LL with OpenCL

[bverka86]

Originally Posted by bverka86

So for example, if 1536K wasn't good enough for M35861153, 2048K performs at least 3x better than 1728K. I would not mind digging into this algorithm, but I've been sick for quite some time and haven't dug in too deeply. Is the entire FFT length selection done in clLucas.cpp?

*****
For a precise example, using M38542223:
Auto-Select:
Iteration 300000 M( 38542223 )C, 0xfd1aedd779d62f55, n = 2240K, clLucas v1.04 err = 0.0308 (33:27 real, 20.0766 ms/iter, ETA 213:02:07)
Manual Select:
Iteration 300000 M( 38542223 )C, 0xfd1aedd779d62f55, n = 2560K, clLucas v1.04 err = 0.0019 (9:39 real, 5.7906 ms/iter, ETA 61:26:39)

[axn]

Originally Posted by bverka86

*****
For a precise example, using M38542223:
Auto-Select:
Iteration 300000 M( 38542223 )C, 0xfd1aedd779d62f55, n = 2240K, clLucas v1.04 err = 0.0308 (33:27 real, 20.0766 ms/iter, ETA 213:02:07)
Manual Select:
Iteration 300000 M( 38542223 )C, 0xfd1aedd779d62f55, n = 2560K, clLucas v1.04 err = 0.0019 (9:39 real, 5.7906 ms/iter, ETA 61:26:39)

Could you also try the following FFTs and see if any of them are faster than 2560?
2187
2304
2401
2500
2592

[bverka86]

Originally Posted by axn

Could you also try the following FFTs and see if any of them are faster than 2560?
2187
2304
2401
2500
2592

I've got to reboot the box running that number anyway (SuperMicro's built in PWM is terribly annoying, switching back to my home-grown pwm program). I'll backup my current checkpoints and run those to 30K (assuming each FFT listed works under 1.04)

**Update - Results of requested FFTs**
2187K:
Starting M38542223 fft length = 2187K
FFT length error.

2304K (thanks for the tip!):
Iteration 30000 M( 38542223 )C, 0xdbd7530295df2924, n = 2304K, clLucas v1.04 err = 0.0175 (0:52 real, 5.1583 ms/iter, ETA 55:10:47)

2401K:
Starting M38542223 fft length = 2401K
FFT length error.

2500K (still faster than 2560K, but not as good as 2304):
Iteration 30000 M( 38542223 )C, 0xdbd7530295df2924, n = 2500K, clLucas v1.04 err = 0.0028 (0:57 real, 5.6499 ms/iter, ETA 60:26:18)

2592K (faster than 2560K, but only slightly):
Iteration 30000 M( 38542223 )C, 0xdbd7530295df2924, n = 2592K, clLucas v1.04 err = 0.0016 (0:57 real, 5.7399 ms/iter, ETA 61:24:04)

clFFT-2.12.0 (if it matters, I have the sources to build 2.10.1 & 2.10.2 if you think those would show further improvement)

Also, given 2187K & 2401K weren't even in terms of K, I tried 2176K & 2402K & 2410K (2048+128 & 2304+92 & 2304+96, could see any as a reasonable typo), but that returned an error of a different nature:
Starting M38542223 fft length = 2176K
OPENCL_V_THROWERROR< CLFFT_NOTIMPLEMENTED > (1753): Failed to clfftCreateDefaultPlan.
terminate called after throwing an instance of 'std::runtime_error'
what(): OPENCL_V_THROWERROR< CLFFT_NOTIMPLEMENTED > (1753): Failed to clfftCreateDefaultPlan.
Aborted

Starting M38542223 fft length = 2402K
OPENCL_V_THROWERROR< CLFFT_NOTIMPLEMENTED > (1753): Failed to clfftCreateDefaultPlan.
terminate called after throwing an instance of 'std::runtime_error'
what(): OPENCL_V_THROWERROR< CLFFT_NOTIMPLEMENTED > (1753): Failed to clfftCreateDefaultPlan.
Aborted

Starting M38542223 fft length = 2410K
OPENCL_V_THROWERROR< CLFFT_NOTIMPLEMENTED > (1753): Failed to clfftCreateDefaultPlan.
terminate called after throwing an instance of 'std::runtime_error'
what(): OPENCL_V_THROWERROR< CLFFT_NOTIMPLEMENTED > (1753): Failed to clfftCreateDefaultPlan.
Aborted

[axn]

Originally Posted by bverka86

Also, given 2187K & 2401K weren't even in terms of K, I tried 2176K & 2402K (2048+128 & 2304+92, could see it as a reasonable typo, couldn't figure an alternate for 2401K), but that returned an error of a different nature:

They weren't typos, but I wasn't sure they'd work in clLucas. Basically, I was looking for all FFTs between 2048 and 4096 which was of the form 2^x*p^y (x, y >=0, p in {3,5,7}). These tend to have efficient implementation in clFFT. 2187 = 3^7, 2401 = 7^4.

