GH-35116: [CI][C++] Enable compile-time AVX2 on some CI platforms (#3…

…6662) AVX2 became mainline on Intel and AMD server CPUs around 2015, so it's unlikely to be unavailable on current cloud platforms: https://en.wikipedia.org/wiki/Advanced_Vector_Extensions#CPUs_with_AVX2 Enable it at least on one Windows and one Linux CI platform. x86 macOS is a legacy platform, so less interesting to exercise there (and I'm not sure the old CPUs in x86 Macs actually support AVX2). Also, fix the buggy AVX2 activation logic in Acero and avoid force-testing AVX2 on incompatible systems. * Closes: #35116 Authored-by: Antoine Pitrou <antoine@python.org> Signed-off-by: Antoine Pitrou <antoine@python.org>
apache · Jul 19, 2023 · 366e808 · 366e808
1 parent 80f77d1
commit 366e808
Show file tree

Hide file tree

Showing 34 changed files with 289 additions and 236 deletions.
diff --git a/.github/workflows/cpp.yml b/.github/workflows/cpp.yml
@@ -67,7 +67,8 @@ jobs:
             image: conda-cpp
             llvm: "14"
             runs-on: ubuntu-latest
-            title: AMD64 Conda C++
+            simd-level: AVX2
+            title: AMD64 Conda C++ AVX2
             ubuntu: "22.04"
           - arch: amd64
             clang-tools: "14"
@@ -85,6 +86,7 @@ jobs:
             ubuntu: "20.04"
     env:
       ARCH: ${{ matrix.arch }}
+      ARROW_SIMD_LEVEL: ${{ matrix.simd-level }}
       CLANG_TOOLS: ${{ matrix.clang-tools }}
       LLVM: ${{ matrix.llvm }}
       UBUNTU: ${{ matrix.ubuntu }}
@@ -175,6 +177,10 @@ jobs:
       ARROW_WITH_ZSTD: ON
       GTest_SOURCE: BUNDLED
     steps:
+      - name: CPU Info
+        run: |
+          sysctl -a | grep cpu
+          sysctl -a | grep "hw.optional"
       - name: Checkout Arrow
         uses: actions/checkout@v3
         with:
@@ -220,7 +226,7 @@ jobs:
           ci/scripts/cpp_test.sh $(pwd) $(pwd)/build
 
   windows:
-    name: AMD64 ${{ matrix.name }} C++17
+    name: ${{ matrix.title }}
     runs-on: ${{ matrix.os }}
     if: ${{ !contains(github.event.pull_request.title, 'WIP') }}
     timeout-minutes: 60
@@ -231,7 +237,8 @@ jobs:
           - windows-2019
         include:
           - os: windows-2019
-            name: Windows 2019
+            simd-level: AVX2
+            title: AMD64 Windows 2019 C++17 AVX2
     env:
       ARROW_BOOST_USE_SHARED: OFF
       ARROW_BUILD_BENCHMARKS: ON
@@ -246,6 +253,7 @@ jobs:
       ARROW_MIMALLOC: ON
       ARROW_ORC: ON
       ARROW_PARQUET: ON
+      ARROW_SIMD_LEVEL: ${{ matrix.simd-level }}
       ARROW_USE_GLOG: OFF
       ARROW_VERBOSE_THIRDPARTY_BUILD: OFF
       ARROW_WITH_BROTLI: OFF

diff --git a/ci/scripts/cpp_build.sh b/ci/scripts/cpp_build.sh
@@ -126,6 +126,7 @@ cmake \
   -DARROW_PARQUET=${ARROW_PARQUET:-OFF} \
   -DARROW_RUNTIME_SIMD_LEVEL=${ARROW_RUNTIME_SIMD_LEVEL:-MAX} \
   -DARROW_S3=${ARROW_S3:-OFF} \
+  -DARROW_SIMD_LEVEL=${ARROW_SIMD_LEVEL:-DEFAULT} \
   -DARROW_SKYHOOK=${ARROW_SKYHOOK:-OFF} \
   -DARROW_SUBSTRAIT=${ARROW_SUBSTRAIT:-ON} \
   -DARROW_TEST_LINKAGE=${ARROW_TEST_LINKAGE:-shared} \

