perf: optimize ShortCode conversion

src/core/benchmark/codec.cc

@@ -197,8 +197,8 @@ static void ShortCodeDeserialize(benchmark::State &state) {
 
 static void ShortCodeToCommonCode(benchmark::State &state) {
 
-    // ShortCode::speed_up(true);
-    ShortCode::speed_up(false);
+    ShortCode::speed_up(true);
+    // ShortCode::speed_up(false);
 
     // ShortCode::fast_decode(4091296);
 
@@ -206,7 +206,9 @@ static void ShortCodeToCommonCode(benchmark::State &state) {
 
     for (auto _ : state) {
 
-        volatile auto kk = short_code.to_common_code();
+        // volatile auto kk = short_code.to_common_code();
+
+        benchmark::DoNotOptimize(short_code.to_common_code());
 
         // if (AllCases::instance().is_available()) {
         // if (ShortCode::stage_ == ShortCode::Stage::FAST) {
@@ -218,6 +220,29 @@ static void ShortCodeToCommonCode(benchmark::State &state) {
 
 }
 
+static void CommonCodeToShortCode(benchmark::State &state) {
+    ShortCode::speed_up(true);
+    // ShortCode::speed_up(false);
+
+    auto common_code = CommonCode::unsafe_create(0x1A9BF0C00);
+
+    // std::vector<CommonCode> samples;
+    // for (auto code : common_code_samples(256)) {
+    //     samples.emplace_back(CommonCode::unsafe_create(code));
+    // }
+
+    for (auto _ : state) {
+
+        // for (auto common_code : samples) {
+            // volatile auto kk = ShortCode(common_code);
+
+            benchmark::DoNotOptimize(ShortCode(common_code));
+
+        // }
+
+    }
+}
+
 // BENCHMARK(CommonCodeSerialize)->Range(8, 256);
 // BENCHMARK(CommonCodeDeserialize)->Range(8, 256);
 // BENCHMARK(CommonCodeSerializeShorten)->Range(8, 256);
@@ -226,7 +251,8 @@ static void ShortCodeToCommonCode(benchmark::State &state) {
 // BENCHMARK(ShortCodeSerialize)->Range(8, 256);
 // BENCHMARK(ShortCodeDeserialize)->Range(8, 256);
 
-BENCHMARK(ShortCodeToCommonCode);
+// BENCHMARK(ShortCodeToCommonCode);
+BENCHMARK(CommonCodeToShortCode);
 
 // static void CommonCodeDecode(benchmark::State &state) {
 //     const auto tmp = str_common_codes(state.range(0));

src/core/common_code/common_code.h

@@ -166,6 +166,15 @@ private:
     // ------------------------------------------------------------------------------------- //
 };
 
+// Ref: https://cplusplus.github.io/CWG/issues/1734.html
+
+// TODO: By definition, the default constructor is deleted, it is not trivial.
+// TODO: But in clang and g++, it is legal, but in msvc it fails.
+static_assert(std::is_trivial_v<CommonCode>);
+
+static_assert(std::is_standard_layout_v<CommonCode>);
+static_assert(std::is_trivially_copyable_v<CommonCode>);
+
 } // namespace klotski::codec
 
 #include "internal/common_code.inl"

src/core/short_code/internal/convert.cc

@@ -50,18 +50,10 @@ static uint32_t check_range(uint32_t head, uint32_t range) noexcept {
     return 0; // pass check
 }
 
-std::mutex ShortCode::busy_ {};
-
-// ShortCode::Stage ShortCode::stage_ = Stage::UNINIT;
-
-// const klotski::cases::RangesUnion *ShortCode::cases_ = &AllCases::instance().fetch();
-const klotski::cases::RangesUnion *ShortCode::cases_ = nullptr;
-
-const klotski::cases::Ranges *ShortCode::ranges_ = nullptr;
-
 uint32_t ShortCode::fast_encode(uint64_t common_code) {
     auto head = common_code >> 32;
-    auto &ranges = AllCases::instance().fetch()[head]; // match available ranges
+    auto &ranges = (*cases_)[head]; // match available ranges
+    // TODO: try to narrow the scope by prefix
     auto target = std::lower_bound(ranges.begin(), ranges.end(), (uint32_t)common_code);
     return ALL_CASES_OFFSET[head] + (target - ranges.begin());
 }
@@ -79,19 +71,22 @@ uint32_t ShortCode::tiny_encode(uint64_t common_code) {
 
     uint32_t offset = 0;
     auto index = RANGES_GLOBAL_OFFSET[prefix];
-    const auto &basic_ranges = BasicRanges::instance().fetch();
+
+    ranges_ = &cases::BasicRanges::instance().fetch();
+    const auto &ranges = *ranges_;
+
     auto target = (uint32_t)common_code; // target range
-    for (; index < basic_ranges.size(); ++index) {
-        auto broken_offset = check_range(head, range_reverse(basic_ranges[index]));
+    for (; index < ranges.size(); ++index) {
+        auto broken_offset = check_range(head, range_reverse(ranges[index]));
         if (!broken_offset) { // valid case
-            if (basic_ranges[index] == target) {
+            if (ranges[index] == target) {
                 break; // found target range
             }
             ++offset; // record sub offset
         } else {
             auto delta = (uint32_t)1 << (32 - broken_offset * 2); // delta to next possible range
-            auto next_min = (basic_ranges[index] & ~(delta - 1)) + delta;
-            while (basic_ranges[++index] < next_min); // located next range
+            auto next_min = (ranges[index] & ~(delta - 1)) + delta;
+            while (ranges[++index] < next_min); // located next range
             --index;
         }
     }
@@ -99,13 +94,6 @@ uint32_t ShortCode::tiny_encode(uint64_t common_code) {
 }
 
