def ifko_writeNT(ATLdir, ARCH, KF0, fko, rout, pre, l1bla, N, M, lda, wnt):
#
# Time the default case
#
warrs = []
fkocmnd.callfko(fko, KF0)
[t0,m0] = cmnd.time(ATLdir, ARCH, pre, l1bla, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print "WNT none : %.2f" % (m0)
for wa in wnt:
KFN = KF0 + " -W " + wa
fkocmnd.callfko(fko, KFN)
[tN,mN] = cmnd.time(ATLdir, ARCH, pre, l1bla, N, M,lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print "WNT %s : %2.f" % (wa, mN)
if mN > m0:
KF0 = KFN
m0 = mN
warrs.append(wa)
return [m0,KF0,warrs]
if (CALLATL != 0):
[time,mf] = l1cmnd.time(ATLdir, ARCH, pre, blas, N, l1atl[j],
CCatl[j], CCFat[j], opt=opt)
assert(time > 0.0)
print "ATL %20.20s : time=%f, mflop=%f" % (pre+l1atl[j], time, mf)
atlT.append(time)
atlMF.append(mf)
if (CALLFKO != 0):
rout = IFKOdir + '/blas/' + pre + blas + '.b'
outf = ATLdir + '/tune/blas/level1/' + blas.upper() + '/fkorout.s'
KF0 = fkocmnd.GetStandardFlags(fko, rout, pre)
KFLAGS = KF0 + ' -o ' + outf + " " + rout
if (os.path.exists(outf)):
os.remove(outf)
fkocmnd.callfko(fko, KFLAGS)
[time,mf] = l1cmnd.time(ATLdir, ARCH, pre, blas, N, 'fkorout.s',
"gcc", "-x assembler-with-cpp", opt=opt)
assert(time > 0.0)
print "FKO %20.20s : time=%f, mflop=%f" % (pre+blas+'.b', time, mf)
print " flags =", KF0
fkoT.append(time)
fkoMF.append(mf)
j += 1
i += 1
print r"OPERATION & gcc+ref & icc+ref & icc+prof & gcc+atlas&icc+atlas& cblas & fko & ifko\\\hline\hline"
form = "%12s& %5.0f & %5.0f & %5.0f & %5.0f & %5.0f & %5.0f & %5.0f & \\\\\\hline"
form2= "%12s& %5.0f & %5.0f & %5.0f & %5.0f*& %5.0f*& %5.0f & %5.0f & \\\\\\hline"
#fkoMF = [1.1, 2.2, 3.3, 4.4, 5.5, 6.6, 7.7, 8.8, 9.9, 10.1];
CCFat[j],
opt=opt)
assert (time > 0.0)
print "ATL %20.20s : time=%f, mflop=%f" % (pre + l1atl[j], time,
mf)
atlT.append(time)
atlMF.append(mf)
if (CALLFKO != 0):
rout = IFKOdir + '/blas/' + pre + blas + '.b'
outf = ATLdir + '/tune/blas/level1/' + blas.upper() + '/fkorout.s'
KF0 = fkocmnd.GetStandardFlags(fko, rout, pre)
KFLAGS = KF0 + ' -o ' + outf + " " + rout
if (os.path.exists(outf)):
os.remove(outf)
fkocmnd.callfko(fko, KFLAGS)
[time, mf] = l1cmnd.time(ATLdir,
ARCH,
pre,
blas,
N,
'fkorout.s',
"gcc",
"-x assembler-with-cpp",
opt=opt)
assert (time > 0.0)
print "FKO %20.20s : time=%f, mflop=%f" % (pre + blas + '.b', time,
mf)
print " flags =", KF0
fkoT.append(time)
fkoMF.append(mf)
def ifko0(l1bla, pre, N, M=None, lda=None):
(IFKOdir, fko) = fkocmnd.GetFKOinfo()
(ATLdir, ARCH) = fkocmnd.FindAtlas(IFKOdir)
rout = IFKOdir + '/blas/' + pre + l1bla + '.b'
#outf = ATLdir + '/tune/blas/level1/' + l1bla.upper() + '/fkorout.s'
outf = ATLdir + kernels.GetBlasPath(l1bla) + '/fkorout.s'
#
# Majedul: calling new info func, info represents the old list
# new data: [npath, red2onePath, vecMethod, vpathinfo, arrtypes] at the end
#
#info = fkocmnd.info(fko, rout)
newinfo = fkocmnd.NewInfo(fko, rout)
info = [newinfo[i] for i in range(11) ]
[npath, red1path, vecm, vpath, arrtypes] = [ newinfo[i] for i in range(11,16)]
ncache = info[0]
vec = info[5]
#(fparrs, fpsets, fpuses, fpurs) = fkocmnd.GetFPInfo(info)
(fparrs, fpsets, fpuses, fpurs) = fkocmnd.GetFPInfo(newinfo)
nfp = len(fparrs)
#
# Findout the default flags (it includes vector, default prefetch and unroll)
#
#KFLAGS = fkocmnd.GetStandardFlags(fko, rout, pre)
KFLAGS = fkocmnd.GetOptStdFlags(fko, rout, pre, 1, URF)
print "\n Default Flag = " + KFLAGS
KFLAGS = KFLAGS + " -o " + str(outf) + " " + rout
mflist = []
testlist = []
