Merge commit for internal changes

 - 02f2c8f - CUDA 9.0 Toolkit is no longer required for pre-built binaries @cliffwoolley
 - 696d5ff - Make only API functions exposed. @JanuszL
 - 26ce6ac - Rework ABI compatibility @cliffwoolley
 - e320b7e - Namespace prefix C API @cliffwoolley
 - 71c7eb7 - logging.warn() is obsolete @cliffwoolley
 - 7bca211 - Fix crash when no files are loaded by reader @JanuszL
 - 9e4cdab - Test importing DALI simultaneously w/ every framework @cliffwoolley

.gitignore

@@ -11,3 +11,4 @@ latex
 valgrind-out.txt
 **/.ipynb_checkpoints
 .dockerignore
+wheelhouse

CMakeLists.txt

@@ -54,7 +54,7 @@ list(APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake/modules)
 include(cmake/Dependencies.cmake)
 
 # CXX flags
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -Wall -Wno-unused-variable -Wno-unused-function -fno-strict-aliasing -O2 -fPIC")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -Wall -Wno-unused-variable -Wno-unused-function -fno-strict-aliasing -O2 -fPIC -fvisibility=hidden")
 
 # Add ptx & bin flags for cuda
 set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} \

Dockerfile

@@ -1,5 +1,5 @@
 #########################################################################################
-##  Stage 2: build DALI wheels using manylinux1 (CentOS 5 derivative)
+##  Stage 2: build DALI wheels on top of the dependencies image built in Stage 1
 #########################################################################################
 ARG DEPS_IMAGE_NAME
 FROM ${DEPS_IMAGE_NAME}

Dockerfile.deps

@@ -1,7 +1,16 @@
 #########################################################################################
-##  Stage 1: build DALI dependencies using manylinux1 (CentOS 5 derivative)
+##  Stage 1: build DALI dependencies
+##     We would prefer to use publicly maintained manylinux1 (CentOS 5 derivative), but
+##     there are CXX11 ABI problems with the rh-devtoolset2 included with it, so we
+##     have switched to a "manylinux3" base that is just a minimally altered manylinux1
+##     built from CentOS 7 (w/ devtoolset3) instead of CentOS 5.  You can switch back
+##     to manylinux1 (or any other base you like) by changing FROM_IMAGE_NAME below
+##     or by passing an argument of the form `--build-arg FROM_IMAGE_NAME=myimage:mytag`
+##     to `docker build -f Dockerfile.deps ... .`
 #########################################################################################
-FROM quay.io/pypa/manylinux1_x86_64
+#ARG FROM_IMAGE_NAME=quay.io/pypa/manylinux1_x86_64
+ARG FROM_IMAGE_NAME=gitlab-dl.nvidia.com:5005/dali/manylinux:manylinux3_x86_64
+FROM ${FROM_IMAGE_NAME}
 
 # Install yum Dependencies
 RUN yum install -y zip
@@ -29,14 +38,17 @@ RUN PROTOBUF_VERSION=3.5.1 && \
     curl -L https://github.com/google/protobuf/releases/download/v${PROTOBUF_VERSION}/protobuf-all-${PROTOBUF_VERSION}.tar.gz | tar -xzf - && \
     cd /protobuf-${PROTOBUF_VERSION} && \
     ./autogen.sh && \
-    ./configure --prefix=/usr/local 2>&1 > /dev/null && \
+    ./configure CXXFLAGS="-fPIC" --prefix=/usr/local --disable-shared 2>&1 > /dev/null && \
     make -j"$(grep ^processor /proc/cpuinfo | wc -l)" install 2>&1 > /dev/null && \
     rm -rf /protobuf-${PROTOBUF_VERSION}
 
 # LMDB
+COPY docker/Makefile-lmdb.patch /tmp
 RUN LMDB_VERSION=0.9.22 && \
     git clone -b LMDB_${LMDB_VERSION} --single-branch https://github.com/LMDB/lmdb && \
     cd /lmdb/libraries/liblmdb && \
+    patch -p3 < /tmp/Makefile-lmdb.patch && \
+    rm -f /tmp/Makefile-lmdb.patch && \
     make -j"$(grep ^processor /proc/cpuinfo | wc -l)" install && \
     rm -rf /lmdb
 
@@ -45,6 +57,7 @@ RUN OPENCV_VERSION=3.1.0 && \
     curl -L https://github.com/opencv/opencv/archive/${OPENCV_VERSION}.tar.gz | tar -xzf - && \
     cd /opencv-${OPENCV_VERSION} && \
     cmake -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=/usr/local \
+          -DBUILD_SHARED_LIBS=OFF \
           -DWITH_CUDA=OFF -DWITH_1394=OFF -DWITH_IPP=OFF -DWITH_OPENCL=OFF -DWITH_GTK=OFF \
           -DBUILD_DOCS=OFF -DBUILD_TESTS=OFF -DBUILD_PERF_TESTS=OFF \
           -DBUILD_opencv_cudalegacy=OFF -DBUILD_opencv_stitching=OFF . && \
@@ -77,9 +90,10 @@ RUN CUDA_VERSION=9.0 && \
 
 # NVJPEG
 RUN NVJPEG_VERSION=9.0 && \
+    NVJPEG_BUILD=9.0.462 && \
     curl -L https://developer.download.nvidia.com/compute/redist/libnvjpeg/cuda-linux64-nvjpeg-${NVJPEG_VERSION}.tar.gz | tar -xzf - && \
     cd /cuda-linux64-nvjpeg/ && \
-    mv lib64/libnvjpeg.so* /usr/local/lib/ && \
+    mv lib64/libnvjpeg*.a* /usr/local/lib/ && \
     mv include/nvjpeg.h /usr/local/include/ && \
     rm -rf /cuda-linux64-nvjpeg
 

README.md

@@ -18,8 +18,9 @@ Note: DALI v0.1 is a pre-release software, which means certain features may not
 
 ## Prerequisities
 
-* Linux
-* [NVIDIA CUDA 9.0](https://developer.nvidia.com/cuda-downloads)
+* Linux x64
+* [NVIDIA Driver](https://www.nvidia.com/drivers) supporting [CUDA 9.0](https://developer.nvidia.com/cuda-downloads) or later
+  - This corresponds to 384.xx and later driver releases.
 * DALI can work with any of the following Deep Learning frameworks:
   - [MXNet](http://mxnet.incubator.apache.org)
     - Version 1.3 beta is required, `mxnet-cu90==1.3.0b20180612` or later

cmake/Dependencies.cmake

@@ -3,16 +3,32 @@
 find_package(CUDA REQUIRED)
 include_directories(${CUDA_INCLUDE_DIRS})
 list(APPEND DALI_LIBS ${CUDA_LIBRARIES})
+list(APPEND DALI_EXCLUDES libcudart_static.a)
+
+# For NVJPEG
+find_package(NVJPEG REQUIRED)
+include_directories(SYSTEM ${NVJPEG_INCLUDE_DIRS})
+list(APPEND DALI_LIBS ${NVJPEG_LIBRARY})
+list(APPEND DALI_EXCLUDES libnvjpeg_static.a)
 
 # For NPP
-find_cuda_helper_libs(nppicom)
-find_cuda_helper_libs(nppicc)
-find_cuda_helper_libs(nppc)
-find_cuda_helper_libs(nppig)
-list(APPEND DALI_LIBS ${CUDA_nppicom_LIBRARY}
-                      ${CUDA_nppicc_LIBRARY}
-                      ${CUDA_nppc_LIBRARY}
-                      ${CUDA_nppig_LIBRARY})
+find_cuda_helper_libs(nppc_static)
+find_cuda_helper_libs(nppicom_static)
+find_cuda_helper_libs(nppicc_static)
+find_cuda_helper_libs(nppig_static)
+list(APPEND DALI_LIBS ${CUDA_nppc_static_LIBRARY}
+                      ${CUDA_nppicom_static_LIBRARY}
+                      ${CUDA_nppicc_static_LIBRARY}
+                      ${CUDA_nppig_static_LIBRARY})
+list(APPEND DALI_EXCLUDES libnppc_static.a
+                          libnppicom_static.a
+                          libnppicc_static.a
+                          libnppig_static.a)
+
+# CULIBOS needed when using static CUDA libs
+find_cuda_helper_libs(culibos)
+list(APPEND DALI_LIBS ${CUDA_culibos_LIBRARY})
+list(APPEND DALI_EXCLUDES libculibos.a)
 
