def NonFPModule(function):
instruction_count = function["instruction_count"]
nonidle_elapsed_time = function["nonidle_elapsed_time"]
fp_addsub = asohelper.get_fp_addsub(function)
fp_muldiv = asohelper.get_fp_muldiv(function)
cpiBase = function["cpiBase"]
cpiMem = function["cpiMem"]
cpiBranchPredictor = function["cpiBranchPredictor"]
cpiOther = nonidle_elapsed_time - cpiBase - cpiMem - cpiBranchPredictor
fp_instructions = fp_addsub + fp_muldiv
non_fp_instructions = instruction_count - fp_instructions
non_fp_fraction = non_fp_instructions / instruction_count #the fraction of non-fp instructions
if instruction_count > 0:
non_fp_time = non_fp_fraction*(cpiBase + cpiMem + cpiBranchPredictor + cpiOther)
else:
non_fp_time = 0
optimizedfunction = copy.deepcopy(function)
if non_fp_time > 0:
#changing properties of function
optimizedfunction["instruction_count"] -=non_fp_instructions
optimizedfunction["time_won_back"] +=non_fp_time
optimizedfunction["nonidle_elapsed_time"] -=non_fp_time
optimizedfunction["cpiBase"] -=non_fp_fraction*cpiBase
optimizedfunction["cpiMem"] -=non_fp_fraction*cpiMem
optimizedfunction["cpiBranchPredictor"] -=non_fp_fraction*cpiBranchPredictor
optimizedfunction["optimizations"].append(dict(optimization="NonFP", timegain=non_fp_time))
optimizedfunction["originalfunction"] = function
return optimizedfunction
def writerooflinestats(outputfile):
output = {}
output["rooflinedata"]=[]
for data in functiondata:
x=0
y=0
fpinstr = (asohelper.get_fp_addsub(data)+asohelper.get_fp_muldiv(data))
if (data["l3miss"]) > 0:
x = float((fpinstr/data["l3miss"])/64) #per byte, so division by 64
if (data["nonidle_elapsed_time"]) > 0:
y = float(fpinstr/data["nonidle_elapsed_time"]*1e6) #GFLOPS
output["rooflinedata"].append([x,y])
functioninfo = copy.deepcopy(functiondata)
for function in functioninfo:
function["nonidle_elapsed_time"]/=1e6
function["cpi"]=getCPI(function)
output["functioninfo"]=functioninfo
output["functionpercentages"]=functionpercentages
output["peakfpperformance"]=getPeakFPPerformance()
output["peakmembandwidth"]=getPeakMemBandwidth()
f = open(outputfile, "w")
f.write(json.dumps(output,indent=4))
f.close()
def VectorizationModule(function):
optimizedfunction = copy.deepcopy(function)
instruction_count = function["instruction_count"]
nonidle_elapsed_time = function["nonidle_elapsed_time"]
fp_addsub = asohelper.get_fp_addsub(function)
fp_muldiv = asohelper.get_fp_muldiv(function)
cpiBase = function["cpiBase"]
cpiMem = function["cpiMem"]
cpiBranchPredictor = function["cpiBranchPredictor"]
cpiOther = nonidle_elapsed_time - cpiBase - cpiMem - cpiBranchPredictor
fp_instructions = fp_addsub + fp_muldiv
non_fp_instructions = instruction_count - fp_instructions
#add instructions
addpd = function["addpd"] #packed double precision
addsd = function["addsd"] #double precision
addss = function["addss"] #single precision
addps = function["addps"] #packed single precision
#sub instructions
subpd = function["subpd"] #packed double precision
subsd = function["subsd"] #double precision
subss = function["subss"] #single precision
subps = function["subps"] #packed single precision
#mul instructions
mulpd = function["mulpd"] #packed double precision
mulsd = function["mulsd"] #double precision
mulss = function["mulss"] #single precision
mulps = function["mulps"] #packed single precision
#div instructions
divpd = function["divpd"] #packed double precision
divsd = function["divsd"] #double precision
divss = function["divss"] #single precision
divps = function["divps"] #packed single precision
#check if number of instructions is correct
assert instruction_count == non_fp_instructions + addpd + addsd + addss + subpd + subsd + subss + subps + mulpd + mulsd + mulss + mulps + divpd + divsd + divss + divps
if instruction_count > 0:
#if we only have addsub or muldiv operations
if fp_addsub == 0 or fp_muldiv == 0:
n_ports = 1 #only one issue port can be used per fp-instruction
else:
n_ports = 2 #the two issue ports can be used simultaneously
newtime = 0
#already vectorized and double precision
#cpibase part used by packed double precision:
cpiBasePD = (
(addpd + subpd + mulpd + divpd) / instruction_count) * cpiBase
newtime += cpiBasePD / n_ports
#vectorization: SD -> PD
#cpibase part used by non-vectorized double precision
cpiBaseSD = (
(addsd + subsd + mulsd + divsd) / instruction_count) * cpiBase
newtime += cpiBaseSD / (2 * n_ports)
#already vectorized and single precision
#cpibase part used by packed single precision:
cpiBasePS = (
(addps + subps + mulps + divps) / instruction_count) * cpiBase
newtime += cpiBasePS / n_ports
#vectorization: SS -> PS
#cpibase part used by non-vectorized single precision
cpiBaseSS = (
(addss + subss + mulss + divss) / instruction_count) * cpiBase
newtime += cpiBaseSS / (4 * n_ports)
#these non_fp instructions can't be vectorized
