Add cpu counter

include/xrpl/beast/core/FunctionProfiler.h

@@ -2,48 +2,79 @@
 #ifndef RIPPLE_BASICS_FUNCTIONPROFILER_H_INCLUDED
 #define RIPPLE_BASICS_FUNCTIONPROFILER_H_INCLUDED
 
-#include <string>
 #include <chrono>
-#include <unordered_map>
-#include <sstream>
-#include <source_location>
 #include <csignal>
+#include <mutex>
+#include <source_location>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <x86intrin.h>
 
 namespace beast {
 
-    void logProfilingResults();
+void
+logProfilingResults();
+
 class FunctionProfiler
 {
+public:
     std::string functionName;
     std::chrono::steady_clock::time_point start;
-public:
+    std::uint64_t cpuCycleStart;
+    inline static std::mutex mutex_;
 
-    inline static std::unordered_map<std::string, std::pair<std::chrono::nanoseconds, std::int64_t>> funcionDurations;
+    struct StatisticData
-    FunctionProfiler(const std::string& tag, std::source_location location = std::source_location::current()): functionName(location.function_name() + tag), start(std::chrono::steady_clock::now()) 
+    {
+        std::chrono::nanoseconds timeInTotal;
+        std::uint64_t cpuCyclesInTotal;
+        std::int64_t count;
+    };
+
+    inline static std::unordered_map<
+        std::string,
+        StatisticData>
+        funcionDurations;
+    FunctionProfiler(
+        std::string const& tag,
+        std::source_location location = std::source_location::current())
+        : functionName(location.function_name() + tag)
+        , start(std::chrono::steady_clock::now())
+        , cpuCycleStart(__rdtsc())
     {
     }
 
     ~FunctionProfiler() noexcept
     {
         auto duration = std::chrono::steady_clock::now() - start;
-        funcionDurations[functionName].first += std::chrono::duration_cast<std::chrono::nanoseconds>(duration);
+        std::lock_guard<std::mutex> lock{mutex_};
-        funcionDurations[functionName].second++;
+        if (funcionDurations[functionName].count ==
+            std::numeric_limits<std::int64_t>::max())
+        {
+            return;
+        }
+        funcionDurations[functionName].timeInTotal += duration;
+        funcionDurations[functionName].cpuCyclesInTotal += (__rdtsc() - cpuCycleStart);
+        funcionDurations[functionName].count++;
     }
 };
 
-
+inline std::string
-inline std::string getProfilingResults()
+getProfilingResults()
 {
+    std::lock_guard<std::mutex> lock{FunctionProfiler::mutex_};
     std::stringstream ss;
     ss << "Function profiling results:" << std::endl;
-    for (const auto& [name, duration] : FunctionProfiler::funcionDurations)
+    ss << "name,time,count" << std::endl;
+    for (auto const& [name, duration] : FunctionProfiler::funcionDurations)
     {
-        ss << "  " << name << ": " << duration.first.count() << " ns" << ", counts: " << duration.second << std::endl;
+        ss << name << "," << duration.first.count() << " ns" << ","
+           << duration.second << std::endl;
     }
 
     return ss.str();
 }
 
-}
+}  // namespace beast
 
-#endif
+#endif

include/xrpl/beast/hash/xxhasher.h

@@ -20,15 +20,17 @@
 #ifndef BEAST_HASH_XXHASHER_H_INCLUDED
 #define BEAST_HASH_XXHASHER_H_INCLUDED
 
+#include <xrpl/beast/core/FunctionProfiler.h>
+
 #include <boost/endian/conversion.hpp>
 
 #include <xxhash.h>
 
 #include <cstddef>
+#include <iostream>
 #include <new>
-#include <type_traits>
-#include <xrpl/beast/core/FunctionProfiler.h>
 #include <span>
+#include <type_traits>
 
 namespace beast {
 
@@ -38,16 +40,29 @@ private:
     // requires 64-bit std::size_t
     static_assert(sizeof(std::size_t) == 8, "");
 
-    XXH3_state_t* state_;
+    struct state_wrapper
+    {
+        XXH3_state_t* state;
+        state_wrapper()
+        {
+            state = XXH3_createState();
+        }
+        ~state_wrapper()
+        {
+            XXH3_freeState(state);
+        }
+    };
 
-    std::array<std::uint8_t, 40> buffer_;
+    XXH3_state_t* state_;
-    std::span<std::uint8_t> readSpan_;
+    inline static thread_local state_wrapper wrapper{};
-    std::span<std::uint8_t> writeSpan_;
+    std::size_t totalSize_ = 0;
-    XXH64_hash_t seed_{};
+    std::chrono::nanoseconds duration_{};
+    std::uint64_t cpuCycles = 0;
 
     static XXH3_state_t*
     allocState()
     {
+        FunctionProfiler _{"-alloc"};
         auto ret = XXH3_createState();
         if (ret == nullptr)
             throw std::bad_alloc();
@@ -65,14 +80,24 @@ public:
 
     xxhasher()
     {
-        state_ = allocState();
+        auto start = std::chrono::steady_clock::now();
-        XXH3_64bits_reset(state_);
+        auto cpuCyclesStart = __rdtsc();
-        writeSpan_ = buffer_;
+        // state_ = allocState();
+        // XXH3_64bits_reset(state_);
+        duration_ += std::chrono::steady_clock::now() - start;
+        cpuCycles += (__rdtsc() - cpuCyclesStart);
+        XXH3_64bits_reset(wrapper.state);
     }
 
