Does TVM support compiling generic code to run on GPUs? For example, if I implement a time-consuming algorithm (which has nothing to do with popular deep learning frameworks), can TVM compile and optimize it to run on GPUs?
Yes, as long as your computation can be described by TVM’s lower level IR (not Relay), we can run it. We have something like sorting and cumsum implemented in TVM.
Is there any doc or tool to help decide which kind of computations can be described by TVM’s lower level IR?