#cpibase part used by non-floating point instructions:
cpiBaseNonFP = (non_fp_instructions / instruction_count) * cpiBase
newtime += cpiBaseNonFP
new_cpiBase = newtime
#add the branch, mem and other component
newtime += cpiMem + cpiBranchPredictor + cpiOther
#the sum of the cpiBase components should be the total cpiBase component
total = cpiBasePD + cpiBaseSD + cpiBasePS + cpiBaseSS + cpiBaseNonFP
assert round(cpiBase) == round(total), "%s != %s" % (
round(cpiBase),
round(total),
)
time_gain = nonidle_elapsed_time - newtime
if time_gain > 0:
#changing properties of function
optimizedfunction["time_won_back"] += time_gain
optimizedfunction["nonidle_elapsed_time"] -= time_gain
optimizedfunction["cpiBase"] = new_cpiBase
optimizedfunction["optimizations"].append(
dict(optimization="Vectorization", timegain=time_gain))
optimizedfunction["originalfunction"] = function
assert time_gain >= 0, "we should not gain negative time"
if fp_instructions == 0:
#if we don't have any floating point instruction, the time_gain should be zero
assert time_gain == 0, "%s != %s" % (
time_gain,
0,
)
return optimizedfunction
def VectorizationModule(function):
optimizedfunction = copy.deepcopy(function)
instruction_count = function["instruction_count"]
nonidle_elapsed_time = function["nonidle_elapsed_time"]
fp_addsub = asohelper.get_fp_addsub(function)
fp_muldiv = asohelper.get_fp_muldiv(function)
cpiBase = function["cpiBase"]
cpiMem = function["cpiMem"]
cpiBranchPredictor = function["cpiBranchPredictor"]
cpiOther = nonidle_elapsed_time - cpiBase - cpiMem - cpiBranchPredictor
fp_instructions = fp_addsub + fp_muldiv
non_fp_instructions = instruction_count - fp_instructions
#add instructions
addpd = function["addpd"] #packed double precision
addsd = function["addsd"] #double precision
addss = function["addss"] #single precision
addps = function["addps"] #packed single precision
#sub instructions
subpd = function["subpd"] #packed double precision
subsd = function["subsd"] #double precision
subss = function["subss"] #single precision
subps = function["subps"] #packed single precision
#mul instructions
mulpd = function["mulpd"] #packed double precision
mulsd = function["mulsd"] #double precision
mulss = function["mulss"] #single precision
mulps = function["mulps"] #packed single precision
#div instructions
divpd = function["divpd"] #packed double precision
divsd = function["divsd"] #double precision
divss = function["divss"] #single precision
divps = function["divps"] #packed single precision
#check if number of instructions is correct
assert instruction_count == non_fp_instructions+addpd+addsd+addss+subpd+subsd+subss+subps+mulpd+mulsd+mulss+mulps+divpd+divsd+divss+divps
if instruction_count > 0:
#if we only have addsub or muldiv operations
if fp_addsub == 0 or fp_muldiv == 0:
n_ports = 1 #only one issue port can be used per fp-instruction
else:
n_ports = 2 #the two issue ports can be used simultaneously
newtime = 0
#already vectorized and double precision
#cpibase part used by packed double precision:
cpiBasePD = ((addpd + subpd + mulpd + divpd) / instruction_count) *cpiBase
newtime += cpiBasePD/n_ports
#vectorization: SD -> PD
#cpibase part used by non-vectorized double precision
cpiBaseSD = ((addsd + subsd + mulsd + divsd) / instruction_count) *cpiBase
newtime += cpiBaseSD / (2*n_ports)
#already vectorized and single precision
#cpibase part used by packed single precision:
cpiBasePS = ((addps + subps + mulps + divps) / instruction_count) *cpiBase
newtime += cpiBasePS / n_ports
#vectorization: SS -> PS
#cpibase part used by non-vectorized single precision
cpiBaseSS = ((addss + subss + mulss + divss) / instruction_count) *cpiBase
newtime += cpiBaseSS / (4*n_ports)
#these non_fp instructions can't be vectorized
#cpibase part used by non-floating point instructions:
cpiBaseNonFP = (non_fp_instructions / instruction_count) *cpiBase
newtime += cpiBaseNonFP
new_cpiBase = newtime
#add the branch, mem and other component
newtime += cpiMem + cpiBranchPredictor + cpiOther
#the sum of the cpiBase components should be the total cpiBase component
total = cpiBasePD + cpiBaseSD + cpiBasePS + cpiBaseSS + cpiBaseNonFP
assert round(cpiBase) == round(total), "%s != %s" % (round(cpiBase), round(total),)
time_gain = nonidle_elapsed_time - newtime
if time_gain > 0:
#changing properties of function
optimizedfunction["time_won_back"] += time_gain
optimizedfunction["nonidle_elapsed_time"] -= time_gain
optimizedfunction["cpiBase"] = new_cpiBase
optimizedfunction["optimizations"].append(dict(optimization="Vectorization", timegain=time_gain))
optimizedfunction["originalfunction"] = function
assert time_gain >= 0, "we should not gain negative time"
if fp_instructions == 0:
#if we don't have any floating point instruction, the time_gain should be zero
assert time_gain == 0, "%s != %s" % (time_gain, 0,)
return optimizedfunction
