void bindThread(uint32_t threadId, uint32_t procGroupId = 0)
{
#if defined(_WIN32)
{
GROUP_AFFINITY affinity = {};
affinity.Group = procGroupId;
#if !defined(_WIN64)
if (threadId >= 32)
{
// In a 32-bit process on Windows it is impossible to bind
// to logical processors 32-63 within a processor group.
// In this case set the mask to 0 and let the system assign
// the processor. Hopefully it will make smart choices.
affinity.Mask = 0;
}
else
#endif
{
affinity.Mask = KAFFINITY(1) << threadId;
}
SetThreadGroupAffinity(GetCurrentThread(), &affinity, nullptr);
}
#else
cpu_set_t cpuset;
pthread_t thread = pthread_self();
CPU_ZERO(&cpuset);
CPU_SET(threadId, &cpuset);
pthread_setaffinity_np(thread, sizeof(cpu_set_t), &cpuset);
#endif
}
void bindThread(SWR_CONTEXT* pContext, uint32_t threadId, uint32_t procGroupId = 0, bool bindProcGroup=false)
{
// Only bind threads when MAX_WORKER_THREADS isn't set.
if (pContext->threadInfo.SINGLE_THREADED || (pContext->threadInfo.MAX_WORKER_THREADS && bindProcGroup == false))
{
return;
}
#if defined(_WIN32)
GROUP_AFFINITY affinity = {};
affinity.Group = procGroupId;
#if !defined(_WIN64)
if (threadId >= 32)
{
// Hopefully we don't get here. Logic in CreateThreadPool should prevent this.
SWR_INVALID("Shouldn't get here");
// In a 32-bit process on Windows it is impossible to bind
// to logical processors 32-63 within a processor group.
// In this case set the mask to 0 and let the system assign
// the processor. Hopefully it will make smart choices.
affinity.Mask = 0;
}
else
#endif
{
// If MAX_WORKER_THREADS is set, only bind to the proc group,
// Not the individual HW thread.
if (!bindProcGroup && !pContext->threadInfo.MAX_WORKER_THREADS)
{
affinity.Mask = KAFFINITY(1) << threadId;
}
else
{
affinity.Mask = KAFFINITY(0);
}
}
if (!SetThreadGroupAffinity(GetCurrentThread(), &affinity, nullptr))
{
SWR_INVALID("Failed to set Thread Affinity");
}
#elif defined(__linux__) || defined(__gnu_linux__)
cpu_set_t cpuset;
pthread_t thread = pthread_self();
CPU_ZERO(&cpuset);
CPU_SET(threadId, &cpuset);
int err = pthread_setaffinity_np(thread, sizeof(cpu_set_t), &cpuset);
if (err != 0)
{
fprintf(stderr, "pthread_setaffinity_np failure for tid %u: %s\n", threadId, strerror(err));
}
#endif
}
// Sets the calling thread's affinity to only run on the processor specified
// in the GCThreadAffinity structure.
// Parameters:
// affinity - The requested affinity for the calling thread. At most one processor
// can be provided.
// Return:
// true if setting the affinity was successful, false otherwise.
bool GCToOSInterface::SetThreadAffinity(GCThreadAffinity* affinity)
{
LIMITED_METHOD_CONTRACT;
assert(affinity != nullptr);
if (affinity->Group != GCThreadAffinity::None)
{
assert(affinity->Processor != GCThreadAffinity::None);
GROUP_AFFINITY ga;
ga.Group = (WORD)affinity->Group;
ga.Reserved[0] = 0; // reserve must be filled with zero
ga.Reserved[1] = 0; // otherwise call may fail
ga.Reserved[2] = 0;
ga.Mask = (size_t)1 << affinity->Processor;
return !!SetThreadGroupAffinity(GetCurrentThread(), &ga, nullptr);
}
else if (affinity->Processor != GCThreadAffinity::None)
{
return !!SetThreadAffinityMask(GetCurrentThread(), (DWORD_PTR)1 << affinity->Processor);
}
// Given affinity must specify at least one processor to use.
return false;
}
void bindThread(SWR_CONTEXT* pContext, uint32_t threadId, uint32_t procGroupId = 0, bool bindProcGroup=false)
{
// Only bind threads when MAX_WORKER_THREADS isn't set.
if (pContext->threadInfo.MAX_WORKER_THREADS && bindProcGroup == false)
{
return;
}
#if defined(_WIN32)
GROUP_AFFINITY affinity = {};
affinity.Group = procGroupId;
#if !defined(_WIN64)
if (threadId >= 32)
{
// Hopefully we don't get here. Logic in CreateThreadPool should prevent this.
