2011/2/24 Jordi Gutiérrez Hermoso
<address@hidden>
On 24 February 2011 08:18, Richard Crozier <
address@hidden> wrote:
> As another example where JIT would be useful, consider the ODE solver
> routines. Your function is evaluated at each time step chosen sequentially
> by the solver, and cannot therefore be vectorized.
JIT, but in an ODE solver your bottleneck would not be in the
iterations the interpreter has to interpret, but in the computations
that happen at each iteration?
I think the idea was that JIT could make the "computations that happen at each iteration" faster not that the JIT makes the solver itself faster. If the solver was the bottleneck, a native oct implementation would probably be preferable to JIT anyway. This would be one of those cases where the types should be easy to annotate at the first solver step.
Again a lot of this is hand waving and speculation on both sides because we don't have a profiler. I wouldn't make a final decision on whether JIT was useful based solely on a hunch. Besides, a clever JIT doen't only have to be about reducing execution time, it could also optimize memory use through transparent chunking and auto-inlining.
--judd