1 /****************************************************************************
2 * Copyright (C) 2014-2016 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
);
226 #error Unsupported platform
232 void bindThread(SWR_CONTEXT
* pContext
, uint32_t threadId
, uint32_t procGroupId
= 0, bool bindProcGroup
=false)
234 // Only bind threads when MAX_WORKER_THREADS isn't set.
235 if (pContext
->threadInfo
.MAX_WORKER_THREADS
&& bindProcGroup
== false)
242 GROUP_AFFINITY affinity
= {};
243 affinity
.Group
= procGroupId
;
248 // Hopefully we don't get here. Logic in CreateThreadPool should prevent this.
249 SWR_REL_ASSERT(false, "Shouldn't get here");
251 // In a 32-bit process on Windows it is impossible to bind
252 // to logical processors 32-63 within a processor group.
253 // In this case set the mask to 0 and let the system assign
254 // the processor. Hopefully it will make smart choices.
260 // If MAX_WORKER_THREADS is set, only bind to the proc group,
261 // Not the individual HW thread.
262 if (!pContext
->threadInfo
.MAX_WORKER_THREADS
)
264 affinity
.Mask
= KAFFINITY(1) << threadId
;
268 SetThreadGroupAffinity(GetCurrentThread(), &affinity
, nullptr);
273 pthread_t thread
= pthread_self();
275 CPU_SET(threadId
, &cpuset
);
277 pthread_setaffinity_np(thread
, sizeof(cpu_set_t
), &cpuset
);
283 uint32_t GetEnqueuedDraw(SWR_CONTEXT
*pContext
)
285 return pContext
->dcRing
.GetHead();
289 DRAW_CONTEXT
*GetDC(SWR_CONTEXT
*pContext
, uint32_t drawId
)
291 return &pContext
->dcRing
[(drawId
-1) % KNOB_MAX_DRAWS_IN_FLIGHT
];
295 bool IDComparesLess(uint32_t a
, uint32_t b
)
297 // Use signed delta to ensure that wrap-around to 0 is correctly handled.
298 int32_t delta
= int32_t(a
- b
);
302 // returns true if dependency not met
304 bool CheckDependency(SWR_CONTEXT
*pContext
, DRAW_CONTEXT
*pDC
, uint32_t lastRetiredDraw
)
306 return pDC
->dependent
&& IDComparesLess(lastRetiredDraw
, pDC
->drawId
- 1);
309 //////////////////////////////////////////////////////////////////////////
310 /// @brief Update client stats.
311 INLINE
void UpdateClientStats(SWR_CONTEXT
* pContext
, DRAW_CONTEXT
* pDC
)
313 if ((pContext
->pfnUpdateStats
== nullptr) || (GetApiState(pDC
).enableStats
== false))
318 DRAW_DYNAMIC_STATE
& dynState
= pDC
->dynState
;
319 SWR_STATS stats
{ 0 };
321 // Sum up stats across all workers before sending to client.
322 for (uint32_t i
= 0; i
< pContext
->NumWorkerThreads
; ++i
)
324 stats
.DepthPassCount
+= dynState
.stats
[i
].DepthPassCount
;
325 stats
.IaVertices
+= dynState
.stats
[i
].IaVertices
;
326 stats
.IaPrimitives
+= dynState
.stats
[i
].IaPrimitives
;
327 stats
.VsInvocations
+= dynState
.stats
[i
].VsInvocations
;
328 stats
.HsInvocations
+= dynState
.stats
[i
].HsInvocations
;
329 stats
.DsInvocations
+= dynState
.stats
[i
].DsInvocations
;
330 stats
.GsInvocations
+= dynState
.stats
[i
].GsInvocations
;
331 stats
.PsInvocations
+= dynState
.stats
[i
].PsInvocations
;
332 stats
.CInvocations
+= dynState
.stats
[i
].CInvocations
;
333 stats
.CsInvocations
+= dynState
.stats
[i
].CsInvocations
;
334 stats
.CPrimitives
+= dynState
.stats
[i
].CPrimitives
;
335 stats
.GsPrimitives
+= dynState
.stats
[i
].GsPrimitives
;
337 for (uint32_t stream
= 0; stream
< MAX_SO_STREAMS
; ++stream
)
339 stats
.SoPrimStorageNeeded
[stream
] += dynState
.stats
[i
].SoPrimStorageNeeded
[stream
];
340 stats
.SoNumPrimsWritten
[stream
] += dynState
.stats
[i
].SoNumPrimsWritten
[stream
];
344 pContext
->pfnUpdateStats(GetPrivateState(pDC
), &stats
);
347 INLINE
void ExecuteCallbacks(SWR_CONTEXT
* pContext
, DRAW_CONTEXT
* pDC
)
349 UpdateClientStats(pContext
, pDC
);
351 if (pDC
->retireCallback
.pfnCallbackFunc
)
353 pDC
->retireCallback
.pfnCallbackFunc(pDC
->retireCallback
.userData
,
354 pDC
->retireCallback
.userData2
,
355 pDC
->retireCallback
.userData3
);
359 // inlined-only version
360 INLINE
int32_t CompleteDrawContextInl(SWR_CONTEXT
* pContext
, DRAW_CONTEXT
* pDC
)
362 int32_t result
= InterlockedDecrement((volatile LONG
*)&pDC
->threadsDone
);
363 SWR_ASSERT(result
>= 0);
367 ExecuteCallbacks(pContext
, pDC
);
369 // Cleanup memory allocations
370 pDC
->pArena
->Reset(true);
373 pDC
->pTileMgr
->initialize();
375 if (pDC
->cleanupState
)
377 pDC
->pState
->pArena
->Reset(true);
382 pContext
->dcRing
.Dequeue(); // Remove from tail
388 // available to other translation modules
389 int32_t CompleteDrawContext(SWR_CONTEXT
* pContext
, DRAW_CONTEXT
* pDC
)
391 return CompleteDrawContextInl(pContext
, pDC
);
394 INLINE
bool FindFirstIncompleteDraw(SWR_CONTEXT
* pContext
, uint32_t& curDrawBE
, uint32_t& drawEnqueued
)
396 // increment our current draw id to the first incomplete draw
397 drawEnqueued
= GetEnqueuedDraw(pContext
);
398 while (IDComparesLess(curDrawBE
, drawEnqueued
))
400 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[curDrawBE
% KNOB_MAX_DRAWS_IN_FLIGHT
];
402 // If its not compute and FE is not done then break out of loop.
403 if (!pDC
->doneFE
&& !pDC
->isCompute
) break;
405 bool isWorkComplete
= pDC
->isCompute
?
406 pDC
->pDispatch
->isWorkComplete() :
407 pDC
->pTileMgr
->isWorkComplete();
412 CompleteDrawContextInl(pContext
, pDC
);
420 // If there are no more incomplete draws then return false.
421 return IDComparesLess(curDrawBE
, drawEnqueued
);
424 //////////////////////////////////////////////////////////////////////////
425 /// @brief If there is any BE work then go work on it.
426 /// @param pContext - pointer to SWR context.
427 /// @param workerId - The unique worker ID that is assigned to this thread.
428 /// @param curDrawBE - This tracks the draw contexts that this thread has processed. Each worker thread
429 /// has its own curDrawBE counter and this ensures that each worker processes all the
431 /// @param lockedTiles - This is the set of tiles locked by other threads. Each thread maintains its
432 /// own set and each time it fails to lock a macrotile, because its already locked,
433 /// then it will add that tile to the lockedTiles set. As a worker begins to work
434 /// on future draws the lockedTiles ensure that it doesn't work on tiles that may
435 /// still have work pending in a previous draw. Additionally, the lockedTiles is
436 /// hueristic that can steer a worker back to the same macrotile that it had been
437 /// working on in a previous draw.
439 SWR_CONTEXT
*pContext
,
442 TileSet
& lockedTiles
,
446 // Find the first incomplete draw that has pending work. If no such draw is found then
447 // return. FindFirstIncompleteDraw is responsible for incrementing the curDrawBE.
448 uint32_t drawEnqueued
= 0;
449 if (FindFirstIncompleteDraw(pContext
, curDrawBE
, drawEnqueued
) == false)
454 uint32_t lastRetiredDraw
= pContext
->dcRing
[curDrawBE
% KNOB_MAX_DRAWS_IN_FLIGHT
].drawId
- 1;
456 // Reset our history for locked tiles. We'll have to re-learn which tiles are locked.
459 // Try to work on each draw in order of the available draws in flight.
460 // 1. If we're on curDrawBE, we can work on any macrotile that is available.
461 // 2. If we're trying to work on draws after curDrawBE, we are restricted to
462 // working on those macrotiles that are known to be complete in the prior draw to
463 // maintain order. The locked tiles provides the history to ensures this.
464 for (uint32_t i
= curDrawBE
; IDComparesLess(i
, drawEnqueued
); ++i
)
466 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[i
% KNOB_MAX_DRAWS_IN_FLIGHT
];
468 if (pDC
->isCompute
) return; // We don't look at compute work.
470 // First wait for FE to be finished with this draw. This keeps threading model simple
471 // but if there are lots of bubbles between draws then serializing FE and BE may
472 // need to be revisited.
473 if (!pDC
->doneFE
) return;
475 // If this draw is dependent on a previous draw then we need to bail.
476 if (CheckDependency(pContext
, pDC
, lastRetiredDraw
))
481 // Grab the list of all dirty macrotiles. A tile is dirty if it has work queued to it.
482 std::vector
<uint32_t> ¯oTiles
= pDC
->pTileMgr
->getDirtyTiles();
484 for (uint32_t tileID
: macroTiles
)
486 // Only work on tiles for this numa node
488 pDC
->pTileMgr
->getTileIndices(tileID
, x
, y
);
489 if (((x
^ y
) & numaMask
) != numaNode
)
494 MacroTileQueue
&tile
= pDC
->pTileMgr
->getMacroTileQueue(tileID
);
496 if (!tile
.getNumQueued())
501 // can only work on this draw if it's not in use by other threads
502 if (lockedTiles
.find(tileID
) != lockedTiles
.end())
511 RDTSC_START(WorkerFoundWork
);
513 uint32_t numWorkItems
= tile
.getNumQueued();
514 SWR_ASSERT(numWorkItems
);
518 if (pWork
->type
== DRAW
)
520 pContext
->pHotTileMgr
->InitializeHotTiles(pContext
, pDC
, tileID
);
523 while ((pWork
= tile
.peek()) != nullptr)
525 pWork
->pfnWork(pDC
, workerId
, tileID
, &pWork
->desc
);
528 RDTSC_STOP(WorkerFoundWork
, numWorkItems
, pDC
->drawId
);
532 pDC
->pTileMgr
->markTileComplete(tileID
);
534 // Optimization: If the draw is complete and we're the last one to have worked on it then
535 // we can reset the locked list as we know that all previous draws before the next are guaranteed to be complete.
536 if ((curDrawBE
== i
) && pDC
->pTileMgr
->isWorkComplete())
538 // We can increment the current BE and safely move to next draw since we know this draw is complete.
540 CompleteDrawContextInl(pContext
, pDC
);
550 // 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.
551 lockedTiles
.insert(tileID
);
557 //////////////////////////////////////////////////////////////////////////
558 /// @brief Called when FE work is complete for this DC.