FWIW, 2304 = 2^8*3^2, 2500 = 2^2*5^4, 2592 = 2^5*3^4

Here it the full list of even FFTs that follow same pattern (between 2048 & 4096K)
2048=2^11
2304=2^8*3^2
2500=2^2*5^4
2560=2^9*5^1
2592=2^5*3^4
2744=2^3*7^3
2916=2^2*3^6
3072=2^10*3^1
3136=2^6*7^2
3200=2^7*5^2
3456=2^7*3^3
3584=2^9*7^1
3888=2^4*3^5
4000=2^5*5^3
4096=2^12

[bverka86]

I'll keep that list of FFT's in mind.

The breakdown into each FFT's "smoothness" is helpful, so thank you!

Sometimes I question using these R9 270X's for LL tests, but then I consider things like M72895313...

R9 270X:
Iteration 18000000 M( 72895313 )C, 0xebe0881ebbc59556, n = 4096K, clLucas v1.04 err = 0.1094 (13:33 real, 8.1293 ms/iter, ETA 123:44:44)

Dual Xeon E5450 (6/8 cores):
[Work thread May 3 10:47] Iteration: 3400000 / 72895313 [4.66%], ms/iter: 21.244, ETA: 17d 02:05

[hurrican50]

On my 290, why does the ETA go from about 150Hours to 180Hours for cllucas when I increase the fanspeed of the GPU? The ETA is correctly adjusted based on the ms/iter, but why does it take a little longer?
All other programs I have seen are either quicker or not effected when I increase the fan speed.

Note: Increasing the fanspeed does not effect the speed of the memory, but it does allow the GPU to run faster, longer.

[hurrican50]

When I increase the Fanspeed on my 290, why does the ETA increase from about 150 hours to 180 hours on cllucas for an LL test? On all other programs this either has no effect, or it makes the program run faster because it allows the GPU to run closer to full speed.

[LaurV]

Originally Posted by hurrican50

When I increase the Fanspeed on my 290, why does the ETA increase from about 150 hours to 180 hours on cllucas for an LL test? On all other programs this either has no effect, or it makes the program run faster because it allows the GPU to run closer to full speed.

Power limit. The fan takes more energy, therefore less is available for computing. The card is enough clever to adjust its ticks to stay in the power requirements. For my Titans, I used to tune the fan curve, to get like 10% more computation. [edit: this was long ago when I was running on air, for water cooling we don't have this problem anymore, and the fans stays at minimum, because there is no fan connected].

[airsquirrels]

I did some testing with clFFT 2.12.1 and the latest Crimson drivers.

Nothing of significant note to report, although as of 2.10 Fury X cards can from > 16384K FFTs...

Iteration 10000 M( 332213573 )C, 0x216c5a1819bd595d, n = 19200K, clLucas v1.04 err = 0.1367 (4:42 real, 28.1697 ms/iter, ETA 2599:26:26)
Iteration 20000 M( 332213573 )C, 0xebc18094b00fad87, n = 19200K, clLucas v1.04 err = 0.1406 (4:42 real, 28.1144 ms/iter, ETA 2594:15:35)

108 days for a 100M test on a Fury X. ($629, 200/220W during test)

[axn]

Originally Posted by airsquirrels

108 days for a 100M test on a Fury X. ($629, 200/220W during test)

clLucas doesn't necessarily choose the best FFT sizes. There might be larger FFTs that can give much higher performance. I would, at minimum, try these alternate FFTs: 18432k, 19208k, 19600k, 20000k, 20480k, 20736k to see if there is a potential for higher performance.

[LaurV]

yes, correct! we would also suggest a try with 32768k, which is the next power of two (however a bit to the limit here) to see how the iteration times behave.