Simd sort

CMakeLists.txt

@@ -544,6 +544,33 @@ hpx_option(
   ADVANCED
 )
 
+hpx_option(
+  HPX_WITH_SIMD_SORT
+  BOOL
+  "Use FetchContent to fetch x86-simd-sort. By default an installed x86-simd-sort will be used. (default: On). It has support only for AVX-512 registers. If the architecture has no AVX-512 registers, simd-sort won't be used"
+  ON
+  CATEGORY "Build Targets"
+  ADVANCED
+)
+hpx_option(
+  HPX_WITH_SIMD_SORT_TAG STRING "x86-simd-sort repository tag or branch"
+  "1b396372e2bc10fd883d49e28dfc2a5ea5d2e194"
+  CATEGORY "Build Targets"
+  ADVANCED
+)
+
+hpx_check_for_avx512f(DEFINITIONS HPX_HAVE_AVX512F)
+
+if(NOT HPX_WITH_AVX512F)
+  set(HPX_WITH_SIMD_SORT Off)
+endif()
+
+if(HPX_WITH_SIMD_SORT)
+  include(HPX_SetupSimdSort)
+  hpx_add_config_define(HPX_HAVE_SIMD_SORT)
+  add_compile_options(-mavx512f)
+endif()
+
 # ##############################################################################
 # HPX CUDA configuration
 # ##############################################################################

cmake/FindSimdSort.cmake

@@ -0,0 +1,42 @@
+# Copyright (c)      2023 Hari Hara Naveen S
+#
+# SPDX-License-Identifier: BSL-1.0
+# Distributed under the Boost Software License, Version 1.0. (See accompanying
+# file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+if(NOT TARGET SimdSort::simdsort)
+  find_path(SIMD_SORT_INCLUDE_DIR avx512-16bit-qsort.hpp
+            HINTS "${SIMD_SORT_ROOT}" ENV SIMD_SORT_ROOT
+                  "${HPX_SIMD_SORT_ROOT}"
+  )
+
+  if(NOT SIMD_SORT_INCLUDE_DIR)
+    hpx_error("Simd Sort not found")
+  endif()
+
+  if(SIMD_SORT)
+    # The call to file is for compatibility with windows paths
+    file(TO_CMAKE_PATH ${SIMD_SORT} SIMD_SORT)
+  elseif("$ENV{SIMD_SORT}")
+    file(TO_CMAKE_PATH $ENV{SIMD_SORT} SIMD_SORT)
+  else()
+    file(TO_CMAKE_PATH "${SIMD_SORT_INCLUDE_DIR}" SIMD_SORT_INCLUDE_DIR)
+    string(REPLACE "/src" "" SIMD_SORT_ROOT "${SIMD_SORT_INCLUDE_DIR}")
+  endif()
+
+  include(FindPackageHandleStandardArgs)
+  find_package_handle_standard_args(
+    SimdSort
+    REQUIRED_VARS SIMD_SORT_INCLUDE_DIR
+    VERSION_VAR SIMD_SORT_VERSION_STRING
+  )
+
+  add_library(SimdSort::simdsort INTERFACE IMPORTED)
+  target_include_directories(
+    SimdSort::simdsort SYSTEM INTERFACE ${SIMD_SORT_INCLUDE_DIR}
+  )
+
+  mark_as_advanced(
+    SIMD_SORT_ROOT SIMD_SORT_INCLUDE_DIR SIMD_SORT_VERSION_STRING
+  )
+endif()

cmake/HPX_AddConfigTest.cmake

@@ -713,3 +713,12 @@ function(hpx_check_for_stable_inplace_merge)
     FILE ${ARGN}
   )
 endfunction()
+
+# ##############################################################################
+function(hpx_check_for_avx512f)
+  add_hpx_config_test(
+    HPX_WITH_AVX512F
+    SOURCE cmake/tests/avx512f.cpp
+    FILE ${ARGN} CXXFLAGS -mavx512f
+  )
+endfunction()

