Float vs Double performance.

ben_wizards

I was playing around the other day and I was curious about a couple things. I changed all of the variables and constants from floats to doubles and got a HUGE (like 5-10%) cpu benefit. I was able to add some functionality because of the extra CPU availability. Ram usage didn't go up noticeably either. Why would this be?

giuliomoro

I do not know see why this should be the case. Actually this ia exactly the opposite of what I would expect.
double computations should be more than twice as expensive than float, because they are not supported in hardware.
What type of operations were you doing?
Could you provide some code snippets?

ben_wizards

giuliomoro ,
Latest build report(cherry picked with the neon-math build)-much less CPU usage, much more stable and returning everything back to float from double give a significant CPU usage drop.

So...Never mind?

ben_wizards

I can send you my code, I can't replicate it outside of my program right now, but I tested it by getting rid of all of the .f extensions and swapping all floats to doubles.

I'm doing lots of sine calculations, using sinf_neon. I've noticed the biggest discrepancy when using the scope.

I haven't updated to the latest scope build yet. I'll do that and report back.