narkohol wrote:
Chainik wrote:

num_streams comes from the RIFE profile
force_fp16 = True, tf32 = True, use_cuda_graph = True
come from helpers.py

What about 'Workspace'?

vsmlrt.py 'performance tuning' header says "increase workspace", but in helpers.py it is set to backend.workspace = None

https://i.postimg.cc/Dy3PQKX9/image.png

"The maximum workspace is set to None for the total memory size of the GPU."

none is the best

2

(13 replies, posted in Using SVP)

official build now supported this:
just add the line to mpv.conf
vf=d3d11vpp=scale=2:scaling-mode=nvidia

or bind to a shortcut

ctrl+g vf toggle d3d11vpp=scale=2:scaling-mode=nvidia

v15.2 gives me a tiny FPS boost
v14test3 ~150fps vs ~ v15.2 ~158fps, also fixed dynamic shape (Performance boost: off) bug

15.2 changelog:

    Upgraded to TensorRT 10.3.0.

    Fixed performance regression of RIFE and SAFA models starting with vs-mlrt v14.test4. This version may still be slightly slower than vs-mlrt v14.test3 under some conditions, however.

Xenocyde wrote:
aloola wrote:

yes, it's the only reason we use v2
from my benchmark, v2 gives a huge fps boost

Why is HAGS ON? Shouldn't turning it off provide an even bigger fps boost?

I want to use Frame Gen in some games that's why I turn it on.

also
HAGS ON gives you better FPS on light tasks, but it's worse on heavy tasks. (in the end, turning it off is still better for RIFE)

Ryzen 5900x + 4080 SUPER
https://cdn.discordapp.com/attachments/290709370600423424/1235626067331645490/firefox_13YuVQIp4S.jpg?ex=66350df3&is=6633bc73&hm=636d98f1a957d216eb9b21e32ce5b80b464105395d375d28be87dc6d4732e4d7&


if you want to benchmark with mpv, add the configs to the bottom of your mpv.conf file.

profile = benchmark # add # to disable benchmark
[benchmark]
osd-msg1 = "FPS: ${estimated-display-fps}"
opengl-swapinterval = 0 # {lock☒}
opengl-dwmflush = no # {lock☒}
vulkan-swap-mode = immediate # When using --gpu-context=winvk, if the number of frames is equal to the refresh rate, try changing to mailbox
d3d11-sync-interval = 0 # {lock☒}
d3d11-flip = no # {lock☒} The swapping effect of flip may clamp the extreme frame rate FPS_max=number of front and rear buffers x refresh rate

hidpi-window-scale = no # {lock☒} Disable DPI scaling first and then the specified magnification will be accurate
window-scale = 1.5 # Manually specify the source zoom ratio (limited by --auto-fit=<value>)
sub = no # {lock☒}
audio = no # {lock☒}
keep-open = always # {lock☒}
# display-fps-override = 24 # Match source framerate as much as possible
video-sync = display-desync # {lock☒}
interpolation = no # {lock☒}

Blackfyre wrote:

So v2 in some instances definitely performs better

yes, it's the only reason we use v2
from my benchmark, v2 gives a huge fps boost

https://cdn.discordapp.com/attachments/290709370600423424/1235391885066108949/firefox_kme87vRahS.jpg?ex=663433d9&amp;is=6632e259&amp;hm=42a5efff4fd00577bec00fb2e10a0a863f43b5dabcb9aa1863601711e0588b6e&amp;

Drakko01 wrote:

If I use mpc-VR (RTX HDR + RTX SuperRes) with MPC-HC or MPC-BE I don't get the best experience, the quality is very good, but I have a kind of stutter in slow camera movements or certain close ups,  no in panning movements. This is not present when I use Madvr. Maybe i got some settings wrong in MpcVR dunno.
BTW I always talking of Realtime Playback just for clarify.

