Differences from matplotlib

fastplot aims to produce the figure that matplotlib and cartopy would. The test suite draws the same figures with both libraries and compares axis limits, ticks, labels, legend and layout geometry, and pixels. The differences below remain, either by design or because matching them would cost more speed than they are worth.

Maps

Topic	Behaviour in fastplot
Earth model	Projections use a sphere. cartopy uses the WGS84 ellipsoid for most of them, so projected x/y values in metres differ slightly. Data, coastlines and gridlines all go through the same projection, so a map is consistent with itself.
Extents on conic maps	set_extent uses the exact bounds of the curved outline. cartopy approximates them, so the aspect ratio of a Lambert map can differ by about 0.3%.
Gridline labels	Placed on the map's four edges and always horizontal. cartopy rotates them along curved gridlines on conic and polar maps.
Contour labels	Placed with matplotlib's rules, on fastplot's own lines, so a label can sit slightly further along a line.
Line ends	Map lines use round caps and are not snapped to pixels.
Thin barbs	Barbs drawn well under 1 pt wide come out slightly lighter.

Plain axes

Topic	Behaviour in fastplot
Text	Glyphs are rasterised from the font outlines by fastplot, not by Agg with FreeType's autohinter. Positions and metrics match; individual pixels differ, which accounts for most of the 0.2 to 1% of pixels that differ on a typical plot.
Rotated text	Text at angles that are not multiples of 90° is placed exactly, but its glyphs are rotated as an image.
Line joins	Always round. matplotlib mitres patch edges, so a corner pixel can differ.
Markers	Scatter markers are positioned to 1/8 of a pixel.
Dense lines	Many segments of one line crossing within a single pixel come out slightly lighter than Agg draws them.

Polar axes

Only full circles are supported; set_thetamin, set_thetamax and set_thetalim raise NotImplementedError. Artists are clipped to the axes' bounding box rather than to the circle.

Output formats

PDF, SVG and EPS files contain the rendered image, losslessly at the savefig dpi, rather than vector paths.

Threads

Drawing figures from several threads at once needs numba's OpenMP or TBB threading layer, which the standard wheels for Linux and Windows include. numba installed with pip on Apple silicon has only the workqueue layer, which cannot run parallel code from two threads at the same time. fastplot detects this and takes a lock, so concurrent drawing is safe there but runs one figure at a time.

Floating point

fastplot's arithmetic gives identical results on every platform. numpy and matplotlib are compiled with fused multiply-add instructions on Apple silicon, so their own results differ in the last bit between platforms. The tests allow for this where they compare against them.