1

编辑：

添加了更多日志信息和推送和弹出相关功能。

设置：Geforce GT520，Windows 64 位（编译为 32 位），Cuda 4.2。

在我的代码片段中，我有工作双端队列，它们为线程提供工作。每个块都有自己的双端队列，可以在其底部推送或弹出动态生成的工作项（popWork 和 pushWork 函数）。如果它自己的双端队列中的工作项数量小于阈值，popWork() 也可以从其他双端队列中窃取工作。

template <class TreeNode, class BV , int iDequeSize , int iFrontSize>
__global__ void traverseTree(const TreeNode* tree_object1, const GPUVertex* vertex_object1, const uint3* tri_object1,
int2* aBvttDeques, int* aiBvttBottoms, unsigned int* auiBvttAges, int *aiBvttDequeFlags , int *piOverflowFlag , 
int2* outputList, unsigned int* outputListIdx , int2* aFrontDeques , int *auiFrontBottoms)
{   
int iTid = threadIdx.x;
int iBid = blockIdx.x;

__shared__ int2 aLocalBvtt[BVTT_DEQUE_SHARED_SIZE];
__shared__ int2 aLocalFront[FRONT_DEQUE_SHARED_SIZE];
__shared__ int iLocalBvttCounter[WORK_STEALING_THREADS];
__shared__ int iLocalFrontCounter[WORK_STEALING_THREADS];
__shared__ unsigned int uiPopDequeIdx;
__shared__ int2 pushOrPopStartIdxAndSize;
__shared__ bool bPopOrPushFlag;
__shared__ unsigned int uiActiveDeques[NDEQUES]; //Contains indices for deques with useful work that can be stolen
__shared__ unsigned int uiActiveDequesIdx;

//Debug
int iRun = 0;
//

while(/*true*/ /*Debug*/iRun++ < 10) //Work loop will continue until cannot pop from bottom or cannot steal work from other deques
{
    //Debug
    if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
    {
        printf("(%d,%d) before sync0\n" , iBid , iTid);
    }

    __syncthreads();

    if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
    {
        printf("(%d,%d) after sync0\n" , iBid , iTid);
    }
    //

    int2 aWork_items[3];
    int2 aFront_item;
    int iBvttCount = 0;
    int iFrontCount = 0;

    if(!popWork<int2 , iDequeSize>(aiBvttDequeFlags , aiBvttBottoms , auiBvttAges , aBvttDeques , iTid , iBid ,
        uiPopDequeIdx , pushOrPopStartIdxAndSize , /*iLocalBvttCounter , iLocalFrontCounter ,*/ bPopOrPushFlag ,
        uiActiveDeques , uiActiveDequesIdx , aWork_items[0]))
    {   //No more work

        //Debug
        if(iTid == 0)
        {
            printf("(%d,%d)no work\n" , iBid , iTid);
        }
        //

        return; 
    }

    //Debug
    if(iTid == 0 && iBid == 0)
    {
        printf("(%d,%d) run=%d work=(%d,%d)\n" , iBid , iTid , iRun , aWork_items[0].x , aWork_items[0].y);
    }
    //

    if(iTid < pushOrPopStartIdxAndSize.y)
    {
        TreeNode node1      = tree_object1[aWork_items[0].x];
        TreeNode node2      = tree_object1[aWork_items[0].y];

        if(aWork_items[0].x == aWork_items[0].y)
        {   //intra-collision test (self collision)
            intraCollision<TreeNode , BV>(iBvttCount , aWork_items , aWork_items[0] , node1 , tree_object1);
        }
        else
        {   //inter-collision test
            //interCollision<TreeNode , BV>(iLocalBvttCounter , aLocalBvtt , outputList , outputListIdx , work_item ,
            //  node1 , node2 , tree_object1 , tri_object1 , aFrontDeques , auiFrontBottoms[iBid]);

            interCollision<TreeNode , BV>(iBvttCount , aWork_items , outputList , outputListIdx , aWork_items[0] ,
                node1 , node2 , tree_object1 , tri_object1 , aFront_item , iFrontCount);
        }
        //__syncthreads();
    }
    iLocalBvttCounter[iTid] = iBvttCount;
    iLocalFrontCounter[iTid] = iFrontCount;

