2

I am writing random generation code in CUDA using the CURAND library. What I read about random generation made me believe that if I use the same seed, I will get the same set of random numbers. But its not the case when I tested it. Please explain what am I doing wrong. I am pasting the code below for reference: 

    curandGenerator_t rand_gen;  
    status = curandCreateGenerator (&rand_gen ,CURAND_RNG_PSEUDO_DEFAULT );  
    if(status != CURAND_STATUS_SUCCESS){  
            printf("Error encountered in generating handle\n");  
    }  
    status = curandSetPseudoRandomGeneratorSeed (rand_gen ,1234ULL);  
    if(status != CURAND_STATUS_SUCCESS){  
            printf("Error encountered in setting seed\n");  
    }  

    for(j=0; j<2; j++){  
            status = curandGenerate(rand_gen,a_d,N);  
            if(status != CURAND_STATUS_SUCCESS){  
                    printf("Error encountered in generating random numbers\n");  
             }

            cudaMemcpy ( a_h , a_d , N * sizeof(unsigned int),cudaMemcpyDeviceToHost);  
            for(i = 0; i < N; i++){  
                    printf("%d : %u\n",i,a_h[i]);  
            }
            printf("-----------%d----------------------\n",j);  
    }  
    status = curandDestroyGenerator(rand_gen);  
    if(status != CURAND_STATUS_SUCCESS){  
            printf("Error encountered in destroying handle\n");  
    }

Output:

0 : 624778773
1 : 3522650202
2 : 2363946744
3 : 1266286439
4 : 3928747533
5 : 3732235839
6 : 1382638835
7 : 3362343509
8 : 48542993
9 : 1225999208
-----------0----------------------
0 : 3356973615
1 : 1004333919
2 : 2916556602
3 : 1213079917
4 : 2705410958
5 : 520650207
6 : 1860816870
7 : 1645310928
8 : 2205755199
9 : 1282999252
-----------1----------------------

talonmies
  • 70,661
  • 34
  • 192
  • 269
user1439690
  • 659
  • 1
  • 11
  • 26

2 Answers2

5

there is a notion of "state" of a pseudo-random generator. For example, Mersenne twister has a state of size about 1024 words while the the default one XORWOW has a state size just a several words (but it also has much smaller period).

Whenever you call 'setPseudoRandomGeneratorSeed' you initialize the state of the generator. Then with subsequent calls to curandGenerate this state will be updated (i.e. to go from one random number to the next one, the state needs to be recomputed) and hence different parts of the random sequence will be generated.

You might also experiment with the driver API - here curandInit() initializes the state for each thread which can be quite expensive. Then the subsequent calls to curand() or curandUniform() etc. will reuse this state. In fact each thread starts from different offset of a random sequence

  • Thank you for the answer. So you mean to say that I need not generate different seeds for different random numbers and that even if I use the same seed I will get different set of random numbers? – user1439690 Sep 04 '12 at 04:14
  • Yes there is no need to generate random seeds for each call to curandGenerate(). You can just set a seed once per program run, ie. using 'time(NULL)' for example –  Sep 04 '12 at 13:37
  • I think I got the explanation. The cuRandGenerator does not reset the offset everytime on seeding and so I get different set of random values. If I create and destroy the generator in the loop, then I get same set of random values. – user1439690 Sep 05 '12 at 06:31
-1

The calling of threads is very random, so even the seed is set at the host, every run will generate different sequences.

rage
  • 11
  • 1