Status of double precision

[misho]

I came to a point in my project (spaceflight dynamics) where I really need to work with double precision. I have spacecraft orbiting around the planet, and when I apply thrusters to manoeuvre among them, I see them moving in discrete "jumps", which I can only attribute to lack of precision. Their orbital positions are to the order of 300 km above earth surface which itself has 6371 km radius (total size of the space "box" roughly 6700km), and their movements should be calculated to within 0.001 of a meter, which I think is at the limit of a single precision calculations.

I found a few posts where users are asking about the status of double precision functionality. I am able to compile my own libraries, and I tried compiling newton.lib using _NEWTON_USE_DOUBLE, but that only gives numerous errors across other components. Fixing this is something I would definitely not dare to embark on

So - is there any hope of implementation of double precision in the latest Newton Dynamics library package? Thanks!

Misho

[Julio Jerez]

let me complete the vehicle and I will add the double precision, I estimated by the end of this week.

[misho]

Thank you so much, Julio!

Misho

[pHySiQuE]

I'm probably about a year away from implementing this, but I look forward to it.

BTW, the vegetation stuff I wrote with Newton is on the cover of Game Engine Gems 3. The technique is described in the chapter I wrote:
http://www.amazon.com/dp/1498755658/?tag=terathon-20

[misho]

Amazing work, sir - but I have to admit that I don't have the book... how does vegetation employ Newton Dynamics?

Misho

[misho]

Hi Julio, et al.

I was wondering if you had a chance looking at the double precision over the weekend? Please let me know!

Thanks,
Misho

[Julio Jerez]

still working on it, it will be by next Saturday.

[misho]

Good stuff! Thanks!

Misho

[misho]

Just so I understand,

You will enable the current github Newton library to use

Code: Select all

#define dFloat double

when a preprocessor define _NEWTON_USE_DOUBLE is used, correct?

Misho

[Julio Jerez]

the problem is that I have to review that dgVector and dgMatrix, there has being some function added that where no duplicated. also there are some funtions that take implicit dgFloat64 and some dgFloat32 so I have to put #ifdef around those so that only the double is active
I will also make the dgVector SSE2
if this works I mat decide to make the double the default. now that all platforms support double there is not reason why no move to that mode only.
It save lot of shenanigan with precision lost, specially in the collision code.
I am wrapping the vehicle and then I will do that.

If is goes well Newton 3,15 will be double precision only.

[misho]

Excellent, thanks!

Misho

[JoeJ]

If is goes well Newton 3,15 will be double precision only.

"only" - meaning removing single precision support?
Halfing performance and doubling memory even for cases where additional precision isn't necessary?

[Julio Jerez]

yes maybe not a good idea yet.

But I am finding that the performance difference is that that great, the memory is a real concern, but with memory almost ten time what is used to be just for of five year ago maybe this is not a big issue either.
The biggest issue is the bandwidth, this is what in fact make the code slower, since moving memory around has not increased in he hardware side.
so maybe you are right and keeping the float and double is still a good idea.

One possible solution could be having all calculation use double by memory saving using floats.
this is still an idea on the air, nothing definite yet

[Stucuk]

Wouldn't you need to keep single precision for support with Phones/etc? Even if most PC's won't have issues i would bet alot of phones use old hardware.

[Julio Jerez]

I think all phone device use double precision now. but like I said maybe that was not a good idea.