Beiträge von Vouk

Yes, please use i.e. YUV 4:2:0 10 bit.

Kleinrotti Seems the high core count works pretty good with AVX2 code.

@All Other experiences?

Please report if ...

• It works
• It crashes
• it is faster than the version before
• it is slower than before
• ...

At some point we have use the latest SDK. At least when FFmpeg doesn't allow us to use older SDKs anymore. It is just a question time, I guess.

In the Voukoder successor I plan to have the FFmpeg binaries as DLL. They would be exchangeable. But I don't want to hack this in the existing voukoder.

I need to perform one FMA operation per sub-pixel. This can be parallelized with CUDA.

I'm not sure how I can archive this using tensor yet.

NOT FOR PRODUCTION USE

Test release: https://github.com/Vouk/voukoder-….0.msi?raw=true

(Use the latest NVIDIA driver version)

These presets are available when compiling FFmpeg with the NVIDIA SDK 10.0.

Voukoder currently uses SDK 9.0 to support even older GPUs or mobile GPUs (Bring back NVENC support for mobile Kepler-series GPUs to Voukoder)

This would mean to drop support for older GPUs.

Exporting a project with 2048 x 1152 and float32 color precision (per channel).

1 pixel per CUDA thread processing 4 floats per FMA in a 32 x 32 block

Unfortunately the cudaMemCopy's are like 10-15ms each. Processing more frames per call would speed this up even more.

System One

AVX2 (i7-7700K)

[11:26:35] Frame #277: vRender: 39 us, vProcess: 41738 us, vEncoding: 19984 us, aRender: 70 us, aEncoding: 268 us, Latency: 64781 us

[11:26:35] Frame #278: vRender: 37 us, vProcess: 44431 us, vEncoding: 19885 us, aRender: 61 us, aEncoding: 12 us, Latency: 66833 us

[11:26:35] Frame #279: vRender: 33 us, vProcess: 41376 us, vEncoding: 18816 us, aRender: 62 us, aEncoding: 310 us, Latency: 62777 us

[11:26:35] Frame #280: vRender: 39 us, vProcess: 43909 us, vEncoding: 18867 us, aRender: 55 us, aEncoding: 218 us, Latency: 65696 us

[11:26:35] Frame #281: vRender: 35 us, vProcess: 43756 us, vEncoding: 20499 us, aRender: 53 us, aEncoding: 231 us, Latency: 66800 us

[11:26:35] Frame #282: vRender: 31 us, vProcess: 43390 us, vEncoding: 20808 us, aRender: 65 us, aEncoding: 315 us, Latency: 66789 us

CUDA (Quadro P2000)

[09:54:33] Frame #288: vRender: 40 us, vProcess: 18746 us, vEncoding: 22134 us, aRender: 68 us, aEncoding: 19 us, Latency: 44891 us

[09:54:33] Frame #289: vRender: 36 us, vProcess: 21211 us, vEncoding: 18160 us, aRender: 72 us, aEncoding: 322 us, Latency: 42256 us

[09:54:34] Frame #290: vRender: 36 us, vProcess: 18531 us, vEncoding: 20253 us, aRender: 59 us, aEncoding: 214 us, Latency: 41408 us

[09:54:34] Frame #291: vRender: 35 us, vProcess: 18369 us, vEncoding: 22336 us, aRender: 70 us, aEncoding: 327 us, Latency: 43288 us

[09:54:34] Frame #292: vRender: 40 us, vProcess: 17668 us, vEncoding: 18668 us, aRender: 63 us, aEncoding: 17 us, Latency: 38560 us

[09:54:34] Frame #293: vRender: 36 us, vProcess: 17704 us, vEncoding: 19705 us, aRender: 71 us, aEncoding: 327 us, Latency: 40145 us

System Two

AVX2 (i7-8700K)

[11:30:00] Frame #70: vRender: 31 us, vProcess: 36414 us, vEncoding: 16255 us, aRender: 1083 us, aEncoding: 10 us, Latency: 55197 us

[11:30:00] Frame #71: vRender: 83 us, vProcess: 40397 us, vEncoding: 15759 us, aRender: 577 us, aEncoding: 244 us, Latency: 59374 us

[11:30:00] Frame #72: vRender: 30 us, vProcess: 36319 us, vEncoding: 15735 us, aRender: 930 us, aEncoding: 357 us, Latency: 54855 us

