Beiträge von Kleinrotti

Can you upload a logfile?

Currently 2pass option for VP9 and VP8 is missing.

I found it. The problem is CUDA enabled in the Premiere Project settings. If I disable it (setting to software only), Voukoder R2 is exactly same as fast as Voukoder 1.2.0 (around 370 fps). Voukoder 1.2.0 doesn't care the CUDA setting in Premiere but in Voukoder R2 for some reason if its enabled the export is much slower.

Latencies are also normal with CUDA disabled:

[20:04:42] Video frame #574: Render: 38 µs, Process: 8 µs, Encoding: 2456 µs
[20:04:42] Video frame #575: Render: 77 µs, Process: 19 µs, Encoding: 2560 µs
[20:04:42] Video frame #576: Render: 79 µs, Process: 19 µs, Encoding: 2813 µs
[20:04:42] Video frame #577: Render: 47 µs, Process: 15 µs, Encoding: 2729 µs
[20:04:42] Video frame #578: Render: 84 µs, Process: 16 µs, Encoding: 2063 µs
[20:04:42] Video frame #579: Render: 58 µs, Process: 13 µs, Encoding: 1721 µs

But in CS6 this didn't work. In my case this doesn't matter because I am not using CS6.

Zitat von Vouk

There are some issues with your system, yes. I have no clue why it is like that. It could be many things:

Maybe ...

• with AMD CPUs at all
• with Threadripper CPUs
• with high core counts
• with specific Premiere settings? Hardware decoding?
• ...

I did not get any hardware donations and I can't afford buying any test equipment. So I can just guess here.

But as you have a working Visual Studio IDE and the voukoder sources I'd recommend that you try to find the issue. Try to find that slow spot.

Try to profile it.

But the question is why Voukoder 1.2.0 is that much faster?

I am debugging right now, maybe I can find something.

What I can currently say is that the ffmpeg version doesn't matter, I tested it with both 4.0 and 4.1.

Voukoder 1.2.0 is also slower in Premiere CS6 (down to also 200 fps)

On my side CS6 is even slower than CC 2018 (around 170 fps but lower latencies).

Here are two screenshot from the Taskmanager (CC 2018 usage and CS6 usage) and the logfile which shows the latencies.

CS6 uses way more RAM than CC 2018 but the CPU and GPU usages are pretty low.

Zitat von Vouk

Can you try Premiere CS6? I'm using it. CC 2019 seems to be slower.

I am using CC 2018, but I can install CS6. Testing it in a few minutes.

Zitat von Vouk

Just to clarify:

Render: 15 µs <- Voukoder has no impact on this, it is entirely premiere

Process: 5 µs <- Voukoder has high impact on this

Encoding: 676 µs <- Voukoder has small impact on this. Mostly libav / FFmpeg

Can you look at my logfile above and tell me if the latencies are normal?

I am changing and debugging some things in the source code right now. Maybe I can find something.

Zitat von Vouk

Rendering is done on the CPU mostly. It also depends on your source pixel format and all effects and filters you have applied.

I have no filters and no effects in my test.

I also have no effects or something else applied. And I think a TR 1950X should have enough power.

As I can see your Render time is about 15 µs average.

Is it normal that this value goes sometimes up to over 5000 µs ?

You can see it in the logfile.

My Gtx 1080ti is reaching with your settings (gpu=0 preset=slow qp=15 rc=constqp) about 200 fps. I don`t think that a 2080ti is more than twice as fast. (my 1080p video has no effects or filters or something else and is located on a SSD raid 0)

I tested I a few more things out and there is some other curious thing. If I change the preset value from nvenc to Slow or medium or fast there is a performance difference in Voukoder 1.2.0 (as it should be).

Slow = 275 fps average

Medium = 285 fps average

Fast = 300 fps average

But in Voukoder R2 all presets have pretty much the same encoding speed and fps values (it never goes over around 200 fps). There has to be something that is slowing down the encoding in voukoder r2.

If you are exporting a video with 2pass enabled in Premiere and Media Encoder at the same time the export will crash when the second pass is starting. I think this appears because of the mb-tree and passlogfile in the temp folder. Both programs are using the same file when exporting.

I think a solution could be that Voukoder checks if the two files already exists and creates the two files with another name like "voukoder-passlogfile-1" and "voukoder-passlogfile-1.mbtree".

Currently x265 2pass option is missing in the gui. Would be nice if it could be implemented.

The R2 logfile looks also pretty strange. I think the log is not complete and missing entries.

Here is are my fps values with 1.2.0 and R2 beta4. Hugh difference. And I attached the R2 log.

I am pretty sure that I am using the same video export settings but also by changing some parameters, R2 is every time much slower.

Here is a recorded video. I had reset Voukoder R2 before.

Both are release builds downloaded from github releases.

Some strange think is also why R2 is using less video engine power but more video copy.

You are using a RTX card right Vouk ? Maybe its only affecting Gtx cards because of a new nvenc version or something?

Settings which are set for fdk aac are still resetting. r2 beta4

I compared the two files from version 1.2.0 and r2 and I can say that the quality of version r2 is a little bit better in rainy scenes than 1.2.0 but really hard to see . So there has to be some difference in the settings but for this loss of speed?

There is also a wrong default value in voukoder r2 with nvenc h264 I noticed. In the GUI ref frames have a default value of 3 but the exported video has 1 ref frame.

1.2.0 settings = h264_nvenc with options: b=2000000,bf=2,maxrate=3000000,preset=slow,profile=high,rc=vbr,refs=3

R2 settings = vcodec: h264_nvenc, voptions: _computeCapability=6.1,_pixelFormat=yuv420p,b=2000000,gpu=0,maxrate=3000000,preset=slow,profile=high,rc=vbr, acodec: libfdk_aac, aoptions: _sampleFormat=s16,profile=aac_low,vbr=3, format: mp4, faststart: 0

And I noticed that Voukoder R2 log is flooded with this:

Calling dl_fn->cuda_dl->cuCtxPopCurrent(&dummy)
                                       
 [2019-04-12T17:31:22] Calling dl_fn->cuda_dl->cuCtxPushCurrent(ctx->cu_context)

I have tested the speed differences between Voukoder 1.2.0 and Voukoder R2 (beta3 and beta4). And I can say that Voukoder R2 is very slow compared to Voukoder 1.2.0 with nvenc. I exported a 50 min video with the same nvenc settings in Voukoder 1.2.0 and Voukoder R2. In version 1.2.0 it takes about 5 minutes, in Voukoder R2 (beta3 and beta4) it takes about 9 minutes. I attached two screenshots from the taskmanager which shows that version 1.2.0 uses more Video engine power I think.