Subversion Repositories Games.Chess Giants

#ifdef _WIN32

#if _WIN32_WINNT < 0x0601

#undef  _WIN32_WINNT

#define _WIN32_WINNT 0x0601 // Force to include needed API prototypes

#endif

#include <windows.h>

// The needed Windows API for processor groups could be missed from old Windows

// versions, so instead of calling them directly (forcing the linker to resolve

// the calls at compile time), try to load them at runtime. To do this we need

// first to define the corresponding function pointers.

extern "C" {

typedef bool(*fun1_t)(LOGICAL_PROCESSOR_RELATIONSHIP,

                      PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD);

typedef bool(*fun2_t)(USHORT, PGROUP_AFFINITY);

typedef bool(*fun3_t)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);

}

#endif

#include <vector>

void prefetch2(void* addr) {

  prefetch(addr);

  prefetch((uint8_t*)addr + 64);

}

namespace WinProcGroup {

#ifndef _WIN32

void bindThisThread(size_t) {}

#else

/// get_group() retrieves logical processor information using Windows specific

/// API and returns the best group id for the thread with index idx. Original

/// code from Texel by Peter Österlund.

int get_group(size_t idx) {

  int threads = 0;

  int nodes = 0;

  int cores = 0;

  DWORD returnLength = 0;

  DWORD byteOffset = 0;

  // Early exit if the needed API is not available at runtime

  HMODULE k32 = GetModuleHandle("Kernel32.dll");

  auto fun1 = (fun1_t)GetProcAddress(k32, "GetLogicalProcessorInformationEx");

  if (!fun1)

      return -1;

  // First call to get returnLength. We expect it to fail due to null buffer

  if (fun1(RelationAll, nullptr, &returnLength))

      return -1;

  // Once we know returnLength, allocate the buffer

  SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;

  ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);

  // Second call, now we expect to succeed

  if (!fun1(RelationAll, buffer, &returnLength))

  {

      free(buffer);

      return -1;

  }

  while (ptr->Size > 0 && byteOffset + ptr->Size <= returnLength)

  {

      if (ptr->Relationship == RelationNumaNode)

          nodes++;

      else if (ptr->Relationship == RelationProcessorCore)

      {

          cores++;

          threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;

      }

      byteOffset += ptr->Size;

      ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);

  }

  free(buffer);

  std::vector<int> groups;

  // Run as many threads as possible on the same node until core limit is

  // reached, then move on filling the next node.

  for (int n = 0; n < nodes; n++)

      for (int i = 0; i < cores / nodes; i++)

          groups.push_back(n);

  // In case a core has more than one logical processor (we assume 2) and we

  // have still threads to allocate, then spread them evenly across available

  // nodes.

  for (int t = 0; t < threads - cores; t++)

      groups.push_back(t % nodes);

  // If we still have more threads than the total number of logical processors

  // then return -1 and let the OS to decide what to do.

  return idx < groups.size() ? groups[idx] : -1;

}

/// bindThisThread() set the group affinity of the current thread

void bindThisThread(size_t idx) {

  // Use only local variables to be thread-safe

  int group = get_group(idx);

  if (group == -1)

      return;

  // Early exit if the needed API are not available at runtime

  HMODULE k32 = GetModuleHandle("Kernel32.dll");

  auto fun2 = (fun2_t)GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");

  auto fun3 = (fun3_t)GetProcAddress(k32, "SetThreadGroupAffinity");

  if (!fun2 || !fun3)

      return;

  GROUP_AFFINITY affinity;

  if (fun2(group, &affinity))

      fun3(GetCurrentThread(), &affinity, nullptr);

}

#endif

} // namespace WinProcGroup