I have a question on the eagle decomposition of hundreds of small metrics using the CUDA.
I need to compute the eigenvalues and eigenvectors of the small (64-by-64) hundreds of actual symmetric matrix (eg 500). I tried to apply it through the arrangement of the chess tournament by jessi method (see for more information).
In this algorithm, 32 threads are defined in each block, whereas each block handles a small matrix, and 32 threads work up to 32 convergence while expanding 32 closed-diagonal elements. However, I am not very satisfied with its performance
I am thinking that where the better algorithm for my question is, i.e. eigen-decomposition of 64-by-64-real symmetric matrices I think the householder's method may be a better option but it is not certain whether it can be implemented efficiently in the CUDA. Very useful information are not online, as most other programmers are more interested in using the CUDA / OpenCL, to dismantle a large matrix rather than a lot of small metrics.
For at least the eagentues, a sample QDA can be found in SDK < P>
Images look broken, but downloads of samples still work. I suggest downloading the full SDK and looking at that example. In addition, this paper may be useful:
Comments
Post a Comment