Not implemented for type Half

[Stonesjtu]

I followed the tutorial to create my cpp_extension. But the module failed to do half-precision computation.

It gives:

RuntimeError: "op" not implemented for 'Half' (operator() at /op_cuda_kernel.cu:146)
frame #0: c10::Error::Error(c10::SourceLocation, std::string const&) + 0x47 (0x7fa561676687 in /miniconda3/lib/python3.7/site-packages/torch/lib/libc10.so)

When submitting a bug report, please include the following information (where relevant):

• OS: Centos
• PyTorch version: 1.3.0
• How you installed PyTorch (conda, pip, source): conda
• Python version: 3.7
• CUDA/cuDNN version: 10/7
• GPU models and configuration: 1080
• GCC version (if compiling from source):7.4.0

[soumith]

cc: @smessmer @ezyang

[ezyang]

Can you post your code? EDIT: Oh are you doing the tutorial directly?

[Stonesjtu]

Yes, I was following the tutorial.
I found that using AT_DISPATCH_FLOATING_TYPES_AND_HALF MACRO should do the magic to support half scalar type, but it is not documented in the tutorial.
Should we add some notes about the different dispatch macros?

[ezyang]

cc @gchanan but our dispatch macros are kind of a mess and not really equipped to handle user code (also, it's not always right to use the _AND_HALF macro, so there's some subtlety on when to use things too.)