ARROW-18162: [C++] Add Arm SVE compiler options (#14515)

As xsimd only supports fixed-size SVE, we have to specify vector size
explicitly on command line. And the binary can only run on hardware
with matched vector size. Otherwise, the code behaviour is undefined.
E.g., `cmake -DARROW_SIMD_LEVEL=SVE256 ..`
According macro `ARROW_HAVE_SVE256` and cmake variable are defined.

We can also leverage compiler auto vectorization to generate size
agnostic SVE code without specifying the vector size.
E.g., `cmake -DARROW_SIMD_LEVEL=SVE ..`

This PR also removes some unused Arm64 arch options.

Authored-by: Yibo Cai <[email protected]>
Signed-off-by: Yibo Cai <[email protected]>

Author: cyb70289

cpp/cmake_modules/DefineOptions.cmake

@@ -185,6 +185,10 @@ takes precedence over ccache if a storage backend is configured" ON)
                        "AVX2"
                        "AVX512"
                        "NEON"
+                       "SVE" # size agnostic SVE
+                       "SVE128" # fixed size SVE
+                       "SVE256" # "
+                       "SVE512" # "
                        "DEFAULT")
 
   define_option_string(ARROW_RUNTIME_SIMD_LEVEL
@@ -196,17 +200,6 @@ takes precedence over ccache if a storage backend is configured" ON)
                        "AVX512"
                        "MAX")
 
-  # Arm64 architectures and extensions can lead to exploding combinations.
-  # So set it directly through cmake command line.
-  #
-  # If you change this, you need to change the definition in
-  # python/CMakeLists.txt too.
-  define_option_string(ARROW_ARMV8_ARCH
-                       "Arm64 arch and extensions"
-                       "armv8-a" # Default
-                       "armv8-a"
-                       "armv8-a+crc+crypto")
-
   define_option(ARROW_ALTIVEC "Build with Altivec if compiler has support" ON)
 
   define_option(ARROW_RPATH_ORIGIN "Build Arrow libraries with RATH set to \$ORIGIN" OFF)

cpp/cmake_modules/SetupCxxFlags.cmake

@@ -110,8 +110,8 @@ elseif(ARROW_CPU_FLAG STREQUAL "ppc")
   endif()
 elseif(ARROW_CPU_FLAG STREQUAL "armv8")
   # Arm64 compiler flags, gcc/clang only
-  set(ARROW_ARMV8_ARCH_FLAG "-march=${ARROW_ARMV8_ARCH}")
-  check_cxx_compiler_flag(${ARROW_ARMV8_ARCH_FLAG} CXX_SUPPORTS_ARMV8_ARCH)
+  set(ARROW_ARMV8_MARCH "armv8-a")
+  check_cxx_compiler_flag("-march=${ARROW_ARMV8_MARCH}+sve" CXX_SUPPORTS_SVE)
   if(ARROW_SIMD_LEVEL STREQUAL "DEFAULT")
     set(ARROW_SIMD_LEVEL "NEON")
   endif()
@@ -485,26 +485,27 @@ if(ARROW_CPU_FLAG STREQUAL "ppc")
 endif()
 
