from pyabc import * import pyabc_split import redirect import sys import os import time import math import main import filecmp global G_C,G_T,latches_before_abs,latches_before_pba,n_pos_before,x_factor,methods,last_winner global last_cex,JV,JP, cex_list,max_bmc, last_cx, pord_on, trim_allowed, temp_dec, abs_ratio, ifbip global if_no_bip """ The functions that are currently available from module _abc are: int n_ands(); int n_pis(); int n_pos(); int n_latches(); int n_bmc_frames(); int prob_status(); 1 = unsat, 0 = sat, -1 = unsolved int cex_get() int cex_put() int run_command(char* cmd); int n_nodes(); int n_levels(); bool has_comb_model(); bool has_seq_model(); bool is_true_cex(); bool is_valid_cex(); return 1 if the number of PIs in the current network and in the current counter-example are equal int n_cex_pis(); return the number of PIs in the current counter-example int n_cex_regs(); return the number of flops in the current counter-example int cex_po(); returns the zero-based output PO number that is SAT by cex int cex_frame(); return the zero-based frame number where the outputs is SAT The last four APIs return -1, if the counter-example is not defined. """ #global variables #________________________________________________ stackno_gabs = stackno_gore = stackno_greg= 0 STATUS_UNKNOWN = -1 STATUS_SAT = 0 STATUS_UNSAT = 1 RESULT = ('SAT' , 'SAT', 'UNSAT', 'UNDECIDED', 'UNDECIDED,', 'ERROR' ) Sat = Sat_reg = 0 Sat_true = 1 Unsat = 2 Undecided = Undecided_reduction = 3 Undecided_no_reduction = 4 Error = 5 Restart = 6 xfi = x_factor = 1 #set this to higher for larger problems or if you want to try harder during abstraction max_bmc = -1 last_time = 0 j_last = 0 seed = 113 init_simp = 1 temp_dec = True ifpord1 = 1 K_backup = init_time = 0 last_verify_time = 20 last_cex = last_winner = 'None' last_cx = 0 trim_allowed = True pord_on = False sec_sw = False sec_options = '' cex_list = [] TERM = 'USL' ifbip = 0 # sets the abtraction method if_no_bip = 0 #True sets it up so it can use bip and reachx commands. abs_ratio = .5 #this controls when abstract stops t_init = 2 #initial time for poor man's concurrency. methods = ['PDR', 'INTRP', 'BMC', 'SIM', 'REACHX', 'PRE_SIMP', 'Simple_prove', 'PDRM', 'REACHM', 'BMC3','Min_Retime', 'For_Retime','REACHP','REACHN','PDR_sd','prove_part_2', 'prove_part_3','verify','sleep','PDRM_sd','prove_part_1', 'run_parallel','INTRPb', 'INTRPm', 'REACHY', 'REACHYc','RareSim','simplify', 'speculate', 'quick_sec', 'BMC_J', 'BMC2', 'extract -a', 'extract', 'PDRa'] #'0.PDR', '1.INTERPOLATION', '2.BMC', '3.SIMULATION', #'4.REACHX', '5.PRE_SIMP', '6.SUPER_PROVE(2)', '7.PDRM', '8.REACHM', 9.BMC3' # 10. Min_ret, 11. For_ret, 12. REACHP, 13. REACHN 14. PDRseed 15.prove_part_2, #16.prove_part_3, 17.verify, 18.sleep, 19.PDRMm, 20.prove_part_1, #21.run_parallel, 22.INTRP_bwd, 23. Interp_m 24. REACHY 25. REACHYc 26. Rarity Sim 27. simplify #28. speculate, 29. quick_sec, 30 bmc3 -S, 31. BMC2 32. extract -a 33. extract 34. pdr_abstract win_list = [(0,.1),(1,.1),(2,.1),(3,.1),(4,.1),(5,-1),(6,-1),(7,.1)] FUNCS = ["(pyabc_split.defer(abc)('&get;,pdr -vt=%f'%t))", "(pyabc_split.defer(abc)('&get;,imc -vt=%f'%(t)))", "(pyabc_split.defer(abc)('&get;,bmc -vt=%f'%t))", "(pyabc_split.defer(simulate)(t))", "(pyabc_split.defer(abc)('reachx -t %d'%t))", "(pyabc_split.defer(pre_simp)())", ## "(pyabc_split.defer(super_prove)(2))", "(pyabc_split.defer(simple)())", "(pyabc_split.defer(pdrm)(t))", "(pyabc_split.defer(abc)('&get;&reachm -vcs -T %d'%t))", "(pyabc_split.defer(abc)('bmc3 -C 1000000 -T %f'%t))", "(pyabc_split.defer(abc)('dr;&get;&lcorr;&dc2;&scorr;&put;dr'))", "(pyabc_split.defer(abc)('dr -m;&get;&lcorr;&dc2;&scorr;&put;dr'))", "(pyabc_split.defer(abc)('&get;&reachp -vr -T %d'%t))", "(pyabc_split.defer(abc)('&get;&reachn -vr -T %d'%t))", ## "(pyabc_split.defer(abc)('&get;,pdr -vt=%f -seed=521'%t))", "(pyabc_split.defer(pdrseed)(t))", "(pyabc_split.defer(prove_part_2)())", "(pyabc_split.defer(prove_part_3)())", "(pyabc_split.defer(verify)(JV,t))", "(pyabc_split.defer(sleep)(t))", "(pyabc_split.defer(pdrmm)(t))", "(pyabc_split.defer(prove_part_1)())", "(pyabc_split.defer(run_parallel)(JP,t,'TERM'))", "(pyabc_split.defer(abc)('&get;,imc -bwd -vt=%f'%t))", ## "(pyabc_split.defer(abc)('int -C 1000000 -F 10000 -K 2 -T %f'%t))", "(pyabc_split.defer(abc)('int -C 1000000 -F 10000 -K 1 -T %f'%t))", "(pyabc_split.defer(abc)('&get;&reachy -v -T %d'%t))", "(pyabc_split.defer(abc)('&get;&reachy -cv -T %d'%t))", "(pyabc_split.defer(simulate2)(t))", "(pyabc_split.defer(simplify)())", "(pyabc_split.defer(speculate)())", "(pyabc_split.defer(quick_sec)(t))", "(pyabc_split.defer(bmc_j)(t))", "(pyabc_split.defer(abc)('bmc2 -C 1000000 -T %f'%t))", "(pyabc_split.defer(extractax)('a'))", "(pyabc_split.defer(extractax)())", "(pyabc_split.defer(pdra)(t))", ] ## "(pyabc_split.defer(abc)('bmc3 -C 1000000 -T %f -S %d'%(t,int(1.5*max_bmc))))" #note: interp given 1/2 the time. allreachs = [4,8,12,13,24,25] reachs = [24] ##allpdrs = [14,7,34,19,0] allpdrs = [14,7,34,19] pdrs = [34,7,14] allbmcs = [2,9,30,31] exbmcs = [2,9] bmcs = [9,30] allintrps = [1,22,23] bestintrps = [23] intrps = [23] allsims = [3,26] sims = [26] allslps = [18] slps = [18] imc1 = [1] JVprove = [7,23,4,24] JV = pdrs+intrps+bmcs+sims #sets what is run in parallel '17. verify' above JP = JV + [27] # sets what is run in '21. run_parallel' above 27 simplify should be last because it can't time out. #_____________________________________________________________ # Function definitions: # simple functions: ________________________________________________________________________ # set_globals, abc, q, x, has_any_model, is_sat, is_unsat, push, pop # ALIASES def initialize(): global xfi, max_bmc, last_time,j_last, seed, init_simp, K_backup, last_verify_time global init_time, last_cex, last_winner, trim_allowed, t_init, sec_options, sec_sw global n_pos_before, n_pos_proved, last_cx, pord_on, temp_dec, abs_time xfi = x_factor = 1 #set this to higher for larger problems or if you want to try harder during abstraction max_bmc = -1 last_time = 0 j_last = 0 seed = 113 init_simp = 1 temp_dec = True K_backup = init_time = 0 last_verify_time = 20 last_cex = last_winner = 'None' last_cx = 0 trim_allowed = True pord_on = False t_init = 2 #this will start sweep time in find_cex_par to 2*t_init here sec_sw = False sec_options = '' cex_list = [] n_pos_before = n_pos() n_pos_proved = 0 abs_time = 10000 def set_abs_method(): """ controls the way we do abstraction, 0 = no bip, 1 = old way, 2 use new bip and -dwr see absab() """ global ifbip print 'Set method of abstraction: \n0 = use vta and no bips, \n1 = old way, \nelse = use ,abs and -dwr' s = raw_input() s = remove_spaces(s) if s == '1': ifbip = 1 #old way elif s == '0': ifbip = 0 #use vta and no bip else: ifbip = 2 #use ,abc -dwr print 'ifbip set to %d. Note engines are set only when read_file is done'%ifbip def ps(): print_circuit_stats() def n_real_inputs(): """This gives the number of 'real' inputs. This is determined by trimming away inputs that have no connection to the logic. This is done by the ABC alias 'trm', which changes the current circuit. In some applications we do not want to change the circuit, but just to know how may inputs would go away if we did this. So the current circuit is saved and then restored afterwards.""" ## abc('w %s_savetempreal.aig; logic; trim; st ;addpi'%f_name) abc('w %s_savetempreal.aig'%f_name) with redirect.redirect( redirect.null_file, sys.stdout ): ## with redirect.redirect( redirect.null_file, sys.stderr ): reparam() n = n_pis() abc('r %s_savetempreal.aig'%f_name) return n def timer(t): btime = time.clock() time.sleep(t) print t return time.clock() - btime def sleep(t): time.sleep(t) return Undecided def abc(cmd): abc_redirect_all(cmd) def abc_redirect( cmd, dst = redirect.null_file, src = sys.stdout ): """This is our main way of calling an ABC function. Redirect, means that we suppress any output from ABC""" with redirect.redirect( dst, src ): return run_command( cmd ) def abc_redirect_all( cmd ): """This is our main way of calling an ABC function. Redirect, means that we suppress any output from ABC, including error printouts""" with redirect.redirect( redirect.null_file, sys.stdout ): with redirect.redirect( redirect.null_file, sys.stderr ): return run_command( cmd ) def convert(t): t = int(t*100) return str(float(t)/100) def set_engines(N=0): """ Called only when read_file is called. Sets the MC engines that are used in verification according to if there are 4 or 8 processors. if if_no_bip = 1, we will not use any bip and reachx engines """ global reachs,pdrs,sims,intrps,bmcs,n_proc,abs_ratio,ifbip if N == 0: N = n_proc = os.sysconf(os.sysconf_names["SC_NPROCESSORS_ONLN"]) if N == 1: reachs = [24] pdrs = [7] ## bmcs = [30] bmcs = [9] intrps = [] sims = [] slps = [18] elif N == 2: reachs = [24] pdrs = [7] bmcs = [30] intrps = [] sims = [] slps = [18] elif N == 4: reachs = [24] pdrs = [34,7] if if_no_bip: allpdrs = pdrs = [7,19] bmcs = [9,30] intrps = [23] sims = [] slps = [18] elif N == 8: reachs = [24] pdrs = [34,7,14] intrps = [23,1] if if_no_bip: allpdrs = pdrs = [7,19] intrps = [23] bmcs = [9,30] sims = [26] #use default slps = [18] def set_globals(): """This sets global parameters that are used to limit the resources used by all the operations bmc, interpolation BDDs, abstract etc. There is a global factor 'x_factor' that can control all of the various resource limiting parameters""" global G_C,G_T,x_factor nl=n_latches() na=n_ands() np = n_pis() #G_C = min(500000,(3*na+500*(nl+np))) G_C = x_factor * min(100000,(3*na+500*(nl+np))) #G_T = min(250,G_C/2000) G_T = x_factor * min(75,G_C/2000) G_T = max(1,G_T) #print('Global values: BMC conflicts = %d, Max time = %d sec.'%(G_C,G_T)) def a(): """this puts the system into direct abc input mode""" print "Entering ABC direct-input mode. Type q to quit ABC-mode" n = 0 while True: print ' abc %d> '%n, n = n+1 s = raw_input() if s == "q": break run_command(s) def remove_spaces(s): y = '' for t in s: if not t == ' ': y = y + t return y def seq_name(f): names = [] f = f + '_' names = [] while len(f)>0: j = f.find('_') if j == -1: break names = names + [f[:j]] ## print names f = f[j+1:] ## print f return names def revert(f,n): l = seq_name(f) for j in range(n): if len(l)>0: l.pop() name = construct(l) return name def construct(l): ll = l name = '' while len(l)>0: name = '_'+ll.pop()+name return name[1:] def process_sat(): l = seq_name(f_name) def read_file_quiet(fname=None): """This is the main program used for reading in a new circuit. The global file name is stored (f_name) Sometimes we want to know the initial starting name. The file name can have the .aig extension left off and it will assume that the .aig extension is implied. This should not be used for .blif files. Any time we want to process a new circuit, we should use this since otherwise we would not have the correct f_name.""" global max_bmc, f_name, d_name, initial_f_name, x_factor, init_initial_f_name, win_list,seed, sec_options global win_list, init_simp, po_map abc('fraig_restore') #clear out any residual fraig_store set_engines() #temporary init_simp = 1 win_list = [(0,.1),(1,.1),(2,.1),(3,.1),(4,.1),(5,-1),(6,-1),(7,.1)] #initialize winning engine list po_map = range(n_pos()) initialize() ## x_factor = 1 ## seed = 223 ## max_bmc = -1 if fname is None: print 'Type in the name of the aig file to be read in' s = raw_input() s = remove_spaces(s) ## print s else: s = fname if s[-4:] == '.aig': f_name = s[:-4] elif s[-5:] == '.blif': f_name = s[:-5] else: f_name = s s = s+'.aig' ## run_command(s) ## print s if s[-4:] == '.aig': run_command('&r %s;&put'%s) #warning: changes names to generic ones. else: #this is a blif file run_command('r %s'%s) abc('st;&get;&put') #changes names to generic ones for doing cec later. run_command('zero;w %s.aig'%f_name) set_globals() init_initial_f_name = initial_f_name = f_name print 'Initial f_name = %s'%f_name abc('addpi') #check_pos() #this removes constant outputs with a warning - #needed when using iso. Need another fix for using iso. run_command('fold') ps() return def read_file(): global win_list, init_simp, po_map read_file_quiet() ## ps() ## init_simp = 1 ## win_list = [(0,.1),(1,.1),(2,.1),(3,.1),(4,.1),(5,-1),(6,-1),(7,.1)] #initialize winning engine list ## po_map = range(n_pos()) def rf(): ## set_engines(4) #temporary read_file() def write_file(s): """this is the main method for writing the current circuit to an AIG file on disk. It manages the name of the file, by giving an extension (s). The file name 'f_name' keeps increasing as more extensions are written. A typical sequence is name, name_smp, name_smp_abs, name_smp_abs_spec, name_smp_abs_spec_final""" global f_name """Writes out the current file as an aig file using f_name appended with argument""" f_name = '%s_%s'%(f_name,s) ss = '%s.aig'%(f_name) print 'WRITING %s: '%ss, ps() abc('w '+ss) def bmc_depth(): """ Finds the number of BMC frames that the latest operation has used. The operation could be BMC, reachability interpolation, abstract, speculate. max_bmc is continually increased. It reflects the maximum depth of any version of the circuit including g ones, for which it is known that there is not cex out to that depth.""" global max_bmc c = cex_frame() if c > 0: b = c-1 else: b = n_bmc_frames() max_bmc = max(b,max_bmc) return max_bmc def null_status(): """ resets the status to the default values but note that the &space is changed""" abc('&get;&put') def set_max_bmc(b): """ Keeps increasing max_bmc which is the maximum number of time frames for which the current circuit is known to be UNSAT for""" global max_bmc max_bmc = max(b,max_bmc) def print_circuit_stats(): """Stardard way of outputting statistice about the current circuit""" global max_bmc i = n_pis() o = n_pos() l = n_latches() a = n_ands() s='ANDs' if a == -1: a = n_nodes() s = 'Nodes' b = max(max_bmc,bmc_depth()) c = cex_frame() if b>= 0: if c>=0: print 'PIs=%d,POs=%d,FF=%d,%s=%d,max depth=%d,CEX depth=%d'%(i,o,l,s,a,b,c) elif is_unsat(): print 'PIs=%d,POs=%d,FF=%d,%s=%d,max depth = infinity'%(i,o,l,s,a) else: print 'PIs=%d,POs=%d,FF=%d,%s=%d,max depth=%d'%(i,o,l,s,a,b) else: if c>=0: print 'PIs=%d,POs=%d,FF=%d,%s=%d,CEX depth=%d'%(i,o,l,s,a,c) else: print 'PIs=%d,POs=%d,FF=%d,%s=%d'%(i,o,l,s,a) def q(): exit() def is_unsat(): if prob_status() == 1: return True else: return False def is_sat(): if prob_status() == 0: return True else: return False def wc(file): """writes so that costraints are preserved explicitly""" abc('&get;&w %s'%file) def rc(file): """reads so that if constraints are explicit, it will preserve them""" abc('&r %s;&put'%file) #more complex functions: ________________________________________________________ #, abstract, pba, speculate, final_verify, dprove3 def timer(s): btime = time.clock() abc(s) print 'time = %f'%(time.clock() - btime) def med_simp(): x = time.time() abc("&get;&scl;&dc2;&lcorr;&dc2;&scorr;&fraig;&dc2;&put;dr") #abc("dc2rs") ps() print 'time = %f'%(time.time() - x) def simplify(): """Our standard simplification of logic routine. What it does depende on the problem size. For large problems, we use the &methods which use a simple circuit based SAT solver. Also problem size dictates the level of k-step induction done in 'scorr' The stongest simplification is done if n_ands < 20000. Then it used the clause based solver and k-step induction where |k| depends on the problem size """ set_globals() abc('&get;&scl;&lcorr;&put') p_40 = False n =n_ands() if n >= 70000: abc('&get;&scorr -C 0;&put') n =n_ands() if n >= 100000: abc('&get;&scorr -k;&put') if (70000 < n and n < 100000): p_40 = True abc("&get;&dc2;&put;dr;&get;&lcorr;&dc2;&put;dr;&get;&scorr;&fraig;&dc2;&put;dr") n = n_ands() ## if n<60000: if n < 80000: abc("&get;&scorr -F 2;&put;dc2rs") else: # n between 60K and 100K abc("dc2rs") n = n_ands() ## if (30000 < n and n <= 40000): if (60000 < n and n <= 70000): if not p_40: abc("&get;&dc2;&put;dr;&get;&lcorr;&dc2;&put;dr;&get;&scorr;&fraig;&dc2;&put;dr") abc("&get;&scorr -F 2;&put;dc2rs") else: abc("dc2rs") n = n_ands() ## if n <= 60000: if n <= 70000: abc('scl -m;drw;dr;lcorr;drw;dr') nn = max(1,n) m = int(min( 70000/nn, 16)) if m >= 1: j = 1 while j <= m: set_size() if j<8: abc('dc2') else: abc('dc2rs') abc('scorr -C 5000 -F %d'%j) if check_size(): break j = 2*j print 'ANDs=%d,'%n_ands(), if n_ands() >= .98 * nands: break continue if not check_size(): print '\n' return get_status() def simulate2(t): """Does rarity simulation. Simulation is restricted by the amount of memory it might use. At first wide but shallow simulation is done, followed by successively more narrow but deeper simulation. seed is globally initiallized to 113 when a new design is read in""" global x_factor, f_name, tme, seed btime = time.clock() diff = 0 while True: f = 100 w = 16 b = 16 r = 500 for k in range(9): #this controls how deep we go f = min(f*2, 3500) w = max(((w+1)/2)-1,1) abc('sim3 -F %d -W %d -N %d -R %d -B %d'%(f,w,seed,r,b)) seed = seed+23 if is_sat(): return 'SAT' if ((time.clock()-btime) > t): return 'UNDECIDED' def simulate(t): abc('&get') result = eq_simulate(t) return result def eq_simulate(t): """Simulation is restricted by the amount of memory it might use. At first wide but shallow simulation is done, followed by successively more narrow but deeper simulation. The aig to be simulated must be in the & space If there are equivalences, it will refine them. Otherwise it is a normal similation seed is globally initiallized to 113 when a new design is read in""" global x_factor, f_name, tme, seed btime = time.clock() diff = 0 while True: f = 5 w = 255 for k in range(9): f = min(f *2, 3500) r = f/20 w = max(((w+1)/2)-1,1) ## abc('&sim3 -R %d -W %d -N %d'%(r,w,seed)) abc('&sim -F %d -W %d -R %d'%(f,w,seed)) seed = seed+23 if is_sat(): return 'SAT' if ((time.clock()-btime) > t): return 'UNDECIDED' def generate_abs(n): """generates an abstracted model (gabs) from the greg file. The gabs file is automatically generated in the & space by &abs_derive. We store it away using the f_name of the problem being solved at the moment. The f_name keeps changing with an extension given by the latest operation done - e.g. smp, abs, spec, final, group. """ global f_name #we have a cex and we use this generate a new gabs file abc('&r %s_greg.aig; &abs_derive; &put; w %s_gabs.aig'%(f_name,f_name)) # do we still need the gabs file if n == 1: #print 'New abstraction: ', ps() return def refine_with_cex(): """Refines the greg (which contains