Merge pull request #459 from DiamonDinoia/spreader-vectorization-v2

Spreader manual vectorization

.github/workflows/C++.yml

@@ -74,7 +74,6 @@ jobs:
 
     steps:
     - uses: actions/checkout@v4
-
     - name: 'Setup MSYS2'
       uses: msys2/setup-msys2@v2
       with:

.github/workflows/python_wheel.yml

@@ -186,6 +186,9 @@ jobs:
 
     steps:
     - uses: actions/checkout@v4
+    - uses: git-for-windows/setup-git-for-windows-sdk
+      with:
+        flavor: minimal
 
     - name: Install GCC and make
       run: C:\msys64\usr\bin\bash.exe -lc "pacman -Sy --noconfirm make mingw-w64-x86_64-toolchain mingw-w64-x86_64-fftw"

CMakeLists.txt

@@ -2,20 +2,24 @@ cmake_minimum_required(VERSION 3.19)
 
 project(finufft VERSION 2.2.0 LANGUAGES C CXX)
 
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
 set(GNU_LIKE_FRONTENDS AppleClang Clang GNU)
 if (CMAKE_CXX_COMPILER_ID IN_LIST GNU_LIKE_FRONTENDS)
     # Set custom compiler flags for gcc-compatible compilers
-    set(CMAKE_CXX_FLAGS_RELEASE "-O3 -DNDEBUG -funroll-loops")
-    set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG -funroll-loops")
+    set(FINUFFT_CXX_FLAGS_RELEASE -O3 -funroll-loops -ffp-contract=fast)
+    set(FINUFFT_CXX_FLAGS_RELWITHDEBINFO -g ${FINUFFT_CXX_FLAGS_RELEASE})
 endif ()
 
 include(CTest)
 
-if (CMAKE_SYSTEM_PROCESSOR MATCHES "ppc|ppc64|powerpc|powerpc64" OR (APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES "ppc|ppc64"))
-    # PowerPC arch does not have -march flag.
-    set(FINUFFT_ARCH_FLAGS "-mtune=native" CACHE STRING "Compiler flags for specifying target architecture.")
-else ()
-    set(FINUFFT_ARCH_FLAGS "-march=native" CACHE STRING "Compiler flags for specifying target architecture.")
+if (CMAKE_CXX_COMPILER_ID IN_LIST GNU_LIKE_FRONTENDS)
+    if (CMAKE_SYSTEM_PROCESSOR MATCHES "ppc|ppc64|powerpc|powerpc64" OR (APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES "ppc|ppc64"))
+        # PowerPC arch does not have -march flag.
+        set(FINUFFT_ARCH_FLAGS "-mtune=native" CACHE STRING "Compiler flags for specifying target architecture.")
+    else ()
+        set(FINUFFT_ARCH_FLAGS "-march=native" CACHE STRING "Compiler flags for specifying target architecture.")
+    endif ()
 endif ()
 set(FINUFFT_FFTW_SUFFIX "OpenMP" CACHE STRING "Suffix for FFTW libraries (e.g. OpenMP, Threads etc.)")
 set(FINUFFT_FFTW_LIBRARIES "DEFAULT" CACHE STRING "Specify a custom FFTW library")
@@ -31,7 +35,7 @@ option(FINUFFT_USE_OPENMP "Whether to use OpenMP for parallelization. If disable
 option(FINUFFT_USE_CUDA "Whether to build CUDA accelerated FINUFFT library (libcufinufft). This is completely independent of the main FINUFFT library" OFF)
 option(FINUFFT_USE_CPU "Whether to build the ordinary FINUFFT library (libfinufft)." ON)
 option(FINUFFT_STATIC_LINKING "Whether to link the static FINUFFT library (libfinufft_static)." ON)
-option(FINUFFT_BUILD_DEVEL "Whether to build developement executables" OFF)
+option(FINUFFT_BUILD_DEVEL "Whether to build development executables" OFF)
 # sphinx tag (don't remove): @cmake_opts_end
 
 if (FINUFFT_USE_CPU)
@@ -47,10 +51,11 @@ if (FINUFFT_USE_CPU)
 
     set(CPM_DOWNLOAD_VERSION 0.38.0)
     set(FFTW_VERSION 3.3.10)
-
+    set(XTL_VERSION 0.7.7)
+    set(XSIMD_VERSION 13.0.0)
     include(cmake/setupCPM.cmake)
     include(cmake/setupFFTW.cmake)
-
+    include(cmake/setupXSIMD.cmake)
 endif ()
 
 if (FINUFFT_BUILD_MATLAB)
@@ -92,8 +97,8 @@ function(enable_asan target)
 
     if (CMAKE_CXX_COMPILER_ID IN_LIST FINUFFT_GNU_LIKE_COMPILERS)
         # Enable only on clang / gcc compilers.
-        target_compile_options(${target} PRIVATE $<$<CONFIG:DEBUG>:-fsanitize=address>)
-        target_link_options(${target} PRIVATE $<$<CONFIG:DEBUG>:-fsanitize=address>)
+        target_compile_options(${target} PRIVATE $<$<CONFIG:DEBUG>:-fsanitize=address -fsanitize=undefined -fsanitize=bounds-strict>)
+        target_link_options(${target} PRIVATE $<$<CONFIG:DEBUG>:-fsanitize=address -fsanitize=undefined -fsanitize=bounds-strict>)
     endif ()
 endfunction()
 
