No, the kernel works well when the dataset is small, say ,1k. when the input dataset is 1M, it works wrong. but indeed there’s nothing related with the dataset??
//delete
Hi, I meet invalide arguments error in kernel. If I comment some random control flow blocks, e.g. a for loop, a device subfunction, then the kernel passes.
I want to ask: is there a limitation for max control flows? Thanks!