From f35d8ea890df9c48eba855ef9474a90391fbb7e1 Mon Sep 17 00:00:00 2001
From: xiaolil1 <xiaoli.liu@intel.com>
Date: Thu, 25 Oct 2018 16:08:39 +0800
Subject: [PATCH] fix bugs for INT8 with work around and debug logs

---
 .../fluid/framework/data_layout_transform.h   |  2 +-
 paddle/fluid/framework/tensor_impl.h          |  4 +-
 paddle/fluid/operators/conv_mkldnn_op.cc      | 70 +++++++++++++++----
 paddle/fluid/operators/dequantize_op.cc       | 17 ++++-
 paddle/fluid/operators/pool_mkldnn_op.cc      | 23 ++++++
 paddle/fluid/operators/requantize_op.cc       | 63 ++++++++++++++---
 paddle/fluid/operators/softmax_mkldnn_op.cc   | 10 ++-
 7 files changed, 161 insertions(+), 28 deletions(-)

diff --git a/paddle/fluid/framework/data_layout_transform.h b/paddle/fluid/framework/data_layout_transform.h
index 90bb206ec6b..763fc18bf1c 100644
--- a/paddle/fluid/framework/data_layout_transform.h
+++ b/paddle/fluid/framework/data_layout_transform.h
@@ -53,7 +53,7 @@ inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) {
 inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
   static const std::map<std::type_index, MKLDNNDataType> dict{
       {std::type_index(typeid(float)), MKLDNNDataType::f32},  // NOLINT
-      {std::type_index(typeid(char)), MKLDNNDataType::s8},    // NOLINT
+      {std::type_index(typeid(signed char)), MKLDNNDataType::s8},    // NOLINT
       {std::type_index(typeid(unsigned char)), MKLDNNDataType::u8},
       {std::type_index(typeid(int16_t)), MKLDNNDataType::s16},
       {std::type_index(typeid(int32_t)), MKLDNNDataType::s32}};
diff --git a/paddle/fluid/framework/tensor_impl.h b/paddle/fluid/framework/tensor_impl.h
index 6d3047c95d6..efdf01504b6 100644
--- a/paddle/fluid/framework/tensor_impl.h
+++ b/paddle/fluid/framework/tensor_impl.h
@@ -25,8 +25,8 @@ inline const T* Tensor::data() const {
   check_memory_size();
   bool valid = std::is_same<T, void>::value ||
                holder_->type() == std::type_index(typeid(T));
-  PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s",
-                 this->holder_->type().name());
+  PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %d",
+                 this->holder_->type());
 
   return reinterpret_cast<const T*>(
       reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc
index 27a4f38e40f..ce5b53a521f 100644
--- a/paddle/fluid/operators/conv_mkldnn_op.cc
+++ b/paddle/fluid/operators/conv_mkldnn_op.cc
@@ -358,7 +358,7 @@ 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));
@@ -373,12 +373,14 @@ for(int i=0; i<50; i++){
 }
 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)
@@ -389,6 +391,7 @@ printf("\n");fflush(stdout);
         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;
         }
     }
 
@@ -398,7 +401,7 @@ printf("\n");fflush(stdout);
         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(
@@ -430,11 +433,20 @@ std::cout<<key_conv_pd<<std::endl;
         weights_md = platform::MKLDNNMemDesc(
             weights_tz, memory::data_type::s8,
             (g == 1) ? chosen_memory_format : mkldnn::memory::format::goihw);
+        auto dst_dt = fuse_relu? paddle::framework::ToMKLDNNDataType(std::type_index(typeid(unsigned char))) : paddle::framework::ToMKLDNNDataType(std::type_index(typeid(signed char)));
+        if(fuse_residual_conn){
+            auto residual = ctx.Input<Tensor>("ResidualData");
+            auto residual_dt = paddle::framework::ToMKLDNNDataType(residual->type());
+            if(dst_dt != residual_dt) 
+                dst_dt = residual_dt;
+        }
         dst_md = platform::MKLDNNMemDesc(
-            dst_tz,
-            fuse_relu? paddle::framework::ToMKLDNNDataType(std::type_index(typeid(unsigned char))) : 
-            paddle::framework::ToMKLDNNDataType(std::type_index(typeid(char))),
+            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);
     }
 