the original problem with the set of FF's that have been abstracted). This uses the current cex to modify the greg file to reflect which regs are in the new current abstraction""" global f_name #print 'CEX in frame %d for output %d'%(cex_frame(),cex_po()) #abc('&r %s_greg.aig; &abs_refine -t; &w %s_greg.aig'%(f_name,f_name)) abc('&r %s_greg.aig;&w %s_greg_before.aig'%(f_name,f_name)) ## run_command('&abs_refine -s -M 25; &w %s_greg.aig'%f_name) run_command('&abs_refine -s; &w %s_greg.aig'%f_name) #print ' %d FF'%n_latches() return def iter_tempor(): na = n_ands() while True: abc('w save.aig') npi = n_pis() print npi abc('tempor -T 5 -F 8') print 'tempor = ', ps() x = trim() ## if n_pis() > 2*npi: ## abc('r save.aig') ## return 'UNDECIDED' abc('dr') print 'retime = ', ps() simplify() trim() print 'simplify -> trim = ', ps() if n_ands() > na: abc('r save.aig') ps() print 'No improvement' return 'UNDECIDED' na = n_ands() ps() if n_latches() == 0: return RESULT[check_sat()] def abstraction_refinement(latches_before,NBF,ratio=.75): """Subroutine of 'abstract' which does the refinement of the abstracted model, using counterexamples found by BMC or BDD reachability""" global x_factor, f_name, last_verify_time, x, win_list, last_winner, last_cex, t_init, j_last, sweep_time global cex_list, last_cx sweep_time = 2 if NBF == -1: F = 2000 else: F = 2*NBF print '\nIterating abstraction refinement' J = slps+intrps+pdrs+bmcs+sims print sublist(methods,J) last_verify_time = t = x_factor*max(50,max(1,2.5*G_T)) t = 1000 #temporary t = abs_time initial_verify_time = last_verify_time = t reg_verify = True print 'Verify time set to %d'%last_verify_time while True: #cex based refinement generate_abs(1) #generate new gabs file from refined greg file set_globals() latches_after = n_latches() if small_abs(ratio): break t = last_verify_time yy = time.time() abc('w %s_beforerpm.aig'%f_name) rep_change = reparam() #new - must do reconcile after to make cex compatible abc('w %s_afterrpm.aig'%f_name) ## if reg_verify: status = verify(J,t) ## else: ## status = pord_1_2(t) ############### if status == Sat_true: print 'Found true cex' reconcile(rep_change) return Sat_true if status == Unsat: return status if status == Sat: abc('write_status %s_after.status'%f_name) reconcile(rep_change) # makes the cex compatible with original before reparam and puts original in work space abc('write_status %s_before.status'%f_name) refine_with_cex() if is_sat(): # if cex can't refine, status is set to Sat_true print 'Found true cex in output %d'%cex_po() return Sat_true else: continue else: break print '**** Latches reduced from %d to %d'%(latches_before, n_latches()) return Undecided_reduction def small_abs(ratio=.75): """ tests is the abstraction is too large""" return ((rel_cost_t([pis_before_abs,latches_before_abs, ands_before_abs])> -.1) or (n_latches() >= ratio*latches_before_abs)) ##def abstract(if_bip=True): ## global ratio ## if if_bip: ## return abstractb(True) #old method using abstraction refinement ## else: ## return abstractb(False) #not using bip and reachx def abstractb(): """ abstracts using N Een's method 3 - cex/proof based abstraction. The result is further refined using simulation, BMC or BDD reachability. abs_ratio is the the limit for accepting an abstraction""" global G_C, G_T, latches_before_abs, x_factor, last_verify_time, x, win_list, j_last, sims global latches_before_abs, ands_before_abs, pis_before_abs, abs_ratio if ifbip < 1: print 'using ,abs in old way' tt = time.time() j_last = 0 set_globals() #win_list = [] latches_before_abs = n_latches() ands_before_abs = n_ands() pis_before_abs = n_real_inputs() abc('w %s_before_abs.aig'%f_name) print 'Start: ', ps() funcs = [eval('(pyabc_split.defer(initial_abstract)())')] # fork off BMC3 and PDRm along with initial abstraction t = 10000 #want to run as long as initial abstract takes. ## J = sims+pdrs+bmcs+intrps J = slps+pdrs+bmcs+intrps J = modify_methods(J,1) ## if n_latches() < 80: ## J = J + [4] funcs = create_funcs(J,t) + funcs mtds = sublist(methods,J) + ['initial_abstract'] #important that initial_abstract goes last m,result = fork_last(funcs,mtds) if is_sat(): print 'Found true counterexample in frame %d'%cex_frame() return Sat_true if is_unsat(): return Unsat ## set_max_bmc(NBF) NBF = bmc_depth() print 'Abstraction good to %d frames'%max_bmc #note when things are done in parallel, the &aig is not restored!!! abc('&r %s_greg.aig; &w initial_greg.aig; &abs_derive; &put; w initial_gabs.aig; w %s_gabs.aig'%(f_name,f_name)) set_max_bmc(NBF) print 'Initial abstraction: ', ps() abc('w %s_init_abs.aig'%f_name) latches_after = n_latches() ## if latches_after >= .90*latches_before_abs: #the following should match similar statement ## if ((rel_cost_t([pis_before_abs, latches_before_abs, ands_before_abs])> -.1) or ## (latches_after >= .75*latches_before_abs)): if small_abs(abs_ratio): abc('r %s_before_abs.aig'%f_name) print "Little reduction!" print 'Abstract time wasted = %0.2f'%(time.time()-tt) return Undecided_no_reduction sims_old = sims sims=sims[:1] #make it so that rarity sim is not used since it can't find a cex result = abstraction_refinement(latches_before_abs, NBF,abs_ratio) sims = sims_old if result <= Unsat: return result ## if n_latches() >= .90*latches_before_abs: ## if ((rel_cost_t([pis_before_abs, latches_before_abs, ands_before_abs])> -.1) or (latches_after >= .90*latches_before_abs)): ## if rel_cost_t([pis_before_abs,latches_before_abs, ands_before_abs])> -.1: if small_abs(abs_ratio): #r is ratio of final to initial latches in absstraction. If greater then True abc('r %s_before_abs.aig'%f_name) #restore original file before abstract. print "Little reduction! ", print 'Abstract time wasted = %0.2f'%(time.time()-tt) result = Undecided_no_reduction return result #new else: write_file('abs') #this is only written if it was not solved and some change happened. print 'Abstract time = %0.2f'%(time.time()-tt) return result def initial_abstract_old(): global G_C, G_T, latches_before_abs, x_factor, last_verify_time, x, win_list set_globals() time = max(1,.1*G_T) abc('&get;,bmc -vt=%f'%time) set_max_bmc(bmc_depth()) c = 2*G_C f = max(2*max_bmc,20) b = min(max(10,max_bmc),200) t = x_factor*max(1,2*G_T) s = min(max(3,c/30000),10) # stability between 3 and 10 cmd = '&get;,abs -bob=%d -stable=%d -timeout=%d -vt=%d -depth=%d'%(b,s,t,t,f) ## print cmd print 'Running initial_abstract with bob=%d,stable=%d,time=%d,depth=%d'%(b,s,t,f) abc(cmd) abc('&w %s_greg.aig'%f_name) ## ps() def initial_abstract(t=100): global G_C, G_T, latches_before_abs, x_factor, last_verify_time, x, win_list, max_bmc, ifbip set_globals() time = max(1,.1*G_T) time = min(time,t) abc('&get;,bmc -vt=%f'%time) set_max_bmc(bmc_depth()) c = 2*G_C f = max(2*max_bmc,20) b = min(max(10,max_bmc),200) t1 = x_factor*max(1,2*G_T) t = max(t1,t) s = min(max(3,c/30000),10) # stability between 3 and 10 cmd = '&get;,abs -bob=%d -stable=%d -timeout=%d -vt=%d -depth=%d'%(b,s,t,t,f) if ifbip == 2: cmd = '&get;,abs -bob=%d -stable=%d -timeout=%d -vt=%d -depth=%d -dwr=%s_vabs'%(b,s,t,t,f,f_name) print 'Using -dwr=%s_vabs'%f_name ## print cmd print 'Running initial_abstract with bob=%d,stable=%d,time=%d,depth=%d'%(b,s,t,f) abc(cmd) bmc_depth() ## pba_loop(max_bmc+1) abc('&w %s_greg.aig'%f_name) return max_bmc def abs_m(): set_globals() y = time.time() nl = n_abs_latches() #initial set of latches c = 2*G_C t = x_factor*max(1,2*G_T) #total time bmd = bmc_depth() if bmd < 0: abc('bmc3 -T 2') #get initial depth estimate bmd = bmc_depth() f = bmd abc('&get') y = time.time() cmd = '&abs_cba -v -C %d -T %0.2f -F %d'%(c,.8*t,bmd) #initial absraction ## print '\n%s'%cmd abc(cmd) b_old = b = n_bmc_frames() f = min(2*bmd,max(bmd,1.6*b)) print 'cba: latches = %d, depth = %d'%(n_abs_latches(),b) ## print n_bmc_frames() while True: if (time.time() - y) > .9*t: break nal = n_abs_latches() cmd = '&abs_cba -v -C %d -T %0.2f -F %d'%(c,.8*t,f) #f is 2*bmd and is the maximum number of frames allowed ## print '\n%s'%cmd abc(cmd) ## print n_bmc_frames() b_old = b b = n_bmc_frames() nal_old = nal nal = n_abs_latches() #nal - nal_old is the number of latches added by cba #b - b_old is the additional time frames added by cba f = min(2*bmd,max(bmd,1.6*b)) #may be this should just be bmd f = max(f,1.5*bmd) print 'cba: latches = %d, depth = %d'%(nal,b) if ((nal == nal_old) and (b >= 1.5*b_old) and b >= 1.5*bmd): """ Went at least bmd depth and saw too many frames without a cex (ideally should know how many frames without a cex) """ print 'Too many frames without cex' break if b > b_old: #if increased depth continue if nal > .9*nl: # try to minimize latches ## cmd = '&abs_pba -v -S %d -F %d -T %0.2f'%(b,b+2,.2*t) cmd = '&abs_pba -v -F %d -T %0.2f'%(b+2,.2*t) ## print '\n%s'%cmd abc(cmd) b = n_bmc_frames() nal_old = nal nal = n_abs_latches() print 'pba: latches = %d, depth = %d'%(nal,b) ## print n_bmc_frames() if nal_old < nal: #if latches increased there was a cex continue if nal > .9*nl: # if still too big return continue ## b = n_bmc_frames() cmd = '&abs_pba -v -F %d -T %0.2f'%(b+2,.2*t) ## print '\n%s'%cmd abc(cmd) b = n_bmc_frames() print 'pba: latches = %d, depth = %d'%(n_abs_latches(),b) ## print n_bmc_frames() print 'Total time = %0.2f'%(time.time()-y) def n_abs_latches(): abc('&w pba_temp.aig') #save the &space abc('&abs_derive;&put') abc('&r pba_temp.aig') return n_latches() def pba_loop(F): n = n_abs_latches() while True: run_command('&abs_pba -v -C 0 -F %d'%F) abc('&w pba_temp.aig') abc('&abs_derive;&put') abc('&r pba_temp.aig') N = n_latches() ## if n == N or n == N+1: ## break ## elif N > n: if N > n: print 'cex found' break def ssm(options=''): """ Now this should be the same as super_prove(1) """ y = time.time() result = prove_part_1() # simplify first if result == 'UNDECIDED': result = ss(options) print 'Total time taken on file %s by function ssm(%s) = %d sec.'%(initial_f_name,options,(time.time() - y)) return result def ssmg(): return ssm('g') def ssmf(): return ssm('f') def ss(options=''): """ Alias for super_sec This is the preferred command if the problem (miter) is suspected to be a SEC problem """ global max_bmc, init_initial_f_name, initial_f_name,win_list, last_verify_time, sec_options sec_options = options print '\n*************Executing speculate************' y = time.time() abc('scl') result = speculate() # if result is 1 then it is a real SAT since we did not do anything before if result > 2: #save the result and read in with /rf so that files are initialized correctly if not '_spec' in f_name: write_file('spec') #make sure we do not overwrite original file read_file_quiet('%s'%f_name) #this resets f_name and initial_f_name etc. print '\n*************Executing super_prove ************' print 'New f_name = %s'%f_name result = sp() if result == 'SAT': result = 'UNDECIDED' #because speculation was done initially. elif result == 1: result = 'SAT' else: result = RESULT[result] print 'Total time taken on file %s by function ss(%s) = %d sec.'%(initial_f_name,options,(time.time() - y)) return result def quick_sec(t): ## fb_name = f_name[:-3]+'New' ## abc('&get;&miter -s %s.aig;&put'%fb_name) ## abc('w %s.%s_miter.aig'%(f_name,fb_name)) quick_simp() verify(slps+ pdrs+bmcs+intrps,t) if is_unsat(): return 'UNSAT' if is_sat(): return 'SAT' else: return'UNDECIDED' def pre_sec(): """ put files to be compared into Old and New aigs. Simplify, but turn off reparameterization so that PIs in Old and New match after simplification. """ global trim_allowed ## trim_allowed = False ## print 'trim allowed = ',trim_allowed print 'First file: ', read_file_quiet() #note - reads into & space and then does &put ps() prs(False) ps() abc('&w Old.aig') print 'Second file: ', read_file_quiet() ps() prs(False) ps() abc('&w New.aig') def cec(): print 'Type in the name of the aig file to be compared against' s = raw_input() s = remove_spaces(s) if not 'aig' in s: s = s+'.aig' run_command("&get;&cec -v %s"%s) def sec(B_part,options): """ Use this for AB filtering and not sup_sec Use pp_sec to make easy names for A and B, namely Old and New. This assumes that the original aig (renamed A_name below) is already read into the working space. Then we form a miter using &miter between two circuits, A_name, and B_name. We then do speculate immediately. Optionally we could simplify A and B and then form the miter and start from there. The only difference in speculate is that &srm2 is used, which only looks at equivalences where one comes from A and one from B. Options are -a and -b which says use only flops in A or in B or both. The switch sec_sw controls what speculate does when it generates the SRM. """ global f_name,sec_sw, A_name, B_name, sec_options yy = time.time() A_name = f_name # Just makes it so that we can refer to A_name later in &srm2 B_name = B_part run_command('&get; &miter -s %s.aig; &put'%B_name) ## abc('orpos') f_name = A_name+'_'+B_name+'_miter' # reflect that we are working on a miter. abc('w %s.aig'%f_name) print 'Miter = ', ps() sec_options = options if sec_options == 'ab': sec_options = 'l' #it will be changed to 'ab' after &equiv sec_sw = True result = speculate() sec_options = '' sec_sw = False if result <= Unsat: result = RESULT[result] else: result = sp() if result == 'SAT': result = 'UNDECIDED' print 'Total time = %d'%(time.time() - yy) return result def filter(opts): global A_name,B_name ## print 'Filtering with options = %s'%opts """ This is for filter which effectively only recognizes options -f -g""" if (opts == '' or opts == 'l'): #if 'l' this is used only for initial &equiv2 to get initial equiv creation return if opts == 'ab': print A_name , print B_name run_command('&ps') run_command('&filter -f %s.aig %s.aig'%(A_name,B_name)) return if not opts == 'f': opts = 'g' run_command('&filter -%s'%opts) def check_if_spec_first(): global sec_sw, A_name, B_name, sec_options, po_map set_globals() t = max(1,.5*G_T) r = max(1,int(t)) abc('w check_save.aig') abc('&w check_and.aig') abc("&get; &equiv3 -v -F 20 -T %f -R %d"%(t,5*r)) filter('g') abc("&srm; r gsrm.aig") print 'Estimated # POs = %d for initial speculation'%n_pos() result = n_pos() > max(50,.25*n_latches()) abc('r check_save.aig') abc('&r check_and.aig') return result def initial_speculate(sec_opt=''): global sec_sw, A_name, B_name, sec_options, po_map set_globals() if sec_options == '': sec_options = sec_opt t = max(1,.5*G_T) r = max(1,int(t)) print 'Initial sec_options = %s'%sec_options if sec_options == 'l': cmd = "&get; &equiv3 -lv -F 20 -T %f -R %d"%(3*t,15*r) else: cmd = "&get; &equiv3 -v -F 20 -T %f -R %d"%(3*t,15*r) print cmd abc(cmd) ## print 'AND space after &equiv3: ', run_command('&ps') if (sec_options == 'l'): if sec_sw: sec_options = 'ab' else: sec_options = 'f' ## print 'A_name: ', ## run_command('r %s.aig;ps'%A_name) ## print 'B_name: ', ## run_command('r %s.aig;ps'%B_name) filter(sec_options) abc('&w initial_gore.aig') ## print 'Running &srm' if sec_sw: print 'miter: ', run_command('&ps') print 'A_name: ', run_command('r %s.aig;ps'%A_name) print 'B_name: ', run_command('r %s.aig;ps'%B_name) cmd = "&srm2 -%s %s.aig %s.aig; r gsrm.aig; w %s_gsrm.aig; &w %s_gore.aig"%(sec_options,A_name,B_name,f_name,f_name) abc(cmd) po_map = range(n_pos()) return else: ## abc('&r %s_gore.aig; &srm ; r gsrm.aig; w %s_gsrm.aig'%(f_name,f_name)) cmd = "&srm; r gsrm.aig; w %s_gsrm.aig; &w %s_gore.aig"%(f_name,f_name) ## print 'Running %s'%cmd abc(cmd) print 'done with &srm' po_map = range(n_pos()) if sec_options == '': if n_pos() > 200: sec_options = 'g' print 'sec_options set to %s'%'g' abc('&r %s_gore.aig'%f_name) filter(sec_options) print 'Running &srm' cmd = "&srm; r gsrm.aig; ps;w %s_gsrm.aig;&ps; &w %s_gore.aig"%(f_name,f_name) print 'Running %s'%cmd abc(cmd) po_map = range(n_pos()) def test_against_original(): '''tests whether we have a cex hitting an original PO''' abc('&w %s_save.aig'%f_name) #we preserve whatever was in the & space abc('&r %s_gore.aig'%f_name) #This is the original abc('testcex') PO = cex_po() ## print 'test_against original gives PO = %d'%PO abc('&r %s_save.aig'%f_name) if PO > -1: ## print 'cex fails an original PO' return True else: abc('write_status %s_status.status'%f_name) return False def set_cex_po(n=0): """ if cex falsifies a non-real PO return that PO first, else see if cex_po is one of the original, then take it next else return -1 which means that the cex is not valid and hence an error. parameter n = 1 means test the &-space """ global n_pos_before, n_pos_proved #these refer to real POs if n == 0: abc('testcex -a -O %d'%(n_pos_before-n_pos_proved)) #test regular AIG space else: abc('testcex -O %d'%(n_pos_before-n_pos_proved)) #test the &-AIG PO = cex_po() ## print 'cex_po = %d, n_pos_before = %d, n_pos_proved = %d'%(PO, n_pos_before, n_pos_proved) if PO >= (n_pos_before - n_pos_proved): #cex_po is not an original ## print '1. cex PO = %d'%PO return PO # after original so take it. if n == 0: abc('testcex -a') #test regular else: abc('testcex') #test &space PO = cex_po() ## print '2. cex PO = %d'%PO cx = cex_get() if PO > -1: if test_against_original(): #this double checks that it is really an original PO cex_put(cx) ## print 'test_against_original was valid' return PO else: print '1. PO not valid' return -1 #error if PO < 0 or PO >= (n_pos_before - n_pos_proved): # not a valid cex because already tested outside original. print '2. PO is not valid' PO = -1 #error ## print '3. cex PO = %d'%PO return PO def speculate(): """Main speculative reduction routine. Finds candidate sequential equivalences and refines them by simulation, BMC, or reachability using any cex found. """ global G_C,G_T,n_pos_before, x_factor, n_latches_before, last_verify_time, trim_allowed, n_pos_before global t_init, j_last, sec_sw, A_name, B_name, sec_options, po_map, sweep_time, sims, cex_list, n_pos_proved,ifpord1 global last_cx last_cx = 0 ifpord1 = 1 initial_po_size = last_srm_po_size = n_pos() initial_sizes = [n_pis(),n_pos(),n_latches(),n_ands()] if sec_sw: print 'A_name = %s, B_name = %s, f_name = %s, sec_options = %s'%(A_name, B_name, f_name, sec_options) elif n_ands()> 36000 and sec_options == '': sec_options = 'g' print 'sec_options set to "g"' def refine_with_cex(): """Refines the gore file to reflect equivalences that go away because of cex""" global f_name abc('write_status %s_before.status'%f_name) abc('&r %s_gore.aig; &resim -m'%f_name) filter(sec_options) run_command('&w %s_gore.aig'%f_name) return def set_cex(lst): """ assumes only one in lst """ for j in range(len(lst)): cx = lst[j] if cx == None: continue else: cex_put(cx) break def generate_srm(): """generates a speculated reduced model (srm) from the gore file""" global f_name, po_map, sec_sw, A_name, B_name, sec_options, n_pos_proved ## print 'Generating' pos = n_pos() ab = n_ands() abc('w %s_oldsrm.aig'%f_name) #save for later purposes if sec_sw: run_command('&r %s_gore.aig; &srm2 -%s %s.aig %s.aig; r gsrm.aig; w %s_gsrm.aig'%(f_name,sec_options,A_name,B_name,f_name)) else: abc('&r %s_gore.aig; &srm ; r gsrm.aig; w %s_gsrm.aig'%(f_name,f_name)) #do we still need to write the gsrm file ## ps() po_map = range(n_pos()) ps() n_pos_proved = 0 return 'OK' n_pos_before = n_pos() n_pos_proved = 0 n_latches_before = n_latches() set_globals() ## t = max(1,.5*G_T)#irrelevant ## r = max(1,int(t)) t = 1000 j_last = 0 J = slps+sims+pdrs+bmcs+intrps J = modify_methods(J,1) funcs = [eval('(pyabc_split.defer(initial_speculate)())')] funcs = create_funcs(J,10000)+funcs #want other functins to run until initial speculate stops mtds = sublist(methods,J) + ['initial_speculate'] #important that initial_speculate goes last fork_last(funcs,mtds) ## ps() if is_unsat(): return Unsat if is_sat(): return Sat_true if n_pos_before == n_pos(): print 'No new outputs. Quitting speculate' return Undecided_no_reduction # return result is unknown ## cmd = 'lcorr;&get;&trim -i;&put;w %s_gsrm.aig'%f_name #print 'Executing %s'%cmd abc('w initial_gsrm.aig') ## ps() ## abc(cmd) print 'Initial speculation: ', ps() if n_latches() == 0: return check_sat() if sec_options == 'l' and sec_sw: sec_options = 'ab' #finished with initial speculate with the 'l' option print "sec_options set to 'ab'" elif sec_options == 'l': sec_options = 'f' print "sec_options set to 'f'" po_map = range(n_pos()) #we need this because the initial_speculate is done in parallel and po_map is not passed back. npi = n_pis() set_globals() if is_sat(): return Sat_true simp_sw = init = True print '\nIterating speculation refinement' sims_old = sims sims = sims[:1] J = slps+sims+pdrs+intrps+bmcs J = modify_methods(J) ## print sublist(methods,J) t = max(50,max(1,2*G_T)) last_verify_time = t print 'Verify time set to %d'%last_verify_time reg_verify = True ref_time = time.time() sweep_time = 2 ifpord1=1 while True: # refinement loop set_globals() yy = time.time() if not init: abc('r %s_gsrm.aig'%f_name) #this is done only to set the size of the previous gsrm. abc('w %s_gsrm_before.aig'%f_name) set_size() result = generate_srm() last_srm_po_size = n_pos() yy = time.time() # if the size of the gsrm did not change after generating a new gsrm # and if the cex is valid for the gsrm, then the only way this can happen is if # the cex_po is an original one. if check_size(): #same size before and after if check_cex(): #valid cex failed to refine possibly if 0 <= cex_po() and cex_po() < (n_pos_before - n_pos_proved): #original PO print 'Found cex in original output = %d'%cex_po() print 'Refinement time = %s'%convert(time.time() - ref_time) return Sat_true elif check_same_gsrm(f_name): #if two gsrms are same, then failed to refine print 'CEX failed to refine' return Error else: print 'not a valid cex' return Error if n_latches() == 0: print 'Refinement time = %s'%convert(time.time() - ref_time) return check_sat() init = False # make it so that next time it is not the first time through if not t == last_verify_time: # heuristic that if increased last verify time, # then try pord_all t = last_verify_time if reg_verify: t_init = (time.time() - yy)/2 #start poor man's concurrency at last cex fime found t_init = min(10,t_init) ## temporary reg_verify = False #will cause pord_all to be used next ## print 'pord_all turned on' t = last_verify_time print 'Verify time set to %d'%t abc('w %s_beforerpm.aig'%f_name) rep_change = reparam() #must be paired with reconcile below if cex abc('w %s_afterrpm.aig'%f_name) if reg_verify: result = verify(J,t) else: result = pord_1_2(t) if result == Unsat: print 'UNSAT' print 'Refinement time = %s'%convert(time.time() - ref_time) return Unsat if result < Unsat: if not reg_verify: set_cex(cex_list) ## if reg_verify: reconcile(rep_change) #end of pairing with reparam() assert (npi == n_cex_pis()),'ERROR: #pi = %d, #cex_pi = %d'%(npi,n_cex_pis()) abc('&r %s_gore.aig;&w %s_gore_before.aig'%(f_name,f_name)) #we are making sure that none of the original POs fail if reg_verify: PO = set_cex_po(0) #testing the regular space else: abc('&r %s_gsrm.aig'%f_name) PO = set_cex_po(1) # test against the &space. print 'cex_PO is %d'%PO if (-1 < PO and PO < (n_pos_before-n_pos_proved)): print 'Found cex in original output = %d'%cex_po() print 'Refinement time = %s'%convert(time.time() - ref_time) return Sat_true if PO == -1: return Error refine_with_cex() #change the number of equivalences continue elif (is_unsat() or n_pos() == 0): print 'UNSAT' print 'Refinement time = %s'%convert(time.time() - ref_time) return Unsat else: #if undecided, record last verification time print 'Refinement returned undecided in %d sec.'%t last_verify_time = t #########################added if reg_verify: #try one last time with parallel POs cex detection (find_cex_par) if not already tried abc('r %s_beforerpm.aig'%f_name) # to continue refinement, need to restore original t_init = min(last_verify_time,(time.time() - yy)/2) #start poor man's concurrency at last cex fime found t_init = min(10,t_init) reg_verify = False t = last_verify_time # = 2*last_verify_time abc('w %s_beforerpm.aig'%f_name) rep_change = reparam() #must be paired with reconcile()below abc('w %s_afterrpm.aig'%f_name) result = pord_1_2(t) #main call to verification if result == Unsat: print 'UNSAT' print 'Refinement time = %s'%convert(time.time() - ref_time) return Unsat if is_sat(): assert result == get_status(),'result: %d, status: %d'%(result,get_status()) set_cex(cex_list) reconcile(rep_change) abc('&r %s_gsrm.aig'%f_name) PO = set_cex_po(1) #testing the & space if (-1 < PO and PO < (n_pos_before-n_pos_proved)): print 'Found cex in original output = %d'%cex_po() print 'Refinement time = %s'%convert(time.time() - ref_time) return Sat_true if PO == -1: return Error refine_with_cex() #change the number of equivalences continue elif is_unsat(): print 'UNSAT' print 'Refinement time = %s'%convert(time.time() - ref_time) return Unsat else: #if undecided, record last verification time last_verify_time = t print 'UNDECIDED' break ################### added else: break sims = sims_old print 'UNDECIDED' print 'Refinement time = %s'%convert(time.time() - ref_time) ## if last_srm_po_size == initial_po_size: #essentially nothing happened. last_srm_po_size will be # POs in last srm. if initial_sizes == [n_pis(),n_pos(),n_latches(),n_ands()]: return Undecided_no_reduction #thus do not write spec file else: #file was changed, so some speculation happened. If we find a cex later, need to know this. write_file('spec') return Undecided_reduction def simple_sat(t=10000): y = time.time() J = [14,2,7,9,30,31,26,5] #5 is pre_simp funcs = create_funcs(J,t) mtds =sublist(methods,J) fork_last(funcs,mtds) result = get_status() if result > Unsat: write_file('smp') result = verify(slps+[14,2,3,7,9,30,31,26],t) print 'Time for simple_sat = %0.2f'%(time.time()-y) return RESULT[result] def simple(t=10000): y = time.time() ## J = [14,1,2,7,9,23,30,5] #5 is pre_simp ## funcs = create_funcs(J,t) ## mtds =sublist(methods,J) ## fork_last(funcs,mtds) ## result = get_status() ## if result > Unsat: ## write_file('smp') ## J = slps+bmcs+pdrs+intrps ## J = modify_methods(J) ## result = verify(J,t) J = slps+bmcs+pdrs+intrps J = modify_methods(J) result = verify(J,t) ## print 'Time for simple = %0.2f'%(time.time()-y) return RESULT[result] def simple_bip(t=1000): y = time.time() J = [14,1,2,30,5] #5 is pre_simp funcs = create_funcs(J,t) mtds =sublist(methods,J) fork_last(funcs,mtds) result = get_status() if result > Unsat: write_file('smp') result = verify(slps+[14,1,2,30],t) print 'Time for simple_bip = %0.2f'%(time.time()-y) return RESULT[result] def simple_prove(t=1000): y = time.time() J = [7,9,23,30,5] funcs = create_funcs(J,t) mtds =sublist(methods,J) fork_last(funcs,mtds) result = get_status() if result > Unsat: write_file('smp') result = verify(slps+[7,9,23,30],t) print 'Time for simple_prove = %0.2f'%(time.time()-y) return RESULT[result] def check_same_gsrm(f): ## return False #disable the temporarily until can figure out why this is there """checks gsrm miters before and after refinement and if equal there is an error""" global f_name abc('r %s_gsrm.aig'%f) ## ps() run_command('miter -c %s_gsrm_before.aig'%f) ## ps() abc('&get; ,bmc -timeout=5') result = True #if the same if is_sat(): #if different result = False abc('r %s_gsrm.aig'%f) ## ps() return result def check_cex(): """ check if the last cex still asserts one of the outputs. If it does then we have an error""" global f_name abc('read_status %s_before.status'%f_name) abc('&r %s_gsrm_before.aig'%f_name) ## abc('&r %s_gsrm.aig'%f_name) run_command('testcex') print 'cex po = %d'%cex_po() return cex_po() >=0 def set_size(): """Stores the problem size of the current design. Size is defined as (PIs, POs, ANDS, FF)""" global npi, npo, nands, nff, nmd npi = n_pis() npo = n_pos() nands = n_ands() nff = n_latches() nmd = max_bmc #print npi,npo,nands,nff def check_size(): """Assumes the problem size has been set by set_size before some operation. This checks if the size was changed Size is defined as (PIs, POs, ANDS, FF, max_bmc) Returns TRUE is size is the same""" global npi, npo, nands, nff, nmd #print n_pis(),n_pos(),n_ands(),n_latches() result = ((npi == n_pis()) and (npo == n_pos()) and (nands == n_ands()) and (nff == n_latches()) ) return result def inferior_size(): """Assumes the problem size has been set by set_size beore some operation. This checks if the new size is inferior (larger) to the old one Size is defined as (PIs, POs, ANDS, FF)""" global npi, npo, nands, nff result = ((npi < n_pis()) or (npo < n_pos()) or (nands < n_ands()) ) return result def quick_verify(n): """Low resource version of final_verify n = 1 means to do an initial simplification first. Also more time is allocated if n =1""" global last_verify_time trim() if n == 1: simplify() if n_latches == 0: return check_sat() trim() if is_sat(): return Sat_true #print 'After trimming: ', #ps() set_globals() last_verify_time = t = max(1,.4*G_T) if n == 1: last_verify_time = t = max(1,2*G_T) print 'Verify time set to %d '%last_verify_time J = [18] + intrps+bmcs+pdrs+sims status = verify(J,t) return status def process_status(status): """ if there are no FF, the problem is combinational and we still have to check if UNSAT""" if n_latches() == 0: return check_sat() return status def get_status(): """this simply translates the problem status encoding done by ABC (-1,0,1)=(undecided,SAT,UNSAT) into the status code used by our python code. -1,0,1 => 3,0,2 """ if n_latches() == 0: return check_sat() status = prob_status() #interrogates ABC for the current status of the problem. # 0 = SAT i.e. Sat_reg = 0 so does not have to be changed. if status == 1: status = Unsat if status == -1: #undecided status = Undecided return status def reparam(): """eliminates PIs which if used in abstraction or speculation must be restored by reconcile and the cex made compatible with file beforerpm""" ## return rep_change = False n = n_pis() ## abc('w t1.aig') abc('&get;,reparam -aig=%s_rpm.aig; r %s_rpm.aig'%(f_name,f_name)) ## abc('w t2.aig') ## abc('testcex') if n_pis() == 0: print 'Number of PIs reduced to 0. Added a dummy PI' abc('addpi') nn = n_pis() if nn < n: print 'Reparam: PIs %d => %d'%(n,nn) rep_change = True return rep_change def reconcile(rep_change): """used to make current cex compatible with file before reparam() was done. However, the cex may have come from extracting a single output and verifying this. Then the cex_po is 0 but the PO it fails could be anything. So testcex rectifies this.""" global n_pos_before, n_pos_proved ## print 'rep_change = %s'%rep_change if rep_change == False: return abc('&r %s_beforerpm.aig; &w tt_before.aig'%f_name) abc('write_status %s_after.status;write_status tt_after.status'%f_name) abc('&r %s_afterrpm.aig;&w tt_after.aig'%f_name) POa = set_cex_po(1) #this should set cex_po() to correct PO. A 1 here means it uses &space to check abc('reconcile %s_beforerpm.aig %s_afterrpm.aig'%(f_name,f_name)) # reconcile modifies cex and restores work AIG to beforerpm abc('write_status %s_before.status;write_status tt_before.status'%f_name) POb = set_cex_po() if POa != POb: abc('&r %s_beforerpm.aig; &w tt_before.aig'%f_name) abc('&r %s_afterrpm.aig; &w tt_after.aig'%f_name) print 'cex PO afterrpm = %d not = cex PO beforerpm = %d'%(POa,POb) assert POa > -1, 'cex did not assert any output' def reconcile_all(lst, rep_change): """reconciles the list of cex's""" global f_name, n_pos_before, n_pos_proved if rep_change == False: return lst list = [] for j in range(len(lst)): cx = lst[j] if cx == None: continue cex_put(cx) reconcile(rep_change) list = list + [cex_get()] return list def try_rpm(): """rpm is a cheap way of doing reparameterization and is an abstraction method, so may introduce false cex's. It finds a minimum cut between the PIs and the main sequential logic and replaces this cut by free inputs. A quick BMC is then done, and if no cex is found, we assume the abstraction is valid. Otherwise we revert back to the original problem before rpm was tried.""" global x_factor if n_ands() > 30000: return set_globals() pis_before = n_pis() abc('w %s_savetemp.aig'%f_name) abc('rpm') result = 0 if n_pis() < .5*pis_before: bmc_before = bmc_depth() #print 'running quick bmc to see if rpm is OK' t = max(1,.1*G_T) #abc('bmc3 -C %d, -T %f'%(.1*G_C, t)) abc('&get;,bmc -vt=%f'%t) if is_sat(): #rpm made it sat by bmc test, so undo rpm abc('r %s_savetemp.aig'%f_name) else: trim() print 'WARNING: rpm reduced PIs to %d. May make SAT.'%n_pis() result = 1 else: abc('r %s_savetemp.aig'%f_name) return result def verify(J,t): """This method is used for finding a cex during refinement, but can also be used for proving the property. t is the maximum time to be used by each engine J is the list of methods to run in parallel. See FUNCS for list""" global x_factor, final_verify_time, last_verify_time, methods set_globals() t = int(max(1,t)) J = modify_methods(J) mtds = sublist(methods,J) print mtds #print J,t F = create_funcs(J,t) (m,result) = fork_break(F,mtds,'US') #FORK here assert result == get_status(),'result: %d, status: %d'%(result,get_status()) return result def dsat_all(t=100,c=100000): print 't=%d,c=%d'%(t,c) N = n_pos() abc('&get') J = range(N) ttt = time.time() J.reverse() abc('w %s_temp.aig'%f_name) for j in J: tt = time.time() abc('r %s_temp.aig'%f_name) run_command('cone -O %d; dc2; dsat -C %d'%(j,c)) if is_unsat(): print 'Output %d is %s'%(j,RESULT[2]), else: print 'Output %d is %s'%(j,RESULT[3]), T = time.time() -tt print 'time = %0.2f'%T if time.time() - tt > t: break print 'Total time = %0.2f'%(time.time() - ttt) def check_sat(): """This is called if all the FF have disappeared, but there is still some logic left. In this case, the remaining logic may be UNSAT, which is usually the case, but this has to be proved. The ABC command 'dsat' is used fro combinational problems""" if not n_latches() == 0: print 'circuit is not combinational' return Undecided ## print 'Circuit is combinational - checking with dsat' abc('&get') #save the current circuit abc('orpos;dsat -C %d'%G_C) if is_sat(): abc('&put') if n_pos() == 1: return Sat_true else: return Undecided_no_reduction #some POs could be unsat. elif is_unsat(): return Unsat else: abc('&put') #restore return Undecided_no_reduction def try_era(s): """era is explicit state enumeration that ABC has. It only works if the number of PIs is small, but there are cases where it works and nothing else does""" if n_pis() > 12: return cmd = '&get;&era -mv -S %d;&put'%s print 'Running %s'%cmd run_command(cmd) def try_induction(C): """Sometimes proving the property directly using induction works but not very often. For 'ind' to work, it must have only 1 output, so all outputs are or'ed together temporarily""" return Undecided_reduction print '\n***Running induction' abc('w %s_temp.aig'%f_name) abc('orpos; ind -uv -C %d -F 10'%C) abc('r %s_savetemp.aig'%f_name) status = prob_status() if not status == 1: return Undecided_reduction print 'Induction succeeded' return Unsat ##def final_verify_recur(K): ## """During prove we make backups as we go. These backups have increasing abstractions done, which can cause ## non-verification by allowing false counterexamples. If an abstraction fails with a cex, we can back up to ## the previous design before the last abstraction and try to proceed from there. K is the backup number we ## start with and this decreases as the backups fails. For each backup, we just try final_verify. ## If ever we back up to 0, which is the backup just after simplify, we then try speculate on this. This often works ## well if the problem is a SEC problem where there are a lot of equivalences across the two designs.""" ## global last_verify_time ## #print 'Proving final_verify_recur(%d)'%K ## last_verify_time = 2*last_verify_time ## print 'Verify time increased to %d'%last_verify_time ## for j in range(K): ## i = K-(j+1) ## abc('r %s_backup_%d.aig'%(initial_f_name,i)) ## if ((i == 0) or (i ==2)): #don't try final verify on original last one ## status = prob_status() ## break ## print '\nVerifying backup number %d:'%i, ## #abc('r %s_backup_%d.aig'%(initial_f_name,i)) ## ps() ## #J = [18,0,1,2,3,7,14] ## J = slps+sims+intrps+bmcs+pdrs ## t = last_verify_time ## status = verify(J,t) ## if status >= Unsat: ## return status ## if i > 0: ## print 'SAT returned, Running less abstract backup' ## continue ## break ## if ((i == 0) and (status > Unsat) and (n_ands() > 0)): ## print '\n***Running speculate on initial backup number %d:'%i, ## abc('r %s_backup_%d.aig'%(initial_f_name,i)) ## ps() ## if n_ands() < 20000: #### pre_simp() ## status = speculate() ## if ((status <= Unsat) or (status == Error)): ## return status ## #J = [18,0,1,2,3,7,14] ## J = slps+sims+intrps+bmcs+pdrs ## t = 2*last_verify_time ## print 'Verify time increased to %d'%last_verify_time ## status = verify(J,t) ## if status == Unsat: ## return status ## else: ## return Undecided_reduction def smp(): abc('smp') write_file('smp') def dprove(): abc('dprove -cbjupr') def trim(): global trim_allowed if not trim_allowed: return ## abc('trm;addpi') reparam() ## print 'exiting trim' def