A multiple k/c sieve for Sierpinski/Riesel problems

[Xentar]

Originally Posted by geoff

Sorry, version 1.4.12 had a bug which could have caused 1 factor in 240 to be missed. If anyone is using it then please upgrade to version 1.4.13.

I used 1.4.12 for about 6 hours - do you think, it is necessary to redo this range?

[jasong]

Originally Posted by geoff

Sorry, version 1.4.12 had a bug which could have caused 1 factor in 240 to be missed. If anyone is using it then please upgrade to version 1.4.13.

111,100 k/sec. Awesome.

[geoff]

Originally Posted by Xentar

I used 1.4.12 for about 6 hours - do you think, it is necessary to redo this range?

No, but if you let me know the range concerned I can find any missed factors faster than fully resieving it -- only the factors p where p=1 (mod 120) have to be checked.

Version 1.4.14 fixes another bug that would have slowed the PPC build down a bit, but probably didn't affect the i386 builds.

[Xentar]

Originally Posted by geoff

No, but if you let me know the range concerned I can find any missed factors faster than fully resieving it -- only the factors p where p=1 (mod 120) have to be checked.

The ranges are 4200-4300 and 4300-4400. I think, I tried version 1.4.12 somewhere between 75 and 85%.. sorry, dont know exactly :(

Btw, could you please try to explain the math to me? I know, that for example 5 mod 3 = 2, but p = 1 mod 120 doesnt make any sense? Or I just dont understand it..

[geoff]

Originally Posted by Xentar

The ranges are 4200-4300 and 4300-4400. I think, I tried version 1.4.12 somewhere between 75 and 85%.. sorry, dont know exactly :(

I'll test those 10% ranges tonight. It takes only 1/32 of the work of the full sieve to check, and I am interested to see how many if any were missed.

Btw, could you please try to explain the math to me? I know, that for example 5 mod 3 = 2, but p = 1 mod 120 doesnt make any sense? Or I just dont understand it..

5 = 2 (mod 3) [or properly, means the same as 5 mod 3 = 2 in this case. The factors p of interest are all of the form 120x+1 for some x.

[Xentar]

Originally Posted by geoff

I'll test those 10% ranges tonight. It takes only 1/32 of the work of the full sieve to check, and I am interested to see how many if any were missed.

OK, could you then please send them as PM to me, so I can check, they are already in the factors-list?

5 = 2 (mod 3) [or properly, means the same as 5 mod 3 = 2 in this case. The factors p of interest are all of the form 120x+1 for some x.

Ok, I understand. So I was just confused by the kind of writing p = 1 (mod 120) :D
So, for any reason, the version 1.4.12 didnt check p = 120x+1 and if this p is a factor, it was missed, right?
Ok, thank you :)

[geoff]

It appears that the 1.4.12 bug is in code that is never used, so there are no missed factors. Once I work out exactly why that piece of code is not needed then I can remove it, and probably some other pieces too, and make the program faster :-)

Originally Posted by Xentar

OK, could you then please send them as PM to me, so I can check, they are already in the factors-list?

Done, just in case.

[Xentar]

Originally Posted by geoff

It appears that the 1.4.12 bug is in code that is never used, so there are no missed factors. Once I work out exactly why that piece of code is not needed then I can remove it, and probably some other pieces too, and make the program faster :-)

Just checked the factors, you send to me. All of them are already in my factors file, so no one was missed.

[geoff]

The main change is the addition of a SSE2 mulmod function for 32-bit machines. The FPU is still used for the division (so it still works mod primes up to 2^62), but uses SSE2 to perform the integer multiplications. It runs about 9% faster on my P4/Celeron. (That was the machine I tuned the code for, it may not do as well on AMD/Core2, all benchmarks welcome.)

[tnerual]

Originally Posted by geoff

The main change is the addition of a SSE2 mulmod function for 32-bit machines. The FPU is still used for the division (so it still works mod primes up to 2^62), but uses SSE2 to perform the integer multiplications. It runs about 9% faster on my P4/Celeron. (That was the machine I tuned the code for, it may not do as well on AMD/Core2, all benchmarks welcome.)

i don't know if it's normal but i have a good speed improve on my amd x2 4200

i use one core, sieving for one k (285728), speed was 2600000 and now is 4650000 it's a 78% improve ... and factor density is the same after 1 day running so it may be true.

if geoff can confirm it's not missing factors, it can give a good boost ...

[Xentar]

Originally Posted by geoff

The main change is the addition of a SSE2 mulmod function for 32-bit machines. The FPU is still used for the division (so it still works mod primes up to 2^62), but uses SSE2 to perform the integer multiplications. It runs about 9% faster on my P4/Celeron. (That was the machine I tuned the code for, it may not do as well on AMD/Core2, all benchmarks welcome.)

about 5 % improve on Core2. (just tested for a few minutes)