Add CUDA_ARCH 80 to the default compile flags

[classicsong]

🚀 Feature

I can not build a dgl whl in a machine equipped with V100 for another machine equipped with A100. The cuda compilation flag is automatically detected. There is no way to specify the cuda compilation flag for A100.

Motivation

Alternatives

Pitch

Additional context

[BarclayII]

You could specify CUDA compiler flags with CUDA_NVCC_FLAGS variable.

If you want to specify compute capability 8.0 you will have to do it manually with -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN='30;35;50;60;70;80' (essentially the list of all architectures separated by colon).

[jermainewang]

@BarclayII Shall we add 80 to the default flag set?

[jermainewang]

@yaox12 I saw you've closed PR4384. Is this still an issue or you will find other resolutions?

[yaox12]

@yaox12 I saw you've closed PR4384. Is this still an issue or you will find other resolutions?

Either Quan's suggestion or specifying -DCUDA_ARCH_NAME=All should work.

[jermainewang]

I think the potential action item is to add a doc page about the compilation flags. As there is a workaround, this will be of low priority.