 if(ARROW_CPU_FLAG STREQUAL "armv8")
-  if(ARROW_SIMD_LEVEL STREQUAL "NEON")
+  if(ARROW_SIMD_LEVEL MATCHES "NEON|SVE[0-9]*")
     set(ARROW_HAVE_NEON ON)
-
-    if(NOT CXX_SUPPORTS_ARMV8_ARCH)
-      message(FATAL_ERROR "Unsupported arch flag: ${ARROW_ARMV8_ARCH_FLAG}.")
-    endif()
-    if(ARROW_ARMV8_ARCH_FLAG MATCHES "native")
-      message(FATAL_ERROR "native arch not allowed, please specify arch explicitly.")
-    endif()
-    set(CXX_COMMON_FLAGS "${CXX_COMMON_FLAGS} ${ARROW_ARMV8_ARCH_FLAG}")
-
     add_definitions(-DARROW_HAVE_NEON)
-
-    if(ARROW_ARMV8_ARCH_FLAG MATCHES "\\+crypto")
-      add_definitions(-DARROW_HAVE_ARMV8_CRYPTO)
-    endif()
-    # armv8.1+ implies crc support
-    if(ARROW_ARMV8_ARCH_FLAG MATCHES "armv8\\.[1-9]|\\+crc")
-      add_definitions(-DARROW_HAVE_ARMV8_CRC)
+    if(ARROW_SIMD_LEVEL MATCHES "SVE[0-9]*")
+      if(NOT CXX_SUPPORTS_SVE)
+        message(FATAL_ERROR "SVE required but compiler doesn't support it.")
+      endif()
+      # -march=armv8-a+sve
+      set(ARROW_ARMV8_MARCH "${ARROW_ARMV8_MARCH}+sve")
+      string(REGEX MATCH "[0-9]+" SVE_VECTOR_BITS ${ARROW_SIMD_LEVEL})
+      if(SVE_VECTOR_BITS)
+        set(ARROW_HAVE_SVE${SVE_VECTOR_BITS} ON)
+        add_definitions(-DARROW_HAVE_SVE${SVE_VECTOR_BITS})
+        # -msve-vector-bits=256
+        set(CXX_COMMON_FLAGS "${CXX_COMMON_FLAGS} -msve-vector-bits=${SVE_VECTOR_BITS}")
+      else()
+        set(ARROW_HAVE_SVE_SIZELESS ON)
+        add_definitions(-DARROW_HAVE_SVE_SIZELSS)
+      endif()
     endif()
+    set(CXX_COMMON_FLAGS "${CXX_COMMON_FLAGS} -march=${ARROW_ARMV8_MARCH}")
   elseif(NOT ARROW_SIMD_LEVEL STREQUAL "NONE")
     message(WARNING "ARROW_SIMD_LEVEL=${ARROW_SIMD_LEVEL} not supported by Arm.")
   endif()

cpp/src/arrow/io/memory_benchmark.cc

@@ -154,7 +154,7 @@ static void StreamReadWrite(void* src, void* dst, size_t size) {
 
 #endif  // ARROW_HAVE_SSE4_2
 
-#ifdef ARROW_HAVE_ARMV8_CRYPTO
+#ifdef ARROW_HAVE_NEON
 
 using VectorType = uint8x16_t;
 using VectorTypeDual = uint8x16x2_t;
@@ -237,7 +237,7 @@ static void StreamReadWrite(void* src, void* dst, size_t size) {
   }
 }
 
-#endif  // ARROW_HAVE_ARMV8_CRYPTO
+#endif  // ARROW_HAVE_NEON
 
 static void PlatformMemcpy(void* src, void* dst, size_t size) { memcpy(src, dst, size); }
 

cpp/src/arrow/util/simd.h

@@ -41,8 +41,4 @@
 #include <arm_neon.h>
 #endif
 
-#ifdef ARROW_HAVE_ARMV8_CRC
-#include <arm_acle.h>
-#endif
-
 #endif

dev/tasks/conda-recipes/arrow-cpp/build-pyarrow.sh

@@ -30,11 +30,6 @@ else
     export PYARROW_WITH_CUDA=0
 fi
 
-# Resolve: Make Error at cmake_modules/SetupCxxFlags.cmake:338 (message): Unsupported arch flag: -march=.
-if [[ "${target_platform}" == "linux-aarch64" ]]; then
-    export PYARROW_CMAKE_OPTIONS="-DARROW_ARMV8_ARCH=armv8-a ${PYARROW_CMAKE_OPTIONS}"
-fi
-
 if [[ "${target_platform}" == osx-* ]]; then
     # See https://conda-forge.org/docs/maintainer/knowledge_base.html#newer-c-features-with-old-sdk
     CXXFLAGS="${CXXFLAGS} -D_LIBCPP_DISABLE_AVAILABILITY"

python/CMakeLists.txt

@@ -126,11 +126,6 @@ if(NOT DEFINED ARROW_RUNTIME_SIMD_LEVEL)
       "MAX"
       CACHE STRING "Max runtime SIMD optimization level")
 endif()
-if(NOT DEFINED ARROW_ARMV8_ARCH)
-  set(ARROW_ARMV8_ARCH
-      "armv8-a"
-      CACHE STRING "Arm64 arch and extensions: armv8-a, armv8-a or armv8-a+crc+crypto")
-endif()
 include(SetupCxxFlags)
 
 # Add common flags