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 18:34:15

Mathias Gaunard wrote:
> On 11/10/2010 17:54, DE wrote:
>> hi there
>> if you focus on simd-aware - and more specifically x86 simd -
>> implementation _DON'T_ read further
>> consider OpenCL as a general way to speed up computations (which uses
>> simd as one of backends as well as gpu shader units etc.)
>> i think it will be much more general and useful
> 
> OpenCL is a possible implementation, albeit we choose to call the
> various SIMD instructions ourselves to have more control on the
> toolchain and the end result.
> OpenCL will however be our main backend for GPU targets which we will
> be supporting in the future. Or maybe we will just target it through
> Clang 
> and LLVM.
> 
> NT2 (also know as the crazy frenchman library), upon which this effort
> is based, only supports x86 (SSE, ..., SSE4, AVX) and PowerPC
> (AltiVec). 
> ARM (NEON, VFP) is being added.
> An effort has been made in its design so that instructions could
> register themselves for a particular (type, cardinal) pair, all of
> which ranked according to a category to select the best candidate. It
> heavily 
> uses meta-programming, including rewritten bits of MPL to augment its
> compile-time performance.
> It also uses expression templates with Proto to detect certain
> patterns, 
> such as fused multiply-add, for which x86 is introducing new
> instructions soon.
> Therefore it is very tunable and extensible.
> 
> The bit I wanted to discuss here, however, is not the implementation,
> but rather the interface that the library provides to the programmer.
> 
> We aim to provide an interface in modern C++ that integrates well with
> the standard library and Boost in order to allow developers to make
> use 
> of SIMD in an easy and fairly high-level fashion, potentially using
> meta-programming to write an algorithm with parametric register and
> cache sizes.
> OpenCL is not an interface that satisfies those criteria.

Have you seen the ct thing coming from Intel?  I just learned about it last week.

http://software.intel.com/en-us/articles/intel-array-building-blocks/

It looks like a container library for vector processing in C++ with a JIT compilation runtime environment as part of the library.  As a guy who appreciates compilers, langauges and libraries there is a lot to like there.  This area is evolving very fast.

Regards,
Luke