#print KFLAGS
#
# Majedul: default and vect case would not be same now. Vspec may be
# worse than NonVec case.
# So, I will choose the best as the default for the later optimization
#
#
# check best scalar xforms, delete any vector flag
#
#j = KFLAGS.find("-V")
#if j != -1 :
#KFn = KFLAGS[0:j-1] + KFLAGS[j+2:]
#
# find out best standard scalar flag
#
KFn = fkocmnd.GetOptStdFlags(fko, rout, pre, 0, URF)
KFn = KFn + " -o " + str(outf) + " " + rout
#print KFn
#
# standard flag without vect
#
KF0 = KFn
fkocmnd.callfko(fko, KF0)
[t0,mf0] = cmnd.time(ATLdir, ARCH, pre, l1bla, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
mflist.append(mf0)
testlist.append("default") ## this is using std flags
print "\n Default Flag = " + KF0
#
# Finding the best path reduction option
#
if npath > 1:
[mfs, KFs] = ifko_PathXform(ATLdir, ARCH, KFn, ncache, fko, rout, pre,
l1bla, N, M, lda, npath, red1path)
mflist.append(mfs)
testlist.append("PathXform")
if (mfs > mf0) :
mf0 = mfs
KF0 = KFs
#
# Finding the best vector option with/without path reduction
#
global isSV;
if SB:
KFv = fkocmnd.GetOptStdFlags(fko, rout, pre, 1, SB, URF)
else:
KFv = fkocmnd.GetOptStdFlags(fko, rout, pre, 1, 0, URF)
print "\n Standad Flag for Vect = " + KFv
KFv = KFv + " -o " + str(outf) + " " + rout
if vec:
if 'v' in skipOpt:
print '\n SKIPPING VECTORIZATION'
else:
[mfv, KFv] = ifko_Vec(ATLdir, ARCH, KFv, ncache, fko, rout, pre,
l1bla, N, M, lda, npath, vecm, vpath)
mflist.append(mfv)
testlist.append("vect")
if (mfv > mf0) :
mf0 = mfv
KF0 = KFv
#
# if we have forceOpt, we will keep vec even if it's not better
#
elif 'sv' in forceOpt or 'vrc' in forceOpt or 'vmmr' in forceOpt:
print '\n FORCING VECTORIZATION'
mf0 = mfv
KF0 = KFv
else: # no vector is selected, skip the SB too #
isSV = 0
#
# choose the better as the ref of later opt
#
KFLAGS = KF0
mf = mf0
print "\n FLAGS so far =", fkocmnd.RemoveFilesFromFlags(l1bla, KFLAGS)
#
# Previous code which is substituted by the above codes
#
#
# Find performance of default case
#
# j = KFLAGS.find("-V")
# if j != -1 :
# KFn = KFLAGS[0:j-1] + KFLAGS[j+2:]
# fkocmnd.callfko(fko, KFn)
# [t,mf] = l1cmnd.time(ATLdir, ARCH, pre, l1bla, N, "fkorout.s",
# "gcc", "-x assembler-with-cpp", opt=opt)
# mflist.append(mf)
# testlist.append("default")
# fkocmnd.callfko(fko, KFLAGS)
# [t,mf] = l1cmnd.time(ATLdir, ARCH, pre, l1bla, N, "fkorout.s",
# "gcc", "-x assembler-with-cpp", opt=opt)
# else :
# fkocmnd.callfko(fko, KFLAGS)
# [t,mf] = l1cmnd.time(ATLdir, ARCH, pre, l1bla, N, "fkorout.s",
# "gcc", "-x assembler-with-cpp", opt=opt)
# testlist.append("default")
# mflist.append(mf)
# mflist.append(mf)
# testlist.append("vect")
#
# Eventually, want to try both -V and scalar, but for now, use -V whenever
# possible
#
# Find if we want to use cache-through writes on any arrays
#
if 'wnt' in skipOpt:
print '\n SKIPPING WNT'
else:
n = len(fpsets)
i = 0
wnt = []
while i < n:
if fpsets[i] > 0 :
# and fpuses[i] == 0:
wnt.append(fparrs[i])
i += 1
if len(wnt) > 0:
[mf,KFLAGS,wnt] = ifko_writeNT(ATLdir, ARCH, KFLAGS, fko, rout, pre,
l1bla, N, M, lda, wnt)
mflist.append(mf)
testlist.append("writeNT")
