Optimize simd::leadingMask 10 times faster

Differential Revision: D51555527
facebookincubator · Nov 27, 2023 · 9915581 · 9915581
1 parent f354679
commit 9915581
Show file tree

Hide file tree

Showing 2 changed files with 60 additions and 8 deletions.
diff --git a/velox/common/base/SimdUtil-inl.h b/velox/common/base/SimdUtil-inl.h
@@ -242,19 +242,68 @@ int32_t indicesOfSetBits(
   return result - originalResult;
 }
 
+namespace detail {
+
 template <typename T, typename A>
-xsimd::batch_bool<T, A> leadingMask(int n, const A&) {
-  constexpr int N = xsimd::batch_bool<T, A>::size;
-  static const auto kMemo = ({
-    std::array<xsimd::batch_bool<T, A>, N> memo;
+struct LeadingMask {
+  LeadingMask() {
     bool tmp[N]{};
     for (int i = 0; i < N; ++i) {
-      memo[i] = xsimd::batch_bool<T, A>::load_unaligned(tmp);
+      memo_[i] = xsimd::batch_bool<T, A>::load_unaligned(tmp);
       tmp[i] = true;
     }
-    memo;
-  });
-  return LIKELY(n >= N) ? xsimd::batch_bool<T, A>(true) : kMemo[n];
+    memo_[N] = xsimd::batch_bool<T, A>::load_unaligned(tmp);
+  }
+
+  xsimd::batch_bool<T, A> operator[](size_t i) const {
+    return memo_[i];
+  }
+
+ private:
+  static constexpr int N = xsimd::batch_bool<T, A>::size;
+  std::array<xsimd::batch_bool<T, A>, N + 1> memo_;
+};
+
+extern const LeadingMask<int32_t, xsimd::default_arch> leadingMask32;
+extern const LeadingMask<int64_t, xsimd::default_arch> leadingMask64;
+
+template <typename T, typename A>
+xsimd::batch_bool<T, xsimd::default_arch> leadingMask(int i, const A&);
+
+template <>
+inline xsimd::batch_bool<int32_t, xsimd::default_arch> leadingMask(
+    int i,
+    const xsimd::default_arch&) {
+  return leadingMask32[i];
+}
+
+template <>
+inline xsimd::batch_bool<float, xsimd::default_arch> leadingMask(
+    int i,
+    const xsimd::default_arch&) {
+  return xsimd::batch_bool<float, xsimd::default_arch>(leadingMask32[i].data);
+}
+
+template <>
+inline xsimd::batch_bool<int64_t, xsimd::default_arch> leadingMask(
+    int i,
+    const xsimd::default_arch&) {
+  return leadingMask64[i];
+}
+
+template <>
+inline xsimd::batch_bool<double, xsimd::default_arch> leadingMask(
+    int i,
+    const xsimd::default_arch&) {
+  return xsimd::batch_bool<double, xsimd::default_arch>(leadingMask64[i].data);
+}
+
+} // namespace detail
+
+template <typename T, typename A>
+xsimd::batch_bool<T, A> leadingMask(int n, const A& arch) {
+  constexpr int N = xsimd::batch_bool<T, A>::size;
+  return detail::leadingMask<T, A>(std::min(n, N), arch);
 }
 
 namespace detail {

diff --git a/velox/common/base/SimdUtil.cpp b/velox/common/base/SimdUtil.cpp
@@ -24,6 +24,9 @@ namespace detail {
 alignas(kPadding) int32_t byteSetBits[256][8];
 alignas(kPadding) int32_t permute4x64Indices[16][8];
 
+const LeadingMask<int32_t, xsimd::default_arch> leadingMask32;
+const LeadingMask<int64_t, xsimd::default_arch> leadingMask64;
+
 } // namespace detail
 
 namespace {