speed expected : volume rendering in cuda

Accelerated Computing CUDA CUDA Programming and Performance
1

hi ,

i need to render a CT image using cuda. the display size is 128*100. what is the speed in frame/sec that i can expect for ray casting approach???  by searching this forum i get a feedback saying rendering in cuda is slower compared to OGL or other techniques. so before i implement i would like to know wht can I expect.

Thx,
– randal

2

frame/sec that i can expect for ray casting approach???

It 100% depends on what level of interactive quality you would like to sustain. Once you do not specify quality you may have any FPS 2D-output can provide.

Let consider the settings to obtain a high quality volumetric ray-tracing:

  1. Interpolation Classification (it’s one of essential requirement for High Quality)

  2. sampling density along ray 16+ -samples per cell

  3. ray-density 4-rayes per pixel on projection plane (bilinear can be used to get final pixel value)

  4. volume size let say 1024x1024x1024

  5. On the fly gradients to have nice Phong Lighting (at least 4 tri-linear interpolation per each sample-event along ray (see point 2 above))

  6. Multi-Transfer-Function support will require 4 more tri-linear interpolation to get interpolated classifications numbers such as opacity, red, green, blue.

  7. Altogether to render single cell of data for single pixel on screen requires 4x4x4x16 = 1024 tri-linear interpolations + 3 bilinear interpolations (see point 3 above)

If you are going to process each voxel of data projected on plane 128x100 pixels then the number of tri-linear interpolations == ( 1024 x 120 x 100 ) X 1024(data thickness) = 12.582.912.000. Well, the apparent optimization is to do an adaptive sampling according to its contribution and it is really where all complexity and all potentials are.

So, there are to answers to your question:

  1. Take number of 12-bln tri-linear interpolations and do the math according to video card you use and you will have a very optimistic assessment since memory performance is not taken into consideration so divide the FPS result by 10.

  2. Measure the performance of available high quality volume rendering solutions; and the threshold of interactive quality is indeed as high as I described.

Stefan