Welcome to Doom9's Forum, THE in-place to be for everyone interested in DVD conversion. Before you start posting please read the forum rules. By posting to this forum you agree to abide by the rules. |
30th July 2018, 13:51 | #801 | Link | |
Registered User
Join Date: Apr 2002
Posts: 756
|
Quote:
And as said before Skylake or whatever lake Intel has, has little to do with AVX support. All the Pentium / Celeron CPU don't support AVX, even thought they are Skylake +. We only have to wait a few more years. We finally have a decent alternative to Intel x86. |
|
30th July 2018, 13:53 | #802 | Link |
Registered User
Join Date: Mar 2002
Posts: 863
|
The effect is similar, but with a bigger gap. The b-frame thing occured in every 3rd frame or so, the golden frame boost happens in every 16th or so frame, and with a way higher quality boost. These factors make the ""pumping" effect more apparent. It could be tuned to be less noticeable, but that would worsen the metrics, so i doubt they would consider it.
|
30th July 2018, 15:11 | #803 | Link | |
Registered User
Join Date: Oct 2009
Posts: 930
|
Quote:
|
|
30th July 2018, 15:21 | #805 | Link | ||
Registered User
Join Date: Jul 2008
Posts: 2
|
Quote:
Quote:
In conclusion, I believe what industry possibly need in the future is AV1 decoder implementation on Vulkan compute, that will work on ARM and Intel-based devices too (in contrary to HSA implementation). However, I think that GPU hybrid implementations will quickly sunset such efforts (attempts to implement AV1 decoder on OpenCL, Vulkan, CUDA, etc.) and make them obsolete. Eventually hybrid implementations will be replaced by more complete hardware solutions, while legacy devices will continue to decode either AV1 on CPU or recieve VP9 or H264 stream for decoding on GPU, exactly as they do now. We seen how this play out with H264->VP9 transition and I don't think this time may be different. Last edited by RussianNeuroMancer; 30th July 2018 at 19:50. |
||
30th July 2018, 21:45 | #806 | Link |
German doom9/Gleitz SuMo
Join Date: Oct 2001
Location: Germany, rural Altmark
Posts: 6,782
|
Then I used the wrong term; let's call it "GOP pumping" instead. DVD GOPs usually have a duration of half a second. And some MPEG-2 (hardware?) encoders prefer the I frame over all the following P and B frames with finer quantization. This effect was quite similar to your Golden Frame issue.
|
30th July 2018, 22:17 | #807 | Link | |
Registered User
Join Date: Mar 2002
Posts: 863
|
Quote:
|
|
31st July 2018, 04:33 | #808 | Link |
Angel of Night
Join Date: Nov 2004
Location: Tangled in the silks
Posts: 9,559
|
ARM64 is a darling in some circles, but it's hardly perfect, and it's even more segmented than the x64 market. You can't get AVC on some x64, but you can't be guaranteed NEON at all (which is like saying you can't get MMX/SSE) on some ARM64. That market is weird. It's also having serious scale-up issues, like everyone else who tries to make silicon go fast.
Xeon Phi has its diehards, but the SPARC route of lots of intermediate-powered chips hasn't set the world on fire. Intel appears to be abandoning the Knights platform, and moving toward doing things with superfast flash (Optane) mixed into a mesh of micro-CPUs. We'll see if that pans out any better. And then, of course, there's GPU. There are people out there who want to ditch the CPU entirely for the GPU, or at least give it an even larger role than it already grabs in today's computing. There's a metric buttload of research on this, everything from new languages to on-the-fly recompiling x86 code into shaders that can run massively parallel, but the current hybrid solution is the current solution for a reason. It works, it's cheap, and it's accessible. Vulcan, OpenCL, and CUDA have proven to be powerful glue languages that specialize in what they can do very well, and fall over when asked to compute loops with side effects. Toss in a chip to do that processing faster, and hey, you have an underpowered CPU in the middle of your GPU! Like fixed-function video decoders, though, the case may be that such a thing could evolve and slowly push out the overkill system CPU, absorbing the CPU into the GPU instead of vice versa (or most likely, both in parallel). It doesn't take a genius to understand that if SIMD width keeps doubling, you're basically starting to process textures. As long as gamers and coin miners keep buying wildly overpowered GPUs for overspecced monitors, though, the state of the art will keep being advanced. Oh, and quantum computing. Google's Bristlecone set a fire in the breasts of many a nerd, but we're probably almost a decade off from getting any derivative in an average joe's hand, even by cloud computing. Researchers are going to be grappling with what the hell to do with these for decades. Some brilliant folks are going to have their name immortalized in the coming years, but quantum has only just graduated from "nuclear fusion" territory. |
31st July 2018, 04:47 | #809 | Link |
Angel of Night
Join Date: Nov 2004
Location: Tangled in the silks
Posts: 9,559
|
Golden frames are technically superior to ref-frames in every way. It's just that no one's actually taken advantage of that superiority, just like as far as I know no one has taken advantage of long-refs in HEVC to minimize the impact of scenecuts, especially repeated scenecuts. The reduced refs HEVC gives you does take some of that flexibility away though.
