RECENT progress in fabrication makes the practical application. Logic Synthesis for Quantum Computing. arxiv: v1 [quant-ph] 8 Jun 2017

Transcription

1 Logi Synthesis for Quntum Computing Mthis Soeken, Mrtin Roetteler, Nthn Wiee, nd Giovnni De Miheli rxiv:76.7v [qunt-ph] 8 Jun 7 Astrt Tody s rpid dvnes in the physil implementtion of quntum omputers ll for slle synthesis methods to mp prtil logi designs to quntum rhitetures. We present synthesis frmework to mp logi networks into quntum iruits for quntum omputing. The synthesis frmework is sed on LUT networks (lookup-tle networks), whih ply key role in stte-of-the-rt onventionl logi synthesis. Estlishing onnetion etween LUTs in LUT network nd reversile single-trget gtes in reversile network llows us to ridge onventionl logi synthesis with logi synthesis for quntum omputing despite severl fundmentl differenes. As result, our proposed synthesis frmework diretly enefits from the sientifi hievements tht were mde in logi synthesis during the pst dedes. We ll our synthesis frmework LUT-sed Hierrhil Reversile Logi Synthesis (LHRS). Input to LHRS is lssil logi network, e.g., represented s Verilog desription; output is quntum network (relized in terms of Clifford+T gtes, the most frequently used gte lirry in quntum omputing). The frmework offers to trde-off the numer of quits for the numer of quntum gtes. In first step, n initil network is derived tht only onsists of single-trget gtes nd lredy ompletely determines the numer of quits in the finl quntum network. Different methods re then used to mp eh singletrget gte into Clifford+T gtes, while iming t optimlly using ville resoures. We demonstrte the effetiveness of our method in utomtilly synthesizing IEEE omplint floting point networks up to doule preision. As mny quntum lgorithms trget sientifi simultion pplitions, they n mke rih use of floting point rithmeti omponents. But due to the lk of quntum iruit desriptions for those omponents, it n e diffiult to find relisti ost estimtion for the lgorithms. Our synthesized enhmrks provide ost estimtes tht llow quntum lgorithm designers to provide the first omplete ost estimtes for host of quntum lgorithms. Thus, the enhmrks nd, more generlly, the LHRS frmework re n essentil step towrds the gol of understnding whih quntum lgorithms will e prtil in the first genertions of quntum omputers. I. INTRODUCTION RECENT progress in frition mkes the prtil pplition of quntum omputers tngile prospet [], [3], [], [5]. However, s quntum omputers sle up to tkle prolems in omputtionl hemistry, mhine lerning, nd ryptonlysis, design utomtion will e neessry to fully leverge the power of this emerging omputtionl model. Quntum iruits differ signifintly in omprison to lssil iruits. This needs to e ddressed y design utomtion tools: A preliminry version of this mnusript hs een presented t the DAC 7 onferene []. M. Soeken nd G. De Miheli re with the Integrted Systems Lortory, EPFL, Lusnne, Switzerlnd. M. Roetteler nd N. Wiee re with Mirosoft Reserh, Redmond, USA. This reserh ws supported y H-ERC--ADG Cyer- Cre, the Swiss Ntionl Siene Foundtion (-698 MAJesty), nd the ICT COST Ation IC5. ) Quntum omputers proess quits insted of lssil its. A quit n e in superposition nd severl quits n e entngled. We trget purely Boolen funtions s input to our synthesis lgorithms. At design stte, it is suffiient to ssume tht ll input vlues re Boolen, even though entngled quits in superposition re eventully ted upon y the quntum hrdwre. ) All opertions on quits esides mesurement, lled quntum gtes, must e reversile. Gtes with multiple fnout known from lssil iruits re therefore not possile. Temporrily omputed vlues must e stored on dditionl helper quits, lled nille. An intensive use of intermedite results therefore inreses the quit requirements of the resulting quntum iruit. Sine quits re limited resoure, the im is to find iruits with possily smll numer of nille. Quntum iruits tht ompute purely Boolen funtion re often referred to s reversile networks. 3) The quntum gtes tht n e implemented y urrent quntum omputers n t on single or t most two quits [3]. Something s simple s n AND opertion n therefore not e expressed y single quntum gte. A universl fult-tolernt quntum gte lirry is the Clifford+T gte set [3]. In this gte set, the T gte is suffiiently expensive in most pprohes to fult tolernt quntum omputing suh tht it is ustomry to neglet ll other gtes when osting quntum iruit [6]. Mpping reversile funtions into networks tht minimize T gtes is therefore entrl hllenge in quntum omputing [7]. ) When exeuting quntum iruit on quntum omputer, ll quits must eventully hold either primry input vlue, primry output vlue, or onstnt. A iruit should not expose intermedite results to output lines s this n potentilly destroy wnted interferene effets, in prtiulr if the iruit is used s suroutine in lrger quntum omputtion. Quits tht nevertheless expose intermedite results re sometimes referred to s grge outputs. It hs reently een shown [8], [9], [] tht hierrhil reversile logi synthesis methods sed on logi network representtions re le to synthesize lrge rithmeti designs. The underlying ide is to mp sunetworks into reversile networks. Hierrhil refers to the ft tht intermedite results omputed y the sunetworks must e stored on dditionl nill quits. If the sunetworks re smll enough, one n lolly pply less effiient reversile synthesis methods tht do not require nill quits nd re sed on Boolen stisfiility [], truth tles [], or deision digrms [3]. However, stte-of-the-rt hierrhil synthesis methods minly

