mfakto: an OpenCL program for Mersenne prefactoring

[LaurV]

Test if I can attach it. (I did not expect that you generate gigabytes of those test results... )

[Rodrigo]

Originally Posted by Bdot

this made me think if non-power-of-two values for GPUSieveSize are allowed at all. Yes they are. But thinking a bit more about the dependencies, I think I found a bug in mfakto and mfaktc:

When using GPUSieveProcessSize=24 and GPUSieveSize that is not divisible by 3, then some FCs may go untested.

(GPUSieveSize * 1024) must be divisible by GPUSieveProcessSize.

Worst case would be GPUSieveProcessSize=24 and GPUSieveSize=4, in which case about 1 in 256 FCs would go untested.

Typical settings of GPUSieveProcessSize=24 and GPUSieveSize=64 leaves 1 in 4096 FCs untested, GPUSieveProcessSize=24 and GPUSieveSize=128 about 1 in 16384.

Unfortunately this is something that the selftest cannot cover without increasing the selftest runtime by a factor of 10 at least.

How many of you use GPU sieving with GPUSieveProcessSize=24 ? What do we want to do with those tests?

I'm using GPUSieveProcessSize=16, as suggested by kracker here.

Going along with his other idea there, GPUSievePrimes is at 92486.

GPUSieveSize has been back at 64 ever since I reported the results with that value at 48.

With these settings, TF of 72xxxxxx exponents has been hovering around 144-146 GHz-d/d.

All of this work done in mfakto x32.

Rodrigo

[kracker]

Originally Posted by Rodrigo

I'm using GPUSieveProcessSize=16, as suggested by kracker here.

Actually, I had recommended 24, "GPUSieveProcessSize to 24 from 16" which 16 is default(I think?)

Anyways it might be safer to stay on 16 (for now)

[Bdot]

Originally Posted by kracker

Actually, I had recommended 24, "GPUSieveProcessSize to 24 from 16" which 16 is default(I think?)

Anyways it might be safer to stay on 16 (for now)

GPUSieveProcessSize=24 is safe as long as GPUSieveSize is a multiple of 3. That is also, what the fix will do - the fix is just in the parameter validation part of reading mfakto.ini.

What is the general opinion if any of the tests need to be repeated? I tried to come up with an estimate how "bad" the bug is ...

My opinion is, that if anyone used GPUSieveProcessSize=24 along with really low GPUSieveSize(say, < 20), then chances of missing a factor (like one in 256 to one in 1024) are too high and the tests should be repeated. For GPUSieveSize=32 and above, missing between one in 4096 and one in 16384 factors is probably not worth the effort.

But I'd like to hear other opinions. If you can provide a guideline for an "acceptable miss probability", then I can compile a list of GPUSieveSize settings that would require a re-test.

[Bdot]

Originally Posted by LaurV

Test if I can attach it. (I did not expect that you generate gigabytes of those test results... )

Thanks a lot for the test results. They show a pretty accurate proportional performance increase from my 7850. No sign for a different calculation base for the high-end GCNs.

However, I do not yet have a good performance measurement of the GPU sieving itself - something to add soon.

[blahpy]

GPUSieveSize=64
GPUSieveProcessSize=16

What should I do? I don't think it's possible to "redo" or "double check" TF is it?

Edit: Or are these settings safe? Upon reading it again it seems more like GPUSieveSize=64 only needs to be divisible by 3 for GPUSieveProcessSize=24?

Edit2: Okay, I read more carefully, ignore this post (I would delete it but can't)

[Bdot]

Originally Posted by blahpy

GPUSieveSize=64
GPUSieveProcessSize=16

What should I do? I don't think it's possible to "redo" or "double check" TF is it?

Edit: Or are these settings safe? Upon reading it again it seems more like GPUSieveSize=64 only needs to be divisible by for GPUSieveProcessSize=24?

Yes, these (default) settings are safe - no need to rerun anything.

[Rodrigo]

Originally Posted by kracker

Actually, I had recommended 24, "GPUSieveProcessSize to 24 from 16" which 16 is default(I think?)

Anyways it might be safer to stay on 16 (for now)

Huh, you're right. I guess I confused it because (according to the info in the mfakto.ini file) the default value is 24:

 
# GPUSieveProcessSize defines how far many bits of the sieve each TF block
# processes (in K bits). Larger values may lead to less wasted cycles by
# reducing the number of times all threads in a warp are not TFing a
# candidate.  However, more shared memory is used which may reduce occupancy.
# Smaller values should lead to a more responsive system (each kernel takes
# less time to execute). GPUSieveProcessSize must be a multiple of 8.
# 
#
# Minimum: GPUSieveProcessSize=8
# Maximum: GPUSieveProcessSize=32  (requires GPUSievePrimes > 310)
#
# Default: GPUSieveProcessSize=24
GPUSieveProcessSize=16

...and so I switched it TO 16, figuring that's what you meant to say.

But, whatever the reason, I seem to have stumbled onto the safest values for the time being.

Rodrigo

[Bdot]

Hmm, the one that you can download with the win package reads:

# GPUSieveProcessSize defines how far many bits of the sieve each TF block
# processes (in K bits). Larger values may lead to less wasted cycles by
# reducing the number of times all threads in a warp are not TFing a
# candidate.  However, more shared memory is used which may reduce occupancy.
# Smaller values should lead to a more responsive system (each kernel takes
# less time to execute). GPUSieveProcessSize must be a multiple of 8.
# 
#
# Minimum: GPUSieveProcessSize=8
# Maximum: GPUSieveProcessSize=32  (requires GPUSievePrimes > 310)
#
# Default: GPUSieveProcessSize=16

GPUSieveProcessSize=16

[kladner]

How many of you use GPU sieving with GPUSieveProcessSize=24 ? What do we want to do with those tests?

I'm using
GPUSieveSize=128
GPUSieveProcessSize=8

I don't think I've ever used 24.

[Rodrigo]

Originally Posted by Bdot

Hmm, the one that you can download with the win package reads:

# GPUSieveProcessSize defines how far many bits of the sieve each TF block
# processes (in K bits). Larger values may lead to less wasted cycles by
# reducing the number of times all threads in a warp are not TFing a
# candidate.  However, more shared memory is used which may reduce occupancy.
# Smaller values should lead to a more responsive system (each kernel takes
# less time to execute). GPUSieveProcessSize must be a multiple of 8.
# 
#
# Minimum: GPUSieveProcessSize=8
# Maximum: GPUSieveProcessSize=32  (requires GPUSievePrimes > 310)
#
# Default: GPUSieveProcessSize=16
 
GPUSieveProcessSize=16

How weird. My backup of the original mfakto.ini shows exactly what you give there.

The only way that I can think of that could lead to the value changing in the comments line, would be if I had made the change manually to the comments line and never actually adjusted the real setting to the 24 that @kracker had recommended. (In which case I never changed it to OR from 16.)

Although I have no memory of it, this is entirely possible: my wife and I spent several days doing some intense (physical and Web) car shopping. I can picture hurriedly making this change on the way out to yet another auto dealership. This must be what happened, there's no other sensible explanation. I plead temporary insanity.

Rodrigo