[bugfix] Link against pytorch library by name rather than path (attempting to fix linking issue).

[nv-dlasalle]

Description

This is meant to fix #3240, which was caused by #3225, which in turn was trying to fix #3220, the issue of having DGL compiled against versions of pytorch with different cuda versions (e.g., the default from pip 1.9.0+cu102 vs. 1.9.0+cu111).

This explicitly links against torch by name, rather than path. This needs more scrutiny and testing, as I can't say I'm an expert on dynamic linking. In my testing below, this appears to achieve the desired result, of allowing pytorch with different CUDA versions to work against the same tensor adapter binary.

Checklist

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [$CATEGORY] (such as [NN], [Model], [Doc], [Feature]])
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• Code is well-documented
• To the best of my knowledge, examples are either not affected by this change,
  or have been fixed to be compatible with this change
• Related issue is referred in this PR

Changes

Prior to #3225, when compiling, cmake would output the following:

-- tensoradapter found PyTorch includes: /home/dominique/.local/lib/python3.6/site-packages/torch/include;/home/dominique/.local/lib/python3.6/site-packages/torch/include/torch/csrc/api/include
-- tensoradapter found PyTorch lib: torch;torch_library;/home/dominique/.local/lib/python3.6/site-packages/torch/lib/libc10.so;/usr/local/cuda/lib64/stubs/libcuda.so;/usr/local/cuda/lib64/libnvrtc.so;/usr/local/cuda/lib64/libnvToolsExt.so;/usr/local/cuda/lib64/libcudart.so;/home/dominique/.local/lib/python3.6/site-packages/torch/lib/libc10_cuda.so

After #3225, but without this PR, it would output:

-- tensoradapter found PyTorch includes: /home/dominique/.local/lib/python3.6/site-packages/torch/include;/home/dominique/.local/lib/python3.6/site-packages/torch/include/torch/csrc/api/include
-- tensoradapter found PyTorch lib: /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch.so

With this PR:

-- tensoradapter found PyTorch includes: /home/dominique/.local/lib/python3.6/site-packages/torch/include;/home/dominique/.local/lib/python3.6/site-packages/torch/include/torch/csrc/api/include
-- tensoradapter found PyTorch lib: torch

When compiling DGL with this PR and torch==1.9.0+cu102:

$ python3
Python 3.6.9 (default, Jan 26 2021, 15:33:00)
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import dgl
Using backend: pytorch
RDFLib Version: 5.0.0
>>> print(dgl._ffi.base.tensor_adapter_loaded)
True

$ ldd ~/.local/lib/python3.6/site-packages/dgl-0.8-py3.6-linux-x86_64.egg/dgl/tensoradapter/pytorch/libtensoradapter_pytorch_1.9.0.so
        linux-vdso.so.1 (0x00007ffcf5bc4000)
        libtorch.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch.so (0x00007f302fa05000)
        libtorch_cpu.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch_cpu.so (0x00007f301c671000)
        libtorch_cuda.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch_cuda.so (0x00007f2fda16d000)
        libc10.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libc10.so (0x00007f2fd9ec9000)
        libnvToolsExt.so.1 => /usr/local/cuda/lib64/libnvToolsExt.so.1 (0x00007f2fd9cc0000)
        libstdc++.so.6 => /usr/lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007f2fd98ec000)
        libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007f2fd96d4000)
        libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f2fd92e3000)
        libgomp-a34b3233.so.1 => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libgomp-a34b3233.so.1 (0x00007f2fd90b9000)
        libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f2fd8e9a000)
        librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007f2fd8c92000)
        libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007f2fd8a8e000)
        libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f2fd86f0000)
        libcudart-80664282.so.10.2 => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libcudart-80664282.so.10.2 (0x00007f2fd846f000)
        /lib64/ld-linux-x86-64.so.2 (0x00007f302fe46000)
        libc10_cuda.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libc10_cuda.so (0x00007f2fd8242000)
        libnvToolsExt-3965bdd0.so.1 => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libnvToolsExt-3965bdd0.so.1 (0x00007f2fd8038000)

Then using the same dgl installation, and changing the pytorch version to 1.9.0+cu111.

$ python3
Python 3.6.9 (default, Jan 26 2021, 15:33:00)
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import dgl
Using backend: pytorch
RDFLib Version: 5.0.0
>>> print(dgl._ffi.base.tensor_adapter_loaded)
True

$ ldd ~/.local/lib/python3.6/site-packages/dgl-0.8-py3.6-linux-x86_64.egg/dgl/tensoradapter/pytorch/libtensoradapter_pytorch_1.9.0.so
        linux-vdso.so.1 (0x00007fff7e3e3000)
        libtorch.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch.so (0x00007f2eaaf9b000)
        libtorch_cpu.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch_cpu.so (0x00007f2e978d7000)
        libtorch_cuda.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch_cuda.so (0x00007f2e976c3000)
        libc10.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libc10.so (0x00007f2e9741f000)
        libnvToolsExt.so.1 => /usr/local/cuda/lib64/libnvToolsExt.so.1 (0x00007f2e97216000)
        libstdc++.so.6 => /usr/lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007f2e96e42000)
        libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007f2e96c2a000)
        libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f2e96839000)
        libtorch_cuda_cpp.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch_cuda_cpp.so (0x00007f2e065b9000)
        libtorch_cuda_cu.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libtorch_cuda_cu.so (0x00007f2dba9bc000)
        libgomp-7c85b1e2.so.1 => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libgomp-7c85b1e2.so.1 (0x00007f2dba792000)
        libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f2dba573000)
        librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007f2dba36b000)
        libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007f2dba167000)
        libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f2db9dc9000)
        libcudart-6d56b25a.so.11.0 => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libcudart-6d56b25a.so.11.0 (0x00007f2db9b40000)
        /lib64/ld-linux-x86-64.so.2 (0x00007f2eab3ca000)
        libc10_cuda.so => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libc10_cuda.so (0x00007f2db9912000)
        libnvToolsExt-24de1d56.so.1 => /home/dominique/.local/lib/python3.6/site-packages/torch/lib/libnvToolsExt-24de1d56.so.1 (0x00007f2db9708000)

[dgl-bot]

To trigger regression tests:

• @dgl-bot run [instance-type] [which tests] [compare-with-branch];
  For example: @dgl-bot run g4dn.4xlarge all dmlc/master or @dgl-bot run c5.9xlarge kernel,api dmlc/master

[BarclayII]

@davidmin7 Could you see if this fix works?

[davidmin7]

Hi @BarclayII,

Yup, this did fix the problem that I was having in #3240. I'll close it as soon as this PR gets merged. Thanks!