Simd(avx512) tests and binary_search tests (frozenca#9)

* prepare simd improvements * fix compile options for release build add cstdint into details.h * use ptrdiff_t instead of in for pointer subst * add intel sde for avx512f testing * fix github action * apt needs sudo * add clang build * fix indentation * use clang-15 for ubuntu * try to fix clang usage * try to use libc++ for clang * we cant use clang on ubuntu without gcc * ok revert libc++ usage * add macos * syntax error * use macos-13 * try to use macos-14 * revert macos to 13 and wrap header memory_resource * wrap std::pmr and std::experimental::pmr * fix error from clang-14 * add experimental link option for clang less or equal 14 --------- Co-authored-by: Alexander B <[email protected]>
alicia-lyu · Jan 15, 2024 · 82b4dd9 · 82b4dd9
1 parent c594ca6
commit 82b4dd9
Show file tree

Hide file tree

Showing 6 changed files with 79 additions and 28 deletions.
diff --git a/.github/workflows/action-cpp.yml b/.github/workflows/action-cpp.yml
@@ -76,3 +76,33 @@ jobs:
         with:
           name: ${{ format('BTree-{0}.{1}', steps.get-version.outputs.prj_ver, steps.container.outputs.value) }}
           path: cmake-build-release/${{ format('BTree-{0}-noarch.???', steps.get-version.outputs.prj_ver) }}
+
+  avx512f-gcc-cmake:
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v3
+      - run: sudo apt-get update -y && sudo apt-get install -yq binutils git make cmake catch2 gcc g++ lsb-release
+      - name: mkdir
+        run: mkdir cmake-build-release
+      - name: cmake cmake-build-release
+        run: cmake -DFORCE_USE_SIMD=ON -DCMAKE_BUILD_TYPE=Release -Bcmake-build-release -H.
+      - name: cmake make
+        run: cmake --build cmake-build-release/ --target all --parallel
+      - name: sde-test
+        uses: petarpetrovt/[email protected]
+      - name: test
+        run: cp cmake-build-release/fc_tests ${SDE_PATH} && cd ${SDE_PATH} && ./sde64 -knl -- ./fc_tests --benchmark-samples=1
+
+  macos-clang-cmake:
+    runs-on: macos-13
+    steps:
+      - uses: actions/checkout@v3
+      - run: brew install git ninja cmake catch2
+      - name: mkdir
+        run: mkdir cmake-build-release
+      - name: cmake cmake-build-release
+        run: cmake -DFORCE_PREFER_BINARY_SEARCH=ON -DCMAKE_BUILD_TYPE=Release -Bcmake-build-release -H.
+      - name: cmake make
+        run: cmake --build cmake-build-release/ --target all --parallel
+      - name: test
+        run: cd cmake-build-release && ctest && cd ..
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -12,6 +12,7 @@ option(FORCE_USE_SIMD "force define FC_USE_SIMD (apllicable only for x86_64)" OF
 option(FORCE_PREFER_BINARY_SEARCH "force define FC_PREFER_BINARY_SEARCH (recommended for clang only)" OFF)
 
 set(CMAKE_CXX_STANDARD 20)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 find_package(VersionHeader)
 set(PROJECT_VERSION "${MAJOR}.${MINOR}.${PATCH}.${COMMITTER_SHORTSHA}")
@@ -103,11 +104,24 @@ target_link_libraries(fc_tests PRIVATE ${CATCH_LIBS_ALIASES} BTree::BTree)
 
 if(MSVC)
     target_compile_options(fc_tests PRIVATE /W4 /WX /nologo /MDd /EHsc /std:c++latest /experimental:module)
-    if (CMAKE_BUILD_TYPE STREQUAL "Release")
-        add_compile_options(/Ox)
-    endif()
+    set(CMAKE_C_FLAGS_RELEASE "/Ox")
+    set(CMAKE_CXX_FLAGS_RELEASE "/Ox")
+    set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "/Ox")
 else()
-    target_compile_options(fc_tests PRIVATE -Wall -Wextra -Wpedantic -Werror -march=native)
+    set(CMAKE_C_FLAGS_RELEASE "-O3")
+    set(CMAKE_CXX_FLAGS_RELEASE "-O3")
+    set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3")
+    if (FORCE_USE_SIMD)
+        target_compile_options(fc_tests PRIVATE -Wall -Wextra -Wpedantic -Werror -mavx512f)
+    else ()
+        target_compile_options(fc_tests PRIVATE -Wall -Wextra -Wpedantic -Werror -march=native -mtune=generic)
+    endif ()
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS_EQUAL 14.0.3.14030022)
+        target_link_options(fc_tests PRIVATE -fexperimental-library)
+    endif()
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "Clang" AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS_EQUAL 14)
+        target_link_options(fc_tests PRIVATE -fexperimental-library)
+    endif()
 endif()
 
