CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW)

[ET_]

Originally Posted by flashjh

Hello all!

I compiled a new windows CUDALucas 2.06 for all CUDA 4.0 - 10.1. Support for win32 ended at CUDA 6.5.

No big changes, removed nmvl requirement and made small updates for new CUDA versions.

2.06 is here

Lib files are here, if you need them

I was able to do basic testing, but please test thoroughly before using for production.

Everyone, let me know what you find that needs to be fixed and what you would like changed.

~Cheers

We would love to see a Linux version, either executable or at least updated source code as well

[kriesel]

Originally Posted by ATH

It is an old Titan Black and it mostly happened when I used CUDA 5.0 version of CUDALucas, so I stopped using that version. But I'm about to scrap this card now, it keeps giving bad results randomly even though when I test it with GPUmemtest and CUDALucas extended self tests and mfaktc extended self tests there are no errors.

Which versions/build dates of CUDALucas do you see this on?
Windows or linux, version?

[ATH]

Windows version of CUDALucas 2.06 but it happened for both CUDA 5.5, 8.0, 10.1 and both for a version I compiled myself and the "official" compiled version from the site.

It must be some very small error on the card even though it passes all benchmarks and RAM tests. I stopped trying to run CUDALucas on it, doing a bit of TF with mfaktc, and it does find factors and find all the factors in the extended mfaktc benchmark.

The only real mem test for GPU I could find was GpuMemTest, but it only tests 4GB out of the 6GB on the card:
http://www.programming4beginners.com/gpumemtest

Others I tried were not real mem tests, but only stress tests of the GPU. I wish there was a gpu mem test you could run for 24+ hours and test all the RAM like the RAM tests for normal RAM. This one only takes ~1 min and have to be restarted again.

[kriesel]

Originally Posted by ATH

The only real mem test for GPU I could find was GpuMemTest, but it only tests 4GB out of the 6GB on the card:
http://www.programming4beginners.com/gpumemtest

Others I tried were not real mem tests, but only stress tests of the GPU. I wish there was a gpu mem test you could run for 24+ hours and test all the RAM like the RAM tests for normal RAM. This one only takes ~1 min and have to be restarted again.

I use the CUDALucas -memtest option. Go for as many 25MB blocks as it will run on the gpu. It needs a bit of space for the gpu program itself. Running that and redirecting the output to disk is a standard part of my startup of any newly acquired gpu. Run it briefly to get a sense of speed, then decide how many iterations to specify based on a total run time computation.
From the 2.05.1 readme:

-memtest s i           s = # of chunks of memory
                       i = number of iterations

                       tests s 25MB chunks of memory doing i repetitions of
                       a 100,000 iteration loop on each of 5 different LL
                       test related sets of data. Each iteration consists
                       of copying a 25MB chunk of data, then re-reading
                       and comparing that copy to the original.

Memory space coverage as complete as possible is necessary. I found on a card I had previously tested only the first 250MB of, that P-1 and CUDALucas were unreliable. Max-practical-address-space testing showed lots of memory errors in the middle third of the gpu ram address space. Locations (blocks with errors, and so conversely blocks without errors) were very repeatable run to run, near 100%.
More details on that gpu at the GPU RIP thread https://www.mersenneforum.org/showth...561#post490561

[chalsall]

Originally Posted by kriesel

I found on a card I had previously tested only the first 250MB of, that P-1 and CUDALucas were unreliable. Max-practical-address-space testing showed lots of memory errors in the middle third of the gpu ram address space. Locations (blocks with errors, and so conversely blocks without errors) were very repeatable run to run, near 100%.

GIMPS. We don't just find prime numbers, we also ensure sanity of kit (for those who care)!

[ATH]

Originally Posted by kriesel

I use the CUDALucas -memtest option. Go for as many 25MB blocks as it will run on the gpu. It needs a bit of space for the gpu program itself. Running that and redirecting the output to disk is a standard part of my startup of any newly acquired gpu. Run it briefly to get a sense of speed, then decide how many iterations to specify based on a total run time computation.

Thanks, I never really noticed that option. It survived one pass with no errors, but I will let it run over night or longer.

