panzerboy
Fractal Lover
Posts: 242
|
|
« Reply #150 on: May 06, 2014, 01:30:18 PM » |
|
I have 3 GiB of RAM and I think thats causing me problems with 1680x1050 8x8 supersampled pictures. Watching the task manager when the mandelbrot starts computing 1,100,000 (ish) k bytes is used. Then the CPU usage drops to 50 % (ish) so I assume Mandel Machine is doing something single threaded. Soon the initialising series approximation message shows on the status bar and the memory increases to 2,300,000 (ish). The render is then much slower than before. Perhaps Mandel Machine is having to search through 1GiB of memory assigned by the series approximation? Another round of approximation is needed to fill in the details and thats where I give up. The memory keeps climbing up to a maximum of 2,385,000 k bytes then the CPU falls to 0 and the memory falls back to 2,315,000. Over and over, perhaps Javaw.exe is hitting an allocation limit, has to garbage collect to free memory and keeps going through this cycle? I've attached the mmf parameter file is you're interested, removed the .txt extension to load into Mandel Machine.
|
|
|
Logged
|
|
|
|
Botond Kósa
|
|
« Reply #151 on: May 10, 2014, 06:10:21 PM » |
|
I tried to render your location and it seems there is a memory leak in my glitch detection algorithm. GC is constantly trying to free up some memory, and this causes the whole application to slow down. Will be fixed in the next release. Thanks for the feedback!
|
|
|
Logged
|
|
|
|
simon.snake
Fractal Bachius
Posts: 640
Experienced Fractal eXtreme plugin crasher!
|
|
« Reply #152 on: May 10, 2014, 10:41:14 PM » |
|
Hi Botond
I have a linux machine (running CentOS 6.5) and I copied Mandel Machine to it with the intention of getting MM running under Wine. Surprisingly, it didn't run (I think it is a problem with the Java Machine you supply) but out of curiosity I installed Sun's Java Runtime and then clicked the jar file, and It burst into life.
There seem to be some issues with the system not detecting all glitches, but I can live with that.
Is the program working perfectly? Not sure. But so far all signs are positive.
|
|
|
Logged
|
To anyone viewing my posts and finding missing/broken links to a website called www.needanother.co.uk, I still own the domain but recently cancelled my server (saving £30/month) so even though the domain address exists, it points nowhere. I hope to one day sort something out but for now - sorry!
|
|
|
SeryZone
Strange Attractor
Posts: 253
Contemplate...
|
|
« Reply #153 on: May 18, 2014, 06:57:32 PM » |
|
Botond, do you try to render my location???
|
|
|
Logged
|
|
|
|
Botond Kósa
|
|
« Reply #154 on: May 18, 2014, 07:43:29 PM » |
|
Which location?
|
|
|
Logged
|
|
|
|
SeryZone
Strange Attractor
Posts: 253
Contemplate...
|
|
« Reply #155 on: May 18, 2014, 07:47:36 PM » |
|
Which location?
My, super-hard and super-dense, that I given to you! It requires about 2,000,000,000 iterations for normal minibrot!
|
|
|
Logged
|
|
|
|
Botond Kósa
|
|
« Reply #156 on: May 18, 2014, 10:30:34 PM » |
|
Oh, I remember, that is a tough location! I could only raise the iteration count to 250 millions (I have 10 GB RAM), and it is still too low to display anything but a flat area.
But stay tuned, the next release of Mandel Machine will be focused around a cleverer series approximation, and one pleasant side-effect will be reduced memory footprint!
|
|
|
Logged
|
|
|
|
SeryZone
Strange Attractor
Posts: 253
Contemplate...
|
|
« Reply #157 on: May 19, 2014, 06:46:14 AM » |
|
Oh, I remember, that is a tough location! I could only raise the iteration count to 250 millions (I have 10 GB RAM), and it is still too low to display anything but a flat area.
But stay tuned, the next release of Mandel Machine will be focused around a cleverer series approximation, and one pleasant side-effect will be reduced memory footprint!
