1 /****************************************************************************
2 * Copyright (C) 2014-2015 Intel Corporation. All Rights Reserved.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 ****************************************************************************/
33 #if defined(__linux__) || defined(__gnu_linux__)
39 #include "common/os.h"
43 #include "rasterizer.h"
44 #include "rdtsc_core.h"
53 uint32_t procGroup
= 0;
54 std::vector
<uint32_t> threadIds
;
59 std::vector
<Core
> cores
;
62 typedef std::vector
<NumaNode
> CPUNumaNodes
;
64 void CalculateProcessorTopology(CPUNumaNodes
& out_nodes
, uint32_t& out_numThreadsPerProcGroup
)
67 out_numThreadsPerProcGroup
= 0;
71 std::vector
<KAFFINITY
> threadMaskPerProcGroup
;
74 std::lock_guard
<std::mutex
> l(m
);
76 static SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX buffer
[KNOB_MAX_NUM_THREADS
];
77 DWORD bufSize
= sizeof(buffer
);
79 BOOL ret
= GetLogicalProcessorInformationEx(RelationProcessorCore
, buffer
, &bufSize
);
80 SWR_ASSERT(ret
!= FALSE
, "Failed to get Processor Topology Information");
82 uint32_t count
= bufSize
/ buffer
->Size
;
83 PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX pBuffer
= buffer
;
85 for (uint32_t i
= 0; i
< count
; ++i
)
87 SWR_ASSERT(pBuffer
->Relationship
== RelationProcessorCore
);
88 for (uint32_t g
= 0; g
< pBuffer
->Processor
.GroupCount
; ++g
)
90 auto& gmask
= pBuffer
->Processor
.GroupMask
[g
];
91 uint32_t threadId
= 0;
92 uint32_t procGroup
= gmask
.Group
;
94 Core
* pCore
= nullptr;
96 uint32_t numThreads
= (uint32_t)_mm_popcount_sizeT(gmask
.Mask
);
98 while (BitScanForwardSizeT((unsigned long*)&threadId
, gmask
.Mask
))
101 KAFFINITY threadMask
= KAFFINITY(1) << threadId
;
102 gmask
.Mask
&= ~threadMask
;
104 if (procGroup
>= threadMaskPerProcGroup
.size())
106 threadMaskPerProcGroup
.resize(procGroup
+ 1);
109 if (threadMaskPerProcGroup
[procGroup
] & threadMask
)
111 // Already seen this mask. This means that we are in 32-bit mode and
112 // have seen more than 32 HW threads for this procGroup
115 SWR_ASSERT(false, "Shouldn't get here in 64-bit mode");
120 threadMaskPerProcGroup
[procGroup
] |= (KAFFINITY(1) << threadId
);
124 PROCESSOR_NUMBER procNum
= {};
125 procNum
.Group
= WORD(procGroup
);
126 procNum
.Number
= UCHAR(threadId
);
128 ret
= GetNumaProcessorNodeEx(&procNum
, (PUSHORT
)&numaId
);
132 if (out_nodes
.size() <= numaId
) out_nodes
.resize(numaId
+ 1);
133 auto& numaNode
= out_nodes
[numaId
];
137 if (nullptr == pCore
)
139 numaNode
.cores
.push_back(Core());
140 pCore
= &numaNode
.cores
.back();
141 pCore
->procGroup
= procGroup
;
143 pCore
->threadIds
.push_back(threadId
);
146 out_numThreadsPerProcGroup
++;
150 pBuffer
= PtrAdd(pBuffer
, pBuffer
->Size
);
154 #elif defined(__linux__) || defined (__gnu_linux__)
156 // Parse /proc/cpuinfo to get full topology
157 std::ifstream
input("/proc/cpuinfo");
160 uint32_t threadId
= uint32_t(-1);
161 uint32_t coreId
= uint32_t(-1);
162 uint32_t numaId
= uint32_t(-1);
164 while (std::getline(input
, line
))
166 if (line
.find("processor") != std::string::npos
)
168 if (threadId
