Merge remote-tracking branch 'remotes/internal/release_1.1.0'

CMakeLists.txt

@@ -10,8 +10,8 @@ endif()
 project(Caffe C CXX)
 
 # ---[ Caffe version
-set(CAFFE_TARGET_VERSION "1.0.0-rc3" CACHE STRING "Caffe logical version")
-set(CAFFE_TARGET_SOVERSION "1.0.0-rc3" CACHE STRING "Caffe soname version")
+set(CAFFE_TARGET_VERSION "1.1.0" CACHE STRING "Caffe logical version")
+set(CAFFE_TARGET_SOVERSION "1.1.0" CACHE STRING "Caffe soname version")
 add_definitions(-DCAFFE_VERSION=${CAFFE_TARGET_VERSION})
 
 # ---[ Using cmake scripts and modules
@@ -26,7 +26,7 @@ include(cmake/Summary.cmake)
 include(cmake/ConfigGen.cmake)
 
 # ---[ Options
-caffe_option(CPU_ONLY  "Build Caffe without CUDA support" OFF) # TODO: rename to USE_CUDA
+caffe_option(CPU_ONLY  "Build Caffe without CUDA support" ON) # 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)

Makefile

@@ -60,6 +60,9 @@ else
 	OTHER_BUILD_DIR := $(DEBUG_BUILD_DIR)
 endif
 
+ifeq ($(CAFFE_PER_LAYER_TIMINGS), 1)
+	COMMON_FLAGS += -DCAFFE_PER_LAYER_TIMINGS
+endif
 
 #################### MLSL ####################
 
@@ -71,17 +74,12 @@ endif
 
 	RETURN_STRING=$(shell ./external/mlsl/prepare_mlsl.sh)
 	MLSL_ROOT=$(firstword $(RETURN_STRING))
-	MLSL_LDFLAGS=$(lastword $(RETURN_STRING))
+	MLSL_LDFLAGS:=-l$(lastword $(RETURN_STRING)) -Wl,-rpath,$(MLSL_ROOT)/intel64/lib
 	COMMON_FLAGS += -DUSE_MLSL=1
 	LIBRARIES += mlsl
 	INCLUDE_DIRS += $(MLSL_ROOT)/intel64/include
 	LIBRARY_DIRS += $(MLSL_ROOT)/intel64/lib
-	IGNORE := $(shell bash -c "source $(MLSL_ROOT)/intel64/bin/mlslvars.sh; env | sed 's/=/:=/' | sed 's/^/export /' > make_mlsl_env")
-	include make_mlsl_env
-
-ifeq ($(CAFFE_PER_LAYER_TIMINGS), 1)
-	COMMON_FLAGS += -DCAFFE_PER_LAYER_TIMINGS
-endif
+	COMMON_FLAGS += -DFOUNDED_MLSL_ROOT=$(MLSL_ROOT)
 
 ifeq ($(CAFFE_MLSL_SHUFFLE), 1)
 	COMMON_FLAGS += -DCAFFE_MLSL_SHUFFLE
@@ -103,8 +101,8 @@ LIBRARY_NAME := $(PROJECT)
 LIB_BUILD_DIR := $(BUILD_DIR)/lib
 STATIC_NAME := $(LIB_BUILD_DIR)/lib$(LIBRARY_NAME).a
 DYNAMIC_VERSION_MAJOR 		:= 1
-DYNAMIC_VERSION_MINOR 		:= 0
-DYNAMIC_VERSION_REVISION 	:= 0-rc3
+DYNAMIC_VERSION_MINOR 		:= 1
+DYNAMIC_VERSION_REVISION 	:= 0
 DYNAMIC_NAME_SHORT := lib$(LIBRARY_NAME).so
 #DYNAMIC_SONAME_SHORT := $(DYNAMIC_NAME_SHORT).$(DYNAMIC_VERSION_MAJOR)
 DYNAMIC_VERSIONED_NAME_SHORT := $(DYNAMIC_NAME_SHORT).$(DYNAMIC_VERSION_MAJOR).$(DYNAMIC_VERSION_MINOR).$(DYNAMIC_VERSION_REVISION)
@@ -336,6 +334,9 @@ endif
 ifdef CUSTOM_CXX
 	CXX := $(CUSTOM_CXX)
 endif
+ifdef CUSTOM_CC
+	CC := $(CUSTOM_CC)
+endif
 
