Comments on: First week of Krita full-time hacking

By: Kevin Kofler

Kevin Kofler — Fri, 05 Feb 2010 15:31:16 +0000

Maybe approximating the atan with something else would be an option? E.g. use some integration formula on 1/(1+x²), or use a polynomial approximation.

By: Boudewijn Rempt

Boudewijn Rempt — Tue, 02 Feb 2010 15:43:05 +0000

My guess would be that to find hardware that works with OpenSolaris, installing it, making Lukas familiar with opensolaris, then learning dtrace could easily eat up a month or two, maybe three. Systemtap might be an option since Lukas is already using fedora. But I doubt that systemtap would help us more than valgrind at this stage. Maybe when we’ve optimized the hell out of our algorithms, memory and cache usages and are faced with weird latencies or io issues.

By: Markus

Markus — Tue, 02 Feb 2010 12:20:04 +0000

Wouldn’t using OpenSolaris with DTrace be better for finding performance bottlenecks?

By: LukasT

LukasT — Tue, 02 Feb 2010 08:44:32 +0000

@chickenpump: yes, we decided to do lookup table for atan2. Problem is that it accept two parameters so it will be 2D lookup table and quite big probably.
But we plan to add lazy initialization, so it maybe quite fast.

By: Brad Hards

Brad Hards — Tue, 02 Feb 2010 07:41:03 +0000

It might be worth tracking which compiler version and compiler options you are using for this. Obviously that could make some difference, although I readily concede it may not be very much in the end.

By: chickenpump

chickenpump — Mon, 01 Feb 2010 23:40:44 +0000

Hi Lukas,

just a tipp for the speedup of the atan2 function:

in the old amiga/atari days we used precomputed memory maps , like x = sinus_array[y] , which should also be possible for arc-tan (split to sin/cos)

this give a huuuge speedup there.

By: Boudewijn Rempt

Boudewijn Rempt — Mon, 01 Feb 2010 15:28:38 +0000

Dmitry,

Wouldn’t that only work for composite_copy or composite_over with 100% opacity? Or do you want to pass the compositeop object to the datamanager?

By: DmitryK

DmitryK — Mon, 01 Feb 2010 09:49:07 +0000

About iterators: iirc, KisRandomAccessor uses memmove for for storing cyclic cache instead of (i%N). I thing this is the easiest thing to fix =)
About bitBlt: we need to implement directBitBlt inside datamanager, that will be used when color spaces coincide. It will share the tiles directly. It should be really fast on tiles3.