 # NVTX for profiling
 if (BUILD_NVTX)
@@ -21,10 +37,6 @@ if (BUILD_NVTX)
   add_definitions(-DDALI_USE_NVTX)
 endif()
 
-find_package(NVJPEG REQUIRED)
-include_directories(SYSTEM ${NVJPEG_INCLUDE_DIRS})
-list(APPEND DALI_LIBS ${NVJPEG_LIBRARY})
-
 # Google C++ testing framework
 if (BUILD_TEST)
   set(BUILD_GTEST ON CACHE INTERNAL "Builds gtest submodule")
@@ -44,6 +56,7 @@ endif()
 find_package(JpegTurbo REQUIRED)
 include_directories(SYSTEM ${JPEG_TURBO_INCLUDE_DIR})
 list(APPEND DALI_LIBS ${JPEG_TURBO_LIBRARY})
+list(APPEND DALI_EXCLUDES libturbojpeg.a)
 
 # OpenCV
 # Note: OpenCV 3.* 'imdecode()' is in the imgcodecs library
@@ -60,6 +73,7 @@ if (OpenCV_FOUND)
 endif()
 include_directories(SYSTEM ${OpenCV_INCLUDE_DIRS})
 list(APPEND DALI_LIBS ${OpenCV_LIBRARIES})
+list(APPEND DALI_EXCLUDES libopencv_core.a;libopencv_imgproc.a;libopencv_imgcodecs.a)
 
 # PyBind
 if (BUILD_PYTHON)
@@ -75,6 +89,7 @@ if (BUILD_LMDB)
     message(STATUS "Found LMDB ${LMDB_INCLUDE_DIR} : ${LMDB_LIBRARIES}")
     include_directories(SYSTEM ${LMDB_INCLUDE_DIR})
     list(APPEND DALI_LIBS ${LMDB_LIBRARIES})
+    list(APPEND DALI_EXCLUDES liblmdb.a)
   else()
     message(STATUS "LMDB not found")
   endif()
@@ -95,6 +110,10 @@ if (PROTOBUF_FOUND)
   endif()
   include_directories(SYSTEM ${PROTOBUF_INCLUDE_DIRS})
   list(APPEND DALI_LIBS ${PROTOBUF_LIBRARY})
+  ################
+  ### Don't exclude protobuf symbols; doing so will lead to segfaults
+  #list(APPEND DALI_EXCLUDES libprotobuf.a)
+  ################
 else()
   message(STATUS "Protobuf not found")
 endif()

cmake/libdali.map

@@ -0,0 +1,4 @@
+DALI_0.1.1 {
+     global:
+         *;
+};

cmake/modules/FindNVJPEG.cmake

@@ -20,7 +20,7 @@ find_path(NVJPEG_INCLUDE_DIR nvjpeg.h
     PATHS ${NVJPEG_ROOT_DIR}
     PATH_SUFFIXES include)
 
-find_library(NVJPEG_LIBRARY libnvjpeg.so nvjpeg
+find_library(NVJPEG_LIBRARY libnvjpeg_static.a nvjpeg
     PATHS ${NVJPEG_ROOT_DIR}
     PATH_SUFFIXES lib lib64)
 

dali/CMakeLists.txt

@@ -18,43 +18,57 @@ set(DALI_TEST_SRCS)
 set(DALI_BENCHMARK_SRCS)
 set(DALI_TF_SRCS)
 