!= uint32_t(-1))
171 if (out_nodes
.size() <= numaId
) out_nodes
.resize(numaId
+ 1);
172 auto& numaNode
= out_nodes
[numaId
];
173 if (numaNode
.cores
.size() <= coreId
) numaNode
.cores
.resize(coreId
+ 1);
174 auto& core
= numaNode
.cores
[coreId
];
176 core
.procGroup
= coreId
;
177 core
.threadIds
.push_back(threadId
);
179 out_numThreadsPerProcGroup
++;
182 auto data_start
= line
.find(": ") + 2;
183 threadId
= std::strtoul(&line
.c_str()[data_start
], &c
, 10);
186 if (line
.find("core id") != std::string::npos
)
188 auto data_start
= line
.find(": ") + 2;
189 coreId
= std::strtoul(&line
.c_str()[data_start
], &c
, 10);
192 if (line
.find("physical id") != std::string::npos
)
194 auto data_start
= line
.find(": ") + 2;
195 numaId
= std::strtoul(&line
.c_str()[data_start
], &c
, 10);
200 if (threadId
!= uint32_t(-1))
203 if (out_nodes
.size() <= numaId
) out_nodes
.resize(numaId
+ 1);
204 auto& numaNode
= out_nodes
[numaId
];
205 if (numaNode
.cores
.size() <= coreId
) numaNode
.cores
.resize(coreId
+ 1);
206 auto& core
= numaNode
.cores
[coreId
];
208 core
.procGroup
= coreId
;
209 core
.threadIds
.push_back(threadId
);
210 out_numThreadsPerProcGroup
++;
213 for (uint32_t node
= 0; node
< out_nodes
.size(); node
++) {
214 auto& numaNode
= out_nodes
[node
];
215 auto it
= numaNode
.cores
.begin();
216 for ( ; it
!= numaNode
.cores
.end(); ) {
217 if (it
->threadIds
.size() == 0)
218 numaNode
.cores
.erase(it
);
224 #elif defined(__CYGWIN__)
226 // Dummy data just to compile
229 core
.threadIds
.push_back(0);
230 node
.cores
.push_back(core
);
231 out_nodes
.push_back(node
);
232 out_numThreadsPerProcGroup
= 1;
236 #error Unsupported platform
242 void bindThread(uint32_t threadId
, uint32_t procGroupId
= 0, bool bindProcGroup
=false)
244 // Only bind threads when MAX_WORKER_THREADS isn't set.
245 if (KNOB_MAX_WORKER_THREADS
&& bindProcGroup
== false)
252 GROUP_AFFINITY affinity
= {};
253 affinity
.Group
= procGroupId
;
258 // Hopefully we don't get here. Logic in CreateThreadPool should prevent this.
259 SWR_REL_ASSERT(false, "Shouldn't get here");
261 // In a 32-bit process on Windows it is impossible to bind
262 // to logical processors 32-63 within a processor group.
263 // In this case set the mask to 0 and let the system assign
264 // the processor. Hopefully it will make smart choices.
270 // If KNOB_MAX_WORKER_THREADS is set, only bind to the proc group,
271 // Not the individual HW thread.
272 if (!KNOB_MAX_WORKER_THREADS
)
274 affinity
.Mask
= KAFFINITY(1) << threadId
;
278 SetThreadGroupAffinity(GetCurrentThread(), &affinity
, nullptr);
280 #elif defined(__CYGWIN__)
287 pthread_t thread
= pthread_self();
289 CPU_SET(threadId
, &cpuset
);
291 pthread_setaffinity_np(thread
, sizeof(cpu_set_t
), &cpuset
);
297 uint32_t GetEnqueuedDraw(SWR_CONTEXT
*pContext
)
299 return pContext
->dcRing
.GetHead();
303 DRAW_CONTEXT
*GetDC(SWR_CONTEXT
*pContext
, uint32_t drawId
)
305 return &pContext
->dcRing
[(drawId
-1) % KNOB_MAX_DRAWS_IN_FLIGHT
];
309 bool IDComparesLess(uint32_t a
, uint32_t b
)
311 // Use signed delta to ensure that wrap-around to 0 is correctly handled.
