It is not surprising that basically everybody (of the respondents — who surely suffer from grave selection bias) is interested in multicore CPUs, but I’m somewhat surprised to see about 2/3 to be interested in GPGPU. The most popular application areas are data analytics, machine learning, and scientific computing with optimisation problems and physical simulations following close up.
The most important algorithmic patterns are iterative numeric algorithms, matrix operations, and —the most popular— standard aggregate operations, such as maps, folds, and scans. (This result most surely suffers from selection bias!)
I am very happy to see that most people who tried Repa or Accelerate got at least some mileage out of them. The most requested backend feature for Repa are SIMD instructions (aka vector instructions) and the most requested feature for Accelerate is support for high-performance CPU execution. I did suspect that and we really like to provide that functionality, but it is quite a bit of work (so will take a little while). The other major request for Accelerate is OpenCL support — we really need some outside help to realise that, as it is a major undertaking.
As far as extending the expressiveness of Accelerate goes, there is strong demand for nested data parallelism and sparse data structures. This also requires quite a bit of work (and is conceptual very hard!), but the good news is that PLS has got a PhD student working on just that!
NB: In the multiple choice questions permitting multiple answers, the percentages given by the Google Docs summary is somewhat misleading.