len.cu

const int N_VECTORS = 1<<20;
const int VECTOR_COMPONENTS = 512;
const int KBlockSize = 256;

__global__
void calculateDistances(const float v[VECTOR_COMPONENTS], float *vectors, float *distances)
{
    int vectorIndex = threadIdx.x + threadIdx.y * KBlockSize;
    if (vectorIndex < N_VECTORS) {
        float *vector = vectors + vectorIndex * VECTOR_COMPONENTS;
        float sum = 0.0f;
        for (int i = 0; i < VECTOR_COMPONENTS; i++) {
            float a = vector[i];
            float b = v[i];
            float diff = a - b;
            sum += diff * diff;
        }

        distances[vectorIndex] = sqrtf(sum);
    }
}

main.cpp

int main()
{
  ...
  float v[VECTOR_COMPONENTS];
  cudaMallocManaged(&vectors, sizeof(float) * N_VECTORS * VECTOR_COMPONENTS);
  cudaMallocManaged(&distance, sizeof(float) * N_VECTORS);
  calculateDistance<<<N_VECTORS/KBlockSize + 1, KBlockSize>>>(v, vectors, distance);
  ...
}