search

IntelLabs / ParallelAccelerator.jl

The ParallelAccelerator package, part of the High Performance Scripting project at Intel Labs
BSD 2-Clause "Simplified" License
294 stars 32 forks source link

Fusion issue #87

Closed ehsantn closed 8 years ago

ehsantn commented 8 years ago

With the recent changes to DomainLambda and mmapInline, logistic regression no longer fuses properly. The data-parallel computation is now six loops instead of one.

ninegua commented 8 years ago

Looks like the removal of mmapInline exposes a bug in ParallelIR fusion at (the wrong order of) checking indices access pattern. It is fixed in 90f6064fbe1a349c1ab25acc924b552849258edf.

ehsantn commented 8 years ago

Thanks! I think one more loop of logistic regression is fused now than before with this fix.