Kalles Fraktaler 2.14

  • 245 Replies
  • 7559 Views

0 Members and 1 Guest are viewing this topic.

Offline saka

  • *
  • Fractal Phenom
  • ****
  • Posts: 45
« Reply #105 on: October 15, 2019, 02:10:57 PM »
- feature: SIMD support.  Compile KF yourself with optimisations for your own specific CPU for a bigger speed boost.

What level of CPU support is it currently compiled against?  Just basic SSE2?

Any thoughts of publishing an AVX2 release also?  It would cover most the recent Intel and AMD CPUs.  If you were going to publish one extra version, AVX2 is probably the sweet spot at the moment.

Out of curiosity, have you benched what the speed difference is with just SIMD support?

donations to support my work are welcome via https://mathr.co.uk/kf/kf.html#donate

Done!

Offline claude

  • *
  • 3f
  • ******
  • Posts: 1546
    • mathr.co.uk
« Reply #106 on: November 27, 2019, 04:46:31 PM »
Working on the next point release (hopefully ready before the end of the year, but no promises).  Progress report:

Fixed so far:
- don't set zoom size when resuming zoom out sequence (now you need to set it manually before resuming) (fixes issue where it was clamped to integer; reported by david)
- clamp half-precision colour output in EXR (fixes issue with NaNs)
- save file format version number with parameters and settings for future proofing
- list 3rd-party software links on the manual
  - exrtact for tile assembly including huge images (better than the old Octave method)
  - previous zoom interpolators that I have made
  - the zoom interpolator I am also working on this week (it works but is very slow so far, <4fps)
  - kf-extras
  - saka's ae movie tools
  - kf movie maker
  - ymm3d
  - Dinkydau's MMIT convertor
  - SeryZone's map visualizer
- add the thread about EXR format details to the user manual

In progress:
- OpenCL for perturbation calculations
  - works for double precision, but isn't as fast as I'd hoped
  - scaled double and floatexp still to do
  - need to investigate batching (like mightymandel) to avoid such pathological memory access patterns
  - breaks headless rendering at the moment, need to fix that asap

Later:
- use rescaled (long) double for series approximation and nanomb1+2 for speed boost
- rotation independent from skew
- check if "boxes in zoom out JPEG" (reported by saka) is reproducible with command-line zoom sequence rendering, fix it if so, otherwise probably not (the GUI zoom out renderer is incomprehensible spaghetti code to me)
« Last Edit: December 05, 2019, 01:59:45 PM by claude, Reason: worked a bit on the manual »

Offline CFJH

  • *
  • Fractal Frankfurter
  • *
  • Posts: 506
« Reply #107 on: December 01, 2019, 06:17:20 PM »
Hello,

currently I'm migrating my "stars" zoom to kf2.14.8. Here I've found some problems:

1. [bug]Zooming out by a custom factor (in my case 33554432, 2^25) and then quickly clicking right mouse button to zoom out produces a program crash (divide by zero). Lower zoom factors (eg 1024) but also entered via "custom" seems to be ok. See attachment kf_divide_by_zero.zip (wine errorbox? don't know how this appears on a real windoof).

2. [suggestion] When aborting a running calculation, the image rendered/displayed so far is often deleted from the screen. In kf 2.12.x the image was still visible

3. [bug] at certain zoom level the coordinates are not visible/editable in the location window (Strg+L). Also C&P from a text file is not possible. This seems to be a problem at higher zoom levels. 1E6000 is still ok, 1E7500 fails. See attachment kf_bug_koordinaten.png

4. [bug] Loading some parameters (created with kf2.12.9) from .kfr or .jpg files the calculation fails on kf2.14 but is on kf2.12.9 ok. I'm getting a blank image after the first reference. See attachments zm49_B07554_2x8star.kfr/jpg. (May be in conjunction with 3.?, bit not all old parameter are failing)


Offline claude

  • *
  • 3f
  • ******
  • Posts: 1546
    • mathr.co.uk
« Reply #108 on: December 03, 2019, 03:39:55 PM »
1. [bug]Zooming out by a custom factor (in my case 33554432, 2^25) and then quickly clicking right mouse button to zoom out produces a program crash (divide by zero). Lower zoom factors (eg 1024) but also entered via "custom" seems to be ok. See attachment kf_divide_by_zero.zip (wine errorbox? don't know how this appears on a real windoof).
I can't reproduce it here.  Your log crashes in WINE gdi32 StretchBlt() function, maybe a WINE bug(?).  I will investigate this in the coming weeks.

