On 14 May 2013, at 23:24, Sébastien Bourdeauducq <sebastien.bourdeaud...@lekernel.net> wrote:
> On 05/14/2013 09:34 PM, toby @ tobyz.net wrote: >> - This function has a sampler available to bring in arbitrary pixels from >> sources > > Hmm, that might work with a pipeline like this: > > "PFPU" generating addresses -> memory controller -> "PFPU" processing the > data -> video DAC > > plus a compiler that splits the code into address generation and processing > parts. Memory addresses that depend on memory content cannot work with an > architecture like that, but it should not be a problem for many graphics > processing algos :) Hmmm. The arbitrary sampling is there for - basics: resizing inputs, handling mismatched aspect ratios - advanced: display wall controller, soft edge blending, warping for projection mapping - crazy: art, ie. as d-fuse we do a lot of 2D cut-up and manipulation of sources live with a load of openGL patches I've written. And I'll be upfront that my interest is in what could be done with the M^3 platform and a software definable pipeline between multiple inputs and outputs. Which leads to the questions - is the present M1 implementation limited to 1:1 pixel mapping between inputs and output, and is there an alternative route you were intending to implement that doesn't have the pixel shader with sampler approach? > (Of course, "memory controller" is an oversimplification, we'll need > multi-ported prefetch caches and what not in order to obtain halfway decent > performance from the Slowtan-6. This paper describes an interesting solution: > https://graphics.stanford.edu/papers/texture_prefetch/) > > Sebastien > > _______________________________________________ > http://lists.milkymist.org/listinfo.cgi/devel-milkymist.org > IRC: #milkymist@Freenode _______________________________________________ http://lists.milkymist.org/listinfo.cgi/devel-milkymist.org IRC: #milkymist@Freenode