2 suffer from two disdvntges. First, they do not expliitly unompute the temporry vlues from the sunetworks nd leve grge outputs. In order to use the network in quntum omputer, one n pply tehnique lled Bennett trik [], whih requires to doule the numer of gtes nd dd one further nill for eh primry output. Seond, urrent lgorithms do not offer stisfying solutions to trde the numer of quits for the numer of T gtes. In ontrst, mny lgorithms optimize towrds the diretion of one extreme [], i.e., the numer of quits is very smll for the ost of very high numer of T gtes or vie vers. This pper presents hierrhil synthesis frmework sed on k-fesile Boolen logi networks, whih find use in onventionl logi synthesis. These re logi networks in whih every gte hs t most k inputs. They re often referred to s k-lut (lookup tle) networks. We show tht there is one-to-one orrespondene etween k-input LUT in logi network nd reversile single-trget gte with k ontrol lines in reversile network. A single-trget gte hs k-input ontrol funtion nd single trget line tht is inverted if nd only if the ontrol funtion evlutes to. The initil reversile network with single-trget gtes n e derived quikly nd provides skeleton for susequent synthesis tht lredy fixes the numer of quits in the finl quntum network. As seond step, eh single-trget gte is mpped into Clifford+T network. We propose different methods for the mpping. A diret method mkes use of the exlusivesum-of-produt (ESOP) representtion of the ontrol funtion tht n e diretly trnslted into multiple-ontrolled Toffoli gtes [5]. Multiple-ontrolled Toffoli gtes re speiliztion of single-trget gtes for whih utomted trnsltions into Clifford+T networks exist. Another method tries to remp single-trget gte into LUT network with fewer numer of inputs in the LUTs, y mking use of temporrily unused quits in the overll quntum network. We show tht neroptiml Clifford+T networks n e preomputed nd stored in dtse if suh LUT networks require suffiiently few gtes. The presented LHRS lgorithm is evluted oth on demi nd industril enhmrks. On the demi EPFL rithmeti enhmrks, we show how the vrious prmeters effet the numer of quits nd the numer of T gtes in the finl quntum network s well s the lgorithm s runtime. We lso used the lgorithm to find quntum networks for severl industril floting point rithmeti networks up to doule preision. From these networks we n derive ost estimtes for their use in quntum lgorithms. This hs een missing informtion in mny proposed lgorithms, nd rithmeti omputtion hs often not een expliitly tken into ount. Our ost estimtes show tht this is misleding s for some lgorithms the rithmeti omputtion ounts for the dominnt ost. Quntum progrmming frmeworks suh s LIQU i [6] or ProjetQ [7] n link in the Clifford+T iruits tht re utomtilly generted y LHRS. The pper is strutured s follows. The next setion introdues definitions nd nottions. Setion III provides the prolem definition nd gives orse outline of the lgorithm, 7 y y y x x x 6 x 7 x 9 x 8 x x x Fig.. A -fesile network with inputs, 3 outputs, nd 3 gtes. seprting it into two steps: synthesizing the mpping, desried in Set. IV nd mpping single-trget gtes, desried in Set. V. Setion VI disusses the results of the experimentl evlution nd Set. VII onludes. A. Some Nottion II. PRELIMINARIES A digrph G = (V, A) is lled simple, if A V V, i.e., there n e t most one r etween two verties for eh diretion. An yli digrph is lled dg. We refer to d (v) = #{w (w, v) A} nd d + (v) = #{w (v, w) A} s in-degree nd out-degree of v, respetively. B. Boolen Logi Networks A Boolen logi network is simple dg whose verties re primry inputs, primry outputs, nd gtes nd whose rs onnet gtes to inputs, outputs, nd other gtes. Formlly, Boolen logi network N = (V, A, F ) onsists of simple dg (V, A) nd funtion mpping F. It hs verties V = X Y G for primry inputs X, primry outputs Y, nd gtes G. We hve d (x) = for ll x X nd d + (y) = for ll y Y. Ars A (X G G Y ) onnet primry inputs nd gtes to other gtes nd primry outputs. Eh gte g G relizes Boolen funtion F (g) : B d (g) B, i.e., the numer of inputs in F (g) oinides with the numer of ingoing rs of g. Exmple : Fig. shows logi network of the enhmrk m85 otined using ABC [8]. It hs inputs, 3 outputs, nd 3 gtes. The gte funtions re not shown ut it n esily e heked tht eh gte hs t most inputs. The fnin of gte or output v G Y, denoted fnin(v), is the set of soure verties of ingoing rs: fnin(v) = {w (w, v) A} () For gte g G, this set is ordered ording to the position of vriles in F (g). For primry output y Y, we hve d (y) =, i.e., fnin(y) = {v} for some v X G. The vertex v is lled driver of y nd we introdue the nottion driver(y) = v. The trnsitive fn-in of vertex v V, denoted tfi(v), is the set ontining v itself, ll primry inputs tht n e rehed from v, nd ll gtes whih re on ny pth from v to the primry inputs. The trnsitive fn-in n e onstruted using the following reursive definition: {v} if v X, tfi(v) = {v} tfi(w) otherwise. () w fnin(v) 3

3 3 () s () Fig.. Reversile iruit for full dder using () single-trget gtes, () 3 Toffoli gtes nd 3 CNOT gtes, nd () Toffoli gte nd 6 CNOT gtes. Exmple : The trnsitive fn-in of output y 3 in the logi network in Fig. ontins {y 3,,,, 5, 9, 3, x,, x,, x 6, x 7, x 8, x 9, x, x }. The driver of y 3 is gte 3. We ll network k-fesile if d (g) k for ll g G. Sometimes k-fesile networks re referred to s k-lut networks (LUT is shorthnd for lookup-tle) nd LUT mpping (see, e.g., [9], [], [], [], [3]) refers to fmily of lgorithms tht otin k-fesile networks, e.g., from homogeneous logi representtions suh s And-inverter grphs (AIGs, []) or Mjority-inverter grphs (MIGs, [5]). Exmple 3: The logi network in Fig. is -fesile. C. Reversile Logi Networks A reversile logi network relizes reversile funtion, whih mkes it very different from onventionl logi networks. The numer of lines, whih orrespond to logil quits, remins the sme for the whole network, suh tht reversile networks re sdes of reversile gtes nd eh gte is pplied to the urrent quit ssignment. The most generl reversile gte we onsider in this pper is singletrget gte. A single-trget gte T ({x,..., x k }, x k+ ) hs n ordered set of ontrol lines {x,..., x k }, trget line x k+, nd ontrol funtion : B k B. It relizes the reversile funtion f : B k+ B k+ with f : x i x i for i k nd f : x k+ x k+ (x,..., x k ). It is known tht ll reversile funtions n e relized y sdes of singletrget gtes [6]. We use the opertor for ontention of gtes. Exmple : Fig. () shows reversile iruit tht relizes full dder using two single-trget gtes, one for eh output. Two dditionl lines, lled nill nd initilized with, re dded to the network to store the result of the outputs. All inputs re kept s output. A multiple-ontrolled Toffoli gte is single-trget gte in whih the ontrol funtion is (tutology) or n e expressed in terms of single produt term. One n lwys deompose single-trget