 uint64_t ShortCode::tiny_decode(uint32_t short_code) { // short code --> common code
-    // speed_up(false);
-
-    // std::lock_guard guard {busy_};
-
-    ranges_ = &cases::BasicRanges::instance().fetch();
-    // stage_ = Stage::TINY;
-
     auto offset_ = std::upper_bound(ALL_CASES_OFFSET.begin(), ALL_CASES_OFFSET.end(), short_code) - 1;
     auto head = offset_ - ALL_CASES_OFFSET.begin(); // head index
     short_code -= *offset_;
@@ -116,11 +104,9 @@ uint64_t ShortCode::tiny_decode(uint32_t short_code) { // short code --> common
 
     /// search for target range
     auto index = RANGES_GLOBAL_OFFSET[prefix];
-    // auto basic_ranges = ranges_;
-    // const auto &basic_ranges = BasicRanges::instance().fetch();
 
+    ranges_ = &cases::BasicRanges::instance().fetch();
     const auto &ranges = *ranges_;
-    // const auto &ranges = BasicRanges::instance().fetch();
 
     for (; index < ranges.size(); ++index) { // traverse basic ranges
         auto broken_offset = check_range(head, range_reverse(ranges[index]));

src/core/short_code/internal/short_code.inl

@@ -9,8 +9,7 @@ namespace klotski::codec {
 // ------------------------------------------------------------------------------------- //
 
 inline ShortCode::ShortCode(const CommonCode common_code) {
-    // TODO: test the affect of CPU branch prediction.
-    if (cases::AllCases::instance().is_available()) {
+    if (fast_) {
         code_ = fast_encode(common_code.unwrap());
     } else {
         code_ = tiny_encode(common_code.unwrap());
@@ -39,19 +38,12 @@ inline bool ShortCode::check(const uint32_t short_code) {
 }
 
 inline void ShortCode::speed_up(const bool fast_mode) {
-    // TODO: keep one way change.
+    ranges_ = &cases::BasicRanges::instance().fetch();
     if (fast_mode) {
-        // cases::AllCases::instance().build();
         std::lock_guard guard {busy_};
         cases_ = &cases::AllCases::instance().fetch();
-        stage_ = Stage::FAST;
-    } else {
-        std::lock_guard guard {busy_};
-
-        // TODO: skip if stage_ is FAST
-
-        ranges_ = &cases::BasicRanges::instance().fetch();
-        stage_ = Stage::TINY;
+        KLSK_MEM_BARRIER;
+        fast_ = true;
     }
 }
 
@@ -73,21 +65,10 @@ inline std::string ShortCode::to_string() const {
 }
 
 inline CommonCode ShortCode::to_common_code() const {
-    // TODO: test the affect of CPU branch prediction.
-    // if (cases::AllCases::instance().is_available()) {
-    //     return CommonCode::unsafe_create(fast_decode(code_));
-    // }
-    // return CommonCode::unsafe_create(tiny_decode(code_));
-
-    switch (stage_) {
-    case Stage::UNINIT:
-        // TODO: do speed up
-        // speed_up(false); // FIXME: slow about 3%
-    case Stage::TINY:
-        return CommonCode::unsafe_create(tiny_decode(code_));
-    case Stage::FAST:
+    if (fast_) {
         return CommonCode::unsafe_create(fast_decode(code_));
     }
+    return CommonCode::unsafe_create(tiny_decode(code_));
 }
 
 // ----------------------------------------------------------------------------------------- //

src/core/short_code/short_code.h

@@ -178,14 +178,19 @@ private:
 
     // ------------------------------------------------------------------------------------- //
 
-    enum class Stage { UNINIT, TINY, FAST };
+    /// Whether fast mode is available.
+    static inline bool fast_ {false};
 
-    static std::mutex busy_;
+    /// Mutex for protecting critical section.
+    static inline std::mutex busy_ {};
 
-    static inline auto stage_ {Stage::UNINIT};
+    /// Static pointer to klotski AllCases data.
+    static inline const cases::RangesUnion *cases_ {nullptr};
 
-    static const cases::Ranges *ranges_;
-    static const cases::RangesUnion *cases_;
+    /// Static pointer to klotski BasicRanges data.
+    static inline std::atomic<const cases::Ranges*> ranges_ {nullptr};
+
+    // ------------------------------------------------------------------------------------- //
 };
 
 } // namespace klotski::codec

src/core/utils/utility.h

@@ -25,6 +25,9 @@
 /// Force function declaration to be inline.
 #define KLSK_INLINE __attribute__ ((always_inline))
 
+/// Prevent reordering for both compiler and processor.
+#define KLSK_MEM_BARRIER std::atomic_thread_fence(std::memory_order_seq_cst)
+
 namespace klotski {
 
 /// Calculate the sum of an array of integers.