 # Linux
 ifeq ($(LINUX), 1)
@@ -492,6 +493,11 @@ ifeq ($(DISABLE_BN_FOLDING), 1)
 	COMMON_FLAGS += -DDISABLE_BN_FOLDING
 endif
 
+# Disable the conv/eltwise/relu layer fusion
+ifeq ($(DISABLE_CONV_SUM_FUSION), 1)
+	COMMON_FLAGS += -DDISABLE_CONV_SUM_FUSION
+endif
+
 # Performance monitoring
 ifeq ($(PERFORMANCE_MONITORING), 1)
 	CXXFLAGS += -DPERFORMANCE_MONITORING
@@ -772,7 +778,7 @@ $(ALL_BUILD_DIRS): | $(BUILD_DIR_LINK)
 
 $(DYNAMIC_NAME): $(OBJS) | $(LIB_BUILD_DIR)
 	@ echo LD -o $@
-	$(Q)$(CXX) -shared -o $@ $(OBJS) $(VERSIONFLAGS) $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_SHARED_HARDENING_FLAGS) $(LDFLAGS)
+	$(Q)$(CXX) -shared -o $@ $(OBJS) $(VERSIONFLAGS) $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MLSL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_SHARED_HARDENING_FLAGS) $(LDFLAGS)
 	@ cd $(BUILD_DIR)/lib; rm -f $(DYNAMIC_NAME_SHORT);   ln -s $(DYNAMIC_VERSIONED_NAME_SHORT) $(DYNAMIC_NAME_SHORT)
 
 $(STATIC_NAME): $(OBJS) | $(LIB_BUILD_DIR)
@@ -804,7 +810,7 @@ $(TEST_ALL_BIN): $(TEST_MAIN_SRC) $(TEST_OBJS) $(GTEST_OBJS) \
 		| $(DYNAMIC_NAME) $(TEST_BIN_DIR)
 	@ echo CXX/LD -o $@ $<
 	$(Q)$(CXX) -std=c++11 $(TEST_MAIN_SRC) $(TEST_OBJS) $(GTEST_OBJS) \
-		-o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) $(LDFLAGS) -l$(LIBRARY_NAME) -Wl,-rpath,$(ORIGIN)/../lib
+		-o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MLSL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) $(LDFLAGS) -l$(LIBRARY_NAME) -Wl,-rpath,$(ORIGIN)/../lib
 
 $(TEST_CU_BINS): $(TEST_BIN_DIR)/%.testbin: $(TEST_CU_BUILD_DIR)/%.o \
 	$(GTEST_OBJS) | $(DYNAMIC_NAME) $(TEST_BIN_DIR)
@@ -816,7 +822,7 @@ $(TEST_CXX_BINS): $(TEST_BIN_DIR)/%.testbin: $(TEST_CXX_BUILD_DIR)/%.o \
 	$(GTEST_OBJS) | $(DYNAMIC_NAME) $(TEST_BIN_DIR)
 	@ echo LD $<
 	$(Q)$(CXX) -std=c++11 $(TEST_MAIN_SRC) $< $(GTEST_OBJS) \
-		-o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) $(LDFLAGS) -l$(LIBRARY_NAME) -Wl,-rpath,$(ORIGIN)/../lib
+		-o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MLSL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) $(LDFLAGS) -l$(LIBRARY_NAME) -Wl,-rpath,$(ORIGIN)/../lib
 
