Merge pull request #153 from sterrettm2/kvsort-openmp

Adds OpenMP based parallelization to key-value sorting

.github/workflows/c-cpp.yml

@@ -135,6 +135,38 @@ jobs:
     - name: Run test suite on SPR
       run: sde -spr -- ./builddir/testexe
 
+  SKX-SKL-openmp:
+
+    runs-on: intel-ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # v4.1.1
+
+    - name: Install dependencies
+      run: |
+        sudo apt update
+        sudo apt -y install g++-10 libgtest-dev meson curl git
+
+    - name: Install Intel SDE
+      run: |
+        curl -o /tmp/sde.tar.xz https://downloadmirror.intel.com/784319/sde-external-9.24.0-2023-07-13-lin.tar.xz
+        mkdir /tmp/sde && tar -xvf /tmp/sde.tar.xz -C /tmp/sde/
+        sudo mv /tmp/sde/* /opt/sde && sudo ln -s /opt/sde/sde64 /usr/bin/sde
+
+    - name: Build
+      env:
+        CXX: g++-10
+      run: |
+        make clean
+        meson setup -Dbuild_tests=true -Duse_openmp=true --warnlevel 2 --werror --buildtype release builddir
+        cd builddir
+        ninja
+
+    - name: Run test suite on SKX and SKL
+      run: |
+        sde -skx -- ./builddir/testexe
+        sde -skl -- ./builddir/testexe
+
   SPR-gcc13-special-cases:
 
     runs-on: intel-ubuntu-latest

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

@@ -11,6 +11,11 @@
 #include "xss-common-qsort.h"
 #include "xss-network-keyvaluesort.hpp"
 
+#if defined(XSS_USE_OPENMP) && defined(_OPENMP)
+#define XSS_COMPILE_OPENMP
+#include <omp.h>
+#endif
+
 /*
  * Parition one ZMM register based on the pivot and returns the index of the
  * last element that is less than equal to the pivot.
@@ -366,7 +371,8 @@ 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 +397,61 @@ 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);
+
+#ifdef XSS_COMPILE_OPENMP
+    if (pivot != smallest) {
+        bool parallel_left = (pivot_index - left) > task_threshold;
+        if (parallel_left) {
+#pragma omp task
+            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 parallel_right = (right - pivot_index) > task_threshold;
+
+        if (parallel_right) {
+#pragma omp task
+            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 +539,33 @@ X86_SIMD_SORT_INLINE void xss_qsort_kv(
             UNUSED(hasnan);
         }
 
-        kvsort_<keytype, valtype>(keys, indexes, 0, arrsize - 1, maxiters);
+#ifdef XSS_COMPILE_OPENMP
+
+        bool use_parallel = arrsize > 10000;
+
+        if (use_parallel) {
+            // This thread limit was determined experimentally; it may be better for it to be the number of physical cores on the system
+            constexpr int thread_limit = 8;
+            int thread_count = std::min(thread_limit, omp_get_max_threads());
+            arrsize_t task_threshold
+                    = std::max((arrsize_t)10000, arrsize / 100);
+
+            // We use omp parallel and then omp single to setup the threads that will run the omp task calls in kvsort_
+            // The omp single prevents multiple threads from running the initial kvsort_ simultaneously and causing problems
+            // Note that we do not use the if(...) clause built into OpenMP, because it causes a performance regression for small arrays
+#pragma omp parallel num_threads(thread_count)
+#pragma omp single
+            kvsort_<keytype, valtype>(
+                    keys, indexes, 0, arrsize - 1, maxiters, task_threshold);
+        }
+        else {
+            kvsort_<keytype, valtype>(
+                    keys, indexes, 0, arrsize - 1, maxiters, 0);
+        }
+#else
+        kvsort_<keytype, valtype>(keys, indexes, 0, arrsize - 1, maxiters, 0);
+#endif
+
         replace_inf_with_nan(keys, arrsize, nan_count);
 
         if (descending) {