cmake/HPX_SetupSimdSort.cmake

@@ -0,0 +1,58 @@
+# Copyright (c) 2023 Hari Hara Naveen S
+#
+# SPDX-License-Identifier: BSL-1.0
+# Distributed under the Boost Software License, Version 1.0. (See accompanying
+# file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+hpx_info(
+  "HPX_WITH_FETCH_SIMD_SORT=${HPX_WITH_FETCH_SIMD_SORT}, x86-simd-sort will be fetched using CMake's FetchContent and installed alongside HPX (HPX_WITH_SIMD_SORT_TAG=${HPX_WITH_SIMD_SORT_TAG})"
+)
+
+include(FetchContent)
+fetchcontent_declare(
+  simdsort
+  GIT_REPOSITORY https://github.com/intel/x86-simd-sort
+  GIT_TAG ${HPX_WITH_SIMD_SORT_TAG}
+)
+
+fetchcontent_getproperties(simdsort)
+if(NOT simdsort_POPULATED)
+  fetchcontent_populate(simdsort)
+endif()
+set(SIMD_SORT_ROOT ${simdsort_SOURCE_DIR})
+
+add_library(simdsort INTERFACE)
+target_include_directories(
+  simdsort SYSTEM INTERFACE $<BUILD_INTERFACE:${SIMD_SORT_ROOT}/src/>
+                            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
+)
+
+install(
+  TARGETS simdsort
+  EXPORT HPXSimdSortTarget
+  COMPONENT core
+)
+
+install(
+  DIRECTORY ${SIMD_SORT_ROOT}/src/
+  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
+  COMPONENT core
+  FILES_MATCHING
+  PATTERN "*.hpp"
+  PATTERN "*.h"
+)
+
+export(
+  TARGETS simdsort
+  NAMESPACE SimdSort::
+  FILE "${CMAKE_CURRENT_BINARY_DIR}/lib/cmake/${HPX_PACKAGE_NAME}/HPXSimdSortTarget.cmake"
+)
+
+install(
+  EXPORT HPXSimdSortTarget
+  NAMESPACE SimdSort::
+  FILE HPXSimdSortTarget.cmake
+  DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/${HPX_PACKAGE_NAME}
+)
+
+add_library(SimdSort::simdsort ALIAS simdsort)

cmake/templates/HPXConfig.cmake.in

@@ -77,6 +77,8 @@ if("${HPX_WITH_DATAPAR_BACKEND}" STREQUAL "SVE")
   endif()
 endif()
 
+include("${CMAKE_CURRENT_LIST_DIR}/HPXSimdSortTarget.cmake")
+
 include("${CMAKE_CURRENT_LIST_DIR}/HPXInternalTargets.cmake")
 include("${CMAKE_CURRENT_LIST_DIR}/HPXTargets.cmake")
 

cmake/tests/avx512f.cpp

@@ -0,0 +1,7 @@
+int main()
+{
+#ifndef __AVX512F__
+    static_assert(false);
+#endif
+    return 0;
+}

libs/CMakeLists.txt

@@ -372,6 +372,10 @@ if("${HPX_WITH_DATAPAR_BACKEND}" STREQUAL "SVE")
   target_link_libraries(hpx_core PUBLIC SVE::sve)
 endif()
 
+if(HPX_WITH_SIMD_SORT)
+  target_link_libraries(hpx_core PUBLIC SimdSort::simdsort)
+endif()
+
 if(HPX_WITH_ITTNOTIFY)
   target_link_libraries(hpx_core PUBLIC Amplifier::amplifier)
 endif()

libs/core/algorithms/CMakeLists.txt

@@ -26,6 +26,7 @@ set(algorithms_headers
     hpx/parallel/algorithms/detail/dispatch.hpp
     hpx/parallel/algorithms/detail/distance.hpp
     hpx/parallel/algorithms/detail/equal.hpp
+    hpx/parallel/algorithms/detail/simd_sort.hpp
     hpx/parallel/algorithms/detail/fill.hpp
     hpx/parallel/algorithms/detail/find.hpp
     hpx/parallel/algorithms/detail/generate.hpp