Strange, Kalles Fraktaler use 6 GB when 400M... So, I guess, if you make save reference to the file, programm will works slowly. it is right?
|
|
|
Logged
|
|
|
|
Botond Kósa
|
|
« Reply #158 on: May 20, 2014, 02:57:27 PM » |
|
Writing the reference orbit to disk could help, but it is unnecessary. In the next version of MM, only the iterations after series approximation will be kept in memory, reducing the memory consumption by at least 90% on average. Efficient series approximation is key to fast perturbation calculation, even more than I previously thought. Take this location for example: http://stardust4ever.deviantart.com/art/Magnum-Opus-Ex-3132-6-Zooms-274474754This was rendered at 3200x3200 resolution in 28 days using Fractal Extreme, using 4-8 core AMD machines. The next version of MM renders it on my dual-core Ivy Bridge laptop in just under 30 seconds. Taking into account the differences in hardware, the speedup is more than 100000x
|
|
|
Logged
|
|
|
|
hapf
Fractal Lover
Posts: 219
|
|
« Reply #159 on: May 20, 2014, 04:28:40 PM » |
|
You skip 90% of iterations on average? With how many coefficients? And what are the max errors you accept?
|
|
|
Logged
|
|
|
|
Kalles Fraktaler
|
|
« Reply #160 on: May 20, 2014, 04:55:11 PM » |
|
That sounds amazing! The location has minimum 377372 and maximum 378486 iterations. The difference is 1114, so I guess you need to calculate at least that amount? KF can skip 310508 iterations, but renders this in 640x360 in 6 minutes, which is "only" 150 times faster than FX, not considering hardware... Long double is used for this location, which breaks KF if the maximum number of iterations is set lower than 310508
|
|
|
Logged
|
|
|
|
laser blaster
Iterator
Posts: 178
|
|
« Reply #161 on: May 20, 2014, 05:10:37 PM » |
|
Amazing speedup! What do you mean only the iterations after series approximation are kept in memory? I thought the amount of iterations you can skip with SA varies from pixel to pixel. Or do you set some minimum threshold of iterations to skip with SA, and if SA is not effective enough for a pixel, you try a new reference point?
|
|
|
Logged
|
|
|
|
Botond Kósa
|
|
« Reply #162 on: May 21, 2014, 01:37:04 AM » |
|
I made Mandel Machine v1.2 available. It is still in development, so this is a beta version. Some of the promised features (reduced memory footprint, true unlimited zoom depth or Pauldelbrot's new glitch correction method) are still missing, and I also had to disable some previously working features temporarily. Those will be available in the final release. This version is focused around two key areas: maximum zoom depth is increased to 3700 (this required the implementation of extended precision routines), and series approximation is adjustable. It can be made more aggressive than before, using up to 130 terms. The cost of initializing SA can expressed as L ref*N terms2, where L ref is the length of the reference orbit and N terms is the number of terms (coefficients) used in the approximation. As this cost is independent of image size, increasing the number of terms above a certain limit makes sense only by huge images where the savings outweigh the increased cost. At small to medium image sizes a right balance has to be found. It might also be automatable in the future. A full list of changes and the download link can be found on my Mandel Machine site: http://web.t-online.hu/kbotond/mandelmachine/
|
|
« Last Edit: May 21, 2014, 02:42:38 AM by Botond Kósa »
|
Logged
|
|
|
|
Dinkydau
|
|
« Reply #163 on: May 21, 2014, 01:53:26 AM » |
|
Awesome, thank you for making a beta available. It's late for me now so I'll be using it tomorrow.
|
|
|
Logged
|
|
|
|
Botond Kósa
|
|
« Reply #164 on: May 21, 2014, 02:04:46 AM » |
|
You skip 90% of iterations on average? With how many coefficients? And what are the max errors you accept?
The number of coefficients is adjustable between 6 and 130. The higher this number is, the more iterations can be skipped. It also depends largely on the actual location. Errors are no longer measured. The number of skipped iterations is determined based on the coefficients only. The basic idea is that if we have N coefficients, X iterations can be skipped safely if the Xth iterations of the coefficients can be divided into two groups that fulfill the following criteria: - The first group contains the first M coefficients (A1,X, A2,X ... AM,X), the second group the remaining N-M coefficients (AM+1,X, AM+2,X ... AN,X)
- Every term based on the coefficients of the first group is several orders of magnitude larger than any term based on the coefficients of the second group. A term for the coefficient Ai,X is Ai,X * delta0i
As long as the coefficients can be divided into these two groups, the perturbed orbit can be approximated with the first M terms safely.
|
|
|
Logged
|
|
|
|
|