-# Build the library
+################################################
+# Build libdali
+################################################
 add_subdirectory(image)
 add_subdirectory(pipeline)
 add_subdirectory(util)
 add_subdirectory(plugin)
 add_subdirectory(c_api)
 
-# Collect files in this directory
-file(GLOB tmp *.cc)
-set(DALI_SRCS ${DALI_SRCS} ${tmp})
-
 set(dali_lib "dali")
 
-if (BUILD_TENSORFLOW)
-  set(customop_lib "dali_tf")
-endif()
+# Collect source files for dali
+file(GLOB tmp *.cc)
+set(DALI_SRCS ${DALI_SRCS} ${tmp})
 
 set(DALI_PROTO_OBJ $<TARGET_OBJECTS:DALI_PROTO>)
-
 if (BUILD_LMDB)
   list(APPEND DALI_PROTO_OBJ $<TARGET_OBJECTS:CAFFE_PROTO> $<TARGET_OBJECTS:CAFFE2_PROTO>)
 endif()
-
 if (BUILD_PROTO3)
   list(APPEND DALI_PROTO_OBJ $<TARGET_OBJECTS:TF_PROTO>)
 endif()
 
-# cuda_add_library(${dali_lib} SHARED ${DALI_SRCS} $<TARGET_OBJECTS:CAFFE_PROTO> $<TARGET_OBJECTS:CAFFE2_PROTO> $<TARGET_OBJECTS:TF_PROTO> $<TARGET_OBJECTS:DALI_PROTO>)
+set(CUDA_LINK_LIBRARIES_KEYWORD PRIVATE)
 cuda_add_library(${dali_lib} SHARED ${DALI_SRCS} ${DALI_PROTO_OBJ})
-target_link_libraries(${dali_lib} ${DALI_LIBS})
 
-# Add install rules
+# Define symbol version script for libdali.so
+set(exports_map_path "${PROJECT_SOURCE_DIR}/cmake")
+set(dali_lib_exports "lib${dali_lib}.map")
+target_link_libraries(${dali_lib} PRIVATE -Wl,--version-script=${exports_map_path}/${dali_lib_exports})
+add_custom_target(${dali_lib_exports} DEPENDS "${exports_map_path}/${dali_lib_exports}")
+add_dependencies(${dali_lib} ${dali_lib_exports})
+
+# Exclude (most) statically linked dali dependencies from the exports of libdali.so
+string(REPLACE ";" ":" exclude_libs "${DALI_EXCLUDES}")
+message(STATUS "Exclude libs '${exclude_libs}'")
+target_link_libraries(${dali_lib} PRIVATE "-Wl,--exclude-libs,${exclude_libs}")
+
+# Link in dali's dependencies
+message(STATUS "Adding dependencies to ${dali_lib}: '${DALI_LIBS}'")
+target_link_libraries(${dali_lib} PRIVATE ${DALI_LIBS})
+
+# libdali installation paths
 install(TARGETS ${dali_lib} DESTINATION lib)
 install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR} DESTINATION
   include FILES_MATCHING PATTERN "*.h" PATTERN "*_test.h" EXCLUDE)
 
+
+################################################
 # Build test suite
+################################################
 if (BUILD_TEST)
   # get all test srcs
   add_subdirectory(test)
@@ -63,16 +77,20 @@ if (BUILD_TEST)
   cuda_add_executable(${test_main_bin} "${DALI_TEST_SRCS}")
 
   # Link to the dali lib
+  message(STATUS "Adding dependencies to ${test_main_bin}: '${dali_lib}'")
   add_dependencies(${test_main_bin} ${dali_lib})
 
   # We'll have to add dependency libs
-  target_link_libraries(${test_main_bin} ${DALI_LIBS} ${dali_lib} gtest)
+  target_link_libraries(${test_main_bin} PRIVATE ${DALI_LIBS} ${dali_lib} gtest)
 
   set_target_properties(${test_main_bin} PROPERTIES
     RUNTIME_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/dali/python/nvidia/dali/test)
 endif()
 
