Register Pressure in Software-Pipelined Loop Nests

Size: px

Start display at page:

Download "Register Pressure in Software-Pipelined Loop Nests"

Valerie Tyler
6 years ago
Views:

1 Rgistr Prssur in Softwr-Piplin Loop Nsts Fst Computtion n Impt on Arhittur Dsign Aln Douillt Gung R. Go {ouillt,ggo}@psl.ul.u Dprtmnt of Eltril & Computr Enginring Univrsity of Dlwr 18 th Intrntionl Workshop on Lngugs n Compilrs for Prlll Computing Hwthorn, Nw York Otor 20 th -22 n, 2005 A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 1 / 43

2 Introution Sintifi Applitions loop nsts ominnt Singl-imnsion Softwr Piplining (SSP) softwr piplins most profitl loop in loop nst high rgistr prssur rgistr llotion is tim-onsuming N for fst mtho to vlut rgistr prssur tt infsil shuls for lling th rgistr llotor msur qulity of rgistr llotion solution giv stimt of rgistr ns for futur rhittur signs A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 2 / 43

3 Outlin 1 Loop Nst Softwr-Piplining 2 Prolm Sttmnt 3 Dfinitions & Issus 4 Fst Rgistr Prssur Computtion 5 Exprimnts 6 Conlusion A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 3 / 43

4 Outlin 1 Loop Nst Softwr-Piplining 2 Prolm Sttmnt 3 Dfinitions & Issus 4 Fst Rgistr Prssur Computtion 5 Exprimnts 6 Conlusion A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 4 / 43

5 Moulo Shuling most populr SWP thniqu wll stui n unrstoo full rry of loop optimiztions singl loop, prlll xution of itrtions nw itrtion issu vry T yls (initition intrvl) FOR J=0,4 END FOR prolog krnl pilog A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 5 / 43

6 Moulo Shuling But... limit to innrmost loop loop trnsformtions to ring ILP or t h rus potntil to innrmost loop not lwys possil A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 6 / 43

7 Singl-Dimnsion Softwr-Piplining (SSP) propos y Rong t l. (CGO 04, PLDI 05) softwr piplins th most profitl loop lvl in loop nst quivlnt to MS if innrmost lvl slt n sn s gnrliztion of MS to loop nsts provn prformn oost vs. MS n tk vntg of loop optimiztions us for MS singl-imnsion / simplifis multi-imnsionl DDG into uni-imnsionl DDG A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 7 / 43

8 SSP Krnl SSP gnrts krnl similr to MS nlos stgs singl initition intrvl T L 1 is th outrmost loop n L n th innrmost S i : numr of stgs t lvl i T mil loop stgs innrmost stgs yls T 2 T 1 A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 8 / 43

9 SSP Il Shul Gnrt using krnl s tmplt nw outrmost itrtion issu vry T yls outrmost itrtions xut in prlll innr itrtions xut squntilly within on outrmost itrtion rsour onflits! A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC 05 9 / 43

10 SSP Il Shul: Exmpl FOR J=1,3 END FOR FOR I=1,4 END FOR T=1 S1=5 S2=3 T rsour onflits A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

11 SSP Finl Shul lys som outrmost itrtions to voi rsour onflits outrmost itrtions xut in groups of S n rsour onflit-fr shul ly Sn itrtions A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

12 SSP Loop Pttrns Pttrns: Outrmost Loop Pttrn Innr Loop Exution Sgmnt Innrmost Loop Pttrn Drining & Filling Pttrn Composition: OLP: ll S krnl stgs ILES: yli omintion of S n onsutiv stgs Krnl OLP ILES OLP ILP DFP... ly... A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

13 SSP Implmnttion loop nst Loop Sltion loop nst Dpnn Simplifition 1 D DDG Shul Constrution krnl Rgistr Allotion rgistr llot krnl Co Gnrtion finl shul A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

14 Outlin 1 Loop Nst Softwr-Piplining 2 Prolm Sttmnt 3 Dfinitions & Issus 4 Fst Rgistr Prssur Computtion 5 Exprimnts 6 Conlusion A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

