Login at Kodi Home

**fritsch** · 2014-12-21, 10:25

High CPU usage is not true. We need just < 4ms to copy the surface on average. It's 10% more load for 1080p24 material.

So for 24fps your CPU will still idle 90% the time of a second.

Matt Devo · (This post was last modified: 2014-12-21, 11:11 by Matt Devo.)

(2014-12-21, 10:15)Sunflux Wrote: Yeah... seems I can either get smooth graduations but high CPU usage, or good CPU usage and banding. When OE RC3 comes out I'll trying VAAPI rendering again, see if I can live with the banding for the CPU usage improvement (I can't quite run 4k videos smoothly with the "Use SW Filter" option enabled right now - I get like 27fps/30fps).

then you've got something else going on, as I can playback 60Mbps 2160p30 h.264 content with <20% CPU and no dropped/skipped frames - with or without the 'Use SW filter' option set

**fritsch** · 2014-12-21, 11:14

For > 1920x1088 we disable the SWFilter. So it's not used for 4K. See here: https://github.com/xbmc/xbmc/blob/master....cpp#L2957

Sunflux · (This post was last modified: 2014-12-21, 11:33 by Sunflux.)

Playing back a video with these properties:

Code:

General

Format                                   : MPEG-4

Format profile                           : Base Media

Codec ID                                 : isom

File size                                : 6.83 GiB

Duration                                 : 38mn 53s

Overall bit rate mode                    : Variable

Overall bit rate                         : 25.1 Mbps

Writing application                      : Lavf52.34.0



Video

ID                                       : 1

Format                                   : AVC

Format/Info                              : Advanced Video Codec

Format profile                           : [email protected]

Format settings, CABAC                   : Yes

Format settings, ReFrames                : 3 frames

Format settings, GOP                     : M=4, N=90

Codec ID                                 : avc1

Codec ID/Info                            : Advanced Video Coding

Duration                                 : 38mn 53s

Bit rate mode                            : Variable

Bit rate                                 : 25.0 Mbps

Maximum bit rate                         : 28.0 Mbps

Width                                    : 3 840 pixels

Height                                   : 2 160 pixels

Display aspect ratio                     : 16:9

Frame rate mode                          : Constant

Frame rate                               : 29.970 fps

Standard                                 : NTSC

Color space                              : YUV

Chroma subsampling                       : 4:2:0

Bit depth                                : 8 bits

Scan type                                : Progressive

Bits/(Pixel*Frame)                       : 0.101

Stream size                              : 6.79 GiB (99%)



Audio

ID                                       : 2

Format                                   : AAC

Format/Info                              : Advanced Audio Codec

Format profile                           : LC

Codec ID                                 : 40

Duration                                 : 38mn 53s

Bit rate mode                            : Variable

Bit rate                                 : 126 Kbps

Channel(s)                               : 2 channels

Channel positions                        : Front: L R

Sampling rate                            : 44.1 KHz

Compression mode                         : Lossy

Stream size                              : 34.9 MiB (0%)

I get low CPU usage all around - like 10-15% - but the SKIP count keeps updating at a rate of about 3 per second, and the framerate shown is around ~27fps. If I change the rescaling algorithm from Auto to one of the HQ ones, the framerate drops to like 14fps (but still with the low CPU usage). Which seems odd to me - why is it running it through the scaler in the first place?

I've tried setting the "use HQ scalers above" setting from 0% to 20% without change... changed view mode from Normal to Original without change...

**fritsch** · (This post was last modified: 2014-12-21, 11:40 by fritsch.)

It needs to downscale! from 4K to 1080p display. This costs too much when doing "Lanczos3 Optimized" which the Auto does in that case, as the target dimension are only 1/4 the original.

I think 24p with that very hight bitrate (you have 28Mbit/s there) is most likely the limit of the celeron HSW GT1. If you share that sample I can test it on my hardware.

What I tested succesfully was: http://distribution.bbb3d.renderfarming....normal.mp4 (bilinear scaling of course).
What all of those cannot do is: http://distribution.bbb3d.renderfarming....normal.mp4

Sunflux · (This post was last modified: 2014-12-21, 11:40 by Sunflux.)

No no, this is happening when I'm running at [email protected] to a 4k TV.

Also, I'm pretty sure "Auto" is selecting one of the LQ scalers, because when I manually select Lanczos3 Optimized (or the Spline one) is when the framerate plummets

**fritsch** · 2014-12-21, 11:41

Makes no sense until I see a Debug Log :-) Keep the value set to 20% while doing so.

and provide: /usr/lib/kodi/kodi-xrandr | pastebinit

If you don't run kodi you have the very lame old acceleration pipeline.

Sunflux · (This post was last modified: 2014-12-21, 11:46 by Sunflux.)

No, I'm not on Kodi yet. Still on the latest "stable" OpenElec - was waiting for RC3 to upgrade. Think that's the problem here?

