Support MKL2017 as default engine

CMakeLists.txt

@@ -29,7 +29,7 @@ include(cmake/ConfigGen.cmake)
 caffe_option(CPU_ONLY  "Build Caffe without CUDA support" OFF) # TODO: rename to USE_CUDA
 caffe_option(USE_OPENMP "Build Caffe with OpenMP support" ON )
 caffe_option(USE_CUDNN "Build Caffe with cuDNN library support" ON IF NOT CPU_ONLY)
-caffe_option(USE_MKL2017_AS_DEFAULT_ENGINE "Use MKL2017 primitives for supported layers" OFF)
+caffe_option(USE_MKL2017_AS_DEFAULT_ENGINE "Use MKL2017 primitives for supported layers" ON)
 caffe_option(USE_MKLDNN_AS_DEFAULT_ENGINE "Use MKL-DNN primitives for supported layers" OFF)
 caffe_option(BUILD_SHARED_LIBS "Build shared libraries" ON)
 caffe_option(BUILD_python "Build Python wrapper" ON)

Makefile.config.example

@@ -43,8 +43,8 @@
 # CPU-only switch (uncomment to build without GPU support).
 CPU_ONLY := 1
 
-# USE_MKL2017_AS_DEFAULT_ENGINE flag is OBSOLETE
-# Put this at the top your train_val.protoxt or solver.prototxt file:
+USE_MKL2017_AS_DEFAULT_ENGINE := 1
+# or put this at the top your train_val.protoxt or solver.prototxt file:
 # engine: "MKL2017" 
 # or use this option with caffe tool:
 # -engine "MKL2017"

docs/release_notes.md

@@ -126,8 +126,6 @@ This Caffe version is seflcontained. This means that newest version of Intel MKL
 * Set layer engine to `MKL2017` in prototxt file (model). Only this specific layer will be accelerated with new primitives.
 * Use -engine = MKL2017 in command line as an option during execution of caffe (training, scoring, benchmark)
 
-Comment: there is obsolete method to compale with `USE_MKL2017_AS_DEFAULT_ENGINE := 1` in `Makefile.config`. This is obsolete solution - not recommended to use anymore.
-
 ### Building for GPU
 Caffe requires the CUDA `nvcc` compiler to compile its GPU code and CUDA driver for GPU operation.
 To install CUDA, go to the [NVIDIA CUDA website](https://developer.nvidia.com/cuda-downloads) and follow installation instructions there. Install the library and the latest standalone driver separately; the driver bundled with the library is usually out-of-date. **Warning!** The 331.* CUDA driver series has a critical performance issue: do not use it.

src/caffe/net.cpp

@@ -189,8 +189,18 @@ void Net<Dtype>::Init(const NetParameter& in_param) {
     }
     // Setup layer.
     const LayerParameter& layer_param = param.layer(layer_id);
-    if (param.engine() != "" && param.layer(layer_id).engine() == "")
-      param.mutable_layer(layer_id)->set_engine(param.engine());
+    if (param.engine() != "") {
+      if (param.layer(layer_id).engine() == "") {
+        param.mutable_layer(layer_id)->set_engine(param.engine());
+      }
+      else {
+        if ((!param.layer(layer_id).engine().compare("MKL2017") && !param.engine().compare("MKLDNN")) 
+           || (!param.layer(layer_id).engine().compare("MKLDNN") && !param.engine().compare("MKL2017"))) {
+          param.mutable_layer(layer_id)->set_engine(param.engine());
+        }
+      }
+    }
+
     if (layer_param.propagate_down_size() > 0) {
       CHECK_EQ(layer_param.propagate_down_size(),
           layer_param.bottom_size())
@@ -490,8 +500,6 @@ template <typename Dtype>
 void Net<Dtype>::CompileNet(const NetParameter& param,
     NetParameter* param_compiled) {
 
-
-
   NetParameter param_temp0;
   param_temp0.CopyFrom(param);
   param_temp0.clear_layer();