I think the lack of G-sync/freesync support causing this.
try the options video -> frame synch -> "Sync video to display" in MPC, it may fix the problem

mpc-BE + mpc-VR (RTX HDR + RTX SuperRes) + SVP RIFE give you the best experience and easy to config.

also for mpv I found this is a good shader for anime/real-life videos https://github.com/cunnyplapper/CuNNy/t … r/mpv/fp16

Drakko01 wrote:

I was always  curious why the devs never updated the TRT Version if there its a reason not to.

there are compatible issues, like crashes on seek, or performance degradation. trt 9.2 seems stable though, I haven't heard any problems with it

unrealit wrote:

Hi!

Sorry for the stupid Question!

But , How can i get Benchmark Values?

And what is HAGS?

What do you mean?

Thanks smile

use the script here.
https://github.com/AmusementClub/vs-mlrt/discussions/19

HAGS = Hardware Accelerated GPU Scheduling
you could google it for more info.

In RIFE with HAGS on performance might drop up to 25%.

unreality wrote:

Hello, I actually noticed that too, fixed 50fps doesn't run at all... very jerky! If I should use 2x24, i.e. 48p, then I will have to endure jerking every time the camera pans?

Is there no solution for this?

you can try the mpv options
video-sync=display-resample
interpolation=yes

# Force video to lock on the display's refresh rate, and change video and audio
# speed to some degree to ensure synchronous playback - can cause problems
# with some drivers and desktop environments.

you don't need this if your monitor support VRR/G-sync

unreality wrote:

Hello everyone,

I would like to play back 50 Hz with MPV (implemented in SVP) because my i9 14900 with DDR5 6200 and RTX 4090 probably can't play 2.5 speed at 4K. So I have to play back at 2x (48p) but the TV does 50hz. is there a possibility? Can't you fix something at MPV?

you should stop at 24x2 48

50fps, 60fps = cost the same power as 24x3

1080@x5 HAGS OFF 4t build 5 in 331.55 seconds
Output 8000 frames in 45.79 seconds (174.70 fps) 4.15 lite
Output 8000 frames in 52.86 seconds (151.35 fps) 4.15

1080@x5 HAGS OFF 4t build 4 in 101.19 seconds
Output 8000 frames in 37.26 seconds (214.68 fps) 4.15 lite
Output 8000 frames in 54.36 seconds (147.16 fps) 4.15

1080@x5 HAGS OFF 4t build 3 in 70.20 seconds
Output 8000 frames in 39.49 seconds (202.59 fps) 4.15 lite
Output 8000 frames in 51.14 seconds (156.42 fps) 4.15

1080@x5 HAGS OFF 4t build 2 in 41.65 seconds
Output 8000 frames in 36.79 seconds (217.45 fps) 4.15 lite
Output 8000 frames in 51.62 seconds (154.99 fps) 4.15

1080@x5 HAGS OFF 4t build 1 in 41.60 seconds
Output 8000 frames in 36.44 seconds (219.55 fps) 4.15 lite
Output 8000 frames in 52.00 seconds (153.83 fps) 4.15

1080@x5 HAGS OFF 4t build 0 in 22.14 seconds
Output 8000 frames in 42.09 seconds (190.09 fps) 4.15 lite
Output 8000 frames in 63.52 seconds (125.95 fps) 4.15

4Kx2 HAGS OFF 4t build 0 in 23.42 seconds
Output 2000 frames in 36.49 seconds (54.80 fps) 4.15 lite
Output 2000 frames in 51.31 seconds (38.98 fps) 4.15

4Kx2 HAGS OFF 4t build 1 in 77.06 seconds
Output 2000 frames in 33.16 seconds (60.32 fps) 4.15 lite
Output 2000 frames in 39.89 seconds (50.14 fps) 4.15

4Kx2 HAGS OFF 4t build 2 in 88.97 seconds
Output 2000 frames in 33.50 seconds (59.69 fps) 4.15 lite
Output 2000 frames in 40.41 seconds (49.49 fps) 4.15

4Kx2 HAGS OFF 4t build 3 in 135.63 seconds
Output 2000 frames in 33.19 seconds (60.26 fps) 4.15 lite
Output 2000 frames in 38.83 seconds (51.51 fps) 4.15

4Kx2 HAGS OFF 4t build 4 in 172.76 seconds
Output 2000 frames in 34.21 seconds (58.46 fps) 4.15 lite
Output 2000 frames in 39.08 seconds (51.17 fps) 4.15

