It appears PR #112 (needed for compatibility with WINDOWS) has also introduced instability and crash when rendering large image esp. with supersampling. From first investigations, the solution seems to be: close all numpy memaps right after access, even in "r+" mode.
As a data point, rendering a 3200 x 3200 image with 5x5 supersampling crashes on my machine.
It appears PR #112 (needed for compatibility with WINDOWS) has also introduced instability and crash when rendering large image esp. with supersampling. From first investigations, the solution seems to be: close all numpy memaps right after access, even in "r+" mode.
As a data point, rendering a 3200 x 3200 image with 5x5 supersampling crashes on my machine.