#
# Find best PFD for each pfarr
#
pfarrs = fparrs
pfsets = fpsets
for arr in pfarrs:
[mf,KFLAGS] = FindPFD(ATLdir, ARCH, KFLAGS, fko, rout, pre,l1bla, N,M,lda,
info, arr)
mflist.append(mf)
testlist.append("pfdist")
KFLAGS = fkocmnd.RemoveRedundantPrefFlags(KFLAGS, pfarrs)
#
# Find best pf type
#
[mf,KFLAGS] = ifko_pftype(ATLdir, ARCH, KFLAGS, ncache, fko, rout, pre,
l1bla, N, M, lda, info, pfarrs, pfsets)
mflist.append(mf)
testlist.append("pftype")
print "\n FLAGS so far =", fkocmnd.RemoveFilesFromFlags(l1bla, KFLAGS)
#
# Find best unroll
#
if URF:
print '\n SKIPPING UNROLL TUNNING : FORCED TO %d' %URF
else:
[mf,KFLAGS] = FindUR(ATLdir, ARCH, KFLAGS, fko, rout, pre, l1bla, N, M,
lda, info)
mflist.append(mf)
testlist.append("unroll")
#
# Find best bet for over speculation
# FIXME: find out the -U and pass it to the function
# FIXME: can't apply Over Spec if there is a memory write inside the loop
#
if isSV:
if l1bla.find("irk1amax") != -1:
print '\n SKIPPING STRONGER BET UNROLLING for IRK1AMAX'
elif l1bla.find("irk2amax") != -1:
print '\n SKIPPING STRONGER BET UNROLLING for IRK2AMAX'
elif l1bla.find("irk3amax") != -1:
print '\n SKIPPING STRONGER BET UNROLLING for IRK3AMAX'
elif l1bla.find("sin") != -1:
print '\n SKIPPING STRONGER BET UNROLLING for SIN'
elif l1bla.find("cos") != -1:
print '\n SKIPPING STRONGER BET UNROLLING for COS'
else:
[mf,KFLAGS] = FindBET(ATLdir, ARCH, KFLAGS, fko, rout, pre, l1bla, N,
M, lda)
mflist.append(mf)
testlist.append("OverSpec")
#
# See if we can apply accumulator expansion
#
# acc = fkocmnd.GetFPAccum(info)
# nacc = len(acc)
# if nacc > 0 and nacc < 3:
# [mf,KFLAGS] = FindAE(ATLdir, ARCH, KFLAGS, fko, rout, pre, l1bla, N, acc)
# mflist.append(mf)
# testlist.append("accexpans")
#
# Majedul: See if we can apply scalar expansion (accexpan + man/min expansion)
#
acc = fkocmnd.GetFPAccum(info)
nacc = len(acc)
if 're' in skipOpt:
print '\n SKIPPING SCALAR EXPANSION'
elif isSV:
print '\n SKIPPING SCALAR EXPANSION: NOT SUPPORTED WITH SV'
elif l1bla.find("iamax") != -1:
print '\n SKIPPING SCALAR EXPANSION FOR IAMAX'
else:
if nacc > 0 and nacc < 3:
[mf,KFLAGS] = FindRE(ATLdir, ARCH, KFLAGS, fko, rout, pre, l1bla, N,
M, lda, acc)
mflist.append(mf)
testlist.append("rdexp")
#
# Majedul: shifted it here to test
#
#
# Find if we want to use cache-through writes on any arrays
#
"""if 'wnt' in skipOpt:
print '\n SKIPPING WNT'
else:
n = len(fpsets)
i = 0
wnt = []
while i < n:
if fpsets[i] > 0 :
# and fpuses[i] == 0:
wnt.append(fparrs[i])
i += 1
if len(wnt) > 0:
[mf,KFLAGS,wnt] = ifko_writeNT(ATLdir, ARCH, KFLAGS, fko, rout, pre,
l1bla, N, wnt)
mflist.append(mf)
testlist.append("writeNT")
#
# Find best PFD for each pfarr
#
pfarrs = fparrs
pfsets = fpsets
for arr in pfarrs:
[mf,KFLAGS] = FindPFD(ATLdir, ARCH, KFLAGS, fko, rout, pre,l1bla, N, M,
lda, info, arr)
mflist.append(mf)
testlist.append("pfdist")
KFLAGS = fkocmnd.RemoveRedundantPrefFlags(KFLAGS, pfarrs)
#
# Find best pf type
#
[mf,KFLAGS] = ifko_pftype(ATLdir, ARCH, KFLAGS, ncache, fko, rout, pre,
l1bla, N, info, pfarrs, pfsets)
mflist.append(mf)
testlist.append("pftype")
print "\n FLAGS so far =", fkocmnd.RemoveFilesFromFlags(l1bla, KFLAGS)
"""