[11:30:00] Frame #73: vRender: 1774 us, vProcess: 47668 us, vEncoding: 70102 us, aRender: 13 us, aEncoding: 245 us, Latency: 121703 us

[11:30:00] Frame #74: vRender: 34 us, vProcess: 40626 us, vEncoding: 15824 us, aRender: 610 us, aEncoding: 8 us, Latency: 58531 us

[11:30:00] Frame #75: vRender: 35 us, vProcess: 40386 us, vEncoding: 15860 us, aRender: 565 us, aEncoding: 234 us, Latency: 58775 us

CUDA (GeForce RTX 2080 TI)

[09:28:47] Frame #1720: vRender: 30 us, vProcess: 12659 us, vEncoding: 13886 us, aRender: 893 us, aEncoding: 332 us, Latency: 29435 us

[09:28:47] Frame #1721: vRender: 38 us, vProcess: 13909 us, vEncoding: 17943 us, aRender: 894 us, aEncoding: 421 us, Latency: 35310 us

[09:28:47] Frame #1722: vRender: 39 us, vProcess: 13063 us, vEncoding: 14418 us, aRender: 558 us, aEncoding: 8 us, Latency: 30184 us

[09:28:47] Frame #1723: vRender: 32 us, vProcess: 13319 us, vEncoding: 14304 us, aRender: 12 us, aEncoding: 343 us, Latency: 29725 us

[09:28:47] Frame #1724: vRender: 51 us, vProcess: 14712 us, vEncoding: 15048 us, aRender: 653 us, aEncoding: 244 us, Latency: 33087 us

[09:28:47] Frame #1725: vRender: 30 us, vProcess: 13147 us, vEncoding: 15400 us, aRender: 570 us, aEncoding: 7 us, Latency: 30813 us

Playing with CUDA right now and already got a speed increase of 69% when using bit depths greater than 8 bit on my development machine.

Still working on it.

The DirectX way doesn't work because there is no texture support of YUVA floats.

Trying CUDA ...

That might take some time.

I have in mind to create a directx texture of the frame and convert it using the gpu, yes. But this is still to be done, and i have it planned for the voukoder successor.

Voukoders processing path for > 8 bit is not optimized yet. Premiere delivers floating point data for high bit depth video. This needs to be converted to a pixel format that FFmpeg understands.

It should be possible to use Voukoder as a VfW codec. But video and audio will be exported separately to two files. You can't have video and audio muxed to one file.

So i guess we'll still need connectors.

This doesn't seem to be voukoders fault.

Zitat

"The request cannot be completed because you have exceeded your quota."

Voukoder has finished doing its work at this point already.

There is most likely an issues with either premiere or youtube.

Just noticed this topic is about importing and not exporting. Voukoder is for exporting only. It is not an importer.

You still have the ability to load the FDK AAC encoder. After all ... if you want to have high quality audio, don't even consider using AAC or MP3 or any other lossy codec. Use either FLAC, ALAC or uncompressed audio.

I will look into this and check if a file export SDK is available.

Das sieht soweit okay aus. Dennoch vermute ich das Problem auf Premiere Seite. Was du noch probieren könntest ist alle Voukoder Presets zu löschen, und wirklich dann ohne Preset nochmal zu exportieren.

You both are basically on the same side.

It is all about finding the best compromise of quality, small file size and fast encoding time. Unfortunately this depends also on your source material and your encoding hardware.

That's why x264 has dozens (x265 has about 200) options to configure. There are no universal "best" settings. As metioned the nvenc encoders work differently with different gpu generations (b-frames). Unfortunately users keep asking for the best settings over and over again.

It is the same with encoder comparisons. The internet is full with stupid comparisons like "Is Voukoder better than <other encoder>?" What does better mean? Faster? Higher quality? With what encoder and what option? There a whole technical essays from professors about it.

But we can find a compromise of settings for getting users a reasonable and satisfactory result, yes. That's what MyPOV did here and that's what iAvoe did with supplying the encoder presets.

By the way: Unfortunately you have no control what quality you get from YouTube. It always transcodes your uploaded video. i.e. it is a huge difference if they transcoded your video to AVC (=h264) or VP9. AVC always has a worse quality and more artifacts. They also do a bitrate based transcode and you'll get the highest bitrate only if the frame size is >= 3200x1800 and the frame rate is >= 48 fps. That's why people asked for voukoder to have a spline based upscaler.

Unfortunately I can't test this with elements. My trial period has expired.