Fine tuning of the evaluation on a per expression basis
Submitted by Gael Guennebaud
Assigned to Nobody
Link to original bugzilla bug (#1159)
Version: 3.3 (current stable)
Description
As motivated by bug #404 (closed), it would be useful to allow tuning evaluation parameters on a per expression basis (or group of expressions). Parameters includes at least:
-
vectorization: ON/OFF/enforce packet size
-
multi-threading: ON/OFF/nb threads
-
unrolling: ON/OFF/limit
This could be addressed through a "device" similar to what is done in Tensor, or by the definition of an evaluation context: