So Dx11 comes with 2 new buffers, giving a total of 3 buffers
- Constant Buffer
- Read/Write Buffer
- Read/Write Structured Buffer
The last 2 are the new ones, that are similar to Constant buffers, only that they can be written to in the Compute and Pixel shaders. Whats the purpose of the structured buffer though? is the setting up of the structure size purely a convenience factor? Because it would be just as easy to fill a normal buffer with the data, and do your own indexing in the shader.
If you dont want the ability to write, do the last 2 hold any benefit over constant buffers?
Can someone suggest some examples of where you would write to a buffer using a pixel shader?
Ive noticed that the geometry shader is actually after the tessellation stage. I find this odd. When would you ever need to create geometry after the mesh has been tessellated? surely would have made more sense to put it before.
DirectX 11 Questions
Hi.
There are even more new buffer types, look better. None of them are anything like constant buffers!!! Recall that constant buffers are very limited in size. All the new types are SRV/UAV, which simply means they are accessed by texture sampling units (or what's the proper name) as all the other textures and buffers. The difference is that you can now scatter (not only gather) to some of them. In another thread here, we mention an example of writing to a buffer in a PS and that would be for Bokeh (using AppendBuffer). Structured buffers are just convenience buffers and I'd say pretty neat.
The fact that the geometry stage is after tessellation is pretty logical, too. Tessellation actually doesn't really duplicate or "spawn" geometry, it just "refines" it (there is some topology involved) and a Domain Shader is simply just a kind of Vertex Shader! You cannot duplicate geometry for a cube map rendering using tessellation any easily, for example. That's why GS comes after DS and it doesn't matter where the input to GS comes from.
There are even more new buffer types, look better. None of them are anything like constant buffers!!! Recall that constant buffers are very limited in size. All the new types are SRV/UAV, which simply means they are accessed by texture sampling units (or what's the proper name) as all the other textures and buffers. The difference is that you can now scatter (not only gather) to some of them. In another thread here, we mention an example of writing to a buffer in a PS and that would be for Bokeh (using AppendBuffer). Structured buffers are just convenience buffers and I'd say pretty neat.
The fact that the geometry stage is after tessellation is pretty logical, too. Tessellation actually doesn't really duplicate or "spawn" geometry, it just "refines" it (there is some topology involved) and a Domain Shader is simply just a kind of Vertex Shader! You cannot duplicate geometry for a cube map rendering using tessellation any easily, for example. That's why GS comes after DS and it doesn't matter where the input to GS comes from.
Author
[color=#1C2837][size=2]There are even more new buffer types, look better.[/quote]
What are the others?
[color=#1C2837][size=2]The fact that the geometry stage is after tessellation is pretty logical, too. Tessellation actually doesn't really duplicate or "spawn" geometry, it just "refines" it (there is some topology involved) and a Domain Shader is simply just a kind of Vertex Shader! You cannot duplicate geometry for a cube map rendering using tessellation any easily, for example. That's why GS comes after DS and it doesn't matter where the input to GS comes from. [/quote]
The reason I find it odd, is that the general workflow for using a geometry shader was to add new geometry on the fly. It seems logical that you would want to add geometry in the geometry shader, and then further refine this with tessellation as necessary. Take for example the case where you wanted to transform points into random 3d shapes,[font="sans-serif"] something like Icosahedron's. If the tessellation was after the GShader, then these could be automatically further refined by the tessellation stage (which of course could still be done in the GShader). Im just trying to understand the benefit of having the tessellation before the GShader. [/font]
What are the others?[/quote]
-> http://msdn.microsof...v=VS.85%29.aspx
The reason I find it odd, is that the general workflow for using a geometry shader was to add new geometry on the fly. It seems logical that you would want to add geometry in the geometry shader, and then further refine this with tessellation as necessary. Take for example the case where you wanted to transform points into random 3d shapes, something like Icosahedron's. If the tessellation was after the GShader, then these could be automatically further refined by the tessellation stage (which of course could still be done in the GShader). Im just trying to understand the benefit of having the tessellation before the GShader. [/quote]
There is huge difference between GS and tessellation purpose. Although GS can be used to do tessellation algorithms, it can do MUCH more and cannot do many things as effectively (and massively) as a tessellator, on the other hand. GS can spawn new geometry of different types. Tessellation just "refines" geometry, that means that it adds new vertices/edges into the existing primitives. But AFAIK, there isn't a way of turning one triangle into two triangles that would not share their vertices in the topology using SM5 tessellation.
