NVIDIA To Devs: Compute/Graphics Toggle Is A Heavyweight Switch

[Mr_Troll]

A couple of weeks ago, NVIDIA published of vademecum of DirectX 12 Do’s and Don’ts that went largely unnoticed. However, it actually contains some interesting information on the tips that NVIDIA gave to developers on how to best use Microsoft’s new lower level API with their existing architecture.

A couple of them, for instance, seem to confirm two stories we reported last month about Maxwell problems with Asynchronous Compute. In case you don’t recall, the reference is to AMD’s Robert Hallock saying that Maxwell can’t perform Async Compute without heavy reliance on slow context switching; a few days later, Tech Report’s David Kanter mentioned that according to Oculus employees, preemption context switching was potentially catastrophic for Maxwell GPUs.

Now, under the Pipeline State Objects (PSOs) section, they were very clear:

 

 

• Don’t toggle between compute and graphics on the same command queue more than absolutely necessary
• This is still a heavyweight switch to make

 

 

 

That’s not all they had to say about compute and graphics tasks – under the Work Submission – Command Lists & Bundles section, NVIDIA warned developers as follows:

 

 

 

• Check carefully if the use of a separate compute command queues really is advantageous
• Even for compute tasks that can in theory run in parallel with graphics tasks, the actual scheduling details of the parallel work on the GPU may not generate the results you hope for
• Be conscious of which asynchronous compute and graphics workloads can be scheduled together

 

 

 

Finally, NVIDIA also gave some advice on how to best use Maxwell and DirectX 12 hardware features. They recommend to use Conservative Rasterization, which right now is only available on Maxwell cards, while they are a bit more cautious about Raster Order Views, the other DX12_1 level feature.

 

• Use hardware conservative raster for full-speed conservative rasterization
• No need to use a GS to implement a ‘slow’ software base conservative rasterization – See https://developer.nvidia.com/content/dont-be-conservative-conservative-rasterization
• Make use of NvAPI (when available) to access other Maxwell features
• Advanced Rasterization features:
  Bounding box rasterization mode for quad based geometry
  New MSAA features like post depth coverage mask and overriding the coverage mask for routing of data to sub-samples
  Programmable MSAA sample locations
• Fast Geometry Shader features:
  Render to cube maps in one geometry pass without geometry amplifications
  Render to multiple viewports without geometry amplifications
  Use the fast pass-through geometry shader for techniques that need per-triangle data in the pixel shader
• New interlocked operations
• Enhanced blending ops
• New texture filtering ops
• Don’t use Raster Order View (ROV) techniques pervasively
• Guaranteeing order doesn’t come for free
• Always compare with alternative approaches like advanced blending ops and atomics

 

 

 

Source : http://wccftech.com/nvidia-devs-computegraphics-toggle-heavyweight-switch/

https://developer.nvidia.com/dx12-dos-and-donts

 

So this just confirms what we all know ... that maxwell and architectures before are not that good at DX 12 as AMD. Not surprising.

 

[Enderman]

nvidia knows how GPUs work better than the devs

which is why the devs should take and appreciate their help/advice

[HumbleMobo]

-snip-

Notice, Nvidias being quite as a church mouse about Pascal?

[zMeul]

oh noes!
not full DX12 hardware that can't do full DX12 specs  <_<

[Mr_Troll]

• Don’t use Raster Order View (ROV) techniques pervasively .... this is a DX 12.1 feature... so are they lying about DX 12.1 compliance aswell?

[themaniac]

i'm sure people will still say that nvidia will fix the performance for dx12 on maxwell and previous gens through drivers but news flash you cant fix a hardware problem with software

[Misanthrope]

nvidia knows how GPUs work better than the devs

which is why the devs should take and appreciate their help/advice

 

Nvidia knows how their GPUs work better and surely want to encourage them to code so they are not at such a disadvantage. They seem to be taking advantage of devs that might make this same assumptions to basically throw AMD under the bus and avoid devs coding for async compute giving their competitor an advantage.

[zMeul]

Don’t use Raster Order View (ROV) techniques pervasively .... this is a DX 12.1 feature... so are they lying about DX 12.1 compliance as well?

to note, these are optional features: Resource binding (three tiers), tiled resources (three tiers), conservative rasterization (three tiers), stencil reference value from Pixel Shader, rasterizer ordered views, typed UAV loads for additional formats, UMA/hUMA support.

