def decorated(*args, **kwargs):
impl.get_runtime().inside_complex_kernel = True
if impl.get_runtime().target_tape:
impl.get_runtime().target_tape.insert(decorated, args)
try:
func(*args, **kwargs)
finally:
impl.get_runtime().inside_complex_kernel = False
def func__(*args, **kwargs):
finishes = False
try:
func(*args, **kwargs)
finishes = True
except ex:
# throws. test passed
pass
except Exception as err_actual:
assert False, 'Exception {} instead of {} thrown'.format(
str(type(err_actual)), str(ex))
if finishes:
assert False, 'Test successfully finished instead of throwing {}'.format(
str(ex))
def run_benchmark():
compile_time = time.time()
func(*args) # compile the kernel first
ti.sync()
compile_time = time.time() - compile_time
ti.stat_write('compilation_time', compile_time)
codegen_stat = _ti_core.stat()
for line in codegen_stat.split('\n'):
try:
a, b = line.strip().split(':')
except:
continue
a = a.strip()
b = int(float(b))
if a == 'codegen_kernel_statements':
ti.stat_write('compiled_inst', b)
if a == 'codegen_offloaded_tasks':
ti.stat_write('compiled_tasks', b)
elif a == 'launched_tasks':
ti.stat_write('launched_tasks', b)
# Use 3 initial iterations to warm up
# instruction/data caches. Discussion:
# https://github.com/taichi-dev/taichi/pull/1002#discussion_r426312136
for i in range(3):
func(*args)
ti.sync()
ti.kernel_profiler_clear()
t = time.time()
for n in range(repeat):
func(*args)
ti.sync()
elapsed = time.time() - t
avg = elapsed / repeat
ti.stat_write('wall_clk_t', avg)
device_time = ti.kernel_profiler_total_time()
avg_device_time = device_time / repeat
ti.stat_write('exec_t', avg_device_time)
def decorated(*args, **kwargs):
func(*args, **kwargs)
def test(*args, **kwargs):
archs = [_ti_core.host_arch()]
for arch in archs:
ti.init(arch=arch)
func(*args, **kwargs)