On Wed, Feb 08, 2012 at 11:08:59PM -0500, al davis wrote: > What Felix is proposing is to do another ordering post- > expansion, where the post-expansion ordering would in effect > bring some subckt nodes down into their parent. I agree that it > should be available as an option, but there are disadvantages, > which I can go into on the developer list.
what i've implemented actually is, collect the incidence matrix from the subckt hierarchy (after expand and in the order you like best) and _then_ pass this matrix to a bandwidth-minimizer before allocating the matrix. needless to say that this needs some changes to iwant_matrix etc. but there were some other details that took me some time to figure out. In fact this post-expansion wastes some memory (at least n^2) and time (isnt bandwidth minimization \in XP?). but please tell me/us about other and more disadvantages. > In this case, the flat version is clearly faster, but I have > other examples where the hierarchical version is faster. > I'm curious about another variant .. to use subckts, but to > make the internal node an external node. I expect this should > run as fast as the flat version. I expect this to work in one example or another (most probably for the transmission line). in general the order and bandwidth still depends on the order the user writes down in the netlist. my goal was rcm (or any other heuristics). this is not about flat or hierarchical. it would be great to have a plugin interface for allocators in the end, what do you think? regards felix _______________________________________________ Gnucap-devel mailing list [email protected] https://lists.gnu.org/mailman/listinfo/gnucap-devel
