cuDLALayerwiseStatsStandalone --std=c++11 whole ./ ../ ../../../Common cuDLA Data Parallel Algorithms Image Processing CUDA CPP11 cudla nvscibuf nvscisync true main.cpp NVSCI 1:CUDA Advanced Topics 1:cuDLA sm60 sm61 sm70 sm72 sm75 sm80 sm86 sm87 sm89 sm90 aarch64 linux aarch64 qnx 6.0 cuDLA Layerwise Statistics Standalone Mode exe