559 INLINE
void CompleteDrawFE(SWR_CONTEXT
* pContext
, DRAW_CONTEXT
* pDC
)
563 if (pContext
->pfnUpdateSoWriteOffset
)
565 for (uint32_t i
= 0; i
< MAX_SO_BUFFERS
; ++i
)
567 if ((pDC
->dynState
.SoWriteOffsetDirty
[i
]) &&
568 (pDC
->pState
->state
.soBuffer
[i
].soWriteEnable
))
570 pContext
->pfnUpdateSoWriteOffset(GetPrivateState(pDC
), i
, pDC
->dynState
.SoWriteOffset
[i
]);
577 InterlockedDecrement((volatile LONG
*)&pContext
->drawsOutstandingFE
);
580 void WorkOnFifoFE(SWR_CONTEXT
*pContext
, uint32_t workerId
, uint32_t &curDrawFE
)
582 // Try to grab the next DC from the ring
583 uint32_t drawEnqueued
= GetEnqueuedDraw(pContext
);
584 while (IDComparesLess(curDrawFE
, drawEnqueued
))
586 uint32_t dcSlot
= curDrawFE
% KNOB_MAX_DRAWS_IN_FLIGHT
;
587 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[dcSlot
];
588 if (pDC
->isCompute
|| pDC
->doneFE
|| pDC
->FeLock
)
590 CompleteDrawContextInl(pContext
, pDC
);
599 uint32_t curDraw
= curDrawFE
;
600 while (IDComparesLess(curDraw
, drawEnqueued
))
602 uint32_t dcSlot
= curDraw
% KNOB_MAX_DRAWS_IN_FLIGHT
;
603 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[dcSlot
];
605 if (!pDC
->isCompute
&& !pDC
->FeLock
)
607 uint32_t initial
= InterlockedCompareExchange((volatile uint32_t*)&pDC
->FeLock
, 1, 0);
610 // successfully grabbed the DC, now run the FE
611 pDC
->FeWork
.pfnWork(pContext
, pDC
, workerId
, &pDC
->FeWork
.desc
);
613 CompleteDrawFE(pContext
, pDC
);
620 //////////////////////////////////////////////////////////////////////////
621 /// @brief If there is any compute work then go work on it.
622 /// @param pContext - pointer to SWR context.
623 /// @param workerId - The unique worker ID that is assigned to this thread.
624 /// @param curDrawBE - This tracks the draw contexts that this thread has processed. Each worker thread
625 /// has its own curDrawBE counter and this ensures that each worker processes all the
628 SWR_CONTEXT
*pContext
,
632 uint32_t drawEnqueued
= 0;
633 if (FindFirstIncompleteDraw(pContext
, curDrawBE
, drawEnqueued
) == false)
638 uint32_t lastRetiredDraw
= pContext
->dcRing
[curDrawBE
% KNOB_MAX_DRAWS_IN_FLIGHT
].drawId
- 1;
640 for (uint64_t i
= curDrawBE
; IDComparesLess(i
, drawEnqueued
); ++i
)
642 DRAW_CONTEXT
*pDC
= &pContext
->dcRing
[i
% KNOB_MAX_DRAWS_IN_FLIGHT
];
643 if (pDC
->isCompute
== false) return;
645 // check dependencies
646 if (CheckDependency(pContext
, pDC
, lastRetiredDraw
))
651 SWR_ASSERT(pDC
->pDispatch
!= nullptr);
652 DispatchQueue
& queue
= *pDC
->pDispatch
;
654 // Is there any work remaining?
