-
Notifications
You must be signed in to change notification settings - Fork 0
Evaluation notes
We do not have any good measure of our productivity from the binomial pricer example. The Vector, Nikola, R and Accelerate implementations was performed by Rolf, Ken and Geoffrey. We thus only had to implement Repa and CUDA ourselves. In addition, Repa was almost a one-to-one mapping from the Data.Vector code by Ken, with some added "force" and inlining statements.
Even though we cannot compare our implementation productivity, we can say a bit about our experience from doing the CUDA implementation:
-
It was hard to comprehend compared to both the R code and Ken's code, as it contains a lot of memory access optimizations etc.
-
It was hard to debug, as the GPU-printf statement doesn't seem to print certain nested loops, or only printing in the first iteration of the loop (we are actually not quite sure what happens with printf's inside loops)