OpenMathLib
Building on Windows with LLVM - docs need update
I tried building OpenBLAS on Windows with the LLVM toolchain to be able to link statically with the native ABI. I got stuck after installing the conda packages, because - at least in my case - conda installs very old versions of packages. The flang compiler (installed version 11.1.0, while 20.1.6 is available) is unable to compile LAPACK.
Is there a location where I can propose some edits for this page? -> http://www.openmathlib.org/OpenBLAS/docs/install/
If you mean through a PR, then that's very welcome - sources are here: https://github.com/OpenMathLib/OpenBLAS/blob/develop/docs/install.md
Happy to discuss here too. You should be able to get conda/mamba/pixi to install the more recent compiler, I have Flang 19.1.7 in a conda-forge based Windows dev environment.
~~Guess it might make sense now to point people to the LLVM release page, like we already do for Windows on Arm - with new releases every two weeks, it may be a bit hard for third-party packagers to keep up~~ The build instructions we currently have should be correct for the 19.1.x series, my impression so far has been that what would primarily need checking&updating is flang's OpenMP support.
conda gave me flang version 11.1.0 from 2017 - that is too old to compile LAPACK. I propose to continue any discussion in https://github.com/OpenMathLib/OpenBLAS/pull/5338
@mennodeij it's best to (a) keep the discussion in the place where it's already happening, and (b) not close an issue before it's actually resolved through either a change on your side or a merged PR.
conda gave me flang version 11.1.0 from 2017
In a fresh environment, or are you installing in an old pre-existing environment? If the former, can you please give a full reproducer so I can try it?
I never used miniconda before, and followed the instructions: install miniconda3 with winget, and then from the anaconda prompt:
conda update -n base conda
conda config --add channels conda-forge
conda install -y cmake flang clangdev perl libflang ninja
These gave me old packages, not only for flang, all of them.
Okay, I can see how that could result in an issue, because it's mixing two channels (defaults and conda-forge) and it installs into an existing environment rather than a fresh one (i.e., step 4 of http://www.openmathlib.org/OpenBLAS/docs/install/#windows_1 seems sub-optimal).
Could you please share the output of conda list? That will probably confirm what I'm thinking the problem is.
I have it at the correct versions now:
# packages in environment at C:\Users\menno\miniconda3:
#
# Name Version Build Channel
anaconda-anon-usage 0.7.1 py313hfc23b7f_100
anaconda_powershell_prompt 1.1.0 haa95532_1
anaconda_prompt 1.1.0 haa95532_1
annotated-types 0.6.0 py313haa95532_0
archspec 0.2.3 pyhd3eb1b0_0
boltons 24.1.0 py313haa95532_0
brotli-python 1.0.9 py313h5da7b33_9
bzip2 1.0.8 h2bbff1b_6
ca-certificates 2025.6.15 h4c7d964_0 conda-forge
certifi 2025.6.15 pyhd8ed1ab_0 conda-forge
cffi 1.17.1 py313h827c3e9_1
charset-normalizer 3.3.2 pyhd3eb1b0_0
clang 20.1.6 default_hb5d4d12_0 conda-forge
clang-20 20.1.6 default_hb5d4d12_0 conda-forge
clang-format 20.1.6 default_hb5d4d12_0 conda-forge
clang-tools 20.1.6 default_hb5d4d12_0 conda-forge
clangdev 20.1.6 default_hb5d4d12_0 conda-forge
clangxx 20.1.6 default_hb5d4d12_0 conda-forge
cmake 4.0.3 hff78f93_0 conda-forge
colorama 0.4.6 py313haa95532_0
compiler-rt 20.1.6 hc790b64_0 conda-forge
compiler-rt_win-64 20.1.6 hc790b64_0 conda-forge
conda 25.5.1 py313haa95532_0
conda-anaconda-telemetry 0.1.2 py313haa95532_1
conda-anaconda-tos 0.2.0 py313haa95532_0
conda-content-trust 0.2.0 py313haa95532_1
conda-libmamba-solver 25.4.0 pyhd3eb1b0_0
conda-package-handling 2.4.0 py313haa95532_0