prs(x=True): global trim_allowed """ If x is set to False, no reparameterization will be done in pre_simp""" trim_allowed = x print 'trim_allowed = ',trim_allowed y = time.clock() pre_simp() print 'Time = %s'%convert(time.clock() - y) write_file('smp') def pre_simp(): """This uses a set of simplification algorithms which preprocesses a design. Includes forward retiming, quick simp, signal correspondence with constraints, trimming away PIs, and strong simplify""" global trim_allowed, temp_dec tt = time.time() set_globals() abc('&get; &scl; &put') if (n_ands() > 200000 or n_latches() > 50000 or n_pis() > 40000): print 'Problem too large, simplification skipped' return 'Undecided' if ((n_ands() > 0) or (n_latches()>0)): trim() ## ps() if n_latches() == 0: return check_sat() ## if n_ands()<70000: ## abs('scorr -C 5000') best_fwrd_min([10,11]) ps() status = try_scorr_constr() if ((n_ands() > 0) or (n_latches()>0)): trim() if n_latches() == 0: return check_sat() status = process_status(status) if status <= Unsat: return status simplify() print 'Simplify: ', ps() if n_latches() == 0: return check_sat() if trim_allowed: t = min(15,.3*G_T) if not '_smp' in f_name: #try this only once on a design try_temps(15) if n_latches() == 0: return check_sat() try_phase() if n_latches() == 0: return check_sat() if ((n_ands() > 0) or (n_latches()>0)): trim() status = process_status(status) print 'Simplification time = %0.2f'%(time.time()-tt) return status def pre_simp2(): """This uses a set of simplification algorithms which preprocesses a design. Includes forward retiming, quick simp, signal correspondence with constraints, trimming away PIs, and strong simplify""" global trim_allowed, temp_dec tt = time.time() set_globals() abc('&get; &scl; &put') if (n_ands() > 200000 or n_latches() > 50000 or n_pis() > 40000): print 'Problem too large, simplification skipped' return 'Undecided' if ((n_ands() > 0) or (n_latches()>0)): trim() ## ps() if n_latches() == 0: return check_sat() if n_ands()<70000: abc('scorr -C 5000') best_fwrd_min([10,11]) ps() status = try_scorr_constr() if ((n_ands() > 0) or (n_latches()>0)): trim() if n_latches() == 0: return check_sat() status = process_status(status) if status <= Unsat: return status simplify() print 'Simplify: ', ps() if n_latches() == 0: return check_sat() if trim_allowed: t = min(15,.3*G_T) if not '_smp' in f_name: #try this only once on a design try_temps(15) if n_latches() == 0: return check_sat() try_phase() if n_latches() == 0: return check_sat() if ((n_ands() > 0) or (n_latches()>0)): trim() status = process_status(status) print 'Simplification time = %0.2f'%(time.time()-tt) return status def try_scorr_constr(): set_size() abc('w %s_savetemp.aig'%f_name) status = scorr_constr() if inferior_size(): abc('r %s_savetemp.aig'%f_name) return status def factors(n): l = [1,] nn = n while n > 1: for i in (2,3,5,7,11,13,17,19,23,29,31,37,41,43,47,53): if not i 45000) or init_simp == 0): if ((n == 1) or (n_ands() > 45000)): return ## init_simp = 0 print 'Trying phase abstraction - Max phase = %d'%n, abc('w %s_phase_temp.aig'%f_name) na = n_ands() nl = n_latches() ni = n_pis() no = n_pos() z = ok_phases(n) print z, if len(z) == 1: return #p = choose_phase() p = z[1] abc('phase -F %d'%p) if no == n_pos(): #nothing happened because p is not mod period print 'Phase %d is incompatible'%p abc('r %s_phase_temp.aig'%f_name) if len(z)< 3: return else: p = z[2] #print 'Trying phase = %d: '%p, abc('phase -F %d'%p) if no == n_pos(): #nothing happened because p is not mod period print 'Phase %d is incompatible'%p abc('r %s_phase_temp.aig'%f_name) return print 'Simplifying with %d phases: => '%p, simplify() trim() ps() cost = rel_cost([ni,nl,na]) print 'New relative cost = %f'%(cost) if cost < -.01: abc('w %s_phase_temp.aig'%f_name) if ((n_latches() == 0) or (n_ands() == 0)): return if n_phases() == 1: #this bombs out if no latches return else: try_phase() return elif len(z)>2: #Try the next eligible phase. abc('r %s_phase_temp.aig'%f_name) if p == z[2]: #already tried this return p = z[2] print 'Trying phase = %d: => '%p, abc('phase -F %d'%p) if no == n_pos(): #nothing happened because p is not mod period print 'Phase = %d is not compatible'%p return ps() print 'Simplify with %d phases: '%p, simplify() trim() ps() cost = rel_cost([ni,nl,na]) print 'New relative cost = %f'%(cost) if cost < -.01: print 'Phase abstraction with %d phases obtained:'%p, print_circuit_stats() abc('w %s_phase_temp.aig'%f_name) if ((n_latches() == 0) or (n_ands() == 0)): return if n_phases() == 1: # this bombs out if no latches return else: try_phase() return abc('r %s_phase_temp.aig'%f_name) return def try_temp(t=15): btime = time.clock() trim() print'Trying temporal decomposition - for max %s sec. '%convert(t), abc('w %s_best.aig'%f_name) ni = n_pis() nl = n_latches() na = n_ands() best = [ni,nl,na] F = create_funcs([18],t) #create a timer function ## F = F + [eval('(pyabc_split.defer(abc)("tempor -s; trm; scr; trm; tempor; trm; scr; trm"))')] F = F + [eval('(pyabc_split.defer(abc)("tempor -s; &get; &trim -o; &put; scr; &get; &trim -o; &put; tempor; &get; &trim -o; &put; scr; &get; &trim -o; &put"))')] for i,res in pyabc_split.abc_split_all(F): break cost = rel_cost(best) print 'cost = %0.2f'%cost if cost < .01: ps() return else: abc('r %s_best.aig'%f_name) def try_temps(t=15): best = (n_pis(),n_latches(),n_ands()) while True: try_temp(t) if ((best == (n_pis(),n_latches(),n_ands())) or n_ands() > .9 * best[2] ): break elif n_latches() == 0: break else: best = (n_pis(),n_latches(),n_ands()) def rel_cost_t(J): """ weighted relative costs versus previous stats.""" if (n_latches() == 0 and J[1]>0): return -10 nli = J[0]+J[1] na = J[2] if ((nli == 0) or (na == 0)): return 100 nri = n_real_inputs() #ri = (float(nri)-float(ni))/float(ni) rli = (float(n_latches()+nri)-float(nli))/float(nli) ra = (float(n_ands())-float(na))/float(na) cost = 10*rli + .5*ra return cost def rel_cost(J): """ weighted relative costs versus previous stats.""" global f_name if (n_latches() == 0 and J[1]>0): return -10 nri = n_real_inputs() ni = J[0] nl = J[1] na = J[2] if (ni == 0 or na == 0 or nl == 0): return 100 ri = (float(nri)-float(ni))/float(ni) rl = (float(n_latches())-float(nl))/float(nl) ra = (float(n_ands())-float(na))/float(na) cost = 1*ri + 5*rl + .2*ra ## print 'Relative cost = %0.2f'%cost return cost def best_fwrd_min(J): global f_name, methods mtds = sublist(methods,J) F = create_funcs(J,0) (m,result) = fork_best(F,mtds) #FORK here print '%s: '%mtds[m], def try_forward(): """Attempts most forward retiming, and latch correspondence there. If attempt fails to help simplify, then we revert back to the original design This can be effective for equivalence checking problems where synthesis used retiming""" abc('w %s_savetemp.aig'%f_name) if n_ands() < 30000: abc('dr') abc('lcorr') nl = n_latches() na = n_ands() abc('w %s_savetemp0.aig'%f_name) abc('r %s_savetemp.aig'%f_name) abc('dr -m') abc('lcorr') abc('dr') if ((n_latches() <= nl) and (n_ands() < na)): print 'Forward retiming reduced size to: ', print_circuit_stats() return else: abc('r %s_savetemp0.aig'%f_name) return return def quick_simp(): """A few quick ways to simplify a problem before more expensive methods are applied. Uses & commands if problem is large. These commands use the new circuit based SAT solver""" na = n_ands() if na < 60000: abc('scl -m;lcorr;drw') else: abc('&get;&scl;&lcorr;&put;drw') print 'Using quick simplification', status = process_status(get_status()) if status <= Unsat: result = RESULT[status] else: ps() ## write_file('smp') #### K_backup = K = K+1 result = 'UNDECIDED' return result def scorr_constr(): """Extracts implicit constraints and uses them in signal correspondence Constraints that are found are folded back when done""" na = max(1,n_ands()) n_pos_before = n_pos() if ((na > 40000) or n_pos()>1): return Undecided_no_reduction abc('w %s_savetemp.aig'%f_name) na = max(1,n_ands()) ## f = 1 f = 18000/na #**** THIS can create a bug 10/15/11 f = min(f,4) f = max(1,f) if n_ands() > 18000: cmd = 'unfold -s -F 2' else: cmd = 'unfold -F %d -C 5000'%f abc(cmd) if n_pos() == n_pos_before: print 'No constraints found' return Undecided_no_reduction if (n_ands() > na): #no constraints found abc('r %s_savetemp.aig'%f_name) return Undecided_no_reduction na = max(1,n_ands()) f = 1 print 'Number of constraints = %d'%((n_pos() - n_pos_before)) abc('scorr -c -F %d'%f) abc('fold') trim() print 'Constrained simplification: ', ps() return Undecided_no_reduction def try_scorr_c(f): """ Trying multiple frames because current version has a bug.""" set_globals() abc('unfold -F %d'%f) abc('scorr -c -F %d'%f) abc('fold') t = max(1,.1*G_T) abc('&get;,bmc3 -vt=%f'%t) if is_sat(): return 0 else: trim() return 1 def input_x_factor(): """Sets the global x_factor according to user input""" global x_factor, xfi print 'Type in x_factor:', xfi = x_factor = input() print 'x_factor set to %f'%x_factor def prove(a): """Proves all the outputs together. If ever an abstraction was done then if SAT is returned, we make RESULT return "undecided". is a == 0 do smp and abs first If a == 1 do smp and spec first if a == 2 do quick simplification instead of full simplification, then abs first, spec second""" global x_factor,xfi,f_name, last_verify_time,K_backup, t_init, sec_options, spec_found_cex spec_first = False max_bmc = -1 abs_found_cex_after_spec = spec_found_cex_after_abs = False if not '_smp' in f_name: #if already simplified, then don't do again if a == 2 : #do quick simplification result = quick_simp() #does not write 'smp' file ## print result else : result = prove_part_1() #do full simplification here if ((result == 'SAT') or (result == 'UNSAT')): return result if a == 1: spec_first = True t_init = 2 abs_found_cex_before_spec = spec_found_cex_before_abs = False ## First phase if spec_first: result = prove_part_3() #speculation done here first else: result = prove_part_2() #abstraction done here first if ((result == 'SAT') or (result == 'UNSAT')): return result ## Second phase if spec_first: #did spec already in first phase t_init = 2 result = prove_part_2() #abstraction done here second if result == 'SAT': abs_found_cex_after_spec = True else: result = prove_part_3() #speculation done here second if result == 'SAT': spec_found_cex_after_abs = True if result == 'UNSAT': return result status = get_status() if result == 'ERROR': status = Error if ('_abs' in f_name and spec_found_cex_after_abs): #spec file should not have been written in speculate f_name = revert(f_name,1) #it should be as if we never did abstraction. print 'f_name = %s'%f_name abc('r %s.aig'%f_name) #restore previous t_init = 2 if not '_rev' in f_name: print 'proving speculation first' write_file('rev') #maybe can get by with just changing f_name print 'f_name = %s'%f_name result = prove(1) #1 here means do smp and then spec if ((result == 'SAT') or (result == 'UNSAT')): return result elif ('_spec' in f_name and abs_found_cex_after_spec): #abs file should not have been written in abstract f_name = revert(f_name,1) #it should be as if we never did speculation. abc('r %s.aig'%f_name) #restore previous t_init = 2 if not '_rev' in f_name: print 'proving abstraction first' write_file('rev') #maybe can get by with just changing f_name result = prove(0) if ((result == 'SAT') or (result == 'UNSAT')): return result else: return 'UNDECIDED' def prove_part_1(): global x_factor,xfi,f_name, last_verify_time,K_backup print 'Initial: ', ps() x_factor = xfi print '\n***Running pre_simp' set_globals() if n_latches() > 0: status = run_par_simplify() else: status = check_sat() if ((status <= Unsat) or (n_latches() == 0)): return RESULT[status] trim() write_file('smp') #need to check that this was not written in pre_simp set_globals() return RESULT[status] def run_par_simplify(): set_globals() t = 1000 funcs = [eval('(pyabc_split.defer(pre_simp)())')] J = pdrs+bmcs+intrps J = modify_methods(J,1) funcs = create_funcs(J,t)+ funcs #important that pre_simp goes last mtds =sublist(methods,J) + ['pre_simp'] fork_last(funcs,mtds) status = get_status() return status def prove_part_2(ratio=.75): """does the abstraction part of prove""" global x_factor,xfi,f_name, last_verify_time,K_backup, trim_allowed,ifbip print'\n***Running abstract' ## print 'ifbip = %d'%ifbip status = abstract(ifbip) #ABSTRACTION done here status = process_status(status) print 'abstract done, status is %d'%status result = RESULT[status] if status < Unsat: print 'CEX in frame %d'%cex_frame() return result #if we found a cex we do not want to trim. trim() return result def prove_part_3(): """does the speculation part of prove""" global x_factor,xfi,f_name, last_verify_time,init_initial_f_name global max_bmc, sec_options if ((n_ands() > 36000) and sec_options == ''): sec_options = 'g' print 'sec_options set to "g"' print '\n***Running speculate' status = speculate() #SPECULATION done here status = process_status(status) print 'speculate done, status is %d'%status result = RESULT[status] if status < Unsat: print 'CEX in frame %d'%cex_frame() return result trim() #if cex is found we doo not want to trim. return result def prove_all(dir,t): """Prove all aig files in this directory using super_prove and record the results in results.txt Not called from any subroutine """ ## t = 1000 #This is the timeoout value xtime = time.time() ## dir = main.list_aig('') results = [] f =open('results_%d.txt'%len(dir), 'w') for name in dir: read_file_quiet(name) print '\n **** %s:'%name, ps() F = create_funcs([18,6],t) #create timer function as i = 0 Here is the timer for i,res in pyabc_split.abc_split_all(F): break tt = time.time() if i == 0: res = 'Timeout' str = '%s: %s, time = %s'%(name,res,convert(tt-xtime)) if res == 'SAT': str = str + ', cex_frame = %d'%cex_frame() str = str +'\n' f.write(str) f.flush() results = results + ['%s: %s, time = %s'%(name,res,convert(tt-xtime))] xtime = tt ## print results f.close() return results def remove_pos(lst): """Takes a list of pairs where the first part of a pair is the PO number and the second is the result 1 = disproved, 2 = proved, 3 = unresolved. Then removes the proved and disproved outputs and returns the aig with the unresolved outputs""" proved = disproved = unresolved = [] for j in range(len(lst)): jj = lst[j] if jj[1] == 2: proved = proved + [jj[0]] if (jj[1] == 1 or (jj[1] == 0)): disproved = disproved +[jj[0]] if jj[1] > 2: unresolved = unresolved +[jj[0]] print '%d outputs proved'%len(proved) if not proved == []: if ((max(proved)>n_pos()-1) or min(proved)< 0): print proved remove(proved) #functions for proving multiple outputs in parallel #__________________________________________________ def prove_only(j): """ extract the jth output and try to prove it""" global max_bmc, init_initial_f_name, initial_f_name, f_name,x #abc('w %s__xsavetemp.aig'%f_name) extract(j,j) set_globals() ps() print '\nProving output %d'%(j) f_name = f_name + '_%d'%j result = prove_1() #abc('r %s__xsavetemp.aig'%f_name) if result == 'UNSAT': print '******** PROVED OUTPUT %d ******** '%(j) return Unsat if result == 'SAT': print '******** DISPROVED OUTPUT %d ******** '%(j) return Sat else: print '******** UNDECIDED on OUTPUT %d ******** '%(j) return Undecided def verify_only(j,t): """ extract the jth output and try to prove it""" global max_bmc, init_initial_f_name, initial_f_name, f_name,x, reachs, last_cex, last_winner, methods ## ps() ## print 'Output = %d'%j extract(j,j) ## ps() set_globals() if n_latches() == 0: result = check_sat() else: f_name = f_name + '_%d'%j # make it so that jabc is not used here reachs_old = reachs reachs = reachs[1:] #just remove jabc from this. res = verify(slps+sims+pdrs+bmcs+intrps,t) #keep the number running at the same time as small as possible. ## res = verify(sims+pdrs+bmcs+intrps,t) #keep the number running at the same time as small as possible. reachs = reachs_old result = get_status() assert res == result,'result: %d, status: %d'%(res,get_status()) if result > Unsat: ## print result ## print '******* %d is undecided ***********'%j return result elif result == Unsat: ## print '******** PROVED OUTPUT %d ******** '%(j) return result elif ((result < Unsat) and (not result == None)): print '******** %s DISPROVED OUTPUT %d ******** '%(last_cex,j) ## print ('writing %d.status'%j), result, get_status() abc('write_status %d.status'%j) last_winner = last_cex return result else: print '****** %d result is %d'%(j,result) return result def verify_range(j,k,t): """ extract the jth thru kth output and try to prove their OR""" global max_bmc, init_initial_f_name, initial_f_name, f_name,x, reachs, last_cex, last_winner, methods extract(j,k) abc('orpos') set_globals() if n_latches() == 0: result = check_sat() else: f_name = f_name + '_%d'%j # make it so that jabc is not used here reachs_old = reachs reachs = reachs[1:] #just remove jabc from this. res = verify(sims+pdrs+bmcs+intrps,t) #keep the number running at the sme time as small as possible. reachs = reachs_old result = get_status() assert res == result,'result: %d, status: %d'%(res,get_status()) if result > Unsat: ## print result ## print '******* %d is undecided ***********'%j return result elif result == Unsat: ## print '******** PROVED OUTPUT %d ******** '%(j) return result elif ((result < Unsat) and (not result == None)): print '******** %s DISPROVED OUTPUT %d ******** '%(last_cex,j) ## print ('writing %d.status'%j), result, get_status() abc('write_status %d.status'%j) last_winner = last_cex return result else: print '****** %d result is %d'%(j,result) return result def prove_n_par(n,j): """prove n outputs in parallel starting at j""" F = [] for i in range(n): F = F + [eval('(pyabc_split.defer(prove_only)(%s))'%(j+i))] #print S #F = eval(S) result = [] print 'Proving outputs %d thru %d in parallel'%(j,j+n-1) for i,res in pyabc_split.abc_split_all(F): result = result +[(j+i,res)] #print result return result def prove_pos_par(t,BREAK): """Prove all outputs in parallel and break on BREAK""" return run_parallel([],t,BREAK) def prove_pos_par0(n): """ Group n POs grouped and prove in parallel until all outputs have been proved""" f_name = initial_f_name abc('w %s__xsavetemp.aig'%f_name) result = [] j = 0 N = n_pos() while j < N-n: abc('r %s__xsavetemp.aig'%f_name) result = result + prove_n_par(n,j) j = j+n if N > j: result = result + prove_n_par(N-j,j) abc('r %s__xsavetemp.aig'%initial_f_name) ps() ## print result remove_pos(result) write_file('group') return def prop_decomp(): """decompose a single property into multiple