About the NNPACK libraries


#1

I didn’t use NNPack libraries during build of TVM. when doing a Auto-tuning al network for NVIDIA GPU , does it have any effect on the performance if I didn’t use NNPack. and there is another question about the auto-tuning for NVIDIA GPU,
def tune_tasks(tasks,
measure_option,
tuner=‘xgb’,
n_trial=1000,
early_stopping=None,
log_filename=‘tuning.log’,
use_transfer_learning=True,
try_winograd=True,
try_spatial_pack_depthwise=False):
try_winograd = True, try_spatial_pack_depthwise=False by default, what do these two arguments used for? and try_spatial_pack_depthwise is not used in other place in the code of tune_relay_cuda.py.

another question: when and in what conditions I should choose ‘xgb’ as the tuner ?

Thanks in advance!


#2

try_spatial_pack_depthwise is added by me. But it will only affect ARM CPU. This schedule is better than current depthwise schedule on ARM CPU. If your hardware is CUDA, this argument will not have any effect.


#3

thank you very much! when try_winograd = True, tvm will use NNPack libraries to auto-tune?


#4

No. It only call TVM’s self winograd schedule.