SWR_REL_ASSERT(false, "Shouldn't get here");
// In a 32-bit process on Windows it is impossible to bind
// to logical processors 32-63 within a processor group.
// In this case set the mask to 0 and let the system assign
// the processor. Hopefully it will make smart choices.
affinity.Mask = 0;
}
else
#endif
{
// If MAX_WORKER_THREADS is set, only bind to the proc group,
// Not the individual HW thread.
if (!pContext->threadInfo.MAX_WORKER_THREADS)
{
affinity.Mask = KAFFINITY(1) << threadId;
}
}
SetThreadGroupAffinity(GetCurrentThread(), &affinity, nullptr);
#else
cpu_set_t cpuset;
pthread_t thread = pthread_self();
CPU_ZERO(&cpuset);
CPU_SET(threadId, &cpuset);
pthread_setaffinity_np(thread, sizeof(cpu_set_t), &cpuset);
#endif
}
ThreadGroupTempAffinity(uint32 core_id)
{
GROUP_AFFINITY NewGroupAffinity;
memset(&NewGroupAffinity, 0, sizeof(GROUP_AFFINITY));
memset(&PreviousGroupAffinity, 0, sizeof(GROUP_AFFINITY));
uint32 currentGroupSize = 0;
while (core_id >= (currentGroupSize = GetMaximumProcessorCount(NewGroupAffinity.Group)))
{
core_id -= currentGroupSize;
++NewGroupAffinity.Group;
}
NewGroupAffinity.Mask = 1ULL << core_id;
SetThreadGroupAffinity(GetCurrentThread(),&NewGroupAffinity,&PreviousGroupAffinity);
}
void bindThread(uint32_t threadId, uint32_t procGroupId = 0, bool bindProcGroup=false)
{
// Only bind threads when MAX_WORKER_THREADS isn't set.
if (KNOB_MAX_WORKER_THREADS && bindProcGroup == false)
{
return;
}
#if defined(_WIN32)
{
GROUP_AFFINITY affinity = {};
affinity.Group = procGroupId;
#if !defined(_WIN64)
if (threadId >= 32)
{
// In a 32-bit process on Windows it is impossible to bind
// to logical processors 32-63 within a processor group.
// In this case set the mask to 0 and let the system assign
// the processor. Hopefully it will make smart choices.
affinity.Mask = 0;
}
else
#endif
{
// If KNOB_MAX_WORKER_THREADS is set, only bind to the proc group,
// Not the individual HW thread.
if (!KNOB_MAX_WORKER_THREADS)
{
affinity.Mask = KAFFINITY(1) << threadId;
}
}
SetThreadGroupAffinity(GetCurrentThread(), &affinity, nullptr);
}
#else
cpu_set_t cpuset;
pthread_t thread = pthread_self();
CPU_ZERO(&cpuset);
CPU_SET(threadId, &cpuset);
pthread_setaffinity_np(thread, sizeof(cpu_set_t), &cpuset);
#endif
}
tt_result_t __thread_bind_numa(IN HANDLE hThread, IN USHORT Node)
{
GROUP_AFFINITY ProcessorMask;
// how to handle the arch in which a numa node includes more than
// 64 cpus??
if (!GetNumaNodeProcessorMaskEx(Node, &ProcessorMask)) {
TT_ERROR("fail to get numa node processor mask");
return TT_FAIL;
}
if (!SetThreadGroupAffinity(hThread, &ProcessorMask, NULL)) {
TT_ERROR("fail to bind thread to numa node[%d]", Node);
return TT_FAIL;
}
return TT_SUCCESS;
}
/*! set the affinity of a given thread */
void setAffinity(HANDLE thread, ssize_t affinity)
{
#if (_WIN32_WINNT >= 0x0601) // FIXME: use getProcAddress to activate this feature only if supported by Windows
int groups = GetActiveProcessorGroupCount();
int totalProcessors = 0, group = 0, number = 0;
for (int i = 0; i<groups; i++) {
int processors = GetActiveProcessorCount(i);
if (totalProcessors + processors > affinity) {
group = i;
number = (int)affinity - totalProcessors;
break;
}
totalProcessors += processors;
}
GROUP_AFFINITY groupAffinity;
groupAffinity.Group = (WORD)group;
groupAffinity.Mask = (KAFFINITY)(uint64(1) << number);
groupAffinity.Reserved[0] = 0;
groupAffinity.Reserved[1] = 0;
groupAffinity.Reserved[2] = 0;
if (!SetThreadGroupAffinity(thread, &groupAffinity, NULL))
THROW_RUNTIME_ERROR("cannot set thread group affinity");
PROCESSOR_NUMBER processorNumber;
