Create a high-level RasterUpdater.write function

[mwtoews]

This function should be best friends with read(), so data from numpy arrays could be symmetrically read/written for multidimensional arrays/bands. See #71 and #83 for background to the high-level read() implementation. This function will make use of write_band, which will remain the same.

The proposed prototype for the RasterUpdater class is:

def write(self, src, bidx=None, window=None)

where

• src: source Numpy array (masked or regular)
  • A 3D array (nbands, nrow, ncol) with optional list of indexes e.g. bidx=[1, 3] for two bands: the first and third
  • A 2D array (nrow, ncol) with band index integer e.g. bidx=1 for the first
• bidx: band index as list (for 3D) or integer (for 2D). Default of None assumes all indexes of the raster, and a length check is done with the 0th dimension of src
• window: same as in write_band

When a mask array is used for src, the band array should be written using the MaskedArray.filled(fill_value=None) function, where fill_value is from nodatavals if not None for the band index, otherwise should this be set from the masked array fill_value. Only in this case, the nodatavals list for the raster could be modified by this function to update it from None to fill_value from the masked array.

Feedback and suggestions are welcome!

[sgillies]

👍

@mwtoews I'd like to point out that we're now releasing the GIL when calling GDALRasterioIO (see https://github.com/mapbox/rasterio/blob/master/rasterio/_io.pyx#L42). If, instead of having write() call write_band() multiple times, we pushed the writes into almost-pure C functions (multi-band versions of the io_*() functions) we could release the GIL once for many reads/writes. Does that make sense to you? Then write_band() could just call write().

[sgillies]

Hey, so we're not duplicating effort: I'm making good progress on io_multi_*() functions that read or write multiple bands in a branch I've dedicated to #96. Along the way I'm learning some new stuff about Cython typed memoryviews and the GIL (all good).

[sgillies]

I closed #96. The io_multi* functions are good to go.

[sgillies]

@mwtoews in the end, I didn't implement mask filling, but we can do that as another issue.