gte T ({x,..., x k }, x k+ ) into sde of Toffoli gtes T (X, x k+ ) T (X, x k+ ) T l (X l, x k+ ), (3) where = l, eh i is produt term or, nd X i {x,..., x k } is the support of i. This deomposition of is lso referred to s ESOP deomposition [7], [8], [9]. ESOP minimiztion lgorithms try to redue l, i.e., the numer of produt terms in the ESOP expression. Smller s () s d x = = H T T T T d x () 3-ontrol Toffoli gte (8 T gtes). T () -ontrol Toffoli gte. d x V T T = S V S H V d x () 3-ontrol Toffoli gte (6 T gtes). Fig. 3. Mpping Toffoli gtes into Clifford+T networks. ESOP expressions led to fewer multiple-ontrolled Toffoli gtes in the deomposition of single-trget gte. If =, we refer to T (, x k+ ) s NOT gte, nd if = x i, we refer to T ({x i }, x k+ ) s CNOT gte. Exmple 5: Fig. () shows the full dder iruit from the previous exmple in terms of Toffoli gtes. Eh single-trget gte is expressed in terms of 3 Toffoli gtes. Positive nd negtive ontrol lines of the Toffoli gtes re drwn s solid nd white dots, respetively. Fig. () shows n lterntive reliztion of the sme output funtion, leit with Toffoli gte. D. Mpping to Quntum Networks Quntum networks re desried in terms of smll lirry of gtes tht intert with one or two quits. One of the most frequently onsidered lirries is the so-lled Clifford+T gte lirry tht onsists of the reversile CNOT gte, the Hdmrd gte, revited H, s well s the T gte, nd its inverse T. Quntum gtes on n quits re represented s n n unitry mtries. We write T to men the omplex onjugte of T, nd use the symol lso for other quntum gtes. The T gte is suffiiently expensive in most pprohes to fult tolernt quntum omputing [6] tht it is ustomry to neglet ll other gtes when osting quntum lgorithm. For more detils on quntum gtes we refer the reder to [3]. Fig. 3() shows one of the mny reliztions of the -ontrol Toffoli gte, whih n e found in [7]. It requires 7 T gtes whih is optimum [6], [3]. Severl works from the literture desrie how to mp lrger multiple-ontrolled Toffoli gtes into Clifford+T gtes (see, e.g., [6], [3], [33], [7]). Fig. 3() shows one wy to mp the 3-ontrol Toffoli gte using diret method s proposed y Breno et l. [3] Given free nill line (tht does not need to e initilized to ), it llows to mp ny multiple-ontrolled Toffoli gte into sequene of -ontrol Toffoli gtes whih n then eh e mpped into the optimum network with T -ount 7. However, the numer of T gtes n e redued y modifying the Toffoli gtes slightly. It n esily e seen tht the network in Fig. 3() is the sme s in Fig. 3(), sine the ontrolled S gte nels the ontrolled S gte nd the V gte nels the V gte. However, the Toffoli gte omined with ontrolled S gte n e relized using only T gtes [3], nd pplying the V to the Clifford+T reliztion nels nother 3 T gtes (see

4 Input : Logi network N, prmeters p Q, prmeters p T Output : Clifford+T network R set N lut_mpping(n, p Q); set R synthesize_mpping(n, p Q); 3 set R synthesize_gtes(r, p T); return R; Algorithm : Overview of the LHRS lgorithm. Fig. 3() nd [35], [7]). In generl, k-ontrolled Toffoli gte n e relized with t most 6(k ) T gtes. If the numer of nill lines is lrger or equl to k, then 8(k ) T gtes suffie [3], [7]. Future improvements to the deomposition of multiple-ontrolled Toffoli gtes into Clifford+T networks will hve n immedite positive effet on our proposed synthesis method. III. MOTIVATION AND PROBLEM DEFINITION A mjor prolem fing quntum omputing is the inility of existing hnd rfted pprohes to generte networks for sientifi opertions tht require resonle numer of quntum its nd gtes. As n exmple, the quntum liner systems lgorithm requires only (logil) quntum its to enode mtrix inversion prolem [36], [37], lerly demonstrting the dvntge tht n e gined y using quntum omputer. However, in prior pprohes, the reiprol step (/x) tht is prt of the lultion n require in exess of 5 quntum its. This mens tht rithmeti my dominte the numer of quits of tht lgorithm [38], diminishing the potentil improvements of quntum lgorithmi implementtion. Similrly, reent quntum hemistry simultion lgorithms n provide improved sling over the est known methods ut t the prie of requiring the moleulr integrls tht define the prolem to e omputed using floting point rithmeti [39]. While floting point ddition ws studied efore [], [], urrently networks do not exist for more omplex floting point opertions suh s exponentil, reiprol squre root, multiplition, nd squring. Without the ility to utomtilly generte iruits for these opertions it will e diffiult tsk to implement suh lgorithms on quntum omputer, to estimte their full osts, or to verify tht the underlying iruitry is orret. In this pper we tkle this hllenge nd ddress the following prolem: Given onventionl omintionl logi network tht represents desired trget funtionlity, find quntum iruit with resonle numer of quits nd numer of T gtes. The lgorithm should e highly onfigurle suh tht insted of single quntum iruit whole design spe of iruits with severl Preto-optiml solutions n e explored. Algorithm outline: Alg. illustrtes the generl outline of the lgorithm. The following susetions provide further detils. Input to the lgorithm is logi network N nd it outputs Clifford+T quntum network R. In ddition to N, two sets of prmeters p Q nd p T re provided tht ontrol detiled ehvior of the lgorithm. The prmeters will e introdued in the following setions nd re summrized in Set. VI-A; ut for now it is suffiient to emphsize the role of the prmeters. Prmeters in p Q n influene oth the x x x y 3 y x x x 3 5 () LUT network () Order:,, 3,, 5. 5 x x x y y x x x x x x 3 5 x x x y y () Reversile network. 