conda-package-streaming 0.11.0 py313haa95532_0
cpp-expected 1.1.0 h214f63a_0
cryptography 45.0.3 py313h51e0144_0
distro 1.9.0 py313haa95532_0
expat 2.7.0 he0c23c2_0 conda-forge
flang 20.1.6 h1b5488d_0 conda-forge
fmt 9.1.0 h6d14046_1
frozendict 2.4.2 py313haa95532_0
idna 3.7 py313haa95532_0
jsonpatch 1.33 py313haa95532_1
jsonpointer 2.1 pyhd3eb1b0_0
krb5 1.21.3 hdf4eb48_0 conda-forge
libarchive 3.7.7 h88ece9c_0 conda-forge
libclang 20.1.6 default_hb5d4d12_0 conda-forge
libclang-cpp 20.1.6 default_hb5d4d12_0 conda-forge
libclang13 20.1.6 default_h6e92b77_0 conda-forge
libcurl 8.14.1 h88aaa65_0 conda-forge
libexpat 2.7.0 he0c23c2_0 conda-forge
libffi 3.4.4 hd77b12b_1
libflang 20.1.6 hac47afa_0 conda-forge
libiconv 1.18 h135ad9c_1 conda-forge
libllvm-c20 20.1.6 h3089188_0 conda-forge
libllvm20 20.1.6 h3089188_0 conda-forge
liblzma 5.8.1 h2466b09_2 conda-forge
liblzma-devel 5.8.1 h2466b09_2 conda-forge
libmamba 2.0.5 hcd6fe79_1
libmambapy 2.0.5 py313h214f63a_1
libmpdec 4.0.0 h827c3e9_0
libsolv 0.7.30 hf2fb9eb_1
libsqlite 3.50.1 hf5d6505_6 conda-forge
libssh2 1.11.1 h9aa295b_0 conda-forge
libuv 1.51.0 h2466b09_0 conda-forge
libxml2 2.13.8 h442d1da_0 conda-forge
libzlib 1.3.1 h2466b09_2 conda-forge
llvm-tools 20.1.6 h112c9dc_0 conda-forge
llvmdev 20.1.6 h3089188_0 conda-forge
lz4-c 1.9.4 h2bbff1b_1
lzo 2.10 hcfcfb64_1001 conda-forge
markdown-it-py 2.2.0 py313haa95532_1
mdurl 0.1.0 py313haa95532_0
menuinst 2.2.0 py313h5da7b33_1
ninja 1.13.0 h79cd779_0 conda-forge
nlohmann_json 3.11.2 h6c2663c_0
openssl 3.5.0 ha4e3fda_1 conda-forge
packaging 24.2 py313haa95532_0
pcre2 10.42 h17e33f8_0 conda-forge
perl 5.32.1.1 7_h57928b3_strawberry conda-forge
pip 25.1 pyhc872135_2
platformdirs 4.3.7 py313haa95532_0
pluggy 1.5.0 py313haa95532_0
pybind11-abi 5 hd3eb1b0_0
pycosat 0.6.6 py313h827c3e9_2
pycparser 2.21 pyhd3eb1b0_0
pydantic 2.10.3 py313haa95532_0
pydantic-core 2.27.1 py313h636fa0f_0
pygments 2.19.1 py313haa95532_0
pysocks 1.7.1 py313haa95532_0
python 3.13.0 hf5aa216_101_cp313 conda-forge
python_abi 3.13 0_cp313
reproc 14.2.4 hd77b12b_2
reproc-cpp 14.2.4 hd77b12b_2
requests 2.32.3 py313haa95532_1
rich 13.9.4 py313haa95532_0
ruamel.yaml 0.18.10 py313h827c3e9_0
ruamel.yaml.clib 0.2.12 py313h827c3e9_0
setuptools 78.1.1 py313haa95532_0
simdjson 3.10.1 h214f63a_0
spdlog 1.11.0 h59b6b97_0
sqlite 3.45.3 h2bbff1b_0
tk 8.6.13 h2c6b04d_2 conda-forge
tqdm 4.67.1 py313h4442805_0
truststore 0.10.0 py313haa95532_0
typing-extensions 4.12.2 py313haa95532_0
typing_extensions 4.12.2 py313haa95532_0
tzdata 2025b h04d1e81_0
ucrt 10.0.22621.0 h57928b3_1 conda-forge
urllib3 2.3.0 py313haa95532_0
vc 14.42 haa95532_5
vc14_runtime 14.44.35208 h818238b_26 conda-forge
vs2015_runtime 14.44.35208 h38c0c73_26 conda-forge
wheel 0.45.1 py313haa95532_0
win_inet_pton 1.1.0 py313haa95532_0
xz 5.8.1 h208afaa_2 conda-forge
xz-tools 5.8.1 h2466b09_2 conda-forge
yaml-cpp 0.8.0 hd77b12b_1
zlib 1.3.1 h2466b09_2 conda-forge
zstandard 0.23.0 py313ha7868ed_2 conda-forge
zstd 1.5.7 hbeecb71_2 conda-forge
Thanks. That is still helpful. Having a mix of packages from defaults (the ones without "conda-forge" behind them) and conda-forge isn't very robust. I'll update the docs for step 4 to be more idiomatic, that should prevent the issue potentially popping up for the next person following those docs.
Looks like using miniforge3 instead of miniconda3 is maybe the easiest solution.
Yes that works indeed. If you're not used to using conda envs, that's a 1:1 change. Better practice is to create a conda environment and not installing things in the base environment.
As you probably guessed by now, I have zero experience with conda, environments etc. Best to make instructions based on that assumption - I have no idea what "not installing things in the base environment" even means 🤷