CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW) - Page 6

frmky · 2009-11-10, 21:28

Here's the profiler output for k in case it's useful. I can send the CSV is you wish.

It appears that in version u, transpose speeds up normalize_kernel but by a bit less than the transpose takes so it's a small net loss. I presume on the GTX260, normalize_kernel was significantly slower before so the transpose saves much more time than it takes.

The C1060s are only running at a clock rate of 1.3GHz, so if your GTX260 is running at 1.44GHz (deviceQuery will tell you) then the difference between our measured times for version u is simply the clock rate of the cards.

msft · 2009-11-11, 07:26

Quote:

Originally Posted by frmky

The C1060s are only running at a clock rate of 1.3GHz, so if your GTX260 is running at 1.44GHz (deviceQuery will tell you)
then the difference between our measured times for version u is simply the clock rate of the cards.

Exactry, I add compile option -DTESRA(Version "k" like code).

Add -DTESRA at .0157 sec/iter for the 2048K FFT and .0415 sec/iter for the 4096K FFT on GTX260.
No -DTESRA at .0131 sec/iter for the 2048K FFT and .0247 sec/iter for the 4096K FFT on GTX260.

Thank you,

frmky · 2009-11-11, 08:57

Excellent! On the Tesla C1060, version v 4096K FFTs times are

without TESRA: 0.0275 sec/iter
with TESRA: 0.0264 sec/iter

so it does match the speed of version k. I've also confirmed that restart works correctly. The 10000 iteration residue matches the mprime "interim Wd8 residue at iteration 10002." Is this just a difference in how iterations are counted in the two programs?

msft · 2009-11-11, 09:35

Quote:

Originally Posted by frmky

I've also confirmed that restart works correctly. The 10000 iteration residue matches the mprime "interim Wd8 residue at iteration 10002." Is this just a difference in how iterations are counted in the two programs?

I'm not touch this part.

frmky · 2009-11-11, 09:45

OK, it's not really an issue anyway. Just have to remember that when comparing interim residues with Prime95. Actually, I should check that the final residues for composites match those of Prime95...

msft · 2009-11-11, 10:20

Quote:

Originally Posted by frmky

OK, it's not really an issue anyway. Just have to remember that when comparing interim residues with Prime95. Actually, I should check that the final residues for composites match those of Prime95...

I check it. Tell me script.(how command ?)

frmky · 2009-11-11, 11:30

I checked it. They do agree with the Res64 of Prime95. I've found a problem, though...

[childers test]$ ./MacLucas_v_T_0 23102129
cutilCheckMsg() CUTIL CUDA error: CUDA Kernel execution failed in file <MacLucasFFTW.cu>, line 1218 : invalid configuration argument.

msft · 2009-11-11, 12:18

Quote:

Originally Posted by frmky

I checked it. They do agree with the Res64 of Prime95. I've found a problem, though...

[childers test]$ ./MacLucas_v_T_0 23102129
cutilCheckMsg() CUTIL CUDA error: CUDA Kernel execution failed in file <MacLucasFFTW.cu>, line 1218 : invalid configuration argument.

I can not reproduce.
100% Repeatability ?
When after machine reboot...

henryzz · 2009-11-11, 17:37

Quote:

Originally Posted by msft

I'm not touch this part.

its a feature of mlucas(actually i suspect its possibly a feature of prime95 as i think glucas might have had the same problem

)
when doing doublechecks of mersenne primes they had to fiddle this

frmky · 2009-11-11, 20:25

Quote:

Originally Posted by msft

I can not reproduce.
100% Repeatability ?
When after machine reboot...

The non-TESRA version works fine, but the TESRA version seems numerically unstable. One compile would work on some GPUs but not others. I'd simply recompile and that would shuffle the GPUs it would work on. When it failed, it did so because the err returned was larger than kErrLimit, so it would keep increasing n until the grid size (n/512) of the rftfsub_kernel call was too large.

I modified the source so that when TESRA is defined, it uses the normalize_kernel from version k rather than breaking it into the three sequential calls. This works correctly on all of the GPUs in the S1070. I'm not sure why breaking it into three kernel calls was causing a problem. Attached is the source as modified.

msft · 2009-11-12, 01:22

Hi, henryzz

Quote:

Originally Posted by henryzz

its a feature of mlucas(actually i suspect its possibly a feature of prime95 as i think glucas might have had the same problem

)
when doing doublechecks of mersenne primes they had to fiddle this

I understand.

Thank you,

2009-11-10, 21:28	#56
frmky Jul 2003 So Cal 101001100111₂ Posts	Here's the profiler output for k in case it's useful. I can send the CSV is you wish. It appears that in version u, transpose speeds up normalize_kernel but by a bit less than the transpose takes so it's a small net loss. I presume on the GTX260, normalize_kernel was significantly slower before so the transpose saves much more time than it takes. The C1060s are only running at a clock rate of 1.3GHz, so if your GTX260 is running at 1.44GHz (deviceQuery will tell you) then the difference between our measured times for version u is simply the clock rate of the cards. Attached Thumbnails Last fiddled with by frmky on 2009-11-10 at 21:43

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Don't DC/LL them with CudaLucas	LaurV	Data	131	2017-05-02 18:41
CUDALucas / cuFFT Performance on CUDA 7 / 7.5 / 8	Brain	GPU Computing	13	2016-02-19 15:53
CUDALucas: which binary to use?	Karl M Johnson	GPU Computing	15	2015-10-13 04:44
settings for cudaLucas	fairsky	GPU Computing	11	2013-11-03 02:08
Trying to run CUDALucas on Windows 8 CP	Rodrigo	GPU Computing	12	2012-03-07 23:20

2009-11-11, 08:57	#58
frmky Jul 2003 So Cal 2,663 Posts	Excellent! On the Tesla C1060, version v 4096K FFTs times are without TESRA: 0.0275 sec/iter with TESRA: 0.0264 sec/iter so it does match the speed of version k. I've also confirmed that restart works correctly. The 10000 iteration residue matches the mprime "interim Wd8 residue at iteration 10002." Is this just a difference in how iterations are counted in the two programs?

2009-11-11, 09:45	#60
frmky Jul 2003 So Cal 2,663 Posts	OK, it's not really an issue anyway. Just have to remember that when comparing interim residues with Prime95. Actually, I should check that the final residues for composites match those of Prime95...

2009-11-11, 11:30	#62
frmky Jul 2003 So Cal 2,663 Posts	I checked it. They do agree with the Res64 of Prime95. I've found a problem, though... [childers test]$ ./MacLucas_v_T_0 23102129 cutilCheckMsg() CUTIL CUDA error: CUDA Kernel execution failed in file <MacLucasFFTW.cu>, line 1218 : invalid configuration argument.