Seed curandGenerateLongLong

I can’t use curandSetPseudoRandomGeneratorSeed with curandGenerateLongLong:

int n = 10;
unsigned long long *randomKeys;
curandGenerator_t gen;
cudaMalloc(&randomKeys, n*sizeof(unsigned long long));
curandCreateGenerator(&gen, curandRngType_t::CURAND_RNG_QUASI_SCRAMBLED_SOBOL64);
curandGenerateLongLong(gen, randomKeys, n);

How can I seed the curandGenerateLongLong, because it gives me the same numbers every time I run it.

“it gives me the same numbers”

i almost want to say this is ‘normal’, within context; if i am not mistaken, it is a sequence generator

“every time”

how many times are we talking about? in essence, are you iterating, and how many times do you wish to iterate?

you could perhaps store a count as pointer, and only accept generated output as input, after said count

If I run the app for the 1st time (with the mentioned codes/setup), it prints for instance 1.
For the 2nd time, it prints 1 too. so it’s completely predictable and needs a mechanism to seed it.

hence, if you start ‘recording’ from ‘result’ 0 the 1st time, and from ‘result’ n the nth time,
result[0] != result[n]

i think the case is common for all random generators - host or device
i doubt whether random generators make provision for iteration; i would consider your case as informally iterating
a simple solution is to offset the ‘start’
curand has a record that is passed to it; can you not offset the start or count there?
if not, manually offset the count/ start

in essence, let the random generator ‘idle’/ run for a while before recording and using the output

as illustration, on the host, not to have the same output each time, have the start as a random number itself

int lint1, lint2;

lint1 = rand();

lint2 = 0;

// simply dump the output

while (lint2 < lint1)
{
lint2++;
lint3 = rand();
}

// start using the output

lint3 = rand();

It seems a possible answer was given in your cross-posting:

http://stackoverflow.com/questions/31488100/seeding-quasi-scrambled-sobol64-in-curand

Since the longlong generator function requires a 64-bit generator, and only the SOBOL sequences are available in 64-bit generator types, and the SOBOL sequences are only quasirandom, it seems that you will have to use such a contrived seeding method.