;
     auto user_src_md = platform::MKLDNNMemDesc(
             {src_tz}, paddle::framework::ToMKLDNNDataType(input->type()), input->format());
@@ -440,13 +413,7 @@ std::cout<<"current op is = "<<key_conv_pd<<std::endl;
             if(dst_dt != residual_dt) 
                 dst_dt = residual_dt;
         }
-        dst_md = platform::MKLDNNMemDesc(
-            dst_tz,// memory::data_type::f32, chosen_memory_format);
-            dst_dt,//paddle::framework::ToMKLDNNDataType(std::type_index(typeid(unsigned char))),
-            chosen_memory_format);
-            //fuse_relu? paddle::framework::ToMKLDNNDataType(std::type_index(typeid(unsigned char))) : 
-            //paddle::framework::ToMKLDNNDataType(std::type_index(typeid(signed char))),
-            //chosen_memory_format);
+        dst_md = platform::MKLDNNMemDesc(dst_tz, dst_dt, chosen_memory_format);
     }
 
     // create a conv primitive descriptor and save it for usage in backward
@@ -495,8 +462,7 @@ std::cout<<"current op is = "<<key_conv_pd<<std::endl;
     // create reorder primitive if the input format is not the preferred one
     auto src_memory_p =
         handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline);
-    std::shared_ptr<mkldnn::memory> weights_memory_p;// = handler.AcquireWeightsMemoryFromPrimitive(
-        //user_weights_memory_p, pipeline, is_test);
+    std::shared_ptr<mkldnn::memory> weights_memory_p;
     if(is_INT8){
         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;
@@ -515,48 +481,18 @@ std::cout<<"current op is = "<<key_conv_pd<<std::endl;
     if(is_INT8){
         if (fuse_residual_conn) {
           auto residual_param = ctx.Input<Tensor>("ResidualData");
-          //auto residual_param_data = residual_param->data<T>();
           PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(),
                             "Output and elementwise parameter need to have the "
                             "same dimension sizes");
-//std::cout<<"output = "<<output<<" residual_param = "<<residual_param<<std::endl;
           output->ShareDataWith(*residual_param);
           auto residual_dt = paddle::framework::ToMKLDNNDataType(residual_param->type());
           if(residual_dt == mkldnn::memory::data_type::u8){
 
               uint8_t* output_data = output->mutable_data<uint8_t>(ctx.GetPlace());
-//std::cout<<"after share output = "<<output<<" residual_param = "<<residual_param<<std::endl;
-//float scale_in_eltwise_data = *(scale_in_eltwise->data<float>());
-printf("residual is u8: this is bottom 1 data\n");
-//unsigned char* f = (unsigned char*)(residual_param_data);
-//for(int i=0; i<50; i++){
-//    printf("%f ", (float)f[i]/scale_in_eltwise_data);
-//}
-//printf("\n");
-//printf("this is output data\n");
-//unsigned char* e = (unsigned char*)(output_data);
-//for(int i=0; i<50; i++){
-//    printf("%f ", (float)e[i]/scale_in_eltwise_data);
-//}
-//printf("\n");
               dst_memory_p =
                   handler.AcquireDstMemoryFromPrimitive(to_void_cast<uint8_t>(output_data));
           } else{
               int8_t* output_data = output->mutable_data<int8_t>(ctx.GetPlace());
-//std::cout<<"after share output = "<<output<<" residual_param = "<<residual_param<<std::endl;
-
-printf("residual is s8 : this is bottom 1 data\n");
-//char* f = (char*)(residual_param_data);
-//for(int i=0; i<50; i++){
-//    printf("%f ", (float)f[i]);
-//}
-//printf("\n");
-//printf("this is output data\n");
-//char* e = (char*)(output_data);
-//for(int i=0; i<50; i++){
-//    printf("%f ", (float)e[i]);
-//}
-//printf("\n");
               dst_memory_p =
                   handler.AcquireDstMemoryFromPrimitive(to_void_cast<int8_t>(output_data));
           }
@@ -571,7 +507,6 @@ printf("residual is s8 : this is bottom 1 data\n");
                   handler.AcquireDstMemoryFromPrimitive(to_void_cast<int8_t>(output_data));
           }
         }
-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) {
@@ -603,8 +538,7 @@ std::cout<<"input fmt = "<<input->format()<<"   input dt = "<<paddle::framework:
           {bias_tz}, platform::MKLDNNGetDataType<float>(), memory::format::x);
       auto user_bias_memory_p =
           handler.AcquireBiasMemory(user_bias_md, to_void_cast<float>(bias_data));
-      std::shared_ptr<mkldnn::memory>  bias_memory_p;// =
-          //handler.AcquireBiasMemoryFromPrimitive(user_bias_memory_p, pipeline);
+      std::shared_ptr<mkldnn::memory>  bias_memory_p;
       if(is_INT8){
           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;
@@ -632,12 +566,6 @@ std::cout<<"input fmt = "<<input->format()<<"   input dt = "<<paddle::framework:
 
