TsFile C++: batch + SIMD + parallel read/write optimization

Batch decode/encode APIs (PLAIN / TS2DIFF / Gorilla) with single-pass
TsBlock decode, AVX2/NEON SIMD paths, a single process-wide worker pool
for chunk-level parallel read and column-parallel write, and batched
NEON statistics. On-disk format unchanged; interoperable with Java/Python.

Author: ColinLeeo

cpp/CMakeLists.txt

@@ -129,7 +129,16 @@ else ()
         if (CMAKE_BUILD_TYPE STREQUAL "Debug")
             set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -g")
         elseif (CMAKE_BUILD_TYPE STREQUAL "Release")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O2")
+            # -flto + MinGW gcc + statically-linked antlr4_static produces
+            # unresolved-reference errors at link time (LTO intermediate objects
+            # can't see the .a's vtable thunks). -march=native is also a poor
+            # default for CI binaries shipped to other machines. Keep both on
+            # Linux/macOS where the optimization actually pays off.
+            if (MINGW OR WIN32)
+                set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O3")
+            else ()
+                set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O3 -march=native -flto")
+            endif ()
         elseif (CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo")
             set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -O2 -g")
         elseif (CMAKE_BUILD_TYPE STREQUAL "MinSizeRel")

cpp/pom.xml

@@ -99,8 +99,8 @@
                                     plugin's generate goal throw an NPE.
                                 -->
                             </options>
-                            <sourcePath />
-                            <targetPath />
+                            <sourcePath/>
+                            <targetPath/>
                         </configuration>
                     </execution>
                     <!-- Compile the test code -->

cpp/src/CMakeLists.txt

@@ -37,6 +37,9 @@ message("cmake using: ENABLE_LZOKAY=${ENABLE_LZOKAY}")
 option(ENABLE_ZLIB "Enable Zlib compression" ON)
 message("cmake using: ENABLE_ZLIB=${ENABLE_ZLIB}")
 
+# ENABLE_SIMD is defined in the top-level CMakeLists.txt
+message("cmake using: ENABLE_SIMD=${ENABLE_SIMD}")
+
 message("Running in src directory")
 if (${COV_ENABLED})
     add_compile_options(-fprofile-arcs -ftest-coverage)
@@ -89,6 +92,13 @@ if (ENABLE_ANTLR4)
     message("Adding ANTLR4 include directory")
 endif()
 
+if (ENABLE_SIMD)
+    add_definitions(-DENABLE_SIMD)
+    list(APPEND PROJECT_INCLUDE_DIR
+            ${CMAKE_SOURCE_DIR}/third_party/simde-0.8.4-rc3
+    )
+endif()
+
 include_directories(${PROJECT_INCLUDE_DIR})
 
 # Mark every translation unit that is compiled into the tsfile library so that
@@ -144,10 +154,17 @@ add_library(tsfile SHARED)
 
 if (${COV_ENABLED})
     message("Enable code cov...")
+    # Apple clang ships coverage runtime via --coverage; libgcov isn't a
+    # standalone library on macOS.  Use --coverage there.
+    if (APPLE)
+        set(COV_LINK_LIB --coverage)
+    else()
+        set(COV_LINK_LIB -lgcov)
+    endif()
     if (ENABLE_ANTLR4)
-        target_link_libraries(tsfile common_obj compress_obj cwrapper_obj file_obj read_obj write_obj parser_obj -lgcov)
+        target_link_libraries(tsfile common_obj compress_obj cwrapper_obj file_obj read_obj write_obj parser_obj ${COV_LINK_LIB})
     else()
-        target_link_libraries(tsfile common_obj compress_obj cwrapper_obj file_obj read_obj write_obj -lgcov)
+        target_link_libraries(tsfile common_obj compress_obj cwrapper_obj file_obj read_obj write_obj ${COV_LINK_LIB})
     endif()
 else()
     message("Disable code cov...")
@@ -171,4 +188,4 @@ set_target_properties(tsfile PROPERTIES SOVERSION ${LIBTSFILE_SO_VERSION})
 install(TARGETS tsfile
         RUNTIME DESTINATION ${LIBRARY_OUTPUT_PATH}
         LIBRARY DESTINATION ${LIBRARY_OUTPUT_PATH}
-        ARCHIVE DESTINATION ${LIBRARY_OUTPUT_PATH})
+        ARCHIVE DESTINATION ${LIBRARY_OUTPUT_PATH})