ones (only for initial single output), by finding single and double literal primes of the outputs.""" if n_pos()>1: return run_command('outdec -v -L 2') if n_pos()>1: ps() def distribute(N,div): """ we are going to verify outputs in groups """ n = N/div rem = N - (div * (N/div)) result = [] for j in range(div): if rem >0: result = result +[n+1] rem = rem -1 else: result = result + [n] return result ####def find_cex_par(tt): #### """prove n outputs at once and quit at first cex. Otherwise if no cex found return aig #### with the unproved outputs""" #### global trim_allowed,last_winner, last_cex, n_pos_before, t_init, j_last, sweep_time #### b_time = time.time() #Wall clock time #### n = n_pos() #### l=remove_const_pos() #### N = n_pos() #### full_time = all_proc = False #### print 'Number of POs: %d => %d'%(n,N) #### if N == 0: #### return Unsat ###### inc = 5 #******* increment for grouping for sweep set here ************* ###### inc = min(12,max(inc, int(.1*N))) #### inc = 1+N/100 ###### if N <1.5*inc: # if near the increment for grouping try to get it below. ###### prove_all_ind() ###### N = n_pos() #### if inc == 1: #### prove_all_ind() #### N = n_pos() #### T = int(tt) #this is the total time to be taken in final verification run before quitting speculation ###### if inc == 10: ###### t_init = 10 ###### t = max(t_init/2,T/20) ###### if N <= inc: ###### t = T ###### print "inc = %d, Sweep time = %s, j_group = %d"%(inc,convert(t),j_last) #### t = sweep_time/2 #start sweeping at last time where cex was found. ###### it used to be t = 1 here but it did not make sense although seemed to work. ###### inc = 2 #### while True: #poor man's concurrency #### N = n_pos() #### if N == 0: #### return Unsat #### #sweep_time controls so that when sweep starts after a cex, it starts at the last sweep time #### t = max(2,2*t) #double sweep time #### if t > .75*T: #### t = T #### full_time = True #### if ((N <= inc) or (N < 13)): #### t = sweep_time = T #### full_time = True #### inc = 1 ###### sweep_time = 2*sweep_time #### if not t == T: #### t= sweep_time = max(t,sweep_time) ###### t = sweep_time ######new heuristic #### if (all_proc and sweep_time > 8): #stop poor man's concurrency and jump to full time. #### t = sweep_time = T #### full_time - True #this might be used to stop speculation when t = T and the last sweep ###### found no cex and we do not prove Unsat on an output #### abc('w %s__ysavetemp.aig'%f_name) #### ps() #### if N < 50: #### inc = 1 #### print "inc = %d, Sweep time = %s, j_last = %d"%(inc,convert(t),j_last) #### F = [] ###### G = [] #### #make new lambda functions since after the last pass some of the functions may have been proved and eliminated. #### for i in range(N): #### F = F + [eval('(pyabc_split.defer(verify_only)(%d,%s))'%(i,convert(T)))] #make time large and let sleep timer control timeouts ###### G = G + [range(i,i+1)] #### ###### #### result = [] #### outcome = '' #### N = len(F) #### rng = range(1+(N-1)/inc) #### rng = rng[j_last:]+rng[:j_last] #pick up in range where last found cex. ###### print 'rng = ', ###### print rng #### k = -1 #### bb_time = time.time() #### for j in rng: #### k = k+1 #keeps track of how many groups we have processed. #### j_last = j #### J = j*inc #### JJ = J+inc #### JJ = min(N,JJ) #### if J == JJ-1: #### print 'Function = %d '%J, #### else: #### print 'Functions = [%d,%d]'%(J,JJ-1) #### Fj = create_funcs([18],t+1) #create timer function as i = 0 Here is the timer #### Fj = Fj + F[J:JJ] #### count = 0 #### fj_time = time.time() #### abc('r %s__ysavetemp.aig'%f_name) #important need to restore aig here so the F refers to right thing when doing verify_only. ###### # because verify_only changes the aig. ###### ps() #### for i,res in pyabc_split.abc_split_all(Fj): #### count = count+1 #### Ji = J+i-1 #gives output number #### if ((res == 0) or (res == 1)): #### abc('read_status %d.status'%Ji) #### res = get_status() #### outcome = 'CEX: Frame = %d, PO = %d, Time = %s'%(cex_frame(),Ji,convert((time.time() - fj_time))) #### break #### if i == 0: #sleep timer expired #### outcome = '*** Time expired in %s sec. Next group = %d to %d ***'%(convert(time.time() - fj_time),JJ,min(N,JJ+inc)) #### break #### elif res == None: #this should not happen #### print res #### print Ji,RESULT[res], #### else: # output Ji was proved #### result = result + [[Ji,res]] #### if count >= inc: #### outcome = '--- all group processed without cex ---' #### all_proc = True #### break #### continue #this can only happen if inc > 1 #### # end of for i loop #### if ((res < Unsat) and (not res == None)): #### break #### else: #### continue # continue j loop #### #end of for j loop #### if k < len(rng): #### t_init = t/2 #next time start with this time. #### else: #### j_last = j_last+1 #this was last j and we did not find cex, so start at next group #### print outcome + ' => ' , #### if ((res < Unsat) and (not res == None)): #### t_init = t/2 #### abc('read_status %d.status'%Ji) #make sure we got the right status file. #### #actually if doing abstraction we could remove proved outputs now, but we do not. -**inefficiency** #### return res #### else: #This implies that no outputs were disproved. Thus can remove proved outputs. #### abc('r %s__ysavetemp.aig'%f_name) #restore original aig #### if not result == []: #### res = [] #### for j in range(len(result)): #### k = result[j] #### if k[1] == 2: #### res = res + [k[0]] ###### print res ###### result = mapp(res,G) #### result = res ###### print result #### remove(result) #remove the outputs that were proved UNSAT. #### #This is OK for both abstract and speculate #### print 'Number of POs reduced to %d'%n_pos() #### if n_pos() == 0: #### return Unsat #### if t>=T: #### return Undecided #### else: #### continue #### return Undecided ####def remap_pos(): #### """ maintains a map of current outputs to original outputs""" #### global po_map #### k = j = 0 #### new = [] #### assert n_pos() == len(po_map), 'length of po_map, %d, and current # POs, %d, don"t agree'%(len(po_map),n_pos()) #### for j in range(len(po_map)): #### N = n_pos() #### abc('removepo -N %d'%k) # this removes the output if it is 0 driven #### if n_pos() == N: #### new = new + [po_map[j]] #### k = k+1 #### if len(new) < len(po_map): ###### print 'New map = ', ###### print new #### po_map = new ####def prove_mapped(): #### """ #### assumes that srm is in workspace and takes the unsolved outputs and proves #### them by using proved outputs as constraints. #### """ #### global po_map ###### print po_map #### po_map.sort() #make sure mapped outputs are in order #### for j in po_map: #put unsolved outputs first #### run_command('swappos -N %d'%j) #### print j #### N = n_pos() #### assert N > len(po_map), 'n_pos = %d, len(po_map) = %d'%(N, len(po_map)) #### run_command('constr -N %d'%(N-len(po_map))) #make the other outputs constraints #### run_command('fold') #fold constraints into remaining outputs. #### ps() #### prove_all_mtds(100) ####def mapp(R,G): #### result = [] #### for j in range(len(R)): #### result = result + G[R[j]] #### return result #_______________________________________ ####def prove_g_pos_split(): #### """like prove_g_pos but quits when any output is undecided""" #### global f_name, max_bmc,x_factor,x #### x = time.clock() #### #input_x_factor() #### init_f_name = f_name #### print 'Beginning prove_g_pos_split' #### prove_all_ind() #### print 'Number of outputs reduced to %d by fast induction with constraints'%n_pos() #### reparam() ###### try_rpm() #### print '********** Proving each output separately ************' #### f_name = init_f_name #### abc('w %s_osavetemp.aig'%f_name) #### n = n_pos() #### print 'Number of outputs = %d'%n #### pos_proved = [] #### J = 0 #### jnext = n-1 #### while jnext >= 0: #### max_bmc = -1 #### f_name = init_f_name #### abc('r %s_osavetemp.aig'%f_name) #### jnext_old = jnext #### extract(jnext,jnext) #### jnext = jnext -1 #### print '\nProving output %d'%(jnext_old) #### f_name = f_name + '_%d'%jnext_old #### result = prove_1() #### if result == 'UNSAT': #### if jnext_old > jnext+1: #### print '******** PROVED OUTPUTS [%d-%d] ******** '%(jnext+1,jnext_old) #### else: #### print '******** PROVED OUTPUT %d ******** '%(jnext_old) #### pos_proved = pos_proved + range(jnext +1,jnext_old+1) #### continue #### if result == 'SAT': #### print 'One of output in (%d to %d) is SAT'%(jnext + 1,jnext_old) #### return result #### else: #### print '******** UNDECIDED on OUTPUTS %d thru %d ******** '%(jnext+1,jnext_old) #### print 'Eliminating %d proved outputs'%(len(pos_proved)) #### # remove outputs proved and return #### f_name = init_f_name #### abc('r %s_osavetemp.aig'%f_name) #### remove(pos_proved) #### trim() #### write_file('group') #### return 'UNDECIDED' #### f_name = init_f_name #### abc('r %s_osavetemp.aig'%f_name) #### if not len(pos_proved) == n: #### print 'Eliminating %d proved outputs'%(len(pos_proved)) #### remove(pos_proved) #### trim() #### write_file('group') #### result = 'UNDECIDED' #### else: #### print 'Proved all outputs. The problem is proved UNSAT' #### result = 'UNSAT' #### print 'Total time = %f sec.'%(time.clock() - x) #### return result ####def group(a,n): #### """Groups together outputs beginning at output n and any contiguous preceeding output #### that does not increase the latch support by a or more""" #### global f_name, max_bmc #### nlt = n_latches() #### extract(n,n) #### nli = n_latches() #### if n == 0: #### return n-1 #### for J in range(1,n+1): #### abc('r %s_osavetemp.aig'%f_name) #### j = n-J #### #print 'Running %d to %d'%(j,n) #### extract(j,n) #### #print 'n_latches = %d'%n_latches() #### #if n_latches() >= nli + (nlt - nli)/2: #### if n_latches() == nli: #### continue #### if n_latches() > nli+a: #### break #### abc('r %s_osavetemp.aig'%f_name) ###### if j == 1: ###### j = j-1 #### print 'extracting [%d-%d]'%(j,n) #### extract(j,n) #### ps() #### return j-1 def extract(n1,n2): """Extracts outputs n1 through n2""" no = n_pos() if n2 > no: return 'range exceeds number of POs' abc('cone -s -O %d -R %d'%(n1, 1+n2-n1)) def remove_intrps(J): global n_proc,ifbip npr = n_proc if 18 in J: npr = npr+1 if len(J) <= npr: return J JJ = [] alli = [23,1,22] # if_no_bip, then this might need to be changed l = len(J)-npr alli = alli[:l] for i in J: if i in alli: continue else: JJ = JJ +[i] return JJ def remove(lst): """Removes outputs in list""" global po_map n_before = n_pos() zero(lst) l=remove_const_pos() print 'n_before = %d, n_list = %d, n_after = %d'%(n_before, len(lst), n_pos()) def check_pos(): """ checks if any POs are 0, and removes them with a warning""" N = n_pos() l=remove_const_pos() if not N == n_pos(): print 'WARNING: some POs are 0 and are rremoved' print '%d POs removed'%(N - n_pos()) def zero(list): """Zeros out POs in list""" for j in list: run_command('zeropo -N %d'%j) def mark_const_pos(ll=[]): """ creates an indicator of which PO are const-0 and which are const-1 does not change number of POs """ L = range(n_pos()) L.reverse() if ll == []: ll = [-1]*n_pos() ind = ll abc('&get') #save original for j in L: n_before = n_pos() abc('removepo -N %d'%j) #removes const-0 output if n_pos() < n_before: ind[j]=0 ## print sumsize(ind) ## ps() abc('&put') for j in L: n_before = n_pos() abc('removepo -z -N %d'%j) # -z removes const-1 output if n_pos() < n_before: ind[j]=1 ## ps() abc('&put') #put back original ## remove_const_pos() print sumsize(ind) return ind def remove_const_pos(): global po_map """removes the 0 pos, but no pis because we might get cexs and need the correct number of pis Should keep tract of if original POs are 0 and are removed. Can this happen outside of prove_all_ind or pord_all which can set proved outputs to 0??? WARNING: This can remove constant 1 nodes too??? """ run_command('&get; &trim -i; &put; addpi') #adds a pi only if there are none po_map = range(n_pos()) def psp(): quick_simp() result = run_parallel([6,21],500,'US') #runs 'run_parallel' and sp() in parallel. ## run_parallel uses JP and TERM to terminate. return result def sp(): global initial_f_name """Alias for super_prove""" print 'Executing super_prove' result = super_prove(0) print '%s is done'%initial_f_name return result def sumsize(L): d = count_less(L,0) u = count_less(L,1)-d s = count_less(L,2) - (d+u) return 'SAT = %d, UNSAT = %d, UNDECIDED = %d'%(s,u,d) def unmap(L,L2,map): mx = max(list(map)) assert mx <= len(L2),'max of map = %d, length of L2 = %d'%(mx,len(L)) for j in range(len(map)): L[j] = L2[map[j]] #expand results of L2 into L return L def create_map(L,N): map = [-1]*N for j in range(len(L)): lj = L[j] for k in range(len(lj)): map[lj[k]] = j return map def mp(op='sp'): L = multi_prove(op,900) return sumsize(L) def multi_prove(op='sp',tt=900): """two phase prove process for multiple output functions""" global max_bmc, init_initial_f_name, initial_f_name,win_list, last_verify_time global f_name_save,name_save x_init = time.time() N = n_pos() L = [-1]*N print 'Mapping for isomorphism: ' iso() #reduces number of POs map = create_map(iso_eq_classes(),N) #creates map into original N = n_pos() r = pre_simp() #pre_simp write_file('smp') NP = n_pos()/N #if NP > 1 then NP unrollings were done. if n_pos() > N: assert NP>=2, 'NP not 2, n_pos = %d, N = %d, NP = %d'%(n_pos(),N,NP) print 'pre_simp done. NP = %d\n\n'%NP #WARNING: if phase abstraction done, then number of POs changed. if r == Unsat: L = [0]*N #all outputs are UNSAT print sumsize(L) return L f_name_save = f_name name_save = '%s_initial_save.aig'%f_name abc('w %s'%name_save) L1 = [-1]*n_pos() # L1 has length after pre_simp ## L1= mark_const_pos(L1) #puts const values in L1 ## print sumsize(L1) #########do second iso here N = n_pos() print 'Mapping for isomorphism: ' iso() #second iso - changes number of POs map2 = create_map(iso_eq_classes(),N) #creates map into previous L2 = [-1]*n_pos() L2 = mark_const_pos(L2) #populates L2 with constants print sumsize(L2) #########second iso done abc('r %s'%name_save) L2 = mprove(L2,op,10) #populates L2 with results print sumsize(L2) time_left = tt - (time.time()-x_init) print '\n\n\n*********time left = %d ********\n\n\n'%time_left N = count_less(L2,0) if N > 0: t = max(100,time_left/N) L2 = mprove(L2,op,t) #populates L2 with more results S = sumsize(L2) T = '%.2f'%(time.time() - x_init) print '%s in time = %s'%(S,T) ########undo second iso L1 = unmap(L1,L2,map2) print 'unmapping for iso: ', print sumsize(L1) ############# if NP > 1: #an unrolling was done L1 = check_L(NP,L1) #map into reduced size before unrolling was done. print 'unmapping for unrolling.', print sumsize(L1) L = unmap(L,L1,map) print 'unmapping for iso', print sumsize(L) return L def check_L(NP,L): """This happens when an unrolling creates additional POs We want to check that L[j] = L[j+N] etc to make sure the PO results agree in all phases, i.e. sat, unsat, or undecided """ N = len(L)/NP #original number of POs for j in range(N): for k in range(NP)[1:]: #k = 1,2,...,NP-1 if (L[j] == 1): break elif L[j+k*N] == 1: L[j] = 1 break elif L[j] == -1: continue #we have to continue to look for a 1 elif L[j] == 0: if L[j+k*N] == -1: L[j] = -1 break continue #have to make sure that all phases are 0 return L[:N] def mprove(L,op='sp',tt=1000): """ 0 = unsat, 1 = sat, -1 = undecided""" global max_bmc, init_initial_f_name, initial_f_name,win_list, last_verify_time global f_name_save,name_save,temp_dec N = len(L) t = tt #controls the amount of time spent on each cone funcs = [eval('(pyabc_split.defer(%s)())'%op)] funcs = create_funcs(slps,t)+funcs mtds =sublist(methods,slps) + [op] res = L for j in range(N): if L[j] > -1: continue #already solved abc('r %s'%name_save) #restore original function x = time.time() name = '%s_cone_%d'%(f_name_save,j) abc('cone -s -O %d'%j) abc('w %s.aig'%name) print '\n********************************************' read_file_quiet(name) print '________%s(%s)__________'%(op,name) temp_dec = False fork_last(funcs,mtds) T = '%.2f'%(time.time() - x) out = get_status() if out == Unsat: res[j] = 0 if out < Unsat: res[j] = 1 print '\n%s: %s in time = %s'%(name,RESULT[out],T) return res def sp1(options = ''): global sec_options sec_options = options return super_prove(1) def super_prove(n=0): """Main proof technique now. Does original prove and if after speculation there are multiple output left if will try to prove each output separately, in reverse order. It will quit at the first output that fails to be proved, or any output that is proved SAT n controls call to prove(n) is n == 0 do smp and abs first, then spec if n == 1 do smp and spec first then abs if n == 2 just do quick simplification instead of full simplification, then abs first, spec second """ global max_bmc, init_initial_f_name, initial_f_name,win_list, last_verify_time, f_name init_initial_f_name = initial_f_name if x_factor > 1: print 'x_factor = %f'%x_factor input_x_factor() max_bmc = -1 x = time.time() ## if n == 2: ## result = prove(2) ## else: ## result = prove(0) result = prove(n) if ((result == 'SAT') or (result == 'UNSAT')): print '%s: total clock time taken by super_prove = %f sec.'%(result,(time.time() - x)) return result elif ((result == 'UNDECIDED') and (n_latches() == 0)): return result print '%s: total clock time taken by super_prove so far = %f sec.'%(result,(time.time() - x)) y = time.time() if n == 2: print 'Entering BMC_VER()' result = BMC_VER() #n=2 is only called from sp2, a super_prove run in parallel. if ((result == 'SAT') and (('_abs' in f_name)or '_spec' in f_name)): #this is because we have done an abstraction and cex is invalid. result = 'UNDECIDED' else: print 'Entering BMC_VER_result' result = BMC_VER_result() #this does backing up if cex is found print 'Total clock time taken by last gasp verification = %f sec.'%(time.time() - y) print 'Total clock time for %s = %f sec.'