Click here to load reader

Marek Perkowski Reversible Logic Synthesis with Garbage Bits Lecture 6

  • View
    217

  • Download
    1

Embed Size (px)

Text of Marek Perkowski Reversible Logic Synthesis with Garbage Bits Lecture 6

  • Slide 1
  • Marek Perkowski Reversible Logic Synthesis with Garbage Bits Lecture 6
  • Slide 2
  • Logic Synthesis for Quantum Pseudo- Binary Logic (Permutation Logic)
  • Slide 3
  • Background Reviev of Reversible logic This approach is mostly for quantum logic realization For optical and CMOS realizations the k*k assumption is not necessarily used
  • Slide 4
  • Notation for Fredkin Gates A 0101 CB PQR A 01 P BC QR A circuit from two multiplexers Its schemata This is a reversible gate, one of many C AC+B BC+AC C A B
  • Slide 5
  • Margolus Gate A 0101 CB PQR A 01 C A B ReversibleConservative 3*3 gate There are many similar gates
  • Slide 6
  • The 3 * 3 Toffoli gate is described by these equations: P = A, Q = B, R = AB C, Toffoli gate is an example of two-through gates, because two of its inputs are given to the output. Toffoli Gate PRQ * + ABC
  • Slide 7
  • * + + AB (b) PQ * + A B C PRQ * + ABC PRQ 0 1 Kerntopf Gate Feynman Toffoli Feynman, Toffoli and Fredkin gates are their own inverses
  • Slide 8
  • The Kerntopf gate is described by equations: P = 1 A B C AB, Q = 1 AB B C BC, R = 1 A B AC. When C=1 then P = A + B, Q = A * B, R = B, so AND/OR gate is realized on outputs P and Q with C as the controlling input value. When C = 0 then P = A * B, Q = A + B, R = A B. 18 different cofactors! Kerntopf Gate Review cofactors if students forgot
  • Slide 9
  • As we see, the 3*3 Kerntopf gate is not a one-through nor a two-through gate. Despite theoretical advantages of Kerntopf gate over classical Fredkin and Toffoli gates, so far there are no published results on realization of this gate. Is it better for no-ancilla synthesis? Kerntopf Gate
  • Slide 10
  • Logic Synthesis for Reversible Logic
  • Slide 11
  • Toffoli Fredkin A B C F1 D F2 Feynman How to build garbage-less circuits GARBAGE BIT 1 GARBAGE BIT 2 We can decrease garbage at the cost of delay and number of gates We create inverse circuit and add spies for all outputs 2 outputs 2 garbages width = 4 delay = 4
  • Slide 12
  • Toffoli Fredki n A B C D Feyn man To ffo li Fred kin A B C D Feyn man copy How to build garbage-less circuits A,B,C,D are original inputs inputs reconstructed This process is informationally reversible It can be in addition thermodynamically reversible F1 from spy F2 from spy 2 outputs no garbage width = 4 delay = 9
  • Slide 13
  • Observations We reduced garbage at the cost of delay and number of gates We were not able to reduce the width of the scratchpad register
  • Slide 14
  • 1. Minimize the garbage 2. Minimize the width of scratchpad register 3. Minimize the total number of gates 4. Minimize the delay Goals of reversible logic synthesis
  • Slide 15
  • Synthesis from (p)KFDDs Preprocessing
  • Slide 16
  • * + * + * + * + ... cici f1f1 f2f2 f3f3 f4f4 k1k1 k2k2 k3k3 k4k4 k5k5 k6k6 Previous levels next levels Other same level
  • Slide 17
  • * + * + * + * + ... f1f1 f2f2 f3f3 k1k1 k2k2 k3k3 k4k4 k5k5 k6k6 Previous levels next levels Other same level f4f4 cici
  • Slide 18
  • Example of converting a decision diagram to reversible circuit
  • Slide 19
  • Starting from Pseudo-Kronecker Functional Decision Diagram
  • Slide 20
  • fg pD nD S S S S S d a b c e 1 d d 1 0 1 1 0 0 1 01 1 0 1 0 (a) * + (b) * + d f G1 G2 g G3 e b c 0 0 1 1 a
  • Slide 21
  • Starting from function-driven Decision Diagram
  • Slide 22
  • S S S y1y1 y3y3 y2y2 0 1 0 1 1 1 0 0 (b) S S S x4x4 x3x3 x1x1 01 0 1 1 10 S x2x2 SS S 0 1 0 1 1 0 0 0 1 (a) y1 =x2 x3 x1x3 y2 =x3 x1 x1x4 y3 =x1 x4 x3x4 y4 =x4 Converting fDDs to reversible circuits Standard BDD Corresponding fDD of Kerntopf Nonlinear preprocessor
  • Slide 23
  • y1 =x2 x3 x1x3 y2 =x3 x1 x1x4 y3 =x1 x4 x3x4 y4 =x4 * + * + * + x1x1 x1 x1 x3x3 x2x2 y1y1 y3y3 y2y2 x4 x4 x4x4 x3 x3 G2 + x3x3 1 0 0 1 1 G1 h G3 Reversible circuit corresponding to fDD together with preprocessor
  • Slide 24
  • Open Problems: Is our set of mapping rules sufficient? (we have currently about 20 rules, but many more can be created). What is the practically best starting point for large functions? KFDD? PKFDD? fDD? How to transform the diagram or the circuit to improve the cost function? What to do in case of rule conflict? How to create rules to decrease garbage?
