VFP in Wince 6.0

Hello, Atul,

What do you mean float to integer truncation?

Is that some kind of casting from float to integer?

I just tested with the following loop.

    double test = 0.0;
    for (int i=0; i<50000; i++)
    {
        test = tan(cos(3.4578236482) + sin(83.9374658) / asin(0.123)) * log(36.123);
        test = (int)log(fabs(test)) * (int)pow(fabs(test), 0.578);
    }

But I don't get any errors about that.

Thank you very much,

Richard

And did you get expected results? Try something simple:

double myDoubleA = 48.4;

double myDoubleB = 48.7;

int myIntA = (int) myDoubleA;

int myIntB - (int) myDoubleB;

Run this with and without VFP and see what values you get for integer variables.

This is just one example. As I said, we have seen failures in some CETK tests and in 1 Graphics demo (others work fine) when we had enabled VFP

Atul

Thanks Atul.

I can see that the result is different.

(VFP disabled) myIntA = 48
(VFP disabled) myIntB = 48

(VFP enabled) myIntA = 48
(VFP enabled) myIntB = 49

For the CETK tests, I got 4 fails running with the DirectDraw Test after enabled the VFP.
But, as it is a minor fault, I would just ignore it.

Would you give me more examples of the FP arithmetic operation error?
Or is there any other FP arithmetic operation will cause errors?
I would like to know all of the errors because the performance of floating calculation is more important for me.

Thanks again.

Richard

I thought the main issue was that the NEON and VFP registers were not stored and restored correctly on context switches. If that is the case then I would expect unpredictable random errors across the board. There have been a few posting on this issue.

I too would like to see this feature (VFP) working under CE 6 (I know it works under CE7). I know there is not enough registers allocated and you technically can not recompile the MS source but I thought there there was a QFE awhile back that might have address this issue. I could be wrong. This would be a common problem for all ARM processors that use this core, not just TI's. I have seen other CE6 BSPs claim a working fix but I do not know the details.

Sorry we don't have a consolidated list of all possible errors. We provide generic BSP to cater to a wide community and therefore we disabled VFP. You can certainly enable it on your platform if it works for all scenarios/applications you are targeting

thanks

Atul

David

Regarding your question about NEON and VFP, MSFT has provided a fix for saving/restoring co-proc registers only. AFAIK, they havent added support for VFPv3 in the CE6 kernel. if you have other information, please let us know.

-Madhvi

Some more background:

http://www.mpc-data.co.uk/docs/MPC-WP02VectorFloatingPointsupportforWindowsEmbeddedCE60.pdf

Looks like this implementation was done for the i.mx31 which is probably a VFPv2.

I understand the VFPv3 (A8) has the added register thus we are out of luck.