bug #1572: use c++11 atomic instead of volatile if c++11 is available, and disable multi-threaded GEMM on non-x86 without c++11.

CMakeLists.txt

@@ -8,6 +8,7 @@ if(${CMAKE_SOURCE_DIR} STREQUAL ${CMAKE_BINARY_DIR})
   message(FATAL_ERROR "In-source builds not allowed. Please make a new directory (called a build directory) and run CMake from there. You may need to remove CMakeCache.txt. ")
 endif()
 
+
 # Alias Eigen_*_DIR to Eigen3_*_DIR:
 
 set(Eigen_SOURCE_DIR ${Eigen3_SOURCE_DIR})
@@ -107,7 +108,7 @@ if(NOT WIN32 OR NOT CMAKE_HOST_SYSTEM_NAME MATCHES Windows)
   option(EIGEN_BUILD_PKGCONFIG "Build pkg-config .pc file for Eigen" ON)
 endif()
 
-set(CMAKE_INCLUDE_CURRENT_DIR ON)
+set(CMAKE_INCLUDE_CURRENT_DIR OFF)
 
 option(EIGEN_SPLIT_LARGE_TESTS "Split large tests into smaller executables" ON)
 
@@ -377,7 +378,7 @@ option(EIGEN_TEST_CXX11 "Enable testing with C++11 and C++11 features (e.g. Tens
 
 set(EIGEN_CUDA_COMPUTE_ARCH 30 CACHE STRING "The CUDA compute architecture level to target when compiling CUDA code")
 
-include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 # Backward compatibility support for EIGEN_INCLUDE_INSTALL_DIR
 if(EIGEN_INCLUDE_INSTALL_DIR)

Eigen/src/Core/products/GeneralMatrixMatrix.h

@@ -108,7 +108,7 @@ static void run(Index rows, Index cols, Index depth,
       // i.e., we test that info[tid].users equals 0.
       // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it.
       while(info[tid].users!=0) {}
-      info[tid].users += threads;
+      info[tid].users = threads;
 
       pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length);
 
@@ -146,7 +146,9 @@ static void run(Index rows, Index cols, Index depth,
       // Release all the sub blocks A'_i of A' for the current thread,
       // i.e., we simply decrement the number of users by 1
       for(Index i=0; i<threads; ++i)
+#if !EIGEN_HAS_CXX11_ATOMIC
         #pragma omp atomic
+#endif
         info[i].users -= 1;
     }
   }

Eigen/src/Core/products/Parallelizer.h

@@ -10,6 +10,10 @@
 #ifndef EIGEN_PARALLELIZER_H
 #define EIGEN_PARALLELIZER_H
 
+#if EIGEN_HAS_CXX11_ATOMIC
+#include <atomic>
+#endif
+
 namespace Eigen {
 
 namespace internal {
@@ -75,8 +79,17 @@ template<typename Index> struct GemmParallelInfo
 {
   GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {}
 
+  // volatile is not enough on all architectures (see bug 1572)
+  // to guarantee that when thread A says to thread B that it is
+  // done with packing a block, then all writes have been really
+  // carried out... C++11 memory model+atomic guarantees this.
+#if EIGEN_HAS_CXX11_ATOMIC
+  std::atomic<Index> sync;
+  std::atomic<int> users;
+#else
   Index volatile sync;
   int volatile users;
+#endif
 
   Index lhs_start;
   Index lhs_length;
@@ -87,7 +100,10 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth,
 {
   // TODO when EIGEN_USE_BLAS is defined,
   // we should still enable OMP for other scalar types
-#if !(defined (EIGEN_HAS_OPENMP)) || defined (EIGEN_USE_BLAS)
+  // Without C++11, we have to disable GEMM's parallelization on
+  // non x86 architectures because there volatile is not enough for our purpose.
+  // See bug 1572.
+#if (! defined(EIGEN_HAS_OPENMP)) || defined(EIGEN_USE_BLAS) || ((!EIGEN_HAS_CXX11_ATOMIC) && !(EIGEN_ARCH_i386_OR_x86_64))
   // FIXME the transpose variable is only needed to properly split
   // the matrix product when multithreading is enabled. This is a temporary
   // fix to support row-major destination matrices. This whole

Eigen/src/Core/util/Macros.h

@@ -537,7 +537,7 @@
 #endif
 #endif
 
-// Does the compiler support type_trais?
+// Does the compiler support type_traits?
 #ifndef EIGEN_HAS_TYPE_TRAITS
 #if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700)
 #define EIGEN_HAS_TYPE_TRAITS 1
@@ -617,6 +617,16 @@
   #endif
 #endif
 
+#ifndef EIGEN_HAS_CXX11_ATOMIC
+  #if    EIGEN_MAX_CPP_VER>=11 && \
+         (__has_feature(cxx_atomic) \
+      || (__cplusplus > 201103L) \
+      || ((__cplusplus >= 201103L) && (EIGEN_COMP_MSVC==0 || EIGEN_COMP_MSVC >= 1700)))
+    #define EIGEN_HAS_CXX11_ATOMIC 1
+  #else
+    #define EIGEN_HAS_CXX11_ATOMIC 0
+  #endif
+#endif
 
 #if defined(EIGEN_CUDACC) && EIGEN_HAS_CONSTEXPR
   // While available already with c++11, this is useful mostly starting with c++14 and relaxed constexpr rules

test/CMakeLists.txt

@@ -387,7 +387,6 @@ if(CUDA_FOUND)
   if(EIGEN_TEST_CUDA_CLANG)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 --cuda-gpu-arch=sm_30")
   endif()
-  cuda_include_directories(${CMAKE_CURRENT_BINARY_DIR})
   set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")
   
   ei_add_test(gpu_basic)
@@ -416,7 +415,6 @@ if (EIGEN_TEST_HIP)
 
       if (${HIP_PLATFORM} STREQUAL "hcc")
 
-	include_directories(${CMAKE_CURRENT_BINARY_DIR})
 	include_directories(${HIP_PATH}/include)
 
 	set(EIGEN_ADD_TEST_FILENAME_EXTENSION  "cu")