CUDA reports 5117M of 6144M GPU memory free.

Initializing memory test using 5075MB of memory on device 0...
Beginning test.

Anyone know why there is only 5117MB / 6144MB free when I'm just on the Windows Desktop not running anything. I can only test 5075 MB of the 5117 MB, so I assume the program uses the difference.

[kriesel]

Originally Posted by ATH

Thanks, I never really noticed that option. It survived one pass with no errors, but I will let it run over night or longer.

Anyone know why there is only 5117MB / 6144MB free when I'm just on the Windows Desktop not running anything. I can only test 5075 MB of the 5117 MB, so I assume the program uses the difference.

My impression is CUDALucas in -memtest mode only needs 50-100MB for itself, and can test whatever of the rest is available.
Three other things that come to mind using up memory: ECC, any other gpu app running, and Windows display. What does NVIDIA-SMI.exe say about it? Sample:

"C:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe"
Mon May 20 19:43:17 2019
+------------------------------------------------------+
| NVIDIA-SMI 353.30     Driver Version: 353.30         |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla C2075         TCC  | 0000:06:00.0     Off |                    0 |
| 60%   87C    P0     0W / 225W |    175MiB /  5375MiB |     99%      Default |
+-------------------------------+----------------------+----------------------+
|   1  Quadro K4000       WDDM  | 0000:42:00.0     Off |                  N/A |
| 50%   85C    P0    60W /  87W |    350MiB /  3072MiB |     99%      Default |
+-------------------------------+----------------------+----------------------+

That Tesla is a 6GB gpu, with ECC enabled, costing 3/4GB. In its initial memory test run of CUDALucas, it said this:

CUDA reports 5316M of 5375M GPU memory free.
Reducing size to 212

Initializing memory test using 5300MB of memory on device 0...

You may want to close (end process) command prompt boxes that previously had gpu apps running in them, in case some memory allocation didn't get cleaned up from the perspective of the OS. I'm curious what's happening with that other GB.


On another gpu (no monitor attached to it),

CUDA reports 10988M of 11264M GPU memory free.

Initializing memory test using 10750MB of memory on device 0...

[ATH]

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 430.39       Driver Version: 430.39       CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX TIT... WDDM  | 00000000:02:00.0  On |                  N/A |
| 59%   55C    P0    97W / 250W |    185MiB /  6144MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

It is only using 185MiB / 6144 MiB, but still only 5117MiB free.

Which is close to 5 GiB = 5*1024 MiB = 5120 MiB, so I thought it might be max for CUDALucas, but I can see your test is using 10750MB...

Anyway I'll keep testing those 5075MB I can test.

[kriesel]

Originally Posted by ATH

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 430.39       Driver Version: 430.39       CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX TIT... WDDM  | 00000000:02:00.0  On |                  N/A |
| 59%   55C    P0    97W / 250W |    185MiB /  6144MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

It is only using 185MiB / 6144 MiB, but still only 5117MiB free.

Which is close to 5 GiB = 5*1024 MiB = 5120 MiB, so I thought it might be max for CUDALucas, but I can see your test is using 10750MB...

Anyway I'll keep testing those 5075MB I can test.

Windows 10 v1709 or later? This is interesting, but may work differently on the display driving gpu's memory usage or igp than discrete or nondisplay gpu. https://www.techjunkie.com/watch-gpu-usage-windows-10/ Use with caution, as I've seen anomalies in its output. For example, on a laptop with an igp set to do the display work, and also a GTX1050Ti, it shows 99% usage of the gpu (igp) on mfakto; 0% on either mfaktc running on the GTX1050Ti; 80MB dedicated each mfaktc process, 24K on the mfakto process (perhaps its footprint is all in system memory).

[AlsXZ]

Hello! Could you help me to find linux binary of last version CUDALucas? I can't see it. Thanks.

[kriesel]

Originally Posted by AlsXZ

Hello! Could you help me to find linux binary of last version CUDALucas? I can't see it. Thanks.

It's my understanding that usually those using linux compile their own. But there's a CUDA 9.1 linux build at https://download.mersenne.ca/CUDALucas/old