312 int32_t delta
= int32_t(a
- b
);
316 // returns true if dependency not met
318 bool CheckDependency(SWR_CONTEXT
*pContext
, DRAW_CONTEXT
*pDC
, uint32_t lastRetiredDraw
)
320 return pDC
->dependent
&& IDComparesLess(lastRetiredDraw
, pDC
->drawId
- 1);
323 // inlined-only version
324 INLINE
int64_t CompleteDrawContextInl(SWR_CONTEXT
* pContext
, DRAW_CONTEXT
* pDC
)
326 int64_t result
= InterlockedDecrement64(&pDC
->threadsDone
);
327 SWR_ASSERT(result
>= 0);
331 // Cleanup memory allocations
332 pDC
->pArena
->Reset(true);
335 pDC
->pTileMgr
->initialize();
337 if (pDC
->cleanupState
)
339 pDC
->pState
->pArena
->Reset(true);
344 pContext
->dcRing
.Dequeue(); // Remove from tail
350 // available to other translation modules
351 int64_t CompleteDrawContext(SWR_CONTEXT
* pContext
, DRAW_CONTEXT
* pDC
)
353 return CompleteDrawContextInl(pContext
, pDC
);
356 INLINE
bool FindFirstIncompleteDraw(SWR_CONTEXT
* pContext
, uint32_t& curDrawBE
, uint32_t& drawEnqueued
)
358 // increment our current draw id to the first incomplete draw
359 drawEnqueued
= GetEnqueuedDraw(pContext
);
360 while (IDComparesLess(curDrawBE
, drawEnqueued
))
362 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[curDrawBE
% KNOB_MAX_DRAWS_IN_FLIGHT
];
364 // If its not compute and FE is not done then break out of loop.
365 if (!pDC
->doneFE
&& !pDC
->isCompute
) break;
367 bool isWorkComplete
= pDC
->isCompute
?
368 pDC
->pDispatch
->isWorkComplete() :
369 pDC
->pTileMgr
->isWorkComplete();
374 CompleteDrawContextInl(pContext
, pDC
);
382 // If there are no more incomplete draws then return false.
383 return IDComparesLess(curDrawBE
, drawEnqueued
);
386 //////////////////////////////////////////////////////////////////////////
387 /// @brief If there is any BE work then go work on it.
388 /// @param pContext - pointer to SWR context.
389 /// @param workerId - The unique worker ID that is assigned to this thread.
390 /// @param curDrawBE - This tracks the draw contexts that this thread has processed. Each worker thread
391 /// has its own curDrawBE counter and this ensures that each worker processes all the
393 /// @param lockedTiles - This is the set of tiles locked by other threads. Each thread maintains its
394 /// own set and each time it fails to lock a macrotile, because its already locked,
395 /// then it will add that tile to the lockedTiles set. As a worker begins to work
396 /// on future draws the lockedTiles ensure that it doesn't work on tiles that may
397 /// still have work pending in a previous draw. Additionally, the lockedTiles is
398 /// hueristic that can steer a worker back to the same macrotile that it had been
399 /// working on in a previous draw.
401 SWR_CONTEXT
*pContext
,
404 TileSet
& lockedTiles
,
408 // Find the first incomplete draw that has pending work. If no such draw is found then
409 // return. FindFirstIncompleteDraw is responsible for incrementing the curDrawBE.
410 uint32_t drawEnqueued
= 0;
411 if (FindFirstIncompleteDraw(pContext
, curDrawBE
, drawEnqueued
) == false)
416 uint32_t lastRetiredDraw
= pContext
->dcRing
[curDrawBE
% KNOB_MAX_DRAWS_IN_FLIGHT
].drawId
- 1;
418 // Reset our history for locked tiles. We'll have to re-learn which tiles are locked.
421 // Try to work on each draw in order of the available draws in flight.
422 // 1. If we're on curDrawBE, we can work on any macrotile that is available.
423 // 2. If we're trying to work on draws after curDrawBE, we are restricted to
424 // working on those macrotiles that are known to be complete in the prior draw to
425 // maintain order. The locked tiles provides the history to ensures this.
426 for (uint32_t i
= curDrawBE
; IDComparesLess(i
, drawEnqueued
); ++i
)
428 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[i
% KNOB_MAX_DRAWS_IN_FLIGHT
];
430 if (pDC
->isCompute
) return; // We don't look at compute work.
432 // First wait for FE to be finished with this draw. This keeps threading model simple
433 // but if there are lots of bubbles between draws then serializing FE and BE may
434 // need to be revisited.
435 if (!pDC
->doneFE
) return;
437 // If this draw is dependent on a previous draw then we need to bail.
438 if (CheckDependency(pContext
, pDC
, lastRetiredDraw
))
443 // Grab the list of all dirty macrotiles. A tile is dirty if it has work queued to it.
