The flip network in staran

The flip ne~Nor k in each array A scramble/ unscramble network is required module of STARAN scrambles and unscrambles to scramble the data when it i s stored into m e m ­ multi-dimensional access (MDA) memory data. o ry and to unscramble the data when it i s r ead The flip network can permute data on transfers from memory. The flip network (Figure 1) does froln memory to PE 's, from PE's to memory, the scrambling an d tins c rambling a nd can also and from PE's to PE' s. Among the a llow able perform a numb e r of other useful per m utations. permutations a r e barrel shifts, barrel shifts on Bau er (Ref. 2) has SIlOWS how a numb er of d a t a substrings, and FFT-butterflies. The n e twork = anipulati ng functions can b o performe d using can be used for such data manipulations as shift ­ t he flip network with app~'opriate PE maSki ng. ing. mirroring (flipping end-for-end). irregula r Here, we show the Co"1I.s truction of the fl ip spr e ad ing, or com pressing and r eplica ting. network an d then a method of i r r gularly sp r e ad­ These manip ulations are useful for sorting. fast in~ a nd compressing data that is f a s te r th a n the Fourier t rans forms, imag e wa rping, a nd solving method shown in R e f. 2. partial d iffer enti a l equati ons on multi-me sh regi ons. Flip Network Construction Intr oduction Notation An earlier paper (Ref. 1) describes the A 2 n-item flip network has 2 n input-data­ multi-dimensional access (MDA) memories in lines label ed with n-bit binary vectors ranging 2 n STARAN. Memory data can be accessed (fete h e d from (00 ... 00) to (11. .. 11). It ha s output ­ or stored) by words, by bit-slices, by byte­ data-lines also l abe led with n-bit binary vectors. slices, etc. MDA memories are built with ordi­ The network has two control inputs: n a r y R AM chips, and data i s scrambled a cer­ tain way when stored in memory so that it can be access ed in various ways .