What's faster float or double on ARM?

[kornalius]

I'd like to know what type is faster: float or double on the ARM? My compiler uses double at this point but I'd like to speed it up a bit more by using float. Will I see a good gain?

I would also like to know if float type is enough for a compiler/interpreter use or I absolutely have to use double type variables?

Btw, my compiler uses no variable types at all. The internal of course do but the user doesn't know about it.

Regards,
Kornalius

[Presto]

I don't think you'll see any gain. Without a floating point processor, any use of float or double is going to be a bit slow. Personally, I prefer using doubles rather than floats, simply because a lot of the functions are designed for doubles (sqrt for example), and the time it would take to typecast the double results to a float would probably negate any potential gain of using the float.

That's just my take on it. I know there are a lot of people here into using fixed-point math. I try to avoid floats/doubles whenever possible, but when I do need that precision, I'm using doubles.

-John

[angedelamort]

float are 4 bytes and double 8 bytes. So the float should be faster and takes less memory. But less precision (but it's enouph).
I have made a floating point emulation class for Brew and if the ppc do the same thing, it should be faster with just float.

And for your information... I don't think casting float to double takes that much times when you call sqrt. You should avoid sqrt instead...

[mlepage]

Make a typedef so you can test between the two. That's good practice anyways because the answer may change from compiler to compiler, or on different devices with various hardware or libraries.

I personally found in one project that float was a bit faster than double. But your project could be different! There are various tradeoffs, and your project may favour some more than others (such as my project).

[kornalius]

[refractor]

On ARMs it really depends what internal format the floating point emulator (FPE) is using. I know that some use double as their internal format, so crunching a lot of floats incurs a lot of translations to/from double on the way in/out of the FPE. That generally means that floats may be processed slower than doubles.

Really, I concur with mlepage; it depends on what you're doing. Benchmark benchmark benchmark..