Topic: JRVR Performance on NUC12 Pro (Read 1558 times)

Hendrik · « **on:** July 04, 2023, 03:05:57 am »

How badly does the NUC12 fare? I would hope without tricubic interpolation it might work.

jmone · « **Reply #1 on:** July 04, 2023, 03:35:58 am »

Good idea! I'll give turning that off a go and report back (but they are watching TV at present!). I just did a quick test with UHD 50fps and while JRVR was reporting sub 20ms on average it was dropping say 10-20 frames per second in big lumps. I'll also see how it goes with std 23.976 material. If that works, I'll setup some profiles

jmone · « **Reply #2 on:** July 04, 2023, 03:57:38 am »

Got a quick test in on one UHD 23.976. Max rendering sits around 36ms and it drops the occasional frame regardless of if tricubic is on or off (it was off by default). I might be able to tune some of the scalling settings to eke out the odd ms reduction. Turning off HDR Peak Detection saves about 5+ms (Max sits around 29ms but spikes to around 36ms as well but did not drop frames on that quick test). I should get some time to see what I can dial in on the NUC12 tmr. Off to look at the PJ now.

haasn · « **Reply #3 on:** July 04, 2023, 11:28:56 am »

Quote from: jmone on July 04, 2023, 03:57:38 am

Got a quick test in on one UHD 23.976. Max rendering sits around 36ms and it drops the occasional frame regardless of if tricubic is on or off (it was off by default). I might be able to tune some of the scalling settings to eke out the odd ms reduction. Turning off HDR Peak Detection saves about 5+ms (Max sits around 29ms but spikes to around 36ms as well but did not drop frames on that quick test). I should get some time to see what I can dial in on the NUC12 tmr. Off to look at the PJ now.

Can you give me a breakdown of where those 36ms are being spent? Maybe we can optimize some low-hanging fruit. I currently only optimize for AMD because that's the only GPU I have in my development rig, unfortunately. Normally AMD and Nvidia are close enough in perf that it doesn't make a difference, but maybe Intel needs some special treatment.

5ms for peak detection sounds rough. Maybe disabling the histogram would help with that? (@Hendrik can you add an option for this?)

Hendrik · « **Reply #4 on:** July 04, 2023, 12:10:02 pm »

Testing has shown that the histogram is making it slow. Some work is underway to investigate.

Hendrik · « **Reply #5 on:** July 04, 2023, 03:48:00 pm »

Setup the NUC12 I had in a box for a while and started testing, can definitely confirm a 5-7ms penalty from turning on peak detect. Will test various performance options and patches proposed by haasn tomorrow once I can hook it up to a 4K display so I can test without scaling being involved.

jmone · « **Reply #6 on:** July 04, 2023, 04:23:38 pm »

Sounds Great!

haasn · « **Reply #7 on:** July 04, 2023, 04:54:11 pm »

I'd appreciate testing of https://code.videolan.org/videolan/libplacebo/-/merge_requests/495 in particular with regards to the magic value of `12`, I have a nagging suspicion it depends on the exact GPU and clip tested.

haasn · « **Reply #8 on:** July 04, 2023, 06:01:05 pm »

Pushed another change to do peak detection after downscaling, instead of at full video res. Should help substantially when viewing 4K HDR sources on non-4K monitors.

Hendrik · « **Reply #9 on:** July 05, 2023, 02:10:18 am »

I've assembled a test build of libplacebo with the latest changes, give it a try for measurements on your NUC12. I'll be doing my own testing a bit later as well.

https://files.jriver-cdn.com/mediacenter/test/jrvr/libplacebo-290-peakdetect-perf.zip

It might be helpful to post a performance breakdown (eg. OSD shot) after the changes as well, so we can see where else the performance might go missing.
Contrast Recovery is a definite candidate, for example.

jmone · « **Reply #10 on:** July 05, 2023, 06:04:39 am »

Sounds great and will do. It will have to be in the morning (Oz Time) though.