 include(CTest)

diff --git a/include/fc/comp.h b/include/fc/comp.h
@@ -22,54 +22,54 @@ using regi = __m512i;
 using regf = __m512;
 using regd = __m512d;
 
-regi broadcast(std::int32_t key) { return _mm512_set1_epi32(key); }
+inline regi broadcast(std::int32_t key) { return _mm512_set1_epi32(key); }
 
-regi broadcast(std::int64_t key) { return _mm512_set1_epi64(key); }
+inline regi broadcast(std::int64_t key) { return _mm512_set1_epi64(key); }
 
-regf broadcast(float key) { return _mm512_set1_ps(key); }
+inline regf broadcast(float key) { return _mm512_set1_ps(key); }
 
-regd broadcast(double key) { return _mm512_set1_pd(key); }
+inline regd broadcast(double key) { return _mm512_set1_pd(key); }
 
-unsigned int cmp(regi key, const std::int32_t *key_ptr) {
+inline unsigned int cmp(regi key, const std::int32_t *key_ptr) {
   regi keys_to_comp =
       _mm512_load_si512(reinterpret_cast<const regi *>(key_ptr));
   return _mm512_cmpgt_epi32_mask(key, keys_to_comp);
 }
 
-unsigned int cmp(regi key, const std::int64_t *key_ptr) {
+inline unsigned int cmp(regi key, const std::int64_t *key_ptr) {
   regi keys_to_comp =
       _mm512_load_si512(reinterpret_cast<const regi *>(key_ptr));
   return _mm512_cmpgt_epi64_mask(key, keys_to_comp);
 }
 
-unsigned int cmp(regf key, const float *key_ptr) {
+inline unsigned int cmp(regf key, const float *key_ptr) {
   regf keys_to_comp = _mm512_load_ps(key_ptr);
   return _mm512_cmp_ps_mask(key, keys_to_comp, _MM_CMPINT_GT);
 }
 
-unsigned int cmp(regd key, const double *key_ptr) {
+inline unsigned int cmp(regd key, const double *key_ptr) {
   regd keys_to_comp = _mm512_load_pd(key_ptr);
   return _mm512_cmp_pd_mask(key, keys_to_comp, _MM_CMPINT_GT);
 }
 
-unsigned int cmp(const std::int32_t *key_ptr, regi key) {
+inline unsigned int cmp(const std::int32_t *key_ptr, regi key) {
   regi keys_to_comp =
       _mm512_load_si512(reinterpret_cast<const regi *>(key_ptr));
   return _mm512_cmpgt_epi32_mask(keys_to_comp, key);
 }
 
-unsigned int cmp(const std::int64_t *key_ptr, regi key) {
+inline unsigned int cmp(const std::int64_t *key_ptr, regi key) {
   regi keys_to_comp =
       _mm512_load_si512(reinterpret_cast<const regi *>(key_ptr));
   return _mm512_cmpgt_epi64_mask(keys_to_comp, key);
 }
 
-unsigned int cmp(const float *key_ptr, regf key) {
+inline unsigned int cmp(const float *key_ptr, regf key) {
   regf keys_to_comp = _mm512_load_ps(key_ptr);
   return _mm512_cmp_ps_mask(keys_to_comp, key, _MM_CMPINT_GT);
 }
 
