def test_scan_int32(self):
in_ary = np.random.randint(0, 10000, ARRAY_SIZE).astype(np.int32)
in_ary_d = cuda.to_device(in_ary)
output = MyScan.scan_gpu(in_ary_d)
out_carry = output.getitem(0)
out_ary = in_ary_d.copy_to_host()
# check last carry
assert np.isclose(out_carry, in_ary.sum()), 'carry return is not sum'
# check array
carry = 0
for i in xrange(ARRAY_SIZE):
assert out_ary[i] == carry, 'output array not correct'
carry += in_ary[i]
def test_scan_fp64(self):
in_ary = np.random.rand(ARRAY_SIZE).astype(np.float64)
in_ary_d = cuda.to_device(in_ary)
output = MyScan.scan_gpu(in_ary_d)
out_carry = output.getitem(0)
out_ary = in_ary_d.copy_to_host()
# check last carry
print out_carry, in_ary.sum()
# self.assertTrue(np.isclose(out_carry, in_ary.sum()),
# 'carry return is not sum')
# check array
carry = 0
for i in xrange(ARRAY_SIZE):
self.assertEqual(out_ary[i], carry, 'output array not correct')
carry += in_ary[i]
def test_recursive_big_scan():
print "running recursive scan test"
MAX_TPB = 512
n = 2e6
n = int(n)
a = np.arange(n).astype(np.int32)
reference = np.empty_like(a)
start = timer()
sum_ref = MyScan.exprefixsumNumba(a, reference, init=0)
end = timer()
dA = cuda.to_device(a)
# e1, e2 = cuda.event(), cuda.event()
# e1.record()
# e2.record()
start2 = timer()
total_sum = MyScan.scan_gpu(dA)
end2 = timer()
dA.copy_to_host(ary=a)
sum_gpu = total_sum.copy_to_host()
print "sum_ref = ", sum_ref
print "sum_gpu = ", sum_gpu
print "CPU took: ", (end - start) * 1000, " ms"
print "Kernel took: ", (end2 - start2) * 1000, " ms"
print (a == reference).all()