    //Debug
    if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
    {
        printf("(%d,%d) before sync1\n" , iBid , iTid);
    }
    //

    __syncthreads();

    //Debug
    if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
    {
        printf("(%d,%d) after sync1\n" , iBid , iTid);
    }
    //

    {
        int iPrefixSum = prefixSum<WORK_STEALING_THREADS>(iLocalFrontCounter , iTid , iFrontCount);
        if(iFrontCount)
        {
            aLocalFront[iPrefixSum] = aFront_item;
        }
        if(iTid == WORK_STEALING_THREADS - 1)
        {
            iLocalFrontCounter[iTid] = iPrefixSum + iFrontCount;
        }

        iPrefixSum = prefixSum<WORK_STEALING_THREADS>(iLocalBvttCounter , iTid , iBvttCount);

        //Debug
        if(iTid == 0 && iBid == 0)
        {
            printf("(%d,%d) nChildren=%d prefixSum=%d\n" , iBid , iTid , iBvttCount , iPrefixSum);
            for(int i = 0 ; i < iBvttCount ; ++i)
            {
                printf("(%d,%d) children %d=(%d,%d)\n" , iBid , iTid , i , aWork_items[i].x , aWork_items[i].y);
            }
        }
        if(iTid == 1 && iBid == 0)
        {
            printf("(%d,%d) nprefixSumt2=%d\n" , iBid , iTid , iPrefixSum);
        }
        //

        for(int i = 0 ; i < iBvttCount ; ++i)
        {
            aLocalBvtt[iPrefixSum + i] = aWork_items[i];
        }
        if(iTid == WORK_STEALING_THREADS - 1)
        {
            iLocalBvttCounter[iTid] = iPrefixSum + iBvttCount;

            //Debug
            if(iBid == 0)
            {
                printf("(%d,%d) totalWork=%d, prefix=%d + count=%d\n" , iBid , iTid , iLocalBvttCounter[iTid] , iPrefixSum , iBvttCount);
            }
            //
        }
    }

    //Debug
    if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
    {
        printf("(%d,%d) before sync2\n" , iBid , iTid);
    }
    //

    __syncthreads();

    //Debug
    if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
    {
        printf("(%d,%d) after sync2\n" , iBid , iTid);
    }
    //

    //TODO: push work back only when memory size is good for better coallescence
    //Push back front to global mem
    if(!pushWork<int2 , iFrontSize>(auiFrontBottoms , aFrontDeques , iTid , iBid ,
        iLocalFrontCounter[WORK_STEALING_THREADS - 1] , aLocalFront , bPopOrPushFlag , pushOrPopStartIdxAndSize))
    {   //overflow
        if(iTid == 0)
        {
            //Debug
            printf("(%d,%d) front overflow\n" , iBid , iTid);
            //
            *piOverflowFlag = 1;
            atomicExch(&aiBvttDequeFlags[iBid] , 0);
        }
        return;
    }

    //Debug
    if(iTid == 0 && iBid == 0)
    {
        printf("(%d,%d) localnWork=%d\n" , iBid , iTid , iLocalBvttCounter[WORK_STEALING_THREADS - 1]);
    }
    //

    //Push back BVTT nodes to global mem
    if(!pushWork<int2 , iDequeSize>(aiBvttBottoms , aBvttDeques , iTid , iBid , iLocalBvttCounter[WORK_STEALING_THREADS - 1] ,
        aLocalBvtt , bPopOrPushFlag , pushOrPopStartIdxAndSize))
    {   //overflow
        if(iTid == 0)
        {
            //Debug
            printf("(%d,%d) bvtt overflow\n" , iBid , iTid);
            //

            *piOverflowFlag = 1;
            atomicExch(&aiBvttDequeFlags[iBid] , 0);
        }
        return;
    }

