CODEWITHSUNDEEP
c++cudanvidia
Ashley Cusack
Ashley Cusack

Reputation: 63

CUDA not returning values when Array > 1769472

am trying to calculate the average 256 sets of data that are 8192 bytes long. I have a kernel that works at 216 data sets but any more and the kernel returns 0 for each average. I am using a very basic reduction system to calculate the average.

Graphics Card: GTX 780 Ti

Here is my code

__global__ void Average(double *Input, int Length, int Sets, double *Average, int N) {
    unsigned int Pos = (blockDim.x * blockIdx.x) + threadIdx.x;
    unsigned int Offset;
    int i = Length / N;
    if (Pos < i * Sets) {
        Offset = ((Pos / i) * Length) + (Pos % i); 
        Input[Offset] += Input[Offset + i];
    }
    __syncthreads();
    if (N == Length) {
        Average[Pos] = Input[Pos*Length] / Length;
    }
}

using namespace std;
int main()
{
    const int Length = 8192;
    const int Sets =256;
    const int Width = Length*Sets;
    double *GPU_Average, *GPU_Data;
    cudaMalloc((void**)&GPU_Average, CameraWidth*sizeof(double)*Sets);
    cudaMalloc((void**)&GPU_Data, CameraWidth*sizeof(double)*Width); 
    double CPU_Data[Width];
    double CPU_Average[Sets];
    for (int i = 0; i < Width; i++) {
        CPU_Data[i] = i;
    }
    cudaMemcpy(GPU_Data, CPU_Data, sizeof(double)*Width, cudaMemcpyHostToDevice);

    int N = 2;
    int Total, Blocks, Threads;
    while (N < Length+1) {
        Total = (Sets*Length) / N;
        if (Total > 1024) {
            Threads = 1024;
            Blocks = Total / Threads;
        }
        else {
            Threads = Total;
            Blocks = 1;
        }
        Average << < Blocks, Threads>> >(GPU_Data, Length, Sets, GPU_Average, N);
        N *= 2;
    }
    cudaMemcpy(CPU_Average, (GPU_Average), sizeof(double)*Sets, cudaMemcpyDeviceToHost);
    return 0;
}

Any help on this matter is appreciated.

Upvotes: 0

Views: 59

Answers (1)

Ashley Cusack
Ashley Cusack

Reputation: 63

I had not realised in my actual code (not the one above) i had wrote

 cudaMalloc((void**)&GPU_Data, Width*sizeof(double)*Width);

instead of

 cudaMalloc((void**)&GPU_Data, sizeof(double)*Width);

this was allocating too much memory and causing errors.

Upvotes: 1

Related Questions