I did some more tests with a noisy interlaced 720x576i videofile source, encoded with x264 CRF=17. The results may be specific for this type of source and my system, but here my findings:
- KNLMeansCL(0,2,3,3.0) is much the same as DGDenoise(strength=0.06) w.r.t. filesize, speed and visual quality
- adding --nr increases(!) the encoding speed slightly (1% …3%)
- adding --nr reduces the filesize significantly (by 30%) and improves visual quality without noticeable loss of details or sharpness (the same file size with the external filter alone would blur the picture / loose details more)
- --nr xxx seems to complement the external filters (KNLMeansCL or DGDenoise) nicely.
- Overall the combination of the external with the internal noise filter wins in every aspect.
These conclusion may be specific for this kind of (low quality) source though.
Thanks, Sharc. I suspect that the encoding is swamping the denoising here with respect to performance timing. I think it's better to test without encoding, in order to compare the denoising performance. Nevertheless, your report on the results of --nr is interesting and valuable.
Anyway, I just unrolled the loops in my kernel, increased the number of registers some more, and told ptxas to use fast math. This increased the frame rate for rat.avs from 110 fps to 184 fps (>60% improvement)! I'll test it a bit and then slipstream it.
So, the current scorecard for me is:
DGDenoise(): 184 fps
KNLMeansCL(d=0): 59 fps
DGDenoise is now 3 times faster than KNLMeansCL on my lowly 1050Ti. It will be fun to see gonca's results.
Might have to wait a while.
I just built this system a month ago
Side note
DGDenoise saturates my cpu during encodes, (it runs at 100%) so no bottleneck
When I OC I check CPU and general MotherBoards temps. I don't recall at actually ever looking at specific temps like NB.
You're running a 980 so your motherboard isn't that old. If it gets too hot it should throttle or reboot automatically, unless you disable that in BIOS.
What kind of temps are you referring too
I can touch the heat sink for 10 seconds, so I'm guessing 65-70 degC. I googled for "X58 northbridge hot" and it's not unusual. Spec allows up to 100 degC but that seems extreme. I may add a spot cooler, or maybe the case fan will be enough. This is a no-name Chinese LGA1366 X58 mobo. Works great though.
I still have a perfectly good Core2 Duo Win10 machine in the closet. I think there's a GTX 720 in there. Can't bring myself to decommission it. One day...off to CPU heaven.
What do you do with all your older systems, gonca? Salvation Army?
My new system replaced my old i7-3930k system and I am about to replace my i5 LG1151 HTPC.
I am going to re purpose parts to build a server and then I might put the remaining parts together and give it away to somebody as is.
Or I might salvage some of the parts to start a second server (can never have too many)
Normally I just give away my old machines, storing them in the outdoor shed isn't a good idea (as I found out when I was younger)
The shed had no heat and wasn't rain proof
Put a computer in there for temporary storage, and when I went to get it, well it was a little rusty, shall we say.
Amazing how many metallic components are inside a computer
By the way, Power Supplies don't work too good after having ice inside of the unit
Yeah, I learned my lesson.
If I want a new system I just store the old spare outside, give it a couple of bad weather days (I am in Toronto, Canada) and there you go.
Darn it, old one is not really usable anymore
admin wrote:The unexpected speed is making me think about porting other filters to CUDA. MV tools?
QTGMC on CUDA would be phantastic
That's why I mentioned mvtools.
I am just about to re-encode all my tapes thanks to the speed improvement with DGDenoise.
Rest assured I will test and probably redo everything once MVtools/QTGMC will be ported to CUDA.
You really deserve a couple of free beer!
QTGMC parameters:
- QTGMC() #default, or QTGMC(InputType=1) for cleaning up progressive sources
- QTGMC(preset="medium") #typical
- QTGMC(preset ="Slow",TR2=2,Sharpness=0.8,EZDenoise=1.0,NoisePreset="Slow",ShowNoise=false,ShowSettings=false,InputType=0) #in few cases, mainly for testing
What kind of beer are we talking about? I have an open mind but I don't like swill...think Budweiser as the epitome of swill. It's easy to make things cold and bubbly, we need character.