655 if (queue
.getNumQueued() > 0)
657 void* pSpillFillBuffer
= nullptr;
658 uint32_t threadGroupId
= 0;
659 while (queue
.getWork(threadGroupId
))
661 ProcessComputeBE(pDC
, workerId
, threadGroupId
, pSpillFillBuffer
);
663 queue
.finishedWork();
669 template<bool IsFEThread
, bool IsBEThread
>
670 DWORD
workerThreadMain(LPVOID pData
)
672 THREAD_DATA
*pThreadData
= (THREAD_DATA
*)pData
;
673 SWR_CONTEXT
*pContext
= pThreadData
->pContext
;
674 uint32_t threadId
= pThreadData
->threadId
;
675 uint32_t workerId
= pThreadData
->workerId
;
677 bindThread(pContext
, threadId
, pThreadData
->procGroupId
, pThreadData
->forceBindProcGroup
);
679 RDTSC_INIT(threadId
);
681 uint32_t numaNode
= pThreadData
->numaId
;
682 uint32_t numaMask
= pContext
->threadPool
.numaMask
;
684 // flush denormals to 0
685 _mm_setcsr(_mm_getcsr() | _MM_FLUSH_ZERO_ON
| _MM_DENORMALS_ZERO_ON
);
687 // Track tiles locked by other threads. If we try to lock a macrotile and find its already
688 // locked then we'll add it to this list so that we don't try and lock it again.
691 // each worker has the ability to work on any of the queued draws as long as certain
692 // conditions are met. the data associated
693 // with a draw is guaranteed to be active as long as a worker hasn't signaled that he
694 // has moved on to the next draw when he determines there is no more work to do. The api
695 // thread will not increment the head of the dc ring until all workers have moved past the
697 // the logic to determine what to work on is:
698 // 1- try to work on the FE any draw that is queued. For now there are no dependencies
699 // on the FE work, so any worker can grab any FE and process in parallel. Eventually
700 // we'll need dependency tracking to force serialization on FEs. The worker will try
701 // to pick an FE by atomically incrementing a counter in the swr context. he'll keep
702 // trying until he reaches the tail.
703 // 2- BE work must be done in strict order. we accomplish this today by pulling work off
704 // the oldest draw (ie the head) of the dcRing. the worker can determine if there is
705 // any work left by comparing the total # of binned work items and the total # of completed
706 // work items. If they are equal, then there is no more work to do for this draw, and
707 // the worker can safely increment its oldestDraw counter and move on to the next draw.
708 std::unique_lock
<std::mutex
> lock(pContext
->WaitLock
, std::defer_lock
);
710 auto threadHasWork
= [&](uint32_t curDraw
) { return curDraw
!= pContext
->dcRing
.GetHead(); };
712 uint32_t curDrawBE
= 0;
713 uint32_t curDrawFE
= 0;
715 while (pContext
->threadPool
.inThreadShutdown
== false)
718 while (loop
++ < KNOB_WORKER_SPIN_LOOP_COUNT
&& !threadHasWork(curDrawBE
))
723 if (!threadHasWork(curDrawBE
))
727 // check for thread idle condition again under lock
728 if (threadHasWork(curDrawBE
))
734 if (pContext
->threadPool
.inThreadShutdown
)
740 RDTSC_START(WorkerWaitForThreadEvent
);
742 pContext
->FifosNotEmpty
.wait(lock
);
745 RDTSC_STOP(WorkerWaitForThreadEvent
, 0, 0);
747 if (pContext
->threadPool
.inThreadShutdown
)
755 RDTSC_START(WorkerWorkOnFifoBE
);
756 WorkOnFifoBE(pContext
, workerId
, curDrawBE
, lockedTiles
, numaNode
, numaMask
);
757 RDTSC_STOP(WorkerWorkOnFifoBE
, 0, 0);
759 WorkOnCompute(pContext
, workerId
, curDrawBE
);
764 WorkOnFifoFE(pContext
, workerId
, curDrawFE
);
768 curDrawBE
= curDrawFE
;
775 template<> DWORD workerThreadMain
<false, false>(LPVOID
) = delete;
777 template <bool IsFEThread
, bool IsBEThread
>
778 DWORD
workerThreadInit(LPVOID pData
)
784 return workerThreadMain
<IsFEThread
, IsBEThread
>(pData
);
788 __except(EXCEPTION_CONTINUE_SEARCH
)
796 template<> DWORD workerThreadInit
<false, false>(LPVOID pData
) = delete;
798 void CreateThreadPool(SWR_CONTEXT
*pContext
, THREAD_POOL
*pPool
)
800 bindThread(pContext
, 0);
803 uint32_t numThreadsPerProcGroup
= 0;
804 CalculateProcessorTopology(nodes
, numThreadsPerProcGroup
);
806 uint32_t numHWNodes
= (uint32_t)nodes
.size();
807 uint32_t numHWCoresPerNode
= (uint32_t)nodes
[0].cores
.size();
808 uint32_t numHWHyperThreads
= (uint32_t)nodes
[0].cores
[0].threadIds
.size();
810 // Calculate num HW threads. Due to asymmetric topologies, this is not
811 // a trivial multiplication.
