high quality volume rendering GPU/CUDA engine Looking for a high quality volume rendering GPU/CUDA e

[stefan_banev]

Hello,

I’m looking for a high quality volume rendering GPU/CUDA engine. GPU is designed for 3D applications so it surely must outperform CPUs based 3d engine. I’ve been failed to implement a competitive volumetric ray tracing/casting via CUDA so I’m looking for such development other people did. I did find CUDA implementations but it was not better then my result and totally inferior to CPUs VR_ray-tracer I’ve played with. To be more specific, I’m looking for GPU/CUDA based volumetric ray-tracing/casting to outperform Fovia’s HDVR or at least to match it. The hardware platform can be ultra-high-end NVIDIA video card such as Quadro FX 5800 vs. dual Xeon W5580 WS. Data sets to render is x4 full body 12-bit CT data 512x512x2300 rendered concurrently/consecutively. The rendering scene is semitransparent film like skin tissue and opaque bones. High quality zoom-in/close-look-up without sub-volume re-sampling & on-the-fly modifiable transfer function are paramount requirements. Single projection plane 1584x1088 pixels or/and the same view field divided by 2x2 windows each for one of 4 datasets. Interactive quality is the primary competitive parameter, the final/static quality is the secondary one. The specific/relevant feedback is highly appreciated.

Thanks,
Stefan

[shawkie]

Hello,

I’m looking for a high quality volume rendering GPU/CUDA engine. GPU is designed for 3D applications so it surely must outperform CPUs based 3d engine. I’ve been failed to implement a competitive volumetric ray tracing/casting via CUDA so I’m looking for such development other people did. I did find CUDA implementations but it was not better then my result and totally inferior to CPUs VR_ray-tracer I’ve played with. To be more specific, I’m looking for GPU/CUDA based volumetric ray-tracing/casting to outperform Fovia’s HDVR or at least to match it. The hardware platform can be ultra-high-end NVIDIA video card such as Quadro FX 5800 vs. dual Xeon W5580 WS. Data sets to render is x4 full body 12-bit CT data 512x512x2300 rendered concurrently/consecutively. The rendering scene is semitransparent film like skin tissue and opaque bones. High quality zoom-in/close-look-up without sub-volume re-sampling & on-the-fly modifiable transfer function are paramount requirements. Single projection plane 1584x1088 pixels or/and the same view field divided by 2x2 windows each for one of 4 datasets. Interactive quality is the primary competitive parameter, the final/static quality is the secondary one. The specific/relevant feedback is highly appreciated.

Thanks,

Stefan

VGStudioMax 2.0 from Volume Graphics has GPU-based volumetric rendering but it does have a limitation. The entire dataset must fit in video memory at once and if the dataset is too large then it is silently re-sampled. For a 512x512x2300 volume I would expect that to be possible on the right video card (you would probably want the GTX 260 1.8GB or the GTX 285 2GB). I’m not sure about rendering 4 such volumes at once though - I think you’d probably be out of luck. Also its a (fairly expensive) commercial package so it might not be what you’re looking for but it might give you an idea of what can be done.

I’d be very interested to see an implementation where very large volumes could be kept in system memory and the onboard memory used as a massive cache. The new zero copy features in CUDA 2.2 might actually make this possible although I think you’d run into problems with datasets over 4GB.

[stefan_banev]

VGStudioMax 2.0 from Volume Graphics has GPU-based volumetric rendering but it does have a limitation. The entire dataset must fit in video memory at once and if the dataset is too large then it is silently re-sampled. For a 512x512x2300 volume I would expect that to be possible on the right video card (you would probably want the GTX 260 1.8GB or the GTX 285 2GB). I’m not sure about rendering 4 such volumes at once though - I think you’d probably be out of luck. Also its a (fairly expensive) commercial package so it might not be what you’re looking for but it might give you an idea of what can be done.

I’d be very interested to see an implementation where very large volumes could be kept in system memory and the onboard memory used as a massive cache. The new zero copy features in CUDA 2.2 might actually make this possible although I think you’d run into problems with datasets over 4GB.

I have made a request for the evaluation copy of the latest VGStudioMax 2.0 and still do not get the feedback. If you do not mind we may run a comparison for interactive image quality for typical single CT dataset 512x512x2300 (12bit) on any high-end computer HW you have access to VGStudioMax 2.0 and I will provide interactive images rendered by HDVR. Once you accept the challenge, we’ll specify the details and the result " images/fps/hardware-specs " can be posted here to let people see how good VGStudioMax 2.0 is. Anyone who would like to take the challenge are welcome as well (for any VR solution)… I’m looking for the best interactive quality VR solution available on market especially CUDA implementation as probably the most flexible GPU based VR solution I’m interested in.