DirectSR is disappointingly unambitious

MfA

Legend
Instead of an API for upscaling TAA replacers, why not experiment a bit and make it suitable for view reprojection and asynchronous timewarp?

Just allow the API to take in the view matrix for frames and the desired reprojected frame, have it return a bitfield with "missing" pixels (filled in with best guess, but application might want to fill in with raytracing).

Even for non VR games ATW might be interesting. Like intel showed with ExtraSS. Blurbusters asked for this too.
 
Yeah I would like that honestly.
When I talked with AMD around FSR 3's launch I mentioned I would really like it for such a system to exist and they were intrigued by it.
 
I would like this too and I suspect something like this would be the next step. Work on DirectSR started some time ago when only Nvidia had frame generation and none except Meta did frame generation with extrapolation. Ideally DirectX needs to be a common api to target different IHV's. I think it's up to IHV's to be ambitious and bring out new features and then for DirectX to incorporate them if those features are well matured and adapted.
 
They have a console too, they should be partnering with Meta not just to bring out xbox branded Quest 3 but also for using Quest 3 with Xbox/PC rendering.

They should be implementing this themselves.
 
I'm sure Sony has an API and implementation for reprojection, why shouldn't Microsoft? To make it a bit more practical this is what I'm saying Microsoft should have done and still do.

- Make DirectSR just a general reprojection API, at the minimum this means allowing view matrix for supplied and requested frames and negative frametime to request extrapolated frames. Minimal changes required, because there is a huge overlap and these applications really belong in a single API.

- Make a built in implementation for ATW, preferably using NPU.

- Work with Meta to get Meta Quest Link (their OpenXR runtime) to use DirectSR for ATW. Other headsets should use it too, but they have a marketing deal with Meta.
 
Last edited:
Microsoft states "DirectSR enables multi-vendor SR through a common set of inputs and outputs, allowing a single code path to exercise DLSS Super Resolution, FidelityFX™ Super Resolution, and XeSS."

Is it correct to state IHV will have different upscaling training/inference models and the resulting data points fed into DirectSR as "inputs" will determine differences in quality/performance?
 
Microsoft states "DirectSR enables multi-vendor SR through a common set of inputs and outputs, allowing a single code path to exercise DLSS Super Resolution, FidelityFX™ Super Resolution, and XeSS."

Is it correct to state IHV will have different upscaling training/inference models and the resulting data points fed into DirectSR as "inputs" will determine differences in quality/performance?

The ability of the various upscalers to process those inputs will determine differences in quality. Question is whether some upscalers require more/better data than others to produce quality output.
 
The ability of the various upscalers to process those inputs will determine differences in quality. Question is whether some upscalers require more/better data than others to produce quality output.
FSR 2 has an optional but highly recommended option provide a "reactive mask" and a "transparency and composition mask". The other upscalers don't accept these inputs at all. DirectSR has the reactive mask and an "ignore history mask" (which is equivalent to FS2's "transparency and composition mask") as optional inputs. Time will tell whether developers bother providing these inputs to DirectSR and the quality of those inputs if they are provided.
 
DirectSR is just a generic API to do TAA drop in with upscaling support. TAAU/FSR/XeSS/DLSS (except 3.5) would all fit the API.

I was originally complaining about it not allowing extrapolation. I should also have complained that the TAA inputs are a very low common denominator, especially when there is significant sampling noise left in lighting. Denoising, anti-aliasing and super-resolution are all related tasks which most efficiently belong in the same algorithm.

Since Microsoft is allergic to extension mechanisms ... their APIs really need to be a bit more forward looking.
 
their APIs really need to be a bit more forward looking.
I guess the main problem with that is that for an API to become common standard the vendors have to agree to its specifications, and with AMD and arguably Intel being so far behind Nvidia in this the only way to make a forward looking API is to use whatever Nvidia needs for their DLSS-D ignoring the other vendors.
Which is probably a road MS doesn't want to take.
 
I guess the main problem with that is that for an API to become common standard the vendors have to agree to its specifications, and with AMD and arguably Intel being so far behind Nvidia in this the only way to make a forward looking API is to use whatever Nvidia needs for their DLSS-D ignoring the other vendors.
Which is probably a road MS doesn't want to take.