     output->set_layout(DataLayout::kMKLDNN);
     output->set_format(GetMKLDNNFormat(*dst_memory_p));
-    //if(is_INT8){
-    //    uint8_t* output_data = output->mutable_data<uint8_t>(ctx.GetPlace()); //work aroud forsum fusion 
-    //    std::cout<<"output_data = "<<output_data<<std::endl;
-    //}
-std::cout<<"input fmt = "<<input->format()<<"  output fmt = "<<output->format()<<"   dst fmt = "<<dst_memory_p->get_primitive_desc().desc().data.format<<"output dt = "<<paddle::framework::ToMKLDNNDataType(output->type())<<"dst dt = "<<dst_memory_p->get_primitive_desc().desc().data.data_type<<std::endl;
-    std::cout<<"this is conv end!!!!!!!!!!!!!!!!!!!!"<<std::endl;
   }
 
  private:

paddle/fluid/operators/dequantize_op.cc

@@ -40,7 +40,6 @@ class DeQuantOpKernel : public framework::OpKernel<T> {
     auto* input = ctx.Input<Tensor>("Input");
     auto* scale = ctx.Input<Tensor>("Scale");
     auto* output = ctx.Output<Tensor>("Output");
-std::cout<<"this is dequant op ***********"<<std::endl;
     auto& dev_ctx =
         ctx.template device_context<platform::MKLDNNDeviceContext>();
     const auto& engine = dev_ctx.GetEngine();
@@ -51,16 +50,6 @@ std::cout<<"this is dequant op ***********"<<std::endl;
     std::vector<float> scale_data = {*(scale->data<float>())};
     std::vector<float> reorder_scale = {1.0f / scale_data[0]};
 
-for(int i=0; i<50; i++){
-    printf("%d ", *(input_data+i));
-}
-printf("\n");fflush(stdout);
-for(int i=0; i<50; i++){
-    printf("%f ", *(input_data+i)/scale_data[0]);
-}
-printf("\n");fflush(stdout);
-std::cout<<"scale = "<<scale_data[0]<<std::endl;
-
     std::vector<primitive> pipeline;
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());

paddle/fluid/operators/pool_mkldnn_op.cc

@@ -72,7 +72,6 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
   void Compute(const paddle::framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                    "It must use CPUPlace.");
-std::cout<<"this is pool op"<<std::endl;
     auto& dev_ctx =
         ctx.template device_context<platform::MKLDNNDeviceContext>();
     const auto& mkldnn_engine = dev_ctx.GetEngine();
@@ -105,16 +104,6 @@ std::cout<<"this is pool op"<<std::endl;
     const T* input_data = input->data<T>();
     T* output_data = output->mutable_data<T>(ctx.GetPlace());
 