15 Motivtion n to trmin fsiility of shuls rgistr llotion is tim-onsuming unfsil shuls / of high rgistr prssur not unommon n to vlut qulity of rgistr llotor how fr from optiml solution? n to vlut tul rgistr ns for rhitturl signs r rgistr fils ig nough? A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

16 Prolm Sttmnt Givn loop nst n n SSP shul for it, vlut th rgistr prssur MxLiv of th finl shul. only rotting rgistrs MxLiv = mximum numr of liv vrils t ny givn yl in th finl shul MxLiv finition similr to th on for MS. A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

17 Upt SSP Implmnttion loop nst Loop Sltion loop nst Dpnn Simplifition 1 D DDG Shul Constrution hoos iffrnt loop lvl ys too high? no OR inrs initition intrvl Rgistr Prssur Evlution krnl Rgistr Allotion rgistr llot krnl Co Gnrtion finl shul A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

18 Outlin 1 Loop Nst Softwr-Piplining 2 Prolm Sttmnt 3 Dfinitions & Issus 4 Fst Rgistr Prssur Computtion 5 Exprimnts 6 Conlusion A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

19 Dfinitions slr liftim: strt: finition yl of th vlu n: kill yl of th vlu omg: numr of outrmost itrtions spnn lssifition glol: onstnt vlus, ignor input & output: prolog n pilog, ignor lol: within sm outrmost itrtion ross-itrtion: twn outrmost itrtions lol strt ross itrtion n omg A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

20 Issus mor omplx liftim pttrns thn MS non-onstnt initition rt strth liftims sm stg my hv iffrnt liftims pttrns stg is not lwys follow y th sm stgs iffrn twn first n lst instn of th sm stg A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

21 Liftims Exmpl A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

22 Outlin 1 Loop Nst Softwr-Piplining 2 Prolm Sttmnt 3 Dfinitions & Issus 4 Fst Rgistr Prssur Computtion 5 Exprimnts 6 Conlusion A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

23 Mtho Ovrviw Mtho Kys: sprt OLP from ILES instns of stgs sprt first from lst instns of stgs sprt lol from ross-itrtion liftims Stps: ount numr of lol liftims in first instn of stgs ount numr of lol liftims in lst instn of stgs ount numr of ross-itrtion liftims in h stg list ll possil omintions of stgs in shul numr of liftims for h omintion in OLP n ILES MxLiv is th highst vlu A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

24 Lol Liftims lst tritionl livnss nlysis omputs for oth first n lst instns of stg s h yl in stg twn 0 n T 1 liv-out st of stg ( = T ) stg of lvl i visit i tims LT lol (s,, first/lst) first lvl 1 lvl 2 lvl 3 A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

25 Cross-Itrtion Liftims n stg n yl of finition n kill irt formul for h yl in OLP twn 0 n T 1 not stg-spifi S f = 1 = 2 f S kill = 0 = 0 kill LT ross LT ross () = v ivs whr omg=3 ((S kill (v) S f (v) + 1) + δ f (, v) + δ kill (, v)) { δf (, v) = 1 if < f (v), 0 othrwis δ kill (, v) = 1 if > kill (v), 0 othrwis A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

26 OLP: lst stgs lol liftims Comintion of first n lst stgs not lwys th sm l 1 s=l n i LT lol (s,, lst) A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

27 OLP: first stgs lol liftims Comintion of first n lst stgs not lwys th sm l n 1 i s=f 1 LT lol (s,, first) A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

28 OLP: ross-itrtion liftims Numr of ross-itrtion liftims intil twn instns of OLP LT ross () A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

29 OLP ount LT olp () = LT ross () + mx i [1,S n] l 1 s=l n i l n 1 i LT lol (s,, lst) + LT lol (s,, first) s=f 1 A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

30 ILES: lst strth lol liftims Liv-out of lst stgs l 1 s=l n LT lol (s, T, lst) A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