Also, could it in any way relate to the PAR bug/issue I discovered?

http://forum.kodi.tv/showthread.php?tid=210815

I had assumed allowing the 1% error would prevent it from scaling entirely, but perhaps because of this it's still scaling from 100% to 100%? At 3840x2160, the calculation error grows to 2 pixels in width.

nolight · 2014-12-21, 11:58

Hi, I recently installed dual boot Chrome OS/OpenELEC on Asus Chromebox with OpenELEC as default OS.

On my Sharp LCD TV LC-32M400M-BK (native resolution 1366x768) it is able to display the OpenELEC on boot up. Connection is direct HDMI.

On my Panasonic plasma TV TH42PV70H (native resolution 1024x768) it is showing up as if no signal. Tried both direct HDMI connection and HDMI connection through chromebox -> AVR -> TV.

Pls help on how I can fix the no display on Panasonic TV. Thanks.

**fritsch** · 2014-12-21, 12:06

(2014-12-21, 11:45)Sunflux Wrote: No, I'm not on Kodi yet. Still on the latest "stable" OpenElec - was waiting for RC3 to upgrade. Think that's the problem here?

Also, could it in any way relate to the PAR bug/issue I discovered?

http://forum.kodi.tv/showthread.php?tid=210815

I had assumed allowing the 1% error would prevent it from scaling entirely, but perhaps because of this it's still scaling from 100% to 100%? At 3840x2160, the calculation error grows to 2 pixels in width.

This is what happens in code:

Code:
float scaleX = fabs(((float)m_sourceWidth - m_destRect.Width())/m_sourceWidth)*100;

    float scaleY = fabs(((float)m_sourceHeight - m_destRect.Height())/m_sourceHeight)*100;

so: 720p to 1080p

abs(720 - 1080) / 720 = 0.5 * 100 = 50
abs(1280 - 1920) / 1280 = 0.5 * 100 = 50

So with a setting of 20% the hq scaler would be chosen

2160p to 1080p

abs(2160-1080) / 2160 = 0.5 * 100 = 50
abs(3840-1920) / 3840 = 50

So Lanczos3 Optimized would also be chosen! when you have set it to 20%.

If you would set it to 60% bilinear would be chosen, but this is not wanted for the 720p case.

I could weight the downscaling code with a specific factor?

In your case with 3840 to 3840 and 2160 to 2160 - a simple setting 10% should be enought to not use lanczos 3

Sunflux · 2014-12-21, 13:01

Okay, I upgraded to RC2 and now the troublesome video plays back just fine in 4k - wasn't aware there would be such a performance difference! Trying to select any of the HQ scalers still results in a horrible framerate, but setting the threshold to 10% as you suggested leaves Auto working fine.

Downscaling the same 4k video to 1080p loses about 1 frame every 2 seconds with any of the HQ scalers, and with or without VAAPI (well, maybe a tiny bit better with VAAPI), but is rock solid on Auto/Bilinear.

Incidentally, any idea why in RC2 I now have the text "There are no settings available for this visualization" behind my Video Settings screen (which is otherwise normal?) Also, how the left-side labels for the "Open Elec" settings menu are all pushed all the way left, with only like a half character visible...

**fritsch** · 2014-12-21, 13:20

Cool.

Remember to install a build that has the GPU hang fixed. I think Matt has documented that in the first post.

Sunflux · (This post was last modified: 2014-12-21, 14:05 by Sunflux.)

Sorry, UI error was the result of me having customized Confluence (and thus requiring me to update the copy and re-apply my changes).

Don't see any reference to a fixed build in the first post or Wiki, but I can probably wait until RC3. The issue is still only when "prefer VAAPI" is enabled, correct?

However, the PAR error I referenced is really messing up video. With the 10% HQ scaler setting, my default preference of Lanczos3 now changes to Nearest Neighbor for 1080p played at 1080p. But due to the PAR issue, it's still rescaling the video at all times (it thinks 1920 wide should be played at 1919, then the 1% aspect error allowance gets it back to 1920).

But even still, and SPECIFICALLY on videos that are not the full 1920x1080, but rather cropped vertically (for example, 1920x800), Nearest Neighbor is causing a noticeable doubling of vertical lines about every 100 pixels or so, despite the fact that it's ultimately playing the file back at its native resolution. Switching to Bilinear results in noticeable softening of the video compared to Nearest Neighbor. Removing the 10% HQ scaler setting lets me use Lanczos3 which looks better, but clearly it's still messing around with the video and isn't giving me 1:1 reproduction.

**fritsch** · (This post was last modified: 2014-12-21, 14:10 by fritsch.)

The solution to this problem is: "Allow errors in ratio to minimise black bars", Choose: 1 or 2% there (Video -> Playback).

I already fixed that in xbmc's master, it's a rounding issue after all: https://github.com/xbmc/xbmc/pull/5922

Sunflux · 2014-12-21, 14:13

I have done that... but for some reason on any file that's less than 1080 pixels high, I'm getting visible resampling artifacts that I can watch come and go as I change resampling algorithms. Note that I do not see the same effect on any video that is the full 1920x1080 frame, just those that are letterboxed.