compile with hardware float

Hi.

I’ve bench CC and GCC in order to see the quality of the produced code. I’ve realised the same test on A BSD server-3.3 Ghz XEON (only to check the difference).
Most of the time GCC in a little bit better than CC.

For Integer test, CC and GCC produce almost the same code quality and the RaspBerry Pi is 2 or 3 times slower than the XEON (it’s quite normal)
But for floating operation, that’s not the same. GCC is 2 time faster than CC. But the Intel gcc compiler is up to 100 times faster on BSD.

My conclusion therefore is that GCC and CC do not use Hardware floating point operation under RISC OS !
I’ve a look at the compiler options but I did not find anything !

Is there any way to do that ?

thank’s

CC only knows how to use the old FPA floating point instruction set, which means all floating point instructions will actually be emulated via the FPEmulator module (very slow).

GCC supports FPA, softfloat, and VFP. IIRC FPA is the only option if you’re linking against the shared C library. Unixlib defaults to softlfloat (which is faster than FPEmulator). VFP must be specifically enabled, since it’s only supported by the newer CPUs:

• -mfpu=vfp will select VFPv2 and automatically switch the CPU to ARM11 – producing code suitable for the Raspberry Pi 1 and later
• -mfpu=neon will select VFPv3 + NEON and automatically switch the CPU to Cortex-A8. The code won’t be compatible with the Pi 1, but it’ll work on all the other modern machines that have NEON.
• -mfpu=neon -funsafe-math-optimizations (and maybe also an optimisation flag like -O3) must be used if you want floating point operations to be automatically vectorised into NEON code – NEON isn’t fully IEEE compliant so you need to tell the compiler that you’re OK with the non-compliance.

Thank you Jeffrey for informations !

VPf is not very fast, -mfpu=neon improve the speed of factor 3.5 far far away from Intel float performance.
Neon should be better. More registers and faster operations, but stranglely -funsafe-math-optimizations do not change anything !

It does not matter. I will just avoid using float :)

I’m not sure if it is still the case, but it is also worth remembering that the Pi isn’t the fastest VFP implementation. Certainly the numbers from the BASIC VFP thread were significantly lower that the (same-ball park clock-speed) i.MX6 system I tested on (basically Pi was roughly half the Titanium, and i.MX6 was closer to Titanium than to Pi).

It has generally been a “rule of thumb” (pardon the bad pun) not to use float with RISC OS due to historical deficiencies and ARM CPUs being stronger with integer maths. Intel CPUs, by contast, have always been floating point kings (it tended to be an area of strength over their rivals at AMD, even when the wind wasn’t blowing in their direction in other areas). As such, you’re probably looking at the weakest part of the Pi, and comparing to the strongest aspect of the Xeon.