c++ - CUDA 调度问题或内核启动中的错误？

Question

在我的程序中，我为内核 lauch 中发布的每个块都有一个工作双端队列。每个块都停留在循环中，在其双端队列中弹出工作，处理它并将动态生成的工作推回。维护了一组双端队列标志，指示哪些双端队列处于活动状态，即有工作。如果双端队列为空，内核进入另一个循环，试图从另一个块的双端队列中窃取工作。当没有更多的双端队列处于活动状态时，达到停止条件。

在测试中，我设置了从 1 个工作项开始的所有双端队列。我的问题是某些块似乎根本没有运行。由于其中一些没有运行，它们保持活动状态，我的程序进入无限循环。

现在到代码。内核和辅助弹出和推送功能：

bool __inline__ __device__ pushBottom(int *aiDequeBottoms , const int &iBid , const unsigned int &uiSize ,
unsigned int &uiPushStartIdx)   
{
int iOldBot = aiDequeBottoms[iBid];
uiPushStartIdx = iOldBot;
iOldBot += uiSize;
if(iOldBot < DEQUE_SIZE)
{
    aiDequeBottoms[iBid] = iOldBot;
    return true;
}
else
{
    return false;
}
}

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 + 2*WORK_STEALING_BATCH_SIZE)
{
    return false;
}

int localTag = oldAge & WORK_STEALING_TAG_MASK;
int size = min(WORK_STEALING_BATCH_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_BATCH_SIZE;
aiDequesBottoms[iBid] = localBot;
unsigned int oldAge = auiDequesAges[iBid];
int oldAgeTop = int(oldAge >> WORK_STEALING_TAG_NBITS);
if(localBot > oldAgeTop)
{
    popStartIdxAndSize.y = WORK_STEALING_BATCH_SIZE;
    popStartIdxAndSize.x = index - WORK_STEALING_BATCH_SIZE;
    return true;
}

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;
    }
}

auiDequesAges[iBid] = newAge;
return false;
}

//----------------------------------------------------------------------------------------------------------------------------
// Function to pop work from deques. Each block try to pop from its own deque. If work isn't available, it try to steal from
// other deques.
//----------------------------------------------------------------------------------------------------------------------------
template <typename Work>
bool __inline__ __device__ popWork(bool *abDequeFlags , int *aiDequesBottoms , unsigned int *auiDequesAges ,
const Work *aDeques , const int &iTid , const int &iBid , unsigned int &uiPopDequeIdx , int2 &popStartIdxAndSize ,
int &iLocalDequeCounter , bool &bPopFlag , unsigned int *uiActiveDeques , unsigned int &uiActiveDequesIdx , Work &work)
{
if(iTid == 0)
{   //Try to pop from block deque
    iLocalDequeCounter = 0;
    bPopFlag = popBottom(aiDequesBottoms , auiDequesAges , iBid , popStartIdxAndSize);
    if(bPopFlag)
    {
        uiPopDequeIdx = iBid;
    }
    else
    {
        abDequeFlags[iBid] = false;
    }
}
__syncthreads();

while(!bPopFlag)
{   //No more work, try to steal some!
    if(iTid == 0)
    {
        uiActiveDequesIdx = 0;
    }
    __syncthreads();

    if(iTid < NDEQUES)
    {
        if(abDequeFlags[iTid] == true) //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 && bPopFlag == false ; ++uiPopDequeIdx)
        {
            bPopFlag = popTop(aiDequesBottoms , auiDequesAges , uiPopDequeIdx , popStartIdxAndSize);
        }
    }
    __syncthreads();

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

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

return true;
}

//----------------------------------------------------------------------------------------------------------------------------
// Function to push work on deques. To achieve better coalescent global memory accesses the input data is assumed to be tight
// packed in shared mem.
//----------------------------------------------------------------------------------------------------------------------------
template <typename Work>
bool __inline__ __device__ pushWork(int *aiDequesBottoms , Work *aDeques , const int &iTid , const int &iBid ,
const unsigned int &uiDequeOutputCounter , Work *aOutputLocalWork)
{
//Transfer to global mem.
unsigned int uiWorkLeft = uiDequeOutputCounter;
unsigned int uiThreadOffset = iTid;

while(uiWorkLeft > 0)
{
    unsigned int uiWorkTransfered = min(WORK_STEALING_BATCH_SIZE , uiWorkLeft);
    unsigned int uiPushStartIdx;
    bool bPushFlag;
    if(iTid == 0)
    {
        bPushFlag = pushBottom(aiDequesBottoms , iBid , uiWorkTransfered , uiPushStartIdx);
    }
    __syncthreads();
    if(!bPushFlag)
    {
        return false;
    }