#
# tesing: re-tune the prefetch distance!
# NOTE: this re-tuning can be omitted just by enabling the comment
#
#"""
#KFLAGS = fkocmnd.SetDefaultPFD(KFLAGS, info)
KFLAGS = fkocmnd.SetDefaultPFD(KFLAGS, newinfo)
#print "default PFD: ", KFLAGS
print "\n TUNING PFD AGAIN: "
for arr in pfarrs:
[mf,KFLAGS] = FindPFD(ATLdir, ARCH, KFLAGS, fko, rout, pre,l1bla, N, M,
lda, info, arr)
KFLAGS = fkocmnd.RemoveRedundantPrefFlags(KFLAGS, pfarrs)
#
# FIXME: it will create problem for the calculaton of % of improvement
#
# if 'pfdist' in testlist:
# j = testlist.index('pfdist')
# mflist[j] = mf
# else:
# mflist.append(mf)
# testlist.append("pfdist")
#KFLAGS = fkocmnd.RemoveRedundantPrefFlags(KFLAGS, pfarrs)
mflist.append(mf)
testlist.append("pfd2")
#"""
#
# Find performance of best case
#
# fkocmnd.callfko(fko, KFLAGS)
# [t,mf] = l1cmnd.time(ATLdir, ARCH, pre, l1bla, N, "fkorout.s",
# "gcc", "-x assembler-with-cpp", opt=opt)
print "\n\n BEST FLAGS FOUND (%.2f) = %s" % (mf,
fkocmnd.RemoveFilesFromFlags(l1bla, KFLAGS))
res = fkocmnd.GetOptVals(KFLAGS, pfarrs, pfsets, acc)
tst = cmnd.test(ATLdir, ARCH, pre, l1bla, N, M, lda, "fkorout.s",
cc="gcc", ccf="-x assembler-with-cpp", opt=optT)
#tst = l1cmnd.silent_test(ATLdir, ARCH, pre, l1bla, N, "fkorout.s",
# cc="gcc", ccf="-x assembler-with-cpp", opt=optT)
return(res, KFLAGS, mf, tst, testlist, mflist)
def FindBET(ATLdir, ARCH, KF0, fko, rout, pre, blas, N, M, lda):
#
# figure out the UR factor from flag
#
j = KF0.find('-U')
if j != -1:
words = KF0[j:].split()
UR0 = int(words[1])
else:
UR0 = 1
#
# set max bet as the UR but not less than 4,
# but not greater than 10
#
#maxbet = UR0 * 2;
maxbet = UR0
if maxbet < 4:
maxbet = 4
if maxbet > 10:
maxbet = 10
#
# reduce tries for sine and cosine
#
if blas.find("sin") != -1:
URm = 8;
elif blas.find("cos") != -1:
URm = 8;
else:
URm = 16
#
# time default case
#
fkocmnd.callfko(fko, KF0)
[t0,m0] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
KFn = KF0
KF1 = KF0
mf1 = m0
UR = 1
SB = 1
print " Finding Over Speculation Factor, UR=%d, mflop= %.2f" % (UR0, m0)
for i in range (2, maxbet):
#
# remove previous bet if exists
#
match = re.search(r'-B\s\d+',KFn) ## should not put space \s at last!
if match:
pat = match.group()
KFn = KFn.replace(pat, '') ## remove all bet
KFn = '-B %d ' %i + KFn
#
# SB * UR should not be greater than 32
#
URN = int (URm/i)
for j in range (1, URN+1):
#
# remove prev unroll factor
#
match = re.search(r'-U\s\d+',KFn) ## should not put space \s at last!
