GH-55313: [C++][Parquet] SIMD: optimize parquet GreaterThan on Neon

cpp/src/parquet/CMakeLists.txt

@@ -204,11 +204,7 @@ append_runtime_avx2_src(PARQUET_SRCS level_comparison_simd_256.cc)
 
 if(ARROW_HAVE_RUNTIME_AVX2)
   # AVX2 is used as a proxy for BMI2.
-  list(APPEND
-       PARQUET_SRCS
-       level_comparison_avx2.cc
-       level_conversion_bmi2.cc
-       bloom_filter_avx2.cc)
+  list(APPEND PARQUET_SRCS level_conversion_bmi2.cc bloom_filter_avx2.cc)
   # We need CMAKE_CXX_FLAGS_RELEASE here to prevent the one-definition-rule
   # violation with -DCMAKE_BUILD_TYPE=MinSizeRel. CMAKE_CXX_FLAGS_RELEASE
   # will force inlining as much as possible.
@@ -218,8 +214,8 @@ if(ARROW_HAVE_RUNTIME_AVX2)
     separate_arguments(RELEASE_FLAGS NATIVE_COMMAND "${CMAKE_CXX_FLAGS_RELEASE}")
     list(APPEND AVX2_FLAGS ${RELEASE_FLAGS})
   endif()
-  set_source_files_properties(level_comparison_avx2.cc bloom_filter_avx2.cc
-                              PROPERTIES COMPILE_OPTIONS "${AVX2_FLAGS}")
+  set_source_files_properties(bloom_filter_avx2.cc PROPERTIES COMPILE_OPTIONS
+                                                              "${AVX2_FLAGS}")
   # WARNING: DO NOT BLINDLY COPY THIS CODE FOR OTHER BMI2 USE CASES.
   # This code is always guarded by runtime dispatch which verifies
   # BMI2 is present.  For a very small number of CPUs AVX2 does not

cpp/src/parquet/level_comparison.cc

@@ -18,11 +18,6 @@
 #include <array>
 
 #include "parquet/level_comparison.h"
-
-#if defined(ARROW_HAVE_RUNTIME_AVX2)
-#  include "parquet/level_comparison_avx2_internal.h"
-#endif
-
 #include "parquet/level_comparison_simd_internal.h"
 
 #define PARQUET_IMPL_NAMESPACE standard
@@ -45,6 +40,8 @@ struct GreaterThanDynamicFunction {
   static constexpr auto targets() {
     return std::array{
         ARROW_DISPATCH_TARGET_NONE(&standard::GreaterThanBitmapImpl)  //
+        ARROW_DISPATCH_TARGET_NEON(&GreaterThanBitmapNeon)            //
+        ARROW_DISPATCH_TARGET_SSE4_2(&GreaterThanBitmapSse42)         //
         ARROW_DISPATCH_TARGET_AVX2(&GreaterThanBitmapAvx2)            //
     };
   }

cpp/src/parquet/level_comparison_simd_128.cc

@@ -17,9 +17,11 @@
 
 #if defined(ARROW_HAVE_NEON)
 #  define FIND_MIN_MAX_PLATFORM FindMinMaxNeon
+#  define GREATER_THAN_BITMAP_PLATFORM GreaterThanBitmapNeon
 #  define ARCH_PLATFORM xsimd::neon64
 #elif defined(ARROW_HAVE_SSE4_2) || defined(ARROW_HAVE_RUNTIME_SSE4_2)
 #  define FIND_MIN_MAX_PLATFORM FindMinMaxSse42
+#  define GREATER_THAN_BITMAP_PLATFORM GreaterThanBitmapSse42
 #  define ARCH_PLATFORM xsimd::sse4_2
 #endif
 
@@ -34,8 +36,14 @@ MinMax FIND_MIN_MAX_PLATFORM(const int16_t* levels, int64_t num_levels) {
   return FindMinMaxSimd<ARCH_PLATFORM>(levels, num_levels);
 }
 
+uint64_t GREATER_THAN_BITMAP_PLATFORM(const int16_t* levels, int64_t num_levels,
+                                      int16_t rhs) {
+  return GreaterThanBitmapSimd<ARCH_PLATFORM>(levels, num_levels, rhs);
+}
+
 }  // namespace parquet::internal
 
