Question

I have a problem that i can’t solve.

The problem is as follows.

CPP code

const int dataSize = 65535;
const int category = 10;
float data[dataSize][category];
const float threshold = 0.5f;

int cnt = 0;

// data array contains any values

for(int i=0;i<dataSize;i++)
{
    if( data[i][9] > threshold )
    {
        data[cnt][0] = data[i][0];
        data[cnt][1] = data[i][1];
        data[cnt][2] = data[i][2];
        data[cnt][3] = data[i][3];
        data[cnt][4] = data[i][4];
        data[cnt][5] = data[i][5];
        data[cnt][6] = data[i][6];
        data[cnt][7] = data[i][7];
        data[cnt][8] = data[i][8];
        data[cnt][9] = data[i][9];
        cnt++;
    }
}

By using this code, I expect ‘data’ array’s element is collected over threshold value.(The element that is not over the threshold is not important to me. Important thing is just over threshold value.)

I wanna code that operates with same result in CUDA.

So I tried to do like this.

CUDA code

__global__ void checkOverThreshold(float *data, float threshold, int *nCount)
{
    int idx = threadIdx.x + blockIdx.x * blockDim.x;

    if( data[idx*10+9] > threshold )
    {
        data[nCount+0] = data[idx*10+0];
        data[nCount+1] = data[idx*10+1];
        data[nCount+2] = data[idx*10+2];
        data[nCount+3] = data[idx*10+3];
        data[nCount+4] = data[idx*10+4];
        data[nCount+5] = data[idx*10+5];
        data[nCount+6] = data[idx*10+6];
        data[nCount+7] = data[idx*10+7];
        data[nCount+8] = data[idx*10+8];
        data[nCount+9] = data[idx*10+9];
        atomicAdd( nCount, 1);
    }
}

....

// kernel function call
checkOverThreshold<<< dataSize / 128, 128 >>>(d_data, treshold, d_count);

But the result of CUDA code is not that I expected.

It contains lots of trash value and even the result is not the same as CPP’s.

I think that the nCount variable’s synchronization problem makes this situation.

But, I have no idea to solve this problem.

Please help my code. Thank you in advance.

Answer 1

This code is broken:

    data[nCount+0] = data[idx*10+0];
    data[nCount+1] = data[idx*10+1];
    data[nCount+2] = data[idx*10+2];
    data[nCount+3] = data[idx*10+3];
    data[nCount+4] = data[idx*10+4];
    data[nCount+5] = data[idx*10+5];
    data[nCount+6] = data[idx*10+6];
    data[nCount+7] = data[idx*10+7];
    data[nCount+8] = data[idx*10+8];
    data[nCount+9] = data[idx*10+9];
    atomicAdd( nCount, 1);

If nCount is modified during all those assignments, nonsense will result. It should be

    int d = atomicAdd(nCount, 1);
    data[d+0] = data[idx*10+0];
    data[d+1] = data[idx*10+1];
    data[d+2] = data[idx*10+2];
    data[d+3] = data[idx*10+3];
    data[d+4] = data[idx*10+4];
    data[d+5] = data[idx*10+5];
    data[d+6] = data[idx*10+6];
    data[d+7] = data[idx*10+7];
    data[d+8] = data[idx*10+8];
    data[d+9] = data[idx*10+9];