CUDA kernel strange behavior, generating random values

Question

I'm trying to simulate a spring-mass system with CUDA. Below is the kernel that updates the position of the particles:

__global__ void timestep(float3 *pos, float3 *pos_antiga, float3 *aceleracao, int numParticulas) {

    int index = threadIdx.x + blockIdx.x * blockDim.x;

    if(index > (numParticulas - 1))
        return;
t
    float3 temp = pos[index];

    pos[index].x = pos[index].x + (pos[index].x - pos_antiga[index].x) * (1.0f - DAMPING) + aceleracao[index].x * TIMESTEP;
    pos[index].y = pos[index].y + (pos[index].y - pos_antiga[index].y) * (1.0f - DAMPING) + aceleracao[index].y * TIMESTEP;
    pos[index].z = pos[index].z + (pos[index].z - pos_antiga[index].z) * (1.0f - DAMPING) + aceleracao[index].z * TIMESTEP;

    pos_antiga[index] = temp;

}

The pos represents the 3D vector of the actual position, pos_antiga is the position in the previous time step, the DAMPING is 0.01 and TIMESTEP is 0.25. I'm using Verlet integration. In a system without any force, aceleracao is zero, so the pos and pos_antigo are the same before and after the kernel calls.

However, after the first iteration, CUDA returns crazy values for some axes, like 1.QNAN and -1.6241e+016. I think it's something related to block and thread sizes. The kernel call is below:

timestep<<<16, 16>>>(pos_d, pos_antiga_d, aceleracao_d, numParticulas);

So, what am I missing?

EDIT: Below is the caller code:

void timestepGPU(vector<Particula> *p) {

// vector<Particula> has all the particles of the system.

      // CPU
    float *pos;
    float *pos_antiga;
    float *aceleracao;

    // GPU
    float *pos_d;
    float *pos_antiga_d;
    float *aceleracao_d;

    // Number of particles
    int numParticulas = p->size();

    // Init
    pos = new float[numParticulas * 3];
    pos_antiga = new float[numParticulas * 3];
    aceleracao = new float[numParticulas * 3];

    // Transfering the values from the class to a plain vector
    vector<Particula>::iterator p_tmp;
    int i = 0;
    for(p_tmp = p->begin(); p_tmp != p->end(); p_tmp++)
    {
        pos[i] = (*p_tmp).getPos().f[0];
        pos[i + 1] = (*p_tmp).getPos().f[1];
        pos[i + 2] = (*p_tmp).getPos().f[2];

        pos_antiga[i] = (*p_tmp).getPosAntiga().f[0];
        pos_antiga[i + 1] = (*p_tmp).getPosAntiga().f[1];
        pos_antiga[i + 2] = (*p_tmp).getPosAntiga().f[2];

        aceleracao[i] = (*p_tmp).getAceleracao().f[0];
        aceleracao[i + 1] = (*p_tmp).getAceleracao().f[1];
        aceleracao[i + 2] = (*p_tmp).getAceleracao().f[2];

        i += 3;
    }

    // Here, I print the particle data BEFORE moving it to GPU
      cout << "PRINT PARTICLE DATA" << endl;
    for(i = 0; i < numParticulas * 3; i += 3) {
        cout << i/3 << " - Pos: " << pos[i] << " " << pos[i + 1] << " " << pos[i + 2] << " | Pos Ant: " << pos_antiga[i] << " " << pos_antiga[i + 1] << " " << pos_antiga[i + 2] << " | Acel: " << aceleracao[i] << " " << aceleracao[i + 1] << " " << aceleracao[i + 2] << endl;
    }
    cout << "END" << endl;

    // GPU
    ErroCUDA(cudaMalloc((void**) &pos_d, numParticulas * 3 * sizeof(float)));
    ErroCUDA(cudaMalloc((void**) &pos_antiga_d, numParticulas * 3 * sizeof(float)));
    ErroCUDA(cudaMalloc((void**) &aceleracao_d, numParticulas * 3 * sizeof(float)));

    // Moving data
    ErroCUDA(cudaMemcpy(pos_d, pos, numParticulas * 3 * sizeof(float), cudaMemcpyHostToDevice));
    ErroCUDA(cudaMemcpy(pos_antiga_d, pos_antiga, numParticulas * 3 * sizeof(float), cudaMemcpyHostToDevice));
    ErroCUDA(cudaMemcpy(aceleracao_d, aceleracao, numParticulas * sizeof(float), cudaMemcpyHostToDevice));

    // Setting number of blocks and threads per block
    unsigned int numThreads, numBlocos;
    calcularGrid(numParticulas, 64, numBlocos, numThreads);
    //cout << numBlocos << "----------" << numThreads << endl;

    // Kernel
    timestep<<<numBlocos, numThreads>>>((float3 *) pos_d, (float3 *) pos_antiga_d, (float3 *) aceleracao_d, numParticulas);
    ErroCUDA(cudaPeekAtLastError());
    cudaDeviceSynchronize();

    // Moving data back to the CPU
    ErroCUDA(cudaMemcpy(pos, pos_d, numParticulas * 3 * sizeof(float), cudaMemcpyDeviceToHost));
    ErroCUDA(cudaMemcpy(pos_antiga, pos_antiga_d, numParticulas * 3 * sizeof(float), cudaMemcpyDeviceToHost));

      // Printing the particles' data AFTER Kernel call. At my GT 4xx, close to the 48th particle, it starts to show crazy values
    cout << "PARTICLE DATA" << endl;
    for(i = 0; i < numParticulas * 3; i += 3) {
        cout << i/3 << " - Pos: " << pos[i] << " " << pos[i + 1] << " " << pos[i + 2] << " | Pos Ant: " << pos_antiga[i] << " " << pos_antiga[i + 1] << " " << pos_antiga[i + 2] << " | Acel: " << aceleracao[i] << " " << aceleracao[i + 1] << " " << aceleracao[i + 2] << endl;
    }
    cout << "END" << endl;

    system("pause");

    i = 0;
    for(p_tmp = p->begin(); p_tmp != p->end(); p_tmp++)
    {
        if((*p_tmp).getMovel())
        {
            (*p_tmp).setPos(Vetor(pos[i], pos[i + 1], pos[i + 2]));
            (*p_tmp).setPosAntiga(Vetor(pos_antiga[i], pos_antiga[i + 1], pos_antiga[i + 2]));
(*p_tmp).setAceleracao(Vetor(0, 0, 0));
        }

        i += 3;
    }

    ErroCUDA(cudaFree(pos_d));
    ErroCUDA(cudaFree(pos_antiga_d));
    ErroCUDA(cudaFree(aceleracao_d));

    free(pos);
    free(pos_antiga);
    free(aceleracao);
}

In my example, attribute p has 100 items (10 x 10 particles). It's a mesh starting at (0, 0, 0) and goes to (20, 20, 20) in 3D space.

Thanks again everyone for the help!

Answer 1

I think your problem is in this line..

 ErroCUDA(cudaMemcpy(aceleracao_d, aceleracao, numParticulas * sizeof(float), cudaMemcpyHostToDevice));

should be ..

  ErroCUDA(cudaMemcpy(aceleracao_d, aceleracao, numParticulas * 3 *sizeof(float), cudaMemcpyHostToDevice));