diff --git a/cpp/cmake_modules/SetupCxxFlags.cmake b/cpp/cmake_modules/SetupCxxFlags.cmake
@@ -62,29 +62,32 @@ if(ARROW_CPU_FLAG STREQUAL "x86")
         "${ARROW_AVX512_FLAG} -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw")
     check_cxx_compiler_flag(${ARROW_SSE4_2_FLAG} CXX_SUPPORTS_SSE4_2)
   endif()
-  check_cxx_compiler_flag(${ARROW_AVX2_FLAG} CXX_SUPPORTS_AVX2)
-  if(MINGW)
-    # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65782
-    message(STATUS "Disable AVX512 support on MINGW for now")
-  else()
-    # Check for AVX512 support in the compiler.
-    set(OLD_CMAKE_REQURED_FLAGS ${CMAKE_REQUIRED_FLAGS})
-    set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} ${ARROW_AVX512_FLAG}")
-    check_cxx_source_compiles("
-      #ifdef _MSC_VER
-      #include <intrin.h>
-      #else
-      #include <immintrin.h>
-      #endif
-
-      int main() {
-        __m512i mask = _mm512_set1_epi32(0x1);
-        char out[32];
-        _mm512_storeu_si512(out, mask);
-        return 0;
-      }"
-                              CXX_SUPPORTS_AVX512)
-    set(CMAKE_REQUIRED_FLAGS ${OLD_CMAKE_REQURED_FLAGS})
+  if(CMAKE_SIZEOF_VOID_P EQUAL 8)
+    # Check for AVX extensions on 64-bit systems only, as 32-bit support seems iffy
+    check_cxx_compiler_flag(${ARROW_AVX2_FLAG} CXX_SUPPORTS_AVX2)
+    if(MINGW)
+      # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65782
+      message(STATUS "Disable AVX512 support on MINGW for now")
+    else()
+      # Check for AVX512 support in the compiler.
+      set(OLD_CMAKE_REQURED_FLAGS ${CMAKE_REQUIRED_FLAGS})
+      set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} ${ARROW_AVX512_FLAG}")
+      check_cxx_source_compiles("
+        #ifdef _MSC_VER
+        #include <intrin.h>
+        #else
+        #include <immintrin.h>
+        #endif
+
+        int main() {
+          __m512i mask = _mm512_set1_epi32(0x1);
+          char out[32];
+          _mm512_storeu_si512(out, mask);
+          return 0;
+        }"
+                                CXX_SUPPORTS_AVX512)
+      set(CMAKE_REQUIRED_FLAGS ${OLD_CMAKE_REQURED_FLAGS})
+    endif()
   endif()
   # Runtime SIMD level it can get from compiler and ARROW_RUNTIME_SIMD_LEVEL
   if(CXX_SUPPORTS_SSE4_2 AND ARROW_RUNTIME_SIMD_LEVEL MATCHES

diff --git a/cpp/src/arrow/CMakeLists.txt b/cpp/src/arrow/CMakeLists.txt
@@ -119,15 +119,15 @@ function(ADD_ARROW_BENCHMARK REL_TEST_NAME)
                 ${ARG_UNPARSED_ARGUMENTS})
 endfunction()
 
-macro(append_avx2_src SRC)
+macro(append_runtime_avx2_src SRC)
   if(ARROW_HAVE_RUNTIME_AVX2)
     list(APPEND ARROW_SRCS ${SRC})
     set_source_files_properties(${SRC} PROPERTIES SKIP_PRECOMPILE_HEADERS ON)
     set_source_files_properties(${SRC} PROPERTIES COMPILE_FLAGS ${ARROW_AVX2_FLAG})
   endif()
 endmacro()
 