812 uint32_t numHWThreads
= 0;
813 for (auto& node
: nodes
)
815 for (auto& core
: node
.cores
)
817 numHWThreads
+= (uint32_t)core
.threadIds
.size();
821 uint32_t numNodes
= numHWNodes
;
822 uint32_t numCoresPerNode
= numHWCoresPerNode
;
823 uint32_t numHyperThreads
= numHWHyperThreads
;
825 if (pContext
->threadInfo
.MAX_NUMA_NODES
)
827 numNodes
= std::min(numNodes
, pContext
->threadInfo
.MAX_NUMA_NODES
);
830 if (pContext
->threadInfo
.MAX_CORES_PER_NUMA_NODE
)
832 numCoresPerNode
= std::min(numCoresPerNode
, pContext
->threadInfo
.MAX_CORES_PER_NUMA_NODE
);
835 if (pContext
->threadInfo
.MAX_THREADS_PER_CORE
)
837 numHyperThreads
= std::min(numHyperThreads
, pContext
->threadInfo
.MAX_THREADS_PER_CORE
);
840 #if defined(_WIN32) && !defined(_WIN64)
841 if (!pContext
->threadInfo
.MAX_WORKER_THREADS
)
843 // Limit 32-bit windows to bindable HW threads only
844 if ((numCoresPerNode
* numHWHyperThreads
) > 32)
846 numCoresPerNode
= 32 / numHWHyperThreads
;
851 // Calculate numThreads
852 uint32_t numThreads
= numNodes
* numCoresPerNode
* numHyperThreads
;
853 numThreads
= std::min(numThreads
, numHWThreads
);
855 if (pContext
->threadInfo
.MAX_WORKER_THREADS
)
857 uint32_t maxHWThreads
= numHWNodes
* numHWCoresPerNode
* numHWHyperThreads
;
858 numThreads
= std::min(pContext
->threadInfo
.MAX_WORKER_THREADS
, maxHWThreads
);
861 if (numThreads
> KNOB_MAX_NUM_THREADS
)
863 printf("WARNING: system thread count %u exceeds max %u, "
864 "performance will be degraded\n",
865 numThreads
, KNOB_MAX_NUM_THREADS
);
868 uint32_t numAPIReservedThreads
= 1;
873 // If only 1 worker threads, try to move it to an available
874 // HW thread. If that fails, use the API thread.
