Looking for a PIV to test my deinterlace filter

[Slyde]

Hi,

I wrote a deinterlace filter for Vdub.

Main routines are coded in assembly and optimised for MMX and SSE.

I've tested it on a Celeron 900@1200, Athlon K7 classic 700@1000 and Athlon XP 1700+ so far.

Seems this filter uses lots of memory bandwith as it is 24% faster than internal Deinterlace (blend fields) on the Athlon XP, but 30% slower on the Athlon Classic.

I'm looking for a PIV box owner as those systems offers huge memory bandwith.

If someone could report the preview time for 5000 or 6000 frames from a AVS script doing nothing than frameserving a vob file in 720*576 with my filter and internal deinterlace in blend mode it would be perfect.

The URL :
http://www.mpeg4-fr.com/Web/files/DeinterlaceInterpolate.vdf

[Ookami]

Hello Slyde,

sorry, this posting is only to bring to your attention that there is already a filter with a very similar (or almost the same) name out there:

Deinterlace PAL interpolation by tHE fISH

http://www.vdfilters.f2s.com/

"Almost the same as Deinterlace PAL by Gunnar Thalin, except that this filter checks the output for remaining interlace lines which are then removed by interpolation."

Direct download http://vdfilters.exit.mytoday.de/download/tHE%20fISH/DeInterPALInterpolation_1_0b1.zip

Thank you for your work on a new VD filter.

Cheers,

Ookie.

[Slyde]

Yeah I know there's at least two or three filters using a similar name, but I don't know how to name it...

Moreover, I've been having a look to the sources of half these filters, very few uses assembly or MMX. On a high bandwith system, my filter should overtake by far Gunnar Thalin or tHE fISH filter, as they are slower than my 1.1 release on my cel 900@1200 box.

I'm thinking of creating a similar filter for AVIsynth with YUY2 color space. I noticed Donald A. Graft has a trick to interpolate in YUY2 CS in his Decomb filter. I will give a closer look to his code as it may be highly accelerated.

For now, PIV owners seems to be very few :(

[Sven Bent]

Seems this filter uses lots of memory bandwith as it is 24% faster than internal Deinterlace (blend fields) on the Athlon XP, but 30% slower on the Athlon Classic.


seem more that its optimized ith SSE :-)
the big difffren on classic vs Xp i woudl more tend to the SSE rather then the increased bus bandwidth

[Slyde]

SSE instructions used are only SSE integer instructions (SSE is mainly floating point which is useless here), implemented in MMX extension since the AMD Athlon K7. The 3Dnow! extension set supported by K6-2 already implement the instructions I'm using, so I tested with them and got NO difference in performances. So memory bandwith must be critical. By the way, if you try to use SSE on a processor that do not support it (eg PII) you will get an nice invalid opcode as a result.

Athlon XP must be faster due to their L2 data prefetch (I'm using prefetchnta wisely to increase performances) and 1:1 L1/L2 cache ratio. As you know Athlon K7/K75 L2 cache is only 1/3, 2/5 or 1/2 the processor speed.

I released v1.1, it is much faster on K7 classic (+5.5% faster than internal blend deinterlace) and a bit faster on other processors.

I'm still unable to find a PIV box user that could test release 1.1 :(

[alkasecond]

Hi I just tried your deinterlacer.My system :P4 1.6A @2.13GHz,
512MB (unfortunately DDR).

I run 3000 frames of mjpeg captured file , 640x480.
Just running deinterlace and compressing back to mjpeg.

Results :
Your filter 1m35s,
built in 1m32s.

I must admit , qualitywise your is much better , here are snapshots:
1500 frame (fairly fast motion)

Vdub's built in deinterlacer:
http://members.rogers.com:81/alkaprim/int0.JPG

and yours:

http://members.rogers.com:81/alkaprim/ext0.JPG

[alkasecond]

Sorry I've got some problems.

[Slyde]

Thank you for testing alkasecond.

I'm a bit surprised it is slower on PIV, I read carrefully the PIV programming guide in order to get maximum performances. I know both ALU and FPU on the PIV are slower than those of the Athlon XP, but I expected the memory bandwith as well as the 512KB L2 cache to help a lot.

[Slyde]

here is the previous release :

http://www.mpeg4-fr.com/Web/files/DeinterlaceInterpolate1.vdf

Seems it is much faster on Athlon XP so I expect it to be also faster on PIV, but I don't know really why for the moment.

Could you test it a few minutes alkasecond ? thanks.