Re: [Bug-apl] Performance optimisations: Results

[Juergen Sauermann]

Hi,

one more thing is proper .configure (see README-2-configure).

In particular VALUE_CHECK_WANTED=no and ASSERT_LEVEL_WANTED=0,
otherwise you get a sequential time component proportional to the result size.

/// Jürgen


On 03/14/2014 07:56 PM, David Lamkins wrote:

Hmm, I hadn't thought of synchronization. That'd suggest that OpenMP infers the presence of shared state.

I'll refresh the patch for you very soon. Basically, though, the only difference should be another pragma for the loop in check_value().

On Mar 14, 2014 10:22 AM, "Juergen Sauermann" <address@hidden> wrote:

Hi David,

sounds like something is wrong. OpenMP states that their parallisation overhead is
4000 or so cycles which is much less than your measurements. Maybe some
unintended sync between the threads?

Could you send me the latest patch?

/// Juergen