D3D11 stream output & compute shader
- Thread starter Novum
- Start date
I forgot about DrawIndirect which was added to DX11. See if that can do what you want. Here's one of the first Google results.
http://gpuexperiments.blogspot.com/2008/12/draw-indirect.html
http://gpuexperiments.blogspot.com/2008/12/draw-indirect.html
Why not write the size out during the stream out to an indirect buffer? then call compute indirect.
During a stream out you can easily add an atomic add to compute the size yourself and output [N, 1, 0, 0] to the indirect instance buffer (or computer version of it).
I have added debug line drawer to compute shaders and cuda this way, an atomic updates number of line, whilst streaming out vertex data (note I don't do it using the stream API so maybe i've overlooked something?)
Deano
During a stream out you can easily add an atomic add to compute the size yourself and output [N, 1, 0, 0] to the indirect instance buffer (or computer version of it).
I have added debug line drawer to compute shaders and cuda this way, an atomic updates number of line, whilst streaming out vertex data (note I don't do it using the stream API so maybe i've overlooked something?)
Deano
Buffer writes are only allowed in pixel and compute shaders as far as I know.
Actually I think your right, what an odd restriction...
Hmm tough problem. Here's a crazy question... is this output something that you could easily generate in a compute shader yourself, or are you using tessellation or something in the fixed-function pipeline to generate the triangle stream?
There are of course some hacks you could do with various performance impacts. For instance you could probably have your GS instead of emiting the primitive, write the primitive data to attributes and emit a single point primitive somewhere on the screen, then have the pixel shader append the triangle to an Append buffer (rather than a stream out buffer) and disable color writes. Your render target would be irrelevant and could probably just be 1x1. This definitely isn't 100% efficient but might work...
There are of course some hacks you could do with various performance impacts. For instance you could probably have your GS instead of emiting the primitive, write the primitive data to attributes and emit a single point primitive somewhere on the screen, then have the pixel shader append the triangle to an Append buffer (rather than a stream out buffer) and disable color writes. Your render target would be irrelevant and could probably just be 1x1. This definitely isn't 100% efficient but might work...
I would like to use the tesselation unit. Emulating that in a compute shader would certainly possible, but not very performant. Using the pixel shader is costly as I said, but I didn't yet try to disable color writes. But I doubt that will make much of a difference.
One thing I could try is to use D3D11_QUERY_SO_STATISTICS and double or triple buffering to hide the latency.
Well make sure to render point primitives - one per polygon that you want to output - and disable color writes. The key here is to just invoke a pixel shader 1:1 with the geometry shader (and try to avoid the rasterizer as much as possible) so you can use UAVs. Silly and not really the best performance but it should work.