Adds OpenMP supports for kv-sort

Author: sterrettm2

Makefile

@@ -1,5 +1,9 @@
 test:
-	meson setup -Dbuild_tests=true --warnlevel 2 --werror --buildtype release builddir
+	meson setup -Dbuild_tests=true -Duse_openmp=false --warnlevel 2 --werror --buildtype release builddir
+	cd builddir && ninja
+
+test_openmp:
+	meson setup -Dbuild_tests=true -Duse_openmp=true --warnlevel 2 --werror --buildtype release builddir
 	cd builddir && ninja
 
 bench:

lib/meson.build

@@ -1,12 +1,18 @@
 libtargets = []
 
+# Add compile flags for OpenMP if enabled
+openmpflags = []
+if get_option('use_openmp')
+  openmpflags = ['-DXSS_USE_OPENMP=true', '-fopenmp']
+endif
+
 if cpp.has_argument('-march=haswell')
   libtargets += static_library('libavx',
     files(
       'x86simdsort-avx2.cpp',
       ),
     include_directories : [src],
-    cpp_args : ['-march=haswell'],
+    cpp_args : ['-march=haswell', openmpflags],
     gnu_symbol_visibility : 'inlineshidden',
     )
 endif
@@ -17,7 +23,7 @@ if cpp.has_argument('-march=skylake-avx512')
       'x86simdsort-skx.cpp',
       ),
     include_directories : [src],
-    cpp_args : ['-march=skylake-avx512'],
+    cpp_args : ['-march=skylake-avx512', openmpflags],
     gnu_symbol_visibility : 'inlineshidden',
     )
 endif

meson.build

@@ -37,6 +37,7 @@ subdir('lib')
 libsimdsort = shared_library('x86simdsortcpp',
                              'lib/x86simdsort.cpp',
                              include_directories : [src, utils, lib],
+                             link_args : [openmpflags],
                              link_with : [libtargets],
                              gnu_symbol_visibility : 'inlineshidden',
                              install : true,

meson_options.txt

@@ -4,5 +4,7 @@ option('build_benchmarks', type : 'boolean', value : false,
   description : 'Build benchmarking suite (default: "false").')
 option('build_ippbench', type : 'boolean', value : false,
   description : 'Add IPP sort to benchmarks (default: "false").')
-option('build_vqsortbench', type : 'boolean', value : false,
-  description : 'Add google vqsort to benchmarks (default: "false").')
+option('build_vqsortbench', type : 'boolean', value : true,
+  description : 'Add google vqsort to benchmarks (default: "true").')
+option('use_openmp', type : 'boolean', value : false,
+  description : 'Use OpenMP to accelerate key-value sort (default: "false").')

scripts/branch-compare.sh

@@ -27,7 +27,7 @@ build_branch() {
         fi
     fi
     cd $dir_name
-    meson setup -Dbuild_benchmarks=true --warnlevel 0 --buildtype release builddir
+    meson setup -Dbuild_benchmarks=true -Duse_openmp=true --warnlevel 0 --buildtype release builddir
     cd builddir
     ninja
     cd ../../

src/xss-common-keyvaluesort.hpp

@@ -366,7 +366,8 @@ X86_SIMD_SORT_INLINE void kvsort_(type1_t *keys,
                                   type2_t *indexes,
                                   arrsize_t left,
                                   arrsize_t right,
-                                  int max_iters)
+                                  int max_iters,
+                                  arrsize_t task_threshold)
 {
     /*
      * Resort to std::sort if quicksort isnt making any progress
@@ -391,14 +392,44 @@ X86_SIMD_SORT_INLINE void kvsort_(type1_t *keys,
     type1_t biggest = vtype1::type_min();
     arrsize_t pivot_index = kvpartition_unrolled<vtype1, vtype2, 4>(
             keys, indexes, left, right + 1, pivot, &smallest, &biggest);
+
+
+#if defined(XSS_USE_OPENMP) && defined(_OPENMP)
+    if (pivot != smallest) {
+        bool parallelLeft = (pivot_index - left) > task_threshold;
+        if (parallelLeft){
+            #pragma omp task if(parallelLeft)
+            kvsort_<vtype1, vtype2>(
+                keys, indexes, left, pivot_index - 1, max_iters - 1, task_threshold);
+        }else{
+            kvsort_<vtype1, vtype2>(
+                keys, indexes, left, pivot_index - 1, max_iters - 1, task_threshold);
+        }
+    }
+    if (pivot != biggest) {
+        bool parallelRight = (right - pivot_index) > task_threshold;
+        
+        if (parallelRight){
+            #pragma omp task if(parallelRight)
+            kvsort_<vtype1, vtype2>(
+                keys, indexes, pivot_index, right, max_iters - 1, task_threshold);
+        }else{
+            kvsort_<vtype1, vtype2>(
+                keys, indexes, pivot_index, right, max_iters - 1, task_threshold); 
+        }
+    }
+#else
+    UNUSED(task_threshold);
+
     if (pivot != smallest) {
         kvsort_<vtype1, vtype2>(
-                keys, indexes, left, pivot_index - 1, max_iters - 1);
+                keys, indexes, left, pivot_index - 1, max_iters - 1, 0);
     }
     if (pivot != biggest) {
         kvsort_<vtype1, vtype2>(
-                keys, indexes, pivot_index, right, max_iters - 1);
+                keys, indexes, pivot_index, right, max_iters - 1, 0);
     }
+#endif
 }
 
 template <typename vtype1,
@@ -486,7 +517,20 @@ X86_SIMD_SORT_INLINE void xss_qsort_kv(
             UNUSED(hasnan);
         }
 
-        kvsort_<keytype, valtype>(keys, indexes, 0, arrsize - 1, maxiters);
+#if defined(XSS_USE_OPENMP) && defined(_OPENMP)
+        bool useParallel = arrsize > 10000;
+        arrsize_t taskThreshold = std::max((arrsize_t) 10000, arrsize / 100);
+        if (useParallel){
+            #pragma omp parallel
+            #pragma omp single
+            kvsort_<keytype, valtype>(keys, indexes, 0, arrsize - 1, maxiters, taskThreshold);
+        }else{
+            kvsort_<keytype, valtype>(keys, indexes, 0, arrsize - 1, maxiters, taskThreshold);
+        }
+#else
+        kvsort_<keytype, valtype>(keys, indexes, 0, arrsize - 1, maxiters, 0);
+#endif    
+        
         replace_inf_with_nan(keys, arrsize, nan_count);
 
         if (descending) {