//////////////////////////////////////////////////////////////////////////
/// @brief If there is any BE work then go work on it.
/// @param pContext - pointer to SWR context.
/// @param workerId - The unique worker ID that is assigned to this thread.
/// @param curDrawBE - This tracks the draw contexts that this thread has processed. Each worker thread
/// has its own curDrawBE counter and this ensures that each worker processes all the
/// draws in order.
/// @param lockedTiles - This is the set of tiles locked by other threads. Each thread maintains its
/// own set and each time it fails to lock a macrotile, because its already locked,
/// then it will add that tile to the lockedTiles set. As a worker begins to work
/// on future draws the lockedTiles ensure that it doesn't work on tiles that may
/// still have work pending in a previous draw. Additionally, the lockedTiles is
/// hueristic that can steer a worker back to the same macrotile that it had been
/// working on in a previous draw.
/// @returns true if worker thread should shutdown
bool WorkOnFifoBE(
SWR_CONTEXT *pContext,
uint32_t workerId,
uint32_t &curDrawBE,
TileSet& lockedTiles,
uint32_t numaNode,
uint32_t numaMask)
{
bool bShutdown = false;
// Find the first incomplete draw that has pending work. If no such draw is found then
// return. FindFirstIncompleteDraw is responsible for incrementing the curDrawBE.
uint32_t drawEnqueued = 0;
if (FindFirstIncompleteDraw(pContext, workerId, curDrawBE, drawEnqueued) == false)
{
return false;
}
uint32_t lastRetiredDraw = pContext->dcRing[curDrawBE % pContext->MAX_DRAWS_IN_FLIGHT].drawId - 1;
// Reset our history for locked tiles. We'll have to re-learn which tiles are locked.
lockedTiles.clear();
// Try to work on each draw in order of the available draws in flight.
// 1. If we're on curDrawBE, we can work on any macrotile that is available.
// 2. If we're trying to work on draws after curDrawBE, we are restricted to
// working on those macrotiles that are known to be complete in the prior draw to
// maintain order. The locked tiles provides the history to ensures this.
for (uint32_t i = curDrawBE; IDComparesLess(i, drawEnqueued); ++i)
{
DRAW_CONTEXT *pDC = &pContext->dcRing[i % pContext->MAX_DRAWS_IN_FLIGHT];
if (pDC->isCompute) return false; // We don't look at compute work.
// First wait for FE to be finished with this draw. This keeps threading model simple
// but if there are lots of bubbles between draws then serializing FE and BE may
// need to be revisited.
if (!pDC->doneFE) return false;
// If this draw is dependent on a previous draw then we need to bail.
if (CheckDependency(pContext, pDC, lastRetiredDraw))
{
return false;
}
// Grab the list of all dirty macrotiles. A tile is dirty if it has work queued to it.
auto ¯oTiles = pDC->pTileMgr->getDirtyTiles();
for (auto tile : macroTiles)
{
uint32_t tileID = tile->mId;
// Only work on tiles for this numa node
uint32_t x, y;
pDC->pTileMgr->getTileIndices(tileID, x, y);
if (((x ^ y) & numaMask) != numaNode)
{
continue;
}
if (!tile->getNumQueued())
{
continue;
}
// can only work on this draw if it's not in use by other threads
if (lockedTiles.find(tileID) != lockedTiles.end())
{
continue;
}
if (tile->tryLock())
{
BE_WORK *pWork;
RDTSC_BEGIN(WorkerFoundWork, pDC->drawId);
uint32_t numWorkItems = tile->getNumQueued();
SWR_ASSERT(numWorkItems);
pWork = tile->peek();
SWR_ASSERT(pWork);
if (pWork->type == DRAW)
{
pContext->pHotTileMgr->InitializeHotTiles(pContext, pDC, workerId, tileID);
}
else if (pWork->type == SHUTDOWN)
{
bShutdown = true;
}
while ((pWork = tile->peek()) != nullptr)
{
pWork->pfnWork(pDC, workerId, tileID, &pWork->desc);
tile->dequeue();
}
RDTSC_END(WorkerFoundWork, numWorkItems);
_ReadWriteBarrier();
pDC->pTileMgr->markTileComplete(tileID);
// Optimization: If the draw is complete and we're the last one to have worked on it then
// we can reset the locked list as we know that all previous draws before the next are guaranteed to be complete.
if ((curDrawBE == i) && (bShutdown || pDC->pTileMgr->isWorkComplete()))
{
// We can increment the current BE and safely move to next draw since we know this draw is complete.
curDrawBE++;
CompleteDrawContextInl(pContext, workerId, pDC);
lastRetiredDraw++;
lockedTiles.clear();
break;
}
if (bShutdown)
{
break;
}
}
else
{
// 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.
lockedTiles.insert(tileID);
}
}
}
return bShutdown;
}
//////////////////////////////////////////////////////////////////////////
/// @brief If there is any BE work then go work on it.
