Geometry shader performance and usability?
Author
Since the geometry shader executes after the vertex shader, if there is a need to work on non-projection-transformed vertices, then either transform-feedback would need to be used, or the projection transform carried out in the geometry shader. I'm wondering what the impact on performance of either of these two options are. Is geometry shader execution of the same thing, such as vector-matrix multiply, at the same speed as it would be in a vertex shader? It would seem to me that the sequential nature of emitting vertices within a primitive in the geometry shader would limit how much that gets parallelized vs the vertex shader... Should one expect any real-world difference on recent hardware (GT280 etc)?
Which brings me to the question of what use the geometry shader really is, other than rendering to all faces of a cubemap in a single pass? Other things I can think of are better done with alternative approaches: tessellation by instancing or tesselators in DX11 hardware, marching cube isosurface extraction by histopyramids (significantly faster than marching cubes by geometry shader), etc. So what are best applications of geometry shaders?
That's one of the reasons why I was thinking that GS is too limited to be of any use. You can't output too many primitives.
I don't know about the GT280 but it was shown to be slow on the Gf8 which was the first generation to have GS.
And also, why include it in GL when tesselators and hull shaders and domain shaders are coming with DX11. They would need to deprecate GS soon.
I don't know about the GT280 but it was shown to be slow on the Gf8 which was the first generation to have GS.
And also, why include it in GL when tesselators and hull shaders and domain shaders are coming with DX11. They would need to deprecate GS soon.
Quote:Original post by V-man
And also, why include it in GL when tesselators and hull shaders and domain shaders are coming with DX11. They would need to deprecate GS soon.
Utter Rubbish.
The GS is going no where because it performs a completely different role; one of information gathering and transformation.
It can see surrounding vertices, post domain shader, allowing you to pass topological information down to pixel shaders.
The most changed stage with the addition of tesselation is the vertex shader which now does very little, effectively passing data through and some skeletal animation related work.
For tesselation of triangles/quads/control points and any adjustments required on that data. The "domain shader" takes on many of the tasks the vertex shader would have done when you perform tesselation.
Tessellation in the GS was always a poor hack until a better solution came around, as indicated by the limited of number of triangles you could produce as output.
The D3D11 pipeline is as follows;
[Input Assembler] => Vertex Shader => Hull Shader => [Tessellator] => Domain Shader => Geometry Shader => Pixel Shader => [Output merger]
Well the stages in square brackets are fixed function elements.
The IA deals with your input, the tesselator performs the vertex generation based on the data passed down and some tessellation factors and the OM is what does the blending etc of the final pixels.
Everything else is programmable via shaders;
- vertex deals with skeletal animation/vertex blending and pass through of data and no longer has to output data in projected space. (technically it didn't with the GS shader around as well but it was generally better to do so)
- hull generates control points for the domain shader stage (max of 32, can generate more, this is done per primative, can see all input data for a primative)
- domain takes the control points from the hull shader and the tessellators output point to create a new vertex to render. Domain shaders should really output projected vertex coords but this can still be left to the geo shader.
- Geometry deals with information and topology of triangles and can pass that data down to the pixel shader
- pixel shader does the final colour work/data output
Every part has its job to do, forunately you can leave out the parts you don't require at any stage.
Tessellation in the GS was always a poor hack until a better solution came around, as indicated by the limited of number of triangles you could produce as output.
The D3D11 pipeline is as follows;
[Input Assembler] => Vertex Shader => Hull Shader => [Tessellator] => Domain Shader => Geometry Shader => Pixel Shader => [Output merger]
Well the stages in square brackets are fixed function elements.
The IA deals with your input, the tesselator performs the vertex generation based on the data passed down and some tessellation factors and the OM is what does the blending etc of the final pixels.
Everything else is programmable via shaders;
- vertex deals with skeletal animation/vertex blending and pass through of data and no longer has to output data in projected space. (technically it didn't with the GS shader around as well but it was generally better to do so)
- hull generates control points for the domain shader stage (max of 32, can generate more, this is done per primative, can see all input data for a primative)
- domain takes the control points from the hull shader and the tessellators output point to create a new vertex to render. Domain shaders should really output projected vertex coords but this can still be left to the geo shader.
- Geometry deals with information and topology of triangles and can pass that data down to the pixel shader
- pixel shader does the final colour work/data output
Every part has its job to do, forunately you can leave out the parts you don't require at any stage.
Author
Quote:It can see surrounding vertices, post domain shader, allowing you to pass topological information down to pixel shaders.
Yes, but again, what are the specific applications?
In regards to the tessellation stage, is it possible to implement in such a framework (fixed stage tessellator) adaptive tessellation such that most tessellation is around object silhouettes (since that is where normal interpolation least hides limited mesh resolution, revealed by angular outlines)? Or is it not flexible enough?
Quote:Original post by Prune
Yes, but again, what are the specific applications?
Layers, TNB calculation, flat-shading normals, everything dealing with silhouette edges, per-triangle data, extrusion, view-dependent generation of primitives, transform-feedback, custom HQ antialiasing of edges.
...
...
Plus, the performance troubles were only on G80, because the internal geom cache was hilariously tiny. The output being sequential is no problem, as long as the data fits in the geom cache. A Mat4 * vec4 on a nVidia card takes 16 cycles.
Try this:
http://www.geeks3d.com/20100210/opengl-3-2-geometry-instancing-culling-on-gpu-demo/
[Edited by - idinev on March 23, 2010 4:25:11 AM]
Using a geometry shader, you can generate the 4 vertices for a particle on the GPU. I'd love to know how you can do that with the tesselator, since each particle is basically a single vertex..
I agree, the GS is here to stay.
I agree, the GS is here to stay.
Point sprites (a specific case of view-dependent generation of primitives) is a very common application of geometry shaders. Geometry shaders would also suit well for general sprites, but common APIs do not support this approach due to hardware penetration of pre-GS cards.
Quote:Original post by Ysaneya
Using a geometry shader, you can generate the 4 vertices for a particle on the GPU. I'd love to know how you can do that with the tesselator, since each particle is basically a single vertex..
I agree, the GS is here to stay.
You can actually expand single points to triangles with the tessellator too, as the output topology (triangle mesh or line list) has nothing to with the input topology (1...32 control points in D3D11). However, this is a bit like swatting flies with a bazooka.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement