Please visit the cudart downloads link to get the *hopefully* bug fixed versions of cudart for CUDA v1.1 and v2.0. The v2.0 one also includes the viewport bug fix.
Although the previous versions worked fine on the majority of systems there were issues with any CUDA 2.0 toolkit install / new forceware drivers. Hopefully these are now sorted out. Please make sure you extract all the files from the zips before you run the .exe's.
Thanks again to Audun for his help on this one. Its nice to see that there are still people out there who know how to use a debugger in native mode.
Works excellent for me now. Steady 25-40fps.. I can't begin to tell you how awesome I think this is :D But I do wonder what it translates to with more realistic scenes (lots of polygons, less reflective surfaces perhaps) and larger resolutions.
ReplyDeleteI also think it would be cool to see how fast it can render scenes with global illumination etc.. This could also be nice as a demo of how cuda can accelerate renderers for software like Blender, 3D Studio Max too.
Now it's working. On NVS 290 I get 2.7-3.5 fps, and on GTX 280 I get 55-79 fps. It is a very nice and quick simulation.
ReplyDeleteI'm glad its working:)
ReplyDeleteActually I am very pleased to see that it is working on cards with different amount of multiprocessors. I have been trying to write the code so it would scale (according to the CUDA docs) but its nice to see that it actually does.
Your GTX280 result of 55-79 is very good news. On my 8800GT with 14 multiprocessors I get 25 to 30. As the GTX280 has 30 multiprocessors this indicates a pretty much linear scaling in speed. This seems to indicate my current implementation is compute bound rather than memory bound.
Having lots of polygons is still a problem. Even with my new acceleration structures its very easy to get caught into a non-coalesced read situation and the framerate falls dramatically. Having textured triangles is actually not at all expensive to ray trace on the nvidia hardware, after all they are pretty much designed to handle textures quickly. My tests with texturing the triangles in my raytracer confirm that there is only a tiny frame rate drop with it on.
Using cuda to accelerate traditional 3d tools would be an interesting exercise. I think you would really need to know how they place their triangle meshes in memory in order to customize a cuda implementation to suit them.