@@ -119,17 +124,18 @@ endfunction()
 # Utility function to set finufft compilation options.
 function(set_finufft_options target)
     set_property(TARGET ${target} PROPERTY POSITION_INDEPENDENT_CODE ON)
-    set_property(TARGET ${target} PROPERTY CMAKE_CXX_STANDARD 14)
+    target_compile_features(${target} PRIVATE cxx_std_17)
+
     enable_asan(${target})
 
     target_compile_options(${target} PRIVATE SHELL:$<$<CONFIG:Release,RelWithDebInfo>:${FINUFFT_ARCH_FLAGS}>)
     if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
         # Enable cx-limited-range on supported compilers
         target_compile_options(${target} PRIVATE $<$<CONFIG:Release,RelWithDebInfo>:-fcx-limited-range>)
     endif ()
-
-    target_include_directories(${target} PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
-    target_include_directories(${target} SYSTEM INTERFACE $<INSTALL_INTERFACE:${CMAKE_INSTALL_PREFIX}/include>)
+    target_compile_options(${target} PRIVATE $<$<CONFIG:Release>:${FINUFFT_CXX_FLAGS_RELEASE}>)
+    target_compile_options(${target} PRIVATE $<$<CONFIG:RelWithDebInfo>:${FINUFFT_CXX_FLAGS_RELWITHDEBINFO}>)
+    target_include_directories(${target} PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}/include")
     if (FINUFFT_USE_OPENMP)
         target_link_libraries(${target} PRIVATE OpenMP::OpenMP_CXX)
         # there are issues on windows with OpenMP and CMake, so we need to manually add the flags
@@ -163,12 +169,13 @@ if (FINUFFT_USE_CPU)
     target_compile_definitions(finufft_f32 PRIVATE SINGLE)
     set_finufft_options(finufft_f32)
     target_link_libraries(finufft_f32 PUBLIC ${FINUFFT_FFTW_LIBRARIES})
+    target_link_libraries(finufft_f32 PRIVATE xsimd)
 
     add_library(finufft_f64 OBJECT ${FINUFFT_PRECISION_DEPENDENT_SOURCES})
     target_compile_definitions(finufft_f64 PRIVATE)
     set_finufft_options(finufft_f64)
     target_link_libraries(finufft_f64 PUBLIC ${FINUFFT_FFTW_LIBRARIES})
-
+    target_link_libraries(finufft_f64 PRIVATE xsimd)
     if (WIN32)
         add_library(finufft_f32_dll OBJECT ${FINUFFT_PRECISION_DEPENDENT_SOURCES})
         target_compile_definitions(finufft_f32_dll PRIVATE SINGLE dll_EXPORTS FINUFFT_DLL)
@@ -193,8 +200,7 @@ if (FINUFFT_USE_CPU)
     if (NOT WIN32)
         target_link_libraries(finufft PUBLIC m)
     endif ()
-    target_include_directories(finufft PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
-    target_include_directories(finufft SYSTEM INTERFACE $<INSTALL_INTERFACE:${CMAKE_INSTALL_PREFIX}/include>)
+    target_include_directories(finufft PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}/include")
 
     add_library(finufft_static STATIC src/utils_precindep.cpp contrib/legendre_rule_fast.cpp)
     set_finufft_options(finufft_static)
@@ -203,8 +209,7 @@ if (FINUFFT_USE_CPU)
     if (NOT WIN32)
         target_link_libraries(finufft_static PUBLIC m)
     endif ()
-    target_include_directories(finufft_static PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
-    target_include_directories(finufft_static SYSTEM INTERFACE $<INSTALL_INTERFACE:${CMAKE_INSTALL_PREFIX}/include>)
+    target_include_directories(finufft_static PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}/include")
 
     file(GLOB FINUFFT_PUBLIC_HEADERS "${CMAKE_CURRENT_SOURCE_DIR}/include/finufft*.h")
     set_target_properties(finufft PROPERTIES PUBLIC_HEADER "${FINUFFT_PUBLIC_HEADERS}")

cmake/setupXSIMD.cmake

@@ -0,0 +1,19 @@
+CPMAddPackage(
+        NAME findxtl
+        GIT_REPOSITORY "https://github.com/xtensor-stack/xtl.git"
+        GIT_TAG ${XTL_VERSION}
+        EXCLUDE_FROM_ALL YES
+        GIT_SHALLOW YES
+        OPTIONS "XTL_DISABLE_EXCEPTIONS YES"
+)
+
+CPMAddPackage(
+        NAME findxsimd
+        GIT_REPOSITORY "https://github.com/xtensor-stack/xsimd.git"
+        GIT_TAG ${XSIMD_VERSION}
+        EXCLUDE_FROM_ALL YES
+        GIT_SHALLOW YES
+        OPTIONS
+            "XSIMD_SKIP_INSTALL YES"
+            "XSIMD_ENABLE_XTL_COMPLEX YES"
+)

devel/CMakeLists.txt

@@ -18,5 +18,7 @@ if (benchmark_ADDED)
 endif()
 
 add_executable(foldrescale foldrescale.cpp)
-target_link_libraries(foldrescale finufft benchmark)
-target_compile_options(foldrescale PRIVATE -mavx2)
+target_link_libraries(foldrescale finufft benchmark xsimd)
+add_executable(padding padding.cpp)
+target_link_libraries(padding finufft xsimd)
+target_compile_options(padding PRIVATE -march=native)