5 (d) Order:,,, 5, 3. Fig.. Simple LUT network to illustrte order heuristis (dshed lines in the single-trget gtes men tht the line is not input to the gte). numer of quits nd T gtes in R, however, their min purpose is to ontrol the numer of quits. Prmeters in p T only ffet the numer of T gtes. The first step in Alg. is to derive LUT mpping from the input logi network. As we will see in Set. IV, one prmeter in p Q is the LUT size for the mpping whih hs the strongest influene on the numer of quits in R. Given the LUT mpping, one n derive reversile logi network in whih eh LUT is trnslted into one or two single-trget gtes. In the lst step, eh of the gtes is mpped into Clifford+T gtes (Set. V). It is importnt to know tht most of the runtime is onsumed y the lst step in Alg., nd tht fter the first two steps the numer of quits for the finl Clifford+T network is lredy known. This llows us to use the lgorithm in n inrementl wy s follows. First, one explores ssignments to prmeters in p Q tht led to desired numer of quits, prtiulrly y evluting different LUT sizes. This n e done y lling the first two steps of the lgorihm with different vlues for the prmeters in p Q. Afterwrds, one n optimize for the numer of T gtes y lling the lst step y smpling the prmeters for p T. IV. SYNTHESIZING THE MAPPING This setion desries how LUT mpping n e trnslted into reversile network. This is the seond step of Alg.. The first step in Alg. pplies onventionl LUT mpping lgorithms nd is not further explined in this pper. The interested reder is referred to the literture [9], [], [], [], [3]. A. Mpping k-luts into Single-trget Gtes Fig. illustrtes the generl ide how k-lut networks re mpped into reversile logi networks omposed of singletrget gtes with ontrol funtions with up to k vriles. Fig. () shows -LUT network with 5 inputs x,..., nd 3 x x x y y

5 5 5 LUTs with nmes to 5. It hs two outputs, y nd y, whih funtions re omputed y LUT 3 nd LUT 5, respetively. A strightforwrd wy to trnslte the LUT network is y using one single-trget gte for eh LUT in topologil order. The trget of eh single-trget gte is -initilized new nill line. The reversile iruit in Fig. () results when pplying suh proedure. With these five gtes, the outputs y nd y re relized t line 8 nd of the reversile iruit. But fter these first five gtes, the reversile iruit hs grge outputs on lines 6, 7, nd 9, indited y, whih ompute the funtions of the inner LUTs of the network. The iruit must e free of grge outputs in order to e implemented on quntum omputer. This is euse the result of the lultion is entngled with the intermedite results nd so they nnot e disrded nd reyled without dmging the results they re entngled with [3]. To void the grge outputs, we n unompute the intermedite results y re-pplying the single-trget gtes for the LUTs in reverse topologil order. This disentngles the quits, reverting them ll to onstnt s. Fig. () shows the omplete reversile iruit; the lst 3 gtes unompute intermedite results t lines 6, 7, nd 9. But we n do etter! One we hve omputed the LUT for primry output tht does not fn in to nother LUT, we n unompute LUTs tht re not used ny longer y other outputs. The unomputed lines restore tht n e used to store the intermedite results of other LUTs insted of reting new nill. For the running exmple, s shown in Fig. (d), we n first ompute output y nd then unompute LUTs nd, s they re not in the logi one of output y. The freed nill n e used for the single-trget gte relizing LUT 3. Compred to the reversile network in Fig. (), this network requires one quit less y hving the ext sme gtes. B. Bounds on the Numer of Anille As we hve seen in the previous setion, the order in whih LUTs re trversed in the LUT network nd trnslted into single-trget gtes ffets the numer of quits. Two questions rise: (i) how mny nille do we need t lest nd t most, nd (ii) wht is good strtegy? We will nswer the first question, nd then disuss the seond one. The exmple order tht ws used in the previous exmple leding to the network in Fig. () illustrtes n upper ound. We n lwys use one nill for eh LUT in the LUT network, postulted in the following lemm. Lemm : When relizing LUT network N = (X Y G, A, F ) y reversile iruit tht uses single-trget gtes for eh LUT, one needs t most G nill lines. The optimized order in Fig. (d) used the ft, tht one n unompute gtes in the trnsitive fn-in one of n output, one the output hs een omputed. This oservtion leds to lemm providing lower ound. Lemm : Given LUT network N = (X Y G, A, F ), let l = mx{# tfi(y) y Y } funtion synthesize_mpping(n = (X Y G, A, F ), p Q) set R empty reversile network; 3 set l ; initilize empty stk C; 5 initilize empty mp m; 6 set S ; 7 set D {driver(y) y Y }; 8 for g G do set r(g) d + (g) [g D]; 9 for x X do dd input line with nme x to R; set m(x) l; set l l + ; 3 end for g topo_order(g, p Q) do 5 set t request_onstnt(); 6 ppend T F (g) (m(fnin(g)), t) to R; 7 set m(g) t; 8 if r(g) = then 9 set S ; unompute_hildren(g); end end 3 for y Y do renme output of line m(driver(y)) in R to y; 5 end 6 return R; 7 funtion request_onstnt() 8 if C is not empty then 9 return C.pop(); 3 else 3 set l l + ; 3 return l; 33 end 3 funtion unompute_hildren(g) 35 for g fnin(g) G do set r(g ) r(g ) ; 36 for g fnin(g) suh tht r(g ) = do 37 unompute_gte(g ); 38 end 39 funtion unompute_gte(g) if g S then return; if g / D then set t m(g); 3 ppend T F (g) (m(fnin(g)), m(g)) to R; C.push(t); 5 set m(t) ; 6 set S S {g}; 7 unompute_hildren(g); Algorithm : Synthesizing LUT mpping into reversile network with single-trget gtes. e the mximum one size over ll outputs. When relizing the LUT network y reversile iruit tht uses single-trget gtes for eh LUT, we need t lest l nill lines. The lower ound inspires the following synthesis strtegy tht minimizes the numer of dditionl lines. One strts y synthesizing iruit for the output with the mximum one. Let s ssume tht this one ontins l LUTs. These LUTs n e synthesized using l single-trget gtes. Note tht ll of these re in ft needed, euse in order to unompute gte, the intermedite vlues of hildren need to e ville. From these l gtes, l gtes n e unomputed (ll exept the LUT omputing the output), nd therefore restores l lines

