def testBigKernelPerformance( self, bits, framework ):
runs = 10
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub(framework, bits, -1)
a_gpu = f_gpu()
b_gpu = f_gpu()
flags = 0
GF2nStub.run("parAdd", a_gpu, b_gpu, flags, runs)
times = GF2nStub.getEllapsedTime_ms()
for time in times:
PerformanceDataLogger().addPerfResult("parAdd small", bits, framework, time)
flags = 1
chunk_size = 32
if GF2nStub.getRegisterSize() == 64:
chunk_size = 64
for num_grids in [2**n for n in range(0,5)]:
for num_threads in [1024, 512, 256, 128]:
num_blocks = (bits+1)/chunk_size/num_grids/num_threads
GF2nStub.setProperty("bn_a", "num_threads", str(num_threads))
GF2nStub.setProperty("bn_a", "num_blocks", str(num_blocks))
GF2nStub.run("parAddLoop", a_gpu, b_gpu, flags, runs)
times = GF2nStub.getEllapsedTime_ms()
for time in times:
PerformanceDataLogger().addPerfResult("parAdd big " + str(num_threads) + " " + str(num_grids), bits, framework, time)
def testAddPerformance(self, bits, framework, func):
runs = 100
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub(framework, bits, -1)
a_gpu = f_gpu()
b_gpu = f_gpu()
flags = 0
if (func == "parAddOwnStream" or
func == "parAdd2OwnStream" or
func == "parAdd4OwnStream" or
func == "parAdd8OwnStream" or
func == 'parAddOwnStream1024Threads' or
func == 'parAddOwnStream512Threads' or
func == 'parAddOwnStream256Threads' or
func == 'parAddOwnStream128Threads') \
and \
framework == "Cuda":
flags = flags | 2
if framework == "Cuda":
GF2nStub.run(func, a_gpu, b_gpu, flags, runs)
else:
GF2nStub.run("add", a_gpu, b_gpu, flags, runs)
times = GF2nStub.getEllapsedTime_ms()
for time in times:
PerformanceDataLogger().addPerfResult(
func, bits, framework, time)
def testCopyHostToDevicePerformance(self, bits):
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub("Cuda", bits, -1)
a_gpu = f_gpu()
b_gpu = f_gpu()
GF2nStub.run("measureKernelLaunchOverhead", a_gpu, b_gpu)
metrics = GF2nStub.getMetrics("bn_a")
PerformanceDataLogger().addPerfResult("copy host -> device", bits,
"Cuda",
metrics["copyToDevice_time"])
def testMySecondMul(self, bits, func):
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub("Cuda", bits)
a_gpu = f_gpu()
b_gpu = f_gpu()
res_gpu = GF2nStub.run(func, a_gpu, b_gpu)
rand_a = GF2nStub.getRandomNumber(bits, 42)
rand_b = GF2nStub.getRandomNumber(bits, 84)
print "a = ", rand_a
print "b = ", rand_b
f_ref = GF2n.GF2n(bits)
a_ref = f_ref(rand_a)
b_ref = f_ref(rand_b)
res_ref = a_ref * b_ref
print "res_ref = ", res_ref
# compare results
self.assertEqual(res_gpu, res_ref)
def testInvWithExpElementPerformance(self, bits, function, framework):
runs = 1
f = GF2nStub.GF2nStub(framework, bits)
a = f()
if framework == "OpenSSL":
res = GF2nStub.run("inverse", a, 0, 0, runs)
else:
res = GF2nStub.run(function, a, 0, 0, runs)
times = GF2nStub.getEllapsedTime_ms()
for time in times:
PerformanceDataLogger().addPerfResult(function, bits, framework,
time)
def testExpPerformance(self, bits, k, function, framework):
runs = 10
rand_irred_poly = GF2nStub.getRandomNumber(bits + 1, 23)
f = GF2nStub.GF2nStub(framework, bits, rand_irred_poly | 1)
a = f()
if framework == "OpenSSL":
res = GF2nStub.run("exp", a, k, 0, runs)
else:
res = GF2nStub.run(function, a, k, 0, runs)
times = GF2nStub.getEllapsedTime_ms()
for time in times:
PerformanceDataLogger().addPerfResult(function + ", k=" + str(k),
bits, framework, time)
def testFrameworkOverhead(self, bits):
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub("Cuda", bits, -1)
a_gpu = f_gpu()
b_gpu = f_gpu()
# measure test run with python framework and prng
start_stub_run = timeit.default_timer()
GF2nStub.run("measureKernelLaunchOverhead", a_gpu, b_gpu)
PerformanceDataLogger().addPerfResult(
"FrameworkOverhead - All", bits, "Cuda",
timeit.default_timer() - start_stub_run)
# measure test run without python framwork and without prng
times = GF2nStub.getEllapsedTime_ms()
PerformanceDataLogger().addPerfResult("FrameworkOverhead - OnlyFunc",
bits, "Cuda", times[0])
def testCopyDeviceToHostPerformance(self, bits):
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub("Cuda", bits, -1)
a_gpu = f_gpu()
b_gpu = f_gpu()
res = GF2nStub.run("parAdd", a_gpu, b_gpu)
res_value = res._value
metrics = GF2nStub.getMetrics("res")
PerformanceDataLogger().addPerfResult("copy device -> host", bits,
"Cuda",
metrics["copyToHost_time"])
def testOpenSSLExponentiation(self, bits, exp):
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub("OpenSSL", bits)
a_gpu = f_gpu()
res_gpu = GF2nStub.run("exp", a_gpu, exp)
# calcualte reference
rand_a = GF2nStub.getRandomNumber(bits, 42)
f_ref = GF2n.GF2n(bits)
a_ref = f_ref(rand_a)
res_ref = a_ref**exp
# compare results
self.assertEqual(res_gpu, res_ref)
def testOpenSSLInverseElement(self, bits):
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub("OpenSSL", bits)
a_gpu = f_gpu()
res_gpu = GF2nStub.run("inverse", a_gpu, 0)
# calcualte reference
rand_a = GF2nStub.getRandomNumber(bits, 42)
f_ref = GF2n.GF2n(bits)
a_ref = f_ref(rand_a)
res_ref = a_ref.inverse()
# compare results
self.assertEqual(res_gpu, res_ref)
def testCudaExponentiation(self, bits, k, func):
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub("Cuda", bits)
a_gpu = f_gpu()
res_gpu = GF2nStub.run(func, a_gpu, k)
# calcualte reference
rand_a = GF2nStub.getRandomNumber(bits, 42)
f_ref = GF2n.GF2n(bits)
a_ref = f_ref(rand_a)
res_ref = a_ref**k
# compare results
self.assertEqual(res_gpu, res_ref)
def testCudaMultiplication(self, bits, func):
# do cuda arithmetic
f_gpu = GF2nStub.GF2nStub("Cuda", bits)
a_gpu = f_gpu()
b_gpu = f_gpu()
res_gpu = GF2nStub.run(func, a_gpu, b_gpu)
# calcualte reference
rand_a = GF2nStub.getRandomNumber(bits, 42)
rand_b = GF2nStub.getRandomNumber(bits, 84)
f_ref = GF2n.GF2n(bits)
a_ref = f_ref(rand_a)
b_ref = f_ref(rand_b)
res_ref = a_ref * b_ref
# compare results
self.assertEqual(res_gpu, res_ref)