[QUOTE=preda;496544]Yes, I'll look into improving loading time (I have an idea, need to implement it)[/QUOTE]

The load time should be better now.

GCD test
 

During PRP-1, one GCD check is done every 1M iterations, if enough MULs (more than 2000) have accumulated.

The GCD is done on the CPU, in parallel with the GPU (i.e. the GPU does not wait and is not stalled). The GCD takes about 45s for 90M exponents, and uses a single CPU thread.

In addition to this GCD, supplementary MULs are done on the GPU. The MULs aren't distributed uniformly, but are concentrated in the first third of the iteration range. So the "iteration time" may become slower in when there are many MULs being done.

On manual quit (ctrl-C) one more GCD is done. This is annoying (having to wait for it on exit), but seems simpler than having to persist&load the accumulated MULs.

[QUOTE=preda;496748]The load time should be better now.[/QUOTE]


My 332M exponent was > 60 days of work, now at 40 days of work, I am keeping ver. 3.5 for speed. I am going to test the new ver. with some smaller exponent.

[QUOTE=preda;496741]You should try doing a "make" in that branch. Probably specifying the target like:
"make cudaowl"

Have a look in the Makefile.
In general you need CUDA toolkit, nvcc.
also a general c++ compiler (g++)

I can't help much on the CUDA side.[/QUOTE]

Thanks for that, I will definitely take a look. However, I currently don't have a Linux system with CUDA GPU. I am just hoping that there is some way to build it for windows so I could try using windows instead of Linux for simplicity and convenience as I am really having trouble mixing AMD and Nvidia GPU in the same system for my ubuntu 16.04 installation.

[QUOTE=xx005fs;496759]Thanks for that, I will definitely take a look. However, I currently don't have a Linux system with CUDA GPU. I am just hoping that there is some way to build it for windows so I could try using windows instead of Linux for simplicity and convenience as I am really having trouble mixing AMD and Nvidia GPU in the same system for my ubuntu 16.04 installation.[/QUOTE]

Mixing AMD & NVIDIA is known to be problematic on Windows too. I experimented with it a bit and decided to segregate them to separate systems. See [URL]https://www.mersenneforum.org/showpost.php?p=494876&postcount=680[/URL] for links to available Windows builds of gpuowl. If you want to do your own build, see [URL]https://www.mersenneforum.org/showpost.php?p=494759&postcount=666[/URL]
If I recall correctly, I got this approach of emulating make with command lines from an example for v2.0 by kracker, including creating the build environment on Windows, in [URL]https://www.mersenneforum.org/showpost.php?p=483209&postcount=356[/URL]

[QUOTE=kriesel;496765]Mixing AMD & NVIDIA is known to be problematic on Windows too. I experimented with it a bit and decided to segregate them to separate systems. See [URL]https://www.mersenneforum.org/showpost.php?p=494876&postcount=680[/URL] for links to available Windows builds of gpuowl. If you want to do your own build, see [URL]https://www.mersenneforum.org/showpost.php?p=494759&postcount=666[/URL]
If I recall correctly, I got this approach of emulating make with command lines from an example for v2.0 by kracker, including creating the build environment on Windows, in [URL]https://www.mersenneforum.org/showpost.php?p=483209&postcount=356[/URL][/QUOTE]

I saw a whole bunch of OpenCL build on the mersenne.ca mirror. I might try to figure out how to build the CUDA version myself later on. It is really difficult for me to separate my AMD GPU with Nvidia GPU because I am using a custom loop and I currently only has 1 PC that's functioning so I guess I would try some methods online to mix and match Nvidia and AMD GPU in Linux. Any advice on building the CUDA branch will definitely be extremely helpful. Also I could obviously use CUDALUCAS for my Nvidia GPUs and GpuOwL for AMD, but I just thought that completing assignments using an algorithm that might not need double checking will certainly be helpful for the future when double checking actually reaches these exponent ranges.

[QUOTE=xx005fs;496781]I saw a whole bunch of OpenCL build on the mersenne.ca mirror. I might try to figure out how to build the CUDA version myself later on. It is really difficult for me to separate my AMD GPU with Nvidia GPU because I am using a custom loop and I currently only has 1 PC that's functioning so I guess I would try some methods online to mix and match Nvidia and AMD GPU in Linux. Any advice on building the CUDA branch will definitely be extremely helpful. Also I could obviously use CUDALUCAS for my Nvidia GPUs and GpuOwL for AMD, but I just thought that completing assignments using an algorithm that might not need double checking will certainly be helpful for the future when double checking actually reaches these exponent ranges.[/QUOTE]
AMD GPU driver is already problematic alone in Linux. I think mixing drivers will worse things.

CUDAowl is unsupported & abandoned
 

[QUOTE=xx005fs;496781] I might try to figure out how to build the CUDA version myself later on. Any advice on building the CUDA branch will definitely be extremely helpful. [/QUOTE]It's worth noting that Preda gave up on CUDAowl development, and sold his NVIDIA gpu bought for such development on linux, and was unable or unwilling to spend the time required to get CUDAowl going on Windows with assistance from people recursively making build attempts on Windows and reporting error messages observed in those attempts. [URL]https://www.mersenneforum.org/showpost.php?p=494414&postcount=644[/URL] [URL]https://www.mersenneforum.org/showpost.php?p=494430&postcount=645[/URL]

[QUOTE=kriesel;496819]It's worth noting that Preda gave up on CUDAowl development, and sold his NVIDIA gpu bought for such development on linux, and was unable or unwilling to spend the time required to get CUDAowl going on Windows with assistance from people recursively making build attempts on Windows and reporting error messages observed in those attempts. [URL]https://www.mersenneforum.org/showpost.php?p=494414&postcount=644[/URL] [URL]https://www.mersenneforum.org/showpost.php?p=494430&postcount=645[/URL][/QUOTE]

Guess LL it is on CUDA. Thanks for the info :D

[QUOTE=xx005fs;496781]It is really difficult for me to separate my AMD GPU with Nvidia GPU because I am using a custom loop and I currently only has 1 PC that's functioning.[/QUOTE]I've found well built older used systems such as HP Z600 or Lenovo D20 or D30 economical on ebay and useful for separating gpu types and separating variables. Plus, another box, more throughput.
Z600 has enough power connectors, power supply capacity, and internal space to run 3 gpus, if one is single-slot width, and one at most requires 150W.
D20/30 have capability of running a couple of double slot width 250W gpus, plus perhaps a third on an extender off pcie x1 or a single slot card such as Q2000 (about to try adding that).
Fully populated, either will definitely affect the utility bill.

[QUOTE=kriesel;496902]I've found well built older used systems such as HP Z600 or Lenovo D20 or D30 economical on ebay and useful for separating gpu types and separating variables. Plus, another box, more throughput.
Z600 has enough power connectors, power supply capacity, and internal space to run 3 gpus, if one is single-slot width, and one at most requires 150W.
D20/30 have capability of running a couple of double slot width 250W gpus, plus perhaps a third on an extender off pcie x1 or a single slot card such as Q2000 (about to try adding that).
Fully populated, either will definitely affect the utility bill.[/QUOTE]

Interesting. But I don't think I will do those large operations with a lot of gpus yet.