%(init_initial_f_name,(time.time() - x)) return result def reachm(t): x = time.clock() abc('&get;&reachm -vcs -T %d'%t) print 'reachm done in time = %f'%(time.clock() - x) return get_status() def reachp(t): x = time.clock() abc('&get;&reachp -rv -T %d'%t) print 'reachm2 done in time = %f'%(time.clock() - x) return get_status() def scorr(): run_command('scorr') ps() def select_undecided(L): res = [] for j in range(len(L)): l = L[j] if l[1] == 'UNDECIDED': res = res + [l[0]] return res ####def execute(L,t): #### """ #### run the files in the list L using ss, sp, ssm each for max time = t #### """ #### funcs1 = [eval('(pyabc_split.defer(ss)())')] #### funcs1 = create_funcs(slps,t)+funcs1 #### mtds1 =sublist(methods,slps) + ['ss'] #### funcs2 = [eval('(pyabc_split.defer(sp)())')] #### funcs2 = create_funcs(slps,t)+funcs2 #### mtds2 =sublist(methods,slps) + ['sp'] #### funcs3 = [eval('(pyabc_split.defer(ssm)())')] #### funcs3 = create_funcs(slps,t)+funcs3 #### mtds3 =sublist(methods,slps) + ['ssm'] #### for j in range(len(L)): #### name = L[j] #### print '\n\n\n\n________ss__________' #### read_file_quiet(name) #### print '****ss****' #### fork_last(funcs1,mtds1) #### print '***Done with ss on %s\n'%name #### print '\n\n******ssm************' #### read_file_quiet(name) #### print '****ssm****' #### fork_last(funcs3,mtds3) #### print '***Done with ssm on %s \n'%name def execute_op(op,L,t): """ run the files in the list L using operation "op", each for max time = t """ funcs = [eval('(pyabc_split.defer(%s)())'%op)] funcs = create_funcs(slps,t)+funcs mtds =sublist(methods,slps) + [op] res = [] for j in range(len(L)): x = time.time() name = L[j] print '\n\n\n\n________%s__________'%op read_file_quiet(name) m,result = fork_last(funcs,mtds) if result == Undecided: result = RESULT[result] T = '%.2f'%(time.time() - x) new_res = [name,result,T] res = res + [new_res] print '\n%s'%new_res return res def x_ops(ops,L,t): """ execute each op in the set of ops on each file in the set of files of L, each for time t""" result = [] for j in range(len(ops)): op = ops[j] result.append('Result of %s'%op) result.append(execute_op(op,L,t)) return result def iso(n=0): if n == 0: run_command('&get;&iso;&put') else: run_command('&get;&iso;iso;&put') def check_iso(N): ans = get_large_po() if ans == -1: return 'no output found' n_iso = count_iso(N) return n_iso def count_iso(N): abc('&get;write_aiger -u file1.aig') #put this cone in & space and write file1 ## print 'PO %d is used'%i n_iso = 0 #start count for i in range(N): abc('permute;write_aiger -u file2.aig') n = filecmp.cmp('file1.aig','file2.aig') print n, n_iso = n_iso+n print 'the number of isomorphisms was %d out of %d'%(n_iso,N) return n_iso def get_large_po(): ## remove_const_pos() #get rid of constant POs NL = n_latches() NO = n_pos() abc('&get') #put the in & space n_latches_max = 0 nl = imax = -1 for i in range(NO): #look for a big enough PO abc('&put;cone -s -O %d;scl'%i) nl = n_latches() if nl >.15*NL: imax = i ## print 'cone %d has %d FF'%(i,nl) break if nl> n_latches_max: n_latches_max = nl imax = i print i,nl if i == NO-1: print 'no PO is big enough' return -1 print 'PO_cone = %d, n_latches = %d'%(imax,nl) def scorro(): run_command('scorr -o') l = remove_const_pos() ps() def drw(): run_command('drw') ps() def dc2rs(): abc('dc2rs') ps() def reachn(t): x = time.clock() abc('&get;&reachn -rv -T %d'%t) print 'reachm3 done in time = %f'%(time.clock() - x) return get_status() def reachx(t): x = time.time() abc('reachx -t %d'%t) print 'reachx done in time = %f'%(time.time() - x) return get_status() def reachy(t): x = time.clock() abc('&get;&reachy -v -T %d'%t) print 'reachy done in time = %f'%(time.clock() - x) return get_status() def create_funcs(J,t): """evaluates strings indexed by J in methods given by FUNCS Returns a list of lambda functions for the strings in FUNCs If J = [], then create provers for all POs""" funcs = [] for j in range(len(J)): k=J[j] funcs = funcs + [eval(FUNCS[k])] return funcs def check_abs(): global init_initial_f_name abc('w %s_save.aig'%init_initial_f_name) ni = n_pis() nl = n_latches() na = n_ands() abc('r %s_smp_abs.aig'%init_initial_f_name) if ((ni == n_pis()) and (nl == n_latches()) and (na == n_ands())): return True else: abc('r %s_save.aig'%init_initial_f_name) return False def modify_methods(J,dec=0): """ adjusts the engines to reflect number of processors""" N = bmc_depth() L = n_latches() I = n_real_inputs() npr = n_proc -dec if 18 in J: #if sleep in J add 1 more processor npr = npr+1 if ( ((I+L<350)&(N>100)) or (I+L<260) or (L<80) ): if not 24 in J: #24 is reachs J = J+[24] # add all reach methods if len(J)>npr: J = remove_intrps(J) #removes only if len(J) Unsat: #still undefined return 'UNDECIDED' else: # (last_name == 'spec' or last_name == 'abs') - the last thing we did was an "abstraction" if result < Unsat: f_name = revert(f_name,1) # revert the f_name back to previous abc('r %s.aig'%f_name) return BMC_VER_result() #recursion here. def try_split(): abc('w %s_savetemp.aig'%f_name) na = n_ands() split(3) if n_ands()> 2*na: abc('r %s_savetemp.aig'%f_name) def time_diff(): global last_time new_time = time.clock() diff = new_time - last_time last_time = new_time result = 'Lapsed time = %.2f sec.'%diff return result def prove_all_ind(): """Tries to prove output k by induction, using other outputs as constraints. If ever an output is proved it is set to 0 so it can't be used in proving another output to break circularity. Finally all zero'ed outputs are removed. Prints out unproved outputs Finally removes 0 outputs """ global n_pos_proved, n_pos_before print 'n_pos_proved = %d'%n_pos_proved n_proved = 0 N = n_pos() ## l=remove_const_pos() ## print '0 valued output removal changed POs from %d to %d'%(N,n_pos()) if n_pos() == 1: return abc('w %s_osavetemp.aig'%f_name) lst = range(n_pos()) ## list.reverse() ## for j in list[1:]: for j in lst: ## abc('zeropo -N 0') abc('swappos -N %d'%j) ## l=remove_const_pos() #may not have to do this if constr works well with 0'ed outputs abc('constr -N %d'%(n_pos()-1)) abc('fold') n = max(1,n_ands()) f = max(1,min(40000/n,16)) f = int(f) ## abc('ind -C 10000 -F %d'%f) abc('ind -C 1000 -F %d'%f) ## run_command('print_status') status = get_status() abc('r %s_osavetemp.aig'%f_name) #have to restore original here if status == Unsat: ## print '+', abc('zeropo -N %d'%j) abc('w %s_osavetemp.aig'%f_name) #if changed, store it permanently if j < n_pos_before - n_pos_proved: n_proved = n_proved + 1 # keeps track of real POs proved. elif status < Unsat: print '-%d'%j, else: print '*%d'%j, l=remove_const_pos() n_pos_proved = n_pos_proved + n_proved print '\nThe number of POs reduced from %d to %d'%(N,n_pos()) print 'n_pos_proved = %d'%n_pos_proved #return status def remove_iso(L): global n_pos_proved, n_pos_before lst = [] for j in range(len(L)): ll = L[j][1:] if len(ll) == 0: continue else: lst = lst + ll zero(lst) n_pos_proved = n_pos_proved + count_less(lst,n_pos_before - n_pos_proved) print 'The number of POs removed by iso was %d'%len(lst) l=remove_const_pos() #can an original PO be zero? def prove_all_iso(): """Tries to prove output k by isomorphism. Gets number of iso-eq_classes as an array of lists. Updates n_pos_proved """ global n_pos_proved, n_pos_before n_proved = 0 N = n_pos() if n_pos() == 1: return print 'n_pos_proved = %d'%n_pos_proved ## run_command('&get;&iso;&put') abc('&get;&iso') L = iso_eq_classes() ## print L remove_iso(L) ## lim = n_pos_before - n_pos_proved ## for j in range(len(L)): ## ll = L[j] ## if len(ll) == 1: ## continue ## if not ll[0] < lim: ## continue ## else: ## n = count_less(ll[1:], lim) #drop the first since it is the representative. ## print n ## n_proved = n_proved + n ## print n, n_proved ## n_pos_proved = n_pos_proved + n_proved print '\nThe number of POs reduced by iso was from %d to %d'%(N,n_pos()) def count_less(L,n): count = 0 for j in range(len(L)): if L[j] < n: count = count + 1 return count def prove_all_mtds(t): """ Tries to prove output k with multiple methods in parallel, using other outputs as constraints. If ever an output is proved it is set to 0 so it can't be used in proving another output to break circularity. Finally all zero'ed ooutputs are removed. """ N = n_pos() ## l=remove_const_pos() ## print '0 valued output removal changed POs from %d to %d'%(N,n_pos()) abc('w %s_osavetemp.aig'%f_name) list = range(n_pos()) for j in list: run_command('swappos -N %d'%j) ## l=remove_const_pos() #may not have to do this if constr works well with 0'ed outputs abc('constr -N %d'%(n_pos()-1)) abc('fold') ## cmd = '&get;,pdr -vt=%d'%t #put in parallel. ## abc(cmd) verify(pdrs+bmcs+intrps+sims,t) status = get_status() abc('r %s_osavetemp.aig'%f_name) if status == Unsat: print '+', abc('zeropo -N %d'%j) abc('w %s_osavetemp.aig'%f_name) #if changed, store it permanently print '%d'%j, assert not is_sat(), 'one of the POs is SAT' #we can do better than this l=remove_const_pos() print '\nThe number of POs reduced from %d to %d'%(N,n_pos()) #return status def prove_all_pdr(t): """Tries to prove output k by pdr, using other outputs as constraints. If ever an output is proved it is set to 0 so it can't be used in proving another output to break circularity. Finally all zero'ed ooutputs are removed. """ N = n_pos() ## l=remove_const_pos() print '0 valued output removal changed POs from %d to %d'%(N,n_pos()) abc('w %s_osavetemp.aig'%f_name) list = range(n_pos()) for j in list: abc('swappos -N %d'%j) ## l=remove_const_pos() #may not have to do this if constr works well with 0'ed outputs abc('constr -N %d'%(n_pos()-1)) abc('fold') cmd = '&get;,pdr -vt=%d'%t #put in parallel. abc(cmd) status = get_status() abc('r %s_osavetemp.aig'%f_name) if status == Unsat: print '+', abc('zeropo -N %d'%j) abc('w %s_osavetemp.aig'%f_name) #if changed, store it permanently print '%d'%j, l=remove_const_pos() print '\nThe number of POs reduced from %d to %d'%(N,n_pos()) #return status def prove_each_ind(): """Tries to prove output k by induction, """ N = n_pos() l=remove_const_pos() print '0 valued output removal changed POs from %d to %d'%(N,n_pos()) abc('w %s_osavetemp.aig'%f_name) list = range(n_pos()) for j in list: abc('cone -s -O %d'%j) n = max(1,n_ands()) f = max(1,min(40000/n,16)) f = int(f) abc('ind -u -C 10000 -F %d'%f) status = get_status() abc('r %s_osavetemp.aig'%f_name) if status == Unsat: print '+', abc('zeropo -N %d'%j) abc('w %s_osavetemp.aig'%f_name) #if changed, store it permanently print '%d'%j, l=remove_const_pos() print '\nThe number of POs reduced from %d to %d'%(N,n_pos()) #return status def prove_each_pdr(t): """Tries to prove output k by PDR. If ever an output is proved it is set to 0. Finally all zero'ed ooutputs are removed. """ N = n_pos() l=remove_const_pos() print '0 valued output removal changed POs from %d to %d'%(N,n_pos()) abc('w %s_osavetemp.aig'%f_name) list = range(n_pos()) for j in list: abc('cone -O %d -s'%j) abc('scl -m') abc('&get;,pdr -vt=%d'%t) status = get_status() abc('r %s_osavetemp.aig'%f_name) if status == Unsat: print '+', abc('zeropo -N %d'%j) abc('w %s_osavetemp.aig'%f_name) #if changed, store it permanently print '%d'%j, l=remove_const_pos() print '\nThe number of POs reduced from %d to %d'%(N,n_pos()) #return status def disprove_each_bmc(t): """Tries to prove output k by PDR. If ever an output is proved it is set to 0. Finally all zero'ed ooutputs are removed. """ N = n_pos() l=remove_const_pos() print '0 valued output removal changed POs from %d to %d'%(N,n_pos()) abc('w %s_osavetemp.aig'%f_name) list = range(n_pos()) for j in list: abc('cone -O %d -s'%j) abc('scl -m') abc('bmc3 -T %d'%t) status = get_status() abc('r %s_osavetemp.aig'%f_name) if status == Sat: print '+', abc('zeropo -N %d'%j) abc('w %s_osavetemp.aig'%f_name) #if changed, store it permanently print '%d'%j, l=remove_const_pos() print '\nThe number of POs reduced from %d to %d'%(N,n_pos()) #return status def pord_1_2(t): """ two phase pord. First one tries with 10% of the time. If not solved then try with full time""" global n_pos_proved, ifpord1, pord_on pord_on = True # make sure that we do not reparameterize after abstract in prove_2 n_pos_proved = 0 if n_pos()<4: return Undecided if ifpord1: print 'Trying each output for %0.2f sec'%(.1*t) result = pord_all(.1*t) #we want to make sure that there is no easy cex. if (result <= Unsat): return result ifpord1 = 0 print 'Trying each output for %0.2f sec'%t #might consider using iso before the second pass of pord_all result = pord_all(t+2*G_T) #make sure there is enough time to abstract pord_on = False #done with pord return result def pord_all(t): """Tries to prove or disprove each output j by PDRM BMC3 or SIM. in time t""" global cex_list, n_pos_proved, last_cx, pord_on, ifpord1 print 'last_cx = %d'%last_cx btime = time.time() N = n_pos() prove_all_ind() ############ change this to keep track of n_pos_proved nn = n_pos() abc('w %s_osavetemp.aig'%f_name) if nn < 4: #Just cut to the chase immediately. return Undecided lst = range(n_pos()) proved = disproved = [] abc('&get') #using this space to save original file. ### Be careful that & space is not changed. ## with redirect.redirect( redirect.null_file, sys.stdout ): ## with redirect.redirect( redirect.null_file, sys.stderr ): cx_list = [] n_proved = 0 lcx = last_cx + 1 lst = lst[lcx:]+lst[:lcx] lst.reverse() n_und = 0 for j in lst: print '\ncone %s. '%j, abc('&put; cone -s -O %d'%j) #puts the &space into reg-space and extracts cone j #requires that &space is not changed. &put resets status. Use &put -s to keep status abc('scl -m') ps() ## print 'running sp2' ### result = run_sp2_par(t) ## J = slps+JV ## result = verify(J,t) ## result = RESULT[result] ## ### ## print 'run_sp2_par result is %s'%result if result == 'UNDECIDED': n_und = n_und + 1 status = Undecided if ((n_und > 1) and not ifpord1): break elif result == 'SAT': status = Sat disproved = disproved + [j] last_cx = j cx = cex_get() cx_list = cx_list + [cx] assert len(cx_list) == len(disproved), cx_list if len(cx_list) > 0: break else: #is unsat here status = Unsat proved = proved + [j] if j < n_pos_before - n_pos_proved: n_proved = n_proved +1 ## n_pos_proved = n_pos_proved + n_proved. #this should not be here because we should start fresh print '\nProved %d outputs'%len(proved) print 'Disproved %d outputs'%len(disproved) print 'Time for pord_all was %0.2f'%(time.time() - btime) NN = len(proved+disproved) cex_list = cx_list if len(disproved)>0: assert status == Sat, 'status = %d'%status n_pos_proved = 0 #we want to reset this because of a bad speculation return Sat else: n_pos_proved = n_pos_proved + n_proved abc('r %s_osavetemp.aig'%f_name) ## abc('&put') # returning original to work spece remove(proved) print '\nThe number of unproved POs reduced from %d to %d'%(N,n_pos()), ps() if n_pos() > 0: return Undecided else: return Unsat def bmc_ss(t): """ finds a set cexs in t seconds starting at 2*N where N is depth of bmc -T 1 The cexs are put in the global cex_list """ global cex_list x = time.time() tt = min(10,max(1,.05*t)) abc('bmc3 -T %0.2f'%tt) N = n_bmc_frames() if N <= max_bmc: return Undecided ## print bmc_depth() ## abc('bmc3 -C 1000000 -T %f -S %d'%(t,int(1.5*max(3,max_bmc)))) run_command('bmc3 -vs -C 1000000 -T %f -S %d'%(t,2*N)) if is_sat(): cex_list = cex_get_vector() #does this get returned from a concurrent process? n = count_non_None(cex_list) print '%d cexs found in %0.2f sec at frame %d'%(n,(time.time()-x),cex_frame()) return get_status() def list_non_None(lst): """ return [i for i,s in enumerate(cex_list) if not s == None]""" L = [] for i in range(len(lst)): if not lst[i] == None: L = L + [i] return L def count_non_None(lst): #return len([i for i,s in enumerate(cex_list) if not s == None] count = 0 for i in range(len(lst)): if not lst[i] == None: count = count + 1 return count def remove_disproved_pos(lst): for i in range(len(lst)): if not lst[i] == None: abc('zeropo -N %d'%i) l=remove_const_pos() def bmc_j(t): """ finds a cex in t seconds starting at 2*N where N is depth of bmc -T 1""" x = time.time() tt = min(5,max(1,.05*t)) abc('bmc3 -T %0.2f'%tt) if is_sat(): print 'cex found in %0.2f sec at frame %d'%((time.time()-x),cex_frame()) return get_status() ## abc('bmc3 -T 1') N = n_bmc_frames() N = max(1,N) ## print bmc_depth() ## abc('bmc3 -C 1000000 -T %f -S %d'%(t,int(1.5*max(3,max_bmc)))) cmd = 'bmc3 -J 2 -D 4000 -C 1000000 -T %f -S %d'%(t,2*N) ## print cmd abc(cmd) if is_sat(): print 'cex found in %0.2f sec at frame %d'%((time.time()-x),cex_frame()) return get_status() def pdrseed(t): """uses the abstracted version now""" ## abc('&get;,treb -rlim=60 -coi=3 -te=1 -vt=%f -seed=521'%t) abc('&get;,treb -rlim=100 -vt=%f -seed=521'%t) def pdrold(t): abc('&get; ,pdr -vt=%f'%t) def pdr(t): abc('&get; ,treb -vt=%f'%t) return RESULT[get_status()] def pdra(t): ## abc('&get; ,treb -rlim=100 -ssize -pre-cubes=3 -vt=%f'%t) abc('&get; ,treb -abs -rlim=100 -gen-cex -vt=%f'%t) return RESULT[get_status()] def pdrm(t): abc('pdr -C 0 -T %f'%t) return RESULT[get_status()] def pdrmm(t): abc('pdr -C 0 -M 298 -T %f'%t) return RESULT[get_status()] def split(n): abc('orpos;&get') abc('&posplit -v -N %d;&put;dc2'%n) trim() def keep_splitting(): for j in range(5): split(5+j) no = n_pos() status = prove_g_pos_split() if status <= Unsat: return status if no == n_pos(): return Undecided def drill(n): run_command('&get; &reachm -vcs -H 5 -S %d -T 50 -C 40'%n) def prove_1(ratio=.75): """ A version of prove called from prove_pos, prove_g_pos, prove_only, prove_g_pos_split when we have speculated and produced multiple outputs. Proves all the outputs together. If ever an abstraction was done then if SAT is returned,we make RESULT return "undecided". """ global x_factor,xfi,f_name,x, initial_f_name x = time.time() max_bmc = -1 print 'Initial: ', ps() x_factor = xfi initial_f_name_save = initial_f_name #needed because we are making local backups here. initial_f_name = '%s_temp'%initial_f_name set_globals() print'\n***Running abstract' status = abstract(ifbip) trim() status = process_status(status) if ((status <= Unsat) or status == Error): if status < Unsat: print 'CEX in frame %d'%cex_frame(), print 'abstract found a cex in initial circuit' print 'Time for proof = %f sec.'%(time.time() - x) initial_f_name = initial_f_name_save return RESULT[status] print 'Time for proof = %f sec.'%(time.time() - x) initial_f_name = initial_f_name_save return RESULT[status] #undecided here print 'Entering direct verificationb' #### status = final_verify_recur(2) status = BMC_VER() return status trim() #### write_file('final') print 'Time for proof = %f sec.'%(time.time() - x) initial_f_name = initial_f_name_save return RESULT[status] def pre_reduce(): x = time.clock() pre_simp() write_file('smp') abstract(ifbip) #### write_file('abs') print 'Time = %f'%(time.clock() - x) def sublist(L,I): # return [s for i,s in enumerate(L) if i in I] z = [] for i in range(len(I)): s = L[I[i]], s = list(s) z = z + s return z #PARALLEL FUNCTIONS """ funcs should look like funcs = [pyabc_split.defer(abc)('&get;,bmc -vt=50;&put'),pyabc_split.defer(super_prove)()] After this is executed funcs becomes a special list of lambda functions which are given to abc_split_all to be executed as in below. It has been set up so that each of the functions works on the current aig and possibly transforms it. The new aig and status is always read into the master when done """ def tf(): result = top_fork() return result def top_fork(J,t): global x_factor, final_verify_time, last_verify_time, methods set_globals() mtds = sublist(methods,J) F = create_funcs(J,t) print 'Running %s in parallel for max %d sec.'