floating-point support

[oberstet]

Hello,

I'm checking out some trivial floating-point stuff and get weird behaviour. Feeling like an idiot, could s.o. pls help?

A)

In Emulator (3rd Ed FP1 SDK), the code at the end outputs

1.1
HAL says that we do NOT have a VFP unit

Both I did expect.

Ok. Now, if I change the initial assignment of var "a" to

double a = .0;

the output changes to

0.0
HAL says that we do NOT have a VFP unit

wtf? I did expect 1.0 !

Now, if I change type to

float a = .0;

the output gets 1.0 as expected.

The same behaviour on E51 device.

Whats going on?

Could it be, that some SW emulation layer for floating point
isn't working with "double"?

I rechecked the behaviour multiple times, since I can't
believe .. or am I fooling myself somehow?


B)
Since the E51 is built on a Freescale MXC300-30 SoC which
seems to have an ARM1136JF-S core (where the F stands for
VFP hardware floating-point), I did expect the HAL:Get to
indicate support for hardware floating point.

However, it does not. Why?

I also compiled everything both with and without the

armfpu vfpv2

in the MMP - does that even apply to GCCE?


Thanks,
Tobias

================

#include <iostream>
#include <hal.h>

// This is a GCCE toolchain workaround needed when compiling with GCCE
// and using main() entry point
#ifdef __GCCE__
#include <staticlibinit_gcce.h>
#endif

int main ()
{
double a = .1;
for (int i = 0; i < 10; ++i) a += .1;

std::cout << a << std::endl;

// adapted from: http://www.symbian.com/developer/tec...ort.guide.html

TInt supportedFp;

TInt HalVfp = HAL::Get(HALData::EHardwareFloatingPoint, supportedFp);
if (HalVfp == KErrNotSupported || supportedFp == EFpTypeNone)
{
std::cout << "HAL says that we do NOT have a VFP unit\n";
}
else
{
std::cout << "HAL says that we have a VFP unit\n";
}

std::cout << "Press a character to exit!";
getchar ();

return 0;
}

[Sorcery-ltd]

Hi,

Ignore HAL, that will just be returning a constant which may or may not have been set correctly on any given platform.

In the emulator floating point is used as normal (and by default), so it is purely the compiler that is determining your results.

Code for the target hardware is compiled without VFP support and is software emulated (very slow for floats and doubles). It is possible to force the compiler and it will work on appropriate hardware - if you search the discussion boards you'll find a very detailed post with the solution (I'm pretty sure the poster is called storsjo although they didn't start the thread, but just searching for FPU or floating point should find it).

Sorcery

[oberstet]

Thx! I've found the thread:
http://discussion.forum.nokia.com/fo...ght=fpu&page=2

If I set

VFP2MODE_OPTION=-mfloat-abi=softfp -mfpu=vfp

in gcce.mk and activate HW FPU in MMP

ARMFPU VFPV2, the app won't even install.

If I set

SOFTVFPMODE_OPTION=-mfloat-abi=softfp -mfpu=vfp

without modifying MMP, the app will install, but will crash on startup with:

"
Main
KERN-EXEC 3
"

(Y-Task:Crash Monitor says the same ..)

This exception is even documented under

"Note, however, that binaries generated with VFP specific code do not work on devices without a VFP unit - any call to VFP specific code results in a KERN-EXEC 3 panic."
http://www.symbian.com/developer/tec...ort.guide.html

So on E51, it seems to be impossible to use HW FPU.