Merge pull request #29 from dave-andersen/master

Hugepages for Linux speed optimization + fixing MacOS build

src/crypto/alt/KeccakF-1600-opt64.c

@@ -24,11 +24,11 @@ typedef unsigned char UINT8;
 
 
 #if defined(__GNUC__)
-#define ALIGN __attribute__ ((aligned(32)))
+#define KALIGN __attribute__ ((aligned(32)))
 #elif defined(_MSC_VER)
-#define ALIGN __declspec(align(32))
+#define KALIGN __declspec(align(32))
 #else
-#define ALIGN
+#define KALIGN
 #endif
 
 #if defined(UseSSE)
@@ -67,7 +67,7 @@ typedef unsigned char UINT8;
     #ifdef UseOnlySIMD64
     #include "KeccakF-1600-simd64.macros"
     #else
-ALIGN const UINT64 rho8_56[2] = {0x0605040302010007, 0x080F0E0D0C0B0A09};
+KALIGN const UINT64 rho8_56[2] = {0x0605040302010007, 0x080F0E0D0C0B0A09};
     #include "KeccakF-1600-simd128.macros"
     #endif
 
@@ -105,18 +105,18 @@ ALIGN const UINT64 rho8_56[2] = {0x0605040302010007, 0x080F0E0D0C0B0A09};
 
     #define ROL6464same(a, o)   _mm_roti_epi64(a, o)
     #define ROL6464(a, r1, r2)  _mm_rot_epi64(a, CONST128( rot_##r1##_##r2 ))
-ALIGN const UINT64 rot_0_20[2]  = { 0, 20};
-ALIGN const UINT64 rot_44_3[2]  = {44,  3};
-ALIGN const UINT64 rot_43_45[2] = {43, 45};
-ALIGN const UINT64 rot_21_61[2] = {21, 61};
-ALIGN const UINT64 rot_14_28[2] = {14, 28};
-ALIGN const UINT64 rot_1_36[2]  = { 1, 36};
-ALIGN const UINT64 rot_6_10[2]  = { 6, 10};
-ALIGN const UINT64 rot_25_15[2] = {25, 15};
-ALIGN const UINT64 rot_8_56[2]  = { 8, 56};
-ALIGN const UINT64 rot_18_27[2] = {18, 27};
-ALIGN const UINT64 rot_62_55[2] = {62, 55};
-ALIGN const UINT64 rot_39_41[2] = {39, 41};
+KALIGN const UINT64 rot_0_20[2]  = { 0, 20};
+KALIGN const UINT64 rot_44_3[2]  = {44,  3};
+KALIGN const UINT64 rot_43_45[2] = {43, 45};
+KALIGN const UINT64 rot_21_61[2] = {21, 61};
+KALIGN const UINT64 rot_14_28[2] = {14, 28};
+KALIGN const UINT64 rot_1_36[2]  = { 1, 36};
+KALIGN const UINT64 rot_6_10[2]  = { 6, 10};
+KALIGN const UINT64 rot_25_15[2] = {25, 15};
+KALIGN const UINT64 rot_8_56[2]  = { 8, 56};
+KALIGN const UINT64 rot_18_27[2] = {18, 27};
+KALIGN const UINT64 rot_62_55[2] = {62, 55};
+KALIGN const UINT64 rot_39_41[2] = {39, 41};
 
 #if defined(UseSimulatedXOP)
     // For debugging purposes, when XOP is not available

src/crypto/alt/KeccakSponge.h

@@ -22,16 +22,16 @@ and related or neighboring rights to the source code in this file.
 #define KeccakMaximumRateInBytes (KeccakMaximumRate/8)
 
 #if defined(__GNUC__)
-#define ALIGN __attribute__ ((aligned(32)))
+#define KALIGN __attribute__ ((aligned(32)))
 #elif defined(_MSC_VER)
-#define ALIGN __declspec(align(32))
+#define KALIGN __declspec(align(32))
 #else
-#define ALIGN
+#define KALIGN
 #endif
 
