You are not logged in. Please login or register.
SmoothVideo Project → Using SVP → New RIFE filter - 3x faster AI interpolation possible in SVP!!!
have you tried
hwdec=dxva2 and dxva2-copy ?
Without VRR everything works fine, no problems. With VRR I am pretty sure I tried them both the other day as well, and it was the same issues.
I'll test them again next time I update the driver and enable GSync/Freesync
VRR (GSync and Freesync Premium) enabled in the Display and NVCP = Crushed blacks and banding with d3d11va-copy running at 24FPS (even outside the VRR range, so VRR is not even working).
have you tried to turn on "G-sync compatible indicator" in Nvidia control panel?
because I tested with various mpv options, and g-sync always turn on in full screen for me. I also LG C2 48''
the green text on top right is a proof that G-sync is active
https://cdn.discordapp.com/attachments/ … 7ab4d&
ah, I forgot. press the green button 7 times on your TV remote will show you VRR information. check if the fps there sync with your video fps
I don't suppose this works with the newly open-sourced ZLUDA for running CUDA software on AMD GPUs?
https://github.com/vosen/ZLUDA
https://www.phoronix.com/review/radeon-cuda-zluda
One particularly key interesting point from the Phoronix review on page 3 is that Blender actually ran faster with CUDA + ZLUDA than its native HIP codepath, so it's within reason that ZLUDA could also be faster than Vulkan for AMD GPUs.
VRR (GSync and Freesync Premium) enabled in the Display and NVCP = Crushed blacks and banding with d3d11va-copy running at 24FPS (even outside the VRR range, so VRR is not even working).
have you tried to turn on "G-sync compatible indicator" in Nvidia control panel?
because I tested with various mpv options, and g-sync always turn on in full screen for me. I also LG C2 48''
the green text on top right is a proof that G-sync is active
https://cdn.discordapp.com/attachments/ … 7ab4d&ah, I forgot. press the green button 7 times on your TV remote will show you VRR information. check if the fps there sync with your video fps
You can visually see the difference in smoothness when GSync is working vs when it's not.
But also, you can check using the “Game Optimiser” settings with the LG Remote. When GSync is working, the frame rate corresponds with the refresh rate shown in the Game Optimiser, but when GSync is not working, it shows the highest refresh rate.
That's how I check it.
Regardless, the visual differences when HDR is enabled in Windows between d3d11va-copy and d3d11va occur whether running SVP or not as I wrote in the last post on the last page. So running at 24FPS default, which is outside the VRR range. When VRR is enabled, even when not triggered, it is causing crushed blacks and banding with d3d11va-copy. At least for me with the configuration I posted.
You can test these if you enable HDR under Display Settings in Windows Settings. Assuming you have GSync enabled in NVCP and the TV.
Test it with Ahsoka Episode 1 in HDR if you have that as the beginning sequence in space and the intro itself will show both. The logo's at the start will show banding with hwdec=d3d11va-copy below, and the dark scene after that has crushed blacks. But when you change to hwdec=d3d11va, both issues will resolve as VRR breaks.
ontop
fullscreen=yes
d3d11-exclusive-fs=yesvo=gpu-next
gpu-api=d3d11
hwdec=d3d11va-copy
video-sync=audio
gpu-context=d3d11
fbo-format=rgba16hftarget-peak=800
target-contrast=inf
target-colorspace-hint=yes
End of above.
I don't suppose this works with the newly open-sourced ZLUDA for running CUDA software on AMD GPUs?
https://github.com/vosen/ZLUDA
https://www.phoronix.com/review/radeon-cuda-zludaOne particularly key interesting point from the Phoronix review on page 3 is that Blender actually ran faster with CUDA + ZLUDA than its native HIP codepath, so it's within reason that ZLUDA could also be faster than Vulkan for AMD GPUs.
^ This comment should be interesting for @Chainik and the devs.
In my case my chinese 165Hz GSync monitor doesnt breaks VRR d3d11va hardware decoder, and cant be out of range because 24fps content synced to 48Hz auto, less/x3Hz hehe
hwdec=d3d11va-copy
What happens if you use hwdec=no ？
What happens if you use hwdec=no ？
Next time I install a new driver, I'll turn on GSync again and test it.
