![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Just to clarify texture bound is roughly analogous to memory bound. An ALU won't execute an instruction until the data has been fetched. Therefore there is a benefit to scheduling it alongside compute bound shaders. And yes pixel, vertex, and compute shaders run on the same hardware for the actual shaders. So while shader A is waiting on texture data to show up, Shader B is active and doing math that likely doesn't have to fetch as much. It's latency hiding when 100% occupancy isn't enough to cover your latency. As compute shaders don't use ROPs, they are thus free to start doing all sorts of interesting stuff. The ROPs MAY be an issue if you start forcing a ton of shadow maps through there. In most cases you will be constrained by how fast you can fetch data, which is what GPUs are designed to alleviate. Keep in mind the warp scheduler will be looking at every available warp to see which one is ready(data fetched) to be scheduled. This is where the interleaving happens. It's mixing shaders that aren't likely to be ready with ones that are.
The final step of any scene will be taking all those samples of the RT, doing any simple per pixel postprocessing, and outputting a simple one sample per pixel image. This can be blur, tone-mapping, compositing, etc. As for the inputs, you could bind your entire packed memory pool to some sort of array you could index in if you really wanted. This is still limited by the APIs, but for GCN you could create a stack and use the entire memory pool as a heap if you really wanted.
All I said was that overdraw will create even more threads than just your resolution. It should be avoided like you said, but it still occurs. You will almost never know how many pixel instances get created by the rasterizer. The amount of threads and in turn warps created
Pretty much confirms what I was saying.
CUDA has dynamic parallelism and DC does not. Along with a number of other features that are missing. Any game that uses CUDA is probably using Physx. So hardware vs software acceleration.
wrong thread.. here let me help you out, you obviously meant THIS thread: Gears of War PC: A complete disaster on GCN1.2 cards in Microsoft's own game + API debut
They're both related to fetching data from memory respective to their shader types, graphics and compute.. That has nothing to do with memory thrashing.
So you can add a bunch of effects and still finish the frame in the same amount of time... Sounds like free to me. As for the clocks I'll have to go check the Nvidia control panel. Last I checked the core and memory clocks were still adjustable.
Can we please not have this back and forth again for like the 6th or 7th time again? Please, PLEASE?
I use YELLOW, though sometimes I think that could be bad idea because links show up as Yellow. But at least you can see it.
ya it's borderlining ADHD or whatever combination of letters.
^ Damage Control is strong with this one. Go make razor 1 arnachist thread considering you have your head so far up your ******. Gaming Evolved is a counterpart of the way its meant to be played not GameWorst.
Like I said, there aren't any async on/off benchmarks for Hitman I could use. I used those numbers just because they were available to show what affect it would likely have in Hitman. How much of an effect depends on the amount of async used. I'd like to apply the same math to Hitman to see what it shows. Cranking the resolution and toggling async should give an idea of the total workload. That would be an interesting number to know.
Very interesting
Ty for sharing.
