Re: [boost] Accelerating algorithms with SIMD - Segmented iterators and alternatives

Subject: Re: [boost] Accelerating algorithms with SIMD - Segmented iterators and alternatives
From: Simonson, Lucanus J (lucanus.j.simonson_at_[hidden])
Date: 2010-10-11 23:23:08

Jeremy Maitin-Shepard wrote:
> Recent versions of compilers are starting to support automatic loop
> vectorization, which can be utilized with minimal, if any, code
>   changes. I imagine that as a library author you are more inclined
> to look for a library solution to this optimization problem, whereas
> compiler authors 
> are more likely to try to implement it as part of the optimizer, but
> it 
> seems to me that implementing it in the compiler is generally
> preferable. 

Even a "good" vectorizing C++ compiler is pretty bad, practically worthless, and while research continues apace, the practical implications of the state of the art hasn't really changed for decades.  The simple fact is that most of the time the compiler just can't make enough useful assumptions about general C code to vectorize it safely.  People write SSE code by hand in assembly (calling intrisics is writing assembly by hand, but more verbose) or accept the scalar code the compiler generates.  Fortran is a different story, it is a language designed to make life easy for the compiler.  The compiler can hardly fail to vectorize fortran code, but other than calling fortran libraries from C interfaces that doesn't help us much.  If you just defined an object model that faithfully duplicates semantics of fortran and let people program that in C++ you could auto-vectorize it in the C++ compiler if it became part of the standard, but that would take years and cost millions of lives, its much easier to do it with a JIT runtime environment in a library.  Since we already have the compiler technology for languages that allow it to work well it seems best to implement such a language as a DSEL that allows us to use that compiler technology from C++ rather than try to implement a vector compiler in a template metaprogram or keep beating our heads against the vectorizing general C code brick wall.  JIT runtime is an end run around standardization and is kind of like going through Belgium to invade France.  You get to Paris without all the fuss.

Regards,
Luke