4Kx2 HAGS OFF 4t build 4 in 566.31 seconds
Output 2000 frames in 33.42 seconds (59.85 fps) 4.15 lite
Output 2000 frames in 39.04 seconds (51.23 fps) 4.15


well just stick with build 3 (default) or build 2 for faster build

has anyone tried the option builder_optimization_level=5?
https://github.com/AmusementClub/vs-mlr … g/v14.test

builder_optimization_level – The builder optimization level which TensorRT should build the engine at. Setting a higher optimization level allows TensorRT to spend longer engine building time searching for more optimization options. The resulting engine may have better performance compared to an engine built with a lower optimization level. The default optimization level is 3. Valid values include integers from 0 to the maximum optimization level, which is currently 5. Setting it to be greater than the maximum level results in identical behavior to the maximum level.

someone here said lvl 5 has better performance.
https://github.com/Blinue/Magpie/discussions/844

Olombo wrote:

Hi

maybe some progress over here....
Setting frc/frame/resize to -1 makes uhd work. But.....what am I doing here? Wiki says "resize to screen if larger". Am I loosing content/resokution by doing so? Other downsides?

what is your display resolution?

default SVP will resize video to display screen's

right now I just need to keep these 3 models:

4.6 for performance
4.16 lite for balance
4.15 for quality

also the mpv VRR bug
looks like d3d11va will cause VRR to stop working, d3d11va-copy and other hwdecs work with VRR

Olombo wrote:

Hi

hardware:
I5-13600
Ddr5 -4800 32gb stock settings
Gainward rtx 4090

MPC-HC
TensorRT
Boost enabled
HAGS off
Win11( anybody else or all using win10??)
Display @47.952
Tried "as Monitor" and "movie x2" with "force exact value" on and off. No difference at all.
AI model....4.6 or 4.15.....no difference. All stuttering

UHD in MPV hase some dropped frames. In MPC is like every fourth frame is dropped. MadVR is almost conpletely passtrough and no difference.

you might try to turn off hardware acceleration (hwdec) in your player. at least in my case, it gave me worse performance on 4K videos.

my point is 4.9 and 4.16 lite have the same performance but 4.16 lite has better quality. then there is no reason for people to use 4.9 even if live-action films are the same quality (no proof or testing yet)

4.16 lite has fewer artifacts and about the same performance as 4.9, so if you are using 4.9, time to switch to 4.16 lite

4.9
https://cdn.discordapp.com/attachments/290709370600423424/1221853277402959872/image.png?ex=6614168a&amp;is=6601a18a&amp;hm=a6b520386b27d4a7e9dd2a2bc914565cf038b10131f18aa32795c5e379e38395&amp;

4.16 lite
https://cdn.discordapp.com/attachments/290709370600423424/1221853645042225242/image.png?ex=661416e1&amp;is=6601a1e1&amp;hm=295620252c5b9c4b02d05386e689306280419b0bd93247016c01bc391e2154a3&amp;


1080@x5 HAGS OFF
Output 8000 frames in 36.37 seconds (219.96 fps) 4.16 lite
Output 8000 frames in 50.77 seconds (157.59 fps) 4.15
Output 8000 frames in 52.62 seconds (152.04 fps) 4.14 lite
Output 8000 frames in 50.25 seconds (159.21 fps) 4.14
Output 8000 frames in 37.24 seconds (214.82 fps) 4.9
Output 8000 frames in 32.78 seconds (244.02 fps) 4.6 (DRAM bottleneck)

the best model atm in terms of quality/performance

Olombo wrote:

Abraxas: how is it possible you can watch 4k x2 UHD with a 4070? I can not get it running without frame drops with a 4090 and mpc-hc?