EDIT:
Also, this goes for everyone, instead of quoting a huge block of text, please just quote the section you want to reply to, or only the important part, and remove everything else.
It makes reading the forum a lot easier.
Now that Nvidia launched the Nvidia App on latest update that turns SDR games into HDR:
https://www.nvidia.com/en-us/software/nvidia-app/
did people started playing around with the RTX HDR & RTX Dynamic Vibrance to watch their non-HDR videos with it ?
You can easily enable it on your video player by drag and dropping it in the NvTrueHDR app:
https://www.nexusmods.com/site/mods/781
>Now that Nvidia launched the Nvidia App on latest update that turns SDR games into HDR:
HDR is too rare even here. Its basically OLED TV and 3-4 monitors. Below HDR 1000 its crap
Is there some comparing of RIFE and Standard svp engine? Cant find. Thanks
Is there some comparing of RIFE and Standard svp engine? Cant find. Thanks
anders.nilsson wrote:Is there some comparing of RIFE and Standard svp engine? Cant find. Thanks
Bad comparison, but okay! Just for sure, generally RIFE are 6-8 times slower than SVP algo, much accurate, expensive!
At movies RIFE (starts 4.4) provides more smoothness and 70% playback time less artefacts (about 88% for 4.14 models)!
Example 2 : SVP algorithm cant do 6x interpolation accurately enough from low frame rate sources (24>144) because every 4/5th frame will blended together or duplicated (content dependant)!
The remaining issue I have with RIFE is the double images for fast motion. What I found is that the higher the frame-rate the less the issue occurs. At x3 frame-rate from the issue is reduced but then I get stutters. But turning on interpolation in mpv removes those stutters and I get the smooth motion back and better fast motion handling, but it still has issues with horizontal slow pans. I like it, but the issue with slow pans annoys me. Maybe if you have a 4090 this might be useful.
but it still has issues with horizontal slow pans.
If i'm using the mpv setting "video-sync=display-resample" this scene has massive stutters:
https://youtu.be/jZvFEtR8RH0?si=59TEOOJ … m&t=18
We should either use "video-sync=audio" like Blackfyre's configs or nothing at all.
We should either use "video-sync=audio" like Blackfyre's configs or nothing at all.
Referring to https://mpv.io/manual/master/
I believe video-sync=audio is the mpv default behaviour.
flowreen91 wrote:We should either use "video-sync=audio" like Blackfyre's configs or nothing at all.
Referring to https://mpv.io/manual/master/
I believe video-sync=audio is the mpv default behaviour.
Yes it is.
dawkinscm wrote:but it still has issues with horizontal slow pans.
If i'm using the mpv setting "video-sync=display-resample" this scene has massive stutters:
https://youtu.be/jZvFEtR8RH0?si=59TEOOJ … m&t=18
We should either use "video-sync=audio" like Blackfyre's configs or nothing at all.
Most configs for mpv on the Internet are out of date at best. But people will copy them verbatim without understanding. Blackfyre's config is his own and in previous comments I noted some unnecessary settings. This is another one, but I only started using "video-sync" myself in the last few days so I'm still learning.
As for the "display-resample" option, I don't need it to be on, but for me it seems to work well in smoothing out some of the remaining micro-stutters even without interpolation=on. But I watch stuff on a large VR screen so I will see micro-stutters that many wont. Because of that, my SVP setup including scripts will be different to most. Everyone's situation is different.
When i tried rife 4.14 v2 (i just copied the onnx file)
it gives an error on screen, and on a log file
[03/02/2024-21:34:57] [W] Could not read timing cache from: C:\Users\Attila\AppData\Roaming\SVP4\cache\Program Files (x86)/SVP 4/rife\models\rife\rife_v4.14.onnx.min64x64_opt1920x1088_max2560x1440_fp16_no-tf32_trt-8502_cudnn_I-fp16_O-fp16_NVIDIA-GeForce-RTX-2080_9c726c00.engine.cache. A new timing cache will be generated and written.
[03/02/2024-21:34:57] [E] Error[4]: [network.cpp::nvinfer1::Network::validate::3100] Error Code 4: Internal Error (input: for dimension number 1 in profile 0 does not match network definition (got min=11, opt=11, max=11), expected min=opt=max=7).)