processorNumber.Group = group;
processorNumber.Number = number;
processorNumber.Reserved = 0;
if (!SetThreadIdealProcessorEx(thread, &processorNumber, NULL))
THROW_RUNTIME_ERROR("cannot set ideal processor");
#else
if (!SetThreadAffinityMask(thread, DWORD_PTR(uint64(1) << affinity)))
THROW_RUNTIME_ERROR("cannot set thread affinity mask");
if (SetThreadIdealProcessor(thread, (DWORD)affinity) == (DWORD)-1)
THROW_RUNTIME_ERROR("cannot set ideal processor");
#endif
}
/*! set the affinity of a given thread */
void setAffinity(HANDLE thread, ssize_t affinity)
{
#if (_WIN32_WINNT >= 0x0601)
int groups = GetActiveProcessorGroupCount();
int totalProcessors = 0, group = 0, number = 0;
for (int i = 0; i<groups; i++) {
int processors = GetActiveProcessorCount(i);
if (totalProcessors + processors > affinity) {
group = i;
number = (int)affinity - totalProcessors;
break;
}
totalProcessors += processors;
}
GROUP_AFFINITY groupAffinity;
groupAffinity.Group = (WORD)group;
groupAffinity.Mask = (KAFFINITY)(uint64(1) << number);
groupAffinity.Reserved[0] = 0;
groupAffinity.Reserved[1] = 0;
groupAffinity.Reserved[2] = 0;
if (!SetThreadGroupAffinity(thread, &groupAffinity, NULL))
throw std::runtime_error("cannot set thread group affinity");
PROCESSOR_NUMBER processorNumber;
processorNumber.Group = group;
processorNumber.Number = number;
processorNumber.Reserved = 0;
if (!SetThreadIdealProcessorEx(thread, &processorNumber, NULL))
throw std::runtime_error("cannot set thread ideal processor");
#else
if (!SetThreadAffinityMask(thread, DWORD_PTR(uint64(1) << affinity)))
throw std::runtime_error("cannot set thread affinity mask");
if (SetThreadIdealProcessor(thread, (DWORD)affinity) == (DWORD)-1)
throw std::runtime_error("cannot set thread ideal processor");
#endif
}
int thread_bind_native(__unused_variable struct cpuid_state_t *state, __unused_variable uint32_t id)
{
#ifdef TARGET_OS_WINDOWS
BOOL ret;
HANDLE hThread = GetCurrentThread();
#if _WIN32_WINNT >= 0x0601
GROUP_AFFINITY affinity;
ZeroMemory(&affinity, sizeof(GROUP_AFFINITY));
affinity.Group = id / 64;
affinity.Mask = 1 << (id % 64);
ret = SetThreadGroupAffinity(hThread, &affinity, NULL);
#else
DWORD mask;
if (id > 32)
return 1;
mask = (1 << id);
ret = SetThreadAffinityMask(hThread, mask);
#endif
return (ret != FALSE) ? 0 : 1;
#elif defined(TARGET_OS_LINUX) || defined(TARGET_OS_FREEBSD)
int ret;
#ifdef CPU_SET_S
size_t setsize = CPU_ALLOC_SIZE(MAX_CPUS);
CPUSET_T *set = CPU_ALLOC(MAX_CPUS);
pthread_t pth;
pth = pthread_self();
CPU_ZERO_S(setsize, set);
CPU_SET_S(id, setsize, set);
ret = pthread_setaffinity_np(pth, setsize, set);
CPU_FREE(set);
#else
size_t bits_per_set = sizeof(CPUSET_T) * 8;
size_t bits_per_subset = sizeof(CPUSET_MASK_T) * 8;
size_t setsize = sizeof(CPUSET_T) * (MAX_CPUS / bits_per_set);
size_t set_id, subset_id;
unsigned long long mask;
CPUSET_T *set = malloc(setsize);
pthread_t pth;
pth = pthread_self();
for (set_id = 0; set_id < (MAX_CPUS / bits_per_set); set_id++)
CPU_ZERO(&set[set_id]);
set_id = id / bits_per_set;
id %= bits_per_set;
subset_id = id / bits_per_subset;
id %= bits_per_subset;
mask = 1ULL << (unsigned long long)id;
((unsigned long *)set[set_id].__bits)[subset_id] |= mask;
ret = pthread_setaffinity_np(pth, setsize, set);
free(set);
#endif
return (ret == 0) ? 0 : 1;
#elif defined(TARGET_OS_MACOSX)
#ifdef USE_CHUD
return (utilBindThreadToCPU(id) == 0) ? 0 : 1;
#else
return 1;
#endif
#else
#error "thread_bind_native() not defined for this platform"
#endif
}
~ThreadGroupTempAffinity()
{
SetThreadGroupAffinity(GetCurrentThread(),&PreviousGroupAffinity,NULL);
}