libs/core/algorithms/include/hpx/parallel/algorithms/detail/simd_sort.hpp

@@ -0,0 +1,62 @@
+//  Copyright (c) 2023 Hari Hara Naveen S
+//
+//  SPDX-License-Identifier: BSL-1.0
+//  Distributed under the Boost Software License, Version 1.0. (See accompanying
+//  file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#pragma once
+
+#ifdef HPX_HAVE_SIMD_SORT
+#include <avx512-32bit-qsort.hpp>
+#include <avx512-64bit-qsort.hpp>
+#endif
+
+#include <type_traits>
+
+namespace hpx::parallel::util {
+    // TODO : add support for _Float16
+    // need compile time test as _Float16 is not always supported
+    template <class T>
+    struct is_simd_sortable
+      : std::integral_constant<bool,
+            std::is_same<uint16_t,
+                typename std::remove_volatile<T>::type>::value ||
+                std::is_same<int16_t,
+                    typename std::remove_volatile<T>::type>::value ||
+                std::is_same<uint32_t,
+                    typename std::remove_volatile<T>::type>::value ||
+                std::is_same<int32_t,
+                    typename std::remove_volatile<T>::type>::value ||
+                std::is_same<float,
+                    typename std::remove_volatile<T>::type>::value ||
+                std::is_same<uint64_t,
+                    typename std::remove_volatile<T>::type>::value ||
+                std::is_same<int64_t,
+                    typename std::remove_volatile<T>::type>::value ||
+                std::is_same<double,
+                    typename std::remove_volatile<T>::type>::value>
+    {
+    };
+
+    template <class T>
+    constexpr bool is_simd_sortable_v = is_simd_sortable<T>::value;
+}    // namespace hpx::parallel::util
+
+#ifdef HPX_HAVE_SIMD_SORT
+namespace hpx::parallel::util {
+    template <typename T>
+    void simd_quicksort(T* arr, int64_t arrsize)
+    {
+        static_assert(hpx::parallel::util::is_simd_sortable_v<T>);
+        return avx512_qsort(arr, arrsize);
+    }
+}    // namespace hpx::parallel::util
+#else
+namespace hpx::parallel::util {
+    template <typename T>
+    void simd_quicksort(T* arr, int64_t arrsize)
+    {
+        static_assert(false);
+    }
+}    // namespace hpx::parallel::util
+#endif

libs/core/algorithms/include/hpx/parallel/algorithms/sort.hpp

@@ -160,6 +160,7 @@ namespace hpx {
 #include <hpx/parallel/algorithms/detail/dispatch.hpp>
 #include <hpx/parallel/algorithms/detail/is_sorted.hpp>
 #include <hpx/parallel/algorithms/detail/pivot.hpp>
+#include <hpx/parallel/algorithms/detail/simd_sort.hpp>
 #include <hpx/parallel/util/compare_projected.hpp>
 #include <hpx/parallel/util/detail/algorithm_result.hpp>
 #include <hpx/parallel/util/detail/chunk_size.hpp>
@@ -172,6 +173,7 @@ namespace hpx {
 #include <exception>
 #include <iterator>
 #include <list>
+#include <memory>
 #include <type_traits>
 #include <utility>
 
@@ -464,6 +466,19 @@ namespace hpx {
                        HPX_MOVE(comp), hpx::identity_v);
         }
     } sort{};
+
+#ifdef HPX_HAVE_SIMD_SORT
+    template <typename RandomIt, typename Comp = hpx::parallel::detail::less,
+        HPX_CONCEPT_REQUIRES_(
+            hpx::traits::is_random_access_iterator_v<RandomIt>)>
+    void tag_invoke(hpx::sort_t, hpx::execution::unsequenced_policy,
+        RandomIt first, RandomIt last, Comp)
+    {
+        return hpx::parallel::util::simd_quicksort(
+            std::addressof(*first), std::distance(first, last));
+    }
+#endif
+
 }    // namespace hpx
 
 #endif    // DOXYGEN