    if(iTid < uiWorkTransfered)
    {
        aDeques[DEQUE_SIZE*iBid + uiPushStartIdx +  uiThreadOffset] = aOutputLocalWork[uiThreadOffset];
    }

    uiThreadOffset += WORK_STEALING_BATCH_SIZE;
    uiWorkLeft -= uiWorkTransfered;
}

return true;
}

void __global__ workKernel(bool *abDequeFlags , int *aiDequesBottoms , unsigned int *auiDequesAges , int2 *aDeques , 
int2 *aOutput , unsigned int *puiOutputCounter)
{
int iTid = threadIdx.x;
int iBid = blockIdx.x;

__shared__ int2 aOutputLocalWork[DEQUE_SHARED_SIZE];
__shared__ unsigned int uiPopDequeIdx;
__shared__ int2 popStartIdxAndSize;
__shared__ int iLocalDequeCounter;
__shared__ bool bPopFlag;
__shared__ unsigned int uiActiveDeques[NDEQUES]; //Contains indices for deques with useful work that can be stolen
__shared__ unsigned int uiActiveDequesIdx;
__shared__ unsigned int uiLastOutputCounter;

int2 work;
int iRun = 0;

while(true) //Work loop will continue until cannot pop from bottom or cannot steal work from other deques
{
    if(!popWork<int2>(abDequeFlags , aiDequesBottoms , auiDequesAges , aDeques , iTid , iBid , uiPopDequeIdx ,
        popStartIdxAndSize , iLocalDequeCounter , bPopFlag , uiActiveDeques , uiActiveDequesIdx , work))
    {   //No more work
        return; 
    }

    //Useful work comes here. For now, just some dummy code for testing.
    if(iRun < 5)
    {   //Just 5 iterations that generate more work
        if(iTid < popStartIdxAndSize.y)
        {
            unsigned int uiNewWorkCounter = 1;
            int iDequeOutputCounter = atomicAdd(&iLocalDequeCounter , uiNewWorkCounter);
            work.x++; work.y++;
            aOutputLocalWork[iDequeOutputCounter] = work;
            __syncthreads();

            if(iTid == 0)
            {
                uiLastOutputCounter = atomicAdd(puiOutputCounter , iLocalDequeCounter);
            }
            __syncthreads();

            if(iTid < iLocalDequeCounter) //assuming iLocalDequeCounter <= blockDim.x
            {
                aOutput[uiLastOutputCounter + iTid] = aOutputLocalWork[iTid];
            }
        }
    }

    //Push back to global mem
    if(!pushWork<int2>(aiDequesBottoms , aDeques , iTid , iBid , iLocalDequeCounter , aOutputLocalWork))
    {   //overflow
        return;
    }
    ++iRun;
}
}

这是测试：

#define NDEQUES 256
#define DEQUE_SIZE 20000

void workStealingWrap(bool *abDequeFlags , int *auiDequesBottoms , unsigned int *auiDequesAges , int2 *aDeques , 
int2 *aOutput , unsigned int *puiOutputCounter)
{
workKernel<<<NDEQUES , WORK_STEALING_THREADS>>>(abDequeFlags , auiDequesBottoms , auiDequesAges , aDeques , aOutput , 
    puiOutputCounter);
CUT_CHECK_ERROR("workKernel");
}


//----------------------------------------------------------------------------------------------------------
// This entry point is for work stealing testing.
//----------------------------------------------------------------------------------------------------------
int main(int argc, char* argv[])
{
//Test 0: All deques start with 1 work item.
bool h_abDequeFlags[NDEQUES];
int h_aiDequesBottoms[NDEQUES];
unsigned int h_auiDequesAges[NDEQUES];
int2 *h_aDeques = (int2*) malloc(sizeof(int2)*NDEQUES*DEQUE_SIZE);
unsigned int h_uiOutputCounter;
int2 *h_aOutput = (int2*) malloc(sizeof(int2)*NDEQUES*DEQUE_SIZE);

for(int i = 0 ; i < NDEQUES ; ++i)
{
    h_abDequeFlags[i] = true;
    h_aiDequesBottoms[i] = 1;
    h_auiDequesAges[i] = 0;
    int2 work; work.x = i ; work.y = i;
    h_aDeques[DEQUE_SIZE*i] = work;
}

bool *d_abDequeFlags;
int *d_auiDequesBottoms;
unsigned int *d_auiDequesAges;
int2 *d_aDeques;
GPUMALLOC((void**)&d_abDequeFlags , sizeof(bool)*NDEQUES);
GPUMALLOC((void**)&d_auiDequesBottoms , sizeof(int)*NDEQUES);
GPUMALLOC((void**)&d_auiDequesAges , sizeof(unsigned int)*NDEQUES);
GPUMALLOC((void**)&d_aDeques , sizeof(int2)*NDEQUES*DEQUE_SIZE);