6 6 dditionl lines lower ound tul upper ound dditionl lines 5 5 lower ound tul upper ound LUT size () Mx: the tul numer of dditionl lines often mthes the upper ound; fter k = 9 inresing the LUT size hs no strong effet on the numer of dditionl lines LUT size () Adder: the tul numer of dditionl lines often mthes the upper ound; the dditionl lines derese lmost linerly when inresing the LUT size. dditionl lines lower ound tul upper ound dditionl lines lower ound tul upper ound LUT size () Divisor: the tul numer of dditionl lines often mthes the lower ound LUT size (d) Log: the upper nd lower ound re very lose to eh other. Note tht for k = 3 the optimum numer of dditionl lines is hieved, sine the funtion hs 3 inputs. Fig. 5. The plots show the upper nd lower ound ording to Lemms nd s well s the tul numer of dditionl lines when synthesizing different rithmeti enhmrks with LUT sizes rnging from 3 to 3. The x-xis shows LUT size nd the y-xis shows the numer of dditionl lines. whih hold onstnt vlue. We n esily see tht the ext numer of required lines my e it lrger, sine ll output vlues need to e kept. Note tht this strtegy unomputes ll LUTs in the trnsitive fn-in one of n output even if it is prt of fn-in one of nother output. Therefore, some LUTs will led to more thn two single-trget gtes in the reversile network. For good trdeoff etween the numer of quits nd T - ount one is interested in the minimum numer of quits suh tht eh LUT is trnslted into t most two single-trget gtes in the reversile network. Finding the minimum numer of nille under suh onstrints reltes to plying the reversile pele gme [] in minimum time using minimum numer of peles. More detils n e found in [3], [], [5], [6]. C. Synthesizing LUT Network Alg. desries in detil how k-lut network N = (X Y G, A, F ) is mpped into reversile network R tht onsists of single-trget gtes with t most k ontrols. The min entry point is the funtion synthesize_mpping (line ). This funtion keeps trk of the urrent numer of lines l, ville nille in stk C, LUT-to-line mpping m : G N tht stores whih LUT gtes re omputed on whih lines in R, nd visited list S (lines 3 6). The referene ounter r(g) heks for eh LUT g how often it is required s input to other LUTs. For driving LUTs, stored in D (line 7), the referene ounter is deresed y. It is initilized with the fn-out size nd llows us to hek if g is not needed ny longer suh tht it n e unomputed (line 8). This is the se whenever the referene ounter is, nd therefore the proess of unomputing is triggered y driving LUTs. Input lines re dded to R in lines 9 3. Input verties re mpped to their line in R using m. In lines single-trget gtes to ompute nd unompute LUTs re dded to R. Eh gte g is visited in topologil order (detils on topo_order follow lter). First, -initilized line t is requested (line 5). Either there is one in C or we get new line y inrementing l. Given t, single-trget gte with ontrol funtion F (g), ontrols m(fnin(g)) = {m(g ) g fnin(g)}, () nd trget line t is dded R (line 6). The LUT-to-line mp is updted ording to the newly dded gte (line 7). Then, if g is driving n output, i.e., r(g) = (line 8), we try to unompute the hildren reursively y lling unompute_hildren (line ). In tht funtion, first the referene ounter is deremented for eh hild g tht is not primry input (line 35). Then, eh hild g tht fterwrds hs referene ount of, is unomputed using unompute_gte (line 37). In there, first it is heked whether the hild hs lredy een

7 7 k-lut (AIG) AIG extrt BDD extrt ESOP ESOP minimiztion ESOP ESOP-sed synthesis Reversile network Clifford+T mpping Clifford+T network () Diret mpping. k-lut (AIG) -LUT mpping -LUT network enough nill not enough nill for eh -LUT -LUT AN-lssifition AN-lss DB lookup diret mpping of k-lut network opt. Clifford+T network Clifford+T network () LUT-sed mpping. Fig. 6. Algorithms to mp single-trget gte into Clifford+T network. visited (line ). If the hild is not driving n output, singletrget gte to unompute the line is dded to R (line 3). The freed line t is dded to C (line ) nd the mpping is lered ordingly (line 5). The hild is stored s visited (line 6), nd the funtion reurs (line 7). After ll gtes hve een omputed, outputs re dded to lines in R (lines 3 5). This proedure is simplified: two or more outputs my shre the sme driving LUT. In this se, one needs dditionl lines nd opy the output result using CNOT gte. With given topologil order of LUTs, the time omplexity of Alg. is liner in the numer of LUTs. As seen in the eginning of this setion, the order in whih LUTs re visited hs n effet on the numer of quits. Therefore, there n e severl strtegies to ompute topologil order on the gtes. This is hndled y the funtion topo_order tht is onfigured y prmeter in p Q. Besides the defult topologil order implied y the implementtion of N (referred to s topo_def in the following), we lso implemented the order topo_tfi_sort, whih is inspired y Lemm : ompute the trnsitive fnin one for eh primry output nd order them y size in desending order. The topologil order is otined using depth-first serh for eh one y not inluding duplites when trversing one. D. The Role of the LUT Size As n e seen from previous disussions, the numer of dditionl lines roughly orresponds to the numer of LUTs. Hene, we re interested in logi synthesis lgorithms tht minimize the numer of LUTs. In lssil logi synthesis the numer of LUT-inputs k needs to e seleted ording to some trget rhiteture. For exmple in FPGA mpping, its vlue is typilly 6 or 7. But for our lgorithm, we n use k s prmeter tht trdes off the numer of quits to the numer of T gtes: If k is smll, one needs mny LUTs to relize the funtion, ut the smll numer of inputs lso limits the numer of ontrol lines when mpping the single-trget gtes into multiple-ontrolled Toffoli gtes. On the ontrry, when k is lrge, one needs fewer LUTs ut the resulting Toffoli gtes re lrger nd therefore require more T gtes. Further, sine for lrger k the LUT funtions re getting more omplex, the runtime to mp single-trget gte into multiple-ontrolled Toffoli gtes inreses. To illustrte the influene of the LUT size we performed the following experiment, illustrted in Fig. 5(). For four enhmrks nd for LUT sizes k from 3 to 3, we omputed LUT mpping using ABC s [8] ommnd if -K k -. The resulting network ws used to ompute oth the upper nd lower ound on the numer of dditionl lines ording to Lemms nd, nd to ompute the tul numer of lines ording to Alg. with topo_def s topologil order. It n e noted tht the tul ound often either mthes the upper ound or the lower ound. In some ses the ounds re very lose to eh other, leving not muh flexiility to improve on the numer of dditionl lines. Further, fter lrger LUT sizes the gin in reduing the numer of lines dereses when inresing the LUT size. It should e pointed out tht for enhmrk Log n optimum numer of dditionl lines n e hieved for k = 