Tvm split code for device

Questions
1

Hi, I want to add a execution mode, the host is cpu and device is riscv. There are some questions about tvm spliting code for host and device:

1)when I set

target = "cuda"
target_host = "llvm"

it generated the proper code, but when I give

target = "llvm --system-lib -mtriple=riscv64-unknown-linux-gnu -mcpu=rocket-rv64"
target_host = "llvm"

it didn’t give device code. So why this happens, I try to find the answer from the source code, but it is really complicated, I read the SplitHostDevice pass code, but I can not find where it decides to generate device code, and the conditons of generating device code.

2)Does the device code just execute loop computation ? What should I do to finish my job ?

Thanks for any reply!

2

IIUC， it should be

if (op->attr_key == attr::thread_extent | op->attr_key == attr::pipeline_exec_scope | op->attr_key == attr::device_scope)

{

return SplitDeviceFunc(GetRef(op));

}

So, try to figure our if such attributes exist in you code.