3090 Compile Ocean Deformable Convolution / Compile Ocean DCN on 3090

see https://github.com/Soarkey/tracking_raw_results/blob/master/compile_ocean_dcn_on_3090.md

1. Environment

• Ubuntu 20
• Python 3.7
• CUDA 11.4
• PyTorch 1.9.0+cu111

2. Modification / Modification

Find the lib/models/dcn/src/deform_conv.py file, replace all AT_CHECK with TORCH_CHECK , and replace all .view operations with .reshape
Operation, then re-execute the compilation python setup.py develop .

Found file lib/models/dcn/src/deform_conv.py, replace all AT_CHECK and .view operations with TORCH_CHECK and .reshape, then recompile python setup.py develop.

3. Possible problems and solutions / References

> ValueError: Unknown CUDA arch (8.6) or GPU not supported

• Solution / Solution

  cpp_extension.py content of 0612ba0ffd2181 in the folder where the conda environment is located from:

  named_arches = collections.OrderedDict([
      ('Kepler+Tesla', '3.7'),
      ('Kepler', '3.5+PTX'),
      ('Maxwell+Tegra', '5.3'),
      ('Maxwell', '5.0;5.2+PTX'),
      ('Pascal', '6.0;6.1+PTX'),
      ('Volta', '7.0+PTX'),
      ('Turing', '7.5+PTX'),
  ])
  supported_arches = ['3.5', '3.7', '5.0', '5.2', '5.3', '6.0', '6.1', '6.2',
                      '7.0', '7.2', '7.5']

To

named_arches = collections.OrderedDict([
    ('Kepler+Tesla', '3.7'),
    ('Kepler', '3.5+PTX'),
    ('Maxwell+Tegra', '5.3'),
    ('Maxwell', '5.0;5.2+PTX'),
    ('Pascal', '6.0;6.1+PTX'),
    ('Volta', '7.0+PTX'),
    ('Turing', '7.5+PTX'),
    ('Ampere', '8.0;8.6+PTX'),
])
supported_arches = ['3.5', '3.7', '5.0', '5.2', '5.3', '6.0', '6.1', '6.2',
                    '7.0', '7.2', '7.5', '8.0', '8.6']

The difference is that with the addition of 8.6 support, 3090 belongs to the sm86 architecture.

• see solution: https://blog.csdn.net/ng323/article/details/116940299)

> undefined symbol: THPVariableClass

• Reason: When importing some third-party packages related to pytorch, if you import the third-party package first, this kind of error is prone to occur. The correct way is to import pytorch first.
• see solution: https://blog.csdn.net/slow122/article/details/116030717