Sparse perturbations in verificarlo

gkiar / research

This repo isn't meant to hold code, but serve as a place for issues and discussions that I can manage on a project board.

GNU General Public License v3.0

0 stars 0 forks source link

Sparse perturbations in verificarlo #2

Closed gkiar closed 3 years ago

gkiar commented 3 years ago

The goal is to instrument a set of libraries and rather than perturbing each flop within an execution, to control the frequency of the perturbations (reducing overhead while still sampling the space of plausible perturbed results). Ideally, this could be controlled via a switch in the specification of the MCA backend.

gkiar commented 3 years ago

This (roughly) should require:

[x] Adding a parameter definition for perturbation sparsity here. For exactness, this should probably be provided and stored as an integer N, and then we would compute a threshold for 1/N on load.
[x] Add a function that samples the random number generator and compares it to our 1/N.
[ ] Use that function in a new if-block here when adding noise.

gkiar commented 3 years ago

Open question: if we are performing both PB and RR perturbations, do we want to randomly sample the perturbation such that it is possible (and perhaps likely) that a given operation will only experience inexactness on either the inputs or outputs, or do we want to do this at the level of the complete operation? Essentially, when we perform MCA will we fully respect the MCA mode we're operating in, or just subsample the possible perturbations within operations for that mode. I think we want to fully respect the MCA mode (perform both PB and RR for every operation we perturb, in this case).

gkiar commented 3 years ago

https://github.com/verificarlo/verificarlo/pull/252