Generate Random Number Within Cuda Kernel - c++

Question

I am trying to use the CuRand library in CUDA. I'm just simply trying to generate a random integer per thread. Below is my results (Clearly not very random):
84
84
84
84
84
5
Please check my code out and tell me what I'm doing wrong, I'm pulling my hair out trying to figure out why this isn't working...
#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>
#include <curand.h>
#include <curand_kernel.h>
__device__ float generate(curandState* globalState, int ind)
{
//int ind = threadIdx.x;
curandState localState = globalState[ind];
float RANDOM = curand_uniform( &localState );
globalState[ind] = localState;
return RANDOM;
}
__global__ void setup_kernel ( curandState * state, unsigned long seed )
{
int id = threadIdx.x;
curand_init ( seed, id, 0, &state[id] );
}
__global__ void addToCount(int N, int *y, curandState* globalState)
{
int id = threadIdx.x + blockIdx.x * blockDim.x;
while (id < N)
{
int number = generate(globalState, id) * 1000000;
printf("%i\n", number);
atomicAdd(&(y[0]), number);
id += blockDim.x * gridDim.x;
}
}
int main(void)
{
int N = 5;
int *y, *d_y;
y = (int*)malloc(N*sizeof(int));
cudaMalloc(&d_y, N * sizeof(int));
cudaMemcpy(d_y, y, N * sizeof(int), cudaMemcpyHostToDevice);
curandState* devStates;
cudaMalloc (&devStates, N * sizeof(curandState));
addToCount<<<2, 5>>>(N, d_y, devStates);
cudaMemcpy(y, d_y, N*sizeof(int), cudaMemcpyDeviceToHost);
printf("%i\n", *y);
}

Answer 1

AS #Robert Crovella mentioned in his comment, you forgot to setup the kernel. The curand states need to be initialized for every thread before they provide actual random numbers. If you change your main to:
int main(void)
{
int N = 5;
int *y, *d_y;
y = (int*)malloc(N*sizeof(int));
cudaMalloc(&d_y, N * sizeof(int));
cudaMemcpy(d_y, y, N * sizeof(int), cudaMemcpyHostToDevice);
curandState* devStates;
cudaMalloc (&devStates, N * sizeof(curandState));
srand(time(0));
/** ADD THESE TWO LINES **/
int seed = rand();
setup_kernel<<<2, 5>>>(devStates,seed);
/** END ADDITION **/
addToCount<<<2, 5>>>(N, d_y, devStates);
cudaMemcpy(y, d_y, N*sizeof(int), cudaMemcpyDeviceToHost);
printf("%i\n", *y);
}
You get nice results with default compilation:
nvcc /tmp/so.cu -o /tmp/so
$ /tmp/so
900981
469952
494161
31968
880329
2777391
$ /tmp/so
525835
742594
750423
117137
66318
2202307
$ /tmp/so
919262
60838
89868
57696
770764
1898428