    //Debug
    if(iTid == 0 && iBid == 0)
    {
        printf("(%d,%d) bot=%d\n\n" , iBid , iTid , aiBvttBottoms[iBid]);
    }
    //

    //Debug
    if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
    {
        printf("(%d,%d) before sync4\n" , iBid , iTid);
    }

    __syncthreads();

    if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
    {
        printf("(%d,%d) after sync4\n" , iBid , iTid);
    }
    //
}

//Debug
if(iTid == 0)
{
    printf("(%d,%d) max iter\n" , iBid , iTid);
}
//
}

popWork() 相关代码：

bool __inline__ __device__ popTop(int *aiDequesBottoms , unsigned int *auiDequesAges , const int &iBid ,
int2 &popStartIdxAndSize)
{
int index;
unsigned int oldAge = auiDequesAges[iBid];
int localBot = aiDequesBottoms[iBid];
index = oldAge >> WORK_STEALING_TAG_NBITS;
if(localBot < index + WORK_STEALING_POP_SIZE + WORK_STEALING_PUSH_SIZE)
{
    return false;
}

int localTag = oldAge & WORK_STEALING_TAG_MASK;
int size = min(WORK_STEALING_POP_SIZE , localBot - index);
unsigned int newAge = (index+size << WORK_STEALING_TAG_NBITS)| localTag;

if(oldAge == atomicCAS(&auiDequesAges[iBid] , oldAge , newAge))
{
    popStartIdxAndSize.x = index;
    popStartIdxAndSize.y = size;
    return true;
}
else
{
    return false;
}
}

bool __inline__ __device__ popBottom(int *aiDequesBottoms , unsigned int *auiDequesAges , const int &iBid ,
int2 &popStartIdxAndSize)
{
int localBot = aiDequesBottoms[iBid];
if(localBot == 0)
{
    return false;
}

int index = localBot;
localBot = localBot - WORK_STEALING_POP_SIZE;
atomicExch(&aiDequesBottoms[iBid] , localBot);

unsigned int oldAge = auiDequesAges[iBid];
int oldAgeTop = int(oldAge >> WORK_STEALING_TAG_NBITS);
if(localBot > oldAgeTop)
{
    popStartIdxAndSize.y = WORK_STEALING_POP_SIZE;
    popStartIdxAndSize.x = index - WORK_STEALING_POP_SIZE;
    return true;
}

atomicExch(&aiDequesBottoms[iBid] , 0);
unsigned int newAge = ((oldAge & WORK_STEALING_TAG_MASK) + 1) % (WORK_STEALING_TAG_MASK + 1);
if(index > oldAgeTop)
{
    if(oldAge == atomicCAS(&auiDequesAges[iBid] , oldAge , newAge))
    {
        popStartIdxAndSize.y = index - oldAgeTop;
        popStartIdxAndSize.x = index - popStartIdxAndSize.y;
        return true;
    }
}

atomicExch(&auiDequesAges[iBid] , newAge);
return false;
}

template <typename Work , int iDequeSize>
bool __inline__ __device__ popWork(int *aiDequeFlags , int *aiDequesBottoms , unsigned int *auiDequesAges ,
const Work *aDeques , const int &iTid , const int &iBid , unsigned int &uiPopDequeIdx , int2 &popStartIdxAndSize ,
/*int &iLocalDequeCounter , int &iLocalFrontCounter ,*/ bool &bPopFlag , unsigned int *uiActiveDeques ,
unsigned int &uiActiveDequesIdx , Work &work)
{
if(iTid == 0)
{   //Try to pop from block deque
    //iLocalDequeCounter = 0;
    //iLocalFrontCounter = 0;

    bPopFlag = popBottom(aiDequesBottoms , auiDequesAges , iBid , popStartIdxAndSize);

    if(bPopFlag)
    {
        uiPopDequeIdx = iBid;
    }
    else
    {
        atomicExch(&aiDequeFlags[iBid] , 0);
    }
}
__syncthreads();

while(!bPopFlag)
{   //No more work, try to steal some (Help, police! We have a burglar here)!
    if(iTid == 0)
    {
        uiActiveDequesIdx = 0;

