On 07.06.2004, at 20:04, William Skaggs wrote:

The main reason not to use convmatrix is that internally it always
does a 5x5 convolution, regardless of the matrix entries.  This means
it should take almost three times as long as the 3x3 convolution in
blur.c; in fact, a little testing on a 5000 x 10000 image shows it
taking over four times as long.  Otherwise using convmatrix would
probably be the right solution.

I agree with Sven here that convolution should be done by the core (probably even support SIMD, where available) and be usable for all plugins and integrated tools. IMHO it would also make sense to offer special functions which allow for 3x3, 4x4, 5x5 and maybe also generic sizes of matrices for optimum speed and cache utilization.


Attachment: PGP.sig
Description: This is a digitally signed message part

Reply via email to