librosa: librosa 0.6.1 install fails during set-up of llvmlite 0.24.0

Solution to librosa installation:

You can install LLVM 7 from the Ubuntu apt repo, i.e., you do need to build it from source. That is quite complex. The simple method is listed as follows.

$ sudo apt-get install llvm-7 llvm-7-dev $ export LLVM_CONFIG=/usr/bin/llvm-config-7 $ pip install librosa

It is much more elegant and straightforward.

Cheers,

Mike

I was able to install librosa on my Raspberry Pi 3B+ only after building llvm from source. With llvm 6.0.1 present, llvmlite 0.24.0 will install, and the librosa install will be successful. It was quite an ordeal to get llvm build. The instructions here are helpful but require some tweaking: https://releases.llvm.org/6.0.1/docs/CMake.html#frequently-used-cmake-variables

Building llvm for a Raspberry Pi 3 required a few modifications to the base system: 1.) 64GB SD card - else the build runs out of disk space (see note further down) 2.) 2GB swap file - else the linker runs out of memory

Assuming the llvm sources have been downloaded and extracted into this directory: ~/Downloads/llvm-6.0.1.src

Create a build directory, e.g. $ mkdir llvm_build_dir $ cd llvm_build_dir

Use the following command from inside the build directory to configure the llvm build (I am not sure the sudo is necessary but didn’t have the patience to try without):

sudo cmake ~/Downloads/llvm-6.0.1.src -DLLVM_TARGETS_TO_BUILD="ARM" -DCMAKE_BUILD_TYPE="Release"

Use the following build command to build and link using all four CPUs: $ sudo cmake --build . -- -j4 This will take several hours. In hindsight, building with the “Release” option results in much smaller llvm file size and may not require the 64GB SD card, but I didn’t confirm that.

The build should go through but if the linker runs out of memory, refer to this website: https://stackoverflow.com/questions/25197570/llvm-clang-compile-error-with-memory-exhausted The following worked for me: sudo ln -s 'which gold' /usr/local/bin/ld

After build completes, install with sudo cmake --build . --target install

Now llvmlite 0.24.0 will install without error: pip3 install llvmlite

Now librosa 0.61 will install without error: pip3 install librosa

However, this is still not quite enough: when trying to import librosa, it still throws an error, namely

ImportError: libf77blas.so.3: cannot open shared object file: No such file or directory

The error is fixed with: sudo apt-get install libatlas-base-dev

It should be possible at this stage to remove the llvm source and build directory to free up many GB on the SD card, but I haven’t tested that yet.

Hopefully sharing this will be useful for other people. Maybe in the future somebody comes up with an easier way to install llvmlite that doesn’t require building llvm first, but for the time being this works.

You can close it,

even though I don’t think there is a satisfactory solution - it is still necessary to go through the pain of build llvm to get llvmlite - but that’s not a librosa issue.

Michael

On Jul 5, 2019, at 16:01, Brian McFee <notifications@github.commailto:notifications@github.com> wrote:

Any objection to closing this out?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://github.com/librosa/librosa/issues/757?email_source=notifications&email_token=ACAKGIO3SHQ4YSFSF4ABUWDP55H23A5CNFSM4FOJJVAKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZJTB5I#issuecomment-508768501, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ACAKGIKVZ6O73YSHBA4Z3CDP55H23ANCNFSM4FOJJVAA.

@kim-sunghoon , everything seems to install properly yet you get this error. I cannot reproduce it since I am on a Raspberry and everything works, as described. The only idea I have is that something went wrong in the llvm build: sudo cmake llvm-7.0.1.src -DLLVM_TARGETS_TO_BUILD=“ARM” -DCMAKE_BUILD_TYPE=“Release" Double-check you have the right target specified. I don’t think this is a librosa issue. You might want to post your question on a llvm-related forum.

If you use berryconda you can simply:

$ conda install numba

and it will just work. The LLVM 7 build needed for llvmlite and Numba to work correctly also requires source level patching as described here https://llvmlite.readthedocs.io/en/latest/admin-guide/install.html#compiling-llvm.

CC: @seibert @sklam

Numba 0.42 is now supported on RPi with berryconda.