In an ideal world Microsoft would form an independent view on the shape of a forward looking upscaler API and not be limited to whatever the IHVs can handle today. The reason Nvidia drives so much of this stuff is that nobody else is doing the work to form an alternative opinion.
 
But wouldn't there be an issue here in that you could (and likely would) end up putting essentially unneccesary requirements on game developers if those inputs are deemed not useful going foward? We can't assume Microsoft themselves unilaterally know what the best direction forward is (if anything we'd likely assume they don't compared to other parties).

Providing these inputs I'd suspect is not "free" for game developers either in terms of workload or even possibly performance (depending on what ends up being required). If you just add a bunch of optional specifications then you go back to just having a fragmenting issue.
 
But wouldn't there be an issue here in that you could (and likely would) end up putting essentially unneccesary requirements on game developers if those inputs are deemed not useful going foward? We can't assume Microsoft themselves unilaterally know what the best direction forward is (if anything we'd likely assume they don't compared to other parties).

Providing these inputs I'd suspect is not "free" for game developers either in terms of workload or even possibly performance (depending on what ends up being required). If you just add a bunch of optional specifications then you go back to just having a fragmenting issue.

Well someone has to define what’s useful. Right now Nvidia is doing that simply by being the trailblazer. There’s already risk of fragmentation. There’s also risk that Nvidia has tunnel vision based on their hardware roadmap and may miss opportunities to improve upscaling tech. If folks aren’t happy with Nvidia setting the bar the only solution is for some other party to step up.
 
The way I understand it is DirectSR will only be a default upscaling methodology across all the different IHVs, so is more a general purpose approach for devs. Studios can also choose the direct implementation approach we see today if they see an added visual/performance benefit provided by any IHV trailblazers. Whether DirectSR is capable or not moving forward there is always another path towards implementing the latest advanced upscaling innovations.
 
The way I understand it is DirectSR will only be a default upscaling methodology across all the different IHVs, so is more a general purpose approach for devs. Studios can also choose the direct implementation approach we see today if they see an added visual/performance benefit provided by any IHV trailblazers. Whether DirectSR is capable or not moving forward there is always another path towards implementing the latest advanced upscaling innovations.
Yeah but this defeats the purpose of having DirectSR which becomes just another upscaling "solution" to support. I doubt that many games would go this way aside from those which will want to implement some new IHV exclusive tech.
 
Yeah but this defeats the purpose of having DirectSR which becomes just another upscaling "solution" to support. I doubt that many games would go this way aside from those which will want to implement some new IHV exclusive tech.

That kind of gets back to what @trinibwoy stated above about MS needs to have a independent, proactive view on moving forward and not tied to MS partnerships/PR priorities.
In an ideal world Microsoft would form an independent view on the shape of a forward looking upscaler API and not be limited to whatever the IHVs can handle today. The reason Nvidia drives so much of this stuff is that nobody else is doing the work to form an alternative opinion.

I can see the development/innovation cycle becoming stale if left up to MS so just grateful there are other implementation options. I don't expect MS to be leading the innovation forefront (ie, AutoHDR vs RTX HDR) as companies will have different strategic objectives and budget appropriately.
 
DirectSR now has support for FidelityFX Super Resolution 3.1 upscaling built-in.
Following our initial announcement of the DirectSR preview, we are pleased to share an update. DirectSR, our standardized super resolution (SR) API for D3D12 titles, now supports FSR 3.1 (upscaler-only), bringing enhanced upscaling capabilities to further improve your gaming experience. Learn more about FSR 3.1 here.
 
DirectSR now has support for FidelityFX Super Resolution 3.1 upscaling built-in.

It seems like MS doesn’t want to build their own upscaler.
 
It seems like MS doesn’t want to build their own upscaler.
They've dipped their toes into the field with AutoSR and might eventually add their own upscaler to the DirectSR runtime. It would be surprising if the next Xbox didn't support AI FSR4 or an AI upscaler of Microsoft's own design. But the obvious next step for DirectSR is to add support for frame generation and denoising; DirectSR won't bring much value to PC until that happens.
 
Back
Top