def time(ATLdir, ARCH, pre, blas, N, M, lda, rout, cc=None, ccf=None, opt=""):
l1blas = kernels.GetFKOlevel1Blas()
l2blas = kernels.GetLevel2Blas()
l3blas = kernels.GetLevel3Blas()
if blas in l1blas:
return l1cmnd.time(ATLdir, ARCH, pre, blas, N, rout, cc, ccf, opt)
elif blas in l2blas:
return l2cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, rout, cc, ccf, opt)
elif blas in l3blas:
return l3cmnd.time(ATLdir, ARCH, pre, blas, N, 1, rout, cc, ccf, opt)
elif blas in kernels.GetAllFKOGEMM():
return l3cmnd.time(ATLdir, ARCH, pre, blas, N, 1, rout, cc, ccf, opt)
else:
print "UNKNOWN KERNEL [FOR CMND] !!! "
return [-1, -1]
def time(ATLdir, ARCH, pre, blas, N, M, lda, rout, cc=None, ccf=None, opt=""):
l1blas = kernels.GetFKOlevel1Blas()
l2blas = kernels.GetLevel2Blas()
l3blas = kernels.GetLevel3Blas()
if blas in l1blas:
return l1cmnd.time(ATLdir, ARCH, pre, blas, N, rout, cc, ccf, opt)
elif blas in l2blas:
return l2cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, rout, cc, ccf,
opt)
elif blas in l3blas:
return l3cmnd.time(ATLdir, ARCH, pre, blas, N, 1, rout, cc, ccf, opt)
elif blas in kernels.GetAllFKOGEMM():
return l3cmnd.time(ATLdir, ARCH, pre, blas, N, 1, rout, cc, ccf, opt)
else:
print "UNKNOWN KERNEL [FOR CMND] !!! "
return [-1, -1]
ICC = 0
CALLREF=1
CALLATL=1
CALLFKO=1
CALLBLA=1
PROFILE=0
#PFLAGS = "-xP -O3 -mp1 -static"
#PFLAGS = "-xW -O3 -mp1 -static -Qoption,ld,-melf_i386 -Qoption,asm,--32"
PFLAGS = "-xW -O3 -mp1 -static"
# print 'l1atl = ', l1atl
for blas in l1routs:
for pre in pres:
if (CALLBLA != 0):
[time,mf] = l1cmnd.time(ATLdir, ARCH, pre, blas, N, l1cblas[i],
opt=opt)
assert(time > 0.0)
print "CBL %20.20s : time=%f, mflop=%f" % (pre+l1cblas[i], time, mf)
blaT.append(time)
blaMF.append(mf)
if (CALLREF != 0):
[time,mf] = l1cmnd.time(ATLdir, ARCH, pre, blas, N, l1refs[i], opt=opt)
assert(time > 0.0)
print "REF %20.20s : time=%f, mflop=%f" % (pre+l1refs[i], time, mf)
refT.append(time)
refMF.append(mf)
if ICC and PROFILE:
[time,mf] = l1cmnd.time(ATLdir, ARCH, pre, blas, N, l1refs[i],
opt=opt, cc="iccprof", ccf=PFLAGS)
CALLATL = 1
CALLFKO = 1
CALLBLA = 1
PROFILE = 0
#PFLAGS = "-xP -O3 -mp1 -static"
#PFLAGS = "-xW -O3 -mp1 -static -Qoption,ld,-melf_i386 -Qoption,asm,--32"
PFLAGS = "-xW -O3 -mp1 -static"
# print 'l1atl = ', l1atl
for blas in l1routs:
for pre in pres:
if (CALLBLA != 0):
[time, mf] = l1cmnd.time(ATLdir,
ARCH,
pre,
blas,
N,
l1cblas[i],
opt=opt)
assert (time > 0.0)
print "CBL %20.20s : time=%f, mflop=%f" % (pre + l1cblas[i], time,
mf)
blaT.append(time)
blaMF.append(mf)
if (CALLREF != 0):
[time, mf] = l1cmnd.time(ATLdir,
ARCH,
pre,
blas,
N,
nargs = len(sys.argv)
if nargs < 4:
print 'USAGE: %s <pre> <blas> <file> [<N> <uopt> <TEST> <CC> <CCF>]' % sys.argv[0]
sys.exit(1)
pre = sys.argv[1]
blas = sys.argv[2]
file = sys.argv[3]
if nargs > 4 :
N = int(sys.argv[4])
if nargs > 5 :
TEST = int(sys.argv[5])
if nargs > 6 :
uopt = sys.argv[6]
if nargs > 7 :
CC = sys.argv[7]
if nargs > 8 :
CCF = sys.argv[8]
#opt = "-X 1 -Y 1 " + uopt
#opt = "-X 1 -Y 1 -Fx 16 -Fy 16 " + uopt
opt = "-X 1 -Y 1 -Fx 32 -Fy 32 " + uopt
[t0,mf] = l1cmnd.time(ATLdir, ARCH, pre, blas, N, file, cc=CC, ccf=CCF, opt=opt)
if TEST:
i = l1cmnd.test(ATLdir, ARCH, pre, blas, N, file, cc=CC, ccf=CCF, opt=teopt)
if i : PF = "FAIL"
else : PF = "PASS"
else : PF = "SKIP"
print "%s%s-%s : time=%e, mflop=%.2f -- %s" % (pre, blas, file, t0, mf, PF)