 #  undef ARCH_PLATFORM
+#  undef GREATER_THAN_BITMAP_PLATFORM
 #  undef FIND_MIN_MAX_PLATFORM
 #endif  // FIND_MIN_MAX_PLATFORM

cpp/src/parquet/level_comparison_simd_256.cc

@@ -17,6 +17,7 @@
 
 #if defined(ARROW_HAVE_RUNTIME_AVX2)
 #  define FIND_MIN_MAX_PLATFORM FindMinMaxAvx2
+#  define GREATER_THAN_BITMAP_PLATFORM GreaterThanBitmapAvx2
 #  define ARCH_PLATFORM xsimd::avx2
 #endif
 
@@ -33,7 +34,13 @@ MinMax FIND_MIN_MAX_PLATFORM(const int16_t* levels, int64_t num_levels) {
   return FindMinMaxSimd<ARCH_PLATFORM>(levels, num_levels);
 }
 
+uint64_t GREATER_THAN_BITMAP_PLATFORM(const int16_t* levels, int64_t num_levels,
+                                      int16_t rhs) {
+  return GreaterThanBitmapSimd<ARCH_PLATFORM>(levels, num_levels, rhs);
+}
+
 }  // namespace parquet::internal
 
 #undef ARCH_PLATFORM
+#undef GREATER_THAN_BITMAP_PLATFORM
 #undef FIND_MIN_MAX_PLATFORM

cpp/src/parquet/level_comparison_simd_internal.h

@@ -25,14 +25,17 @@ namespace parquet::internal {
 
 #if defined(ARROW_HAVE_NEON)
 MinMax FindMinMaxNeon(const int16_t* levels, int64_t num_levels);
+uint64_t GreaterThanBitmapNeon(const int16_t* levels, int64_t num_levels, int16_t rhs);
 #endif
 
 #if defined(ARROW_HAVE_SSE4_2) || defined(ARROW_HAVE_RUNTIME_SSE4_2)
 MinMax FindMinMaxSse42(const int16_t* levels, int64_t num_levels);
+uint64_t GreaterThanBitmapSse42(const int16_t* levels, int64_t num_levels, int16_t rhs);
 #endif
 
 #if defined(ARROW_HAVE_RUNTIME_AVX2)
 MinMax FindMinMaxAvx2(const int16_t* levels, int64_t num_levels);
+uint64_t GreaterThanBitmapAvx2(const int16_t* levels, int64_t num_levels, int16_t rhs);
 #endif
 
 }  // namespace parquet::internal

cpp/src/parquet/level_comparison_simd_kernel_internal.h

@@ -23,6 +23,7 @@
 #include <cstdint>
 #include <limits>
 
+#include "arrow/util/endian.h"
 #include "parquet/level_comparison.h"
 
 namespace parquet::internal {
@@ -53,4 +54,26 @@ MinMax FindMinMaxSimd(const int16_t* levels, int64_t num_levels) {
   return out;
 }
 
+/// Build a bitmap (LSB first, little-endian) with bit i set iff levels[i] > rhs.
+///
+/// \param[in] num_levels must be in [0, 64].
+template <typename Arch>
+uint64_t GreaterThanBitmapSimd(const int16_t* levels, int64_t num_levels, int16_t rhs) {
+  using batch = xsimd::batch<int16_t, Arch>;
+  constexpr int64_t kLanes = static_cast<int64_t>(batch::size);
+  static_assert(kLanes <= 64);
+
+  const batch vrhs(rhs);
+  uint64_t mask = 0;
+  int64_t i = 0;
+  for (; i + kLanes <= num_levels; i += kLanes) {
+    const auto gt = batch::load_unaligned(levels + i) > vrhs;
+    mask |= static_cast<uint64_t>(gt.mask()) << i;
+  }
+  for (; i < num_levels; ++i) {
+    mask |= static_cast<uint64_t>(levels[i] > rhs) << i;
+  }
+  return ::arrow::bit_util::ToLittleEndian(mask);
+}
+
 }  // namespace parquet::internal