Merge pull request #632 from janden/fix_626

Add source install instructions for gpu
flatironinstitute · Feb 13, 2025 · 04fbd67 · 04fbd67
2 parents 8433686 + 7f445d6
commit 04fbd67
Show file tree

Hide file tree

Showing 2 changed files with 12 additions and 2 deletions.
diff --git a/docs/python.rst b/docs/python.rst
@@ -14,7 +14,7 @@ If you would like to compile from source, you can tell ``pip`` to compile the li
   pip install --no-binary finufft finufft
 
 By default, this will use the ``-march=native`` flag when compiling the library, which should result in improved performance.
-Note that ``finufft`` has to be specified twice (first as an argument to ``--no-binary`` and second as the package the is to be installed). This option also allows you to switch out the default FFT library (FFTW) for DUCC0 using::
+Note that ``finufft`` has to be specified twice (first as an argument to ``--no-binary`` and second as the package that is to be installed). This option also allows you to switch out the default FFT library (FFTW) for DUCC0 using::
 
   pip install --no-binary finufft finufft --config-settings=cmake.define.FINUFFT_USE_DUCC0=ON finufft
 

diff --git a/docs/python_gpu.rst b/docs/python_gpu.rst
@@ -4,7 +4,17 @@ Python interface (GPU)
 Quick-start examples
 --------------------
 
-As mentioned in the :ref:`Python GPU installation instructions <install-python-gpu>`, the easiest way to install the Python interface for cuFINUFFT is to run ``pip install cufinufft``.
+As mentioned in the :ref:`Python GPU installation instructions <install-python-gpu>`, the easiest way to install the Python interface for cuFINUFFT is to run::
+
+  pip install cufinufft
+
+If you would like to compile from source, you can tell ``pip`` to compile the library from source with the option ``--no-binary`` using the command::
+
+  pip install --no-binary cufinufft cufinufft
+
+This will often result in improved performance since the build will be optimized for your particular architecture.
+In particular, it can take advantage of newer CUDA features if you have a recent version of CUDA installed.
+Note that ``cufinufft`` has to be specified twice (first as an argument to ``--no-binary`` and second as the package that is to be installed).
 
 *Note*: The interface to cuFINUFFT has changed between versions 1.3 and 2.2.
 Please see :ref:`Migration to cuFINUFFT v2.2<cufinufft_migration>` for details.