On 06/22/2017 02:42 AM, Nikolay Nikolov wrote:
On 06/22/2017 01:21 AM, James Richters wrote:
putimage can be accelerated, although it would still have to do a
memory copy.
Like this?
https://github.com/Zaaphod/ptcpas/compare/Zaaphod_Custom?expand=1#diff-fb31461e009ff29fda5c35c5115978b4
This is amazingly faster. I ran a test of just ptcgraph.putimage()
in a loop, putting the same image over and over 1000 times and timing
it. The original ptcgraph.putimage() took 18.017 seconds. After I
applied this, the same loop took 1.056 seconds. Quite an
improvement! It's still nowhere near as fast as just drawing stuff
with ptcgraph directly, but for doing a memory copy of the entire
screen, it's very fast
Yes, that's a good start. That was exactly what I meant :)
I have an idea on how I could speed it up even further....
If I set up a second array with 1 bit per pixel, then (somehow)
aggpas could set bits in this array to 1 whenever it changed a
corresponding bit. Now by analyzing the 'pixel changed' array one
word at a time, (or maybe longword or qword at a time) I could just
skip over all the words that =0 and when I come across a word that <>
0 I could do a binary search of that word to only change the pixels
that need to be changed. If very little on the screen has changed,
this would be quite a bit faster because the pixel changed array
would be 1/16 the size of the full buffer.
The only way this would be of any benefit though is if aggpas set the
bits in the 'pixel changed' array while it was changing the pixels of
the buffer, because at that time it already has the array position
and the fact that something changed available. If I had to analyze
the buffer separately and create the 'pixels changed' array, it would
take too long.
That sounds like a little bit of a special case - it'll work where
you're using putimage for a large area, that has very few pixels set.
Perhaps just reimplementing the general algorithm in inline asm, by
using SSE (or MMX) vector instructions would be the fastest, but maybe
it's not worth the pain and the pascal implementation is fast enough
for you. Just experiment and see what works best :)
Btw, I looked at your code again and saw a quick and cheap optimization
- just move the case statement (case BitBlt of) outside the inner loop
(for i:=X to X1 do), so the value of BitBlt is not checked once every
pixel, but once per row.
Nikolay
_______________________________________________
fpc-pascal maillist - fpc-pascal@lists.freepascal.org
http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-pascal