cpp/src/common/CMakeLists.txt

@@ -22,21 +22,15 @@ aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR} common_SRC_LIST)
 aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR}/allocator common_allocator_SRC_LIST)
 aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR}/container common_container_SRC_LIST)
 aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR}/tsblock common_tsblock_SRC_LIST)
-aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR}/mutex common_mutex_SRC_LIST)
 aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR}/datatype common_datatype_SRC_LIST)
 
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
-add_library(common_obj OBJECT ${common_SRC_LIST} 
+add_library(common_obj OBJECT ${common_SRC_LIST}
     ${common_allocator_SRC_LIST}
     ${common_container_SRC_LIST}
-    ${common_tsblock_SRC_LIST} 
-    ${common_mutex_SRC_LIST} 
+    ${common_tsblock_SRC_LIST}
     ${common_datatype_SRC_LIST})
 
-if (ENABLE_ANTLR4)
-    target_compile_definitions(common_obj PRIVATE ENABLE_ANTLR4)
-endif()
-
 # install header files recursively
 file(GLOB_RECURSE HEADERS "${CMAKE_CURRENT_SOURCE_DIR}/*.h")
 copy_to_dir(${HEADERS} "common_obj")

cpp/src/common/allocator/alloc_base.h

@@ -82,35 +82,43 @@ class ModStat {
     }
     void init();
     void destroy();
-    INLINE void update_alloc(AllocModID mid, int32_t size) {
+    INLINE void update_alloc(AllocModID mid, int64_t size) {
 #ifdef ENABLE_MEM_STAT
         ASSERT(mid < __LAST_MOD_ID);
         ATOMIC_FAA(get_item(mid), size);
 #endif
     }
-    void update_free(AllocModID mid, uint32_t size) {
+    void update_free(AllocModID mid, uint64_t size) {
 #ifdef ENABLE_MEM_STAT
         ASSERT(mid < __LAST_MOD_ID);
-        ATOMIC_FAA(get_item(mid), 0 - size);
+        ATOMIC_FAA(get_item(mid), -static_cast<int64_t>(size));
 #endif
     }
     void print_stat();
 
+    int64_t get_stat(int8_t mid) {
+#ifdef ENABLE_MEM_STAT
+        if (stat_arr_ != NULL && mid < __LAST_MOD_ID)
+            return ATOMIC_FAA(get_item(mid), 0LL);
+#endif
+        return 0;
+    }
+
 #ifdef ENABLE_TEST
-    int32_t TEST_get_stat(int8_t mid) { return ATOMIC_FAA(get_item(mid), 0); }
+    int64_t TEST_get_stat(int8_t mid) { return ATOMIC_FAA(get_item(mid), 0LL); }
 #endif
 
    private:
-    INLINE int32_t* get_item(int8_t mid) {
-        return &(stat_arr_[mid * (ITEM_SIZE / sizeof(int32_t))]);
+    INLINE int64_t* get_item(int8_t mid) {
+        return &(stat_arr_[mid * (ITEM_SIZE / sizeof(int64_t))]);
     }
 
    private:
     static const int32_t ITEM_SIZE = CACHE_LINE_SIZE;
     static const int32_t ITEM_COUNT = __LAST_MOD_ID;
-    int32_t* stat_arr_;
+    int64_t* stat_arr_;
 
-    STATIC_ASSERT((ITEM_SIZE % sizeof(int32_t) == 0), ModStat_ITEM_SIZE_ERROR);
+    STATIC_ASSERT((ITEM_SIZE % sizeof(int64_t) == 0), ModStat_ITEM_SIZE_ERROR);
 };
 
 /* base allocator */