        //Debug
        /*if(iBid == 6 || iBid == 1)
        {
            printf("bid=%d dequeFlags:[%d,%d,%d,%d,%d,%d,%d,%d]\n" , iBid , aiDequeFlags[0] , aiDequeFlags[1] ,
                aiDequeFlags[2] , aiDequeFlags[3] , aiDequeFlags[4] , aiDequeFlags[5] , aiDequeFlags[6] ,
                aiDequeFlags[7]);

            printf("bId=%d dequesCounts:[%d,%d,%d,%d,%d,%d,%d,%d]\n" , iBid , aiDequesBottoms[0] ,
                aiDequesBottoms[1] , aiDequesBottoms[2] , aiDequesBottoms[3] , aiDequesBottoms[4] ,
                aiDequesBottoms[5] , aiDequesBottoms[6] , aiDequesBottoms[7]);
        }*/
        //
    }
    __syncthreads();

    if(iTid < NDEQUES)
    {
        if(aiDequeFlags[iTid] == 1) //assuming iTid >= NDEQUES
        {   //Set this deque for a work stealing atempt.
            unsigned int uiIdx = atomicAdd(&uiActiveDequesIdx,1);
            uiActiveDeques[uiIdx] = iTid;
        }
    }
    __syncthreads();

    if(iTid == 0)
    {   //Try to steal until succeeds or there are no more deques left to search
        bPopFlag = false;
        for(uiPopDequeIdx = 0 ; uiPopDequeIdx < uiActiveDequesIdx; ++uiPopDequeIdx)
        {
            bPopFlag = popTop(aiDequesBottoms , auiDequesAges , uiPopDequeIdx , popStartIdxAndSize);
            if(bPopFlag)
            {
                atomicExch(&aiDequeFlags[iBid] , 1);
                break;
            }
        }
    }
    __syncthreads();

    if(uiActiveDequesIdx == 0)
    { //No more work to steal. End.
        return false;
    }
    __syncthreads();
}

//Get poped data
if(iTid < popStartIdxAndSize.y) //assuming number of threads >= WORK_SIZE
{
    work = aDeques[uiPopDequeIdx*iDequeSize + popStartIdxAndSize.x + iTid];

    //Debug
    /*if(iTid == 20 && iBid == 0)
    {
        printf("work=(%d,%d) deque=%d dSize=%d start=%d final=%d\n" , work.x , work.y , uiPopDequeIdx , iDequeSize ,
            popStartIdxAndSize.x , uiPopDequeIdx*iDequeSize + popStartIdxAndSize.x + iTid);
    }*/
    //
}

return true;
}

pushWork() 相关代码：

template<int iDequeSize>
bool __inline__ __device__ pushBottom(int *aiDequeBottoms , const int &iBid , int2 &pushStartIdxAndSize)    
{
int iOldBot = aiDequeBottoms[iBid];
pushStartIdxAndSize.x = iOldBot;
iOldBot += pushStartIdxAndSize.y;

if(iOldBot < iDequeSize)
{
    atomicExch(&aiDequeBottoms[iBid] , iOldBot);
    return true;
}

return false;
}


template <typename Work , int iDequeSize>
bool __inline__ __device__ pushWork(int *aiDequesBottoms , Work *aDeques , const int &iTid , const int &iBid ,
const unsigned int &uiDequeOutputCounter , Work *aOutputLocalWork , bool &bPushFlag , int2 &pushStartIdxAndSize)
{
if(uiDequeOutputCounter == 0)
{
    return true;
}

pushStartIdxAndSize.y = uiDequeOutputCounter;

if(iTid == 0)
{
    bPushFlag = pushBottom<iDequeSize>(aiDequesBottoms , iBid , pushStartIdxAndSize);
}

//Debug
if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
{
    printf("(%d,%d) before push sync\n" , iBid , iTid);
}
//
__syncthreads();