-macro(append_avx512_src SRC)
+macro(append_runtime_avx512_src SRC)
   if(ARROW_HAVE_RUNTIME_AVX512)
     list(APPEND ARROW_SRCS ${SRC})
     set_source_files_properties(${SRC} PROPERTIES SKIP_PRECOMPILE_HEADERS ON)
@@ -254,8 +254,8 @@ if(ARROW_JEMALLOC)
                               PROPERTIES SKIP_UNITY_BUILD_INCLUSION ON)
 endif()
 
-append_avx2_src(util/bpacking_avx2.cc)
-append_avx512_src(util/bpacking_avx512.cc)
+append_runtime_avx2_src(util/bpacking_avx2.cc)
+append_runtime_avx512_src(util/bpacking_avx512.cc)
 
 if(ARROW_HAVE_NEON)
   list(APPEND ARROW_SRCS util/bpacking_neon.cc)
@@ -425,11 +425,11 @@ list(APPEND
      compute/row/row_internal.cc
      compute/util.cc)
 
-append_avx2_src(compute/key_hash_avx2.cc)
-append_avx2_src(compute/key_map_avx2.cc)
-append_avx2_src(compute/row/compare_internal_avx2.cc)
-append_avx2_src(compute/row/encode_internal_avx2.cc)
-append_avx2_src(compute/util_avx2.cc)
+append_runtime_avx2_src(compute/key_hash_avx2.cc)
+append_runtime_avx2_src(compute/key_map_avx2.cc)
+append_runtime_avx2_src(compute/row/compare_internal_avx2.cc)
+append_runtime_avx2_src(compute/row/encode_internal_avx2.cc)
+append_runtime_avx2_src(compute/util_avx2.cc)
 
 if(ARROW_COMPUTE)
   # Include the remaining kernels
@@ -464,8 +464,8 @@ if(ARROW_COMPUTE)
        compute/kernels/vector_select_k.cc
        compute/kernels/vector_sort.cc)
 
-  append_avx2_src(compute/kernels/aggregate_basic_avx2.cc)
-  append_avx512_src(compute/kernels/aggregate_basic_avx512.cc)
+  append_runtime_avx2_src(compute/kernels/aggregate_basic_avx2.cc)
+  append_runtime_avx512_src(compute/kernels/aggregate_basic_avx512.cc)
 endif()
 
 if(ARROW_FILESYSTEM)

diff --git a/cpp/src/arrow/acero/CMakeLists.txt b/cpp/src/arrow/acero/CMakeLists.txt
@@ -19,7 +19,7 @@ add_custom_target(arrow_acero)
 
 arrow_install_all_headers("arrow/acero")
 
-macro(append_acero_avx2_src SRC)
+macro(append_acero_runtime_avx2_src SRC)
   if(ARROW_HAVE_RUNTIME_AVX2)
     list(APPEND ARROW_ACERO_SRCS ${SRC})
     set_source_files_properties(${SRC} PROPERTIES SKIP_PRECOMPILE_HEADERS ON)
@@ -56,8 +56,8 @@ set(ARROW_ACERO_SRCS
     union_node.cc
     util.cc)
 
-append_acero_avx2_src(bloom_filter_avx2.cc)
-append_acero_avx2_src(swiss_join_avx2.cc)
+append_acero_runtime_avx2_src(bloom_filter_avx2.cc)
+append_acero_runtime_avx2_src(swiss_join_avx2.cc)
 
 set(ARROW_ACERO_SHARED_LINK_LIBS)
 set(ARROW_ACERO_STATIC_LINK_LIBS)

