Performance improvement for TMS tileset generation

[hokieg3n1us]

I've been utilizing the TMS tileset rendering feature for a couple weeks, after seeing Tom White's article, using his branch that was recently merged. I'd been attempting to render the OSM 1 billion point dataset to a TMS tileset for demonstration purposes, but repeatedly failing on my machine (Xeon E3-1505M v6, 64 GB Ram). I investigated and identified some performance improvements that I've included in this pull request:

• Remove duplicate calls to gen_super_tiles
• Cache results of rasterize_func instead of running twice, during both calculate_zoom_level_stats and render_super_tile
• Moved calculate_zoom_level_stats to a serial implementation instead of dask.bag.map to reduce memory overhead that caused large renders to fail

Using this branch, I was able to successfully render the OSM 1 billion point dataset. The elapsed times for each level are below.

Level	Super Tile Count	Calculating Stats (s)	Rendering (s)	Total Time (s)
0	1	46.70	16.39	63.08
1	1	42.98	9.47	52.45
2	1	43.79	8.82	52.61
3	1	41.64	9.68	51.32
4	1	41.99	11.97	53.97
5	4	65.52	25.96	91.48
6	16	155.82	81.61	237.43
7	64	515.14	314.10	829.25
8	256	2130.61	737.57	2868.18
9	1024	8646.66	1456.87	10103.54

[brendancol]

@hokieg3n1us Great PR and thanks for the benchmarks. Looks like we just need to get CI passing then we can merge.

As a side note, I've forked tiles.py into xarray-spatial along with a bunch of the analysis tools from the spatial and geo modules. The plan is to continue development of analysis related tools there and have datashader focus on the rasterization pipeline. @jbednar can comment more on this.

@jbednar I think we should merge (once CI passes) and then I will patch https://github.com/makepath/xarray-spatial/blob/master/xrspatial/tiles.py

[hokieg3n1us]

@brendancol @jbednar

I dug into the CI failures, starting by examining the version differences being pulled in by the environment.yml between the last successful build and the build associated with this PR, since I couldn't see any code change to Datashader that'd link to the failures we're seeing.

I believe I've narrowed the cause to changes to Matplotlib present in this PR.

I'm not sure what the proper solution is, but I can confirm that by manually setting my version of Matplotlib to 3.1.3 instead of 3.2.0, the tests passed successfully.

[philippjfr]

I think we should merge and deal with the CI failures separately. I'll push up a PR for those tomorrow.

[jbednar]

Thanks!!