 # Target for extension-less symlinks to tool binaries with extension '*.bin'.
 $(TOOL_BUILD_DIR)/%: $(TOOL_BUILD_DIR)/%.bin | $(TOOL_BUILD_DIR)
@@ -825,12 +831,12 @@ $(TOOL_BUILD_DIR)/%: $(TOOL_BUILD_DIR)/%.bin | $(TOOL_BUILD_DIR)
 
 $(TOOL_BINS): %.bin : %.o | $(DYNAMIC_NAME)
 	@ echo CXX/LD -o $@
-	$(Q)$(CXX) $< -o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) -l$(LIBRARY_NAME) $(LDFLAGS) \
+	$(Q)$(CXX) $< -o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MLSL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) -l$(LIBRARY_NAME) $(LDFLAGS) \
 		-Wl,-rpath,$(ORIGIN)/../lib
 
 $(EXAMPLE_BINS): %.bin : %.o | $(DYNAMIC_NAME)
 	@ echo CXX/LD -o $@
-	$(Q)$(CXX) $< -o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) -l$(LIBRARY_NAME) $(LDFLAGS) \
+	$(Q)$(CXX) $< -o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MLSL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) -l$(LIBRARY_NAME) $(LDFLAGS) \
 		-Wl,-rpath,$(ORIGIN)/../../lib
 
 proto: $(PROTO_GEN_CC) $(PROTO_GEN_HEADER)

Makefile.config.example

@@ -76,8 +76,11 @@ USE_MKLDNN_AS_DEFAULT_ENGINE := 1
 # already. 
 # BOOST_ROOT := 
 
-# Use remove batch norm optimization to boost inferrence
+# Use remove batch norm optimization to boost inference
 DISABLE_BN_FOLDING := 0
+
+#Use conv/eltwise/relu layer fusion to boost inference.
+DISABLE_CONV_SUM_FUSION := 0
 # Intel(r) Machine Learning Scaling Library (uncomment to build
 # with MLSL for multi-node training)
 # USE_MLSL :=1

Makefile.dlcp

@@ -42,6 +42,7 @@ endif
 endif
 
 ifeq ($(USE_MLSL), 1)
+ifneq ($(ENABLE_WEIGHT_GRAD_COMPRESSION), 0)
 ifeq ($(origin DLCPROOT), environment)
 dlcp: dlcproot_set
 else
@@ -53,6 +54,10 @@ dlcp:
 endif
 endif
 else
+dlcp:
+	@echo "disabling weight grad compression if ENABLE_WEIGHT_GRAD_COMPRESSION is set to 0"
+endif
+else
 dlcp:
 	@echo "disabling weight grad compression if USE_MLSL is not 1"
 endif

cmake/Dependencies.cmake

@@ -94,6 +94,10 @@ if(USE_OPENCV)
   add_definitions(-DUSE_OPENCV)
 endif()
 
+if(CAFFE_PER_LAYER_TIMINGS)
+  add_definitions("-DCAFFE_PER_LAYER_TIMINGS")
+endif()
+
 # ---[ MLSL
 if(USE_MLSL)
   if (NOT CPU_ONLY)
@@ -116,13 +120,11 @@ if(USE_MLSL)
   endif()
   message(STATUS "Machine Learning Scaling Library (MLSL) found (${MLSL_ROOT}/intel64)")
   add_definitions("-DUSE_MLSL=1")
+  add_definitions("-DFOUNDED_MLSL_ROOT=${MLSL_ROOT}")
   include_directories(SYSTEM "${MLSL_ROOT}/intel64/include")
   link_directories(SYSTEM "${MLSL_ROOT}/intel64/lib")
   list(APPEND Caffe_LINKER_LIBS mlsl)
 
