Archit ect ure s w it h

Size: px

Start display at page:

Download "Archit ect ure s w it h"

Sharlene Newton
5 years ago
Views:

1 Inst ruct ion Select ion for Com pilers t hat Target Archit ect ure s w it h Echo Inst ruct ions Philip Brisk Ani Nahapetian Majid philip@cs.ucla.edu ani@cs.ucla.edu Sarrafzadeh m ajid@cs.ucla.edu Em bedded and Reconfigurable Syst em s Lab Com put er Science Depart m ent Universit y of California, Los Angeles

2 Outline Code Com pression LZ7 7 Com pression Echo Inst ruct ions Com piler Fram ew ork Experim ent al Result s Sum m ary

3 Code Com pression W hy Com press a Program? Silicon Require m ent s for on-chip ROM Pow e r Consum pt ion Cost t o Fabricat e Cost Paid by t he Consum er Are t here Cost s t o Decom pression? Perform a nce Overhe ad Dedicat e d Hardw a re (or Soft w are) Longer CPU Pipelines Incre ase d Bra nch M ispredict ion Pe na lt y

4 LZ77 Com pression To Com press a St ring, Ident ify Repeat ed Subst rings and Replace Each w it h a Point er (Offset, Lengt h of Sequence) ABCDBCABCDBACABCDBADAABCDBDC 2 8 ABCDBCABCDBACABCDBADAABCDBDC 2 8 one chara ct er Lengt h

5 Echo Instructions Decom presses LZ7 7 -com pressed Program s w it h M inim al Hardw are Requirem ent s 2 Dedicat e d Regist ers: R 1, R 2 1 Decre m e nt e r w it h = 0 Test ( NOR) Echo( Offset, N) 1. Save PC and N in R 1 a nd R 2 2. Branch t o PC Offse t 3. Execut e t he next N inst ruct ions 4. Re t urn t o t he Call Point 5. Re st ore PC from R 1 (Fraser, M icrosoft 0 2 ) (La u, CASES 0 3 )

6 Substring Matching $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 (Fraser, SCC $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 $Echo(240, 5) $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 Ech o(304, 5) $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1

7 Reschedule/Renam e $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 $10 $5 + $4 $11 $9 * $6 $6 $9 * $10 $10 $11 / $6 $10 $ $11 $7 * $8 $1 $2 + $3 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 Renam e $4 : $3 $5 : $2 $6 : $8 $9 : $7 $10 : $1 $11 : $11 (Cooper, PLDI 9 9 ) (Debray, TOPLAS 0 0 ) Reschedule (Lau, CASES 0 3 ) $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1

8 DFG Isom orphism $1 $2 + $3 $11 $7 * $8 $8 $7 * $1 $1 $11 / $8 $1 $8 + 1 $10 $5 + $4 $11 $9 * $6 $6 $9 * $10 $10 $11 / $6 $10 $ $11 $7 * $8 $1 $2 + $3 $8 $7 * $1 $1 $11 / $8 $1 $ * / + * Our Approach Represe nt Sequence s a s Da t a Flow Graphs Ide nt ify Repeat ed Isom orphic Subgra phs Replace Subgraphs w it h Echo Inst ruct ions

9 Isom orphic Subgraph Identification Edge Cont ract ion (Kast ner, ICCAD 0 1 ) Conside r a New Subgraph for Each DFG Edge

10 Isom orphic Subgraph Identification Com put e an Independent Set for Each Edge Type NP-Com ple t e Problem It erat ive Im provem ent Algorit hm - (Kirovski, DAC 9 8 )

11 Isom orphic Subgraph Identification Replace M ost Frequent ly Occurring Pat t ern w it h a Tem plat e Original DFG Edge Dat a Dependencies Incident on Te m plat es Dat a Dependencies t hat Cross Tem plat e Boundaries

12 Isom orphic Subgraph Identification Edge Cont ract ion in t he Presence of Tem plat es Gene ra t e Ne w Tem pla t es Along Bold Edges Test for Tem pla t e Equivale nce is DAG Isom orphism Used t he Publicly Available VF2 Algorit hm

13 Isom orphic Subgraph Identification

14 Isom orphic Subgraph Identification Replace M ost Frequent ly Occurring Pat t ern w it h a Tem plat e Original DFG Edge Dat a Dependencies Incident on Te m plat es Dat a Dependencies t hat Cross Tem plat e Boundaries

15 Isom orphic Subgraph Identification Replace M ost Frequent ly Occurring Pat t ern w it h a Tem plat e Original DFG Edge Dat a Dependencies Incident on Te m plat es Dat a Dependencies t hat Cross Tem plat e Boundaries

16 Register Allocation Isom orphic Tem plat es M ust Have Ident ical Usage of Regist ers Regist er s Shuffle or Spill Code

17 Register Allocation Code Reuse Const raint s M ay W ork Against Code Size Each Te m plat e Elim inat es 3 Inst rs. 5 Shuffle/Spill Ops. are Required The General Problem is Very Com plicat ed Exist ing Allocat ion Te chniques Are Not Applicable Regist er s Shuffle or Spill Code Present St at us: The Alloca t or is a W ork-in-

