[RESOLVED] i7 2600k

[sandy_geser]

Hi

I've recently got an i7 2600k (win7 64bit, , Nvidia 9600gt). I've tested dgavcnv
vs directshow (ffdshow) using an avs script for x264 encoding of a blu-ray source.

seems to be a huge difference in 1st pass speeds (exact same x264.exe variables on both):

using DGAVCNV x86 .dgi: encoded 5001 frames, 29.93 fps, 5067.45 kb/s

using directshow with ffdshow: encoded 5001 frames, 83.85 fps, 5068.75 kb/s


with my older cpu (win7 64bit, i7 950, 9500GT) difference was a lot more minimal. around ~40fps
with dgavcnv and ~47fps with directshow/ffdshow.

Seems like 1st pass is slower than it was with the older cpu.

further info - this is for encode resized to 720p. resizing isn't through dgavcnv but one
the common avs resizers. It might be faster if I resized on the dgi line but then I can only
crop in increaments of 4 which is not accurate enough for my taste. driver installed at the time of this test:
8.17.12.6658 from 1/7/2011.

Further more, I was was also trying to use the x64 DGAVCNV version but avisynth fails to load the x64 DGDecodeNV.dll.
Could be that it's cause avisynth isn't x64 ? I don't think there is an x64 version. or am I
doing something wrong in the process?

I'm not a huge expert so help would be much appreciated!

[admin]

DGAVCNV?

I'm interested in the raw decoding speeds for the two scenarios but not so much transcoding speeds, because so many other variables enter. Please give the raw decoding speed numbers. You use DGIndexNV with the Disable DIsplay option as described in the user manual.

Yes, you need 64-bit Avisynth to run the 64-bit executables.

[sandy_geser]

DGAVCNV?

I mean DGDecNV

I'm interested in the raw decoding speeds for the two scenarios but not so much transcoding speeds, because so many other variables enter. Please give the raw decoding speed numbers. You use DGIndexNV with the Disable DIsplay option as described in the user manual.

I use DGIndexNV.exe to frameserve the source to a dgi file which I then use to transcode with x264 via avisynth. the variables are very few really when it comes to blurays. it's pretty much just crop/resize and no additional filters. I'm mostly interested in transcoding speeds but I'm happt to report what you want...not sure how to check that but I loaded the source with display disabled, hit F5 and let it run for a few minutes which shows fps on the info window. I tried a few parts and it seems to hit somewhere between 39-46fps.

As for the "other scenario" I assume you mean with ffdshow ? I have no idea how to do that.

Yes, you need 64-bit Avisynth to run the 64-bit executables.

is there even one ? don't think I've seen such on sourceforge and also most avisynth plugins are x86 so not sure how effective that is.

[Didée]

There are 64bit-builds of Avisynth out there, but none of them are "official". What they have in common is that they're not fully stable (though it's not quite sure if that's a problem of the Avisynth core, or of the same-unofficial x64-plugins, or both.) For most usual tasks, Avisynth x64 should be quite safe to use. Problems start to come up when you're using very "ressource hungry" scripts.

Regarding the speeds you're observing: generally, your 9600GT is not up to par with your CPU. The 9600GT uses an older VP engine (not sure ATM if VP2 or VP3). Glancing thoroughthe DGDecNV benchmark thread, it seems the 9600GT will decode high-bitrate stuff at 45~50 fps, give or take. Obviously, you cannot encode faster than you can decode.

Recent Nvidia cards use the VP4 engine, which is able to decode in the speed range of 100fps. (Clearly above 100 for moderate bitrates, maybe slightly below for high-bitrate.)
Note that the VP engine is an independent decoder chip on the card, and is shared between different models. A 50-bucks GT240 will decode at the same speed as a 500-bucks GTX580.

Still, those 29.xx fps seem a little low, given that you see 39~46 fps within DGIndexNV. But OTOH, don't underestime the effort needed for resizing - 1080p has lots of pixels. And it can make a big difference *exactly which* resizing filter you're using. Example without encoding, just screen rendering in VirtualDub:

BlankClip(width=1920,height=1080).LanczosResize(taps=8) ==> ~40 fps
BlankClip(width=1920,height=1080).LanczosResize(taps=2) ==> ~195 fps

You see, even when the decoding effort is zero (blankclip), an unsuited resizer can cut the speed down severely. (So much for "the variables are very few");)

[sandy_geser]

There are 64bit-builds of Avisynth out there, but none of them are "official". What they have in common is that they're not fully stable (though it's not quite sure if that's a problem of the Avisynth core, or of the same-unofficial x64-plugins, or both.) For most usual tasks, Avisynth x64 should be quite safe to use. Problems start to come up when you're using very "ressource hungry" scripts.

thanks for explaining. I'd have to try and see if I can find such unofficial builds for avsynth and the main plugins I use.

Regarding the speeds you're observing: generally, your 9600GT is not up to par with your CPU. The 9600GT uses an older VP engine (not sure ATM if VP2 or VP3). Glancing thoroughthe DGDecNV benchmark thread, it seems the 9600GT will decode high-bitrate stuff at 45~50 fps, give or take. Obviously, you cannot encode faster than you can decode.

Recent Nvidia cards use the VP4 engine, which is able to decode in the speed range of 100fps. (Clearly above 100 for moderate bitrates, maybe slightly below for high-bitrate.)
Note that the VP engine is an independent decoder chip on the card, and is shared between different models. A 50-bucks GT240 will decode at the same speed as a 500-bucks GTX580.

aha! as I mentioned I'm far from being an expert which is why I was asking. I was actually thinking of getting a newer card. GTS450 or GT240. so might just do that now.

Still, those 29.xx fps seem a little low, given that you see 39~46 fps within DGIndexNV. But OTOH, don't underestime the effort needed for resizing - 1080p has lots of pixels. And it can make a big difference *exactly which* resizing filter you're using. Example without encoding, just screen rendering in VirtualDub:

BlankClip(width=1920,height=1080).LanczosResize(taps=8) ==> ~40 fps
BlankClip(width=1920,height=1080).LanczosResize(taps=2) ==> ~195 fps

You see, even when the decoding effort is zero (blankclip), an unsuited resizer can cut the speed down severely. (So much for "the variables are very few");)

I usually use Spline36Resize for HD res transocodes (like 1080p bluray to 720p) and LanczosResize for SD res transcodes (like DVD). For both I use default parms. Don't ask me why I use those, guess someone down the line said I should...again, not an expert
Specifically for this test I used:

dgsource("test.dgi")
trim(10000,15000)
crop(0,22,-0,-22)
Spline36Resize(1280,688)

[admin]

Didée as usual has explained things well.

The only thing I might add is to point out that if you are choosing encoding parameters and resolution such that your encoding rate is above your GPU decoding rate (e.g., HD->SD transcode), then you likely will be bottlenecked by the decoder. In such a case, you want faster decoding, so do indeed upgrade your video card to a VP4+ engine.