-  if(CAFFE_PER_LAYER_TIMINGS)
-    add_definitions("-DCAFFE_PER_LAYER_TIMINGS")
-  endif()
   if(CAFFE_MLSL_SHUFFLE)
     add_definitions("-DCAFFE_MLSL_SHUFFLE")
   endif()
@@ -134,28 +136,30 @@ endif()
 
 # ---| DLCP
 if(USE_MLSL)
-  if(DEFINED ENV{DLCPROOT})
-    set(DLCPROOT_DIR $ENV{DLCPROOT})
-    if(NOT ${DLCPROOT_DIR} STREQUAL "")
-      set(DLCPROOT_INCLUDE_DIR "${DLCPROOT_DIR}/include/")
-      set(DLCPROOT_LIB_DIR "${DLCPROOT_DIR}/lib/")
-      if(EXISTS ${DLCPROOT_INCLUDE_DIR}/dl_compression.h AND EXISTS ${DLCPROOT_LIB_DIR}/libdlcomp.so)
-        message(STATUS "Found DLCP: ${DLCPROOT_DIR}")
+  if (ENABLE_WEIGHT_GRAD_COMPRESSION OR NOT DEFINED ENABLE_WEIGHT_GRAD_COMPRESSION)
+    if(DEFINED ENV{DLCPROOT})
+      set(DLCPROOT_DIR $ENV{DLCPROOT})
+      if(NOT ${DLCPROOT_DIR} STREQUAL "")
+        set(DLCPROOT_INCLUDE_DIR "${DLCPROOT_DIR}/include/")
+        set(DLCPROOT_LIB_DIR "${DLCPROOT_DIR}/lib/")
+        if(EXISTS ${DLCPROOT_INCLUDE_DIR}/dl_compression.h AND EXISTS ${DLCPROOT_LIB_DIR}/libdlcomp.so)
+          message(STATUS "Found DLCP: ${DLCPROOT_DIR}")
+          set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DENABLE_WEIGHT_GRAD_COMPRESSION")
+          list(APPEND Caffe_LINKER_LIBS "${DLCPROOT_LIB_DIR}/libdlcomp.so")
+          include_directories(SYSTEM ${DLCP_INCLUDE_DIR})
+        else()
+          message(STATUS "DLCP not found. DLCP_INCLUDE_DIR = ${DLCPROOT_INCLUDE_DIR} DLCP_LIB_DIR = ${DLCPROOT_LIB_DIR}")
+          message(WARNING "weight grad compression is disabled.")
+        endif()
+      endif()
+    else()
+      Download_DLCP()
+      if(DLCP_CXX)
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DENABLE_WEIGHT_GRAD_COMPRESSION")
-        list(APPEND Caffe_LINKER_LIBS "${DLCPROOT_LIB_DIR}/libdlcomp.so")
-        include_directories(SYSTEM ${DLCP_INCLUDE_DIR})
-      else()
-        message(STATUS "DLCP not found. DLCP_INCLUDE_DIR = ${DLCPROOT_INCLUDE_DIR} DLCP_LIB_DIR = ${DLCPROOT_LIB_DIR}")
-        message(WARNING "weight grad compression is disabled.")
+        list(APPEND Caffe_LINKER_LIBS "${DLCP_LIBDIR}/libdlcomp.so")
+        include_directories(SYSTEM ${DLCP_INCLDIR})
       endif()
     endif()
-  else()
-    Download_DLCP()
-    if(DLCP_CXX)
-      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DENABLE_WEIGHT_GRAD_COMPRESSION")
-      list(APPEND Caffe_LINKER_LIBS "${DLCP_LIBDIR}/libdlcomp.so")
-      include_directories(SYSTEM ${DLCP_INCLDIR})
-    endif()
   endif()
 endif()
 