3, euse in this se k mthes the numer of inputs of the funtion. Consequently, the LUT mpping hs s mny gtes s the numer of outputs. V. MAPPING SINGLE-TARGET GATES For the following disussion it is importnt to understnd the representtion of the logi network tht is given s input to Alg. nd the k-lut network tht from results the first step. The input network is given s gte-level logi network, i.e., ll gtes re simple logi gtes. In our experimentl evlution nd urrent implementtion the logi network is given s AIG, i.e., logi network omposed of AND gtes nd inverters. The LUT network is represented y nnotting in the gtelevel netlist (i) whih nodes re LUT outputs nd (ii) whih nodes re LUT inputs for eh LUT. As result, the funtion of LUT is impliitly represented s sunetwork in the gtelevel logi network. A. Diret Mpping The ide of diret mpping is to represent the LUT funtion s ESOP expression, optimize it, nd then trnslte the optimized ESOP into Clifford+T network. Fig. 6() illustrtes the omplete diret mpping flow. As desried ove, LUT funtion is represented in terms of multi-level AIG. In order to otin -level ESOP expression for the LUT funtion, one needs to ollpse the network. This proess is lled over extrtion nd two tehniques lled AIG extrt nd BDD extrt will e desried in this setion. The numer of produt terms in the resulting ESOP expression is typilly fr from optiml nd is redued using ESOP minimiztion. The optimized ESOP expression is first trnslted into reversile network with multiple-ontrolled Toffoli gtes s desried in

8 8 Setion II-D nd then eh multiple-ontrolled Toffoli gte is mpped into Clifford+T with the mpping desried in [7]. ESOP over extrtion: There re severl wys to extrt n ESOP expression from n AIG tht represents the sme funtion. Our implementtion uses two different methods. The hoie of the method hs n influene on the initil ESOP expression nd therefore ffets oth the runtime of the lgorithm nd the numer of T gtes in the finl network. The method AIG extrt omputes n ESOP for eh node in the AIG in topologil order. The finl ESOP n then e red from the output node. First, ll primry inputs x i re ssigned the ESOP expression x i. The ESOP expression of n AND gte is omputed y onjoining oth ESOP expressions of the hildren, tking into onsidertion possile omplementtion. Therefore, the numer of produt terms for the AND gte n e s lrge s the produt of the numer of terms of the hildren. The finl ESOP n e preoptimized y removing terms tht our twie, e.g., x x x x =, or y merging terms tht differ in single literl, e.g., x x x = x x [9]. AIG extrt is implemented similr to the ommnd &esop in ABC [8]. We were le to inrese the performne of our implementtion y limiting the numer of inputs to 3 its, whih is suffiient in our pplition, nd y using ue hshing [7]. The method BDD extrt first expresses the LUT funtion in terms of n AIG, gin y trnslting eh node into BDD in topologil order. From the BDD Pseudo-Kroneker expression [8], [9] is extrted using the lgorithm presented in [5]. A Pseudo-Kroneker expression is speil se of n ESOP expression. For the extrted expression, it n e shown tht it is minimum in the numer of produt terms with respet to hosen vrile order. Therefore, it provides good strting point for ESOP minimiztion. In our experiments we notied tht AIG extrt ws lwys superior nd BDD extrt did not show ny dvntge. BDD extrt my e dvntgeous for lrger LUT sizes. ESOP minimiztion: In our implementtion we use exorism [9] to minimize the numer of terms in the ESOP expression. Exorism is heuristi minimiztion lgorithm tht pplies lol rewriting of produt terms using the EX- ORLINK opertion [5]. In order to introdue this opertion, we need to define the nottion of distne. For eh produt term vrile n either pper s positive literl, s negtive literl, or not t ll. The distne of two produt terms is the numer of vriles with different pperne in eh term. For exmple, the two produt terms x nd x x hve distne, sine x does not pper in the first produt term nd ppers s negtive literl in the seond one. It n e shown tht two produt terms with distne k n e rewritten s n equivlent ESOP expression with k produt terms in k! different wys. The EXORLINK-k opertion is proedure to enumerte ll k! replements for produt term pir with distne k. Applying the EXORLINK opertion to produt term pirs with distne of less thn immeditely leds to redution of the numer of produt terms in n ESOP expression. In ft, s desried ove, suh heks re lredy pplied when reting the initil over. Applying the EXORLINK- opertion does not inrese the numer of x x x 5 Fig. 7. Reversile network from Fig. (d). Additionl lines whih hve onstnt vlue re drwn thiker. These lines n e used s dditionl resoures when mpping single-trget gtes. produt terms ut n derese it, if produt terms in the replement n e omined with other terms. The sme pplies for distnes lrger thn, ut it n lso led to n inrese in the numer of produt terms. This n sometimes e helpful to espe lol minim. Exorism implements defult minimiztion heuristi, referred to s def in the following, tht pplies different omintions nd sequenes of EXORLINK-k opertions for k. We hve modified the heuristi y just omitting the EXORLINK- opertions, referred to s def_wo in the following. B. LUT-sed Mpping This setion desries mpping tehnique tht exploits two oservtions: (i) when mpping single-trget gte there my e dditionl lines ville with onstnt vlue; nd (ii) for single-trget gtes with few ontrol lines (e.g., up to ) one n preompute ner-optiml Clifford+T networks nd store them in dtse. Fig. 6() illustrtes the LUT-sed mpping flow. The ide is to pply -LUT mpping on the ontrol funtion of the single-trget gte nd use ville - vlued lines to store intermedite results from inner LUTs in the mpping. If enough -vlues lines re ville, eh of the single-trget gtes resulting from this mpping is diretly trnslted into ner-optimum Clifford+T network tht is looked up in dtse. The size of the dtse is ompressed y mking use of Boolen funtion lssifition sed on ffine input trnsformtions nd output omplementtion. This proedure requires no dditionl lines eing dded to the iruit. If the proedure nnot e pplied due to too few -vlued lines, diret mpping is used s fllk. Aville -nill resoures: Fig. 7 shows the sme reversile network s in Fig. (d) tht is otined from the initil k-lut mpping. However, dditionl lines tht re -vlued re drwn thiker. We n see tht for the reliztion of the first single-trget gte, there re three -vlued lines, ut for the lst single-trget gte there is only one -vlued line. This informtion n esily e otined from the reversile network resulting from Alg.. The following holds in generl. Let g = T (X, t) e single-trget gte with X = k nd let there e l ville - vlued lines, esides -vlued line for trget line t. If n e relized s -LUT network with t most l + LUTs, then we n relize it s reversile network with l singletrget gtes tht relizes funtion on trget line t suh tht 3 x x x y y

