def get_result(desc, poly, attrs=None, profiling=True, need_compare=True):
backend = _get_backend(desc)
mod = composite.build(desc, attrs, poly=poly)
if not need_compare:
return True
input_for_mod, expect, output_indexes = gen_json_data(desc)
output = utils.mod_launch(mod, input_for_mod, output_indexes)
# In profiling mode, mod_launch will return compute outputs and profiling value, only compute outputs needed here
if isinstance(output, tuple) and len(output) > 0 and isinstance(
output[-1], dict):
output = output[0]
output = output if isinstance(output, (list, tuple)) else [output]
expect = expect if isinstance(expect, (list, tuple)) else [expect]
output = list(output)
expect = list(expect)
for i, _ in enumerate(expect):
if expect[i].dtype == "complex128" or expect[i].dtype == "complex64":
final_shape = functools.reduce(lambda x, y: x * y, output[i].shape)
flattern_output = output[i].reshape((final_shape, ))
output_real = []
output_imag = []
for k, _ in enumerate(flattern_output):
if k % 2 == 0:
output_real.append(flattern_output[k])
else:
output_imag.append(flattern_output[k])
output[i] = np.vectorize(complex)(output_real, output_imag)
output[i] = output[i].reshape(expect[i].shape)
if len(output) != len(expect):
raise RuntimeError(
"output and expect have different length, {} vs {}".format(
len(output), len(expect)))
compare_tolerance = get_compare_tolerance(desc, output_indexes)
compare_res = list(map(_compare_func, output, expect, compare_tolerance))
if not all(compare_res):
source = (mod.imported_modules[0]
if backend == "cuda" else mod).get_source()
logging.debug(source)
_dump_info(desc, attrs, poly, input_for_mod, output, expect)
logging.warning("Compare results: %s", str(compare_res))
return False
if profiling and backend in ["cuda", "cpu"]:
ctx = tvm.context(backend, 0)
has_complex = False
for i in input_for_mod:
if i.dtype == "complex64" or i.dtype == "complex128":
has_complex = True
break
if has_complex == False:
inputs = to_tvm_nd_array(input_for_mod, ctx)
target_profiling(mod, *inputs, target=backend, repeat_time=1000)
return True
def target_profiling(mod, *args, target="cuda", repeat_time=1, device_id=0, need_warm_up=True):
"""Do profiling on gpu/cpu for op"""
ctx = tvm.context(target, device_id)
if target == "llvm":
ftimer = mod.time_evaluator(mod.entry_name, ctx, number=repeat_time,
repeat=3, min_repeat_ms=1000)
if need_warm_up:
mod.time_evaluator(mod.entry_name, ctx, number=1000)(*args)
else:
ftimer = mod.time_evaluator(mod.entry_name, ctx, number=repeat_time)
tcost = ftimer(*args).mean
print("{}: exec={} ms/op".format(ctx, tcost * 1000))
return tcost
def gpu_schedule_Mul(outs):
"""
gpu schedule for mul.
Args:
outs (tvm.tensor.Tensor): outputs of compute.
Returns:
sch (schedule.Schedule): The created schedule.
"""
device = 'cuda'
ctx = tvm.context(device, 0)
if not ctx.exist:
raise SystemError("Skip because %s is not enabled" % device)
with tvm.target.create(device):
sch = topi.cuda.schedule_broadcast(outs)
return sch
def default_schedule(outs):
"""
default schedule function.
Args:
outs (Union[tvm.tensor.Tensor, list[tvm.tensor.Tensor]]): outputs of compute.
Returns:
sch (schedule.Schedule): The created schedule.
"""
if not isinstance(outs, tvm.tensor.Tensor) and not isinstance(outs, list):
raise ValueError(
"outs should be list of akg.tvm.tensor.Tensor or akg.tvm.tensor.Tensor"
)
device = 'cuda'
ctx = tvm.context(device, 0)
if not ctx.exist:
raise SystemError("Skip because %s is not enabled" % device)
outs_list = [outs] if isinstance(outs, tvm.tensor.Tensor) else outs
with tvm.target.create(device):
sch = tvm.create_schedule(outs_list[0].op)
outputs_tensor = Queue()
outputs_tensor.put(outs_list[0])
op_list = []
while not outputs_tensor.empty():
out = outputs_tensor.get()
if out.op not in op_list and isinstance(out.op,
tvm.tensor.ComputeOp):
op_list.append(out.op)
for input_tensor in out.op.input_tensors:
outputs_tensor.put(input_tensor)
for op in op_list:
stage = sch[op.output(0)]
bx, tx = stage.split(op.axis[0], factor=DEFAULT_GPU_THREAD)
stage.bind(bx, tvm.thread_axis("blockIdx.x"))
stage.bind(tx, tvm.thread_axis("threadIdx.x"))
return sch
def gpu_profiling(mod, *args, repeat_time=1, device_id=0):
"""Do profiling on gpu for cuda op"""
ctx = tvm.context("cuda", device_id)
ftimer = mod.time_evaluator(mod.entry_name, ctx, number=repeat_time)
tcost = ftimer(*args).mean
print("{}: exec={} sec/op".format(ctx, tcost))