875 if (numCoresPerNode
< numHWCoresPerNode
)
879 else if (numHyperThreads
< numHWHyperThreads
)
883 else if (numNodes
< numHWNodes
)
889 pPool
->numThreads
= 0;
890 SET_KNOB(SINGLE_THREADED
, true);
896 // Save HW threads for the API if we can
897 if (numThreads
> numAPIReservedThreads
)
899 numThreads
-= numAPIReservedThreads
;
903 numAPIReservedThreads
= 0;
907 pPool
->numThreads
= numThreads
;
908 pContext
->NumWorkerThreads
= pPool
->numThreads
;
910 pPool
->inThreadShutdown
= false;
911 pPool
->pThreadData
= (THREAD_DATA
*)malloc(pPool
->numThreads
* sizeof(THREAD_DATA
));
914 if (pContext
->threadInfo
.MAX_WORKER_THREADS
)
916 bool bForceBindProcGroup
= (numThreads
> numThreadsPerProcGroup
);
917 uint32_t numProcGroups
= (numThreads
+ numThreadsPerProcGroup
- 1) / numThreadsPerProcGroup
;
918 // When MAX_WORKER_THREADS is set we don't bother to bind to specific HW threads
919 // But Windows will still require binding to specific process groups
920 for (uint32_t workerId
= 0; workerId
< numThreads
; ++workerId
)
922 pPool
->pThreadData
[workerId
].workerId
= workerId
;
923 pPool
->pThreadData
[workerId
].procGroupId
= workerId
% numProcGroups
;
924 pPool
->pThreadData
[workerId
].threadId
= 0;
925 pPool
->pThreadData
[workerId
].numaId
= 0;
926 pPool
->pThreadData
[workerId
].coreId
= 0;
927 pPool
->pThreadData
[workerId
].htId
= 0;
928 pPool
->pThreadData
[workerId
].pContext
= pContext
;
929 pPool
->pThreadData
[workerId
].forceBindProcGroup
= bForceBindProcGroup
;
930 pPool
->threads
[workerId
] = new std::thread(workerThreadInit
<true, true>, &pPool
->pThreadData
[workerId
]);
932 pContext
->NumBEThreads
++;
933 pContext
->NumFEThreads
++;
938 pPool
->numaMask
= numNodes
- 1; // Only works for 2**n numa nodes (1, 2, 4, etc.)
940 uint32_t workerId
= 0;
941 for (uint32_t n
= 0; n
< numNodes
; ++n
)
943 auto& node
= nodes
[n
];
944 uint32_t numCores
= numCoresPerNode
;
945 for (uint32_t c
= 0; c
< numCores
; ++c
)
947 if (c
>= node
.cores
.size())
952 auto& core
= node
.cores
[c
];
953 for (uint32_t t
= 0; t
< numHyperThreads
; ++t
)
955 if (t
>= core
.threadIds
.size())
960 if (numAPIReservedThreads
)
962 --numAPIReservedThreads
;
966 SWR_ASSERT(workerId
< numThreads
);
968 pPool
->pThreadData
[workerId
].workerId
= workerId
;
969 pPool
->pThreadData
[workerId
].procGroupId
= core
.procGroup
;
970 pPool
->pThreadData
[workerId
].threadId
= core
.threadIds
[t
];
971 pPool
->pThreadData
[workerId
].numaId
= n
;
972 pPool
->pThreadData
[workerId
].coreId
= c
;
973 pPool
->pThreadData
[workerId
].htId
= t
;
974 pPool
->pThreadData
[workerId
].pContext
= pContext
;
976 pPool
->threads
[workerId
] = new std::thread(workerThreadInit
<true, true>, &pPool
->pThreadData
[workerId
]);
977 pContext
->NumBEThreads
++;
978 pContext
->NumFEThreads
++;
987 void DestroyThreadPool(SWR_CONTEXT
*pContext
, THREAD_POOL
*pPool
)
989 if (!pContext
->threadInfo
.SINGLE_THREADED
)
991 // Inform threads to finish up
992 std::unique_lock
<std::mutex
> lock(pContext
->WaitLock
);
993 pPool
->inThreadShutdown
= true;
995 pContext
->FifosNotEmpty
.notify_all();
998 // Wait for threads to finish and destroy them
999 for (uint32_t t
= 0; t
< pPool
->numThreads
; ++t
)
1001 pPool
->threads
[t
]->join();
1002 delete(pPool
->threads
[t
]);
1005 // Clean up data used by threads
1006 free(pPool
->pThreadData
);