ArmMathM0 is a ready-to-link library (GNU archive) of floating-point math for ARM Cortex-M0 and M0+ that replaces the functions provided with the compiler. Compared to the math library included with GCC, this library is both faster and smaller.
ArmMathM0 includes add, subtract, multiply, divide, and square
root for both float and double. It also has float implementations for
expf(), logf(), sinf(), cosf(), sincosf(), tanf(), asinf(),
acosf(), atanf(), and atan2f().
The plan is to build out the float functions before
implementing these functions for double. The higher functions in the
standard library are built on floating-point arithmetic, so those original
functions will go faster with ArmMathM0 arithmetic (roughly 2x).
ArmMathM0 has full support for Infinity and NAN, with support for denormal (tiny) numbers as a compile-time option. It is compatible with the ARM Run-time ABI, so it simply links in and satisifes compiler calls to floating-point arithmetic.
All arithmetic (including square root) provides the exact or closest representable result. This means the maximum error is 0.5 unit last place (ULP), which happens if the exact answer is at the midpoint between two adjacent representable values.
All transcendental functions have an error of less than 1 ULP. This means when the exact result is between two adjacent representable values, it might not choose the closest one. Accuracy for trig functions is valid for inputs up to ±255 radians.
This project was built with Microchip Studio, and the project files are included. The build folders Denormals and NoDenormals have makefiles and option settings to build their respective versions. The point of the NoDenormals build is to save on code size; there is no performance difference with normal numbers.