-for(int i=0; i<50; i++){
-    printf("%f ",(float) *(input_data+i));
-}
-printf("\n");fflush(stdout);
-for(int i=0; i<50; i++){
-    printf("%f ", *(input_data+i)/14.4791);
-}
-printf("\n");fflush(stdout);
-
-
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
 
@@ -143,8 +132,6 @@ printf("\n");fflush(stdout);
 
       mkldnn::memory::data_type dt = paddle::framework::ToMKLDNNDataType(input->type());
 
-std::cout<<"input type = "<<dt<<std::endl;
-
       auto src_md = platform::MKLDNNMemDesc(
           src_tz, dt, input_format);
 
@@ -203,25 +190,12 @@ std::cout<<"input type = "<<dt<<std::endl;
                           .data.format;
     }
 
-printf("befor submit!!!!!!!!!!!\n");
-for(int i=0; i<50; i++){
-    printf("%f ", *(output_data+i)/14.4791);
-}
-printf("\n");fflush(stdout);
-
     // push primitive to stream and wait until it's executed
     std::vector<mkldnn::primitive> pipeline{*(pool_p.get())};
     stream(stream::kind::eager).submit(pipeline).wait();
 
-printf("after submit!!!!!!!!!!!\n");
-for(int i=0; i<50; i++){
-    printf("%f ", *(output_data+i)/14.4791);
-}
-printf("\n");fflush(stdout);
-
     output->set_layout(DataLayout::kMKLDNN);
     output->set_format(output_format);
-std::cout<<"input fmt = "<<input->format()<<"  output fmt = "<<output->format()<<"output dt = "<<paddle::framework::ToMKLDNNDataType(output->type())<<std::endl;
   }
 
  private:

paddle/fluid/operators/quantize_op.cc

@@ -37,7 +37,6 @@ class QuantOpKernel : public framework::OpKernel<T> {
     auto* input = ctx.Input<Tensor>("Input");
     auto* scale = ctx.Input<Tensor>("Scale");
     auto* output = ctx.Output<Tensor>("Output");
-std::cout<<"this is quantize op!!!!!!!!!!!!!!"<<std::endl;
     auto& dev_ctx =
         ctx.template device_context<platform::MKLDNNDeviceContext>();
     const auto& engine = dev_ctx.GetEngine();
@@ -50,32 +49,6 @@ std::cout<<"this is quantize op!!!!!!!!!!!!!!"<<std::endl;
     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);

paddle/fluid/operators/requantize_op.cc

@@ -38,7 +38,6 @@ class ReQuantOpKernel : public framework::OpKernel<T> {
     auto* input = ctx.Input<Tensor>("Input");
     //auto* scale = ctx.Input<Tensor>("Scale");
     auto* output = ctx.Output<Tensor>("Output");
-std::cout<<"this is requantize op!!!!!!!!!!"<<std::endl;
     auto& dev_ctx =
         ctx.template device_context<platform::MKLDNNDeviceContext>();
     const auto& engine = dev_ctx.GetEngine();
@@ -75,49 +74,12 @@ std::cout<<"this is requantize op!!!!!!!!!!"<<std::endl;
     auto reorder_pd = std::shared_ptr<reorder::primitive_desc>(
         new reorder::primitive_desc(src_pd, dst_pd, attri));   
 
