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.
Alright, neither do I.
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.
Alright, neither do I.
The DirectX way doesn't work because there is no texture support of YUVA floats.
Trying CUDA ...
You may also want to try Tensor. If it IS possible to leverage Tensor in this use case, Tensor should be WAY faster than Cuda.
If you make any good progress I'd love to know.
The DirectX way doesn't work because there is no texture support of YUVA floats.
Trying CUDA ...
But this is still to be done, and i have it planned for the voukoder successor.
Voukoder successor? When's that happening?
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.
Is this the kind of thing a GPU could process?
This is pretty normal when using 10 Bit instead of 8 Bit. Encoding takes a lot longer then.
But I'm not being limited by my hardware, it just isn't using it. The encoder in question can already encode a 4k60 10 bit video in real time with no frame drops, while playing a game, it's definitely not struggling in Premiere.
I'm a bit hesitant to call this a bug, maybe it is though.
With Identical settings, NVENC HEVC only uses about 15% of my GPU's video encode when 10 bit is selected, NVENC HEVC uses 90% of my GPU's video encode when 8 bit is selected. In 8 bit I get a real time export, while 10 bit takes 4-5 times longer for what seems to be no reason. Since I kind of need 10 bit for my export, this is a pretty big setback for the longer videos I render, since 30 minutes becomes 2 and a half hours.
Edit: I'm aware this could be Premiere being slow with handing frames to Voukoder.
I completely forgot re-encoding does lose quality each encode, total brainfart. For now it saves me an hour, so I'm not too torn up about it.
Well that's true but you'll lose a lot more quality that way.
The files are recorded USING NVENC in the first place.
A week later and I suddenly had the most incredible idea, use NVENC to do the FFmpeg part. It is WAY faster, like 5 minutes for 6 minutes of video.
Unfortunately there is no way to ask Premiere to send the raw frames in a certain color space.
Most pixel formats are either:
- bt709
- non-bt709
Hm okay.
check you media player, it might have poorly designed decoder, or added filters
You can try use a different media player like mpv which is now the best option unlike potplayer's broken color schemes
Not the issue, I've already completely solved my problems. This issue is on Adobe, but may be solvable by Vouk.
Make sure you are using the correct video levels (limited/full) and bt709 for hd video or bt601 for sd.
That wasn't the issue. Had to do with the way Premiere ingests content it recognizes as HDR. Premiere converts it, then I have to convert it with a preset, then export to Voukoder where it gets converted again. At some point the colors shift.
Solutions are use FFmpeg to make a new video that Premiere reads as SDR, or use the built in HEVC encoder.
Alternatively, it may be possible to bypass this issue if you can add Bt2020 to the video tab in the Export menu of Premiere. I have no clue if it would work, or if it's even possible to implement that.
I meant the Inferno would act as the second monitor.
Oh, that may work depending on audio.
DLSS 2.0 and NVENC 2.0 made my 2070 Super a 1080 Ti on roids, and it was definitely worth it. I'm really excited to see more programs adopt DLSS and Tensor Core support. Because 1080p upscaled to 2160p is basically identical with a 70%ish boost to fps.
Yeah, I don't have 2 monitors yet. And that's a big upgrade from a 970.
Anyways, thanks for keeping this thread alive, it's really helpful. Maybe when I actually start making content I'll make a video form of this guide.
Note that your recordings from NVIDIA are probably 4:2:0, as are youtube, and I don't think youtube goes over 60fps
Yeah they're 4:2:0. I don't intend to record 144hz, but the port on the Ninja Inferno would have to be capable of taking that signal.
I honestly think waiting for my hardware upgrade and then either using Premiere's HEVC encoder or the FFmpeg method will be the best solution, at least until HDMI 2.1 comes to capture cards. I'll have to test render times both ways though.
As another option, if you can afford it, the Ninja Inferno is what I use to capture HDR footage natively to ProRes via HDMI. In order to get a format Voukoder can work with properly, I just click the "rec709" button in the recorder to let it format the input signal as rec709 without modifying the input colors at all. It'll look all washed out on the monitor while recording, like it does in premiere when you import that format, but it'll work great with Voukoder!
Does the Ninja Inferno require a capture PC to use? Even if it does, it's gonna need HDMI 2.1 support, because I'll be running games at 4k 144hz 10 bit 4:4:4.
95-100% utilization on the CPU for an hour. I am working from a hard drive, in the future when I have more SSD space, I may end up temporarily moving relevant videos to an SSD for the editing and exporting process.
Not sure about how the Ninja Inferno works out for me. But regardless I'm getting a 4950x/5950x (depending on what AMD calls it) and a 3090 soon, so I won't have to resort to only using Voukoder NVENC for convenience.
If there's one thing I've noticed in my time preparing to make 100% HDR gaming content on my channel, it's that 99% of resources available are complex and poorly built. Premiere suffers from a lack of GPU acceleration while editing and specifically exporting HDR, and issues with Rec709 and Bt2020. FFmpeg uses Command Prompt which can be confusing for people who don't speak command line. The other editing softwares have pros and cons as well, but are a little worse for editing. Point is, everyone falls a little short, and it's annoying.
It appears to have worked, uploading to Youtube to compare side by side as i'm posting this.
Although this sort of defeats the purpose of why I personally use Voukoder (NVENC HDR). It takes my 3600 2 hours to render my videos with the normal Premiere HEVC encoder, 20 minutes for NVENC HEVC with Voukoder. This method took 1 hour to finish in FFmpeg and another 20 in Premiere, so while it IS faster I still have to choke slam my poor CPU to do it.
Thanks for the help on this, thanks for the incredible post, but it looks like i'll be waiting for a CPU upgrade!