[Numpy-discussion] numpy arrays, data allocation and SIMD alignement
cournape at gmail.com
Thu Aug 9 11:14:05 EDT 2007
On 8/9/07, Stefan van der Walt <stefan at sun.ac.za> wrote:
> It doesn't really matter where the memory allocation occurs, does it?
> As far as I understand, the underlying fftw function has some flag to
> indicate when the data is aligned. If so, we could expose that flag
> in Python, and do something like
> x = align16(data)
> _fft(x, is_aligned=True)
> I am not intimately familiar with the fft wrappers, so maybe I'm
> missing something more fundamental.
You can do that, but this is only a special case of what I have in
mind. For example, what if you want to call functions which are
relatively cheap, but called many times, and want an aligned array ?
Going back and forth would be a huge waste. Also, having aligned
buffers internally (in C_, even for non array data, can be useful (eg
filters, and maybe even core numpy functionalities like ufunc,
etc...). Another point I forgot to mention before is that we can
define a default alignment which would already be SIMD friendly (as
done on Mac OS X or FreeBSD by default malloc) for *all* numpy arrays
at 0 cost: for fft, this means that most arrays would already by as
wanted, meaning a huge boost of performances for free.
Basically, the functionalities would be more usable in C, without too
much constraint, because frankly, the implementation is not difficult:
I have something almost ready, and the patch is 7kb, including code to
detect platform dependent aligned allocator. The C code can be tested
really easily (since it is independent of python).
More information about the NumPy-Discussion