Port of NNEDI under new v2.6 AVS API

[TheFluff]

Originally Posted by Aktan

I think it is still related to Intel compiler.

What do you base that on? Most of the NNEDI3 performance-critical code is in hand-written assembler in this version. In the Vapoursynth version (which uses mostly compiler intrinsics in C++ rather than pure asm) you could argue that the compiler could have a big impact, but not in this version. In fact it's been repeatedly pointed out in this very thread that having a special DLL with compiler optimizations for AVX2 enabled is completely pointless - it has essentially no performance impact at all, since there's nothing performance critical that the compiler can actually optimize.

[Aktan]

Originally Posted by TheFluff

What do you base that on? Most of the NNEDI3 performance-critical code is in hand-written assembler in this version. In the Vapoursynth version (which uses mostly compiler intrinsics in C++ rather than pure asm) you could argue that the compiler could have a big impact, but not in this version. In fact it's been repeatedly pointed out in this very thread that having a special DLL with compiler optimizations for AVX2 enabled is completely pointless - it has essentially no performance impact at all, since there's nothing performance critical that the compiler can actually optimize.

I will admit it was just a bad guess as I'm not that familiar with compilers. I guess I won't be able to compile it then since there was hand written ASM. I based it on the fact that in the past Intel did selectively optimize to only Intel chips in their compiler.

Edit: Just curious, if I use the Release_Intel_W7_Core2_AVX DLL or even the Release_Intel_W7_Core2_SSE4.2 DLL and set opt to 5, would AVX2 still be used? I tried that and I still get the same slow speed. I was expecting it to throw an error like invalid opt mode, but I was just guessing.

[TheFluff]

What happens if you use threads=1?

BTW, if you're willing to try Vapoursynth, you can try znedi3, an attempt at making NNEDI3 on the CPU competitive with the OpenCL version running on the GPU. IIRC it was something like 50-100% faster than the original VS NNEDI3 on 8-bit input, but I don't know if anyone ever tested it on Ryzen.

[Aktan]

Originally Posted by TheFluff

What happens if you use threads=1?

BTW, if you're willing to try Vapoursynth, you can try znedi3, an attempt at making NNEDI3 on the CPU competitive with the OpenCL version running on the GPU. IIRC it was something like 50-100% faster than the original VS NNEDI3 on 8-bit input.

I'll try both in a bit.

Edit: threads=1:

opt	FPS
4	~4
5	0.11 (~1 frame every 9 seconds)

[pinterf]

It's not Ryzen. This one:

nnedi3_rpow2(rfactor=4, cshift="lanczosresize", fwidth=2880, fheight=1920, ep0=2, nsize=0, nns=4, qual=2, pscrn=4, opt=5)

is giving me 0.03 fps for the first 3-4 frames with opt=0 or opt=5
and >3 fps with opt=4

x64 version freshly recompiled with VS2017 15.5
Intel i7-7700 (AVX2), Win10 x64, Avisynth+ 2591 (dev)

[Aktan]

That's interesting, I guess now to figure out which of the param is actually causing it.

[jpsdr]

Originally Posted by Aktan

Edit: Just curious, if I use the Release_Intel_W7_Core2_AVX DLL or even the Release_Intel_W7_Core2_SSE4.2 DLL and set opt to 5, would AVX2 still be used?

No differences in the code, just compiler options, all versions will behave the same.

[jpsdr]

Odd... I don't remember so much differences when testing... But on my tests, i've never changed ep0 and always tested pscrn with 2 or 0.
Btw, with 0, very small frames (480p) to have some frames without having to wait hours...
Don't have time for now, but will redo tests later.

[pinterf]

Missing else

https://github.com/jpsdr/NNEDI3/blob...edi3.cpp#L1725

https://github.com/jpsdr/NNEDI3/blob...edi3.cpp#L1734

fps is OK again for AVX2

(And a request: the AVX2 asm files were missing from the sln file had to add them manually)

[jpsdr]

I was about to said that i'll check if i didn't messed-up something, but you beats me...

About ASM files :
Yes it's normal they are missing, because i want to keep the project on github on VS2010 (if you want to use more than VS2010, you can just upgrade the project), but VS2010 will not be able to to compile the AVX2 asm files, so it will result on a incorrect project if they were included. This is why the AVX2 asm are on others files.

[Aktan]

Awesome catch! Yep, changing prescreen to original (1) where the bug doesn't exist gave me results comparable to AVX.

[jpsdr]

New version, see first post, and i've also added on it a part about the multi-threading.

[jpsdr]

There is issue with the Intel versions.

I'll update the release files on github, removing the Intel versions, and keeping only VS version, and adding an VS AVX2 version. Wait at least 24h to check/re-download the files.

[jpsdr]

Trashed Intel version, file updated, redownload it.

[jpsdr]

New version, see first post, updated also the Multi-treading text part.

[mp3dom]

It seems there's a bug in YUY2 colorspace that create a green vertical bar on the right side of the image under some circumstances (both x86 and x64):

colorbars(width=1416,height=1080,pixel_type="yuy2")
nnedi3(0)

[Taurus]

Originally Posted by mp3dom

It seems there's a bug in YUY2 colorspace that create a green vertical bar on the right side of the image under some circumstances (both x86 and x64):

colorbars(width=1416,height=1080,pixel_type="yuy2")
nnedi3(0)

But only if the width is at some "odd" resolutions...not devidable with mod16.
Older versions show the same behaviour.
Just a short test..

[jpsdr]

Thanks for reporting, i've found what was wrong.
I want to finish something else i'm working on before making new releases of some filters. So, several days before i'll make new builds, but it's fixed on Github.

[mp3dom]

Originally Posted by jpsdr

So, several days before i'll make new builds, but it's fixed on Github.

Thanks

[jpsdr]

New version, see first post.