Optimize simd::leadingMask 10 times faster (#7736)

Summary: Pull Request resolved: #7736

Reviewed By: oerling

Differential Revision: D51555527

fbshipit-source-id: a4fc6d60f23cbe31c7fb434ad9b56266e9663265

velox/common/base/SimdUtil-inl.h

@@ -242,19 +242,72 @@ 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;
+  xsimd::batch_bool<T, A> memo_[N + 1];
+};
+
+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 reinterpret_cast<
+      xsimd::batch_bool<float, xsimd::default_arch>::register_type>(
+      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 reinterpret_cast<
+      xsimd::batch_bool<double, xsimd::default_arch>::register_type>(
+      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 {

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 {

velox/common/base/benchmarks/SimdUtilBenchmark.cpp

@@ -0,0 +1,66 @@
+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "velox/common/base/SimdUtil.h"
+
+#include <folly/Benchmark.h>
+#include <folly/init/Init.h>
+
+#include <random>
+
+namespace facebook::velox {
+namespace {
+
+#define VELOX_BENCHMARK(_type, _name, ...) \
+  _type _name(FOLLY_PP_STRINGIZE(_name), __VA_ARGS__)
+
+template <typename T>
+class LeadingMask {
+ public:
+  LeadingMask(const char* name, std::default_random_engine& gen) {
+    std::uniform_int_distribution<> dist(xsimd::batch<T>::size + 1);
+    for (int i = 0; i < kSize; ++i) {
+      inputs_[i] = dist(gen);
+    }
+    folly::addBenchmark(__FILE__, name, [this] { return run(); });
+  }
+
+ private:
+  unsigned run() {
+    xsimd::batch_bool<T> ans = {};
+    for (int i = 0; i < kSize; ++i) {
+      ans = ans ^ simd::leadingMask<T>(inputs_[i]);
+    }
+    folly::doNotOptimizeAway(ans);
+    return kSize;
+  }
+
+  static constexpr int kSize = 4 << 10;
+  int8_t inputs_[kSize];
+};
+
+} // namespace
+} // namespace facebook::velox
+
+int main(int argc, char* argv[]) {
+  using namespace facebook::velox;
+  folly::Init follyInit(&argc, &argv);
+  std::default_random_engine gen(std::random_device{}());
+  VELOX_BENCHMARK(LeadingMask<int32_t>, leadingMaskInt32, gen);
+  VELOX_BENCHMARK(LeadingMask<int64_t>, leadingMaskInt64, gen);
+  folly::runBenchmarks();
+  return 0;
+}