444 std::vector
<uint32_t> ¯oTiles
= pDC
->pTileMgr
->getDirtyTiles();
446 for (uint32_t tileID
: macroTiles
)
448 // Only work on tiles for for this numa node
450 pDC
->pTileMgr
->getTileIndices(tileID
, x
, y
);
451 if (((x
^ y
) & numaMask
) != numaNode
)
456 MacroTileQueue
&tile
= pDC
->pTileMgr
->getMacroTileQueue(tileID
);
458 if (!tile
.getNumQueued())
463 // can only work on this draw if it's not in use by other threads
464 if (lockedTiles
.find(tileID
) != lockedTiles
.end())
473 RDTSC_START(WorkerFoundWork
);
475 uint32_t numWorkItems
= tile
.getNumQueued();
476 SWR_ASSERT(numWorkItems
);
480 if (pWork
->type
== DRAW
)
482 pContext
->pHotTileMgr
->InitializeHotTiles(pContext
, pDC
, tileID
);
485 while ((pWork
= tile
.peek()) != nullptr)
487 pWork
->pfnWork(pDC
, workerId
, tileID
, &pWork
->desc
);
490 RDTSC_STOP(WorkerFoundWork
, numWorkItems
, pDC
->drawId
);
494 pDC
->pTileMgr
->markTileComplete(tileID
);
496 // Optimization: If the draw is complete and we're the last one to have worked on it then
497 // we can reset the locked list as we know that all previous draws before the next are guaranteed to be complete.
498 if ((curDrawBE
== i
) && pDC
->pTileMgr
->isWorkComplete())
500 // We can increment the current BE and safely move to next draw since we know this draw is complete.
502 CompleteDrawContextInl(pContext
, pDC
);
512 // This tile is already locked. So let's add it to our locked tiles set. This way we don't try locking this one again.
513 lockedTiles
.insert(tileID
);
519 void WorkOnFifoFE(SWR_CONTEXT
*pContext
, uint32_t workerId
, uint32_t &curDrawFE
)
521 // Try to grab the next DC from the ring
522 uint32_t drawEnqueued
= GetEnqueuedDraw(pContext
);
523 while (IDComparesLess(curDrawFE
, drawEnqueued
))
525 uint32_t dcSlot
= curDrawFE
% KNOB_MAX_DRAWS_IN_FLIGHT
;
526 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[dcSlot
];
527 if (pDC
->isCompute
|| pDC
->doneFE
|| pDC
->FeLock
)
529 CompleteDrawContextInl(pContext
, pDC
);
538 uint32_t curDraw
= curDrawFE
;
539 while (IDComparesLess(curDraw
, drawEnqueued
))
541 uint32_t dcSlot
= curDraw
% KNOB_MAX_DRAWS_IN_FLIGHT
;
542 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[dcSlot
];
544 if (!pDC
->isCompute
&& !pDC
->FeLock
)
546 uint32_t initial
= InterlockedCompareExchange((volatile uint32_t*)&pDC
->FeLock
, 1, 0);
549 // successfully grabbed the DC, now run the FE
550 pDC
->FeWork
.pfnWork(pContext
, pDC
, workerId
, &pDC
->FeWork
.desc
);
560 //////////////////////////////////////////////////////////////////////////
561 /// @brief If there is any compute work then go work on it.
562 /// @param pContext - pointer to SWR context.
563 /// @param workerId - The unique worker ID that is assigned to this thread.
564 /// @param curDrawBE - This tracks the draw contexts that this thread has processed. Each worker thread
565 /// has its own curDrawBE counter and this ensures that each worker processes all the
568 SWR_CONTEXT
*pContext
,
572 uint32_t drawEnqueued
= 0;
573 if (FindFirstIncompleteDraw(pContext
, curDrawBE
, drawEnqueued
) == false)
578 uint32_t lastRetiredDraw
= pContext
->dcRing
[curDrawBE
% KNOB_MAX_DRAWS_IN_FLIGHT
].drawId
- 1;
580 for (uint64_t i
= curDrawBE
; IDComparesLess(i
, drawEnqueued
); ++i
)
582 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[i
% KNOB_MAX_DRAWS_IN_FLIGHT
];
583 if (pDC
->isCompute
== false) return;
585 // check dependencies
586 if (CheckDependency(pContext
, pDC
, lastRetiredDraw
))
591 SWR_ASSERT(pDC
->pDispatch
!= nullptr);
592 DispatchQueue
& queue
= *pDC
->pDispatch
;
594 // Is there any work remaining?