9 9 x f () T f x f( x) x U A U f U UA X p () T g x p f(a x+ ) = g( x) Fig. 8. A single-trget gte for T f nd single trget gte for n T g, whih n e onstruted from T f, sine f nd g re AN-equivlent. ll inner LUTs ompute nd unompute their results on the l ville -vlued lines. This synthesis proedure similr ut muh simpler thn Alg., sine is single-output funtion. Therefore, the numer of dditionl lines required for the inner LUTs nnot e improved y different topologil order nd is solely determined y the numer of LUTs in the mpping. Ner-optiml -input single-trget gtes: There exists n Boolen funtions over n vriles, i.e.,, 6, 56, nd for n =,, 3, nd, respetively. Consequently, there re s mny different single-trget gtes T ({x,..., x n }, x n+ ). For this numer, it is possile to preompute optimum or neroptiml Clifford+T networks for eh of the single-trget gtes using ext or heuristi optimiztion methods for Clifford+T gtes (see, e.g., [6], [5], [53], [5]), nd store them in dtse. Mpping single-trget gtes resulting from the LUTsed mpping tehnique desried in this setions is therefore very effiient. However, the numer of funtions n e redued signifintly when using ffine funtion lssifition. Next, we review ffine funtion lssifition nd show tht two optimum Clifford+T networks for two single-trget gtes with ffine equivlent funtions hve the sme T -ount. For Boolen funtion f(x,..., x n ), let us use the nottion f( x), where x = (x,..., x n ) T. We sy tht two funtions f nd g re ffine equivlent [55], if there exists n n n invertile mtrix A GL n (B) nd vetor B n suh tht g( x) = f(a x + ) for ll x B n. (5) We sy tht f nd g re ffine equivlent under negtion [56], if either (5) holds or g( x) = f(a x + ) for ll x. For the ske of revity, we sy tht f is AN-equivlent to g in the reminder. AN-equivlene is n equivlene reltion nd llows to prtition the set of n into muh smller sets of funtions. For n =,, 3, nd, there only, 3, 6, nd 8 lsses of AN-equivlent funtions (see, e.g., [56], [55], [57]. And ll input Boolen funtions fll into only 6 lsses of AN-equivlent funtions! The dtse solution proposed in this mpping n therefore sle for 5 vriles given fst wy to lssify funtions. Before we disuss lssifition lgorithms, the following lemm shows tht ANequivlent funtions preserve T -ost. Lemm 3: Let f nd g e two n-vrile AN-equivlent funtions. Then the T -ount in Clifford+T networks relizing T f nd T g is the sme. Proof: Sine f nd g re AN-equivlent, there exists A GL n (B), B n, nd p B suh tht g( x) = p f(a x + ) for ll x. It is possile to rete reversile iruit tht tkes x A x + using only CNOT gtes for A nd NOT gtes for (see, e.g., [58]). The output n e inverted using NOT gte. Fig. 8 illustrtes the proof. The suiruit U A relizes # 35f 3 # # 356 # 33f 7 # 7f # 3f 8 # f 7 # 7 79 # ff # 7f # 3f 6 # f 3 # 7 3 # f 7 # 7 7 # 3 5 # # 7 7 # f # 7 6 # 3 7 # 5 # 3 # 7,65 -input funtions 7,87 3-input funtions 7,68 -input funtions 3,996,638 8,6,7 7,7 8,56 5,68 75,8 3,5 99,9 69,398 9, 83,59 6,6 75, 6,79 38,66 8,66 79,83,666,38,75,6 % % % 6% 8% % Fig. 9. This plot shows the distriution of AN-equivlene lsses mong ll -LUTs tht hve een disovered in ll our experimentl results (see Set. VI). The y-xis shows from the ottom to the top ll, 5, nd 7 nononstnt AN-equivlene lsses of, 3, nd vriles. The numer right to the truth tle is the T -ount in the est known Clifford+T reliztion of the orresponding single-trget gte. The x-xis shows the frequeny perentge with respet to other lsses tht hve the sme numer of vriles. The smll numer next to the r shows the frequeny in solute vlues. the liner trnsformtion using CNOT gtes, U relizes the input inversions using NOT gtes, nd X p represents NOT gte, if p =, otherwise the identity. In order to mke use of the lgorithm we need to ompute n optimum or ner-optiml iruit for one representtive in eh equivlene lss for up to vriles. To mth n ritrry single-trget gte with ontrol funtion of up to vriles in the dtse, one needs to derive it s representtive. Algorithms s presented in [59] n e used for this purpose. Fig. 9 lists ll the AN-equivlent lsses for vriles; the lss ontining the onstnt funtions is omitted. It shows how often they re disovered in -LUT in ll our experimentl evlutions. Also, it shows the numer of T gtes in the est-known Clifford+T networks of the orresponding single-trget gte. The lsses #, #, nd # our most frequently. These lsses ontin mong others x x, x x, nd x x x, nd re therefore ontrol funtions of the multiple-ontrolled Toffoli gtes. The tle n guide to the lsses whih enefit most from optimizing their T - ount. Reversile nd Clifford+T networks of the est-known reliztions n e found t quntumli.sttionq.om. C. Hyrid Mpping LUT-sed mpping nnot e pplied if the numer of ville -vlued lines is insuffiient. As fllk, the -LUT network is omitted nd diret mpping is pplied on the k- LUT network, therefore not using ny of the -vlued lines

