For a better crop

kassandro · 14th July 2003, 10:48

If frames are not properly aligned, i.e. the address of the first pixel of a row is not a multiple of 16, than various filters, especially those which use sse are up to 20% slower. Most filters don't change alignment, but crop does it and usually crop is applied quite early leading even to more slowdown. I suggest do add an option like "align=16", to crop. Then crop has to do a bitblt if one wants to crop by an amount, which is not a multiple of 16.

sh0dan · 14th July 2003, 19:58

Couldn't hurt. I'll try it out.

sh0dan · 15th July 2003, 21:24

It will be in the next version. It's implemented as a bool for simplicity. It aligns if not 16 byte aligned on SSE2 capable machines and 8 byte aligned on other machines.

ARDA · 16th July 2003, 21:47

a little off topic
@sh0dan

I've been playing a little with chromas in YV12; doing a plugin that
uses SSE2.
I was testing with an interlaced source and I applied my
filter after SeparateFields and before weave;it was then I
discovered that the inicial address was not 16 byte aligned
after SeparateFields.
I had been doing a lot of tests and in almost all situations,I had no problems
to be able to use movdqa to read or write (except crop ).There must be others
cases in which this problem appears;it would be usefull if you could point them.
By now I opened a branch verifying...if ( int (srcpu) & 15 || int (srcpv) & 15){}
and in such case went back to ISSE routine.
(it is less expensive than making a lot of controls).

asking?...or asking

Is it possible to make something similar you've done with crop yesterday?

Thanks in advance
Arda

sh0dan · 16th July 2003, 21:58

I can sum up the safe things:

* pitch will be at least up to the next mod 16 (luma), mod 8 (chroma)
* env->NewVideoFrame(vi) will return a 16 byte aligned frame (pointer).
* env->NewVideoFrame(vi,32) will return a 32 (luma), 16 (chroma) byte aligned frame (pointer).

You can always create a new frame and BitBlt your material into that one (as I do in crop). You do however loose speed on that account too.

ARDA · 17th July 2003, 08:16

@sh0dan

quote:
--------------------------------------------------------------------------------
Originally posted by ARDA
By now I opened a branch verifying...if ( int (srcpu) & 15 || int (srcpv) & 15){}
and in such case went back to ISSE routine.
(it is less expensive than making a lot of controls).
--------------------------------------------------------------------------------

quote:
--------------------------------------------------------------------------------
Originally posted by ARDA
Is it possible to make something similar you've done with crop yesterday?
--------------------------------------------------------------------------------

quote:
--------------------------------------------------------------------------------
Originally posted by sh0dan
You can always create a new frame and BitBlt your material into that one (as I do in crop).
You do however loose speed on that account too.
--------------------------------------------------------------------------------

Sorry, in my own post was the answer.I was waiting a miracle.

Thanks
Arda

14th July 2003, 10:48	#1 \| Link
kassandro Registered User Join Date: May 2003 Location: Germany Posts: 502	For a better crop If frames are not properly aligned, i.e. the address of the first pixel of a row is not a multiple of 16, than various filters, especially those which use sse are up to 20% slower. Most filters don't change alignment, but crop does it and usually crop is applied quite early leading even to more slowdown. I suggest do add an option like "align=16", to crop. Then crop has to do a bitblt if one wants to crop by an amount, which is not a multiple of 16.

14th July 2003, 19:58	#2 \| Link
sh0dan Retired AviSynth Dev ;) Join Date: Nov 2001 Location: Dark Side of the Moon Posts: 3,480	Couldn't hurt. I'll try it out. __________________ Regards, sh0dan // VoxPod

15th July 2003, 21:24	#3 \| Link
sh0dan Retired AviSynth Dev ;) Join Date: Nov 2001 Location: Dark Side of the Moon Posts: 3,480	It will be in the next version. It's implemented as a bool for simplicity. It aligns if not 16 byte aligned on SSE2 capable machines and 8 byte aligned on other machines. __________________ Regards, sh0dan // VoxPod

16th July 2003, 21:58	#5 \| Link
sh0dan Retired AviSynth Dev ;) Join Date: Nov 2001 Location: Dark Side of the Moon Posts: 3,480	I can sum up the safe things: * pitch will be at least up to the next mod 16 (luma), mod 8 (chroma) * env->NewVideoFrame(vi) will return a 16 byte aligned frame (pointer). * env->NewVideoFrame(vi,32) will return a 32 (luma), 16 (chroma) byte aligned frame (pointer). You can always create a new frame and BitBlt your material into that one (as I do in crop). You do however loose speed on that account too. __________________ Regards, sh0dan // VoxPod

Thread Tools	Search this Thread
Show Printable Version Email this Page	Search this Thread: Advanced Search
Display Modes
Linear Mode Switch to Hybrid Mode Switch to Threaded Mode

16th July 2003, 21:47	#4 \| Link
ARDA Registered User Join Date: Nov 2001 Posts: 291	a little off topic @sh0dan I've been playing a little with chromas in YV12; doing a plugin that uses SSE2. I was testing with an interlaced source and I applied my filter after SeparateFields and before weave;it was then I discovered that the inicial address was not 16 byte aligned after SeparateFields. I had been doing a lot of tests and in almost all situations,I had no problems to be able to use movdqa to read or write (except crop ).There must be others cases in which this problem appears;it would be usefull if you could point them. By now I opened a branch verifying...if ( int (srcpu) & 15 \|\| int (srcpv) & 15){} and in such case went back to ISSE routine. (it is less expensive than making a lot of controls). asking?...or asking Is it possible to make something similar you've done with crop yesterday? Thanks in advance Arda

17th July 2003, 08:16	#6 \| Link
ARDA Registered User Join Date: Nov 2001 Posts: 291	@sh0dan quote: -------------------------------------------------------------------------------- Originally posted by ARDA By now I opened a branch verifying...if ( int (srcpu) & 15 \|\| int (srcpv) & 15){} and in such case went back to ISSE routine. (it is less expensive than making a lot of controls). -------------------------------------------------------------------------------- quote: -------------------------------------------------------------------------------- Originally posted by ARDA Is it possible to make something similar you've done with crop yesterday? -------------------------------------------------------------------------------- quote: -------------------------------------------------------------------------------- Originally posted by sh0dan You can always create a new frame and BitBlt your material into that one (as I do in crop). You do however loose speed on that account too. -------------------------------------------------------------------------------- Sorry, in my own post was the answer.I was waiting a miracle. Thanks Arda