diff --git a/cpp/src/arrow/acero/bloom_filter.cc b/cpp/src/arrow/acero/bloom_filter.cc
@@ -123,7 +123,7 @@ void BlockedBloomFilter::InsertImp(int64_t num_rows, const T* hashes) {
 void BlockedBloomFilter::Insert(int64_t hardware_flags, int64_t num_rows,
                                 const uint32_t* hashes) {
   int64_t num_processed = 0;
-#if defined(ARROW_HAVE_AVX2)
+#if defined(ARROW_HAVE_RUNTIME_AVX2)
   if (hardware_flags & arrow::internal::CpuInfo::AVX2) {
     num_processed = Insert_avx2(num_rows, hashes);
   }
@@ -134,7 +134,7 @@ void BlockedBloomFilter::Insert(int64_t hardware_flags, int64_t num_rows,
 void BlockedBloomFilter::Insert(int64_t hardware_flags, int64_t num_rows,
                                 const uint64_t* hashes) {
   int64_t num_processed = 0;
-#if defined(ARROW_HAVE_AVX2)
+#if defined(ARROW_HAVE_RUNTIME_AVX2)
   if (hardware_flags & arrow::internal::CpuInfo::AVX2) {
     num_processed = Insert_avx2(num_rows, hashes);
   }
@@ -181,7 +181,7 @@ void BlockedBloomFilter::Find(int64_t hardware_flags, int64_t num_rows,
                               bool enable_prefetch) const {
   int64_t num_processed = 0;
 
-#if defined(ARROW_HAVE_AVX2)
+#if defined(ARROW_HAVE_RUNTIME_AVX2)
   if (!(enable_prefetch && UsePrefetch()) &&
       (hardware_flags & arrow::internal::CpuInfo::AVX2)) {
     num_processed = Find_avx2(num_rows, hashes, result_bit_vector);
@@ -202,7 +202,7 @@ void BlockedBloomFilter::Find(int64_t hardware_flags, int64_t num_rows,
                               bool enable_prefetch) const {
   int64_t num_processed = 0;
 
-#if defined(ARROW_HAVE_AVX2)
+#if defined(ARROW_HAVE_RUNTIME_AVX2)
   if (!(enable_prefetch && UsePrefetch()) &&
       (hardware_flags & arrow::internal::CpuInfo::AVX2)) {
     num_processed = Find_avx2(num_rows, hashes, result_bit_vector);

diff --git a/cpp/src/arrow/acero/bloom_filter.h b/cpp/src/arrow/acero/bloom_filter.h
@@ -17,13 +17,14 @@
 
 #pragma once
 
-#if defined(ARROW_HAVE_AVX2)
+#if defined(ARROW_HAVE_RUNTIME_AVX2)
 #include <immintrin.h>
 #endif
 
 #include <atomic>
 #include <cstdint>
 #include <memory>
+
 #include "arrow/acero/partition_util.h"
 #include "arrow/acero/util.h"
 #include "arrow/memory_pool.h"
@@ -203,7 +204,7 @@ class ARROW_ACERO_EXPORT BlockedBloomFilter {
 
   void SingleFold(int num_folds);
 
-#if defined(ARROW_HAVE_AVX2)
+#if defined(ARROW_HAVE_RUNTIME_AVX2)
   inline __m256i mask_avx2(__m256i hash) const;
   inline __m256i block_id_avx2(__m256i hash) const;
   int64_t Insert_avx2(int64_t num_rows, const uint32_t* hashes);

diff --git a/cpp/src/arrow/acero/bloom_filter_avx2.cc b/cpp/src/arrow/acero/bloom_filter_avx2.cc
@@ -16,14 +16,13 @@
 // under the License.
 
 #include <immintrin.h>
+
 #include "arrow/acero/bloom_filter.h"
 #include "arrow/util/bit_util.h"
 
 namespace arrow {
 namespace acero {
 
-#if defined(ARROW_HAVE_AVX2)
-
 inline __m256i BlockedBloomFilter::mask_avx2(__m256i hash) const {
   // AVX2 translation of mask() method
   //
@@ -132,7 +131,5 @@ int64_t BlockedBloomFilter::Insert_avx2(int64_t num_rows, const uint64_t* hashes
   return InsertImp_avx2(num_rows, hashes);
 }
 
-#endif
-
 }  // namespace acero
 }  // namespace arrow