Direct3D feature levels discussion

From a hardware perspective I never thought it made much sense that UAV read/writes should only be possible in the pixel shader. It all executes on the same compute units.

But obviously there must have been a reason from at least one IHV to not allow it. My guess would be Intel.

 

It would be interesting to know if the NDA version of NVAPI exposes UAV capabilities for D3D11.x API family. What is for sure is that Geforce do not support TIR before Maxwell 2.0. If it was a "political" choice or an API design/implementation issue to not expose UAV caps on D3D11.x only Microsoft knows

Please, tell me moar D:

 

Unrelated to this specific feature, but there are restrictions in Tier 1/Tier 2 that do not correspond to anything from hardware that I know of from the "main three" desktop GPU vendors. Speculate away

I could be wrong/forgetting some obscure hardware of course, but that's less fun to talk about.

Why would you think it was Intel? Check the hardware docs - IIRC there are basically no restrictions around UAVs really - the write path for UAVs (data port) is quite separate from the ROP path so there's no esoteric restrictions on shader stages or counts, etc. Our current restrictions are more related to format conversions (i.e. typed UAVs) in that path since - as noted - it is separate from the ROP hardware.

GCN obviously doesn't have any issues here either, so my guess is NVIDIA on most UAV count/stage-related restrictions.

In any case there are non-hardware reasons for not having UAV accesses at other stages, namely it starts to expose implementation-specific side effects (ex. post transform vertex cache) and in the presence of that, there haven't been any "killer applications" for UAVs at other stages, other than debuggers which could already do it at all stages via driver paths IIRC. It's possible that they simply didn't expose it for API cost/benefit reasons.

 

No respect for qualcomm...

 

So true.

http://blogs.msdn.com/cfs-filesyste...es/00-00-01-20-72-images/1854.directx12v5.png

There are four lights vendors.

 

I see 5 lights.

 

Everyone always forget power-vr and ARM ) :

 

At least for the moment they aren't in any Windows devices. The only Windows ARM devices have been NVIDIA and Qualcomm powered, hence Qualcomm as the fourth Musketeer.

 

Yes, actually they are not in the consumer market.. But I can guess they are at least interested to DX12 too, at least ARM (and don't forget about VIA, it is still not dead xD ). Plus, there is still a Windows NT family for embedded and industrial systems running both ARM and x86..
It would be interesting to know what is Qualcomm doing. Actually it does not have any DX12 capable system on the market as far I know, moreover they have a VLIW architecture inherited from the ex-ATI/AMD mobile division, and I don't even know if Adreno supports virtual memory...

 

Adreno 400 probably isn't DX12 capable. But Adreno 500 certainly will be.

 

Intel GMA 500, 600, 3600 and 3650 were PowerVR based and were used with x86 ATOM CPUs. Desktop Windows (and desktop DirectX) was used with these.

 

Adreno 4xx is 11_1 and qualcomm is used in all windows phones. Guess what the next windows phone will have (with D3D12 support too I bet)?

 

They're not a desktop GPU vendor of course. The only point I'm making is that they are important enough to MS that they got a seat at the big kids table, as it were. So there may be a quirk or two in DirectX 12 and its feature levels that are designed to accommodate Qualcomm.

 

Are we sure that Adreno 400 Series supports tiled resources and virtual memory? Anyway, WP10 is expected AFTER Windows 10 for PC RTM, so some modification can still occur for mobile hardware (maybe some less hardware and shader features to reduce mobile GPU complexity and power requirements).

 

Well they claim to support "feature level 11_2" which I assume their marketing got confused and meant 11_1 with tiled resources (http://www.anandtech.com/show/8035/qualcomm-snapdragon-805-performance-preview). Perhaps Ryan could get some verification from qualcomm/ms whether or not they foresee adreno 4xx having DX12 drivers.

I'd be shocked if WP10 didn't have DX12 sooner rather than later (tbh I'd be a little shocked if it wasn't at launch). Although my personal experience with qualcomm and drivers have not been good so who knows.

 

It would have been a stupid move not to involve all the GPU manufacturers for the API design from the beginning. Mobile is more important than desktop nowadays. It doesn't matter that only a few mobile companies have DX12 mobile GPUs in the retail market right now, since DX12 itself is not even out yet. New generation of mobile GPUs will be released for the holiday season market flagship phones. And I am sure that the mobile GPU vendors have already started the hardware design of their next generation as well. It would benefit Microsoft to ensure that as many as possible support DX12.

It's not a hard requirement. For example last gen consoles had unified memory systems, with CPU virtual memory only. GPUs used pure physical memory adresses.