LabVIEW
Help with algorithm?
Azzazel,
I agree with chilly charly that you should definitely initialize the shift registers, else results from previous calculations will leak through.
I don't think you can expect miracles, but it seem to me that combining the two arrays into a complex array speeds things up by about 30% on my PIII. I've attached a crude modification to show the approach, modify as needed (it migh even contain bugs. Check the result with a known good dataset).
For more consistent timing results, you should really isolate the calculation part from the other stuff. In your example, there is no guarantee that the random generator for the second array executes before or after the first time stamp. Same on the right side. You might include the FP updates of some of the indicators (=slow!) before the second timestamp. See my example for a possibility.
Of course YMMV 😉
I agree with chilly charly that you should definitely initialize the shift registers, else results from previous calculations will leak through.
I don't think you can expect miracles, but it seem to me that combining the two arrays into a complex array speeds things up by about 30% on my PIII. I've attached a crude modification to show the approach, modify as needed (it migh even contain bugs. Check the result with a known good dataset).
For more consistent timing results, you should really isolate the calculation part from the other stuff. In your example, there is no guarantee that the random generator for the second array executes before or after the first time stamp. Same on the right side. You might include the FP updates of some of the indicators (=slow!) before the second timestamp. See my example for a possibility.
Of course YMMV 😉
OK, I quickly looked at it again over my morning coffee and there are two improvements that give at least 20% combined speedup.
(1) Major: Change the indexing in the inner loop to avoid the second indexing of the rotated array(see attached).
(2) A minor improvement seems to be gained if the inverse tangent is done after the transposition (saving one transpose operation).
Since these arrays are only for display, you might ask yourself if it is sufficient to do all calculations in single precision (SGL).
On my rig, it gives another 40% speedup and you still get about 6 good significant digits. This is plenty for e.g. 16bit grayscale! 🙂
The attached modification implements the above mentioned two improvements but is also set to SGL. It can easily be switched back to DBL by changing the representation of the single diagram constant (zero) in the outer loop and the representation of the array indicators.
Combined, the attached solution in SGL mode is about 2.25x faster that my earlier solution above. I am sure quite a few more things could be tweaked to improve the speed.
(1) Major: Change the indexing in the inner loop to avoid the second indexing of the rotated array(see attached).
(2) A minor improvement seems to be gained if the inverse tangent is done after the transposition (saving one transpose operation).
Since these arrays are only for display, you might ask yourself if it is sufficient to do all calculations in single precision (SGL).
On my rig, it gives another 40% speedup and you still get about 6 good significant digits. This is plenty for e.g. 16bit grayscale! 🙂
The attached modification implements the above mentioned two improvements but is also set to SGL. It can easily be switched back to DBL by changing the representation of the single diagram constant (zero) in the outer loop and the representation of the array indicators.
Combined, the attached solution in SGL mode is about 2.25x faster that my earlier solution above. I am sure quite a few more things could be tweaked to improve the speed.