-ALIGN typedef struct spongeStateStruct {
-    ALIGN unsigned char state[KeccakPermutationSizeInBytes];
-    ALIGN unsigned char dataQueue[KeccakMaximumRateInBytes];
+KALIGN typedef struct spongeStateStruct {
+    KALIGN unsigned char state[KeccakPermutationSizeInBytes];
+    KALIGN unsigned char dataQueue[KeccakMaximumRateInBytes];
     unsigned int rate;
     unsigned int capacity;
     unsigned int bitsInQueue;

src/currency_core/miner.cpp

@@ -81,7 +81,7 @@ namespace currency
     CRITICAL_REGION_LOCAL(m_aliace_to_apply_in_block_lock);
     if(m_bc.get_alias_info(m_alias_to_apply_in_block.m_alias, dummy))
     {
-      LOG_PRINT_L0("Alias \"" << m_alias_to_apply_in_block.m_alias << "\" got used bye someone else, canceled." );
+      LOG_PRINT_L0("Alias \"" << m_alias_to_apply_in_block.m_alias << "\" got used by someone else, canceled." );
       m_alias_to_apply_in_block = AUTO_VAL_INIT(m_alias_to_apply_in_block);
       return false;
     }
@@ -256,14 +256,14 @@ namespace currency
     {
       LOG_PRINT_CYAN("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!", LOG_LEVEL_0);
       LOG_PRINT_L0(ENDL << "**********************************************************************" << ENDL 
-        << "NOTICE: Each block in blockchain have a vote, that takes into account "<< ENDL 
-        << "when calculating the reward amount for project team(once a day)." << ENDL 
+        << "NOTICE: Each block in the blockchain has a vote that is taken into account "<< ENDL 
+        << "when calculating the reward amount for project team (once per day)." << ENDL 
         << ENDL
-        << "Be sure to specify an option that expresses your attitude to work that make the project developers."  << ENDL 
+        << "Be sure to specify an option that expresses your attitude to work by the project developers."  << ENDL 
         << "If you support the project, leave donations enabled. If you disagree with the actions of the team,"
-        << "vote against donations.(by entering command \"set_donations false\")"  << ENDL 
+        << "vote against donations (by entering command \"set_donations false\")."  << ENDL 
         << ENDL 
-        << "By default, if you didn't refuse it explicitely, donations set to enabled." << ENDL 
+        << "By default, if you don't disable them explicitly, donations will be enabled." << ENDL 
         << ENDL
         << "**********************************************************************");
       LOG_PRINT_CYAN("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!", LOG_LEVEL_0);
@@ -276,13 +276,13 @@ namespace currency
     CRITICAL_REGION_LOCAL(m_threads_lock);
     if(is_mining())
     {
-      LOG_ERROR("Starting miner but it's already started");
+      LOG_ERROR("Starting miner canceled:  miner already running");
       return false;
     }
 
     if(!m_threads.empty())
     {
-      LOG_ERROR("Unable to start miner because there are active mining threads");
+      LOG_ERROR("Starting miner canceled:  there are still active mining threads");
       return false;
     }
 

src/miner/simpleminer.cpp

@@ -14,6 +14,7 @@
 #include "currency_core/account.h"
 #include "currency_core/currency_format_utils.h"
 #include "rpc/core_rpc_server_commands_defs.h"
+#include <sys/mman.h>
 
 using namespace epee;
 namespace po = boost::program_options;
@@ -101,6 +102,8 @@ namespace mining
     m_pass = command_line::get_arg(vm, arg_pass);
     m_hi = AUTO_VAL_INIT(m_hi);
     m_last_job_ticks = 0;
+    m_fast_scratchpad_pages = 0;
+    m_fast_scratchpad = NULL;
 
     return true;
   }
@@ -142,7 +145,7 @@ namespace mining
         crypto::hash h = currency::null_hash;
 	currency::get_blob_longhash(blob, h, m_job.prev_hi.height+1, [&](uint64_t index) -> crypto::hash&
 						{
-	return m_scratchpad[index%m_scratchpad.size()];
+	return m_fast_scratchpad[index%m_scratchpad.size()];
 	});
 