-for(int i=0; i<50; i++){
-    printf("%d ", *(input_data+i));
-}
-printf("\n");fflush(stdout);
-//for(int i=0; i<50; i++){
-//    printf("%f ", *(input_data+i)/107.426);
-//}
-//printf("\n");fflush(stdout);
-std::cout<<"scale = "<<scale_data[0]<<std::endl;
-//for(int i=0; i<50; i++){
-//    printf("%f ", *(output_data+i)/107.426);
-//}
-//printf("\n");fflush(stdout);
-
-//    int is_sum = false;//ctx.Attr<int>("is_sum");
-//    if(is_sum){
-//std::cout<<"input fmt = "<<input->format()<<"  output fmt = "<<output->format()<<"output dt = "<<paddle::framework::ToMKLDNNDataType(output->type())<<std::endl;
-//        output_data = (uint8_t*)input_data;
-//std::cout<<"input fmt = "<<input->format()<<"  output fmt = "<<output->format()<<"output dt = "<<paddle::framework::ToMKLDNNDataType(output->type())<<std::endl;
-//
-//printf("after*************\n");
-//for(int i=0; i<50; i++){
-//    printf("%f ", *(output_data+i)/107.426);
-//}
-//printf("\n");fflush(stdout);
-//
-//    } else{
-        auto reorder_p= std::shared_ptr<reorder>(new reorder(*reorder_pd, *src_memory_p, dst_memory));
-        pipeline.push_back(*reorder_p);
-        stream(stream::kind::eager).submit(pipeline).wait();
-//    }
-//uint8_t* output_data_2 = output->mutable_data<uint8_t>(ctx.GetPlace());
-//for(int i=0; i<50; i++){
-//    printf("%f ", *(output_data_2+i)/107.426);
-//}
-//printf("\n");fflush(stdout);
-for(int i=0; i<50; i++){
-    printf("%d ", *(output_data+i));
-}
-printf("\n");fflush(stdout);
+    auto reorder_p= std::shared_ptr<reorder>(new reorder(*reorder_pd, *src_memory_p, dst_memory));
+    pipeline.push_back(*reorder_p);
+    stream(stream::kind::eager).submit(pipeline).wait();
+
     output->set_layout(DataLayout::kMKLDNN);
     output->set_format(GetMKLDNNFormat(dst_memory));
-std::cout<<"input fmt = "<<input->format()<<"  output fmt = "<<output->format()<<"output dt = "<<paddle::framework::ToMKLDNNDataType(output->type())<<std::endl;
   }
 };
 

paddle/fluid/operators/softmax_mkldnn_op.cc

@@ -132,13 +132,6 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
 
     const T* input_data = flattened_input.data<T>();
     T* output_data = flattened_output.mutable_data<T>(ctx.GetPlace());
-printf("this is softmax!!!!!!!!!!!!!!\n");
-for(int i=0; i<50; i++){
-    printf("%f ", (float)*(input_data+i));
-}
-printf("\n");fflush(stdout);
-
-std::cout<<"input fmt = "<<input->format()<<"   input dt = "<<paddle::framework::ToMKLDNNDataType(input->type())<<"  output fmt = "<<output->format()<<"   output dt = "<<paddle::framework::ToMKLDNNDataType(output->type())<<std::endl;
 
     std::vector<int> src_tz = paddle::framework::vectorize2int(flattened_dims);
     std::vector<int> dst_tz = src_tz;
@@ -170,7 +163,7 @@ std::cout<<"input fmt = "<<input->format()<<"   input dt = "<<paddle::framework:
     std::vector<primitive> pipeline{
         *(static_cast<softmax_forward::primitive*>(softmax_p.get()))};
     stream(stream::kind::eager).submit(pipeline).wait();
-std::cout<<"input fmt = "<<input->format()<<"   input dt = "<<paddle::framework::ToMKLDNNDataType(input->type())<<"  output fmt = "<<output->format()<<"   output dt = "<<paddle::framework::ToMKLDNNDataType(output->type())<<std::endl;
+
     const bool is_test = ctx.Attr<bool>("is_test");
     if (!is_test) {
       T threshold = exp(-64);

paddle/fluid/platform/mkldnn_helper.h

@@ -70,7 +70,6 @@ inline mkldnn::memory::desc MKLDNNMemDesc(const std::vector<int>& dims,
                                           mkldnn::memory::data_type data_type,
                                           mkldnn::memory::format format) {
   mkldnn::memory::dims tz = dims;
-std::cout<<"this is MKLDNNMemDesc"<<"   data_type"<<data_type<<"   format"<<format<<std::endl;
   return mkldnn::memory::desc({tz}, data_type, format);
 }