  • Slide 25
  • Use of Reversibility Observation: Every synthesis method can be executed forward and backwards Sometimes solution backwards can be simpler
  • Slide 26
  • A B C D 0 0 0 0 0 1 0 0 1 0 0 0 1 1 0 1 0 0 0 1 0 1 1 0 0 1 1 1 1 0 0 0 1 0 0 1 1 0 1 0 1 1 1 1 0 0 1 1 0 1 1 1 1 0 1 1 X Y Z V 0 0 1 1 1 0 1 1 0 0 1 0 1 0 0 0 0 1 1 1 0 0 0 1 0 1 1 0 1 1 1 0 0 0 1 1 1 0 1 0 0 1 1 1 0 1 0 1 1 1 0 0 0 1 0 0 Function F Function F -1 X Y Z V 0 0 0 0 0 1 0 0 1 0 0 0 1 1 0 1 0 0 0 1 0 1 1 0 0 1 1 1 1 0 0 0 1 0 0 1 1 0 1 0 1 1 1 1 0 0 1 1 0 1 1 1 1 0 1 1 A B C D 0 1 0 0 0 1 1 0 0 0 1 0 0 0 1 1 1 1 0 1 0 1 1 1 0 1 1 0 0 1 1 0 1 1 0 0 1 1 0 0 0 1 1 1 1 0 1 1 0 0 1 0 1 0 0 0 Forward Function Inverse Function Several methods exist to calculate inverse from forward function
  • Slide 27
  • X Toffoli Fredkin A B C Z Y D V Feynman Toffoli Fredkin A B C Z Y D V Feynman X Circuit for function F Circuit for function F -1 F*F -1 =I
  • Slide 28
  • Method of realization of F 1.1. Find function inverse F -1 of function F. 2.2. Synthesize F -1 using any method for reversible gates 3.3. Draw the schematics of F -1 from gates. 4.4. In the schematics replace every gate by its reverse and change the direction of signals. 5.5. The new schematics is the realization of F.
  • Slide 29
  • Fredkin A B C =Z V Y Z Z Y D V Feynman X X=G1 Feynman G2=Z V 0 X YZYZ YZYZ Y+ Z ZVZV YZYZ ZZ 0 Y Y Y Fan-out gate We were not able to find the best realization of F -1 shown earlier Another realization of F -1
  • Slide 30
  • Compositions and Decompositions
  • Slide 31
  • Composition Methods Composition methods with matching-based cell selection and complexity measures have been presented by: Dietmeyer Schneider Wojcik Michalski Jozwiak, Chojnacki and Volf DeMicheli library matching Kravets and Sakallah
  • Slide 32
  • Uses library of 1*1,2*2 and 3*3 cells For every 3*3 function, a ready solution cascade is stored - see the results of Kerntopf and Storme-DeVos For every reversible gate, all cofactors can be stored (constants and garbage) and used in NPN matching. Gates with more cofactors (like Kerntopf) are better
  • Slide 33
  • Can be applied to both classical reversible and quantum logic All intermediate functions calculated in terms of input variables Uses equivalence mapping transformations based on NPN-equivalence Synthesis from inputs to outputs and from outputs to inputs, backtracking and look-ahead strategies
  • Slide 34
  • How to select the cell, its alignment and constants? Select the gate that provides smallest complexity evaluation of the remaining functions and intermediate functions. This works for both forward and backward transformations I proposed PPRM for matching because it is easy to implement. Other representations and measures can be used for matching - De Micheli, Dietmeyer, Jozwiak. Solve exor equations to find inverse functions and input values. Sometimes Kmaps are easier to use.
