Hi,

I was wondering if anyone has interest or worked on implementing

statistics functions in CUDA. In particular various flavor of student

t-test http://en.wikipedia.org/wiki/Student’s_t-test

Brahim

Hi,

I was wondering if anyone has interest or worked on implementing

statistics functions in CUDA. In particular various flavor of student

t-test http://en.wikipedia.org/wiki/Student’s_t-test

Brahim

To compute those sort of descriptive statistical tests, you need to calculate the underlying population mean, variance, etc, which the GPU is pretty well suited to. The SDK contains an optimal implementation (memory bandwidth limited, which translates to over 100Gb/s on nVidia’s current fastest cards) of a generalized algorithm called a parallel reduction which can be pretty easily adapted to compute sums, variance, etc. on the GPU. From there it should pretty trivial to implement the final test on the CPU using the results of the reductions run on the GPU.

I think some1 was working on that “R” library or watever… dont remember any further details…

There is the “gputools” package for the R language, available from CRAN. There are

several data-mining utilities, as well as CUDA equivalents of the “lm”, “glm” and “qr”

commands - generalized linear modeling, linear modeling and q-r decomposition,

respectively. We are testing a new release, which should also be on CRAN in the

next few weeks.

If there is something you would really like to see, please feel free to post.

Regards,

MA