Shear Warp + SIMD

After finally sitting down and putting some thought into how to use SIMD with the Shear Warp algorithm, it doesn't seem that applicable. Unlike Ray Casting where there was additional processing going on, this algorithm basically consists of copying each voxel to it's pixel in the output image, handling the transparency value of the voxel.

The warp part of the algorithm involves a few calculations, however compared to generating the image it only adds a very small amount of time to the render, so I didn't feel it was worth spending time on.

I implemented some SIMD code to aid the calculations involved in writing a transparent voxel to the output image, but only succeeded in slowing things down.

I looked into processing 4 voxels at once. This would involve doing calculations for the red, green and blue of each pixel at the same time. This offers no benefits over processing single voxels though, as there are no common calculations happening to all four.

On Monday, I will start the report. My plan has altered slightly after the tutor meeting to include design sections for the two algorithms and the program. Also I missed out the Evaluation section.

Polishing

I spent today fixing various bugs and generally improving the program, to get it to the point where I am happy to call the coding finished. The only remaining thing is to try some SIMD on the Shear Warp algorithm over the next couple of days. I have been getting render times in the region of 0.05 seconds on some volumes, however transparency still causes slowdown so if I can speed up the general calculation with SIMD it will be helpful.

On Monday I will begin the report, and will ideally get around 4000 words written over the week, although it is hard to plan it out exactly.

I have written a list of the sections I intend to include in the report, to talk about with my tutor tomorrow.

Abstract
Introduction
Overview of Volume Rendering Techniques
Overview of Parallel Processing
Implementation of Ray Casting
- Basic
- Parallel Processing
- SIMD
Implementation of Shear Warp
- Basic
- Parallel Processing
- SIMD
Implementation of Rendering Program
Conclusion
References
Appendices
- Render time test results

RLE Fixed

My work so far tonight has seen some nice results.

Earlier I reported that enabled the RLE to work both forwards and backwards had slowed it down considerably. As it turns out I had simply broken it, which I have now fixed.



On the above image, a standard single-thread render took 1.32 seconds. With RLE enabled, it took 0.64 seconds. This is slower than previous times quoted for the skull, as this render contains a lot of non-empty voxels.

In order to improve this further I have added a setting where voxels that are surrounded by other solid voxels and therefore are not visible are considered empty, and not rendered at all. This lengthens the pre-processing of the volume, though not by an unacceptable amount. The rendering time for the above image was reduced further to 0.13 seconds.

I have not been able to test it on a multi threaded render yet.

I also added a linear interpolation sampling when warping and scaling the render to fill the screen. This helps to remove some artifacts and give a generally smoother (though blurry) appearance. Render time with this option enabled increased slightly to 0.15 seconds.