Logical blend operations, double precision (64-bit) floating point operations, minimum floating point precision (10 or 16 bit).

and these are required features for 12_0: Resource Binding Tier 2, Tiled Resources Tier 2 (Texture2D), Typed UAV Loads (additional formats).

and 12_1: Conservative Rasterization Tier 1, Rasterizer Ordered Views.

[Bouzoo]

nvidia knows how GPUs work better than the devs

which is why the devs should take and appreciate their help/advice

 

The question is are they giving the instructions so their GPUs don't fall short of AMD's?

[Dabombinable]

oh noes!

not full DX12 hardware that can't do full DX12 specs  <_<

As per usual, the first generation of cards that support the latest DirectX API suck at actually supporting it.

[Stefan1024]

i'm sure people will still say that nvidia will fix the performance for dx12 on maxwell and previous gens through drivers but news flash you cant fix a hardware problem with software

If they use the marketshart to force all developers not to use featuers the nvidia cards can't handle well, it is somehow "solved".

And I'm sure they try to do so usong heavy PR.

[Kinda Bottlenecked]

 

• Don’t use Raster Order View (ROV) techniques pervasively .... this is a DX 12.1 feature... so are they lying about DX 12.1 compliance aswell?

 

I'm pretty sure its supported but probably really taxing to use it excessively. 

[Mr_Troll]

 

typo sorry...

 

[escape093]

NVIDIA published of vademecum of DirectX 12 Do’s and Don’ts that went largely unnoticed

 

This is literally my first time seeing this term.

[Jerakl]

This is literally my first time seeing this term.

 

Me too

[Notional]

Nvidia just sucks at async compute, and Kepler and Maxwell always will, no matter the driver updates they will get. People still don't understand that the efficiency of these these two architectures (especially Maxwell) came at huge compromises.

AMD's GCN cards are going to wreck most NVidia stuff in DX12 that is out right now. Based on certain powerpoint presentations, we might see Pascal suck at async compute too, but of course, that is speculation.

 

We know most devs are going to focus a lot on async compute, solely for the sake of consoles, which are GCN based, so we will probably see this efficiency translate to AMD, whereas we might see NVidia specific code paths, where async compute is limited or completely off. Maybe even using ROV or other 12.1 stuff, that could help NVidia a little bit.

 

Also this is hilarious:

 

• Don’t rely on being able to allocate all GPU memory in one go

• Depending on the underlying GPU architecture the memory may or may not be segmented​

 

That's the 970 right there. What a shit card, to require specific coding from devs, just because NVidia wanted to overcharge for defective chips (sure it's not a bug it's a feature, but what a shitty feature it is).

[Jerakl]

I don't understand.

 

heeeeeeelp

[WereCat]

So basicaly:

Do / Dont do this so that the game runs as best as possible on our NVIDIA cards.

 

I want to see Do and Dont to devs from AMD as well then.

[patrickjp93]

 

• Don’t use Raster Order View (ROV) techniques pervasively .... this is a DX 12.1 feature... so are they lying about DX 12.1 compliance aswell?

 

No, but they know they'll be the target of antitrust litigation if they don't give this warning since AMD has no support of 12.1 features.

[Enderman]

The question is are they giving the instructions so their GPUs don't fall short of AMD's?

it would probably increase performance on any gpu to follow those steps

its not like nvidia GPUs and AMD gpuswork completely differently...after all they both do the same work, render frames for a game

[Enderman]

Nvidia knows how their GPUs work better and surely want to encourage them to code so they are not at such a disadvantage. They seem to be taking advantage of devs that might make this same assumptions to basically throw AMD under the bus and avoid devs coding for async compute giving their competitor an advantage.

all the stuff listed is not "nvidia specific" or proprietary

you're just making up conspiracy theories lol

[ivan134]

[don_svetlio]

TL:DR - our compute sucks - please don't use it

[Misanthrope]

all the stuff listed is not "nvidia specific" or proprietary

you're just making up conspiracy theories lol

 

Not being nvidia means nothing if it performs a lot worst: The reverse would be AMD not recommending opengl, it's not Nvidia specific but their cards perform much worst with it so of course they wouldn't want devs to use that.

[MahtXL]

The question is are they giving the instructions so their GPUs don't fall short of AMD's?

pretty much yes.