595 if (queue
.getNumQueued() > 0)
597 void* pSpillFillBuffer
= nullptr;
598 uint32_t threadGroupId
= 0;
599 while (queue
.getWork(threadGroupId
))
601 ProcessComputeBE(pDC
, workerId
, threadGroupId
, pSpillFillBuffer
);
603 queue
.finishedWork();
609 template<bool IsFEThread
, bool IsBEThread
>
610 DWORD
workerThreadMain(LPVOID pData
)
612 THREAD_DATA
*pThreadData
= (THREAD_DATA
*)pData
;
613 SWR_CONTEXT
*pContext
= pThreadData
->pContext
;
614 uint32_t threadId
= pThreadData
->threadId
;
615 uint32_t workerId
= pThreadData
->workerId
;
617 bindThread(threadId
, pThreadData
->procGroupId
, pThreadData
->forceBindProcGroup
);
619 RDTSC_INIT(threadId
);
621 uint32_t numaNode
= pThreadData
->numaId
;
622 uint32_t numaMask
= pContext
->threadPool
.numaMask
;
624 // flush denormals to 0
625 _mm_setcsr(_mm_getcsr() | _MM_FLUSH_ZERO_ON
| _MM_DENORMALS_ZERO_ON
);
627 // Track tiles locked by other threads. If we try to lock a macrotile and find its already
628 // locked then we'll add it to this list so that we don't try and lock it again.
631 // each worker has the ability to work on any of the queued draws as long as certain
632 // conditions are met. the data associated
633 // with a draw is guaranteed to be active as long as a worker hasn't signaled that he
634 // has moved on to the next draw when he determines there is no more work to do. The api
635 // thread will not increment the head of the dc ring until all workers have moved past the
637 // the logic to determine what to work on is:
638 // 1- try to work on the FE any draw that is queued. For now there are no dependencies
639 // on the FE work, so any worker can grab any FE and process in parallel. Eventually
640 // we'll need dependency tracking to force serialization on FEs. The worker will try
641 // to pick an FE by atomically incrementing a counter in the swr context. he'll keep
642 // trying until he reaches the tail.
643 // 2- BE work must be done in strict order. we accomplish this today by pulling work off
644 // the oldest draw (ie the head) of the dcRing. the worker can determine if there is
645 // any work left by comparing the total # of binned work items and the total # of completed
646 // work items. If they are equal, then there is no more work to do for this draw, and
647 // the worker can safely increment its oldestDraw counter and move on to the next draw.
648 std::unique_lock
<std::mutex
> lock(pContext
->WaitLock
, std::defer_lock
);
650 auto threadHasWork
= [&](uint32_t curDraw
) { return curDraw
!= pContext
->dcRing
.GetHead(); };
652 uint32_t curDrawBE
= 0;
653 uint32_t curDrawFE
= 0;
655 while (pContext
->threadPool
.inThreadShutdown
== false)
658 while (loop
++ < KNOB_WORKER_SPIN_LOOP_COUNT
&& !threadHasWork(curDrawBE
))
663 if (!threadHasWork(curDrawBE
))
667 // check for thread idle condition again under lock
668 if (threadHasWork(curDrawBE
))
674 if (pContext
->threadPool
.inThreadShutdown
)
680 RDTSC_START(WorkerWaitForThreadEvent
);
682 pContext
->FifosNotEmpty
.wait(lock
);
685 RDTSC_STOP(WorkerWaitForThreadEvent
, 0, 0);
687 if (pContext
->threadPool
.inThreadShutdown
)
695 RDTSC_START(WorkerWorkOnFifoBE
);
696 WorkOnFifoBE(pContext
, workerId
, curDrawBE
, lockedTiles
, numaNode
, numaMask
);
697 RDTSC_STOP(WorkerWorkOnFifoBE
, 0, 0);
699 WorkOnCompute(pContext
, workerId
, curDrawBE
);
704 WorkOnFifoFE(pContext
, workerId
, curDrawFE
);
708 curDrawBE
= curDrawFE
;
715 template<> DWORD workerThreadMain
<false, false>(LPVOID
) = delete;
717 template <bool IsFEThread
, bool IsBEThread
>
718 DWORD
workerThreadInit(LPVOID pData
)
724 return workerThreadMain
<IsFEThread
, IsBEThread
>(pData
);
728 __except(EXCEPTION_CONTINUE_SEARCH
)
736 template<> DWORD workerThreadInit
<false, false>(LPVOID pData
) = delete;
738 void CreateThreadPool(SWR_CONTEXT
*pContext
, THREAD_POOL
*pPool
)
743 uint32_t numThreadsPerProcGroup
= 0;
744 CalculateProcessorTopology(nodes
, numThreadsPerProcGroup
);
746 uint32_t numHWNodes
= (uint32_t)nodes
.size();
747 uint32_t numHWCoresPerNode
= (uint32_t)nodes
[0].cores
.size();
748 uint32_t numHWHyperThreads
= (uint32_t)nodes
[0].cores
[0].threadIds
.size();
750 // Calculate num HW threads. Due to asymmetric topologies, this is not
751 // a trivial multiplication.