-unsigned int cmp(const double *key_ptr, regd key) {
+inline unsigned int cmp(const double *key_ptr, regd key) {
   regd keys_to_comp = _mm512_load_pd(key_ptr);
   return _mm512_cmp_pd_mask(keys_to_comp, key, _MM_CMPINT_GT);
 }
@@ -81,12 +81,12 @@ template <CanUseSimd K> struct SimdTrait {
 };
 
 template <CanUseSimd K, bool less>
-inline std::int32_t get_lb_simd(K key, const K *first, const K *last) {
-  const auto len = static_cast<std::int32_t>(last - first);
+inline std::ptrdiff_t get_lb_simd(K key, const K *first, const K *last) {
+  const auto len = static_cast<std::ptrdiff_t>(last - first);
   const K *curr = first;
   auto key_broadcast = broadcast(key);
   int mask = 0;
-  int offset = 0;
+  ptrdiff_t offset = 0;
   while (offset < len) {
     if constexpr (less) {
       mask = ~cmp(key_broadcast, curr);
@@ -109,12 +109,12 @@ inline std::int32_t get_lb_simd(K key, const K *first, const K *last) {
 }
 
 template <CanUseSimd K, bool less>
-inline std::int32_t get_ub_simd(K key, const K *first, const K *last) {
-  const auto len = static_cast<std::int32_t>(last - first);
+inline std::ptrdiff_t get_ub_simd(K key, const K *first, const K *last) {
+  const auto len = static_cast<std::ptrdiff_t>(last - first);
   const K *curr = first;
   auto key_broadcast = broadcast(key);
   int mask = 0;
-  int offset = 0;
+  ptrdiff_t offset = 0;
   while (offset < len) {
     if constexpr (less) {
       mask = cmp(curr, key_broadcast);

diff --git a/include/fc/details.h b/include/fc/details.h
@@ -2,6 +2,7 @@
 #define FC_DETAILS_H
 
 #include <concepts>
+#include <cstdint>
 
 namespace frozenca {
 

diff --git a/include/fc/disk_btree.h b/include/fc/disk_btree.h
@@ -12,7 +12,7 @@ class DiskBTreeBase
     : public BTreeBase<K, V, Fanout, Comp, AllowDup, AllocatorFixed> {
 public:
   using Base = BTreeBase<K, V, Fanout, Comp, AllowDup, AllocatorFixed>;
-  using Node = Base::node_type;
+  using Node = typename Base::node_type;
 
 private:
   MemoryMappedFile mm_file_;

diff --git a/include/fc/disk_fixed_alloc.h b/include/fc/disk_fixed_alloc.h
@@ -8,14 +8,20 @@
 #include <concepts>
 #include <cstdint>
 #include <iostream>
+#if defined(__clang__) && __clang_major__ < 15
+#include <experimental/memory_resource>
+namespace stdpmr = std::experimental::pmr;
+#elif defined(__clang__) || (__GNUC__)
 #include <memory_resource>
+namespace stdpmr = std::pmr;
+#endif
 #include <stdexcept>
 #include <type_traits>
 
 namespace frozenca {
 
 template <typename T>
-class MemoryResourceFixed : public std::pmr::memory_resource {
+class MemoryResourceFixed : public stdpmr::memory_resource {
   T *pool_ = nullptr;
   std::size_t pool_size_ = 0;
   T *free_ = nullptr;
@@ -72,7 +78,7 @@ class MemoryResourceFixed : public std::pmr::memory_resource {
   }
 
   [[nodiscard]] bool
-  do_is_equal(const std::pmr::memory_resource &other) const noexcept override {
+  do_is_equal(const stdpmr::memory_resource &other) const noexcept override {
     if (this == &other) {
       return true;
     }
@@ -83,7 +89,7 @@ class MemoryResourceFixed : public std::pmr::memory_resource {
 };
 
 template <typename T> class AllocatorFixed {
-  std::pmr::memory_resource *mem_res_;
+  stdpmr::memory_resource *mem_res_;
 
 public:
   template <typename Other> struct rebind {
@@ -93,7 +99,7 @@ template <typename T> class AllocatorFixed {
   using value_type = T;
 
   explicit AllocatorFixed(
-      std::pmr::memory_resource *mem_res = std::pmr::get_default_resource())
+      stdpmr::memory_resource *mem_res = stdpmr::get_default_resource())
       : mem_res_{mem_res} {}
 
   template <typename Other>
@@ -110,7 +116,7 @@ template <typename T> class AllocatorFixed {
                          std::alignment_of_v<T>);
   }
 
-  [[nodiscard]] std::pmr::memory_resource *
+  [[nodiscard]] stdpmr::memory_resource *
   get_memory_resource() const noexcept {
     return mem_res_;
   }