  • Slide 35
  • 0 0 1 1 a + + a c b x y b c ab c ab a a Second stage Second stage of decomposition: Fredkin gate First stage First stage of decomposition: Feynman gate Decompositional synthesis of Toffoli Gate from Fredkin and Feynman Gates s t Third stage Third stage of decomposition: Feynman gate Using PPRM representation allows for NPN matching and good evaluation of the remaining function complexity Direction of synthesis
  • Slide 36
  • 0 0 1 1 a + + a c b x y b c ab c ab a a Second stage Second stage of composition: Reversible Expansion for Fredkin gate Third stage Third stage of composition: Feynman gate Compositional synthesis of Toffoli Gate from Fredkin and Feynman Gates s t First stage First stage of composition: Feynman gate NPN matching takes care of permutations and inverters Direction of synthesis
  • Slide 37
  • Composition From inputs to outputs What to do if the initial function is not reversible?
  • Slide 38
  • A B C 0 0 0 0 0 1 0 1 0 0 1 1 1 0 0 1 0 1 1 1 0 1 1 1 X Y Z 0 0 0 0 0 1 0 1 0 0 1 1 1 0 0 1 0 1 1 1 1 1 1 0 h=A B g=A*B G 0 0 - 1 0 - 0 1 - Toffoli Z=A*B Restrict to C=0 3. Toffoli gate assumed. 4. New functions X,Y,Z created. 5. C=0 restriction taken 6. Functions h and g of X,Y,Z can be now realizedComposition 1. These are inputs 2. These are outputs
  • Slide 39
  • Toffoli Feynman A B C X=A Y=B =0 Z=AB G =X=A g=Z=AB h=X Y=A B Created first Calculated in terms of new variables X,Y,Z and selected Feynman gate Compositional synthesis of non-reversible function of half-adder garbage IDEA: Look to outputs, modify near inputs, reduce the difference
  • Slide 40
  • Adder is composed from half-adders 2 garbage bits and one constant Compositional synthesis is here top-down, hierachical, like in classical CMOS
  • Slide 41
  • Toffoli Feynman A B C X=A Y=B =0 Z=AB G =X=A g=Z=AB h=X Y=A B Another variant assumes Kerntopf gate
  • Slide 42
  • Heuristics for finding simplest reversible function during composition CCN CN A B 0 A A B (Sum) A and B (Carry) Toffoli Feynman A B AB 0 AB 00 10 01 10 Mapping of a half-adder 00 01 11 10 This is what we would like to have But this is not reversible, A B A B A * B
  • Slide 43
  • CCN CN A B 0 A A B (Sum) A and B (Carry) Toffoli Feynman A B AB 0 Mapping of a half-adder AB 00 10 01 10 00 01 11 10 A B A B A * B The simplest way to separate them is to add variable A Heuristics for finding simplest reversible function during composition
  • Slide 44
  • CCN CN A B 0 A A B (Sum) A and B (Carry) Toffoli Feynman A B AB 0 AB 000 010 101 110 Mapping of a half-adder 00 01 11 10 This is what we would like to have But this is not reversible, because more outputs than inputs A * B A B A B A Heuristics for finding simplest reversible function during composition
  • Slide 45
  • CCN CN A B 0 A A B (Sum) A and B (Carry) Toffoli Feynman A B AB 0 AB C 001 011 100 111 000 010 101 110 Mapping of a half-adder Positive Davio Gate Now function is reversible of A,B,C arguments. But we still need to decompose it to known gates A * B C A B C A B A Heuristics for finding simplest reversible function during composition
  • Slide 46
  • Search is defined by: selecting function to be realized selecting a gate type selecting its forward or backwards matching selecting other signals to match selecting constant Garbage functions are becoming less and less undefined in the process of decomposition They can be used for synthesis We adopt classical AI search algorithms (depth-first, breadth-first, A*,best bound, etc) Search
  • Slide 47
  • Search cannot be avoided Trade-off between quality of solution and search time The only real improvement is possible through better selection heuristics and better implementation of cell library. Search
  • Slide 48
  • b 0101 a c Q g1 = a b + b c a R P 0101 1 c G2 = b + c + start forwardbackward H= c+b forward + G3 * c G4 Library matching cooperation of various methods Illustration of cooperation of various methods to design the Kerntopf gate from Fredkin, Toffoli, Feynman gates and inverters. One fan-out gate for input a, inverter for c and two fan-out gates for c are not shown.
