I was wondering just how realistic is it to process strings, rather than numerics, on the GPU? Specifically, what I'm interested in is using C++ AMP to perform comparisons between an array of strings and a target string.
I've started with basics such as passing a wchar_t* strings[] into a function, but it turns out that you cannot even create a view with a type smaller than an int!
So my question is - are there any best practices out there, or is this generally a bad idea? I'm also interested in things like warp divergence - for instance, how efficient would it be to compute string lengths on a large array?
You can work with chars in C++ AMP as per this blog post: http://blogs.msdn.com/b/nativeconcurrency/archive/2012/01/17/c-amp-it-s-got-character-but-no-char.aspx
IMO warp divergence is no different in string processing as it would be in other algorithms, so I wouldn't pre-worry about that aspect of things. First get it right, then get it fast, then tune it to be faster.
In September we will post on our blog a string processing sample that shows performance benefits with C++ AMP over a CPU multi-core implementation - staty tuned for that. In short, yes it can be worth offloading string manipulation algorithms to accelerators such as the GPU.