hashencoder.cu error with some specific gpu models

[gukai330]

Thank you for your work!

Had this issue when I am running your code on the server with different GPU models, the script failed when running with some specific GPU models (like tesla P100,), but worked fine with some other GPU models(like tesla v100)

hashencoder.cu(25): error: no instance of overloaded function "atomicAdd" matches the argument list
argument types are: (__half *, c10::Half)

ninja: build stopped: subcommand failed.

cuda version: 11.3.1
cudnn version: 8.2.0.53

[gukai330]

I find the issue in instant ngp #52),
I am trying with the solution

[ashawkey]

Currently --ff only supports GPU with arch >= 70, so P100 (arch == 60) is unable to compile hashencoder and ffmlp.
However, you can still use --tcnn backend in P100, although the speed will be slower compared to V100.

[ashawkey]

I'm not sure if it is related to windows, but this error message does come from low GPU arch.

[ashawkey]

You can check the CUDA programming guide here:
The 16-bit __half floating-point version of atomicAdd() is only supported by devices of compute capability 7.x and higher.

[gukai330]

I'm not sure if it is related to windows, but this error message does come from low GPU arch.

thank you for your explanation, I misunderstand the error message. this should have nothing to do with windows