• January 21, 2022, 12:51:06 PM

Author Topic: (Problem) test cases for potential "scaled long double" rendering  (Read 619 times)

0 Members and 1 Guest are viewing this topic.

Offline claude

  • 3f
  • ******
  • Posts: 2159
    • mathr.co.uk
(Problem) test cases for potential "scaled long double" rendering
« on: January 23, 2018, 11:05:42 AM »
currently kf has these number types for these (approximate) ranges of values, for quadratic Mandelbrot:

double: 1e-300
scaled double: 1e-600
long double: 1e-4900
floatexp: everything beyond that

I would like to add (between long double and floatexp)

scaled long double: 1e-9800

but I don't have any locations to test with: so if you have any nice locations between 1e-4000 and 1e-12000 that preferably don't need super high iteration counts, please post them in this thread.

thanks!


Linkback: https://fractalforums.org/index.php?topic=754.0

Offline claude

  • 3f
  • ******
  • Posts: 2159
    • mathr.co.uk
Re: test cases for potential "scaled long double" rendering
« Reply #1 on: January 23, 2018, 06:40:24 PM »
simple location near the tip of the needle at 1e6000 attached.  this is rendered correctly in kf-2.13 (unreleased) with floatexp and derivatives for de in 8mins (10refs).  so far I haven't managed to get scaled long double to give a good image (I think a problem with the derivatives and colouring more than anything else), but it takes 13mins (72refs).  So I think it might not be worth it.

similarly in kf-2.13 scaled double takes 5m40 and long double takes 21s for Dinkydau's Ssssss.kfr, at 4e533. edit kf-2.12 takes 10s with scaled double and 15s with long double - so it's definitely something to do with (long)double conversions for derivatives...

I will run some tests in kf-2.12 in case the derivatives and type conversions are slowing it down too much to make a fair comparison (scaled double needs long double derivatives to avoid overflow, similarly scaled long double needs floatexp derivatives).

but in the meantime my conclusion so far is scaling isn't worth it as an optimisation, as it's both slower and more complicated...
« Last Edit: January 23, 2018, 06:51:58 PM by claude, Reason: kf-2.12 benchmark »

Offline Dinkydau

  • Uploader
  • *
  • Posts: 316
    • DeviantART gallery
Re: test cases for potential "scaled long double" rendering
« Reply #2 on: January 23, 2018, 07:36:24 PM »
Do you count the time to compute references? Of course it's slower when more references are used.

There are two locations that I have in your range. One is from the image "3rd Order Evolution Of Tilings":
https://fractalforums.org/fractal-image-gallery/18/3rd-order-evolution-of-tilings/214/msg1150
The other one is something I still need to render so please use it only to test.
« Last Edit: January 23, 2018, 09:02:29 PM by Dinkydau »

Offline claude

  • 3f
  • ******
  • Posts: 2159
    • mathr.co.uk
Re: test cases for potential "scaled long double" rendering
« Reply #3 on: January 23, 2018, 08:18:03 PM »
Yes I counted total time in the status bar - I haven't instrumented timing for separate phases, is quite tricky to do...  still confused why scaling needs so many more references / so many more glitches detected...

Trying again to implement scaled long double in kf-2.12 this time, without derivatives.  Should be a lot simpler...

Thanks for the locations, I have downloaded them so you may delete the attachments if you wish...

Offline claude

  • 3f
  • ******
  • Posts: 2159
    • mathr.co.uk
Re: test cases for potential "scaled long double" rendering
« Reply #4 on: January 24, 2018, 12:25:17 AM »
Progress: got scaled long double working in kf-2.12, and benchmarks are much more promising (not as slow as expected, but series approximation underskips a lot at my 1e6000 location):

scaled long double: 4m40 11refs 19941skip (first ref) 25% slower
forced floatexp: 3m40 11refs 24926skip (first ref)

Appearance seems correct too.   Though non-analytic DE colouring looks poopy (attached).

Other locations show a big improvement, however, so I hope to release kf-2.12.9 very soon with this feature:

Dinkydau's 1e4770: scaled (not deep enough, so regular unscaled) long double 6.3x faster
scaled long double: 12m45 for 100% of 1ref (I didn't wait longer) 1896167skip (first ref) (1m25s for reference before pixels)
forced floatexp: 27m30 for 33% of 1ref (didn't wait) 1896167skip (first ref) (same skip as above) (estimated 80mins for 100% of 1 ref)

Dinkydau's 7e8382: scaled long double 5x faster
scaled long double: 1h12m15 for 5% of 1ref (didn't wait) 17910782skip (first ref took 55m before pixel computations started, this is included in the total previous) estimated 400mins for 100% of 1 ref
forced floatexp: 1h12m20 for 1% of 1ref (didn't wait) 11885108skip (first ref took 53m before pixels) (massive underskipping here means I couldn't wait for more pixels) estimated 2000mins for 100% of 1 ref

all tests at 640x360 with default settings (other than toggling "use floatexp always"), on quad core amd64 desktop at 3GHz

Offline Dinkydau

  • Uploader
  • *
  • Posts: 316
    • DeviantART gallery
Re: test cases for potential "scaled long double" rendering
« Reply #5 on: January 24, 2018, 01:21:04 AM »
Yes, the E4770 location is not deep enough exactly because I optimized it to avoid floatexp, after I asked about it in the kalles fraktaler thread.

Offline claude

  • 3f
  • ******
  • Posts: 2159
    • mathr.co.uk
Re: test cases for potential "scaled long double" rendering
« Reply #6 on: January 24, 2018, 01:23:18 AM »
I released 2.12.9, please test!


xx
Computing bounds on the potential function and it's derivative

Started by Laser Blaster on Fractal Mathematics And New Theories

15 Replies
752 Views
Last post April 04, 2020, 07:46:42 AM
by gerrit
question
Mandelbulber - Strange "tasselation" problem with GPU but not with CPU

Started by andrics96 on Noob's Corner

7 Replies
425 Views
Last post December 15, 2021, 12:50:34 AM
by mclarekin
xx
But how long IS the coastline of brittian?

Started by DeusDarker on Fractal Mathematics And New Theories

6 Replies
312 Views
Last post October 10, 2020, 06:19:47 AM
by fractower
question
Newton Raphson - how long wilt it run ?

Started by CFJH on Kalles Fraktaler

13 Replies
838 Views
Last post September 11, 2019, 08:38:32 AM
by claude
xx
Neural Style Transfer enhanced Mandelbrot zoom

Started by schizo on Fractal movie gallery

6 Replies
989 Views
Last post December 22, 2020, 09:31:14 PM
by schizo