reorganzize parallel.h to use parallelizm if USE_OPENMP is set, removing gnu_parallel stuff

Author: bingmann

.travis.yml

@@ -12,33 +12,31 @@ env:
 matrix:
   include:
   # gcc-5 builds
-    - env: CMAKE_CC="gcc-5" CMAKE_CXX="g++-5" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release   -DUSE_OPENMP=ON   -DUSE_GNU_PARALLEL=OFF"
+    - env: CMAKE_CC="gcc-5" CMAKE_CXX="g++-5" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DUSE_OPENMP=ON"
       addons: &gcc5
         apt:
           sources:
             - ubuntu-toolchain-r-test
           packages:
             - g++-5
-    - env: CMAKE_CC="gcc-5" CMAKE_CXX="g++-5" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release   -DUSE_OPENMP=ON   -DUSE_GNU_PARALLEL=ON"
-      addons: *gcc5
 
   # gcc-7 builds
-    - env: CMAKE_CC="gcc-7" CMAKE_CXX="g++-7" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DUSE_OPENMP=ON -DUSE_GNU_PARALLEL=OFF"
+    - env: CMAKE_CC="gcc-7" CMAKE_CXX="g++-7" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DUSE_OPENMP=ON"
       addons: &gcc7
         apt:
           sources:
             - ubuntu-toolchain-r-test
           packages:
             - g++-7
-    - env: CMAKE_CC="gcc-7" CMAKE_CXX="g++-7" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DUSE_OPENMP=ON -DUSE_GNU_PARALLEL=ON"
+    - env: CMAKE_CC="gcc-7" CMAKE_CXX="g++-7" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DUSE_OPENMP=OFF"
       addons: *gcc7
-    - env: CMAKE_CC="gcc-7" CMAKE_CXX="g++-7" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=RelWithAssert -DUSE_OPENMP=ON -DUSE_GNU_PARALLEL=OFF"
+    - env: CMAKE_CC="gcc-7" CMAKE_CXX="g++-7" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=RelWithAssert -DUSE_OPENMP=ON"
       addons: *gcc7
-    - env: CMAKE_CC="gcc-7" CMAKE_CXX="g++-7" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=RelWithAssert -DUSE_OPENMP=ON -DUSE_GNU_PARALLEL=ON"
+    - env: CMAKE_CC="gcc-7" CMAKE_CXX="g++-7" CMAKE_FLAGS=""     CMAKE_ARGS="-DCMAKE_BUILD_TYPE=RelWithAssert -DUSE_OPENMP=OFF"
       addons: *gcc7
 
   # clang build
-    - env: CMAKE_CC="clang-3.9"   CMAKE_CXX="clang++-3.9" CMAKE_FLAGS="" CMAKE_ARGS="-DCMAKE_BUILD_TYPE=RelWithAssert -DUSE_GNU_PARALLEL=OFF"
+    - env: CMAKE_CC="clang-3.9"   CMAKE_CXX="clang++-3.9" CMAKE_FLAGS="" CMAKE_ARGS="-DCMAKE_BUILD_TYPE=RelWithAssert"
       addons:
         apt:
           sources:
@@ -48,7 +46,7 @@ matrix:
             - clang-3.9
             - g++-5
             - libiomp-dev
-    - env: CMAKE_CC="clang-5.0"   CMAKE_CXX="clang++-5.0" CMAKE_FLAGS="" CMAKE_ARGS="-DCMAKE_BUILD_TYPE=RelWithAssert -DUSE_GNU_PARALLEL=OFF"
+    - env: CMAKE_CC="clang-5.0"   CMAKE_CXX="clang++-5.0" CMAKE_FLAGS="" CMAKE_ARGS="-DCMAKE_BUILD_TYPE=RelWithAssert"
       addons:
         apt:
           sources:

CMakeLists.txt

@@ -77,16 +77,6 @@ option(STXXL_RUN_SLOW_TESTS "Run slow tests of extra stxxl tests and tool progra
 
 option(USE_OPENMP "Use OpenMP for multi-core parallelism" ON)
 
-if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
-  # FIXME: autodetect, n/a with libc++
-  option(USE_GNU_PARALLEL "Use GNU parallel extensions for multi-core parallelism" ON)
-
-  # GNU parallel mode but NOT OpenMP is impossible.
-  if(USE_GNU_PARALLEL AND NOT USE_OPENMP)
-    set(USE_GNU_PARALLEL OFF)
-  endif()
-endif()
-
 option(STXXL_TRY_COMPILE_HEADERS "Test stxxl header files for self-sufficiency: try to compile them." OFF)
 
 option(USE_VALGRIND "Run tests with valgrind, pre-initialize some memory buffers." OFF)
@@ -403,12 +393,7 @@ check_cxx_source_compiles(
 ###############################################################################
 # optional OpenMP parallelization
 
-if(USE_GNU_PARALLEL AND NOT USE_OPENMP)
-  message(FATAL_ERROR "Configuration selects GNU parallel mode but not OpenMP."
-    "This is impossible.")
-endif()
-
-if(USE_OPENMP OR USE_GNU_PARALLEL)
+if(USE_OPENMP)
 
   include(FindOpenMP)
 
@@ -428,26 +413,7 @@ else()
   message(STATUS "OpenMP disabled in STXXL (no parallelism is used).")
   set(STXXL_PARALLEL 0)
 
-endif(USE_OPENMP OR USE_GNU_PARALLEL)
-
-###############################################################################
-# optional GNU parallel STL mode
-
-if(USE_GNU_PARALLEL)
-
-  if(NOT OPENMP_FOUND)
-    message(FATAL_ERROR "OpenMP not found. Try compilation without GNU parallel mode.")
-  else()
-    check_include_file_cxx(parallel/algorithm HAVE_PARALLEL_ALGORITHM_H)
-
-    if(NOT HAVE_PARALLEL_ALGORITHM_H)
-      message(FATAL_ERROR "GNU parallel mode header not found. Try compilation without parallel mode.")
-    else()
-      set(STXXL_WITH_GNU_PARALLEL 1)
-    endif()
-  endif()
-
-endif(USE_GNU_PARALLEL)
+endif(USE_OPENMP)
 
 ###############################################################################
 # test for additional includes and features used by some stxxl_tool components

doc/install.dox

@@ -208,9 +208,9 @@ $ cmake -DCMAKE_BUILD_TYPE=Release ...
 \endverbatim
 The mode mostly changes CXXFLAGS.
 
-- Some parts of STXXL have been parallelized using the __gnu_parallel (aka MCSTL) library. Currently, with CMake one can only use the newer __gnu_parallel library by defining
+- Some parts of STXXL have been parallelized using OpenMP (with code from MCSTL and the GNU parallel library)
 \verbatim
-$ cmake -DUSE_GNU_PARALLEL=ON ...
+$ cmake -DUSE_OPENMP=ON ...
 \endverbatim
 when building. <b>Parallel is now ON by default for gcc</b>, if it can be detected. The cmake script will check availability of the corresponding header files.
 

extlib/foxxll

@@ -470,7 +470,7 @@ void merge_runs(RunType** in_runs, size_t nruns,
 // end of STL-style merging
 
 #else
-        STXXL_THROW_UNREACHABLE();
+        FOXXLL_THROW_UNREACHABLE();
 #endif
     }
     else

include/stxxl/bits/algo/sort.h

@@ -69,11 +69,6 @@
 // cmake:   -DUSE_PARALLEL=ON
 // effect:  enable parallelized algorithms using OpenMP like multiway_merge and sort
 
-#cmakedefine STXXL_WITH_GNU_PARALLEL ${STXXL_WITH_GNU_PARALLEL}
-// default: off
-// cmake:   -DUSE_GNU_PARALLEL=ON
-// effect:  explicitly enables use of __gnu_parallel algorithms
-
 #cmakedefine STXXL_STD_RANDOM ${STXXL_STD_RANDOM}
 // default: off
 // cmake:   detection of <random>

include/stxxl/bits/config.h.in

@@ -71,7 +71,7 @@ class ext_merger
     using tree_type = parallel_merger_adapter<self_type, CompareType, Arity>;
 #else
     //! type of embedded loser tree
-    using tree_type = loser_tree<self_type, CompareWithMin, Arity>;
+    using tree_type = loser_tree<self_type, CompareType, Arity>;
 #endif
 
     //! size type of total number of item in merger

include/stxxl/bits/containers/pq_ext_merger.h

@@ -56,7 +56,7 @@ class int_merger
     using tree_type = parallel_merger_adapter<self_type, CompareType, MaxArity>;
 #else
     //! type of embedded loser tree
-    using tree_type = loser_tree<self_type, CompareWithMin, MaxArity>;
+    using tree_type = loser_tree<self_type, CompareType, MaxArity>;
 #endif
 
     //! type of sequences in which the values are stored: memory arrays

include/stxxl/bits/containers/pq_int_merger.h

@@ -47,62 +47,40 @@
 #define STXXL_PARALLEL_MULTIWAY_MERGE 0
 #endif
 
-#if defined(STXXL_PARALLEL_MODE) && ((__GNUC__ * 10000 + __GNUC_MINOR__ * 100) < 40400)
-#undef STXXL_PARALLEL_MULTIWAY_MERGE
-#define STXXL_PARALLEL_MULTIWAY_MERGE 0
-#endif
-
-#if !defined(STXXL_PARALLEL_MULTIWAY_MERGE)
+#if STXXL_PARALLEL && !defined(STXXL_PARALLEL_MULTIWAY_MERGE)
 #define STXXL_PARALLEL_MULTIWAY_MERGE 1
 #endif
 
 #if !defined(STXXL_NOT_CONSIDER_SORT_MEMORY_OVERHEAD)
 #define STXXL_NOT_CONSIDER_SORT_MEMORY_OVERHEAD 0
 #endif
 
-#if STXXL_WITH_GNU_PARALLEL
-#include <parallel/algorithm>
-#else
-#include <algorithm>
-#endif
-
 #include <tlx/algorithm/multiway_merge.hpp>
 #include <tlx/algorithm/parallel_multiway_merge.hpp>
 
 namespace stxxl {
 
 inline unsigned sort_memory_usage_factor()
 {
-#if STXXL_PARALLEL && !STXXL_NOT_CONSIDER_SORT_MEMORY_OVERHEAD && defined(STXXL_PARALLEL_MODE)
-    return (__gnu_parallel::_Settings::get().sort_algorithm == __gnu_parallel::MWMS && omp_get_max_threads() > 1) ? 2 : 1;   //memory overhead for multiway mergesort
+#if STXXL_PARALLEL && !STXXL_NOT_CONSIDER_SORT_MEMORY_OVERHEAD
+    return (omp_get_max_threads() > 1) ? 2 : 1;   //memory overhead for multiway mergesort
 #else
     return 1;                                                                                                                //no overhead
 #endif
 }
 
 inline void check_sort_settings()
 {
-#if STXXL_PARALLEL && defined(STXXL_PARALLEL_MODE) && !defined(STXXL_NO_WARN_OMP_NESTED)
-    static bool did_warn = false;
-    if (!did_warn) {
-        if (__gnu_parallel::_Settings::get().sort_algorithm != __gnu_parallel::MWMS) {
-            if (omp_get_max_threads() <= 2) {
-                did_warn = true;  // no problem with at most 2 threads, no need to check again
-            }
-            else if (!omp_get_nested()) {
-                LOG1 << "Inefficient settings detected. To get full potential from your CPU it is recommended to set OMP_NESTED=TRUE in the environment.";
-                did_warn = true;
-            }
-        }
-    }
+#if STXXL_PARALLEL && !defined(STXXL_NO_WARN_OMP_NESTED)
+    // nothing to check, maybe in future?
 #else
     // nothing to check
 #endif
 }
 
 inline bool do_parallel_merge()
 {
-#if STXXL_PARALLEL_MULTIWAY_MERGE && defined(STXXL_PARALLEL_MODE)
+#if STXXL_PARALLEL_MULTIWAY_MERGE
     return !stxxl::SETTINGS::native_merge && omp_get_max_threads() >= 1;
 #else
     return false;
@@ -114,23 +92,9 @@ inline bool do_parallel_merge()
 //! parallelism is optional.
 namespace potentially_parallel {
 
-#if STXXL_WITH_GNU_PARALLEL
-
-using __gnu_parallel::sort;
-using __gnu_parallel::random_shuffle;
-
-#elif STXXL_PARALLEL
-
-using std::sort;
-using std::random_shuffle;
-
-#else
-
 using std::sort;
 using std::random_shuffle;
 
-#endif
-
 /*! Multi-way merging dispatcher.
  * \param seqs_begin Begin iterator of iterator pair input sequence.
  * \param seqs_end End iterator of iterator pair input sequence.

include/stxxl/bits/parallel.h

@@ -1113,7 +1113,7 @@ class basic_runs_merger
 
 // end of STL-style merging
 #else
-            STXXL_THROW_UNREACHABLE();
+            FOXXLL_THROW_UNREACHABLE();
 #endif //STXXL_PARALLEL_MULTIWAY_MERGE
         }
         else
@@ -1296,7 +1296,7 @@ class basic_runs_merger
             }
 // end of STL-style merging
 #else
-            STXXL_THROW_UNREACHABLE();
+            FOXXLL_THROW_UNREACHABLE();
 #endif //STXXL_PARALLEL_MULTIWAY_MERGE
         }
         else