Still, if you want to achieve your workflow, then just first expand your points into icosahedrons (without a tessellation stage!!!) and then feed back your newly generated geometry to the tessellation stage for refinement/displacement/whatever. And then perhaps continue with yet another GS that'd "duplicate" them for each side of cubemap at once (or not). There will not be a great performance hit, all data will stay on GPU and the host will just issue two draw calls (pixels will be rasterised just once, of course, at the very end).
Constant buffers and the other buffer types are not really the same at all. Constant buffers are intended for small amounts heterogeneous data, regular buffers are intended for large amounts of homogeneous data.
Structured buffers are for when you have a buffer containing a user-defined structure of data that you'd like to look up by index. Without structured buffer implementing this would require lots of tedious and error-prone format conversions, unpacking, and address calculations in the shader.
pcmaster already mentioned that you can use an AppendStructuredBuffer in a pixel shader to push out data from a subset of your pixels, which I used in a sample to implement a bokeh effect using point sprites. Another example is AMD's order independent transparency demo, where instead of writing out pixel colors to a render target they used atomic operations on buffers to implement per-pixel linked lists.
The geometry shader is directly tied to both the stream out and rasterization stages, both of which require fully-formed primitive and not un-tessellated patches. Also most common and well-suited use cases for geometry shaders are generating fins, generating point sprites, and rendering geometry to multiple cube map faces/shadow map cascades in a single draw call. You would never want to do any of those things before tessellation. You also wouldn't want to just expand points to arbitrary geometry...expansion in a geometry shader can be very expensive and you want to minimize the number of output vertices as much as possible.
Structured buffers are for when you have a buffer containing a user-defined structure of data that you'd like to look up by index. Without structured buffer implementing this would require lots of tedious and error-prone format conversions, unpacking, and address calculations in the shader.
pcmaster already mentioned that you can use an AppendStructuredBuffer in a pixel shader to push out data from a subset of your pixels, which I used in a sample to implement a bokeh effect using point sprites. Another example is AMD's order independent transparency demo, where instead of writing out pixel colors to a render target they used atomic operations on buffers to implement per-pixel linked lists.
The geometry shader is directly tied to both the stream out and rasterization stages, both of which require fully-formed primitive and not un-tessellated patches. Also most common and well-suited use cases for geometry shaders are generating fins, generating point sprites, and rendering geometry to multiple cube map faces/shadow map cascades in a single draw call. You would never want to do any of those things before tessellation. You also wouldn't want to just expand points to arbitrary geometry...expansion in a geometry shader can be very expensive and you want to minimize the number of output vertices as much as possible.
One further convenience with the StructuredBuffer is that it can be used with the Append/Consume functionality with a whole structure. So if you have a particular data structure that you are using (as a particle state for example) then you can append and consume directly with complete structures instead of trying to manage the individual pieces of data.
In addition to the other points made about the geometry shader, don't forget that it can also reduce data as well as introduce it. After the tessellation is performed, if you want to cull unnecessary primitives before they get rasterized then the geometry shader can make the decision not to pass that primitive along. The GS can also change the topology type, so even if you tessellate triangles, then you can still convert them to lines or points if you want... It is one of the more flexible pipeline stages, and usually can be used for some unconventional and/or creative algorithms.
In addition to the other points made about the geometry shader, don't forget that it can also reduce data as well as introduce it. After the tessellation is performed, if you want to cull unnecessary primitives before they get rasterized then the geometry shader can make the decision not to pass that primitive along. The GS can also change the topology type, so even if you tessellate triangles, then you can still convert them to lines or points if you want... It is one of the more flexible pipeline stages, and usually can be used for some unconventional and/or creative algorithms.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
