def test_add(self):
"""
This seems to be about the amount of work to balance out the overhead
by the overloaded one
"""
cr = compiler.compile_isolated(add, [types.int32])
cfunc = cr.entry_point
disp = registry.CPUOverloaded(add)
disp.add_overload(cr)
x = 321
def python():
add(x)
def pycfunc():
cfunc(x)
def overloaded():
disp(x)
print(add)
print(utils.benchmark(python, maxsec=.5))
print(utils.benchmark(pycfunc, maxsec=.5))
print(utils.benchmark(overloaded, maxsec=.5))
def test_use_car_move(self):
tyctx = typing.Context()
tyctx.insert_class(Car, self.carattrs)
cgctx = CPUContext(tyctx)
cgctx.insert_class(Car, self.carattrs)
car_object = types.Object(Car)
argtys = (car_object, types.int32)
flags = compiler.Flags()
cr = compiler.compile_extra(tyctx, cgctx, use_car_move, args=argtys,
return_type=None, flags=flags, locals={})
func = cr.entry_point
if cr.typing_error:
raise cr.typing_error
car1 = Car(value=123)
car2 = Car(value=123)
self.assertEqual(use_car_move(car1, 321), func(car2, 321))
def bm_python():
use_car_move(car1, 321)
def bm_numba():
func(car2, 321)
python = utils.benchmark(bm_python, maxsec=.1)
numba = utils.benchmark(bm_numba, maxsec=.1)
print(python)
print(numba)
def test_array_overhead(self):
"""
The time to set two array element seems to be more expensive than
the overhead of the overloaded call.
"""
cr = compiler.compile_isolated(array_overhead, [types.int32[::1]])
cfunc = cr.entry_point
disp = registry.CPUOverloaded(array_overhead)
disp.add_overload(cr)
self.assertEqual(cr.signature.args[0].layout, 'C')
x = numpy.zeros(shape=2, dtype='int32')
def python():
array_overhead(x)
def pycfunc():
cfunc(x)
def overloaded():
disp(x)
print(array_overhead)
print(utils.benchmark(python, maxsec=.5))
print(utils.benchmark(pycfunc, maxsec=.5))
print(utils.benchmark(overloaded, maxsec=.5))
def test_array_overhead(self):
"""
The time to set two array element seems to be more expensive than
the overhead of the overloaded call.
"""
cr = compiler.compile_isolated(array_overhead, [types.int32[::1]])
cfunc = cr.entry_point
disp = registry.CPUDispatcher(array_overhead)
disp.add_overload(cr)
self.assertEqual(cr.signature.args[0].layout, 'C')
x = np.zeros(shape=2, dtype='int32')
def python():
array_overhead(x)
def pycfunc():
cfunc(x)
def overloaded():
disp(x)
print(array_overhead)
print(utils.benchmark(python, maxsec=.5))
print(utils.benchmark(pycfunc, maxsec=.5))
print(utils.benchmark(overloaded, maxsec=.5))
def test_add(self):
"""
This seems to be about the amount of work to balance out the overhead
by the overloaded one
"""
cr = compiler.compile_isolated(add, [types.int32])
cfunc = cr.entry_point
disp = registry.CPUDispatcher(add)
disp.add_overload(cr)
x = 321
def python():
add(x)
def pycfunc():
cfunc(x)
def overloaded():
disp(x)
print(add)
print(utils.benchmark(python, maxsec=.5))
print(utils.benchmark(pycfunc, maxsec=.5))
print(utils.benchmark(overloaded, maxsec=.5))
def run(mod):
name = mod.__name__[len(BENCHMARK_PREFIX):]
print('running', name, end=' ...\n')
bmr = benchmark(mod.python_main)
python_best = bmr.best
print('\tpython', python_best, 'seconds')
bmr = benchmark(mod.numba_main)
numba_best = bmr.best
print('\tnumba', numba_best, 'seconds')
print('\tspeedup', python_best / numba_best)
return name, numba_best / python_best
def run(mod):
name = mod.__name__[len(BENCHMARK_PREFIX):]
print('running', name, end=' ...\n')
bmr = benchmark(mod.python_main)
python_best = bmr.best
print('\tpython', python_best, 'seconds')
bmr = benchmark(mod.numba_main)
numba_best = bmr.best
print('\tnumba', numba_best, 'seconds')
print('\tspeedup', python_best / numba_best)
return name, numba_best / python_best
def run(mod):
name = mod.__name__[len(BENCHMARK_PREFIX) :]
print("running", name, end=" ...\n")
bmr = benchmark(mod.python_main)
python_best = bmr.best
print("\tpython", python_best, "seconds")
bmr = benchmark(mod.numba_main)
numba_best = bmr.best
print("\tnumba", numba_best, "seconds")
print("\tspeedup", python_best / numba_best)
return name, numba_best / python_best
def test_loop_nest(self):
"""
Test bug that decref the iterator early.