/// @param pContext - pointer to SWR context.
/// @param workerId - The unique worker ID that is assigned to this thread.
/// @param curDrawBE - This tracks the draw contexts that this thread has processed. Each worker thread
/// has its own curDrawBE counter and this ensures that each worker processes all the
/// draws in order.
/// @param lockedTiles - This is the set of tiles locked by other threads. Each thread maintains its
/// own set and each time it fails to lock a macrotile, because its already locked,
/// then it will add that tile to the lockedTiles set. As a worker begins to work
/// on future draws the lockedTiles ensure that it doesn't work on tiles that may
/// still have work pending in a previous draw. Additionally, the lockedTiles is
/// hueristic that can steer a worker back to the same macrotile that it had been
/// working on in a previous draw.
void WorkOnFifoBE(
SWR_CONTEXT *pContext,
uint32_t workerId,
volatile uint64_t &curDrawBE,
std::unordered_set<uint32_t>& lockedTiles)
{
// Find the first incomplete draw that has pending work. If no such draw is found then
// return. FindFirstIncompleteDraw is responsible for incrementing the curDrawBE.
if (FindFirstIncompleteDraw(pContext, curDrawBE) == false)
{
return;
}
uint64_t lastRetiredDraw = pContext->dcRing[curDrawBE % KNOB_MAX_DRAWS_IN_FLIGHT].drawId - 1;
// Reset our history for locked tiles. We'll have to re-learn which tiles are locked.
lockedTiles.clear();
// Try to work on each draw in order of the available draws in flight.
// 1. If we're on curDrawBE, we can work on any macrotile that is available.
// 2. If we're trying to work on draws after curDrawBE, we are restricted to
// working on those macrotiles that are known to be complete in the prior draw to
// maintain order. The locked tiles provides the history to ensures this.
for (uint64_t i = curDrawBE; i < GetEnqueuedDraw(pContext); ++i)
{
DRAW_CONTEXT *pDC = &pContext->dcRing[i % KNOB_MAX_DRAWS_IN_FLIGHT];
if (pDC->isCompute) return; // We don't look at compute work.
// First wait for FE to be finished with this draw. This keeps threading model simple
// but if there are lots of bubbles between draws then serializing FE and BE may
// need to be revisited.
if (!pDC->doneFE) break;
// If this draw is dependent on a previous draw then we need to bail.
if (CheckDependency(pContext, pDC, lastRetiredDraw))
{
return;
}
// Grab the list of all dirty macrotiles. A tile is dirty if it has work queued to it.
std::vector<uint32_t> ¯oTiles = pDC->pTileMgr->getDirtyTiles();
for (uint32_t tileID : macroTiles)
{
MacroTileQueue &tile = pDC->pTileMgr->getMacroTileQueue(tileID);
// can only work on this draw if it's not in use by other threads
if (lockedTiles.find(tileID) == lockedTiles.end())
{
if (tile.getNumQueued())
{
if (tile.tryLock())
{
BE_WORK *pWork;
RDTSC_START(WorkerFoundWork);
uint32_t numWorkItems = tile.getNumQueued();
if (numWorkItems != 0)
{
pWork = tile.peek();
SWR_ASSERT(pWork);
if (pWork->type == DRAW)
{
InitializeHotTiles(pContext, pDC, tileID, (const TRIANGLE_WORK_DESC*)&pWork->desc);
}
}
while ((pWork = tile.peek()) != nullptr)
{
pWork->pfnWork(pDC, workerId, tileID, &pWork->desc);
tile.dequeue();
}
RDTSC_STOP(WorkerFoundWork, numWorkItems, pDC->drawId);
_ReadWriteBarrier();
pDC->pTileMgr->markTileComplete(tileID);
// Optimization: If the draw is complete and we're the last one to have worked on it then
// we can reset the locked list as we know that all previous draws before the next are guaranteed to be complete.
if ((curDrawBE == i) && pDC->pTileMgr->isWorkComplete())
{
// We can increment the current BE and safely move to next draw since we know this draw is complete.
curDrawBE++;
lastRetiredDraw++;
lockedTiles.clear();
break;
}
}
else
{
// 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.
lockedTiles.insert(tileID);
}
}
}
}
}
}
