PDA

View Full Version : Resize a bufferobject after transform feedback



ropel
07-04-2007, 02:25 AM
Hi all,

I have a transform feedback running. Problem is that I need to allocate a bunch of memory, which will definitely be sufficient for the (unknown amount of) values that I get back from the feedback.

Afterward I would like to reduce this amount of memory to fit the required size. How can I do that. Will glBufferSubData reallocate the memory?

Thanks

Zengar
07-04-2007, 08:53 AM
I don't think it is possible... You could create another buffer object and copy the data, but that would be very innefective. We need a server-side buffer copy :)

ropel
07-05-2007, 01:58 AM
We sure do. I'm very surpised that it is not possible. That would keep me with a buffer which is way to big. But I can't guess the amount of geometry the transform feedback will create.

So really, there is no reasonable solution for now?!

Rob Barris
07-05-2007, 11:42 AM
If you could copy the interesting data subset to a new buffer object efficiently, would that solve this problem ?

ropel
07-06-2007, 12:47 AM
Well, yes, that would be a solution. But actually I would like to reduce size to gain efficiency. However if the copying takes too much time/memory, than it wouldn't be worth the hassle.

Korval
07-06-2007, 11:01 AM
But I can't guess the amount of geometry the transform feedback will create.You can't? Not at all?

I mean, it is your shader doing the transforms. You're sending it vertex data. Unless the algorithm is non-deterministic, I don't see how you can't even make a reasonable estimation of the space it will take.

Brolingstanz
07-06-2007, 06:50 PM
yeah, you should definitely have some reasonable upper bound on this thing. And I would guess even feedback has some fairly hefty limitations on the size of the output anyway (haven't looked at the spec on it yet).

AlexN
07-06-2007, 08:56 PM
I don't see any limitations on the total size of the output with transform feedback. Whether or not a reasonable limit could be estimated in advance for this particular problem, a mechanism for direct copies between buffer objects would be nice to have.

Brolingstanz
07-06-2007, 09:03 PM
you're right, just had a peek. There is a limitation on the size of the vertex though, so if you have a bound on the number of verts, the total size probably isn't much of a stretch.

What I didn't see is something similar to DrawAuto in d3d10, where you can turn around and render from a buffer without having to query the number of primitives written. Don't know if that's worth considering, but it seems pretty handy.

Jan
07-07-2007, 01:35 AM
AFAIK Geometry Shaders have an output limit (though it is pretty high).

Haven't worked with GSes yet, one question i was asking myself: is it possible to "recursively" emit triangles? I mean, do the vertices, that are generated in the GS pass through the VS and the emitted triangles through the GS again, so that i could emit another triangle from a triangle, that was already generated earlier?

Or does only "original" geometry pass through the GS?

Jan.

AlexN
07-07-2007, 12:52 PM
You would have to issue a second draw call on the results of transform feedback to use the Geometry Shader recursively. I think the new ATI cards don't have a limit on the total number of output vertices per input primitive, and even the G80 can output at most 1024 floats per input so you can get quite a bit of data expansion.

Buffer object copies would be good! I'm sure all GPUs are quite capable of copying memory around quickly, and there's a hole in that we can do this for PIXEL_PACK and PIXEL_UNPACK but not for just copying between buffers.

Korval
07-08-2007, 04:33 PM
is it possible to "recursively" emit triangles?Nope. And it's probably a good idea to keep it that way (for the sake of hardware efficiency. Only Intel right now would have the hardware to put a stack in the geometry shader). Fortunately, recursive algorithms can be re-written iteratively, so it's not that big of a deal.