+
+################################################
 # Build benchmark suite
+################################################
 if (BUILD_BENCHMARK)
   # get benchmark main
   add_subdirectory(benchmark)
@@ -81,53 +99,50 @@ if (BUILD_BENCHMARK)
   cuda_add_executable(${benchmark_bin} "${DALI_BENCHMARK_SRCS}")
 
   # Link to the dali lib
+  message(STATUS "Adding dependencies to ${benchmark_bin}: '${dali_lib}'")
   add_dependencies(${benchmark_bin} ${dali_lib})
 
-  target_link_libraries(${benchmark_bin} ${DALI_LIBS} ${dali_lib} benchmark pthread)
+  target_link_libraries(${benchmark_bin} PRIVATE ${DALI_LIBS} ${dali_lib} benchmark pthread)
 
   set_target_properties(${benchmark_bin} PROPERTIES
     RUNTIME_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/dali/python/nvidia/dali/test)
 endif()
 
-# Build the python bindings
-if (BUILD_PYTHON)
-  # Get all python srcs
-  add_subdirectory(python)
-endif()
 
+################################################
 # Build the TensorFlow op
+################################################
 if (BUILD_TENSORFLOW)
-  # Get the TF compile et linking flags
-  #execute_process(
-  #  COMMAND python -c "from __future__ import print_function; import tensorflow as tf; print(\" \".join(tf.sysconfig.get_compile_flags()), end=\"\")"
-  #  OUTPUT_VARIABLE TF_CFLAGS)
-
-  #execute_process(
-  #  COMMAND python -c "from __future__ import print_function; import tensorflow as tf; print(\" \".join(tf.sysconfig.get_link_flags()), end=\"\")"
-  # OUTPUT_VARIABLE TF_LFLAGS)
 
   execute_process(
     COMMAND python -c "from __future__ import print_function; import sys; paths = sys.path; f = [p for p in paths if \"-packages\" in p] ; print(f[0], end=\"\")"
     OUTPUT_VARIABLE DIST_PACKAGES_DIR)
 
   add_subdirectory(tensorflow)
 
-  #set(customop_lib "dali_tf")
-  cuda_add_library(${customop_lib} SHARED ${DALI_TF_SRCS})
+  set(customop_lib "dali_tf")
 
-  # TF <= 1.3
-  # target_compile_options(${customop_lib} PRIVATE -D_GLIBCXX_USE_CXX11_ABI=0 -I${TF_INC} -I${TF_INC}/external/nsync/public -lcudart -L /usr/local/cuda/lib64 -L${TF_LIB} -ltensorflow_framework)
+  cuda_add_library(${customop_lib} SHARED ${DALI_TF_SRCS})
 
   separate_arguments(TF_CFLAGS UNIX_COMMAND "-I${DIST_PACKAGES_DIR}/tensorflow/include -D_GLIBCXX_USE_CXX11_ABI=0")
   target_compile_options(${customop_lib} PRIVATE ${TF_CFLAGS})
 
+  message(STATUS "Adding dependencies to ${customop_lib}: '${dali_lib}'")
   add_dependencies(${customop_lib} ${dali_lib})
 
-  target_link_libraries(${customop_lib} -L${DIST_PACKAGES_DIR}/tensorflow)
-  target_link_libraries(${customop_lib} -ltensorflow_framework)
-  target_link_libraries(${customop_lib} ${DALI_LIBS} ${dali_lib})
+  target_link_libraries(${customop_lib} PRIVATE -L${DIST_PACKAGES_DIR}/tensorflow -ltensorflow_framework)
+  target_link_libraries(${customop_lib} PRIVATE ${dali_lib})
 
   set_target_properties(${customop_lib} PROPERTIES
     LIBRARY_OUTPUT_DIRECTORY "${PROJECT_BINARY_DIR}/dali/python/nvidia/dali/plugin")
 endif()
 
+
+################################################
+# Build the DALI python bindings
+################################################
+if (BUILD_PYTHON)
+  # Get all python srcs
+  add_subdirectory(python)
+endif()
+