Firstly I did a quick test to check the speed increase of SIMD on when adding vectors of 4 floats. At first SIMD was actually taking longer, but that turned out to be a side effect of debug mode. Results were better in release mode:
to add 100,000,000 vectors:
0.797 seconds (normal)
0.328 seconds (SIMD)
SIMD was about 2.4 times faster.
Next I used SIMD to aid with the calculation of the trilinear interpolation of the volume data sample, as it involves quite a few operations to blend the 8 values together. In a test, a render of 44.3 seconds was reduced to 39.8 seconds. Not a massive reduction but certainly better than nothing.
This shows that SIMD speedups are certainly possible, and that my methods for using it are correct. After tomorrow's meeting I will see how well this works on the multi-core PCs in uni, and then make use of it in the actual rendering code, the aim being to have looked at it to a decent extent before next week.
In terms of the actual meeting I don't have a lot to discuss this week. I'm pretty happy with my understanding of SIMD and the results achieved so far so it's just a case of continuing as I am.
No comments:
Post a Comment