%(mtds,t) (m,result) = fork_last(F,mtds) #FORK here return get_status() def run_sp2_par(t): """ Runs the single method simple, timed for t seconds.""" global cex_list,methods J = slps+[6] print sublist(methods,J) funcs = create_funcs(J,t) y = time.time() for i,res in pyabc_split.abc_split_all(funcs): ## print 'i,res = %d,%s'%(i,res) t = time.time()-y if i == 0: print 'sleep timer expired in %0.2f'%t return 'UNDECIDED' else: ## print i,res if res == 'UNSAT': print 'Simple_prove proved UNSAT in %0.2f sec.'%t return 'UNSAT' elif res == 'UNDECIDED': print 'Simple_prove proved UNDECIDED in %0.2f sec.'%t return 'UNDECIDED' else: print 'Simple_prove found cex in %0.2f sec.'%t return 'SAT' def run_parallel(J,t,BREAK): """ Runs the listed methods J, each for time = t, in parallel and breaks according to BREAK = subset of '?USLB'""" global cex_list, methods mtds = sublist(methods,J) F = create_funcs(J,t) #if J = [] we are going to create functions that process each output separately. #if 18, then these are run in parallel with sleep if ((J == []) ): result = fork_break(F,mtds,BREAK) ## #redirect here to suppress printouts. ## with redirect.redirect( redirect.null_file, sys.stdout ): ## with redirect.redirect( redirect.null_file, sys.stderr ): ## result = fork_break(F,mtds,BREAK) elif 'L' in BREAK: result = fork_last(F,mtds) elif 'B' in BREAK: result = fork_best(F,mtds) else: result = fork_break(F,mtds,BREAK) return result def fork_all(funcs,mtds): """Runs funcs in parallel and continue running until all are done""" global methods y = time.time() for i,res in pyabc_split.abc_split_all(funcs): status = prob_status() t = time.time()-y if not status == -1: #solved here if status == 1: #unsat print '%s proved UNSAT in %f sec.'%(mtds[i],t) else: print '%s found cex in %f sec. - '%(mtds[i],t), if not mtds[i] == 'REACHM': print 'cex depth at %d'%cex_frame() else: print ' ' continue else: print '%s was undecided in %f sec. '%(mtds[i],t) return i,res def fork_break(funcs,mtds,BREAK): """ Runs funcs in parallel and breaks according to BREAK <= '?US' If mtds = 'sleep' or [], we are proving outputs in parallel Saves cex's found in cex_list in case we are proving POs. """ global methods,last_verify_time,seed,cex_list,last_winner,last_cex seed = seed + 3 # since parallel processes do not chenge the seed in the prime process, we need to change it here cex_list = lst = [] y = time.time() #use wall clock time because parent fork process does not use up compute time. for i,res in pyabc_split.abc_split_all(funcs): status = get_status() t = time.time()-y lm = len(mtds) if ((lm < 2) and not i == 0): # the only single mtds case is where it is 'sleep' M = 'Output %d'%(i-lm) else: M = mtds[i] last_winner = M if M == 'sleep': print 'sleep: time expired in %s sec.'%convert(t) assert status >= Unsat,'status = %d'%status break if ((status > Unsat) and '?' in BREAK): #undecided break elif status == Unsat: #unsat print '%s: UNSAT in %s sec.'%(M,convert(t)) if 'U' in BREAK: break elif status < Unsat: #status == 0 - cex found if M in methods: if methods.index(M) in exbmcs+allreachs+allpdrs+[1]: #set the known best depth so far. [1] is interp set_max_bmc(n_bmc_frames()) last_cex = M print '%s: -- cex in %0.2f sec. at depth %d => '%(M,t,cex_frame()), cex_list = cex_list+[cex_get()] #accumulates multiple cex's and puts them on list. if len(cex_list)>1: print 'len(cex_list): %d'%len(cex_list) if 'S' in BREAK: break else: continue return i,status def fork_best(funcs,mts): """ fork the functions, If not solved, take the best result in terms of AIG size""" global f_name n = len(mts)-1 y = time.time() m_best = -1 best_size = [n_pis(),n_latches(),n_ands()] ## print best_size abc('w %s_best_aig.aig'%f_name) for i,res in pyabc_split.abc_split_all(funcs): status = prob_status() ## print i, ## ps() ## print i,res, #ps() if not status == -1: #solved here m = i t = time.time()-y if status == 1: #unsat print '%s proved UNSAT in %f sec.'%(mtds[i],t) else: print '%s found cex in %f sec. - '%(mtds[i],t), break else: cost = rel_cost(best_size) ## print i,cost if cost < 0: best_size = [n_pis(),n_latches(),n_ands()] ## print best_size m_best = i ## print m_best abc('w %s_best_aig.aig'%f_name) abc('r %s_best_aig.aig'%f_name) return m_best,res def take_best(funcs,mts): """ fork the functions, If not solved, take the best result in terms of AIG size""" global f_name n = len(mts)-1 y = time.time() m_best = -1 best_size = 1000000 abc('w %s_best_aig.aig'%f_name) for i,res in pyabc_split.abc_split_all(funcs): if n_ands() < best_size: best_size = n_ands() m_best = i abc('w %s_best_aig.aig'%f_name) abc('r %s_best_aig.aig'%f_name) return m_best,res def fork_last(funcs,mtds): """ fork the functions, and take first definitive answer, but if last method ends first, then kill others""" n = len(mtds)-1 m = -1 y = time.time() lst = '' print mtds #print 'starting fork_last' for i,res in pyabc_split.abc_split_all(funcs): #print i,res status = prob_status() if not status == -1: #solved here m = i t = int(time.time()-y) if status == 1: #unsat print '%s proved UNSAT in %d sec.'%(mtds[i],t) else: print '%s found cex in %s sec. - '%(mtds[i],convert(t)), break elif i == n: t = int(time.time()-y) m = i print '%s: %d sec.'%(mtds[i],t) ps() break elif mtds[i] == 'sleep': res = Undecided t = time.time()-y print 'sleep timer expired in %0.2f'%t break lst = lst + ', '+mtds[i] return m,res def fork(funcs,mtds): """ runs funcs in parallel This keeps track of the verify time when a cex was found, and if the time to find the cex was > 1/2 allowed time, then last_verify_time is increased by 2""" global win_list, methods, last_verify_time,seed beg_time = time.time() i,res = fork_break(funcs,mtds,'US') #break on Unsat of Sat. t = time.time()-beg_time #wall clock time because fork does not take any compute time. if t > .4*last_verify_time: ## if t > .15*last_verify_time: ##### temp t = last_verify_time = last_verify_time + .1*t #print 'verify time increased to %s'%convert(t) assert res == get_status(),'res: %d, status: %d'%(res,get_status()) return i,res def save_time(M,t): global win_list,methods j = methods.index(M) win_list = win_list + [(j,t)] #print win_list def summarize(lst): result = [0]*10 for j in range(len(lst)): k = lst[j] result[k[0]]=result[k[0]]+k[1] return result def top_n(lst,n): result = [] ll = list(lst) #makes a copy m = min(n,len(ll)) for i in range(m): mx_index = ll.index(max(ll)) result = result + [mx_index] ll[mx_index] = -1 return result def super_pre_simp(): while True: nff = n_latches() print 'Calling pre_simp' pre_simp() if n_latches() == nff: break #______________________________ #new synthesis command ####def synculate(t): #### """ #### Finds candidate sequential equivalences and refines them by simulation, BMC, or reachability #### using any cex found. If any are proved, then they are used to reduce the circuit. The final aig #### is a new synthesized circuit where all the proved equivalences are merged. #### If we put this in a loop with increasing verify times, then each time we work with a simpler model #### and new equivalences. Should approach srm. If in a loop, we can remember the cex_list so that we don't #### have to deal with disproved equivalences. Then use refine_with_cexs to trim the initial equivalences. #### If used in synthesis, need to distinguish between #### original outputs and new ones. Things to take care of: 1. a PO should not go away until it has been processes by merged_proved_equivalences #### 2. Note that &resim does not use the -m option where as in speculation - m is used. It means that if #### an original PO isfound to be SAT, the computation quits becasue one of the output #### miters has been disproved. #### """ #### global G_C,G_T,n_pos_before, x_factor, n_latches_before, last_verify_time, f_name,cex_list, max_verify_time #### #### #### def refine_with_cexs(): #### """Refines the gores file to reflect equivalences that go away because of cexs in cex_list""" #### global f_name #### abc('&r %s_gores.aig'%f_name) #### for j in range(len(cex_list)): #### cex_put(cex_list[j]) #### run_command('&resim') #put the jth cex into the cex space and use it to refine the equivs #### abc('&w %s_gores.aig'%f_name) #### return #### #### def generate_srms(): #### """generates a synthesized reduced model (srms) from the gores file""" #### global f_name, po_map #### abc('&r %s_gores.aig; &srm -sf; r gsrms.aig; w %s_gsrms.aig'%(f_name,f_name)) #### print 'New srms = ',ps() #### po_map = range(n_pos()) #### return 'OK' #### #### def merge_proved_equivalences(): #### #this only changes the gores file. #### run_command('&r %s_gores.aig; &equiv_mark -vf %s_gsrms.aig; &reduce -v; &w %s_gores.aig'%(f_name,f_name,f_name)) #### return #### #### def generate_equivalences(): #### set_globals() #### t = max(1,.5*G_T) #### r = max(1,int(t)) #### cmd = "&get; &equiv2 -C %d -F 200 -T %f -S 1 -R %d"%((G_C),t,r) #### abc(cmd) #### #run_command('&ps') #### eq_simulate(.5*t) #### #run_command('&ps') #### cmd = '&semi -W 63 -S 5 -C 500 -F 20 -T %d'%(.5*t) #### abc(cmd) #### #run_command('&ps') #### run_command('&w %s_gores.aig'%f_name) #### #### l=remove_const_pos() #makes sure no 0 pos to start #### cex_list = [] #### n_pos_before = n_pos() #### n_latches_before = n_latches() ###### print 'Generating equivalences' #### generate_equivalences() ###### print 'Generating srms file' #### generate_srms() #this should not create new 0 pos ###### if n_pos()>100: ###### removed #### l=remove_const_pos() #### n_pos_last = n_pos() #### if n_pos_before == n_pos(): #### print 'No equivalences found. Quitting synculate' #### return Undecided_no_reduction #### print 'Initial synculation: ',ps() ###### ps() #### set_globals() #### simp_sw = init = True #### simp_sw = False #temporary #### print '\nIterating synculation refinement' #### abc('w initial_sync.aig') #### max_verify_time = t #### print 'max_verify_time = %d'%max_verify_time #### """ #### in the following loop we increase max_verify_time by twice time spent to find last cexs or Unsat's #### We iterate only when we have proved cex + unsat > 1/2 n_pos. Then we update srms and repeat. #### """ #### while True: # refinement loop #### t = max_verify_time #this may have been increased since the last loop ###### print 'max_verify_time = %d'%max_verify_time #### set_globals() #### if not init: #### generate_srms() #generates a new gsrms file and leaves it in workspace ###### print 'generate_srms done' #### if n_pos() == n_pos_before: #### break #### if n_pos() == n_pos_last: #if nothing new, then quit if max_verification time is reached. #### if t > max_verify_time: #### break #### if simp_sw: #Warning: If this holds then simplify could create some 0 pos #### na = n_ands() #### simplify() #### while True: #### npo = n_pos() ###### print 'npos = %d'%npo #### merge_proved_equivalences() #So we need to merge them here. Can merging create more??? #### generate_srms() #### if npo == n_pos(): #### break #### if n_ands() > .7*na: #if not significant reduction, stop simplification #### simp_sw = False #simplify only once. #### if n_latches() == 0: #### return check_sat() #### n_pos_last = n_pos() #### init = False # make it so that next time it is not the first time through #### syn_par(t) #### if (len(cex_list)+len(result)) == 0: #nothing happened aand ran out of time. #### break #### abc('w %s_gsrms.aig'%f_name) #### #print 'No. of cexs after syn_parallel = %d'%len(cex_list) #### merge_proved_equivalences() #changes the underlying gores file by merging fanouts of proved eqs #### #print 'merge done' #### refine_with_cexs() #changes the gores file by refining the equivalences in it using cex_list. #### #print 'refine_with_cexs done' #### continue #### extract(0,n_pos_before) #get rid of unproved outputs #### return #### ####def syn_par(t): #### """prove n outputs at once and quit at first cex. Otherwise if no cex found return aig #### with the unproved outputs""" #### global trim_allowed,max_verify_time, n_pos_before #### global cex_list, result #### b_time = time.time() #### n = n_pos() #### if n == n_pos_before: #### return #### mx = n_pos() #### if n_pos() - n_pos_before > 50: #### mx = n_pos_before + 50 #### r = range(n_pos_before, mx) #### N = max(1,(mx-n_pos_before)/2) #### abc('w %s__ysavetemp.aig'%f_name) #### F = [eval(FUNCS[18])] #create a timer function #### #print r #### for i in r: #### F = F + [eval('(pyabc_split.defer(verify_only)(%d,%d))'%(i,t))] #### cex_list = result = [] #### outcome = '' #### #redirect printout here ###### with redirect.redirect( redirect.null_file, sys.stdout ): ###### with redirect.redirect( redirect.null_file, sys.stderr ): #### for i,res in pyabc_split.abc_split_all(F): #### status = get_status() ###### print i #### if i == 0: #timed out #### outcome = 'time expired after = %d'%(time.time() - b_time) #### break #### if status < Unsat: #### cex_list = cex_list + [cex_get()] #### if status == Unsat: #### result = result + [r[i-1]] #### if (len(result)+len(cex_list))>= N: #### T = time.time() - b_time #### if T > max_verify_time/2: #### max_verify_time = 2*T #### break #### continue #### if not outcome == '': #### print outcome ###### print 'cex_list,prove_list = ',cex_list,result #### abc('r %s__ysavetemp.aig'%f_name) #restore initial aig so that pos can be 0'ed out #### if not result == []: # found some unsat's ###### min_r = min(result) ###### max_r = max(result) ###### no = n_pos() ###### assert (0 <= min_r and max_r < no), 'min_r, max_r, length = %d, %d, %d'%(min_r,max_r,len(result)) #### zero(result) #### return #### #print "Number PO's proved = %d"%len(result) #### ####def absec(n): #### #abc('w t.aig') #### for j in range(n): #### print '\nFrame %d'%(j+1) #### run_command('absec -F %d'%(j+1)) #### if is_unsat(): #### print 'UNSAT' #### break #### #### ####""" #### we might be proving some original pos as we go, and on the other hand we might have some equivalences that we #### can't prove. There are two uses, in verification #### verification - we want to remove the proved pos whether they are original or not. But if a cex for an original, then need to #### remember this. #### synthesis - the original outputs need to be kept and ignored in terms of cex's - supposedly they can't be proved. ####""" #### ####""" Experimental""" #### ####def csec(): #### global f_name #### if os.path.exists('%s_part0.aig'%f_name): #### os.remove('%s_part0.aig'%f_name) #### run_command('demiter') #### if not os.path.exists('%s_part0.aig'%f_name): #### return #### run_command('r %s_part0.aig'%f_name) #### ps() #### run_command('comb') #### ps() #### abc('w %s_part0comb.aig'%f_name) #### run_command('r %s_part1.aig'%f_name) #### ps() #### run_command('comb') #### ps() #### abc('w %s_part1comb.aig'%f_name) #### run_command('&get; &cec %s_part0comb.aig'%(f_name)) #### if is_sat(): #### return 'SAT' #### if is_unsat(): #### return 'UNSAT' #### else: #### return 'UNDECIDED' ########################### #### we will verify outputs ORed in groups of g[i] #### here we take div = N so no ORing ## div = max(1,N/1) ## g = distribute(N,div) ## if len(g) <= 1: ## t = tt ## g.reverse() #### print g ## x = 0 ## G = [] ## for i in range(div): ## y = x+g[i] ## F = F + [eval('(pyabc_split.defer(verify_range)(%d,%d,%s))'%(x,y,convert(t)))] ## G = G + [range(x,y)] ## x = y #### print G ########################### """ These commands map into luts and leave the result in mapped format. To return to aig format, you have to do 'st' """ def sop_balance(k=4): '''minimizes LUT logic levels ''' ## kmax = k kmax=min(k+2,15) abc('st; if -K %d;ps'%kmax) print nl(), ## for i in range(1): ## abc('st; if -K %d;ps'%kmax) ## run_command('ps') kmax=min(k+2,15) abc('st; if -g -C %d -K %d -F 2;ps'%(10,kmax)) #balance print nl(), for i in range(1): abc('st;dch; if -C %d -K %d;ps'%(10,kmax)) print nl(), def speedup(k=4): run_command('speedup;if -K %d'%k) print nl() def speed_tradeoff(k=4): print nl(), best = n_nodes() abc('write_blif %s_bestsp.blif'%f_name) L_init = n_levels() while True: L_old = n_levels() L = L_old -1 abc('speedup;if -D %d -F 2 -K %d -C 11'%(L,k)) if n_nodes() < best: best = n_nodes() abc('write_blif %s_bestsp.blif'%f_name) if n_levels() == L_old: break print nl(), continue abc('r %s_bestsp.blif'%f_name) return def map_lut_dch(k=4): '''minimizes area ''' abc('st; dch; if -a -F 2 -K %d -C 11; mfs -a -L 50 ; lutpack -L 50'%k) def map_lut_dch_iter(k=8): ## print 'entering map_lut_dch_iter with k = %d'%k best = n_nodes() abc('write_blif %s_best.blif'%f_name) ## abc('st;dch;if -a -K %d -F 2 -C 11; mfs -a -L 1000; lutpack -L 1000'%k) ## if n_nodes() < best: ## abc('write_blif %s_best.blif'%f_name) ## best = n_nodes() ## print nl(), ## else: ## abc('r %s_best.blif'%f_name) ## best = n_nodes() ## abc('write_blif %s_best.blif'%f_name) ## print 'best = %d'%best n=0 while True: map_lut_dch(k) if n_nodes()< best: best = n_nodes() ## print 'best=%d'%best n = 0 abc('write_blif %s_best.blif'%f_name) print nl(), continue else: n = n+1 if n>2: break abc('r %s_best.blif'%f_name) def dmitri_iter(k=8): best = 100000 n=0 while True: dmitri(k) if n_nodes()< best: best = n_nodes() ## print '\nbest=%d'%best n = 0 abc('write_blif %s_best.blif'%f_name) continue else: n = n+1 if n>2: break abc('r %s_best.blif'%f_name) ## run_command('cec -n %s.aig'%f_name) print nl() def map_lut(k=4): '''minimizes edge count''' for i in range(5): abc('st; if -e -K %d; ps; mfs ;ps; lutpack -L 50; ps'%(k)) print nl(), def extractax(o=''): abc('extract -%s'%o) def nl(): return [n_nodes(),n_levels()] def dc2_iter(th=.999): abc('st') while True: na=n_ands() abc('dc2') print n_ands(), ## print nl(), if n_ands() > th*na: break ## print n_ands() def adc2_iter(th=.999): abc('st;&get') while True: na=n_ands() abc('&dc2;&put') ## print n_ands(), if n_ands() > th*na: break print n_ands() def try_extract(): ## abc('dc2;dc2') print 'Initial: ', ps() na = n_ands() ## abc('w %s_savetemp.aig'%f_name) #no need to save initial aig since fork_best will return initial if best. J = [32,33] mtds = sublist(methods,J) F = create_funcs(J,0) (m,result) = take_best(F,mtds) #FORK here if not m == -1: print 'Best extract is %s: '%mtds[m], ps() ## if (n_ands() < na): ## return ## else: ## abc('r %s_savetemp.aig'%f_name) def speedup_iter(k=8): abc('st;if -K %d'%k) run_command('ps') abc('write_blif %s_bests.blif'%f_name) run_command('ps') best = n_levels() print 'n_levels before speedup = %d'%n_levels() n=0 while True: nl() abc('speedup;if -K %d'%k) if n_levels() < best: best = n_levels() abc('write_blif %s_bests.blif'%f_name) n=0 else: n = n+1 if n>2: break abc('r %s_bests.blif'%f_name) print 'n_levels = %d'%n_levels() def jog(n=16): """ applies to a mapped blif file""" run_command('eliminate -N %d;fx'%n) run_command('if -K %d'%(n/2)) run_command('fx') def perturb_f(k=4): abc('st;dch;if -g -K %d'%(k)) ## snap() abc('speedup;if -K %d -C 10'%(k)) jog(5*k) ## snap() ## abc('if -a -K %d -C 11 -F 2;mfs -a -L 50;lutpack -L 50'%k def perturb(k=4): abc('st;dch;if -g -K %d'%k) ## snap() abc('speedup;if -K %d -C 10'%(k)) def preprocess(k=4): n_initial = n_nodes() abc('write_blif %s_temp_initial.blif'%f_name) ## abc('st;dc2') abc('w %s_temp_initial.aig'%f_name) ni = n_pis() + n_latches() res = 1 if ni >= 101: abc('st;if -a -F 2 -K %d'%k) return res ## dc2_iter() abc('st;if -a -K %d'%k) # to get plain direct map if n_nodes() > n_initial: abc('r %s_temp_initial.blif'%f_name) res = 1 #plain n_plain = n_nodes() ## print nl() abc('write_blif %s_temp_plain.blif'%f_name) #clp abc('st;clp; if -a -K %d'%k) ## print nl() abc('write_blif %s_temp_clp.blif'%f_name) n_clp = n_nodes() #clp_lutmin abc('r %s_temp_initial.blif'%f_name) abc('st;clp;lutmin -K %d;'%k) abc('write_blif %s_temp_clp_lut.blif'%f_name) n_clp_lut = n_nodes() ## print nl() if n_plain <= min(n_clp,n_clp_lut): abc('r %s_temp_plain.blif'%f_name) res = 1 elif n_clp < n_clp_lut: abc('r %s_temp_clp.blif'%f_name) res = 1 else: abc('r %s_temp_clp_lut.blif'%f_name) res = 1 ## print nl() return res def snap(): ## abc('fraig;fraig_store') abc('fraig_store') def snap_bestk(k): abc('write_blif %s_temp.blif'%f_name) unsave_bestk(k) snap() abc('r %s_temp.blif'%f_name) def cec_it(): """ done because &r changes the names. Can't use -n because rfraig_store reorders pis and pos.""" abc('write_blif %s_temp.blif'%f_name) abc('&r %s.aig;&put'%f_name) run_command('cec %s_temp.blif'%f_name) abc('r %s_temp.blif'%f_name) def save_bestk(b,k): ## if os.access('%s_best%d.blif'%(f_name,k),os.R_OK): ## res = get_bestk(k) ## else: """ saves the best, returns bestk and if not best, leaves blif unchanged""" res = b if n_nodes() < res: res = n_nodes() abc('write_blif %s_best%d.blif'%(f_name,k)) print 'best%d = %d'%(k,res) return res ## unsave_bestk(k) def unsave_bestk(k): abc('r %s_best%d.blif'%(f_name,k)) def unsnap(k=4): ## snap() abc('fraig_restore') map_lut_dch(k) ## abc('fraig_restore;if -a -F 2 -C 11 -K %d'%k) def map_until_conv(k=4): kk = 2*k # make sure that no residual results are left over. if os.access('%s_best%d.blif'%(f_name,k),os.R_OK): os.remove('%s_best%d.blif'%(f_name,k)) if os.access('%s_best%d.blif'%(f_name,kk),os.R_OK): os.remove('%s_best%d.blif'%(f_name,kk)) tt = time.time() #get initial map and save map_lut_dch(k) bestk = save_bestk(100000,k) print nl() ## snap() res = preprocess() #get best of initial, clp, and lutmin versions print nl() ## map_lut_dch(k) ## ### ## bestk = save_bestk(bestk,k) ## map_iter(k) ## bestk = save_bestk(bestk,k) ## ### map_lut_dch_iter(kk) #initialize with mapping with 2k input LUTs bestkk = save_bestk(100000,kk) snap() unsnap(k) #have to do snap first if want current result snapped in. # unsnap fraigs snapshots and does map_lut_dch at end print nl() bestk = save_bestk(bestk,k) abc('r %s_bestk%d.blif'%(f_name,k)) map_iter(k) #1 bestk = save_bestk(bestk,k) while True: print 'Perturbing with %d-Lut'%kk ## snap() map_lut_dch_iter(kk) ## snap() bestkk_old = bestkk bestkk = save_bestk(bestkk,kk) if bestkk >= bestkk_old: break ## snap() ## jog(kk) ## perturb_f(k) ## snap() ## perturb_f(k) ## snap() ## unsave_bestk(k) ## map_lut_dch(k+1) ## snap() ## snap_bestk(k) snap() unsnap(k) #fraig restore and map ## bestk = save_bestk(bestk,k) ## snap() bestk_old = bestk map_iter(k) bestk = save_bestk(bestk,k) if bestk >= bestk_old: break continue abc('fraig_restore') #dump what is left in fraig_store unsave_bestk(k) print '\nFinal size = ', print nl() print 'time for %s = %.02f'%(f_name,(time.time()-tt)) ## cec_it() def get_bestk(k=4): abc('write_blif %s_temp.blif'%f_name) unsave_bestk(k) res = n_nodes() abc('r %s_temp.blif'%f_name) return res def map_iter(k=4): tt = time.time() bestk = get_bestk(k) ## bestk = n_nodes() ## bestk = save_bestk(bestk,k) ## abc('st;dch;if -a -F 2 -K %d -C 11; mfs -a -L 1000; lutpack -L 1000'%k)#should be same as Initial ## map_lut_dch_iter(k) #### map_lut_dch(k) bestk = save_bestk(bestk,k) n=0 unsave_bestk(k) while True: ## snap() perturb(k) # ## snap() perturb(k) ## snap_bestk(k) ## unsnap(k) ## bestk = save_bestk(bestk,k) ## snap() ## map_lut_dch(k+1) ## abc('if -K %d'%(k+1)) ## snap() ## unsnap(k) old_bestk = bestk ## print old_bestk map_lut_dch_iter(k) bestk = save_bestk(bestk,k) print bestk if bestk < old_bestk: n=0 # keep it up continue elif n == 2: #perturb break else: n = n+1 print '%d-perturb'%n ## snap() ## unsave_bestk(k) unsave_bestk(k) def map_star(k=4): tt = time.time() map_until_conv(k) abc('write_blif %s_best_star.blif'%f_name) best = n_nodes() while True: jog(2*k) map_until_conv(k) if n_nodes() >= best: break else: best = n_nodes() abc('write_blif %s_best_star.blif'%f_name) abc('r %s_best_star.blif'%f_name) print 'SIZE = %d, TIME = %.2f for %s'%(n_nodes(),(time.time() - tt),f_name) def decomp_444(): abc('st; dch; if -K 10 -S 444') abc('write_blif -S 444 %s_temp.blif; r %s_temp.blif'%(f_name,f_name)) def dmitri(k=8): ## abc('w t.aig') ## dc2_iter() ## print 'first iter done: %d'%n_ands() ## abc('dc2rs') #### dc2_iter() ## print 'second iter done: %d'%n_ands() ## sop_balance(k) ## abc('w t_before.aig') ## run_command('cec -n t.aig') ## speedup_iter(k) ## print 'n_levels after speedup = %d'%n_levels() ## abc('write_blif %s_save.blif'%f_name) ## nn=n_levels() abc('st;dch; if -g -K %d'%(k)) ## print 'n_levels after sop balance = %d'%n_levels() ## if n_levels() > nn: ## run_command('r %s_save.blif'%f_name) ## print 'n_levels = %d'%n_levels() ## print 'final n_levels = %d'%n_levels() ## print 'sop_balance done: ', ## print nl() ## run_command('st;w t_after.aig') ## run_command('cec -n t.aig') abc('if -G %d '%k) ## print 'after if -G %d: '%k, ## print nl() ## run_command('cec -n t.aig') abc('cubes') ## print 'after cubes: ', ## print nl() ## run_command('cec -n t.aig') abc('addbuffs -v') ## print 'after addbuffs: ', print nl(), ## run_command('cec -n t.aig') def lut(): dc2_iter() abc('extract -a') print nl() dc2_iter() ## ps() sop_balance(6) map_lut_dch() map_lut() print nl() ## run_command('ps') ################################## gate level abstraction """ Code for using for abstraction """ def bip_abs(t=100): """ t is ignored here""" set_globals() time = max(1,.1*G_T) abc('&get;,bmc -vt=%f'%time) set_max_bmc(bmc_depth()) c = 2*G_C f = max(2*max_bmc,20) b = min(max(10,max_bmc),200) t1 = x_factor*max(1,2*G_T) t = max(t1,t) s = min(max(3,c/30000),10) # stability between 3 and 10 ## cmd = '&get;,abs -bob=%d -stable=%d -timeout=%d -vt=%d -depth=%d -dwr=vabs'%(b,s,t,t,f) cmd = '&get;,abs -timeout=%d -vt=%d -dwr=%s_vabs'%(t,t,f_name) print 'Running %s'%cmd ## abc(cmd) run_command(cmd) bmc_depth() abc('&w %s_greg.aig'%f_name) return max_bmc def check_frames(): abc('read_status vta.status') return n_bmc_frames() def gate_abs(t): """ Do gate-level abstraction for F frames """ r = 100 *(1 - abs_ratio) run_command('orpos; &get;&vta -dv -A %s_vabs.aig -P 2 -T %d -R %d; &vta_gla;&gla_derive; &put'%(f_name,t,r)) ## write_file('abs') def gla_abs(t): """ Do gate-level abstraction for F frames """ r = 100 *(1 - abs_ratio) run_command('orpos; &get;&gla_cba -C 0 -T %d -F 0 -R %d; &gla_derive; &put'%(t,r)) def sizeof(): return [n_pis(),n_pos(),n_latches(),n_ands()] def abstract(ifb=2): global abs_ratio ## print 'ifb = %d'%ifb if ifb == 0: #new way using gate_abs and no bip return abstracta(False) elif ifb == 1: #old way using ,abs assert ifb == ifbip, 'call to abstract has ifb not = global ifbip' return abstractb() else: #new way using ,abs -dwr -- (bip_abs) return abstracta(True) def abstracta(if_bip=True): """ if_bip = 0 it uses a new abstraction based on &vta (gate level abstraction) and no bip operations Right now, if we do not prove it with abstraction in the time allowed, we abandon abstraction and go on with speculation if_bip = 1, we use ,abs -dwr """ global G_C, G_T, latches_before_abs, x_factor, last_verify_time, x, win_list, j_last, sims global latches_before_abs, ands_before_abs, pis_before_abs, abs_ratio ## n_vabs = 0 latches_before_abs = n_latches() ands_before_abs = n_ands() pis_before_abs = n_real_inputs() tt = time.time() print 'using abstracta, ', ## print 'if_bip = %d'%if_bip ## latch_ratio = abs_ratio ## t = 100 t = 1000 #temporary t = abs_time if if_bip == 0: t = 1000 #timeout on vta t = abs_time tt = time.time() if n_pos() > 1 and if_bip == 0: abc('orpos') print 'POs ORed together, ', initial_size = sizeof() abc('w %s_before_abs.aig'%f_name) # 25 below means that it will quit if #FF+#ANDS > 75% of original ## funcs = [eval("(pyabc_split.defer(abc)('orpos;&get;&vta -d -R 25'))")] #right now we need orpos if if_bip: print 'using bip_abs' mtds = ['bip_abs'] funcs = [eval('(pyabc_split.defer(bip_abs)(t))')] else: print 'using gate_abs' mtds = ['gate_abs'] funcs = [eval('(pyabc_split.defer(gate_abs)(t))')] funcs = funcs + [eval('(pyabc_split.defer(monitor_and_prove)())')] J = [34,30] if n_ands()> 500000: #if greater than this, bmc_j may take too much memory. J = [34] ## J=[] funcs = funcs + create_funcs(J,1000) mtds = mtds + ['monitor_and_prove'] + sublist(methods,J) print 'methods = ', print mtds vta_term_by_time=0 for i,res in pyabc_split.abc_split_all(funcs): ## print i,res if i == 0: #vta ended first print 'time taken = %0.2f'%(time.time() - tt) if is_sat(): print 'vta abstraction found cex in frame %d'%cex_frame() return Sat if is_unsat(): print 'vta abstraction proved UNSAT' return Unsat else: #undecided if time.time() - tt < .95*t: print 'abstraction terminated but not by timeout' vta_term_by_time = 0 break else: print 'abstraction terminated by a timeout of %d'%t ## print 'final abstraction: ', ## ps() vta_term_by_time=1 break if i == 1: #monitor and prove ended first (sleep timed out) ## print i,res if res > Unsat: #we abandon abstraction ## print 'final abstraction: ', ## ps() ## print 'Trying to verify final abstraction' ## result = verify([7,9,19,23,24,30],100) #do this if if_bip==0 ## if result == Unsat: ## print 'Abstraction proved valid' ## return result ## else: ## print 'Abstract time wasted = %0.2f'%(time.time()-tt) ## abc('r %s_before_abs.aig'%f_name) ## result = Undecided_no_reduction ## return result ## elif res == Undecided_no_reduction: print 'monitor and prove timed out or little reduction' abc('r %s_before_abs.aig'%f_name) return Undecided_no_reduction else: if not initial_size == sizeof(): #monitor and prove should not return SAT in this case' assert not is_sat(), 'monitor_and_prove returned SAT on abstraction!' print 'time taken = %0.2f'%(time.time() - tt) if is_unsat(): return Unsat elif is_sat(): return Sat else: abc('r %s_before_abs.aig'%f_name) return Undecided_no_reduction else: #one of the engines got an answer print 'time taken = %0.2f'%(time.time() - tt) if is_unsat(): print 'Initial %s proved UNSAT'%mtds[i] return Unsat if is_sat(): print 'Initial %s proved SAT'%mtds[i] return Sat else: # an engine failed here print 'Initial %s terminated without result'%mtds[i] ## return Undecided continue if vta_term_by_time == 0 and if_bip == 0: #vta timed out itself print 'Trying to verify final abstraction', ps() result = verify([7,9,19,23,24,30],100) if result == Unsat: print 'Abstraction proved valid' return result # should do abstraction refinement here if if_bip==1 if if_bip == 0: print 'abstraction no good - restoring initial simplified AIG' abc('r %s_before_abs.aig'%f_name) return Undecided_no_reduction else: if is_sat(): print 'Found true counterexample in frame %d'%cex_frame() return Sat_true if is_unsat(): return Unsat ## set_max_bmc(NBF) NBF = bmc_depth() print 'Abstraction good to %d frames'%max_bmc #note when things are done in parallel, the &aig is not restored!!! abc('&r %s_greg.aig; &w initial_greg.aig; &abs_derive; &put; w initial_gabs.aig; w %s_gabs.aig'%(f_name,f_name)) set_max_bmc(NBF) print 'Initial abstraction: ', ps() abc('w %s_init_abs.aig'%f_name) latches_after = n_latches() ## if latches_after >= .90*latches_before_abs: #the following should match similar statement ## if ((rel_cost_t([pis_before_abs, latches_before_abs, ands_before_abs])> -.1) or ## (latches_after >= .75*latches_before_abs)): if small_abs(abs_ratio): abc('r %s_before_abs.aig'%f_name) print "Little reduction!" print 'Abstract time wasted = %0.2f'%(time.time()-tt) return Undecided_no_reduction sims_old = sims sims=sims[:1] #make it so that rarity sim is not used since it can't find a cex result = abstraction_refinement(latches_before_abs, NBF,abs_ratio) sims = sims_old if result <= Unsat: return result ## if n_latches() >= .90*latches_before_abs: ## if ((rel_cost_t([pis_before_abs, latches_before_abs, ands_before_abs])> -.1) or (latches_after >= .90*latches_before_abs)): ## if rel_cost_t([pis_before_abs,latches_before_abs, ands_before_abs])> -.1: if small_abs(abs_ratio): #r is ratio of final to initial latches in absstraction. If greater then True abc('r %s_before_abs.aig'%f_name) #restore original file before abstract. print "Little reduction! ", print 'Abstract time wasted = %0.2f'%(time.time()-tt) result = Undecided_no_reduction return result #new else: write_file('abs') #this is only written if it was not solved and some change happened. print 'Abstract time = %0.2f'%(time.time()-tt) return result def monitor_and_prove(): """ monitor and prove. whenever a new vabs is found, try to verify it """ global ifbip #write the current aig as vabs.aig so it will be regularly verified at the beginning. ## print 'Entering monitora_and_prove' print ifbip run_command('w %s_vabs.aig'%f_name) if ifbip == 0: run_command('w vabs.aig') initial_size = sizeof() print 'initial size = ', print initial_size funcs = [eval('(pyabc_split.defer(read_and_sleep)())')] t = 1000 # do not want to timeout verification engines. t = abs_time J = [9,19,23,24,34] #engines BMC3,PDRMsd,INTRPm,REACHY - engines for first time through when no abstraction funcs = funcs + create_funcs(J,t) mtds = ['read_and_sleep'] + sublist(methods,J) print 'methods = %s'%mtds #a return of Undecided means that abstraction might be good and calling routine will check this while True: time_done = abs_bad = 0 for i,res in pyabc_split.abc_split_all(funcs): ## print i,res if i == 0: # read and sleep terminated if res == False: #found new abstraction abs_bad = 0 #new abs starts out good. if not initial_size == sizeof() and n_latches() > abs_ratio * initial_size[2]: return Undecided_no_reduction else: break elif res == True: # read and sleep timed out time_done = 1 ## print 'read_and_sleep timed out' if abs_bad: return Undecided_no_reduction else: #abs is still good. Let other engines continue return Undecided #calling routine handles >Unsat all the same right now. else: assert False, 'something wrong. read and sleep did not return right thing' if i > 0: #got result from one of the verify engines ## print 'method %s found SAT in frame %d'%(mtds[i],cex_frame()) if is_unsat(): print 'Parallel method %s proved UNSAT on current abstraction'%mtds[i] return Unsat if is_sat(): #abstraction is not good yet. print 'Parallel method %s found SAT on current abstraction in frame %d'%(mtds[i],cex_frame()) ## print 'n_vabs = %d'%n_vabs if initial_size == sizeof():# the first time we were working on an aig before abstraction return Sat ## print 'current abstraction invalid' abs_bad = 1 break #this kills off other verification engines working on bad abstraction else: #one of the engines undecided for some reason - failed? print 'Parallel %s terminated without result on current abstraction'%mtds[i] continue if abs_bad and not time_done: #here we wait until have a new vabs. print 'current abstraction bad, waiting for new one' ## print 'waiting for new abstraction' abc('r %s_abs.aig'%f_name) #read in the abstraction to destroy is_sat. res = read_and_sleep(5) #this will check every 5 sec, until abs_time sec has passed without new abs if res == False: #found new vabs. Now continue if vabs small enough ## print 'n_vabs = %d'%n_vabs if (not initial_size == sizeof()) and n_latches() > abs_ratio * initial_size[2]: return Undecided_no_reduction else: continue elif res ==True: #read_and_sleep timed out ## print 'read_and_sleep timed out' return Undecided_no_reduction else: break #this should not happen elif abs_bad and time_done: print 'current abstraction bad, time has run out' return Undecided_no_reduction elif time_done: #abs is good here print 'current abstraction still good, time has run out' return Undecided #this will cause calling routine to try to verify the final abstraction #right now handles the same as Undecided_no_reduction-if time runs out we quit abstraction else: #abs good and time not done print 'current abstraction still good, time has not run out' #we want to continue but after first time, we use expanded set of engines. funcs = [eval('(pyabc_split.defer(read_and_sleep)())')] funcs = funcs + create_funcs(J,t) #use old J first time mtds = ['read_and_sleep'] + sublist(methods,J) if initial_size == sizeof(): print 'methods = %s'%mtds J = [7,9,19,23,24,30] #first time reflects that 7 and 30 are already being done continue #will try with new vabs def read_and_sleep(t=5): """ keep looking for a new vabs every 5 seconds. This is usually run in parallel with &vta -d """ #t is not used at present tt = time.time() T = 1000 #if after the last abstraction, no answer, then terminate T = abs_time set_size() name = '%s_vabs.aig'%f_name ## if ifbip > 0: ## name = '%s_vabs.aig'%f_name #### print 'name = %s'%name while True: if time.time() - tt > T: #too much time between abstractions print 'read_and_sleep timed out in %d sec.'%T return True if os.access('%s'%name,os.R_OK): abc('r %s'%name) abc('w %s_vabs_old.aig'%f_name) os.remove('%s'%name) if not check_size(): print '\nNew abstraction: ', ps() set_size() abc('w %s_abs.aig'%f_name) return False #if same size, keep going. print '.', sleep(5) ####################################################