752 uint32_t numHWThreads
= 0;
753 for (auto& node
: nodes
)
755 for (auto& core
: node
.cores
)
757 numHWThreads
+= (uint32_t)core
.threadIds
.size();
761 uint32_t numNodes
= numHWNodes
;
762 uint32_t numCoresPerNode
= numHWCoresPerNode
;
763 uint32_t numHyperThreads
= numHWHyperThreads
;
765 if (KNOB_MAX_WORKER_THREADS
)
767 SET_KNOB(HYPERTHREADED_FE
, false);
770 if (KNOB_HYPERTHREADED_FE
)
772 SET_KNOB(MAX_THREADS_PER_CORE
, 0);
775 if (KNOB_MAX_NUMA_NODES
)
777 numNodes
= std::min(numNodes
, KNOB_MAX_NUMA_NODES
);
780 if (KNOB_MAX_CORES_PER_NUMA_NODE
)
782 numCoresPerNode
= std::min(numCoresPerNode
, KNOB_MAX_CORES_PER_NUMA_NODE
);
785 if (KNOB_MAX_THREADS_PER_CORE
)
787 numHyperThreads
= std::min(numHyperThreads
, KNOB_MAX_THREADS_PER_CORE
);
790 #if defined(_WIN32) && !defined(_WIN64)
791 if (!KNOB_MAX_WORKER_THREADS
)
793 // Limit 32-bit windows to bindable HW threads only
794 if ((numCoresPerNode
* numHWHyperThreads
) > 32)
796 numCoresPerNode
= 32 / numHWHyperThreads
;
801 if (numHyperThreads
< 2)
803 SET_KNOB(HYPERTHREADED_FE
, false);
806 // Calculate numThreads
807 uint32_t numThreads
= numNodes
* numCoresPerNode
* numHyperThreads
;
808 numThreads
= std::min(numThreads
, numHWThreads
);
810 if (KNOB_MAX_WORKER_THREADS
)
812 uint32_t maxHWThreads
= numHWNodes
* numHWCoresPerNode
* numHWHyperThreads
;
813 numThreads
= std::min(KNOB_MAX_WORKER_THREADS
, maxHWThreads
);
816 if (numThreads
> KNOB_MAX_NUM_THREADS
)
818 printf("WARNING: system thread count %u exceeds max %u, "
819 "performance will be degraded\n",
820 numThreads
, KNOB_MAX_NUM_THREADS
);
823 uint32_t numAPIReservedThreads
= 1;
828 // If only 1 worker threads, try to move it to an available
829 // HW thread. If that fails, use the API thread.
