Buffers and pointers (Py3.4)
Rob Gaddi
rgaddi at highlandtechnology.invalid
Tue Apr 18 13:09:41 EDT 2017
On 04/17/2017 11:47 AM, Chris Angelico wrote:
> On Tue, Apr 18, 2017 at 3:58 AM, Rob Gaddi
> <rgaddi at highlandtechnology.invalid> wrote:
>> If I were writing this as a C extension, getting that information from any
>> buffer object would be trivial, but that changes my project from a pure
>> Python wrapper using ctypes to a mixed language project with compile
>> dependencies and mental headaches.
>>
>> buffertype = c_uint8 * size
>> return addressof(buffertype.from_buffer(buf, offset))
>>
>> works but is inefficient and woefully inelegant.
>
> Have you considered Cython? It can massively simplify a lot of the
> work in doing this kind of thing.
>
> ChrisA
>
I did, and it definitely seems easier than grunting it up from pure C,
but it still adds a compile step and extra tools, and I found the
documentation on working with arbitrary memory to be lacking. I know C,
I know Python, I'm definitely unclear as to how to use Cython to split
the difference. Somewhere I've got the book and really need to get
around to reading it.
I actually got to an implementation that isn't awful. Creating new
objects that represent each DMA transfer (and wrap a Structure to be
passed to the library) takes about 9 us per object, including a call to:
def setbuffer(self, buf, start=0, end=None):
"""Set the local memory.
Parameters
buf - The local memory, an object implementing the buffer
protocol such as a numpy.ndarray or a bytearray. Can
be None, which invalidates the transfer.
start - The index of the first buffer element to transfer.
Defaults to 0.
end - The index of the last buffer element to transfer. Defaults
to None for len(buf).
"""
if buf is None:
self._localmem = None
self._desc.ptr = 0
return
basemv = memoryview(buf)
mv = basemv[start:end]
if not mv.contiguous:
raise ValueError("noncontiguous buffer")
self._localmem = mv
self._desc.size = mv.nbytes
self._desc.ptr = addressof(c_uint8.from_buffer(mv))
This could obviously be made faster by dropping some of the safety
checks and flexibility, but I feel like the speed hit is worth not
getting segfaults down the line because someone gets the clever idea to
stride the array. Plus the objects can be persistent, just calling
setbuffer to update for the new destination, which takes a lot of that
time back off.
--
Rob Gaddi, Highland Technology -- www.highlandtechnology.com
Email address domain is currently out of order. See above to fix.
More information about the Python-list
mailing list