if match:
pat = match.group()
KFn = KFn.replace(pat, '') ## remove all bet
KFn = '-U %d ' %j + KFn
fkocmnd.callfko(fko, KFn)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " SB=%d, UR=%d, mflop= %.2f" % (i, j, mf)
if mf > mf1*1.001:
mf1 = mf
KF1 = KFn
SB = i
UR = j
if mf1 > m0:
KFn = KF1
print " Over Speculation : SB=%d, UR=%d, mf=%2.f, KFn=%s" % (SB, UR,
mf1, KFn)
else:
print " No Over Speculation Selected!\n"
KFn = KF0
mf1 = m0
return [mf1, KFn]
def ifko_Vec(ATLdir, ARCH, KF0, ncache, fko, rout, pre, blas, N, M, lda,
npath, vm, vpath):
""" This function will try all vectorization methods incorporating other
"""
#
# add vector flag if not in the flags
#
if KF0.find('-V') == -1:
KF0 = ' -V' + KF0
#
# flag for speculation
#
global isSV
#
# find out the standard vec flags
#
#if SB:
# KF0 = GetOptStdFlags(fko, rout, pre, 1, SB);
#else:
# KF0 = GetOptStdFlags(fko, rout, pre, 1);
#print KF0
#
# delete all flags which are associated with -p/rc/mmr
#
j = KF0.find('-p')
if j != -1:
match = re.search(r'-p\s\d+\s',KF0) ## single instance
if match:
rem = match.group()
KF0 = KF0.replace(rem, '') ## removing the -p val
j = KF0.find('-rc')
if j != -1:
KF0 = KF0.replace('-rc','')
j = KF0.find('-mmr')
if j != -1:
KF0 = KF0.replace('-mmr','')
print "\n Finding best vectorization:"
#
# check with other xforms
# new vec info = maxmin, redcom, spec
#
m0 = 0
t0 = 0
KFn = KF0
if npath > 1:
#if vm[0] or vm[1] or vm[2]:
if vm[0] :
if 'mmr' in skipOpt or 'vrc' in forceOpt or 'sv' in forceOpt:
print ' SKIPPING MMR+V'
else:
KF1 = ' -mmr' + KF0
fkocmnd.callfko(fko, KF1)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " MMR+V, mflop = %.2f" % ( mf)
if mf > m0:
m0 = mf
t0 = t
KFn = KF1
#if vm[3] :
if vm[1] :
if 'rc' in skipOpt or 'vmmr' in forceOpt or 'sv' in forceOpt:
print ' SKIPPING RC+V'
else:
KF1 = ' -rc' + KF0
fkocmnd.callfko(fko, KF1)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " RC+V, mflop = %.2f" % ( mf)
if mf > m0:
m0 = mf
t0 = t
KFn = KF1
#
# check for speculative
#
#if vm[4] :
if vm[2] :
if 'sv' in skipOpt or 'vmmr' in forceOpt or 'vrc' in forceOpt:
print ' SKIPPING SPECULATIVE VECTORIZATION'
else:
for i in range(npath):
if vpath[i] :
KF1 = ' -p %d' %(i+1) + KF0
if SB:
KF1 = ' -B %d ' %(SB) + KF1
fkocmnd.callfko(fko, KF1)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
if SB:
print " V+SP%d SB=%d, mflop = %.2f" % (i+1, SB, mf)
else:
print " V+SP%d, mflop = %.2f" % (i+1, mf)
if mf > m0:
m0 = mf
t0 = t
KFn = KF1
isSV = 1 ## SV is superior, so is applied from now
else: ## no path to reduce
fkocmnd.callfko(fko, KF0)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " V, mflop = %.2f" % (mf)
if mf > m0:
m0 = mf
t0 = t
KFn = KF0
KF0 = KFn
return [m0,KFn]
def ifko_pftype(ATLdir, ARCH, KF0, ncache, fko, rout, pre, blas, N, M, lda,
info, pfarrs, pfsets):
#
# Time the default case
#
fkocmnd.callfko(fko, KF0)
[t0,m0] = cmnd.time(ATLdir, ARCH, pre, blas, N, M,lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
assert(t0 > 0.0)
print " base mf = %f.2" % m0
KF = KF0
#
# If there are output arrays, varying inst used for writes
#
SETS = 0
for i in pfsets:
if i:
SETS = 1
break
if SETS:
#
# Try using 3DNow's prefetchw for all writes
#
KF = KF0 + " -Paw 3"
fkocmnd.callfko(fko, KF)
[t1,m1] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
if (m1 > 0.0):