//Debug
if(iBid == 0 && (iTid == 0 || iTid == WORK_STEALING_THREADS - 1))
{
    printf("(%d,%d) after push sync\n" , iBid , iTid);
}
//

if(!bPushFlag)
{
    return false;
}

//Transfer to global mem.
//unsigned int uiWorkLeft = uiDequeOutputCounter;
unsigned int uiThreadOffset = iTid;

//Debug
//int iRun = 0;
//

while(uiThreadOffset < uiDequeOutputCounter)
{
    //Debug
    /*if(iTid == 0 && iBid == 0)
    {
        printf("workLeft=%d bot=%d\n" , uiWorkLeft , aiDequesBottoms[iBid]);
    }*/
    //
        //Debug
        /*if(iBid == 0)
        {
            printf("tid=%d run=%d pushStartIdx.x=%d final=%d\n" , iTid , iRun , pushStartIdxAndSize.x ,
                iDequeSize*iBid + pushStartIdxAndSize.x +  iTid);
        }*/
        //

        //aDeques[iDequeSize*iBid + pushStartIdxAndSize.x +  iTid] = aOutputLocalWork[uiThreadOffset];

    atomicExch(&(aDeques[iDequeSize*iBid + pushStartIdxAndSize.x + uiThreadOffset].x) , aOutputLocalWork[uiThreadOffset].x);
    atomicExch(&(aDeques[iDequeSize*iBid + pushStartIdxAndSize.x + uiThreadOffset].y) , aOutputLocalWork[uiThreadOffset].y);

    uiThreadOffset += blockDim.x;

    //Debug
    //++iRun;
    //

    //__threadfence();
}

return true;
}

我的问题是线程似乎没有在第一次迭代后等待同步点 (__syncthreads())。这由生成的日志说明。此外，此问题仅发生在发布版本上。调试构建在这方面工作正常。

(0,0) before sync0
(0,95) before sync0
(0,0) after sync0
(0,95) after sync0
(0,95) before sync1
(0,0) run=1 work=(0,0)
(0,0) before sync1
(0,0) after sync1
(0,95) after sync1
(0,0) nChildren=3 prefixSum=0
(0,95) totalWork=3, prefix=3 + count=0
(0,0) children 0=(1,2)
(0,95) before sync2
(0,0) children 1=(1,1)
(0,0) children 2=(2,2)
(0,1) nprefixSumt2=3
(0,0) before sync2
(0,0) after sync2
(0,95) after sync2
(0,0) localnWork=3
(0,95) before push sync
(0,0) before push sync
(0,0) after push sync
(0,95) after push sync
(0,95) before sync4
(0,95) after sync4 //PROBLEM HERE! (0,95) SHOULD STAY WAINTING IN SYNC4.
(0,0) bot=3

(0,0) before sync4
(0,95) before sync0 //NOW THREADS ARE SYNC ON DIFFERENT POINTS.
(0,0) after sync4
(0,95) after sync0

感谢您的关注。

4

2 回答 2

1

早期return破坏了这里的统一性。从技术上讲，从那里开始，所有__syncthreads()代码都在条件代码中，因此具有未定义的行为。请注意，如果它从未被执行，这甚至是正确的，因为它仍然会影响编译器对重新收敛点的选择。

如果某些线程无法工作，您仍然必须让它们保持活动状态才能在__syncthreads().

您可以通过消除返回并设置每个线程的“跳过代码”标志来测试这个理论。除了 __syncthreads() 之外的所有代码都应该在if (!skip_flag) {...}条件语句内。我相信这会解决这个问题（除非有不相关的问题）。

或者，您可以通过查看说明深入了解目标代码cuobjdump -sass并检查再收敛点的位置SSY。我相信你会找到同步点超过 next 的地方__syncthreads()，可能一直到内核结束。

于 2012-09-20T21:06:41.990 回答
0

只是一个观察：您的内核中有多个同步点。

可能是您的循环中有多个 __syncthreads() 。线程以不同的速度执行并且它们在不同的同步点同步？有些在第一个同步点等待，有些在下一个同步点。因此一切都崩溃了。在循环结束时同步可以解决这个问题。

于 2012-09-21T07:45:57.997 回答