✨ [FEATURE REQUEST] Hardware Accelerated Square Root

[SizzinSeal]

Requested Feature

Hardware accelerated square root through. The Cortex A9 on the V5 Brain has instructions to compute the square root of floating-point numbers through the VSQRT instruction.

Current Implementation

Currently, GCC uses the sqrt implementation of libm which does not make use of this instruction.

Motivation

Computing the square root of a floating-point number is an expensive operation. With the use of more expensive algorithms like live-regenerating motion profiling and Monte Carlo Localization on the rise in VRC, it makes sense to get the V5 to perform as well as it possibly can. The better it performs, the less PROS users have to worry about performance limitations.

Features of VSQRT

• Supports single-precision and double-precision floating-point numbers
• Can produce "Invalid Operation" or "Inexact" exceptions (e.g, square root of -1, square root of 1 Googol)

Why doesn't GCC use this instruction automatically?

libm uses errno. Yup, that simple. GCC can't inline VSQRT because it does not set errno. This would be solved with -fno-math-errno but this is a bad idea because libraries like Eigen may check errno. Instead, libm functions would need to be overridden.

Possible implementations

We can override the sqrt provided by libm with a custom implementation that uses the VSQRT instruction, but also sets errno.

[djava]

You're asking to set -fno-math-errno in PROS? No.

[Exortions]

You're asking to set -fno-math-errno in PROS? No.

Why not? Does pros use errno in a lot of places?

[SizzinSeal]

PROS doesn't but something like Eigen might. I've discussed this with @djava and he's convinced me this is not the route to take

[Exortions]

👍

[SizzinSeal]

@Rocky14683 has this been completed? I didn't see a commit to PROS which implements this

[Rocky14683]

i thought u said its not the route to take

[SizzinSeal]

Sorry, I should have been more specific. While we can't use -fno-math-errno, we can implement our own sqrt operation which uses the VSQRT instruction, but also sets errno, so behavior is still the same as previously, just much faster

[ion098]

Re @SizzinSeal

While we can't use -fno-math-errno, we can implement our own sqrt operation which uses the VSQRT instruction, but also sets errno, so behavior is still the same as previously, just much faster

Doesn't newlib already do this? https://github.com/bminor/newlib/blob/363357c023ce01e936bdaedf0f479292a8fa4e0f/newlib/libm/machine/arm/e_sqrt.c#L30-L41

[SizzinSeal]

Perhaps it does after the libm update PR. I've only analyzed binaries created using PROS 4.1.1, I'll find out