On 03/03/2011 05:46 AM, Mathias Gaunard wrote:
[snip] Imagine something similar to the mpn component of the GMP library, but with a generic and easy-to-use C++ interface. Through genericity, you could make the algorithms accept ranges of integers and have them work regardless of digit size, with the digit chosen at the discretion of the user. If the user provides a SIMD type as a digit, you'll get optimized SIMD code for free, for example, in a completely externalized way. Now *that* would be awesome.
This whole explicitly passing simd<T> in place of T, which seems to be the same thing done in nt2, in order to make use of certain optimizations seems like a poor interface choice. In almost all cases, the user logically cares about a container of T, and wants to work with individual elements, and the library should ensure that SIMD operations are used whenever possible. In particular, the design of the Eigen library, at least in this respect, seems more appropriate. Perhaps I am just misinterpreting the intended semantics of using simd<T>, but if it in fact it results in the logical value_type still being T, then the question is why you would even need to specify simd at all, rather than having it always be used. I also don't see how the optimization can be externalized in the (quite common) case of doing more than element-wise operations.