Why does OpenBLAS not support sofftp with ARMv7

[jainanshul]

As per #363 (comment) softfp is not supported with ARMv7. I tested it on the device and I do get a segmentation fault. Wondering why is softfp only supported for ARMv5?

[brada4]

ARMv5 is C-only. You can compile it with exact -march flag for your architecture to get optimal code. ARMv7 kernels gets all enhancements from use of VFP which you dont have. Also you must measure if libm or gcc builtins are faster (-fno-builtins vs -fbuiltins). Probably it is best to avoid double precision alltogether as one double precision instruction will be emulated in 30-40 int32 instructions.

[xianyi]

For ARMv7, we write hard fp ABI in assembly kernels. If we support softfp ABI, we must change these assembly manually. It can be done but lots of work.

[jainanshul]

Thanks for your replies. I change the target to ARMV5 and I don't see any crash with softfp. However the performance is abysmal 27secs to load and predict using caffe with softp openblas vs 7secs with hardfp openblas.

Does OpenBLAS support 64bit ARM architecture?

[xianyi]

The ARMv5 is naive C kernels. Therefore, it is very slow.

We support ARM 64bit including Cortex-A57.

[brada4]

Yes, performance is abysmal as your CPU lacks floating point support and everything is done in integer ALU. 4 instructions to emulate floating point instruction is very very reasonable.

You can try -fbuiltins - maybe GCC can optimize something, but will not get better than 3x slowdown.

Example: http://www.jhauser.us/arithmetic/SoftFloat.html find 32bit mul in integer c. 4 instructions according to my brain's optimistic C compiler (or 3 if you really really dont do anything about NaNs.

[xianyi]

@brada4 , there are three FP ABI mode soft, softfp, and hard.

The soft uses software FPU (as you mentioned) and soft calling convention.
The softfp uses hardware FPU and compatible soft calling convention.
The hard uses hardware FPU and different calling convention. The hard calling convention is passing the floating point argument by FP registers.

OpenBLAS uses hard FP ABI mode. It conflicts with soft and softfp.

[brada4]

...which is a call to libm which may or may not use FPU which may or may not exist.

[brada4]

Actually "does not support" claim is now void as per @jainanshul
It is hard to tell why is it so slow without pieces of code. softfp slows down all calls involving floating point arguments. Depending of GCC flags you chose it might be much more or much less of such calls. Without code or compiler arguments - there is no issue.

[jainanshul]

@brada4 well openBLAS crashes with ARMv7 and softfp support so I would still say openBLAS doesn't support softfp with ARMv7.

[brada4]

ARMv7 is name of Debian architecture, i.e hardfp
ARMv5 means any ARM, just like Prescott means Any x86.

[lygstate]

What's the point of ARMV6?

[martin-frbg]

What's the point of ARMV6?

Older models of Raspberry Pi and similar (support was added in early 2013) ?

[lygstate]

Does it also use the hard ABI?

[martin-frbg]

I'm not that experienced with ARM coding, but the assembly used in the v6-specific routines under kernel/arm look suspiciously similar to the v7 counterparts (and both were written by the same developer), so probably yes.

[epicstar]

@xianyi what is the current timeline for this feature to be complete? Thanks for working on this btw.

[martin-frbg]

Expected to be resolved through #1221, so closing