31 ILES: first strth lol liftims Liv-out of first stgs f n 2 s=f 1 LT lol (s, T, first) A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

32 ILES: strth ross-itrtion liftims liv-out ross-itrtion liftims from OLP LT ross (T) A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

33 ILES: lol liftims S n onsutiv stgs (yli) mx l [2,n] ( mx i 0 [0,S l 1] ( Sn 1 i=0 LT lol (f l + (i 0 + i)%s l,, first) )) A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

34 ILES ount LT ils () = LT ross (T) + l 1 s=l n LT lol (s, T, lst) f n 2 + LT lol (s, T, first) s=f 1 + mx l [2,n] ( mx i 0 [0,S l 1] ( Sn 1 i=0 LT lol (f l + (i 0 + i)%s l,, first) )) A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

35 MxLiv FtCovr olp = mx ( LTolp () ) [0,T 1] FtCovr ils = mx (LT ils ()) [0,T 1] MxLiv = mx(ftcovr ils, FtCovr olp ) A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

36 Outlin 1 Loop Nst Softwr-Piplining 2 Prolm Sttmnt 3 Dfinitions & Issus 4 Fst Rgistr Prssur Computtion 5 Exprimnts 6 Conlusion A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

37 Exprimntl Frmwork ORC2.1 ompilr 1.4Ghz Itnium worksttion, 1GB RAM Livrmor, SPEC2000 FP, NPB 2.2 nhmrks 127 loop nsts 328 tst ss A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

38 Running Tim in th orr of 1/1000 s qurti running tim 3 orrs of mgnitu fstr thn rgistr llotor spup inrss s loop gts pr fst nough to us in SSP frmwork Spup (log sl) Bnhmrks A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

39 MxLiv Avrg: INT=42 FP=15 rgistr prssur too high for 43% of loop nsts of pth 4 1 Amnility Rtio Dpth A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

40 FP/INT Comprison FP rgistr prssur stl s loop gts pr prssur nvr xs 64 rgistrs INT rgistr prssur inrss s loop gts pr loop ovrhs rry inxs longr liv rngs Extr FP Rgistr Prssur pr Lvl Lvl 3 or highr Lvl 2 Lvl 1 0 Bnhmrks A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

41 Rgistr Fil Siz mx FP rgistr fil siz: 64 il INT/FP rtio: 2 77% n 67% of loop nsts of pth 4 n 5 woul om fsil Totl Rgistr Prssur (FP+INT) Totl Rgistr Prssur FP/INT Rtio FP/INT Rtio 32 Bnhmrks 0 A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

42 Outlin 1 Loop Nst Softwr-Piplining 2 Prolm Sttmnt 3 Dfinitions & Issus 4 Fst Rgistr Prssur Computtion 5 Exprimnts 6 Conlusion A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

43 Conlusion SSP softwr-piplins loop nsts t th most profitl lvl rgistr prssur is howvr vry high n for fst mtho to vlut rgistr prssur tt infsil shuls rly msur qulity of rgistr llotor giv n stimt of th tul rgistr ns for futur rhittur signs propos solution ls with issus spifi to loop nst SWP n SSP is vry fst n n us for rgistr llotion futur work inrmntl solution to intgrt to th shulr iffrnt rhitturs: lustr VLIW,... A. Douillt, G.R. Go (Univ. of Dlwr) Rgistr Prssur in SWP Loop Nsts LCPC / 43

Danfoss Compressors. Beverage Coolers R134a R404A/R507 R V 50/60 Hz MBP. Beverage Coolers V 50/60 Hz MBP.

Danfoss Compressors. Beverage Coolers R134a R404A/R507 R V 50/60 Hz MBP. Beverage Coolers V 50/60 Hz MBP. Bvrag oolrs 3a 0A/07 0 0-0 V 0/60 Hz BP ST - SI T--g S AKI O IVI POSSIB 87 8 ain wining ain wining a Start wining ain wining a Start wining lip Washr rommt slv omprssor as a Stl pin omprssor as Start wining