10 TABLE I PARAMETERS FOR LHRS Prmeter Vlues Desription Prmeters p Q tht ffet the numer of quits (nd T gtes) LUT size {3,..., 3} Mximum numer of inputs to LUTs in the LUT mpping, defult: 6 Topo. order topo_def, topo_tfi_sort Heuristi in whih order to trverse LUTs in the LUT mpping (Set. IV-C), defult: topo_def re iters. (init) {,..., } Numer of itertions for glol re reovery heuristi sed on ext re [], [6], defult: flow iters. (init) {,..., } Numer of itertions for lol re reovery heursiti sed on re flow [6], [6], defult: Prmeters p T tht only ffet the numer of T gtes Mpping diret, hyrid ESOP extrtion AIG extrt, BDD extrt ESOP heuristi none, def, def_wo Mpping method (Set. V), defult: hyrid ESOP extrtion method (Set. V-A), defult: AIG extrt ESOP minimiztion heuristi (Set. V-A), none mens tht no optimiztion is pplied, defult: def_wo Prmeters in hyrid mpping method re iters. {,..., } see ove flow iters. {,..., } see ove SAT-sed opt. true, flse Enles whether -LUT networks in the hyrid mpping re postoptimized using the SAT-sed tehnique desried in [63], defult: flse t ll. The ide of hyrid synthesis is to merge -LUTs into lrger LUTs. By merging LUTs in the network, the numer of LUTs is deresed y nd therefore one fewer -vlued line is required. Our lgorithm for hyrid synthesis merges the output LUT with one of its hildren, therey generting lrger output LUT. This proedure is repeted until the LUT network is smll enough to mth the numer of - vlued lines. The topmost LUT is then mpped using diret mpping, while the remining LUTs re trnslted using the LUT-mpping tehnique. VI. EXPERIMENTAL EVALUATION We hve implemented LHRS s ommnd lhrs in the open soure reversile logi synthesis frmework RevKit [6]. All experiments hve een rried out on n Intel Xeon CPU E5-68 v3 t.5 GHz with 6 GB of min memory running Linux. nd g 5.. A. LHRS Configurtion Tle I gives n overview of ll prmeters tht n e given s input to LHRS. The prmeters re split into two groups. Prmeters in the upper hlf hve minly n influene on the numer of lines nd re used to synthesize the initil reversile network tht is omposed of single-trget gtes (Set. IV). Prmeters in the lower hlf only influene the numer of T gtes y hnging how single-trget gtes re mpped into The soure ode n e found t githu.om/msoeken/irkit TABLE II EPFL ARITHMETIC BENCHMARKS Benhmrk Inputs Outputs Originl Best-LUT AIG nodes Levels LUTs Levels Adder Brrel shifter ,336 5 Divisor 8 8,76,7 368,8 Hypotenuse 56 8,335,8,6,53 Log 3 3 3,6 6,57 9 M 3, Multiplier 8 8 7,6 7,93 9 Sine 5 5,6 5,9 55 Squre-root 8 6,68 5,58 3,77,6 Squre 6 8 8,8 5 3,6 7 Clifford+T networks. Eh prmeter is shown with possile vlues nd some explntion text. B. Benhmrks We used oth demi nd industril enhmrks for our evlution. As demi enhmrks we used the rithmeti instnes of the EPFL omintionl logi synthesis enhmrks [65], whih re ommonly used to evlute logi synthesis lgorithms. In order to investigte the influene of the initil logi representtion, we used different reliztions of the enhmrks: (i) the originl enhmrk desription in terms of n AIG, lled Originl, nd (ii) the est known 6-LUT network wrt. the numer of lines, lled Best-LUT. Further sttistis out the enhmrks re given in Tle II. All experimentl results nd synthesis outomes for the demi enhmrks n e viewed nd downloded from quntumli.sttionq.om. As ommeril enhmrks we used Verilog netlists of severl rithmeti floting point designs in hlf (6-it), single (3-it), nd doule (6-it) preision. For synthesis ll Verilog files were trnslted into AIGs nd optimized for size using ABC s resyn sript. C. Experiments for EPFL Benhmrks We evluted LHRS for oth reliztions (Originl nd Best-LUT) for ll rithmeti enhmrks with LUT size of 6,, nd 6. Tle III lists ll results. We hose LUT size 6, euse it is typil hoie for FPGA mpping, nd therefore we expet tht LUT mpping lgorithms perform well for this size. We hose 6, sine we notied in our experiments tht it is the lrgest LUT size for whih LHRS performs resonly fst for most of the enhmrks. LUT size hs een hosen, sine it is roughly in etween the other two. These onfigurtions llow to synthesize 6 different initil single-trget gte networks for eh enhmrk, therefore spnning 6 optimiztion points in Preto set. For eh of these onfigurtions, we hose different onfigurtions of prmeters in p T, sed on vlues to the mpping method nd the ESOP heuristi ({diret, hyrid} {def, def_wo}). The experiments onfirm the oservtion of Set. IV-D: A lrger LUT size leds to smller numer quits. In some see lsi.epfl.h/enhmrks, version 7.

11 TABLE III EXPERIMENTAL EVALUATION OF LHRS ON THE EPFL ARITHMETIC BENCHMARKS Behmrk LUT size def, diret def_wo, diret def, hyrid def_wo, hyrid quits T -ount runtime T -ount runtime T -ount runtime T -ount runtime Adder 6 Best-LUT 8,.3,87.,63.8,7.8 Originl 55 6,67.9 6,5.8,66.5,66.5 Best-LUT 5 8,675.,3. 3, ,67. Originl 9,357.5,33.,86.5,86. 6 Best-LUT 3 5, ,96.37,699.9,797.8 Originl 63 33,.9 39, ,.7 5,.5 Brrel shifter 6 Best-LUT 8 3,5. 3,5. 7,.5 7,. Originl 58 5,9.6 5,9.5 76, ,883.6 Best-LUT 7 5,986. 5,986.,656.35, Originl 58 5,9.5 5,9.6 76, , Best-LUT 595,69., , ,53.37 Originl 58 5,8.6 5,8. 78, ,67.5 Divisor 6 Best-LUT 3,399 36,.77 3, , , Originl,389 75, , , ,98.3 Best-LUT 3,6 5, , , , Originl,55 875, ,87.63,73,7 5.36,73, Best-LUT 3,7 3,85, ,89, , ,3.7 Originl,87,5,585.79,7, ,96,7 5.65,96,7 5.8 Hypotenuse 6 Best-LUT,6 3,87, ,67, ,5,57. 5,5,5.68 Originl 7,8,3,6 8.6,3, ,57, ,57, Best-LUT 36,3 5,9, ,83, ,893, ,893, Originl 3,87,35,66 3.,3, ,3, ,3, Best-LUT 3,336,39,83 5,9.,99,9,38.3,73, ,76, Originl 39,3,3, ,3,693,595. 8,8, ,8, Log 6 Best-LUT 6,65 66, ,68.8,363, ,363, Originl 7,6 5, , , , Best-LUT 3,38,36,8 5.87,5,8 35.8,3,9 9.3,33, Originl,875 3,8, ,3, ,58, ,59, Best-LUT,5 56,589,96 5,8.7 59,587,6 6,.,6,,66.7,83, Originl,35 6,7,767 9,9.6 6,6,3 7,.5 6,368,3,7.36 6,7, Mx 6 Best-LUT,36 56,9.3 55,97.5 9,89.3 9,89.9 Originl,33 7, ,8.3 6, ,933.9 Best-LUT 8 95, , ,7.3 5,7.3 Originl 9 9,669.5, ,.88 88,.8 6 Best-LUT 75,8, ,77, ,88. 3,.39 Originl , , ,758. 9,7.9 Multiplier 6 Best-LUT 5,8 683, , , , Originl 5,86 386, , , ,65.87 Best-LUT 3,8,8, ,,53 6.,368,8 8.69,368,8 8.7 Originl 3,5,377, ,377,5.,59,55 5.,59, Best-LUT,55 9,98,9,7.3,,88,8.93 3,8, ,9,87.7 Originl,85 8,7,69,6.5 8,83, ,89, ,9, Sine 6 Best-LUT,77,885.7,56.6,3.36,3.3 Originl,68 87, ,87.6 3,66. 3,66.9 Best-LUT , , , ,79.67 Originl 7 8, ,85.5,7,87 5.5,7, Best-LUT 8 3,73,77 6,579. 3,953,575,.36,759, ,79,53.6 Originl 58 3,8,38,. 3,93, ,7, ,8, Squre-root 6 Best-LUT 3, 368, ,593. 7, , Originl 8, 79, , , , Best-LUT,87 59,6.3 5, , , Originl 7,89 33, , , , Best-LUT,63 6,58, ,7,98 6.,, ,8, Originl 7,86,5, ,79, ,8,39.39,8,.3 Squre 6 Best-LUT 3,39 99, , , , Originl,58 95,9.5 95, , ,75.35 Best-LUT,88 53, ,. 86, , 7.7 Originl 3,355 6,. 6,3.,3, ,3, Best-LUT,33 3,96,8 6,98.,,6 7,888.93,33,37 8,99.69,355,9, Originl,66,9,99 9,866.,7,766 7,8.6 3,75,569 8,8.6 3,56,6,97.58