|
1st August 2018, 01:50 | #812 | Link | |
Registered User
Join Date: Jun 2013
Posts: 95
|
Quote:
The encoder will be immediately interesting for its speed. The quality will (hopefully) come later. And by quality, I mean subjective quality.
__________________
https://github.com/MoSal |
|
1st August 2018, 10:02 | #813 | Link | |
Registered User
Join Date: Jul 2018
Posts: 80
|
I'm trill by this new video codec (AV1), but also with the image (av1-avif) brought by the same Alliance.
This AVIF image codec it's better that WebP. You can compare the image AV1-AVIF with other formats here. (Select the AV1-2018 version) You can also compare the video AV1 with other formats here. It's very good that it begins new implementation of AV1, but how an encoder (rav1e) can make this statement? Quote:
About the AV1 roadmap We just complete phase 1. In 1 year we complete phase 2. In 2 to 3 years we complete phase 3. In 4 to 5 years we complete phase 4. Last edited by marcomsousa; 1st August 2018 at 10:30. |
|
1st August 2018, 10:16 | #814 | Link |
Registered Developer
Join Date: Mar 2010
Location: Hamburg/Germany
Posts: 10,346
|
Its safe because its written in Rust, which has fundamental type and memory safety built-in.
__________________
LAV Filters - open source ffmpeg based media splitter and decoders |
1st August 2018, 21:18 | #815 | Link | |||
Registered User
Join Date: Nov 2009
Location: Northeast Ohio
Posts: 447
|
Quote:
Quote:
---------------------------------------------------------------- Quote:
And of course, most of the lower core-count Ryzen parts (4c/8t and less) also have a relatively powerful iGPU as well (there's only a few 4core Ryzen parts that lack an iGPU, and they're all on desktop anyway). Also, technically this isn't a "finally" as AMD was very competitive with Intel in the first half of the 2000s (Athlon XP to Athlon 64 x2) as well as with the late 2000's Phenom II parts (particularly the Phenom II x6). It's only been since 2011 with Intel's release of Sandy Bridge and AMD's "Faildozer" that Intel became so absolutely dominant in the CPU space. ...unless you were talking about the ISA of x86, in which case you can largely ignore everything I just said. :P It is worth pointing out though that 64bit x86 (that is, "x86-64") is actually an AMD creation and therefore both Intel and AMD cross-license with each-other. But in other non-x86 related news, both AMD and Nvidia not only make ARM CPUs but are also part of the RISC-V Foundation. Last edited by Nintendo Maniac 64; 1st August 2018 at 21:50. |
|||
2nd August 2018, 19:04 | #816 | Link | |
Moderator
Join Date: Jan 2006
Location: Portland, OR
Posts: 4,770
|
Quote:
For example, all the strobing concerns mentioned above. That's not an issue in the bitstream format at all, just about encoder implementation. It'll get addressed with better psychovisual tuning via adaptive quantization. FWIW, my read on Golden Frames is that it started as an alternative way to get some of the advantages of b-frames and multiple reference frames while staying clear of IP around those technologies. |
|
2nd August 2018, 19:57 | #817 | Link | |
Registered User
Join Date: Jun 2013
Posts: 95
|
Quote:
__________________
https://github.com/MoSal |
|
2nd August 2018, 22:42 | #818 | Link |
Registered User
Join Date: Jan 2006
Posts: 41
|
Right, but there is --tune=psychovisual option already, which enables optimizations that improve visual and hurt metrics (using cdef-dist for RDO), but this was already implemented for libaom.
OTOH it doesn't even have inter prediction yet, which is the reason it performs so badly. |
|
|