     ~xxhasher() noexcept
     {
-        XXH3_freeState(state_);
+        // profiler_.functionName = "xxhasher-" + std::to_string(totalSize_);
+        // auto start = std::chrono::steady_clock::now();
+        if (0)
+        {
+            FunctionProfiler _{"-free"};
+            XXH3_freeState(state_);
+        }
     }
 
     template <
@@ -80,10 +105,13 @@ public:
         std::enable_if_t<std::is_unsigned<Seed>::value>* = nullptr>
     explicit xxhasher(Seed seed)
     {
-        state_ = allocState();
+        auto start = std::chrono::steady_clock::now();
-        XXH3_64bits_reset_withSeed(state_, seed);
+        auto cpuCyclesStart = __rdtsc();
-        seed_ = seed;
+        // state_ = allocState();
-        writeSpan_ = buffer_;
+        // XXH3_64bits_reset_withSeed(state_, seed);
+        XXH3_64bits_reset_withSeed(wrapper.state, seed);
+        duration_ += std::chrono::steady_clock::now() - start;
+        cpuCycles += (__rdtsc() - cpuCyclesStart);
     }
 
     template <
@@ -91,33 +119,53 @@ public:
         std::enable_if_t<std::is_unsigned<Seed>::value>* = nullptr>
     xxhasher(Seed seed, Seed)
     {
-        state_ = allocState();
+        auto start = std::chrono::steady_clock::now();
-        XXH3_64bits_reset_withSeed(state_, seed);
+        auto cpuCyclesStart = __rdtsc();
-        seed_ = seed;
+        // state_ = allocState();
-        writeSpan_ = buffer_;
+        // XXH3_64bits_reset_withSeed(state_, seed);
+        XXH3_64bits_reset_withSeed(wrapper.state, seed);
+        duration_ += std::chrono::steady_clock::now() - start;
+        cpuCycles += (__rdtsc() - cpuCyclesStart);
     }
 
     void
     operator()(void const* key, std::size_t len) noexcept
     {
-        FunctionProfiler _{"-size-" + std::to_string(len)};
+        auto start = std::chrono::steady_clock::now();
-        //XXH3_64bits_update(state_, key, len);
+        auto cpuCyclesStart = __rdtsc();
-        if (len >= writeSpan_.size())
+        totalSize_ += len;
-        {
+        // FunctionProfiler _{"-size-" + std::to_string(len)};
-            exit(-1);
+        // XXH3_64bits_update(state_, key, len);
-        }
+        XXH3_64bits_update(wrapper.state, key, len);
-        std::memcpy(writeSpan_.data(), key, len);
+        duration_ += std::chrono::steady_clock::now() - start;
-        readSpan_ = std::span<std::uint8_t>(std::begin(buffer_), readSpan_.size() + len);
+        cpuCycles += (__rdtsc() - cpuCyclesStart);
-        writeSpan_ = std::span<std::uint8_t>(std::begin(buffer_) + readSpan_.size(), buffer_.size() - readSpan_.size());
     }
 
     explicit
     operator std::size_t() noexcept
     {
-        FunctionProfiler _{"-digest"};
+        auto start = std::chrono::steady_clock::now();
-        // XXH3_64bits_update(state_, readSpan_.data(), readSpan_.size());
+        // auto ret =  XXH3_64bits_digest(state_);
-        // return XXH3_64bits_digest(state_);
+        auto ret = XXH3_64bits_digest(wrapper.state);
-        return XXH64(readSpan_.data(), readSpan_.size(), seed_);
+        duration_ += std::chrono::steady_clock::now() - start;
+
+        if (FunctionProfiler::funcionDurations
+                ["xxhasher-" + std::to_string(totalSize_)]
+                    .count == std::numeric_limits<std::int64_t>::max())
+        {
+            return ret;
+        }
+        std::lock_guard<std::mutex> lock{FunctionProfiler::mutex_};
+        FunctionProfiler::funcionDurations
+            ["xxhasher-" + std::to_string(totalSize_)]
+                .timeInTotal += duration_;
+        FunctionProfiler::funcionDurations
+            ["xxhasher-" + std::to_string(totalSize_)]
+                .cpuCyclesInTotal += cpuCycles;
+        FunctionProfiler::funcionDurations
+            ["xxhasher-" + std::to_string(totalSize_)]
+                .count++;
+        return ret;
     }
 };