RLibm is a math library that provides correctly rounded result for all inputs. Currently, RLibm supports a number of elementary functions for bfloat16, posit16, and float representations. rlibm is generated using the technique described from this paper.
To compile the math library, please follow the instructions below. This compilation instruction conrates separate math library for each of the available representations.
If you want to compile the math library for posit16, you have to install SoftPosit. Please follow the instructions from the SoftPosit GitLab.
- Clone the rlibm repository
git clone https://github.com/rutgers-apl/rlibm.git
- Create an environment variable SOFTPOSITPATH that points to the directory of SoftPosit:
export SOFTPOSITPATH=<path-to-softposit-directory>
- Build the math library
- If you want to build all the math libraries, simply use make rule
cd rlibm
make
- If you want to build math libraries for each representation separately, you can use these make rule
cd rlibm
make bfloat16mlib
make posit16mlib
make floatmlib
- (Optional) Test that the math library does produce the correct value. This step requires MPFR installed.
- To run the correctness bench suite for all math libraries, run
./runLibTest.sh
- The correctness bench suite for math library is located in the libtest folder.
cd libtest/bfloat16
make
./runAll.sh
The math library will be located in the lib directory:
- bfloat16MathLib.a : math library for bfloat16
- floatMathLib.a : math library for float
- posit16MathLib.a : math library for posit16.
The header files for each library is located in the include directory:
- bfloat16_math.hpp : header for bfloat16 math library
- float_math.h : header for float math library
- posit16_math.h : header for posit16 math library
If you want to use the bfloat16 math library, you need to also include bfloat16.hpp which constains our custom bfloat16 class.
You can use our library in the code similar to how standard math library is used, except our function names start with "rlibm_":
test.cpp:
#include "float_math.h"
int main() {
float result = rlibm_cospi(1.5f);
return 0;
}
To build the program, include the math library in the compilation command:
g++ test.cpp ../../lib/floatMathLib.a -lm -o test
Currently, RLibm uses some functions from the default math library for range reduction, such as to decompose a floating point value into the integral part and fractional part.