On 18 September 2012 15:19, Christoph Groth <c...@falma.de> wrote: > Hi, > > Thanks a lot for the recently added generic memoryview support, it's > just what we needed to optimize an important part of our program. > > However, I believe that there's a problem with Cython's support for > objects which do not provide "strides" information (because they are > C-contiguous). > > The buffer API documentation [1] is a bit ambiguous on this, but it says > > "If strides is NULL, the array is interpreted as a standard > n-dimensional C-array." > > In any case, numpy happily accepts such buffers and computes the > "strides" itself. > > A thread [2] on Python's devel mailing list also supports the view that > setting strides to NULL is OK even if strides were requested. > > You might ask: why not simply provide strides when they have been > requested? The problem is to find space for them. malloc'ing or > PyMem_Alloc'ing memory for strides each time when a buffer is requested > is slow, and reserving space for strides within the object is wasteful > when "shape" is all that's needed and you have _many_ small objects > which expose the buffer interface. > > The provided patch works for our use case, but someone with a good > understanding of Cython's memoryview support should finish it. I > believe the patch will break more complicated buffer use cases. > > [1] http://docs.python.org/dev/c-api/buffer.html > [2] http://thread.gmane.org/gmane.comp.python.devel/134387 > > Thank you for considering fixing this issue in Cython, > > Christoph > > > --- > Cython/Utility/MemoryView_C.c | 19 ++++++++++++------- > 1 file changed, 12 insertions(+), 7 deletions(-) > > diff --git a/Cython/Utility/MemoryView_C.c b/Cython/Utility/MemoryView_C.c > index 84e55da..c2524b0 100644 > --- a/Cython/Utility/MemoryView_C.c > +++ b/Cython/Utility/MemoryView_C.c > @@ -207,12 +207,6 @@ static int __Pyx_ValidateAndInit_memviewslice( > goto fail; > } > > - if (!buf->strides) { > - PyErr_SetString(PyExc_ValueError, > - "buffer does not supply strides necessary for memoryview."); > - goto fail; > - } > - > for(i=0; i<ndim; i++) { > spec = axes_specs[i]; > > @@ -320,8 +314,19 @@ __Pyx_init_memviewslice(struct __pyx_memoryview_obj > *memview, > goto fail; > } > > + if (buf->strides) { > + for (i = 0; i < ndim; i++) { > + memviewslice->strides[i] = buf->strides[i]; > + } > + } else { > + Py_ssize_t stride = buf->itemsize; > + for (i = ndim - 1; i >= 0; i--) { > + memviewslice->strides[i] = stride; > + stride *= buf->shape[i]; > + } > + } > + > for (i = 0; i < ndim; i++) { > - memviewslice->strides[i] = buf->strides[i]; > memviewslice->shape[i] = buf->shape[i]; > if (buf->suboffsets) { > memviewslice->suboffsets[i] = buf->suboffsets[i]; > -- > 1.7.10.4 > > _______________________________________________ > cython-devel mailing list > cython-devel@python.org > http://mail.python.org/mailman/listinfo/cython-devel
Hey Christoph, Thanks for the patch, it looks fine to me. If you want, you could make a pull request and I'll merge it. Mark _______________________________________________ cython-devel mailing list cython-devel@python.org http://mail.python.org/mailman/listinfo/cython-devel