18 Isom orphic Subgraph Identification Aft er Regist er Allocat ion, Replace Subgraphs w it h Echo Inst ruct ions Echo Echo Echo

19 Our Goal is t o Evaluat e t he Effect iveness of Subgraph Ident ificat ion Experim ental Fram ework Built Subgraph Ident ificat ion int o t he M achine-suif Fram ew ork Pass Pla ced Be t w een Inst ruct ion Select ion and Regist e r Alloca t ion Curre nt Im ple m e nt a t ion Support s Alpha as Target Allow s for Fut ure Int e grat ion w it h Sim plescalar Sim ulat or

20 Experim ental Methodology W it hout Allocat ion in Place, W e Cannot : Est im at e W here Shuffle /Spill Code W ill be Insert ed at Tem plat e Boundaries Det erm ine W hich Copy Inst ruct ions W ill be Coalesce d But W e Can: M ake Assum pt ions Regarding t he St art ing Point for Regist er Allocat ion

21 Two Approaches to Coalescing Pessim ist ic Coalescing (M ost Allocat ors) Begin w it h All Copy Inst ruct ions in Pla ce Coalesce Copies W hen Safe Opt im ist ic Coalescing (Park & M oon, PACT 9 8 ): Init ially Coalesce ALL Copy Inst ruct ions Re-Int roduce Coale sced Copies t o Avoid Spilling Live Ranges W he ne ver Possible

22 Assum ptions and Measurem ent Pessim ist ic Assum pt ion No Copy Inst ruct ions a re Coalesced Opt im ist ic Assum pt ion ALL Copy Inst ruct ions are Coale sced Com put e t he Num ber of DFG Operat ions Before and Aft er Com pression St ep PU: Pessim ist ic, Uncom pressed OU: Opt im ist ic, Uncom pressed PC: Pessim ist ic, Com pressed Com pressed OC: Opt im ist ic,

23 Benchm arks Taken from t he M ediabench and M ibench Applicat ion Suit es Be nchm a r k ADPCM low fish Epic G7 2 1 JPEG M PEG2 Dec M PEG2 Enc Pegw it Descript ion Adapt ive Different Sym ial m Pulse et ric Code Block M odulat Cipher ion w it h Variable Key Lengt h Im age Dat a Com pression Ut ilit y Voice Com pression Im age Com pression and Decom pre ssion M PEG2 Decoder M PEG2 Encoder Public Key Encrypt ion and Aut hent icat ion

24 Experim ental Results PU: Pessim ist ic, Uncom pressed OU: Opt im ist ic, Uncom pressed PC: Pessim ist ic, Com pressed Com pressed OC: Opt im ist ic, DFG Ops PU PC OU OC 0 ADPCM Blowfish Epic G721 JPEG MPEG2 Dec. MPEG2 Enc. Pegwit

25 Runtim e Considerations Algorit hm Ran Efficient ly ( a few seconds) for M ost Benchm arks Several Not able Except ions Four Com m on Fe at ure s Large DFGs User-De fine d M acros Unrolled Loops Cyclic Shift ing of Pa ra m et ers sha1.c (Pegw it ) One DFG Com pilat ion Tim e W as in Excess of 3 Hrs

26 Runtim e Considerations sha1.c # define R0 ( v, w, x, y, z, i) { z + = ; w = } void SHA1 Transform ( unsigned long st at e[5], ) { unsigned long a = st at e[0 ], b = st at e[1 ], c = st at e[2 ], d = st at e[3], e = st at e[4 ] ; } R0 (a, b, c, d, e, 0 ); R0 (e, a, b, c, d, 1 ); R0 (d, e, a, b, c, 2 ); R0 (c, d, e, a, b, 3 ); R0 (b, c, d, e, a, 4 ); R0 (a, b, c, d, e, 5 ); R0 (a, b, c, d, e, 1 5 );

27 Com pilat ion Tim e W as Reduced t o Seconds Runtim e Considerations sha1.c # define R0 ( v, w, x, y, z, i) { z + = ; w = } void SHA1 Transform ( unsigned long st at e[5], ) { unsigned long a = st at e[0 ], b = st at e[1 ], c = st at e[2 ], d = st at e[3], e = st at e[4 ], t m p; } for( unsigned long i = 0 ; i < 1 6 ; i+ + ) { R0 (a, b, c, d, e, i); t m p = e; e = d; d = c; c = b; b = a; a = t m p; }

28 Conclusion Echo Inst ruct ions Com pression a t a M inim al Hardw are Cost Perform a nce Overhead is Tw o Branches pe r Echo Com piler Opt im izat ion Ident ify Redunda ncy via Subgra ph Isom orphism New Cha llenge s for Regist er Allocat ion Experim ent s Significant Redunda ncy Observed in

Nov Julien Michel

Nov Julien Michel Int roduct ion t o Free Ene rgy Calculat ions ov. 2005 Julien Michel Predict ing Binding Free Energies A A Recept or B B Recept or Can we predict if A or B will bind? Can we predict the stronger binder?