Merge "Relu optimization related to ICL-84."

Author: Yu, Chong

src/caffe/layers/mkldnn_relu_layer.cpp

@@ -188,10 +188,31 @@ void MKLDNNReLULayer<Dtype>::InitReLUBwd(const vector<Blob<Dtype>*>& top
     if (top_diff_is_prv) {
       shared_ptr<MKLDNNMemoryDescriptor<Dtype, /* is_diff */ true> > mem_descr
         = get_mkldnn_prv_descriptor<Dtype, /* is_diff */ true>(top[0]);
+#ifdef DEBUG
+      memory::format bwd_prv_top_diff_mfmt = static_cast<memory::format>(mem_descr->prv_memory_pd()->desc().data.format);
+      LOG(INFO) << "MKLDNNReLULayer<Dtype>::InitReLUBwd: memory format of prv top diff is: " << bwd_prv_top_diff_mfmt;
+#endif
       top_diff_md.reset(new memory::desc(mem_descr->prv_memory_pd()->desc()));
       usr_diff_mpd = mem_descr->usr_memory_pd();
       prv_diff_mpd = mem_descr->prv_memory_pd();
     } else {
+      bool bottom_data_is_prv = (const_cast<Dtype*>(bottom[0]->prv_data()) != NULL);
+      if (bottom_data_is_prv) {
+          shared_ptr<MKLDNNMemoryDescriptor<Dtype, false> > mem_descr
+              = get_mkldnn_prv_descriptor<Dtype, false>(bottom[0]);
+#ifdef DEBUG
+          memory::format fwd_prv_bottom_data_mfmt = static_cast<memory::format>(mem_descr->prv_memory_pd()->desc().data.format);
+          LOG(INFO) << "MKLDNNReLULayer<Dtype>::InitReLUBwd: memory format of prv bottom data is: " << fwd_prv_bottom_data_mfmt;
+          LOG(INFO) << "MKLDNNReLULayer<Dtype>::InitReLUBwd: Reorder the usr top diff to the format of prv bottom data! (Performance consideration)";              
+#endif
+          prv_diff_mpd = mem_descr->prv_memory_pd();
+          //top[0]->prv_data() is empty, however top[0]->get_prv_diff_descriptor() has value.
+          //Find root cause in the mkldnn_memory: create_output_memory() and sync_before_write() functions.
+          //But that a major fix, will lead the nan in the AlexNet training.
+          //So need investigation further, however, this will fix ICL-84.
+          top[0]->set_prv_diff_descriptor(NULL);
+      }
+
       top_diff_md.reset(new memory::desc({{n, ic, ih, iw}}, mpcsn, memory::format::nchw));
       usr_diff_mpd.reset(new memory::primitive_desc(*top_diff_md, cpu_engine));
     }

src/caffe/mkldnn_memory.cpp

@@ -374,6 +374,16 @@ void MKLDNNMemoryDescriptor<Dtype, is_diff>::sync_before_write(bool inplace)
             this->_blob->set_prv_data_descriptor(this->get_shared_ptr(), this->conversion_needed() ? false : true);
         }
     }
+    //Fix me: this->conversion_needed() == false means diff/data is in the CPU, no need to set the prv_diff/data_descriptor
+    /*
+    if ((!inplace) && (this->conversion_needed())) {
+        if (is_diff) {
+            this->_blob->set_prv_diff_descriptor(this->get_shared_ptr(), false);
+        } else {
+            this->_blob->set_prv_data_descriptor(this->get_shared_ptr(), false);
+        }
+    }
+    */
 }
 
 template <typename Dtype, bool is_diff>
@@ -420,7 +430,7 @@ shared_ptr<primitive> MKLDNNMemoryDescriptor<Dtype, is_diff>::create_input(bool
 template <typename Dtype, bool is_diff>
 shared_ptr<memory> MKLDNNMemoryDescriptor<Dtype, is_diff>::create_output_memory(bool inplace)
 {
-    // TODO: need to iptimize code
+    // TODO: need to optimize code
     shared_ptr<memory> omem = create_output_memory(this->_blob);
     if(!inplace) {
         if(is_diff) {
@@ -429,6 +439,16 @@ shared_ptr<memory> MKLDNNMemoryDescriptor<Dtype, is_diff>::create_output_memory(
             this->_blob->set_prv_data_descriptor(this->get_shared_ptr(), this->conversion_needed() ? false : true);
         }
     }
+    /*
+    //Fix me: this->conversion_needed() == false means diff/data is in the CPU, no need to set the prv_diff/data_descriptor
+    if ((!inplace) && (this->conversion_needed())) {
+        if (is_diff) {
+            this->_blob->set_prv_diff_descriptor(this->get_shared_ptr(), false);
+        } else {
+            this->_blob->set_prv_data_descriptor(this->get_shared_ptr(), false);
+        }
+    }
+    */
     return omem;
 }