Merge branch 'master' of https://github.com/flatironinstitute/finufft

Author: ahbarnett

docs/python.rst

@@ -9,10 +9,19 @@ The easiest way to install is to run::
   pip install finufft
 
 which downloads and installs the latest precompiled binaries from PyPI.
+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 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::
+
+  pip install --no-binary finufft finufft --config-settings=cmake.define.FINUFFT_USE_DUCC0=ON finufft
+
 If you have ``pytest`` installed, you can test it with::
 
   pytest python/finufft/test
-  
+
 or, without having ``pytest`` you can run the older-style eyeball check::
 
   python3 python/finufft/test/run_accuracy_tests.py

src/cuda/memtransfer_wrapper.cu

@@ -423,7 +423,8 @@ void freegpumemory(cufinufft_plan_t<T> *d_plan)
 */
 {
   utils::WithCudaDevice device_swapper(d_plan->opts.gpu_device_id);
-  auto &stream = d_plan->stream;
+  // Fixes a crash whewre the plan itself is deleted before the stream
+  const auto stream = d_plan->stream;
 
   CUDA_FREE_AND_NULL(d_plan->fw, stream, d_plan->supports_pools);
   CUDA_FREE_AND_NULL(d_plan->fwkerhalf1, stream, d_plan->supports_pools);