tree 965ea25d84b36e106c38cd05ed13f0ffcf9ef650
parent 11276da5473a47c1269e386d46ccf6183226d97f
author Ben Vanik <ben.vanik@gmail.com> 1691083138 -0700
committer GitHub <noreply@github.com> 1691083138 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJky+GCCRBK7hj4Ov3rIwAATZkIAIqdmfAGY8SgM/PCFSWFrD5H
 SyB0eEnyDq+vDwMtw8ExV71yKDl+F3rjgFAxlcoYt8nm/2PAl3Yh1lrFsPvyCp4l
 ZZEgM3b107dIM7uT0Ji/OFEKo3kPKirkbjj/7eOGpUNd0nwmesJJWJVRW7u7YkZf
 w45X3dOqh42HGJQAX2LFCN8dpoYtq37ZoYVte9kdIt4d+7tyxEuth3NMfhkZnF2O
 s8Pi1COXSkk8SD9X2QbL3aEoQJr8soC1yEW1TrzwHmb15VQzGzfW2W2q/FZOl7lG
 uc5ShFXc3d4QiEvxGm8drrMowY/lq/yaoh+IWlfFjW5K07lfTcTcmsHf0ra4VEw=
 =dqvs
 -----END PGP SIGNATURE-----
 

Adding `--module_mode=mmap` tooling flag. (#14563)

This uses the platform memory mapping support - if available - to map
VMFB file contents into memory on-demand. This can skew profiling
results as it introduces both high warm-up costs and high per-iteration
variance as the OS is free to discard/reload pages at any time. It can
be useful for very large models though that otherwise exceed available
physical memory. On certain HAL targets this can avoid additional wired
memory but most (GPUs) will still require fully loading the contents
into memory and mapping will be less efficient but be friendlier to host
resource usage.

This isn't a great API but is good enough for our tools for now. I still
plan on exposing proper memory objects but that'll likely happen as part
of sparse residency/file queue operations in the HAL.