Program to find prime numbers

People have mentioned a couple of the building blocks toward doing this efficiently, but nobody's really put the pieces together. The sieve of Eratosthenes is a good start, but with it you'll run out of memory long before you reach the limit you've set. That doesn't mean it's useless though -- when you're doing your loop, what you really care about are prime divisors. As such, you can start by using the sieve to create a base of prime divisors, then use those in the loop to test numbers for primacy.

When you write the loop, however, you really do NOT want to us sqrt(i) in the loop condition as a couple of answers have suggested. You and I know that the sqrt is a "pure" function that always gives the same answer if given the same input parameter. Unfortunately, the compiler does NOT know that, so if use something like '<=Math.sqrt(x)' in the loop condition, it'll re-compute the sqrt of the number every iteration of the loop.

You can avoid that a couple of different ways. You can either pre-compute the sqrt before the loop, and use the pre-computed value in the loop condition, or you can work in the other direction, and change i<Math.sqrt(x) to i*i<x. Personally, I'd pre-compute the square root though -- I think it's clearer and probably a bit faster--but that depends on the number of iterations of the loop (the i*i means it's still doing a multiplication in the loop). With only a few iterations, i*i will typically be faster. With enough iterations, the loss from i*i every iteration outweighs the time for executing sqrt once outside the loop.

That's probably adequate for the size of numbers you're dealing with -- a 15 digit limit means the square root is 7 or 8 digits, which fits in a pretty reasonable amount of memory. On the other hand, if you want to deal with numbers in this range a lot, you might want to look at some of the more sophisticated prime-checking algorithms, such as Pollard's or Brent's algorithms. These are more complex (to put it mildly) but a lot faster for large numbers.

There are other algorithms for even bigger numbers (quadratic sieve, general number field sieve) but we won't get into them for the moment -- they're a lot more complex, and really only useful for dealing with really big numbers (the GNFS starts to be useful in the 100+ digit range).

Try this:

void prime_num(long num)
{

    // bool isPrime = true;
    for (long i = 0; i <= num; i++)
    {
        bool isPrime = true; // Move initialization to here
        for (long j = 2; j < i; j++) // you actually only need to check up to sqrt(i)
        {
            if (i % j == 0) // you don't need the first condition
            {
                isPrime = false;
                break;
            }
        }
        if (isPrime)
        {
            Console.WriteLine ( "Prime:" + i );
        }
        // isPrime = true;
    }
}

You only need to check odd divisors up to the square root of the number. In other words your inner loop needs to start:

for (int j = 3; j <= Math.Sqrt(i); j+=2) { ... }

You can also break out of the function as soon as you find the number is not prime, you don't need to check any more divisors (I see you're already doing that!).

This will only work if num is bigger than two.

No Sqrt

You can avoid the Sqrt altogether by keeping a running sum. For example:

int square_sum=1;
for (int j=3; square_sum<i; square_sum+=4*(j++-1)) {...}

This is because the sum of numbers 1+(3+5)+(7+9) will give you a sequence of odd squares (1,9,25 etc). And hence j represents the square root of square_sum. As long as square_sum is less than i then j is less than the square root.

You can do this faster using a nearly optimal trial division sieve in one (long) line like this:

Enumerable.Range(0, Math.Floor(2.52*Math.Sqrt(num)/Math.Log(num))).Aggregate(
    Enumerable.Range(2, num-1).ToList(), 
    (result, index) => { 
        var bp = result[index]; var sqr = bp * bp;
        result.RemoveAll(i => i >= sqr && i % bp == 0); 
        return result; 
    }
);

The approximation formula for number of primes used here is π(x) < 1.26 x / ln(x). We only need to test by primes not greater than x = sqrt(num).

Note that the sieve of Eratosthenes has much better run time complexity than trial division (should run much faster for bigger num values, when properly implemented).