Pyvkfft

-----------------------------
* Include doc in manifest

-----------------------------
* Based on VkFFT 1.3.1
* Add support for more than 3 FFT dimensions (defaults to 8
in pyvkfft, can be changed when installing)
* Add options to manually or automatically tune the FFT performance
for the used GPUs.
* Add pyvkfft-benchmark script.
* The VkFFT source is now included as a git subproject
* Actually use cuda_stream parameter in the pyvkfft.fft interface
* Take into account current cuda device when automatically
caching VkFFTApp using the pyvkfft.fft interface(26)
This enable using multiple GPUs in a multi-threaded approach.
* The pyvkfft-test will ignore the PoCL OpenCL platform if
another is available and unless it is explicitly selected.

-----------------------------
* Fix MANIFEST.in including vkfft_cuda.cu for python setup.py sdist

-----------------------------
* VkFFT 1.2.33, now using Rader algorithm for better performance
with many non-radix sizes.
* Fix R2C tests when using numpy (scipy unavailable) [19]
* Add support for F-ordered arrays (C2C and R2C)
* Allow selection of backend for non-systematic pvkfft-test
* Add parameter to select the OpenCL platform in pyvkfft-test
* For pyopencl, transforms will use the array's queue by default
instead of the application's (a warning will be written to
notify this change when they differ; it can be disabled using
config.WARN_OPENCL_QUEUE_MISMATCH). A queue can also be supplied
to the fft() and ifft() methods.
(from isuruf, https://github.com/vincefn/pyvkfft/pull/17)
* Fix simple fft interface import when only pycuda is used
* Add cuda_driver_version, cuda_compile_version, cuda_runtime_version
functions.
* Add simpler interface to run benchmarks, using separate processes.
* add pyvkfft-test-suite for long tests (up to 30 hours) for validation
before new releases.

-----------------------------
* Correct the dtype of the returned array for fft.rfftn() and fft.irfftn()
in the case of an inplace transform
* Pycuda: cast the gpudata pointer to int for comparisons
* Fix TestFFT colour attribute default value

-----------------------------
* Added accuracy unit tests, which can be used systematically
using the 'pyvkfft-test' installed script
* An extensive testing is now made before official releases,
evaluating all type of transforms (c2c, r2c, dct, 1, 2 and 3D,
in and out-of-place, norm 0 and 1), different GPUs, both OpenCL
and CUDA, etc... Comparison is made against pyfftw, scipy or numpy.
* Update to VkFFT 1.2.21, with support for DCT types 1, 2, 3 and 4,
also fixing a number of issues (see closed issues at
https://github.com/DTolm/VkFFT/issues), and passing all tests
on different GPUs (OpenCL and CUDA, AMD and nVidia)
* Raise a RuntimeError if the VkFFTApp initialisation or the
GPU kernel launch fails, with the corresponding VkFFT error.
* [BUG] Correct inverse FFT calculation using pyvkfft.fft.ifftn()
* Installation from source using 'python setup.py install' is now
disabled - 'pip install' should always be used.
* Added config.USE_LUT and config.FFT_CACHE_NB variables, which
can be used to modify the default behaviour, and can also be set
e.g. with the PYVKFFT_USE_LUT environment variable.