Half function undefined during compilation

[5had3z]

Description

I've implemented a plugin for grid sample 2D from the pytorch repo and used it within my own codebase. However when I try to add it to the Tensor RT OSS codebase to contribute, compilation fails with undefined identifiers to all the half functions used, i.e. error: identifier "__hmax" is undefined. I have noticed that I don't think I can see any other half math intrinsics used in the rest of the codebase, is that something that is to be avoided?

Environment

TensorRT Version: OSS-master
GPU Type: RTX 3090
Nvidia Driver Version: 455.32.00
CUDA Version: 11.1
CUDNN Version:
Operating System + Version: Ubuntu 20.04
Python Version (if applicable):
TensorFlow Version (if applicable):
PyTorch Version (if applicable):
Baremetal or Container (if container which image + tag):

Relevant Files

https://github.com/5had3z/TensorRT

Steps To Reproduce

Just compile as per the readme in the standard Tensor RT OSS instructions.

[pranavm-nvidia]

You'd probably want to guard the intrinsics with SM version checks. For example: https://github.com/NVIDIA/TensorRT/blob/master/plugin/common/common.cuh#L67

[5had3z]

Is there a good way to determine the earliest SM version which supports a particular function? The math intrinsics documentation doesn't specify as far as I can tell. Also is it still required that I support SM<5.3 since they are being depricated and removed in future versions of Tensor RT?
I've been without a PSU for the past few weeks so have only resumed working on this now.

[5had3z]

Is there any particular way about going about testing for older SM arch? I've got things compiling now, however its a bit messy trying to cover float and half and recasting things if it isn't supported (especially without if constexpr). The plugin works fine in my own application with a 3090, I'm just not sure how to validate it properly as a unit test and with other GPUs.

I also haven't played with signing commits off before (previously haven't needed to do so). I've cooked some of the commits, should I be able to fix this by rebasing?

[pranavm-nvidia]

Is there a good way to determine the earliest SM version which supports a particular function?

There's a table here which shows what functionality each compute capability supports.

I also haven't played with signing commits off before (previously haven't needed to do so). I've cooked some of the commits, should I be able to fix this by rebasing?

You should be able to retroactively sign commits with: git commit --amend --signoff

CC'ing @rajeevsrao regarding how to test on older SMs,

[5had3z]

Interestingly __hmin and __hmax only work when CUDA_ARCH >= 800, otherwise returns undefined when using CUDA_ARCH >= 530, even when it should be supported from then onwards. Other functions such as __hdiv or __habs work corrrectly when using CUDA_ARCH >= 530.

[nvpohanh]

Closing for now due to >14 days with no response. Please feel free to reopen if the issue still exists. Thanks