print " prefetchw speedup: %.3f !" % (m1/m0)
#
# Try temporal prefetch
#
KF1 = KF0 + " -Paw 0"
fkocmnd.callfko(fko, KF1)
[t2,m2] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " prefetchw speedup: %.3f !" % (m2/m0)
if m2 > m1:
m1 = m2
KF = KF1
if m1 > m0:
m0 = m1
KF0 = KF
#
# If we've got read-only arrays, try a varying pref inst for reads
#
READO = 0
for i in pfsets:
if not i:
READO = 1
break
if READO:
KF = KF0 + " -Par 0"
fkocmnd.callfko(fko, KF)
[t1,m1] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " prefetchr speedup: %.3f !" % (m1/m0)
if m1 > m0:
KF0 = KF
m0 = m1
#
# For each array, find best cache level to fetch to
#
NC = info[0]
LS = info[1]
for arr in pfarrs:
j = 1
while j < NC:
KF = " -P %s %d %d" % (arr, j, LS[j])
KF = KF0 + KF
fkocmnd.callfko(fko, KF)
[t1,m1] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " lvl %d %s speedup: %.3f !" % (j, arr, m1/m0)
if m1 > m0:
KF0 = KF
m0 = m1
j += 1
return [m0,KF0]
def ifko_PathXform(ATLdir, ARCH, KF0, ncache, fko, rout, pre, blas, N, M, lda,
npath, red1p):
""" This function will apply some basic transformation on scalar code
like: fall-thru xform, rc, mmr etc
should only call this when npath > 1
"""
#
# checkout the default performance
#
#print KF0
fkocmnd.callfko(fko, KF0)
[t0,m0] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
#
# must have multiple paths
#
if npath == 1 :
print 'Single path! No Xform can be applied!!'
sys.exit(1);
#
# apply fall-thru optimization and see which fall-thru path is better
#
KFn = KF0
print "\n Finding best fall-thru path:"
for i in range(npath):
KF1 = ' -p %d' %(i+1) + KF0
fkocmnd.callfko(fko, KF1)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " Path = %d, mflop = %.2f" % (i+1, mf)
if mf > m0:
m0 = mf
t0 = t
KFn = KF1
#
# apply various reduction methods
# [mmr, maxr, minr, rc] => [mmr, rc]
# newinfo = [maxr, minr, rc] => [mmr, rc]
#
#if red1p[0] or red1p[1] or red1p[2] or red1p[3]:
if red1p[0] or red1p[1] or red1p[2]:
print "\n Finding best path reduction:"
#if red1p[0] or red1p[1] or red1p[2] :
if red1p[0] or red1p[1] :
if 'mmr' in skipOpt:
print ' SKIPPING MMR'
else:
KF1 = ' -mmr' + KF0
fkocmnd.callfko(fko, KF1)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " MMR, mflop = %.2f" % ( mf)
if mf > m0:
m0 = mf
t0 = t
KFn = KF1
#if red1p[3] :
if red1p[2] :
#
# skipping rc for cos kernel, there is a problem in cos!
#
#if 'rc' in skipOpt or blas.find("cos") != -1 :
if 'rc' in skipOpt != -1 :
print ' SKIPPING RC'
else:
KF1 = ' -rc' + KF0
fkocmnd.callfko(fko, KF1)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " RC, mflop = %.2f" % ( mf)
if mf > m0:
m0 = mf
t0 = t
KFn = KF1
KF0 = KFn
return [m0, KF0]
def FindAE(ATLdir, ARCH, KF0, fko, rout, pre, blas, N, M, lda, acc, maxlen=6):
#
# Time the default case
#
fkocmnd.callfko(fko, KF0)
[t0,m0] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
mf0 = m0
#
# Find present unrolling factor, and remove it from flags
#
j = KF0.find("-U ")
assert(j != -1)
words = KF0[j:].split()
ur = int(words[1])
if j > 0: KFN = KF0[0:j-1]
else : KFN = ""
for word in words[2:] :
KFN = KFN + " " + word
KF0 = KFN
print " Finding AccumExpan, UR=%d, mflop= %.2f" % (ur, m0)
mfB = 0.0
urB = ur
aeB = 1
nacc = len(acc)
for ac in acc:
i = 2
while i <= maxlen:
if ur >= i :
KFLAG = KFN + " -U %d -AE %s %d" % (ur, ac, i)