Quote
2. [suggestion] When aborting a running calculation, the image rendered/displayed so far is often deleted from the screen. In kf 2.12.x the image was still visible
Recent KF recolours the image when it is done because the in-progress colouring is just an approximation (especially with slopes/screen-space de colouring).  I can try to make it not do this on abort.  Thanks for the idea.

Quote
3. [bug] at certain zoom level the coordinates are not visible/editable in the location window (Strg+L). Also C&P from a text file is not possible. This seems to be a problem at higher zoom levels. 1E6000 is still ok, 1E7500 fails. See attachment kf_bug_koordinaten.png
This is a bit annoying, but editing 1000s of digits in a tiny box is not really productive anyway.  I'm not sure what the Windows limit for this widget is, maybe there is a buffer overflow that I need to fix.  You can copy/paste from Notepad or other text editor to/from KF, using keyboard shortcuts Ctrl-X and Ctrl-V in the main KF window (need to check if CRLF line endings are handled correctly in WINE on Linux as *nix uses just LF, let me know if it works or not).  I'm not sure what is the minimum text you need to have for Ctrl-V in KF to recognize as a location, but if you copy from KF, paste in editor, edit in editor, copy from editor, paste in KF, it should work.  Maybe I can try to use a better widget for the digits, but it's not such a high priority right now....

Quote
4. [bug] Loading some parameters (created with kf2.12.9) from .kfr or .jpg files the calculation fails on kf2.14 but is on kf2.12.9 ok. I'm getting a blank image after the first reference. See attachments zm49_B07554_2x8star.kfr/jpg. (May be in conjunction with 3.?, bit not all old parameter are failing)
Enabling "use floatexp always" it seems to render ok.  This must be a bug with "scaled long double" rendering.  Will investigate.  Thanks for the report.

Offline Adam Majewski

  • *
  • Fractal Flamingo
  • ****
  • Posts: 330
« Reply #109 on: December 04, 2019, 12:06:41 PM »
What about moving Fractal type from Menu/iterations to new menu item? ( for example Actions/Fractal type)
For me as a simple user it is hard to find it


Offline claude

  • *
  • 3f
  • ******
  • Posts: 1546
    • mathr.co.uk
« Reply #110 on: December 04, 2019, 04:51:59 PM »
It's documented at the very start of the manual so I think it is ok how it is.
https://mathr.co.uk/kf/manual.html#quick-start

Offline CFJH

  • *
  • Fractal Frankfurter
  • *
  • Posts: 506
« Reply #111 on: December 05, 2019, 11:18:18 PM »
.......
This is a bit annoying, but editing 1000s of digits in a tiny box is not really productive anyway.
......

This was a result of (4.). Alter loading the parameters file the render failed, I tried to enter copy&paste  the coordinates via the location window.
On older kf versions, this really worked, but it took minutes until the window opens (on E30000 zooms)

Offline claude

  • *
  • 3f
  • ******
  • Posts: 1546
    • mathr.co.uk
« Reply #112 on: December 31, 2019, 08:20:47 PM »
kf-2.14.9 released get it at https://mathr.co.uk/kf/kf.html#kf-2.14.9 you can donate to support my work via https://mathr.co.uk/kf/kf.html#donate

  -   feature: new formula `z^2 exp(2 a / z) + c` (suggested by gerrit); set seed to `0` (default) for singular orbit; set seed to `a` for critical orbit (large `|a|` bugs at zooms <= 1e3); no Newton zooming yet; thread: https://fractalforums.org/f/28/t/3234
  -  feature: use minimum `|z|` pixels for new references (suggested by quaz0r) (option disabled by default as it is not always faster)
  -  feature: save file format version number with parameters and settings
  -  feature: documentation improvements (references to third-party software and EXR map file semantics)
  -  bugfix: don't set zoom size when resuming zoom out sequence (now you need to set it manually before resuming) (fixes issue where it was clamped to integer; reported by david)
  -  bugfix: clamp half-precision colour output in EXR (fixes issue with NaNs)
  -  bugfix: interpolate neighbourhood when ignoring isolated glitches (appearance should be smoother, especially analytic DE was broken before)
  -  bugfix: avoid assertion failure when loading a parameter with fractal type out of range (instead unknown formulas appear incorrectly as Mandelbrot) (reported by gerrit)
  -  bugfix: remove -pipe from Makefile to allow builds with smaller RAM
  -  bugfix: updated Windows/MSYS2 build instructions (thanks to Patrick Owen)
  -  upgrade to boost 1.72.0
  -  upgrade to glm 0.9.9.6
  -  update openexr repository url
  -  update pixman build to skip demos
  -  enable fat gmp build (with runtime CPU detection for faster reference calculations)
  -  feature: experimental OpenCL support for GPU rendering (not built by default, enable it at build time with `make clean && make OPENCL=1`)

Offline gerrit

  • *
  • 3f
  • ******
  • Posts: 2165
« Reply #113 on: December 31, 2019, 11:33:15 PM »
For a=z0=-0.5 I get below result. Should be as 2nd image.

