README

This is a conservative autoscheduler that compute_root most Funcs except for the trivial ones (think of it as a -O1 optimizer for Halide). It recognizes large reduction patterns and use rfactor or atomic to parallelize on associative reduction when there's not enough parallelism in the pure variable domain. This strategy works reasonably well for gradient pipelines, and is suitable as a default option for decent but not optimal performance. This is also currently the only autoscheduler that generates GPU schedules.

Running some benchmarks in the app directory gives the following statistics (all use halide_reuse_device_allocations(nullptr, true) for GPU)

app	manual (CPU)	gradient-autoscheduler (CPU)	manual (GPU)	gradient-autoscheduler (GPU)
bilateral filter	7.93 ms	12.92 ms	0.29 ms	1.05 ms
camera_pipe	8823.33 us	25126 us	605.03 us	3347.44 us
lens_blur	7.77 ms	22.41 ms	0.73 ms	5.60 ms
local_laplacian	42.29 ms	128.31 ms	0.81 ms	14.30 ms
nl_means	145.003 ms	out-of-memory	N/A	82.93 ms
conv_layer	15.46 ms	6.89 ms	N/A	1.90 ms
stencil_chain	18.86 ms	21.46 ms	N/A	6.35 ms

Tested on a 8 core Intel CPU (16 with HT) and TITAN Xp.

See test/autoschedulers/li2018 for examples of using this autoscheduler.