[vsipl++] More SSAR optimizations

[Jules Bergmann]

Don, This looks good, please check it in.  thanks, -- Jules

Don McCoy wrote:
> The attached patch splits the Kernel 1 processing class into two parts.  
> The new base class is responsible for most the setup that is applicable 
> to images with the same geometry.  Its constructor also computes the 
> dimensions of the final output image.  The benefit to the derived class 
> is that it can now pre-allocate the remaining memory needed for 
> processing, including the creation of the Fftm objects, which includes a 
> potentially lengthy planning process.
> 
> Also of note, this "pre-processing" phase allows two equations to be 
> reduced (at run-time that is) to simple multiplications, which can then 
> be vectorized by the SIMD unit.  See equations 62 and 68.  The setup for 
> these equations is expensive in part because they involve two 
> vector-matrix multiplies (one along the rows and one along the columns) 
> which results in a hard-to-optimize memory access pattern.  As this 
> portion is now done outside the computational loop, the cost is less of 
> an issue.  It should be possible to use the resulting matrices (that I'm 
> correctly calling 'filters') on any incoming radar data.
> 
> An explicit loop at eq. 65 was also removed.
> 
> The good news: These simple changes realized a 1.5x performance 
> improvement over the current (SVN head) version!
> 

Woo-hoo!



-- 
Jules Bergmann
CodeSourcery
jules at codesourcery.com
(650) 331-3385 x705