nvptx target missing in 0.33 when installed using pip

[jvesely]

Hi, has the PTX target been removed in 0.33?

Trying to use target triple nvptx64-nvidia-cuda results in:

cls = <class 'llvmlite.binding.targets.Target'>, triple = 'nvptx64-nvidia-cuda'

    @classmethod
    def from_triple(cls, triple):
        """
        Create a Target instance for the given triple (a string).
        """
        with ffi.OutputString() as outerr:
            target = ffi.lib.LLVMPY_GetTargetFromTriple(triple.encode('utf8'),
                                                        outerr)
            if not target:
>               raise RuntimeError(str(outerr))
E               RuntimeError: No available targets are compatible with triple "nvptx64-nvidia-cuda"

../.local/lib/python3.7/site-packages/llvmlite/binding/targets.py:197: RuntimeError

Reverting to 0.32.1 fixes the issue and my program runs without issues.
Rebuilding 0.33.0 from source using system LLVM also works without any changes to the original program.

[stuartarchibald]

Thanks for the report. Do you have a reproducer for this please? Also, where did you get llvmlite from? I cannot reproduce locally:

In [1]: import llvmlite                                                                                                         

In [2]: llvmlite.__version__                                                                                                    
Out[2]: '0.33.0'

In [3]: import llvmlite.binding as ll                                                                                           

In [4]: ll.initialize()                                                                                                         

In [5]: ll.initialize_all_targets()                                                                                             

In [6]: from llvmlite.binding.targets import Target                                                                             

In [7]: Target.from_triple('nvptx64-nvidia-cuda')                                                                               
Out[7]: <llvmlite.binding.targets.Target at 0x7facb55b7730>

[stuartarchibald]

Also, the llvmdev package, against which llvmlite is linked has:

$ llvm-config --version
9.0.1

$ llvm-config --targets-built
AMDGPU NVPTX WebAssembly X86

[jvesely]

The above example doesn't work when llvmlite is installed from pip:

$ pip3 install -U --user llvmlite
Collecting llvmlite
  Using cached llvmlite-0.33.0-cp37-cp37m-manylinux1_x86_64.whl (18.3 MB)
Installing collected packages: llvmlite
Successfully installed llvmlite-0.33.0
$ python3
Python 3.7.7 (default, May  3 2020, 11:13:24) 
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import llvmlite
>>> llvmlite.__version__
'0.33.0'
>>> import llvmlite.binding as ll
>>> ll.initialize()
>>> ll.initialize_all_targets()
>>> from llvmlite.binding.targets import Target
>>> Target.from_triple('nvptx64-nvidia-cuda')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/orome/.local/lib/python3.7/site-packages/llvmlite/binding/targets.py", line 197, in from_triple
    raise RuntimeError(str(outerr))
RuntimeError: No available targets are compatible with triple "nvptx64-nvidia-cuda"
>>>

[stuartarchibald]

Thanks for confirming, llvmlite 0.33 moved to llvm 9, I think the additional targets didn't get added to the build script in the wheel build path. Conda version: https://github.com/numba/llvmlite/blob/master/conda-recipes/llvmdev/build.sh#L31-L32 Wheel: https://github.com/numba/llvmlite/blob/master/conda-recipes/llvmdev_manylinux1/build.sh#L31

[esc]

I prepped a PR applying the suggested fix here: #602