Merge branch 'release_1.0.0_rc2' of ssh://git-ccr-1.devtools.intel.com:29418/dl_framework-intel_caffe

Author: daisyden

LICENSE

@@ -1,3 +1,4 @@
+
 COPYRIGHT
 
 All modification made by Intel Corporation: © 2017 Intel Corporation.

Makefile.mkldnn

@@ -8,6 +8,10 @@ MKLDNN_COMMIT := `cat ${CAFFE_ROOTDIR}/mkldnn.commit`
 MKLDNN_CXX := $(CXX)
 MKLDNN_CC := $(CC)
 
+RETURN_STRING=$(shell ./external/mkl/prepare_mkl.sh)
+MKLROOT=$(firstword $(RETURN_STRING))
+MKL_ROOTDIR := $(MKLROOT)
+
 # We do this because earlier versions of CMake have problems with ccache
 ifneq (,$(findstring ccache,$(CXX)))
 	MKLDNN_CXX := $(lastword $(CXX))
@@ -18,7 +22,7 @@ ifneq (,$(findstring ccache,$(CC)))
 endif
 
 MKLDNN_GITHUB := https://github.com/01org/mkl-dnn.git
-MKLDNN_CMAKE_FLAGS += $(MKLDNN_SRCDIR) -DCMAKE_INSTALL_PREFIX=$(CAFFE_ROOTDIR)/$(MKLDNN_INSTALLDIR) -B$(CAFFE_ROOTDIR)/$(MKLDNN_BUILDDIR) -DCMAKE_CXX_COMPILER="$(MKLDNN_CXX)" -DCMAKE_C_COMPILER="$(MKLDNN_CC)"
+MKLDNN_CMAKE_FLAGS += $(MKLDNN_SRCDIR) -DCMAKE_INSTALL_PREFIX=$(CAFFE_ROOTDIR)/$(MKLDNN_INSTALLDIR) -DMKLROOT=${MKL_ROOTDIR} -B$(CAFFE_ROOTDIR)/$(MKLDNN_BUILDDIR) -DCMAKE_CXX_COMPILER="$(MKLDNN_CXX)" -DCMAKE_C_COMPILER="$(MKLDNN_CC)"
 
 ifeq ("$(wildcard $(MKLDNN_INSTALLDIR)/include/mkldnn.hpp)", "")
 mkldnn_download:

README.md

@@ -71,3 +71,6 @@ Please cite Caffe in your publications if it helps your research:
 
 ***
  *Other names and brands may be claimed as the property of others
+
+
+

external/mkl/prepare_mkl.sh

@@ -74,10 +74,11 @@ echo $VERSION_LINE  # Return Version Line
 # MKL
 DST=`dirname $0`
 OMP=0 
-VERSION_MATCH=20170101
-ARCHIVE_BASENAME=mklml_lnx_2017.0.2.20170110.tgz
+VERSION_MATCH=20170425
+ARCHIVE_BASENAME=mklml_lnx_2018.0.20170425.tgz
 MKL_CONTENT_DIR=`echo $ARCHIVE_BASENAME | rev | cut -d "." -f 2- | rev`
-GITHUB_RELEASE_TAG=self_containted_MKLGOLD_u2
+GITHUB_RELEASE_TAG=1.0.0
+
 MKLURL="https://github.com/intel/caffe/releases/download/$GITHUB_RELEASE_TAG/$ARCHIVE_BASENAME"
 # there are diffrent MKL lib to be used for GCC and for ICC
 reg='^[0-9]+$'

include/caffe/layers/accuracy_layer.hpp

@@ -76,7 +76,7 @@ class AccuracyLayer : public Layer<Dtype> {
   // If there are two top blobs, then the second blob will contain
   // accuracies per class.
   virtual inline int MinTopBlobs() const { return 1; }
-  virtual inline int MaxTopBlos() const { return 2; }
+  virtual inline int MaxTopBlobs() const { return 2; }
 
