Fix warnings, build adjustments in CMake (#209)

* Fix warnings, build adjustments in CMake

* fix cuda variable-declaration order problems

* Remove gpu flavors below 30, no longer compatible

.gitignore

@@ -19,3 +19,4 @@ rabit
 dmlc-core
 *.db
 *.bak
+build

cmake/Cuda.cmake

@@ -5,16 +5,6 @@ endif()
 include(CheckCXXCompilerFlag)
 check_cxx_compiler_flag("-std=c++11"   SUPPORT_CXX11)
 
-
-# Known NVIDIA GPU achitectures mshadow can be compiled for.
-# This list will be used for CUDA_ARCH_NAME = All option
-if(${CUDA_VERSION} GREATER 7.5)
-  set(mshadow_known_gpu_archs "20 21(20) 30 35 50 52 60 61")
-else()
-  set(mshadow_known_gpu_archs "20 21(20) 30 35 50 52")
-endif()
-
-
 ################################################################################################
 # A function for automatic detection of GPUs installed  (if autodetection is enabled)
 # Usage:
@@ -167,17 +157,14 @@ macro(mshadow_cuda_compile objlist_variable)
   endforeach()
   if(UNIX OR APPLE)
     list(APPEND CUDA_NVCC_FLAGS -Xcompiler -fPIC)
-    if(SUPPORT_CXX11)
-      list(APPEND CUDA_NVCC_FLAGS -Xcompiler -fPIC --std=c++11)
-    endif()
   endif()
 
   if(APPLE)
     list(APPEND CUDA_NVCC_FLAGS -Xcompiler -Wno-unused-function)
   endif()
-  
+
   set(CUDA_NVCC_FLAGS_DEBUG "${CUDA_NVCC_FLAGS_DEBUG} -G -lineinfo")
-  
+
   if(MSVC)
     # disable noisy warnings:
     # 4819: The file contains a character that cannot be represented in the current code page (number).
@@ -191,6 +178,14 @@ macro(mshadow_cuda_compile objlist_variable)
     endforeach(flag_var)
   endif()
 
+  # If the build system is a container, make sure the nvcc intermediate files
+  # go into the build output area rather than in /tmp, which may run out of space
+  if(IS_CONTAINER_BUILD)
+    set(CUDA_NVCC_INTERMEDIATE_DIR "${CMAKE_CURRENT_BINARY_DIR}")
+    message(STATUS "Container build enabled, so nvcc intermediate files in: ${CUDA_NVCC_INTERMEDIATE_DIR}")
+    list(APPEND CUDA_NVCC_FLAGS "--keep --keep-dir ${CUDA_NVCC_INTERMEDIATE_DIR}")
+  endif()
+
   cuda_compile(cuda_objcs ${ARGN})
 
   foreach(var CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_RELEASE CMAKE_CXX_FLAGS_DEBUG)
@@ -232,7 +227,16 @@ endfunction()
 ###  Non macro section
 ################################################################################################
 
-find_package(CUDA 5.5 QUIET)
+# Try to prime CUDA_TOOLKIT_ROOT_DIR by looking for libcudart.so
+if(NOT CUDA_TOOLKIT_ROOT_DIR)
+  find_library(CUDA_LIBRARY_PATH libcudart.so PATHS ENV LD_LIBRARY_PATH PATH_SUFFIXES lib lib64)
+  if(CUDA_LIBRARY_PATH)
+    get_filename_component(CUDA_LIBRARY_PATH ${CUDA_LIBRARY_PATH} DIRECTORY)
+    set(CUDA_TOOLKIT_ROOT_DIR "${CUDA_LIBRARY_PATH}/..")
+  endif()
+endif()
+
+find_package(CUDA 5.5 QUIET REQUIRED)
 find_cuda_helper_libs(curand)  # cmake 2.8.7 compartibility which doesn't search for curand
 
 if(NOT CUDA_FOUND)
@@ -245,6 +249,18 @@ include_directories(SYSTEM ${CUDA_INCLUDE_DIRS})
 list(APPEND mshadow_LINKER_LIBS ${CUDA_CUDART_LIBRARY}
                               ${CUDA_curand_LIBRARY} ${CUDA_CUBLAS_LIBRARIES})
 