     // create a conv primitive descriptor and save it for usage in backward
@@ -486,7 +498,7 @@ 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++){
@@ -503,17 +515,48 @@ std::cout<<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);
-          if(fuse_relu){
+          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));
           }
@@ -563,7 +606,7 @@ std::cout<<"input fmt = "<<input->format()<<"   input dt = "<<paddle::framework:
       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++){
@@ -589,7 +632,10 @@ 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;
   }
@@ -612,7 +658,7 @@ 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 d7504972bce..9d4886454f4 100644
--- a/paddle/fluid/operators/dequantize_op.cc
+++ b/paddle/fluid/operators/dequantize_op.cc
@@ -49,6 +49,17 @@ std::cout<<"this is dequant op ***********"<<std::endl;
     float* output_data = output->mutable_data<float>(ctx.GetPlace());
     //T scale_data = *(scale->data<T>());
     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());
@@ -58,7 +69,7 @@ std::cout<<"this is dequant op ***********"<<std::endl;
 
     mkldnn::primitive_attr attri;
     int mask = 0;
-    attri.set_output_scales(mask, scale_data);
+    attri.set_output_scales(mask, reorder_scale);
 
     auto src_md = platform::MKLDNNMemDesc(
             {src_tz}, src_dt, src_fmt); 
@@ -75,6 +86,10 @@ std::cout<<"this is dequant op ***********"<<std::endl;
         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);
+    stream(stream::kind::eager).submit(pipeline).wait(); 
+
+    output->set_layout(DataLayout::kMKLDNN);
+    output->set_format(GetMKLDNNFormat(dst_memory));
 
   }
 };
diff --git a/paddle/fluid/operators/pool_mkldnn_op.cc b/paddle/fluid/operators/pool_mkldnn_op.cc
index d8799287ee1..cdb748f05d7 100644
--- a/paddle/fluid/operators/pool_mkldnn_op.cc
+++ b/paddle/fluid/operators/pool_mkldnn_op.cc
@@ -105,6 +105,16 @@ 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());
 
@@ -193,12 +203,25 @@ 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:
diff --git a/paddle/fluid/operators/requantize_op.cc b/paddle/fluid/operators/requantize_op.cc
index 6decbded6b0..a453534a560 100644
--- a/paddle/fluid/operators/requantize_op.cc
+++ b/paddle/fluid/operators/requantize_op.cc
@@ -36,9 +36,9 @@ class ReQuantOpKernel : 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 = 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();
@@ -47,18 +47,18 @@ class ReQuantOpKernel : public framework::OpKernel<T> {
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
     mkldnn::memory::data_type src_dt = paddle::framework::ToMKLDNNDataType(input->type());
-    mkldnn::memory::data_type dst_dt = paddle::framework::ToMKLDNNDataType(output->type());
+    mkldnn::memory::data_type dst_dt = mkldnn::memory::data_type::u8;//paddle::framework::ToMKLDNNDataType(output->type());
     mkldnn::memory::format src_fmt = memory::format::nhwc;//input->format();
     mkldnn::memory::format dst_fmt = memory::format::nhwc;//output->format();
 
     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());
     //T scale_data = *(scale->data<T>());
-    std::vector<T> scale_data = {*(scale->data<T>())};
+    std::vector<float> scale_data = {0.9999999}; //{*(scale->data<float>())};
 
     mkldnn::primitive_attr attri;
     int mask = 0;
-    attri.set_output_scales(mask, scale_data);
+    attri.set_output_scales(mask,scale_data);// scale_data);
     //attri.set_int_output_round_mode(round_nearest); //FIX ME
 
     auto src_md = platform::MKLDNNMemDesc(
@@ -70,13 +70,54 @@ class ReQuantOpKernel : public framework::OpKernel<T> {
     auto dst_md = platform::MKLDNNMemDesc(
             {dst_tz}, dst_dt, dst_fmt);
     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));    
-    auto reorder_p= std::shared_ptr<reorder>(new reorder(*reorder_pd, *src_memory_p, dst_memory));
-    pipeline.push_back(*reorder_p);
+        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);
+    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;
   }
 };
 
@@ -113,4 +154,4 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(requantize, ops::ReQuantOp, ops::ReQuantOpMaker, paddle::framework::DefaultGradOpDescMaker<true>);
 
-REGISTER_OP_KERNEL(requantize, MKLDNN, ::paddle::platform::CPUPlace, ops::ReQuantOpKernel<float>);
+REGISTER_OP_KERNEL(requantize, MKLDNN, ::paddle::platform::CPUPlace, ops::ReQuantOpKernel<int8_t>);
diff --git a/paddle/fluid/operators/softmax_mkldnn_op.cc b/paddle/fluid/operators/softmax_mkldnn_op.cc
index 01819f53e3a..6388a424737 100644
--- a/paddle/fluid/operators/softmax_mkldnn_op.cc
+++ b/paddle/fluid/operators/softmax_mkldnn_op.cc
@@ -16,6 +16,7 @@ limitations under the License. */
 #include "mkldnn.hpp"
 #include "paddle/fluid/operators/softmax_op.h"
 #include "paddle/fluid/platform/mkldnn_helper.h"
+#include "paddle/fluid/framework/data_layout_transform.h"
 
 namespace paddle {
 namespace operators {
@@ -131,6 +132,13 @@ 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;
@@ -162,7 +170,7 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
     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);
-- 
GitLab