[03/02/2024-21:34:57] [E] Error[2]: [builder.cpp::nvinfer1::builder::Builder::buildSerializedNetwork::751] Error Code 2: Internal Error (Assertion engine != nullptr failed. )
[03/02/2024-21:34:57] [E] Engine could not be created from network
[03/02/2024-21:34:57] [E] Building engine failed
[03/02/2024-21:34:57] [E] Failed to create engine from model or file.
[03/02/2024-21:34:57] [E] Engine set up failed
rtx 2080 driver 554.61
build in older versions of rife (like 4.4) works
edit : thank you, i missed that i should create a folder. Sorry for that. I have searched my error on the forum and this one wasn't listed ^^'
edit2: actually my 8gb rtx 2080 is not compatible with 4.14 or 4.14 lite, it tries to allocate exactly 8gb vram, but it can't as the vram already contains a little bit of something
[03/02/2024-23:38:14] [E] Error[2]: [virtualMemoryBuffer.cpp::nvinfer1::StdVirtualMemoryBufferImpl::resizePhysical::145] Error Code 2: OutOfMemory (no further information)
[03/02/2024-23:38:14] [E] Error[1]: [virtualMemoryBuffer.cpp::nvinfer1::StdVirtualMemoryBufferImpl::resizePhysical::132] Error Code 1: Cuda Driver (invalid argument)
[03/02/2024-23:38:14] [W] [TRT] Requested amount of GPU memory (8589934592 bytes) could not be allocated. There may not be enough free memory for allocation to succeed.
[03/02/2024-23:38:14] [W] [TRT] Skipping tactic 1 due to insufficient memory on requested size of 8589934592 detected for tactic 0x0000000000000001.
Try decreasing the workspace size with IBuilderConfig::setMemoryPoolLimit().
edit3: enabling "performance boost" fixes the memory problem
When i tried rife 4.14 v2 (i just copied the onnx file)
To use the v2 versions you must not put them inside the v1 folder.
It should look something like this file structure:
The parent folder "C:\Program Files (x86)\SVP 4\rife\models\":
https://gyazo.com/5a4401fd17ff3d1f009120a5db754c2b
The v2 only folder:
"C:\Program Files (x86)\SVP 4\rife\models\rife_v2\"
https://gyazo.com/d7ec471252d535602bc4ece31bfaee27
The v1 only folder:
"C:\Program Files (x86)\SVP 4\rife\models\rife\"
https://gyazo.com/f2a125fec5f6a18e7e3eef28e1f85999
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
When i tried rife 4.14 v2 (i just copied the onnx file)
it gives an error on screen, and on a log file
[03/02/2024-21:34:57] [W] Could not read timing cache from: C:\Users\Attila\AppData\Roaming\SVP4\cache\Program Files (x86)/SVP 4/rife\models\rife\rife_v4.14.onnx.min64x64_opt1920x1088_max2560x1440_fp16_no-tf32_trt-8502_cudnn_I-fp16_O-fp16_NVIDIA-GeForce-RTX-2080_9c726c00.engine.cache. A new timing cache will be generated and written.
[03/02/2024-21:34:57] [E] Error[4]: [network.cpp::nvinfer1::Network::validate::3100] Error Code 4: Internal Error (input: for dimension number 1 in profile 0 does not match network definition (got min=11, opt=11, max=11), expected min=opt=max=7).)
[03/02/2024-21:34:57] [E] Error[2]: [builder.cpp::nvinfer1::builder::Builder::buildSerializedNetwork::751] Error Code 2: Internal Error (Assertion engine != nullptr failed. )
[03/02/2024-21:34:57] [E] Engine could not be created from network
[03/02/2024-21:34:57] [E] Building engine failed
[03/02/2024-21:34:57] [E] Failed to create engine from model or file.
[03/02/2024-21:34:57] [E] Engine set up failedrtx 2080 driver 554.61
build in older versions of rife (like 4.4) works
edit : thank you, i missed that i should create a folder. Sorry for that. I have searched my error on the forum and this one wasn't listed ^^'
edit2: actually my 8gb rtx 2080 is not compatible with 4.14 or 4.14 lite, it tries to allocate exactly 8gb vram, but it can't as the vram already contains a little bit of something
[03/02/2024-23:38:14] [E] Error[2]: [virtualMemoryBuffer.cpp::nvinfer1::StdVirtualMemoryBufferImpl::resizePhysical::145] Error Code 2: OutOfMemory (no further information)
[03/02/2024-23:38:14] [E] Error[1]: [virtualMemoryBuffer.cpp::nvinfer1::StdVirtualMemoryBufferImpl::resizePhysical::132] Error Code 1: Cuda Driver (invalid argument)
[03/02/2024-23:38:14] [W] [TRT] Requested amount of GPU memory (8589934592 bytes) could not be allocated. There may not be enough free memory for allocation to succeed.
[03/02/2024-23:38:14] [W] [TRT] Skipping tactic 1 due to insufficient memory on requested size of 8589934592 detected for tactic 0x0000000000000001.
Try decreasing the workspace size with IBuilderConfig::setMemoryPoolLimit().edit3: enabling "performance boost" fixes the memory problem
because of this bug https://github.com/AmusementClub/vs-mlrt/issues/72
if you use model v2 v4.7+ you need to turn performance boost on else you will get the error.
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.
So after more testing I pretty much have a fixed mpv/SVP configuration for VR. Although knowing me, it won't stop me from tinkering. None of this is strictly necessary because SVP works pretty well out of the box.
For SVP I use Rife 4.13v2 with 12% SCT at 60fps fixed. I'm also using a later version of vsmlrt which has some extra code for playback buffering but this is not compatible with the default SVP helper scripts so I modified it to work.
For mpv the difference betwen using hwdec and not using it is about 1-2% in favor of not using it. YMMV. I also use interpolation with some motion blur. Overall motion was smooth enough before, but fast motion is now almost perfectly fluid so I am very happy with the results. Looking at the SVP manual, is it still the case that SVP runs after decoding but before rendering meaning that mpv is interpolating the output of SVP?
For SVP I use Rife. I also use interpolation with some motion blur.
What i imagine when people enable interpolation from mpv config on top of the movie already being fully interpolated by RIFE:
https://gyazo.com/8ce62ce348609112c567448cb406a223
I pretty much have a fixed mpv/SVP configuration for VR.
Please share your mpv.conf solution, other VR users might be interested.
What i imagine when people enable interpolation from mpv config on top of the movie already being fully interpolated by RIFE:
https://gyazo.com/8ce62ce348609112c567448cb406a223
lol But as I said yesterday, I don't need interpolation to be enabled and it probably doesn't have room to do much. I now use it because it needs to be active in order to set up motion-blur.
Please share your mpv.conf solution, other VR users might be interested.
Not unless they have OLED lenses because OLED has an issue with smooth movement. I'm not particularly susceptible to it myself but even I can see that the extra config reduces the OLED micro-stutters. LED doesn't have the same issues. The rest is mostly standard stuff from my Plex mpv config.
#start
input-ipc-server=mpvpipe
hwdec=no
vo=gpu-next
profile=high-quality
#This is standard stuff most of which is not needed because it's the default but I like to play around with them 
gpu-api=auto
gpu-context=auto
linear-downscaling=yes
correct-downscaling=yes
dither=ordered
dither-depth=8
temporal-dither=yes
deband=yes
pause="no"
no-hidpi-window-scale
#This is config specific to my situation and as some have already seen, it can cause problems even for other VR users because they all use LED lenses.
video-sync=display-resample
interpolation=yes
tscale=sphinx
tscale-blur=0.65
swapchain-depth=1
#standard shader stuff
glsl-shaders="~~/shaders/CfL_Prediction.glsl"
glsl-shaders-append="~~/shaders/adaptive-sharpen.glsl"
After that there's shader stuff that won't work because they aren't designed to work the way I use them because I modified the code. So I haven't included them here because they are only used for special circumstances. I also haven't included any color settings because every display is different.
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+
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 shapeyou can test it yourself with performance boost off + model v2 4.7+
Oh right, good to know. After all this time still learning new things. Thanks
SmoothVideo Project → Using SVP → New RIFE filter - 3x faster AI interpolation possible in SVP!!!
Powered by PunBB, supported by Informer Technologies, Inc.