  protected:
   /**

include/caffe/layers/base_conv_layer.hpp

@@ -153,6 +153,13 @@ class BaseConvolutionLayer : public Layer<Dtype> {
           pad_.cpu_data()[0], pad_.cpu_data()[1],
           stride_.cpu_data()[0], stride_.cpu_data()[1],
           dilation_.cpu_data()[0], dilation_.cpu_data()[1], col_buff);
+    } else if (!force_nd_im2col_ && num_spatial_axes_ == 3) {
+      im3d2col_cpu(data, conv_in_channels_,
+          conv_input_shape_.cpu_data()[1], conv_input_shape_.cpu_data()[2], conv_input_shape_.cpu_data()[3],
+          kernel_shape_.cpu_data()[0], kernel_shape_.cpu_data()[1], kernel_shape_.cpu_data()[2],
+          pad_.cpu_data()[0], pad_.cpu_data()[1], pad_.cpu_data()[2],
+          stride_.cpu_data()[0], stride_.cpu_data()[1], stride_.cpu_data()[2],
+          dilation_.cpu_data()[0], dilation_.cpu_data()[1], dilation_.cpu_data()[2], col_buff);
     } else {
       im2col_nd_cpu(data, num_spatial_axes_, conv_input_shape_.cpu_data(),
           col_buffer_shape_.data(), kernel_shape_.cpu_data(),
@@ -167,6 +174,13 @@ class BaseConvolutionLayer : public Layer<Dtype> {
           pad_.cpu_data()[0], pad_.cpu_data()[1],
           stride_.cpu_data()[0], stride_.cpu_data()[1],
           dilation_.cpu_data()[0], dilation_.cpu_data()[1], data);
+    } else if (!force_nd_im2col_ && num_spatial_axes_ == 3) {
+      col2im3d_cpu(col_buff, conv_in_channels_,
+          conv_input_shape_.cpu_data()[1], conv_input_shape_.cpu_data()[2], conv_input_shape_.cpu_data()[3],
+          kernel_shape_.cpu_data()[0], kernel_shape_.cpu_data()[1], kernel_shape_.cpu_data()[2],
+          pad_.cpu_data()[0], pad_.cpu_data()[1], pad_.cpu_data()[2],
+          stride_.cpu_data()[0], stride_.cpu_data()[1], stride_.cpu_data()[2],
+          dilation_.cpu_data()[0], dilation_.cpu_data()[1], dilation_.cpu_data()[2], data);
     } else {
       col2im_nd_cpu(col_buff, num_spatial_axes_, conv_input_shape_.cpu_data(),
           col_buffer_shape_.data(), kernel_shape_.cpu_data(),

include/caffe/layers/mkldnn_layers.hpp

@@ -100,6 +100,9 @@ class SoftmaxWithLossLayer : public LossLayer<Dtype> {
   virtual inline int MinTopBlobs() const { return 1; }
   virtual inline int MaxTopBlobs() const { return 2; }
 
+  virtual inline int ExactNumBottomBlobs() const { return -1; }
+  virtual inline int MinBottomBlobs() const { return 2; }
+  virtual inline int MaxBottomBlobs() const { return 3; }
  protected:
   virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top);

include/caffe/layers/softmax_loss_layer.hpp

@@ -203,6 +203,10 @@ template <typename Dtype>
 class MKLDNNPrimitive {
 public:
     explicit MKLDNNPrimitive():aprimitive(), mkldnn_stream() {}
+
+    //API for initializing with shared_ptr<primitive>
+    MKLDNNPrimitive(shared_ptr<primitive> aprimitive_input) {this->aprimitive = aprimitive_input;}
+
     virtual ~MKLDNNPrimitive() {}
     void reset(primitive* pprimitive) { this->aprimitive.reset(pprimitive);}
     shared_ptr<primitive> aprimitive;

include/caffe/mkldnn_base.hpp

@@ -372,13 +372,16 @@ int collectAndCheckLayerData(bool collect_step,
         }
 
         if (bottom_need_backward[i].size() > 0 && bottom_need_backward[i][0]) {
-            getFileName(file_name, false, "FwrdBtmDat", i);
-            checkData(file_name, bottom_vecs[i][0]->cpu_data(),
-                layers[i]->type(), output_dir,
-                &erronous_layers);
-            checkAllNans(bottom_vecs[i][0]->cpu_diff(),
-                bottom_vecs[i][0]->count(), "bottom.diff",
-                layers[i]->type(), &erronous_layers);
+          // We check data only for out-of-place computations
+          if (bottom_vecs[i][0] != top_vecs[i][0]) {
+              getFileName(file_name, false, "FwrdBtmDat", i);
+              checkData(file_name, bottom_vecs[i][0]->cpu_data(),
+                  layers[i]->type(), output_dir,
+                  &erronous_layers);
+          }
+          checkAllNans(bottom_vecs[i][0]->cpu_diff(),
+              bottom_vecs[i][0]->count(), "bottom.diff",
+              layers[i]->type(), &erronous_layers);
         }
 
         checkAllNans(top_vecs[i][0]->cpu_diff(),