[cairo] Large image support?

[Ian Britten]

Simon Budig wrote:

>> 2) Creating large images
>>     We have workflows where users need to generate very large
>>     high-res images (eg: 50000x50000 pixels).  Again in this case,
>>     simply trying to allocate a Cairo image surface that size will
>>     likely exceed the available memory.

> In Gimp images internally are stored in relatively small "tiles".
> 
> for (all tiles) {
>    surface = cairo_image_surface_create_for_data (tmp_buf,
>                                                   CAIRO_FORMAT_A8,
>                                                   width, height,
>                                                   stride);
> 
>    cairo_surface_set_device_offset (surface, -x0, -y0);
>    cr = cairo_create (surface);
>     [drawing code]
>    cairo_destroy (cr);
>    cairo_surface_destroy (surface);
> }
> 
> i.e. we render the same drawing operations multiple times to different
> tiles, adjusting the device offset accordingly.

Ya, we had kicked around a similar idea here already...  A couple of
issues:
- Our 'canvas' just receives the data once, and isn't really in a
   position to trigger multiple redraws.  However, I might be able to
   come up with something (It's solely our problem)
- A large portion of our data is vector, not raster.  I can see
   cases where a tile may end up lacking data, due to the main part
   of the vector data being prematurely discarded for not falling
   within the tile (eg: Point data that draws extra labels, etc).

So, what about a slightly different approach?
I know from my PDF usage that Cairo maintains the vector data
internally somehow (non-raster).  Could I maybe:
- Render all my data to a PDF (or similar) surface first, as I
   currently do.  Then, once done:
- Make a temporary image tile
- Iterate over the PDF surface an appropriate number of times.
- On each iteration, copy from the PDF surface to my tile.
- Write/concatenate the tile to the output image file.
My (limited) understanding is that Cairo should/will flatten and
rasterize the PDF data when copied to an image surface, thus
giving me the raster pixels I need.  As well, since my main
target surface is PDF, it won't allocate that 50000x50000 pixel
buffer, bypassing that problem.

Thoughts?  Problems?
As always, many thanks for any feedback!
Ian