830 if (numCoresPerNode
< numHWCoresPerNode
)
834 else if (numHyperThreads
< numHWHyperThreads
)
838 else if (numNodes
< numHWNodes
)
844 pPool
->numThreads
= 0;
845 SET_KNOB(SINGLE_THREADED
, true);
851 // Save HW threads for the API if we can
852 if (numThreads
> numAPIReservedThreads
)
854 numThreads
-= numAPIReservedThreads
;
858 numAPIReservedThreads
= 0;
862 pPool
->numThreads
= numThreads
;
863 pContext
->NumWorkerThreads
= pPool
->numThreads
;
865 pPool
->inThreadShutdown
= false;
866 pPool
->pThreadData
= (THREAD_DATA
*)malloc(pPool
->numThreads
* sizeof(THREAD_DATA
));
869 if (KNOB_MAX_WORKER_THREADS
)
871 bool bForceBindProcGroup
= (numThreads
> numThreadsPerProcGroup
);
872 uint32_t numProcGroups
= (numThreads
+ numThreadsPerProcGroup
- 1) / numThreadsPerProcGroup
;
873 // When MAX_WORKER_THREADS is set we don't bother to bind to specific HW threads
874 // But Windows will still require binding to specific process groups
875 for (uint32_t workerId
= 0; workerId
< numThreads
; ++workerId
)
877 pPool
->pThreadData
[workerId
].workerId
= workerId
;
878 pPool
->pThreadData
[workerId
].procGroupId
= workerId
% numProcGroups
;
879 pPool
->pThreadData
[workerId
].threadId
= 0;
880 pPool
->pThreadData
[workerId
].numaId
= 0;
881 pPool
->pThreadData
[workerId
].coreId
= 0;
882 pPool
->pThreadData
[workerId
].htId
= 0;
883 pPool
->pThreadData
[workerId
].pContext
= pContext
;
884 pPool
->pThreadData
[workerId
].forceBindProcGroup
= bForceBindProcGroup
;
885 pPool
->threads
[workerId
] = new std::thread(workerThreadInit
<true, true>, &pPool
->pThreadData
[workerId
]);
887 pContext
->NumBEThreads
++;
888 pContext
->NumFEThreads
++;
893 pPool
->numaMask
= numNodes
- 1; // Only works for 2**n numa nodes (1, 2, 4, etc.)
895 uint32_t workerId
= 0;
896 for (uint32_t n
= 0; n
< numNodes
; ++n
)
898 auto& node
= nodes
[n
];
899 uint32_t numCores
= numCoresPerNode
;
900 for (uint32_t c
= 0; c
< numCores
; ++c
)
902 if (c
>= node
.cores
.size())
907 auto& core
= node
.cores
[c
];
908 for (uint32_t t
= 0; t
< numHyperThreads
; ++t
)
910 if (t
>= core
.threadIds
.size())
915 if (numAPIReservedThreads
)
917 --numAPIReservedThreads
;
921 SWR_ASSERT(workerId
< numThreads
);
923 pPool
->pThreadData
[workerId
].workerId
= workerId
;
924 pPool
->pThreadData
[workerId
].procGroupId
= core
.procGroup
;
925 pPool
->pThreadData
[workerId
].threadId
= core
.threadIds
[t
];
926 pPool
->pThreadData
[workerId
].numaId
= n
;
927 pPool
->pThreadData
[workerId
].coreId
= c
;
928 pPool
->pThreadData
[workerId
].htId
= t
;
929 pPool
->pThreadData
[workerId
].pContext
= pContext
;
931 if (KNOB_HYPERTHREADED_FE
)
935 pContext
->NumBEThreads
++;
936 pPool
->threads
[workerId
] = new std::thread(workerThreadInit
<false, true>, &pPool
->pThreadData
[workerId
]);
940 pContext
->NumFEThreads
++;
941 pPool
->threads
[workerId
] = new std::thread(workerThreadInit
<true, false>, &pPool
->pThreadData
[workerId
]);
946 pPool
->threads
[workerId
] = new std::thread(workerThreadInit
<true, true>, &pPool
->pThreadData
[workerId
]);
947 pContext
->NumBEThreads
++;
948 pContext
->NumFEThreads
++;
958 void DestroyThreadPool(SWR_CONTEXT
*pContext
, THREAD_POOL
*pPool
)
960 if (!KNOB_SINGLE_THREADED
)
962 // Inform threads to finish up
963 std::unique_lock
<std::mutex
> lock(pContext
->WaitLock
);
964 pPool
->inThreadShutdown
= true;
966 pContext
->FifosNotEmpty
.notify_all();
969 // Wait for threads to finish and destroy them
970 for (uint32_t t
= 0; t
< pPool
->numThreads
; ++t
)
972 pPool
->threads
[t
]->join();
973 delete(pPool
->threads
[t
]);
976 // Clean up data used by threads
977 free(pPool
->pThreadData
);