fkocmnd.callfko(fko, KFLAG)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " '%s' AE=%d, UR=%d, mflop= %.2f" % (ac, i, ur, mf)
if mf > mfB:
mfB = mf
urB = ur
aeB = i
if i < ur or ur%i :
j = ((ur+i-1) / i)*i
if j != ur:
KFLAG = KFN + " -U %d -AE %s %d" % (j, ac, i)
fkocmnd.callfko(fko, KFLAG)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " '%s' AE=%d, UR=%d, mflop= %.2f" % (ac, i, j, mf)
if mf > mfB:
mfB = mf
urB = j
aeB = i
j = (ur / i) * i
if j and j != ur:
KFLAG = KFN + " -U %d -AE %s %d" % (j, ac, i)
fkocmnd.callfko(fko, KFLAG)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " '%s' AE=%d, UR=%d, mflop= %.2f" % (ac, i, j, mf)
if mf > mfB:
mfB = mf
urB = j
aeB = i
i += 1
if mfB > mf0*1.001:
KFN = KFN + " -U %d -AE %s %d" % (urB, ac, aeB)
mf0 = mfB
print " AE=%d, UR=%d, mfB=%2.f, KFN=%s" % (aeB, urB, mfB, KFN)
return[mfB, KFN]
def FindRE(ATLdir, ARCH, KF0, fko, rout, pre, blas, N, M, lda, acc, maxlen=6):
""" Exactly same as FindAE but the fko command is changed
"""
#
# Time the default case
# FIXED: need to save the default flag also. If RE is no better than previous
# one, restore the default flag
#
fkocmnd.callfko(fko, KF0)
[t0,m0] = cmnd.time(ATLdir, ARCH, pre, blas, N, M,lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
mf0 = m0
#
# Find present unrolling factor, and remove it from flags
#
j = KF0.find("-U ")
assert(j != -1)
words = KF0[j:].split()
ur = int(words[1])
if j > 0: KFN = KF0[0:j-1]
else : KFN = ""
for word in words[2:] :
KFN = KFN + " " + word
#KF0 = KFN ## don't! lost the previous -U
print " Finding ReduceExpan, UR=%d, mflop= %.2f" % (ur, m0)
mfB = 0.0
urB = ur
aeB = 1
nacc = len(acc)
for ac in acc:
i = 2
while i <= maxlen:
if ur >= i :
KFLAG = KFN + " -U %d -RE %s %d" % (ur, ac, i)
fkocmnd.callfko(fko, KFLAG)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " '%s' RE=%d, UR=%d, mflop= %.2f" % (ac, i, ur, mf)
if mf > mfB:
mfB = mf
urB = ur
aeB = i
if i < ur or ur%i :
j = ((ur+i-1) / i)*i
if j != ur:
KFLAG = KFN + " -U %d -RE %s %d" % (j, ac, i)
fkocmnd.callfko(fko, KFLAG)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " '%s' RE=%d, UR=%d, mflop= %.2f" % (ac, i, j, mf)
if mf > mfB:
mfB = mf
urB = j
aeB = i
j = (ur / i) * i
if j and j != ur:
KFLAG = KFN + " -U %d -RE %s %d" % (j, ac, i)
fkocmnd.callfko(fko, KFLAG)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N,M,lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " '%s' RE=%d, UR=%d, mflop= %.2f" % (ac, i, j, mf)
if mf > mfB:
mfB = mf
urB = j
aeB = i
i += 1
if mfB > mf0*1.001:
KFN = KFN + " -U %d -RE %s %d" % (urB, ac, aeB)
mf0 = mfB
#
# check for altimate result
#
#print mfB, KFN
#print mf0, KF0
if mfB > m0*1.001:
print " RE=%d, UR=%d, mfB=%2.f, KFN=%s" % (aeB, urB, mfB, KFN)
else:
KFN = KF0 ## restore original flags
mfB = m0
print " KFN = KF0 = %s" % KFN
return[mfB, KFN]
def FindUR(ATLdir, ARCH, KF0, fko, rout, pre, blas, N, M, lda, info, UR0=1, URN=64):
print " Finding best unroll:"
#
# Get rid of default unrolling so we can add our own
#
# default max unroll, URN is 64
# but for now, we consider 32 for cos
#
if blas.find("cos") != -1 :
URN = 32;
#
# if speculation is applied, max_unroll is applied to 16
#
if isSV:
if blas.find("sin") != -1 :
URN = 8;
elif blas.find("cos") != -1 :
URN = 5; # FIXME: bitvec exceeds the datatype limit
else:
URN = 16
#
# if SB is specified, SB*UR can be MaxUnroll
#
if SB:
URN = int (URN+SB-1)/SB
# MAX_UR = 32
# if blas.find("amax") != -1 :
# if SB:
# URN = int((MAX_UR+SB-1)/SB) ### just testing!!
# else:
# URN = MAX_UR
# if blas.find("iamax") != -1 :
# if SB:
# URN = int((MAX_UR+SB-1)/SB)
# else:
# URN = MAX_UR
# if blas.find("nrm2") != -1 : ## limit the blind unrolling for nrm2...
# if SB:
# URN = int((MAX_UR+SB-1)/SB)
# else:
# URN = MAX_UR
j = KF0.find("-U ")
if j != -1:
words = KF0[j:].split()
if (j != 0): KF0 = KF0[0:j]
else : KF0 = ""
for word in words[2:]:
KF0 = KF0 + " " + word
UR = UR0
mf0 = 0
URB = 1
while UR <= URN:
KFn = KF0 + " -U %d" % (UR)
fkocmnd.callfko(fko, KFn)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M,lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " UR=%d, mflop=%.2f" % (UR, mf)
#
# Demand that higher unrollings get at least 1% better
#
if (mf >= mf0*1.01):
URB = UR
mf0 = mf
#
# for SV, we need to increament by 1
#
if isSV:
UR += 1
else:
UR *= 2
#UR += 1
KF0 = KF0 + " -U %d" % URB
print "\n BEST Unroll Factor = %d" %URB
return [mf0,KF0]
def FindPFD(ATLdir, ARCH, KF0, fko, rout, pre, blas, N, M, lda, info, arr,
pfd0=0, pfdN=2048, pfdinc=0):
#
# Figure out the prefetch cache level
#
st = "-P %s " % (arr)
j = KF0.find(st)
if j == -1:
j = KF0.find("-P all ")
assert(j != -1)
j += 7
else:
j += 4 + len(arr)
words = KF0[j:].split()
pflvl = int(words[0])
LS = info[1][pflvl];
if not pfd0: pfd0 = LS
if not pfdinc: pfdinc = LS
print "\n Finding PFD for %s in [%d:%d:%d]" % (arr, pfd0, pfdN, pfdinc)
ipd = int(words[1])
mfM = 0.0
pfdM = 0
#
# Majedul: delete previous default pref sch
#
# match = re.search(r'-Ps\s\w\s\w\s\d+\s\d+',KF0) ## find -Ps b A 0 1
# if match:
# pat = match.group()
# KF0 = KF0.replace(pat, '') ## remove the all occurance
# KF0 = KF0 + " -Ps b A 0 1"
#
# will delete the -P all 0 128
#
# match = re.search(r'-P\s\w+\s[-]?\d+\s\d+',KF0) ## find -P X 0 1
# if match:
# pat = match.group()
# KF0 = KF0.replace(pat, '') ## remove all occurance
#
# Scope very short PFD
#
if pfd0 >= LS and LS > 32:
pfd = 32
while pfd <= LS:
KFn = KF0 + " -P %s %d %d" % (arr, pflvl, pfd)
fkocmnd.callfko(fko, KFn)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " %s : PFD = %d mflop = %.2f" % (arr, pfd, mf)
if mf > mfM*1.0001:
mfM = mf
pfdM = pfd
pfd += 8
pfd = pfd0
while pfd <= pfdN:
KFn = KF0 + " -P %s %d %d" % (arr, pflvl, pfd)
fkocmnd.callfko(fko, KFn)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
# mfs.append(mf)
print " %s : PFD = %d mflop = %.2f" % (arr, pfd, mf)
#
# Demand higher PFD is at least 1.0001 faster
#
if mf > mfM*1.0001:
mfM = mf
pfdM = pfd
pfd += LS
print "\n BEST prefetch distance = %d (%.2f)" % (pfdM, mfM)
#
# Try not prefetching array at all
#
KFn = KF0 + " -P %s -1 0" % (arr)
# print KFn
fkocmnd.callfko(fko, KFn)
[t,mf] = cmnd.time(ATLdir, ARCH, pre, blas, N, M, lda, "fkorout.s",
"gcc", "-x assembler-with-cpp", opt=opt)
print " %s : NO PREFETCH: mflop = %.2f" % (arr, mf)
if mf >= mfM:
KF0 = KF0 + " -P %s -1 0" % (arr)
mfM = mf
elif pfdM != ipd:
KF0 = KF0 + " -P %s %d %d" % (arr, pflvl, pfdM)
return [mfM,KF0]