May it be the model of 4090 why it does not work for some of us? Gainward is mine...others with stock OC may be faster

my 3070ti with HAGS off can run 4k@24 x 2 with a few frame drops. so 3080 should be a minimum to be able to smoothly run 4k@24 x 2 with HAGS off

what's your PC specs: CPU, RAM speed, bus, RAM timing?

is there any comparison for frame.resize ?
I mean if it downscale to 1080p, why then starting from 4k and not start with 1080p ? You're not really playing 4K then

you don't need to resize it to 1080p, 4090 should do perfectly fine for 3040 x 1710@24 x3

* with 4.6 v2 model

It looks like x3 4k with RIFE is impossible if your RAM is DRAM4

x2 cost 25~28GB/s, my max memory BW is 31GB/s

https://cdn.discordapp.com/attachments/290709370600423424/1220054721440776312/image.png?ex=660d8b81&amp;is=65fb1681&amp;hm=f673d41b28c29cf8872076b7f0f28dd2d94790f9051d9d10dd3c48d1badf7678&amp;

based on the picture you might need RAM5 6000+ to run RIFE 4k x3
https://www.servethehome.com/wp-content … 68x648.jpg

it's not your GPU fault, the cause is the DDRAM's BW bottleneck (this was confirmed) or even the Vapoursynth bottleneck, we don't know for sure.

instead of x3 you can try a better quality model like Rife 4.9 or 4.15 x2

Olombo wrote:

Changing the copy back has no effect. I recognized that also mpv has drops. I just had a mistake while testing (used 1080p file....works easy of course. Only UHD has frame drops).

So....2 different players both with frame drops. Which rife model I use makes no real difference. Any recommendation how to proceed?

there are some options to boost the performance:
- turn off HAGS (in windows)
- pre-resize video resolution to lower res (in SVP menu)

Olombo wrote:

Ok....I check transcoding. This works with x2 for up to 60fps. Also mpv player works fine with installed 4.14
Any Idea what could cause mpc to stutter but lets mpv work perfect?

try to turn off the Hardware Decoder in LAV filter or change it to something not -copy.

in my case hardware decoder copy-back works badly with UHD+ videos, stuttering even without SVP/RIFE.



my benchmark RTX 4080

4kx2 HAGS ON
Output 8000 frames in 128.49 seconds (62.26 fps) 4.6
Output 8000 frames in 139.04 seconds (57.54 fps) 4.9
Output 8000 frames in 177.04 seconds (45.19 fps) 4.14
Output 8000 frames in 179.05 seconds (44.68 fps) 4.14 lite
Output 8000 frames in 181.29 seconds (44.13 fps) 4.15

1080@x5 HAGS OFF
Output 8000 frames in 50.77 seconds (157.59 fps) 4.15
Output 8000 frames in 50.25 seconds (159.21 fps) 4.14
Output 8000 frames in 37.24 seconds (214.82 fps) 4.9
Output 8000 frames in 32.78 seconds (244.02 fps) 4.6

*The benchmark is just for comparison performance between models, not for real-world video*

framo wrote:
flowreen91 wrote:

If done correctly, restart SVP and then you should be able to see selectable AI models with (v2) in their name like this:
https://gyazo.com/9d6d1360852e8e92f834425e69803cb0

im using ncnn/vulkan and theres no drop down menu for me, am i doing sthg wrong?
i downloaded the v4.14 model and named it "rife_v.4.14f" , put the name in the rife_trt_model section in application settings. how can i make sure it is using the correct model?

Download the files here
https://github.com/styler00dollar/Vapou … ter/models

https://cdn.discordapp.com/attachments/290709370600423424/1214794469782851644/image.png?ex=65fa6883&amp;is=65e7f383&amp;hm=ee42f5e828693a3931fce3d6c945f50a616acc78912ef7b5ef096da8c78268e4&amp;

dawkinscm wrote:
aloola wrote:

because of this bug https://github.com/AmusementClub/vs-mlrt/issues/72

if you use model v4.7+ you need to turn performance boost on else you will get the error.

Isn't this a separate issue to the TensorRT bug you refer to?  I don't think the use of dynamic shapes reduces the memory footprint enough to help @oriento who is literally running out of memory. Using performance mode reduces the amount of memory being used. Lower number Rife models also use less memory. Maybe for a 20 series card a user should do both by using the SVP default Rife 4.9 which works almost as well as later Rife models but is less taxing on the GPU.

it's the same
performance boost off = dynamic shape
performance boost on = static shape

you can test it yourself with performance boost off + model v2 4.7+