comaniac opened a new pull request #10122:
URL: https://github.com/apache/tvm/pull/10122
#6225 reported that some math functions (e.g., `pow` and `tanh`) are not
supported in `cuda_fp16.h` and result in nvcc error due to undefined reference.
The solution in #6225 is hard coded these two functions as a patch. However,
other functions such as `erf`, `tan`, `atan` also have this issue, as reported
by #6349. This PR adds the support of these functions with some refactors.
Meanwhile, the unit test failed to cache this issue because of the incorrect
logic. This PR also fixes them.
1. The unary unit test skips all unary ops with float16 on CUDA GPU instead
of just Vulkan.
2. The unary unit test never creates float16 Relay graph. Note that
`relay.var("x", TensorType(shape), dtype="float16")` will create a *float32*
var, because when a complete TensorType is given as a type annotation, the
dtype parameter will be ignored (we should error out in this case...)
cc @vinx13 @Lunderberg @junrushao1994
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