+# Known NVIDIA GPU achitectures mshadow can be compiled for.
+# This list will be used for CUDA_ARCH_NAME = All option
+if(CUDA_ARCH_ALL)
+  set(mshadow_known_gpu_archs "${CUDA_ARCH_ALL}")
+else()
+  if(${CUDA_VERSION} GREATER 7.5)
+    set(mshadow_known_gpu_archs "30 35 50 52 60 61")
+  else()
+    set(mshadow_known_gpu_archs "30 35 50 52")
+  endif()
+endif()
+
 # cudnn detection
 if(USE_CUDNN)
   detect_cuDNN()

cmake/mshadowUtils.cmake

@@ -0,0 +1,2 @@
+include("${CMAKE_CURRENT_LIST_DIR}/Utils.cmake")
+

mshadow/extension/broadcast.h

@@ -92,7 +92,7 @@ broadcast_scalar(const expr::Exp<SrcExp, DType, etype> &src, Shape<dimdst> shape
   TypeCheckPass<ExpInfo<SrcExp>::kDim == 1>
                 ::Error_Expression_Does_Not_Meet_Dimension_Req();
   typedef ShapeCheck<1, SrcExp> ShapeCheckDim1SrcExp;
-  CHECK_EQ(ShapeCheckDim1SrcExp::Check(src.self())[0], 1)
+  CHECK_EQ(ShapeCheckDim1SrcExp::Check(src.self())[0], 1U)
       << "broadcast_scalar, source need to be scalar expression";
   return BroadcastScalarExp<SrcExp, DType, dimdst>(src.self(), shape);
 }

mshadow/extension/broadcast_with_axis.h

@@ -193,13 +193,13 @@ const TShape &axes, const TShape &sizes) {
 template<typename SrcExp, typename DType, int etype, typename TShape>
 inline BroadcastWithMultiAxesExp<SrcExp, DType, ExpInfo<SrcExp>::kDim>
 broadcast_to(const Exp<SrcExp, DType, etype> &src, const TShape &target_shape) {
-  static const int dimsrc = ExpInfo<SrcExp>::kDim;
+  static const size_t dimsrc = ExpInfo<SrcExp>::kDim;
   CHECK_EQ(target_shape.ndim(), dimsrc);
   std::vector<index_t> axes_vec, sizes_vec;
   Shape<dimsrc> src_shape = ShapeCheck<dimsrc, SrcExp>::Check(src.self());
-  for (int i = 0; i < dimsrc; ++i) {
+  for (size_t i = 0; i < dimsrc; ++i) {
     if (src_shape[i] != target_shape[i]) {
-      CHECK_EQ(src_shape[i], 1) << "broadcasting axis must have size 1, received shape="
+      CHECK_EQ(src_shape[i], 1U) << "broadcasting axis must have size 1, received shape="
         << src_shape << " target_shape=" << target_shape;
       axes_vec.push_back(i);
       sizes_vec.push_back(target_shape[i]);

mshadow/tensor_cpu-inl.h

@@ -475,7 +475,7 @@ inline void VectorDot(Tensor<Device, 1, DType> dst,
                       const Tensor<Device, 1, DType> &rhs) {
   CHECK_EQ(lhs.size(0), rhs.size(0))
       << "VectorDot: Shape mismatch";
-  CHECK_EQ(dst.size(0), 1)
+  CHECK_EQ(dst.size(0), 1U)
       << "VectorDot: expect dst to be scalar";
   expr::BLASEngine<Device, DType>::SetStream(lhs.stream_);
   mshadow::expr::BLASEngine<Device, DType>::dot(