CAcarterschonwald·Dec 20, 2014
I can totally get struct of arrays format in haskell (Heck, the Unboxed.* modules in the vector package are basically that http://hackage.haskell.org/package/vector)
I can define the SOA format as something like (Vector Bool, Vector Bool, Vector a, Vector b), or something similarly along those lines. That allows a pretty reasonable branchless vectorized formulation with only a relatively small ~2x space overhead for most pointwise operations.
I actually wrote a prototype of a simd backed vectorized arithmetic lib for haskell that does the right thing over a year ago, https://github.com/wellposed/vector-vectorized, but the fact of the matter is that pointwise simd is honestly kinda boring, it really only is useful for streaming/linear time workloads. What more interesting is SIMD on super linear time algorithms like FFT and (Dense) Matrix
Read the rest
multiply, along with on compressed structures and audio/video. In all of these latter cases, SIMD shuffle matters to! And I've not yet seen a high level language that does a good treatment of supporting simd shuffles.
https://github.com/flame/blis/blob/master/kernels/x86_64/avx...
is a very very good example of how shuffle/permutation heavy interesting SIMD gets. (The associated project, BLIS, is a modern, amazingly well engineered sequel to blas, written in C, thats easy to customize to new architectures).
I definitely agree that GHC's current SIMD isnt very useful, it lacks nice support for shuffles, and its a bit weird (it naively just bundles up the LLVM simd api). I hope to work with a few folks to fix that up in time for GHC 7.12, I actually spent a bit of time this summer trying to plan out some design ideas.