cmake/Misc.cmake

@@ -10,6 +10,12 @@ if(DISABLE_BN_FOLDING)
 	message(STATUS "Bn folding is disabled!")
 	add_definitions("-DDISABLE_BN_FOLDING")
 endif()
+
+if(DISABLE_CONV_SUM_FUSION)
+	message(STATUS "conv/eltwise/relu fusion is disabled!")
+	add_definitions("-DDISABLE_CONV_SUM_FUSION")
+endif()
+
 # --[ If user doesn't specify build type then assume release
 if("${CMAKE_BUILD_TYPE}" STREQUAL "")
   set(CMAKE_BUILD_TYPE Release)

docker/README.md

@@ -17,7 +17,7 @@ docker run -ti caffe:cpu caffe --version
 ```
 which should show a message like:
 ```
-caffe version 1.0.0-rc3
+caffe version 1.1.0
 ```
 
 One can also build and run the Caffe tests in the image using:

examples/pycaffe/tune_model.py

@@ -10,7 +10,13 @@ def genOptimalModel(net, mkldnn_direct_time_map, mkldnn_winograd_time_map, optim
     for index in range(0, len(net.layer)):
         l = net.layer[index]
         if l.type == "Convolution":
-            if mkldnn_winograd_time_map[l.name] < mkldnn_direct_time_map[l.name]:
+            if len(l.convolution_param.kernel_size) == 0:
+                continue
+            kernel_size = l.convolution_param.kernel_size[0]
+            stride = 1
+            if len(l.convolution_param.stride) != 0:
+                stride = l.convolution_param.stride[0]
+            if mkldnn_winograd_time_map[l.name] < mkldnn_direct_time_map[l.name] and kernel_size == 3 and stride == 1 and l.convolution_param.num_output % 16 ==0:
                 l.convolution_param.conv_algorithm = "winograd"
             else:
                 l.convolution_param.conv_algorithm = "direct"

external/mkl/prepare_mkl.sh

@@ -74,12 +74,11 @@ echo $VERSION_LINE  # Return Version Line
 # MKL
 DST=`dirname $0`
 OMP=0 
-VERSION_MATCH=20171007
-ARCHIVE_BASENAME=mklml_lnx_2018.0.1.20171007.tgz
+VERSION_MATCH=20171227
+ARCHIVE_BASENAME=mklml_lnx_2018.0.1.20171227.tgz
 MKL_CONTENT_DIR=`echo $ARCHIVE_BASENAME | rev | cut -d "." -f 2- | rev`
-GITHUB_RELEASE_TAG=1.0.6
 
-MKLURL="https://github.com/intel/caffe/releases/download/$GITHUB_RELEASE_TAG/$ARCHIVE_BASENAME"
+MKLURL="https://github.com/01org/mkl-dnn/releases/download/v0.12/$ARCHIVE_BASENAME"
 # there are diffrent MKL lib to be used for GCC and for ICC
 reg='^[0-9]+$'
 VERSION_LINE=`GetVersionName $MKLROOT`

include/caffe/filler.hpp

@@ -230,7 +230,7 @@ template <typename Dtype> class MSRAFiller : public Filler<Dtype> {
                FillerParameter_VarianceNorm_FAN_OUT) {
       n = fan_out;
     }
-    Dtype std = sqrt(Dtype(2) / n);
+    Dtype std = this->filler_param_.scale() * sqrt(Dtype(2) / n);
     caffe_rng_gaussian<Dtype>(blob->count(), Dtype(0), std,
                               blob->mutable_cpu_data());
     CHECK_EQ(this->filler_param_.sparse(), -1)

include/caffe/layers/base_conv_layer.hpp

@@ -63,8 +63,9 @@ class BaseConvolutionLayer : public Layer<Dtype> {
 
   virtual inline int MinBottomBlobs() const { return 1; }
   virtual inline int MinTopBlobs() const { return 1; }
+#ifdef DISABLE_CONV_SUM_FUSION
   virtual inline bool EqualNumBottomTopBlobs() const { return true; }
-
+#endif
  protected:
   // Split Reshape into two parts
   // Part 1 for normal blob reshape