avx2 optimized

CMakeLists.txt

@@ -1,11 +1,12 @@
 cmake_minimum_required(VERSION 3.31)
 project(yggm)
+set(CMAKE_CXX_STANDARD 20)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
 find_package(OpenSSL REQUIRED)
 file(GLOB ${PROJECT_NAME}SOURCES *.cpp)
 file(GLOB ${PROJECT_NAME}HEADERS *.h)
 add_executable(${PROJECT_NAME} sources/main.cpp)
-set(CXX_ADDITIONAL_FLAGS "-mavx2 -fomit-frame-pointer -ftree-vectorize -ftree-slp-vectorize -fdelete-null-pointer-checks -fno-exceptions -fno-rtti -pthread -fomit-frame-pointer")
+set(CXX_ADDITIONAL_FLAGS "-mavx2 -fomit-frame-pointer -ftree-vectorize -ftree-slp-vectorize -fdelete-null-pointer-checks -fno-exceptions -fno-rtti -pthread")
 set(CMAKE_CXX_FLAGS_RELEASE "-march=native -O3 -ffast-math -pipe -Wall -Wextra -Wpedantic -Wconversion -Wuninitialized -Wsign-conversion -flto=full")
-target_link_libraries(${PROJECT_NAME} pthread)
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} ${CXX_ADDITIONAL_FLAGS}")
-target_link_libraries(${PROJECT_NAME} sodium)
+target_link_libraries(${PROJECT_NAME} pthread sodium)
-#target_link_libraries(${PROJECT_NAME} OpenSSL::SSL OpenSSL::Crypto)

sources/main.cpp

@@ -1,4 +1,3 @@
-#include <sodium.h>
 #include <iostream>
 #include <sstream>
 #include <fstream>
@@ -6,16 +5,17 @@
 #include <vector>
 #include <array>
 #include <string>
+#include <random>
 #include <memory.h>
 #include <thread>
 #include <regex>
-#include <atomic>
+#include <sodium.h>
 #include <arpa/inet.h>
 #include <immintrin.h>
 struct option {
     unsigned int proc = 0;
     bool log = true;
-    int high = 20;
+    int high = 14;
     std::string outputfile;
 };
 static option conf;
@@ -48,12 +48,12 @@ void displayConfig() {
     std::cout << " Threads: " << conf.proc << ", " << "high addresses (2" << std::setw(2) << std::setfill('0') << std::hex << conf.high << "+)" << std::dec << std::endl;
 }
 using Address = unsigned char[16];
-std::string getAddress(const Address& rawAddr) {
+[[gnu::unused]] std::string getAddress(const Address& rawAddr) noexcept {
     char ipStrBuf[46];
     inet_ntop(AF_INET6, rawAddr, ipStrBuf, 46);
     return std::string(ipStrBuf);
 }
-inline std::string keyToString(const unsigned char* key) {
+inline std::string keyToString(const unsigned char* key) noexcept {
     std::string result;
     result.resize(64);
     const char* hexDigits = "0123456789abcdef";
@@ -68,38 +68,68 @@ struct KeysBox {
     Key PublicKey;
     Key PrivateKey;
 };
-/*
+[[gnu::always_inline]] [[gnu::unused]] inline void bitwiseInverse(const unsigned char* key, Key& inverted) noexcept {
-[[gnu::always_inline]] inline void bitwiseInverse(const unsigned char* key, Key& inverted) noexcept {
     __m256i chunk = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(key));
     chunk = _mm256_xor_si256(chunk, _mm256_set1_epi8(0xFF));
     _mm256_storeu_si256(reinterpret_cast<__m256i*>(inverted), chunk);
 }
-*/
+[[nodiscard]] inline unsigned char zeroCounter(unsigned int x) noexcept {
+#if defined(__GNUC__) || defined(__clang__)
+    if (x == 0)
+        return 32;
+    return static_cast<unsigned char>(__builtin_clz(x));
+#else
+    unsigned char zeros = 0;
+    for (uint32_t mask = 0x80000000; mask; mask >>= 1) {
+        if (x & mask) break;
+        zeros++;
+    }
+    return zeros;
+#endif
+}
 [[nodiscard]] inline unsigned char getZeros(const Key& value) noexcept {
     unsigned char leadZeros = 0;
 #pragma unroll
-    for (unsigned char i = 0; i < 6; i++) {
+    for (unsigned char i = 0; i < 32; i += 4) {
-        if (value[i] == 0x00) {
+        unsigned int word =
-            leadZeros += 8;
+            (static_cast<unsigned int>(value[i]) << 24) |
+            (static_cast<unsigned int>(value[i + 1]) << 16) |
+            (static_cast<unsigned int>(value[i + 2]) << 8) |
+            (static_cast<unsigned int>(value[i + 3]));
+        if (word == 0) {
+            leadZeros += 32;
         } else {
-            unsigned char byte = value[i];
+            leadZeros += zeroCounter(word);
-            while (!(byte & 0x80)) {
+            break;
-                leadZeros++;
-                byte <<= 1;
-            }
-            return leadZeros;
         }
     }
     return leadZeros;
 }
+inline long long xorshift64(unsigned long& state) {
+    state ^= state << 21;
+    state ^= state >> 35;
+    state ^= state << 4;
+    return static_cast<long long>(state * 2685821657736338717);
+}
+inline void randombytesavx2(unsigned char* buf, unsigned char size, unsigned long& state) noexcept {
+#pragma unroll
+    for (unsigned char x = 0; x < size; x += 32) {
+        _mm256_storeu_si256((__m256i*) & buf[x], _mm256_set_epi64x(xorshift64(state), xorshift64(state), xorshift64(state), xorshift64(state)));
+    }
+}
 void miner_thread() noexcept {
     KeysBox keys;
     unsigned char ones = 0;
+    unsigned char seed[32];
+    std::random_device rd;
+    unsigned long state = static_cast<unsigned long>(rd());
     while (true) {
-        crypto_sign_ed25519_keypair(keys.PublicKey, keys.PrivateKey);
+        randombytesavx2(seed, sizeof(seed) / sizeof(seed[0]), state);
+        crypto_sign_ed25519_seed_keypair(keys.PublicKey, keys.PrivateKey, seed);
         ones = getZeros(keys.PublicKey);
         if (ones > conf.high) {
-            printf("\nAddr:\t2%x/%d\nPK:\t%s\nSK:\t%s\n", conf.high = ones, ones, keyToString(keys.PublicKey).c_str(), keyToString(keys.PrivateKey).c_str());
+            conf.high = ones;
+            printf("\nAddr:\t2%x/%d\nPK:\t%s\nSK:\t%s\n", ones, ones, keyToString(keys.PublicKey).c_str(), keyToString(keys.PrivateKey).c_str());
         }
     }
 }
@@ -112,7 +142,7 @@ void startThreads() noexcept {
         thread.join();
     }
 }
-int main(int argc, char* argv[]) {
+int main(int argc, char* argv[]) noexcept {
     if (argc >= 2) {
         int res = -1;
         for (int i = 1;; ++i) {
@@ -132,6 +162,7 @@ int main(int argc, char* argv[]) {
             }
         }
     }
+    if (sodium_init() < 0) return -1;
     displayConfig();
     startThreads();
     return 0;