Offline claude

  • *
  • 3f
  • ******
  • Posts: 1546
    • mathr.co.uk
« Reply #114 on: January 01, 2020, 12:30:30 AM »
|-0.5| counts as large, see release notes (bugs at shallow zooms < 1e3)

Offline gerrit

  • *
  • 3f
  • ******
  • Posts: 2165
« Reply #115 on: January 01, 2020, 02:34:39 AM »
|-0.5| counts as large, see release notes (bugs at shallow zooms < 1e3)
Does this apply to critical orbit only? Singular orbit seems fine regarding that.
Attached other singular bug: should be minis in noisy area. Shows up when you zoom into them.

Offline gerrit

  • *
  • 3f
  • ******
  • Posts: 2165
« Reply #116 on: January 01, 2020, 05:47:34 AM »
Another test location. While blobs appear.
Also, my kfr file was called k400.kfr, but in my directory I found a file called "k400.2020-01-01T01-44-46-794.kfr" which must have been generated automatically somehow?

Offline claude

  • *
  • 3f
  • ******
  • Posts: 1546
    • mathr.co.uk
« Reply #117 on: January 01, 2020, 04:16:56 PM »
Does this apply to critical orbit only?
I think so.  Once you zoom past 1e3 the automatic glitch detection/correction should work as expected, but getting there is admittedly a pain in the butt. For a future release I plan to add "plain" (unperturbed) rendering for low zooms instead of fixed c=0+0i reference which should fix this issue.

Attached other singular bug: should be minis in noisy area. Shows up when you zoom into them.
For the noisy areas, try enabling "Glitch low tolerance" in the iterations dialog.  If that reliably fixes this kind of issue I can adjust the default glitch tolerance factor for this formula.  Or I can try to implement your improved glitch detection method instead of Pauldelbrot's criterion.

While blobs appear.
Perhaps try forcing long double always in case these are precision related and the extra 10 bits will be enough?  Or maybe overflow-related, in which case try long double or even floatexp.  Or if they are glitch-correction related try glitch low tolerance.  All a bit experimental so far, thanks for testing!

I found a file called "k400.2020-01-01T01-44-46-794.kfr" which must have been generated automatically somehow?
By request (in a previous version) I made Ctrl-S/Save generate timestamped files to avoid accidental overwrite, you can change this in the settings (Advanced -> Save overwrites existing file)

Offline gerrit

  • *
  • 3f
  • ******
  • Posts: 2165
« Reply #118 on: January 01, 2020, 08:47:57 PM »
I tried "glitch low tolerance" and "long double always" for the 2 examples (no mini interior/white circles) I posted to no effect.
If white circles are "escaped in 1 iteration" they may shrink with a bigger escape radius.

Offline gerrit

  • *
  • 3f
  • ******
  • Posts: 2165
« Reply #119 on: January 01, 2020, 08:57:44 PM »
[Large |a| bug]Once you zoom past 1e3 the automatic glitch detection/correction should work as expected, but getting there is admittedly a pain in the butt. For a future release I plan to add "plain" (unperturbed) rendering for low zooms instead of fixed c=0+0i reference which should fix this issue.
How about changing variables z-->z+a so c=0 becomes the critical orbit (so have critical/singular versions as separate formulas again), or is that too simplistic?


xx
Kalles Fraktaler 2.13

Started by claude on Kalles Fraktaler

90 Replies
4692 Views
Last post March 20, 2019, 05:32:26 AM
by claude
xx
Kalles Fraktaler 2 + GMP

Started by claude on Kalles Fraktaler

232 Replies
11741 Views
Last post September 12, 2019, 10:27:53 PM
by claude
xx
Another version of Kalles Fraktaler?

Started by greentexas on Kalles Fraktaler

9 Replies
1104 Views
Last post September 22, 2017, 02:59:16 PM
by greentexas
xx
Kalles Fraktaler Movie Maker64

Started by Bill Snowzell on Fractal movie gallery

0 Replies
300 Views
Last post February 01, 2018, 09:29:04 PM
by Bill Snowzell
question
Kalles Fraktaler tiled rendering

Started by Chronicler1701 on Kalles Fraktaler

5 Replies
444 Views
Last post September 06, 2018, 07:43:16 PM
by gerrit