TOGPU(d_abDequeFlags , h_abDequeFlags , sizeof(bool)*NDEQUES);
TOGPU(d_auiDequesBottoms , h_aiDequesBottoms , sizeof(int)*NDEQUES);
TOGPU(d_auiDequesAges , h_auiDequesAges , sizeof(unsigned int)*NDEQUES);
TOGPU(d_aDeques , h_aDeques , sizeof(int2)*NDEQUES*DEQUE_SIZE);

int2 *d_aOutput;
unsigned int *d_puiOutputCounter;
GPUMALLOC((void**)&d_aOutput , sizeof(int2)*NDEQUES*DEQUE_SIZE);
GPUMALLOC((void**)&d_puiOutputCounter , sizeof(unsigned int));
GPUMEMSET(d_aOutput , -1 , sizeof(int2)*NDEQUES*DEQUE_SIZE);
GPUMEMSET(d_puiOutputCounter , 0 , sizeof(unsigned int));

workStealingWrap(d_abDequeFlags , d_auiDequesBottoms , d_auiDequesAges , d_aDeques , d_aOutput , d_puiOutputCounter);

FROMGPU(h_aOutput , d_aOutput , sizeof(int2)*NDEQUES*DEQUE_SIZE);
FROMGPU(&h_uiOutputCounter , d_puiOutputCounter , sizeof(unsigned int));
assert(h_uiOutputCounter == NDEQUES);
for(int i = 0 ; i < NDEQUES*DEQUE_SIZE ; ++i)
{
    int2 work = h_aOutput[i];
    if(i < NDEQUES)
    {
        assert(work.x >= 1 && work.x < NDEQUES*5 && work.y >= 1 && work.y < NDEQUES*5);
    }
    else
    {
        assert(work.x == -1 && work.y == -1);
    }
}

GPUFREE(d_abDequeFlags);
GPUFREE(d_auiDequesBottoms);
GPUFREE(d_auiDequesAges);
GPUFREE(d_aDeques);
GPUFREE(d_aOutput);
GPUFREE(d_puiOutputCounter);

safeFree(h_aDeques);
safeFree(h_aOutput);
}

使用 NSight 调试此代码我已经验证只有前 8 个块正在运行。我想知道这是否是一个调度问题并且 popWork 轮询正在消耗所有资源，或者它只是我程序中的一个错误。任何帮助将不胜感激。

score 1 · Accepted Answer

我在代码中发现了一些问题。我没有运行它，所以我不知道它在多大程度上是相关的。

popTop缺少执行。我假设它可以成功也可以失败，结果会影响bPopFlag. 我假设您可以uiActiveDequesIdx在bPopFlag仍然为零的情况下具有非零值。

在这种情况下，popWork你有一个while(!bPopFlag)循环。想象两个经线到达__syncthreads()循环中的最后一个。现在，第一个 warp 将检查uiActiveDequesIdx（假设非零），回到循环的开头，并且iTid线程设置uiActiveDequesIdx为 0。现在第二个 warp 从最后一个恢复执行__syncthreads()，检查uiActiveDequesIdx（现在是 0）并退出环形。从这一点开始，你的经线已经分道扬镳__syncthreads()，坏事将会发生。
在pushWork中，bPushFlag是本地寄存器。它仅由iTid == 0线程更改，但随后由所有线程读取。您可能会读取未初始化的值并进一步获得分歧的扭曲__syncthreads()。

可能还有更多我没有看到的问题。特别是，我担心您__threadfence()在全局内存中设置标志时可能会跳过（查看编程指南了解它的作用）

此外，请仔细检查其他循环是否存在与我上面报告的问题类似的同步问题——其中一些扭曲仍在循环的旧迭代中，而另一些则已经在新迭代中。我认为一个很好的通用方法是，如果您有一个取决于某些共享值的非发散循环，请始终将 a__syncthreads()放在循环的末尾，并且仅当您对代码进行一些后期细粒度优化时，尝试删除它.

c++ - CUDA 调度问题或内核启动中的错误？

1 回答 1

Related

Reference