  • Slide 49
  • Adaptations of functional decomposition methods
  • Slide 50
  • G H... G H G H Ashenhurst-like parallel serial New Curtis Curtis-like
  • Slide 51
  • During graph coloring of an incompatibility graph for nodes with minterms of bound set, two conditions are satisfied: there must be 4 colors every color must be used exactly 2 times. G H... We sacrifice this wire for garbage These two signals should be encoded using 4 symbols Bound set Curtis Decomposition
  • Slide 52
  • Probability of finding such coloring is increased by having more dont cares More dont cares are created by repeating variables in bound and free sets. This principle is the base of all decompositions of reversible functions and relations. Curtis Decomposition
  • Slide 53
  • Levelized Structures
  • Slide 54
  • Two-Dimensional Lattice Diagrams for reversible logic
  • Slide 55
  • Three Types of General Expansions f 01 A f0f0 f1f1 f and A f 0 and f 1 Forward Shannon This is a classical Shannon that you know
  • Slide 56
  • Three Types of General Expansions g,h and A g 1 A+h 0 A 10 1 0 g1g1 h A g 1 A+h 0 A Reverse Shannon (b) hoho g This is a reverse Shannon that you do not know Observe that this operator re-introduces the variable A in the middle, the variable that has been reduced in Shannon.
  • Slide 57
  • Three Types of General Expansions g 1 A+h 0 A A g h 10 g 0 A+h 1 A g, h, and A g 0 A+h 1 A and g 1 A+h 0 A Reversible Shannon This is a reversible Shannon that you do not know
  • Slide 58
  • 01 1010 - - 0 1 1 1 1001 f garbage g 1010 Z fg X garbage gf h i Y 1 garbage hfgfgh YZ X - - X YZ X 11 1 1 00 0 1 X 00 01 11 10 0101 Shannon lattice uses only half of each gate
  • Slide 59
  • - - 0 1 1 1 - - 1 0 - - X Y Z 1 0 0 1 1 1 - - X - - 1 1 - - - - 1 0 - - 1 - - - - 1 - - - 0 - - 0 - - - = 1 1 0 - - - - - - - - 0 1 - - =1=0=1=0 - - 1 - - - - - - 0 - - YZ This method is fully algorithmic Variable order selection problem Expansion type selection problem
  • Slide 60
  • Applications of levelized expansion method This method can be used to create arbitrary circuits, not only lattices Any constraint on layout size or shape can be imposed. last-resort approachThe expansion method can be used as the last-resort approach when other methods cannot find solution, in order to reduce the number of variables It introduces garbage and constants, but this is unavoidable when functions are not balanced When realizing multi-output functions start from those that are balanced or closest to balanced.
  • Slide 61
  • Conclusions New concepts: (1) Reversible Shannon Expansion for k*k binary Fredkin Gates (k>2) and generalizations - reversible decision diagrams, (2) Reversible Fredkin Lattice structures for logic based on binary Fredkin gates, (3) Generalized Composition-Decomposition Methods (4) Adaptations of Ashenhurst/Curtis decompositions. (5) Levelized expansions for Shannon and Davio-like gates. (6) adaptation of BDD-based, decomposition-based and technology mapping methods from standard binary logic
  • Slide 62
  • Research topics (1) Multiple-valued reversible gates (2) Multiple-valued quantum logic and synthesis methods (3) Fuzzy reversible logic and synthesis methods (4) Ashenhurst/Curtis-like decompositions of multi-valued reversible logic (5) Levelized expansions for multi-valued Shannon and Davio-like gates. (6) Decision Diagrams for binary and MV reversible logic (7) Genetic Algorithm combined with search for quantum logic (8) Regular structures for MV unate, symmetric, threshold and other functions (9) Visual Software End of Lecture 6