Profiling floating point value ranges for reconfigurable implementation

Reconfigurable architectures offer potential for performance enhancement by specializing the implementation of floating-point arithmetic. This paper presents FloatWatch, a dynamic execution profiling tool designed to identify where an application can benefit from reduced precision or reduced range in floating-point computations. FloatWatch operates on x86 binaries, and generates a profile output file recording, for each instruction and line of source code, the overall range of floating-point values, the bucketised sub-ranges of values, and the maximum difference between 64-bit and 32-bit executions. We present results from the tool on a suite of four benchmark codes. Our tool indicates potential performance loss due to denormal values, and helps to identify opportunities for using dual fixed-point arithmetic representation which has proved effective for reconfigurable designs. Our results show that applications often have highly modal value distributions, offering promise for aggressive floating-point arithmetic optimisations.