Hendrik · « **Reply #11 on:** July 05, 2023, 07:33:00 am »

On a test on my end, on linux, because thats what my NUC runs, I can get stable 24p playback just fine. Performance sits at 20-22ms with peak detect and contrast recovery.
To reach 60 fps, you have to turn off contrast recovery, and even then its a bit hit or miss to stay consistent below 16ms with peak detect on. Might be better on Windows, we'll see once more tests come in. Still looking for some knobs to make it more consistent for 4k60 content.

jmone · « **Reply #12 on:** July 05, 2023, 04:27:01 pm »

Here is a STD UHD BD test on the NUC12 under Win11 - This is dropping some frames (in 1's and 2's) - More to come (the TV just started getting used) - It looks like Peak Detection is eating up most of the performance.

jmone · « **Reply #13 on:** July 05, 2023, 05:13:36 pm »

Here are two more screen shots of the OSD for comparison (does not seem to drop frames)
- Turn Recovery Off
- Turn HDR and Recovery Off

jmone · « **Reply #14 on:** July 05, 2023, 05:27:38 pm »

One more from the same UHD BD. This time just HDR Peak Deteaction Off

Note: Tricubic On/Off does not seem to really impact the performance on these tests.

jmone · « **Reply #15 on:** July 05, 2023, 05:28:30 pm »

...and for good measure UHD 59.94 with most things disabled plays fine (just)

Hendrik · « **Reply #16 on:** July 05, 2023, 06:49:55 pm »

Are you sure you grabbed the test build? This does not seem any different from the numbers you reported before.

I'll probably update the included version with the main build by the end of the week as well.

jmone · « **Reply #17 on:** July 05, 2023, 07:20:24 pm »

The zip file was "libplacebo-290-peakdetect-perf.zip" with 5 files in it all Date Created of 5/07/2023 @ 9:07 am with and extra file I'd not seen before called libplacebo-290.pdb

jmone · « **Reply #18 on:** July 05, 2023, 07:43:02 pm »

All the ones I tested were over the opening parts of Casino Royal. It varies abit by title, this one is from the middle of Bullet Train.

jmone · « **Reply #19 on:** July 05, 2023, 08:13:51 pm »

FWIW - also check for intel updates (updated BIOS), then went into the BIOS and maxed the GPU Memory Allocations, bumped up the power package. No difference.

jmone · « **Reply #20 on:** July 05, 2023, 08:25:22 pm »

Also tried unchecking "Allow HW decoder direct rendering on mismatched size". No change to the overall stats but it now seems to play std UHD without dropping the odd frame.

Hendrik · « **Reply #21 on:** July 06, 2023, 02:33:16 pm »

Build 31 implements an option to disable the histogram in peak detection. Currently it's only used to refine the peak value in some edge cases, and peak detect with no histogram is still massively superior to no peak detect, so give it a try how much of an impact it has for you.

jmone · « **Reply #22 on:** July 06, 2023, 04:54:13 pm »

Thanks. Not seeing any impact vs the prior tests. Attached are more shots from Casino Royal this time with Histogram Off, and with both Historgram and Tricubic Off.

jmone · « **Reply #23 on:** July 07, 2023, 03:04:42 am »

Anything you want me to try or double check regarding settings / versions etc (it's just odd I don't see any differences)?

JimH · « **Reply #24 on:** July 07, 2023, 07:29:21 am »

I wonder if that's related to the crashes mattkahn reported in the last release thread.

Hendrik · « **Reply #25 on:** July 07, 2023, 07:31:34 am »

Did some investigating, and it seems like d3d11 lacks some features that make the peak detect shader faster. When I was testing on Linux I was using Vulkan which has those features.
d3d12 would have them, but it doesn't seem likely someone is going to add that.

Long term maybe we can create interoperability between d3d11 and vulkan so that decode and deinterlace can still use it, and rendering in vulkan, but not soon.

jmone · « **Reply #26 on:** July 07, 2023, 04:16:11 pm »

Thanks - no probs, I was not sure if it was "just me".

FenceMan · « **Reply #27 on:** July 07, 2023, 07:30:51 pm »

Potential issue with the newest posted files to test on NUC -

(https://yabb.jriver.com/interact/index.php/topic,136463.msg945409.html#msg945409)

I am thinking those were causing issues with my 4070 GPU (very high render times and massive skipped frames). I went back to the one posted before that which seems to have fixed things.

Possible something you did caused and issue with Nvidia GPUs?

FenceMan · « **Reply #28 on:** July 08, 2023, 05:29:47 am »

Disregard I cannot seem to reproduce so much have been something else.

INTERACT FORUM

Author Topic: JRVR Performance on NUC12 Pro (Read 1558 times)