If the bug occurs, a segfault should occur
"""
pyfunc = loop_nest_3
cres = compile_isolated(pyfunc, (), flags=forceobj)
cfunc = cres.entry_point
self.assertEqual(pyfunc(5, 5), cfunc(5, 5))
def bm_pyfunc():
pyfunc(5, 5)
def bm_cfunc():
cfunc(5, 5)
print(utils.benchmark(bm_pyfunc))
print(utils.benchmark(bm_cfunc))
def test_sum1d_pyobj(self):
pyfunc = usecases.sum1d
cr = compile_isolated(pyfunc, (types.int32, types.int32),
flags=force_pyobj_flags)
cfunc = cr.entry_point
ss = -1, 0, 1, 100, 200
es = -1, 0, 1, 100, 200
for args in itertools.product(ss, es):
self.assertEqual(pyfunc(*args), cfunc(*args), args)
args = 0, 500
def bm_python():
pyfunc(*args)
def bm_numba():
cfunc(*args)
print(utils.benchmark(bm_python, maxsec=.1))
print(utils.benchmark(bm_numba, maxsec=.1))
def test_sum1d_pyobj(self):
pyfunc = usecases.sum1d
cr = compile_isolated(pyfunc, (types.int32, types.int32),
flags=force_pyobj_flags)
cfunc = cr.entry_point
ss = -1, 0, 1, 100, 200
es = -1, 0, 1, 100, 200
for args in itertools.product(ss, es):
self.assertEqual(pyfunc(*args), cfunc(*args), args)
args = 0, 500
def bm_python():
pyfunc(*args)
def bm_numba():
cfunc(*args)
print(utils.benchmark(bm_python, maxsec=.1))
print(utils.benchmark(bm_numba, maxsec=.1))
def test_binary_ufunc_performance(self):
pyfunc = _make_binary_ufunc_usecase('add')
arraytype = types.Array(types.float32, 1, 'C')
cr = compile_isolated(pyfunc, (arraytype, arraytype, arraytype))
cfunc = cr.entry_point
nelem = 5000
x_operand = np.arange(nelem, dtype=np.float32)
y_operand = np.arange(nelem, dtype=np.float32)
control = np.empty_like(x_operand)
result = np.empty_like(x_operand)
def bm_python():
pyfunc(x_operand, y_operand, control)
def bm_numba():
cfunc(x_operand, y_operand, result)
print(utils.benchmark(bm_python, maxsec=.1))
print(utils.benchmark(bm_numba, maxsec=.1))
assert np.allclose(control, result)
def test_binary_ufunc_performance(self):
pyfunc = _make_binary_ufunc_usecase('add')
arraytype = types.Array(types.float32, 1, 'C')
cr = compile_isolated(pyfunc, (arraytype, arraytype, arraytype))
cfunc = cr.entry_point
nelem = 5000
x_operand = np.arange(nelem, dtype=np.float32)
y_operand = np.arange(nelem, dtype=np.float32)
control = np.empty_like(x_operand)
result = np.empty_like(x_operand)
def bm_python():
pyfunc(x_operand, y_operand, control)
def bm_numba():
cfunc(x_operand, y_operand, result)
print(utils.benchmark(bm_python, maxsec=.1))
print(utils.benchmark(bm_numba, maxsec=.1))
assert np.allclose(control, result)
def test_overhead(self):
"""
This will show higher overhead due to unboxing in the native version.
"""
cr = compiler.compile_isolated(overhead, [types.int32])
cfunc = cr.entry_point
disp = registry.CPUDispatcher(overhead)
disp.add_overload(cr)
x = 321
def python():
overhead(x)
def pycfunc():
cfunc(x)
def overloaded():
disp(x)
print(overhead)
print(utils.benchmark(python, maxsec=.5))
print(utils.benchmark(pycfunc, maxsec=.5))
print(utils.benchmark(overloaded, maxsec=.5))
def benchmark_refct_speed():
def pyfunc(x, y, t):
"""Swap array x and y for t number of times
"""
for i in range(t):
x, y = y, x
return x, y
cfunc = nrtjit(pyfunc)
x = np.random.random(100)
y = np.random.random(100)
t = 10000
def bench_pyfunc():
pyfunc(x, y, t)
def bench_cfunc():
cfunc(x, y, t)
python_time = utils.benchmark(bench_pyfunc)
numba_time = utils.benchmark(bench_cfunc)
print(python_time)
print(numba_time)
def test_overhead(self):
"""
This will show higher overhead due to unboxing in the native version.
"""
cr = compiler.compile_isolated(overhead, [types.int32])
cfunc = cr.entry_point
disp = registry.CPUOverloaded(overhead)
disp.add_overload(cr)
x = 321
def python():
overhead(x)
def pycfunc():
cfunc(x)
def overloaded():
disp(x)
print(overhead)
print(utils.benchmark(python, maxsec=.5))
print(utils.benchmark(pycfunc, maxsec=.5))
print(utils.benchmark(overloaded, maxsec=.5))
def benchmark_refct_speed():
def pyfunc(x, y, t):
"""Swap array x and y for t number of times
"""
for i in range(t):
x, y = y, x
return x, y
cfunc = nrtjit(pyfunc)
x = np.random.random(100)
y = np.random.random(100)
t = 10000
def bench_pyfunc():
pyfunc(x, y, t)
def bench_cfunc():
cfunc(x, y, t)
python_time = utils.benchmark(bench_pyfunc)
numba_time = utils.benchmark(bench_cfunc)
print(python_time)
print(numba_time)
tol = 1.0e-6
error = 1.0
for j in range(n):
A[j, 0] = 1.0
Anew[j, 0] = 1.0
it = 0
while error > tol and it < iter_max:
error = fn(A, Anew)
# swap A and Anew
tmp = A
A = Anew
Anew = tmp
it += 1
def python_main():
run(jocabi_relax_core)
def numba_main():
run(numba_jocabi_relax_core)
if __name__ == '__main__':
print(benchmark(python_main))
print(benchmark(numba_main))
count += 2
return count
@jit("intp()", nopython=True)
def euler():
triangle = 1
index = 1
while factorCount(triangle) < 1001:
index += 1
triangle += index
return triangle
answer = 842161320
def numba_main():
result = euler()
assert result == answer
def python_main():
result = py_euler()
assert result == answer
if __name__ == '__main__':
print(benchmark(python_main))
print(benchmark(numba_main))
