ReadAsArray in python/numpy gdal bindings should take pointer to preallocated memory

[jaredrubinaz]

When using the C++ side of gdal with the numpy/python bindings. I would expect that a preallocated array would be generated from numpy and a pointer to the beginning of the memory buffer would be passed to gdal.

Currently the BandReadAsArray method in gdal_array.py is not even using the buf_obj that is passed in. If a buf_obj is passed in it overrides it with a new numpy array. This makes no sense. This method should just use the buf_obj

129 if buf_obj is not None: 130 buf_obj = numpy.zeros( shape, typecode ) 131 132 band_str = band.ReadRaster( xoff, yoff, win_xsize, win_ysize, 133 buf_xsize, buf_ysize, datatype ) 134 ar = numpy.fromstring(band_str,dtype=typecode) 135 ar = numpy.reshape(ar, [buf_ysize,buf_xsize])

allocates the memory for the

RasterIO call, can one not have numpy allocate the memory and then pass

a pointer to the gdal/swig bindings?

[dhelfman]

Note: #3173 describes a crash that can occur as a result of the GDAL Python bindings not taking a buffer to pre-allocated memory.

[dhelfman]

#2658 is also related and proposes a specific solution.

[Even Rouault]

• r17895, r17897 : Python bindings : make Band|Dataset.ReadAsArray() take into account preallocated array (#2658, #3028)
• r17896 : Test fix for #2658, #3028

The fix required quite a few changes in the gdal_array module (I swig'ified it, removed swig/python/extensions/numpydataset.cpp which was unused, etc..). Testing by others appreciated.