Merge 7da04ab into c435525

STEllAR-GROUP · Aug 29, 2023 · 0c93a3b · 0c93a3b
2 parents c435525 + 7da04ab
commit 0c93a3b
Show file tree

Hide file tree

Showing 8 changed files with 215 additions and 0 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -544,6 +544,37 @@ hpx_option(
   ADVANCED
 )
 
+hpx_option(
+  HPX_WITH_FETCH_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" "v2.0"
+  CATEGORY "Build Targets"
+  ADVANCED
+)
+
+# simd sort supports only avx-512 registers
+if(__AVX512F__)
+  set(HPX_WITH_SIMD_SORT ON)
+  if(HPX_WITH_FETCH_SIMD_SORT)
+    include(HPX_SetupSimdSort)
+  else()
+    find_package(SimdSort)
+  endif()
+else()
+  set(HPX_WITH_SIMD_SORT OFF)
+  set(HPX_WITH_FETCH_SIMD_SORT OFF)
+endif()
+
+# debug message to be removed before merge
+message(WARNING "HPX_WITH_SIMD_SORT : ${HPX_WITH_SIMD_SORT}")
+message(WARNING "HPX_WITH_FETCH_SIMD_SORT : ${HPX_WITH_FETCH_SIMD_SORT}")
+
 # ##############################################################################
 # HPX CUDA configuration
 # ##############################################################################

diff --git a/cmake/FindSimdSort.cmake b/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 ${SIMD_SORT_INCLUDE_DIR}
+  )
+
+  mark_as_advanced(
+    SIMD_SORT_ROOT SIMD_SORT_INCLUDE_DIR SIMD_SORT_VERSION_STRING
+  )
+endif()
diff --git a/cmake/HPX_SetupSimdSort.cmake b/cmake/HPX_SetupSimdSort.cmake
@@ -0,0 +1,67 @@
+# Copyright (c) 2023 Hari Hara Naveen S
+#
+# SPDX-License-Identifier: BSL-1.0
+# Distributed under the Boost Software License, Version 1.0. (See accompanying
+# file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+if(HPX_WITH_FETCH_SIMD_SORT)
+  if(FETCHCONTENT_SOURCE_DIR_SIMD_SORT)
+    hpx_info(
+      "HPX_WITH_FETCH_SIMD_SORT=${HPX_WITH_FETCH_SIMD_SORT}, x86-simd-sort will be used through CMake's FetchContent and installed alongside HPX (FETCHCONTENT_SOURCE_DIR_SIMD_SORT=${FETCHCONTENT_SOURCE_DIR_SIMD_SORT})"
+    )
+  else()
+    hpx_info(
+      "HPX_WITH_FETCH_SIMD_SORT=${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})"
+    )
+  endif()
+
+  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)
+
+endif()
diff --git a/cmake/templates/HPXConfig.cmake.in b/cmake/templates/HPXConfig.cmake.in
@@ -72,6 +72,8 @@ 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")
 

diff --git a/libs/CMakeLists.txt b/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()

diff --git a/libs/core/algorithms/CMakeLists.txt b/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
@@ -41,6 +42,7 @@ set(algorithms_headers
     hpx/parallel/algorithms/detail/sample_sort.hpp
     hpx/parallel/algorithms/detail/search.hpp
     hpx/parallel/algorithms/detail/set_operation.hpp
+    hpx/parallel/algorithms/detail/simd_sort.hpp
     hpx/parallel/algorithms/detail/spin_sort.hpp
     hpx/parallel/algorithms/detail/transfer.hpp
     hpx/parallel/algorithms/detail/upper_lower_bound.hpp

diff --git a/libs/core/algorithms/include/hpx/parallel/algorithms/detail/simd_sort.hpp b/libs/core/algorithms/include/hpx/parallel/algorithms/detail/simd_sort.hpp
@@ -0,0 +1,50 @@
+//  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)
+
+#include <type_traits>
+
+#pragma once
+
+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
+
+#if (HPX_WITH_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
+#endif
diff --git a/libs/core/algorithms/include/hpx/parallel/algorithms/sort.hpp b/libs/core/algorithms/include/hpx/parallel/algorithms/sort.hpp
@@ -159,6 +159,7 @@ namespace hpx {
 #include <hpx/parallel/algorithms/detail/advance_to_sentinel.hpp>
 #include <hpx/parallel/algorithms/detail/dispatch.hpp>
 #include <hpx/parallel/algorithms/detail/is_sorted.hpp>
+#include <hpx/parallel/algorithms/detail/simd_sort.hpp>
 #include <hpx/parallel/algorithms/detail/pivot.hpp>
 #include <hpx/parallel/util/compare_projected.hpp>
 #include <hpx/parallel/util/detail/algorithm_result.hpp>
@@ -172,6 +173,7 @@ namespace hpx {
 #include <exception>
 #include <iterator>
 #include <list>
+#include <memory>
 #include <type_traits>
 #include <utility>
 
@@ -464,6 +466,21 @@ namespace hpx {
                        HPX_MOVE(comp), hpx::identity_v);
         }
     } sort{};
+
+#if (HPX_WITH_SIMD_SORT)
+    template <typename RandomIt, typename Comp = hpx::parallel::detail::less,
+        HPX_CONCEPT_REQUIRES_(hpx::traits::is_random_iterator_v<RandomIt>&&
+                hpx::is_invocable_v<Comp, hpx::traits::iter_value_t<RandomIt>,
+                    hpx::traits::iter_value_t<RandomIt>>)>
+    parallel::util::detail::algorithm_result_t<ExPolicy> tag_invoke(hpx::sort_t,
+        hpx::execution::unsequenced_policy, RandomIt first, RandomIt last,
+        Comp comp = Comp())
+    {
+        return hpx::parallel::util::simd_quicksort(
+            std::addressof(*first), std::distance(first, last));
+    }
+#endif
+
 }    // namespace hpx
 
 #endif    // DOXYGEN