Unknown Soldier
Veteran
When is Mantle going to officially be released?
AMD Mantle API [updating]
- Thread starter MarkoIt
- Start date
Jawed
Legend
Yeah, just like other compute APIs.
Maybe but Mantle is built for more than just graphics, one of its strength is its ability to overlap compute and graphics more efficiently than other API
I'm pretty sure that uses Tiled Resources, not bindless. It's a DirectX game, right? There's no bindless in DirectX, and in GL it's an NVIDIA extension...
As for "gr", I'm guessing "graphics" as opposed to the compute and dma queues.
Using GigapixelLandscape Textures in Dragon Commander: Lessons Learned
http://twvideo01.ubm-us.net/o1/vault/gdceurope2013/Presentations/824810SwenVincke.pdf
http://twvideo01.ubm-us.net/o1/vault/gdceurope2013/Presentations/824810SwenVincke.pdf
http://www.amd.com/us/products/desktop/graphics/oem-solutions/Pages/amd-radeon-hd-r5-series.aspx#4
Was this already mentioned? AMD R5 (OEM) specs confirm Mantle-support coming to VLIW-based Radeons, too, apparently
Was this already mentioned? AMD R5 (OEM) specs confirm Mantle-support coming to VLIW-based Radeons, too, apparently
If thats accurate then its pretty big news. It means Mantle definitely isn't a GCN exclusive API as many have suspected and thus its less low level than some hoped bit more portable than some believed.
The rest of the product marketing on the page neglects to mention Mantle except for the GCN-based card or flat-out states it's for the R5 240 only. Every statement given by AMD thus far has kept it GCN-only.
An error in filling out the table seems like the best guess.
An error in filling out the table seems like the best guess.
It contradicts what was said at the AMD conference though... it wouldn't be surprising to see some subset of Mantle running on it (or really any GPU from the past little while) but there are details that I suspect might not work, although I'm not really an expert.
The VLIW bit is less likely to cause problems since they are apparently just using HLSL for now, but I'm not sure how much of the submission model stuff (dynamically-indexed descriptor sets, multiple async queues and sync primitives, etc) came with GCN.
I vaguely recall one of the presentations from the AMD event saying that the descriptor stuff wouldn't work as-is pre-GCN and that would make some amount of sense... by my understanding in GCN the scalar unit reads the relevant data for the texture unit and passes it from registers to the sampler. Not sure if/how that would be done pre-GCN but you would not want to blow vector instructions with it. Thus I'd assume the sampler read the relevant data directly from memory instead, but someone can correct me on that (doubtful I'll have time to read through those now-outdated docs).
Interesting if true, but I'm skeptical as well. At the very least they'd probably have hardware tiers for GCN vs. previous if that were true, in which case some of the simplicity benefits would be lost.
The VLIW bit is less likely to cause problems since they are apparently just using HLSL for now, but I'm not sure how much of the submission model stuff (dynamically-indexed descriptor sets, multiple async queues and sync primitives, etc) came with GCN.
I vaguely recall one of the presentations from the AMD event saying that the descriptor stuff wouldn't work as-is pre-GCN and that would make some amount of sense... by my understanding in GCN the scalar unit reads the relevant data for the texture unit and passes it from registers to the sampler. Not sure if/how that would be done pre-GCN but you would not want to blow vector instructions with it. Thus I'd assume the sampler read the relevant data directly from memory instead, but someone can correct me on that (doubtful I'll have time to read through those now-outdated docs).
Interesting if true, but I'm skeptical as well. At the very least they'd probably have hardware tiers for GCN vs. previous if that were true, in which case some of the simplicity benefits would be lost.
I skimmed the evergreen/cayman docs and it looks like resource descriptors still came from physical registers that were set using command buffer packets.
Ah interesting, so the shader <-> sampler interface sounds similar then, just I guess those registers in practice only get filled by some sort of pre-shader descriptor/memory streamer thing. I can't imagine they'd waste vector ALU slots loading that data... (all sort of academic at this point, but still curious heh).
Like EA commanded, everything else is on hold 'till DICE fixes the more immediate issues with BF4.
Mantle patch is their job, despite Frostbite-studio being independent now, so it's on hold too