Hi,
Recently I submitted a PR https://github.com/dmlc/tvm/pull/4056/files regarding to cuda fp16 codegen issue. I passed the task_python_integration.sh on my own machine with CUDA10. However, it failed at CI integration test for
/tmp/tmpkz8vof0p/my_kernel.cu(4): error: identifier “__hgt” is undefined
/tmp/tmpkz8vof0p/my_kernel.cu(8): error: identifier “__hlt” is undefined
/tmp/tmpkz8vof0p/my_kernel.cu(12): error: no suitable conversion function from “const volatile __half” to “int” exists
/tmp/tmpkz8vof0p/my_kernel.cu(12): error: no suitable conversion function from “const volatile __half” to “int” exists
I noticed we are using cuda8 for CI test, but according to the documentation here, those functions should be available.
Can someone please take a look?