def _op_build_ascend(opnames, computes, args, custom_schedule, device,
kernel_name, attrs):
tmp_rst = op_build_to_func(opnames, computes, args, custom_schedule,
device, kernel_name, attrs)
if tmp_rst is not None:
try:
_api_internal._BuildToModule(tmp_rst, _get_target(device))
except Exception:
logging.error(traceback.format_exc())
return None
return True
def build(inputs, args, target=None, shape_params=None, name="default_function",
binds=None, attrs=None, polyhedral=False, aicpu=False):
tmp_rst = build_to_func(inputs, args, shape_params=shape_params, name=name, binds=binds,
attrs=attrs, polyhedral=polyhedral, aicpu=aicpu)
tmp_target = target if target is not None else 'cce'
return _api_internal._BuildToModule(tmp_rst, tmp_target)
def op_build(opnames, computes, args, custom_schedule, device, kernel_name,
attrs):
"""op_build"""
if device in ("aicore", "aicpu"):
tmp_rst = op_build_to_func(opnames, computes, args, custom_schedule,
device, kernel_name, attrs)
return _api_internal._BuildToModule(tmp_rst)
if device == "cuda":
cuda_path = os.path.realpath(MS_CUDA_KERNEL_PATH)
if not os.path.isdir(cuda_path):
os.makedirs(cuda_path)
if not opnames:
logging.error("no opname given.")
return None
schedule_name = 'gpu_schedule_' + opnames[0]
schedule_func = getattr(akg.ms.gpu, schedule_name)
if not isinstance(schedule_func,
(types.FunctionType, typing.Callable)):
logging.error("no schedule func found %s", str(schedule_name))
return None
ptx_file = os.path.realpath(MS_CUDA_KERNEL_PATH + kernel_name + ".ptx")
if os.path.exists(ptx_file):
os.remove(ptx_file)
try:
with open(ptx_file, 'at') as file:
fcntl.flock(file.fileno(), fcntl.LOCK_EX)
file.seek(0, 2)
if file.tell() == 0:
s = schedule_func(computes)
foo = akg.tvm.build(s, args, device, name=kernel_name)
ptx_code = foo.imported_modules[0].get_source("ptx")
file.write(ptx_code)
json_file = os.path.realpath(MS_CUDA_KERNEL_PATH +
kernel_name + ".json")
kernel_info = (ptx_code, json_file, kernel_name)
gpu_utils.save_gpu_params(s, args, kernel_info)
os.chmod(ptx_file, 0o400)
except Exception:
logging.error(traceback.format_exc())
return None
return True
logging.error("Not support device %s.", device)
return None
def compilewithjson(json_str):
tmp_rst = compilewithjson_to_func(json_str)
if isinstance(tmp_rst, bool):
return tmp_rst
return _api_internal._BuildToModule(tmp_rst)
def _build(desc_s, desc_d, attr=None):
if desc_d['process'] == 'cuda':
func = tvm.get_global_func("composite_with_json")
return func(desc_s, attr)
rst = _build_to_func(desc_s, desc_d, attr)
return _api_internal._BuildToModule(rst)
def _build(desc_s, desc_d, attrs=None, poly=False, use_repo=True):
if desc_d['process'] == 'cuda':
return _build_to_gpu_func(desc_s, desc_d, attrs, poly)
rst = _build_to_func(desc_s, desc_d, attrs, use_repo)
return _api_internal._BuildToModule(rst)