Path: blob/master/examples/example6/extra_information.txt
612 views
i7. HotFloorplan:12a) As mentioned above, one of the inputs to HotFloorplan is a3floorplan description file that has information about the4blocks to be floorplanned. The maximum aspect ratio constraint is set5to 1:3 except when a block has a bigger aspect ratio in the6base floorplan. In that case, the block's aspect7ratio itself forms the upper limit. Also,8the connectivity section of 'ev6.desc' currently lists 139major interconnects that we thought were important at the10architecture level. They are assumed to be of equal wire11density. Also, the wire_density field of the floorplan12description can be used to assign weights for the different13wires according to their contribution to performance. In14'ev6.desc', the assignment of these weights is uniform15across all wires.16b) HotFloorplan includes a first-order wire-delay model for17wires in the global and intermediate metal layers. It is18a simple stand-alone model that converts wire length to19wire delay and can be used in any tool, whether the thermal20model, floorplan model, or any other simulator. It is21adapted from Otten and Brayton, DAC'98, and related work.22The interfaces for the model are provided 'wire.[ch]'. An23appropriate process technology node can be chosen by using24the TECHNODE #define in 'wire.h'. After that, the model is25a simple call to the function 'wirelength2delay'.26c) HotFloorplan can be configured through many command line options.27The 'hotspot.config' file clearly lists them and what they do.28Among the options, following merit further discussion:29i) Annealing parameters:30The meaning of each of these can be easily obtained from any31standard text on VLSI CAD that deals with simulated32annealing (e.g.: Sarrafzadeh and Wong, "An Introduction to33VLSI Physical Design", McGraw-Hill 1996). The original34DAC '86 paper by Wong and Liu on floorplanning is a very good35starting point.36ii) Compaction ratio:37During floorplanning, empty spaces can arise in a floorplan.38Sometimes, these spaces are just an artifact of the discrete39nature of the aspect ratio function of the individual40blocks. Typically, about 50% of the maximum possible number41of dead spaces is due to this artifact. Such dead spaces42are very tiny and hence can be safely ignored without any43impact on the thermal model. In fact, ignoring such blocks44reduces the size of the problem input to the thermal solver.45Since the thermal model's solver algorithm is a higher order46polynomial function in the number of functional blocks,47ignoring the tiny blocks improves the performance of the48floorplanner significantly. The 'compact_ratio' option49sets a threshold on the area of such tiny blocks. The50default is 0.5% of the area of the encompassing 'parent51rectangle'. So, dead spaces that are smaller than 0.5% of52the parent rectangle are ignored during floorplanning. This53results in a maximum error in core area of about 0.46% for54our experiments. This setting, combined with the 'n_orients'55setting that determines the magnitude of discreteness of56the blocks' aspect ratio function, can be tuned to manage57the trade-off between the speed and accuracy of the58floorplanner. It is to be noted that in HotFloorplan, the59empty spaces that are not compacted are named in the form60_0, _1 and so on. It can be seen that the 'output.flp'61generated from the steps discussed above has blocks named62in that form.63iii) Weights for the metric (objective function):64HotFloorplan uses an objective function that is of the form65lambdaA * A + lambdaT * T + lambdaW * W where A, T and W are66the area, temperature and wire length respectively. The67weights lambdaA lambdaT and lambdaW can be set through68configuration parameters. Please note that for HotFloorplan,69A is in the order or hundreds of mm, T is in the order of70hundreds of Kelvin and W is in the order of tens of mm. In71combining them to a single metric, the weights have to be72assigned in a manner that not only takes into account the73desired importance of the variables A, T and W but also74compensates for the mismatch in their units of measurement.75The default weights in HotFloorplan are chosen based on this76principle. If one is interested in objective functions that77are not in the above-mentioned form, it is quite easy to78make the change in HotFloorplan. The design of the simulated79annealing algorithm is such that any arbitrary metric (objective80function) can be incorporated. In HotFloorplan, this is done81just by changing the return statement of the82'flp_evaluate_metric' function at the beginning of 'flp.c' into83an appropriate objective function of A, T and W.84858687