 	if( currency::check_hash(h, m_job.difficulty))
@@ -237,6 +240,7 @@ namespace mining
 
       uint64_t get_job_start_time = epee::misc_utils::get_tick_count();
       get_job(); /* Next version:  Handle this asynchronously */
+      update_fast_scratchpad();
       uint64_t get_job_end_time  = epee::misc_utils::get_tick_count();
       if ((get_job_end_time - get_job_start_time) > 1000) 
       {
@@ -336,6 +340,7 @@ namespace mining
       if (job_submit_failures == 5)
       {
         LOG_PRINT_L0("Too many submission failures.  Something is very wrong.");
+	free_fast_scratchpad();
         return false;
       }
       if (job_submit_failures > 3)
@@ -353,6 +358,7 @@ namespace mining
 	LOG_PRINT_L0("avg hr: " << hash_rate);
       }
     }
+    free_fast_scratchpad();
     return true;
   }
   //----------------------------------------------------------------------------------------------------------------------------------
@@ -388,6 +394,51 @@ namespace mining
     LOG_PRINT_L0("Scratchpad received ok, size: " << (m_scratchpad.size()*32)/1024 << "Kb, heigh=" << m_hi.height);
     return true;
   }
+  //----------------------------------------------------------------------------------------------------------------------------------
+  void simpleminer::free_fast_scratchpad()
+  {
+    if (m_fast_scratchpad) 
+    {
+      if (m_fast_mmapped)
+      {
+        munmap(m_fast_scratchpad, m_fast_scratchpad_pages*4096);
+      } else {
+	free(m_fast_scratchpad);
+      }
+      m_fast_scratchpad = NULL;
+    }
+  }
+
+  void simpleminer::update_fast_scratchpad()
+  {
+
+    /* Check size - reallocate fast scratch if necessary */
+    size_t cur_scratchpad_size = m_scratchpad.size() * sizeof(crypto::hash);
+    size_t cur_scratchpad_pages = (cur_scratchpad_size / 4096) + 1;
+    if (cur_scratchpad_pages > m_fast_scratchpad_pages)
+    {
+      free_fast_scratchpad();
+      void *addr = MAP_FAILED;
+      size_t mapsize = cur_scratchpad_pages * 4096;
+#ifdef MAP_HUGETLB
+      addr = mmap(0, mapsize, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS|MAP_HUGETLB|MAP_POPULATE, 0, 0);
+      if (addr == MAP_FAILED)
+      {
+        LOG_PRINT_L0("Unable to mmap huge pages.  Enable for faster operation.");
+      }
+#endif
+      if (addr == MAP_FAILED) {
+        addr = malloc(mapsize);
+      } else {
+        m_fast_mmapped = true;
+      }
+      m_fast_scratchpad_pages = cur_scratchpad_pages;
+      m_fast_scratchpad = (crypto::hash *)addr;
+    }
+
+    memcpy(m_fast_scratchpad, &m_scratchpad[0], m_scratchpad.size() * sizeof(crypto::hash));
+  }
+
   //----------------------------------------------------------------------------------------------------------------------------------
   bool  simpleminer::pop_addendum(const addendum& add)
   {

src/miner/simpleminer.h

@@ -43,11 +43,16 @@ namespace mining
     bool pop_addendum(const addendum& add);
     bool push_addendum(const addendum& add);
     void worker_thread(uint64_t start_nonce, uint32_t nonce_offset, std::atomic<uint32_t> *result, std::atomic<bool> *do_reset, std::atomic<bool> *done);
+    void update_fast_scratchpad();
+    void free_fast_scratchpad();
 
     std::vector<mining::addendum> m_blocks_addendums; //need to handle splits without re-downloading whole scratchpad
     height_info_native m_hi;
     std::vector<crypto::hash> m_scratchpad;
+    crypto::hash *m_fast_scratchpad;
+    uint32_t m_fast_scratchpad_pages;
     uint64_t m_last_job_ticks;
+    bool m_fast_mmapped;
     uint32_t m_threads_total;
     std::atomic<uint64_t> m_hashes_done;
     std::string m_pool_session_id;