mfakto: an OpenCL program for Mersenne prefactoring

[KyleAskine]

I am going to start it tonight, and will let everyone know at around 5am tomorrow morning (EDT).

[KyleAskine]

Cayman (6950 with Bios changed to 6970):

Selftest statistics
number of tests 318416
successful tests 318416

selftest PASSED!

[Bdot]

Originally Posted by KyleAskine

Cayman (6950 with Bios changed to 6970):

Selftest statistics
number of tests 318416
successful tests 318416

selftest PASSED!

Very nice , thanks!

I will do some "paperwork" and packaging for releasing v0.13. Until then, 0.13pre5 should be stable enough for "productive work".

[KyleAskine]

Also, just so you know, it passed on my 6570 as well (but it took 15 hours(!!))!


Selftest statistics
number of tests 318416
successful tests 318416

selftest PASSED!

[kracker]

Passed on my 7770.

Selftest statistics
number of tests 287358
successful tests 287358

selftest PASSED!

EDIT: Strange, did I miss some tests?

[Bdot]

Originally Posted by kracker

Passed on my 7770.

Selftest statistics
number of tests 287358
successful tests 287358

selftest PASSED!

EDIT: Strange, did I miss some tests?

Thanks also for your results. No, you did not miss anything. You tested with GPU sieving, Kyle tested with CPU sieving. There are 3 less relevant kernels that I did not enable for GPU sieving. One is used for TF from 2^58 to 2^60, which is the only range where my mongomery implementation shines. The others are older kernels that are too slow to be selected in any real TF. I should probably remove them.

BTW, I successfully tested both GPU and CPU sieving for both HD5770 and HD7850.

On the 5770, I found GPUSievePrimes of 60k-70k to be the optimum. The 7850 seems to peak at ~110k. This also shows that the GPU sieve on VLIW5 is not very efficient yet (only 40% occupation on average). But this is something for the next mfakto version(s).

[kracker]

Originally Posted by Bdot

Thanks also for your results. No, you did not miss anything. You tested with GPU sieving, Kyle tested with CPU sieving. There are 3 less relevant kernels that I did not enable for GPU sieving. One is used for TF from 2^58 to 2^60, which is the only range where my mongomery implementation shines. The others are older kernels that are too slow to be selected in any real TF. I should probably remove them.

BTW, I successfully tested both GPU and CPU sieving for both HD5770 and HD7850.

On the 5770, I found GPUSievePrimes of 60k-70k to be the optimum. The 7850 seems to peak at ~110k. This also shows that the GPU sieve on VLIW5 is not very efficient yet (only 40% occupation on average). But this is something for the next mfakto version(s).

On my APU, ~60k seems to be best, I still yet have to tinker on m 7770... Are you planning more improvements or should we do benchmark time?

[Bdot]

Originally Posted by kracker

On my APU, ~60k seems to be best, I still yet have to tinker on m 7770... Are you planning more improvements or should we do benchmark time?

I found a cheap small thing for VLIW5 (e.g. your APU), but this is the last change I will do to the TF part. I'm now checking a few of the other features and already fixed the wait time display when CPU sieving.

You can start tweaking your 7770 and send the benchmark to James , the last build is just around the corner and will look the same for GCN.

[kracker]

Originally Posted by Bdot

I found a cheap small thing for VLIW5 (e.g. your APU), but this is the last change I will do to the TF part. I'm now checking a few of the other features and already fixed the wait time display when CPU sieving.

You can start tweaking your 7770 and send the benchmark to James , the last build is just around the corner and will look the same for GCN.

Looks like 10k is best for me for my 7770. Although, do I need to set it to 82485 for the benchmark? on James's benchmark form it says: "(set GPUSievePrimes=82485)"

[Bdot]

Originally Posted by kracker

Looks like 10k is best for me for my 7770. Although, do I need to set it to 82485 for the benchmark? on James's benchmark form it says: "(set GPUSievePrimes=82485)"

10k or 110k? If it's really 10k, then something is wrong ... which TF test are you doing, what are GPUSieveSize and GPUSieveProcessSize and what's the reported GHz? I think you should be getting around 140-145 GHzdays/day for M62... 2⁷⁰ to 2⁷³, or 125-130 for M62... 2⁷³ to 2⁸². Yes, GCN has a 10% performance step right there.

And for the reporting to James ... I don't see why you should not use the fastest possible value. Would be crippling devices that have their optimal point somewhere else ... Maybe James can comment on that.

[kracker]

Originally Posted by Bdot

10k or 110k? If it's really 10k, then something is wrong ... which TF test are you doing, what are GPUSieveSize and GPUSieveProcessSize and what's the reported GHz? I think you should be getting around 140-145 GHzdays/day for M62... 2⁷⁰ to 2⁷³, or 125-130 for M62... 2⁷³ to 2⁸². Yes, GCN has a 10% performance step right there.

100k. Sorry.

And yes, at ~M62, I'm getting 148 GHz days.

EDIT: When I go to the 332M range 73 to 74 bits, it drops to ~120 GHz...