The floating point encode step is a tricky bit to get right. Issue #10 shows that numerical stability issues can make a big difference in some edge cases.

Encoding each unique floating point value as an element of a 4098 square matrix should require only a few hundred of these matrices. This should be doable in a reasonable amount of time.

References

• WebGL Insights
• What Every Computer Scientist Should Know About Floating-Point Arithmetic
• Floating Point Guide: Error Propagation
• Random ASCII: Floating Point
• IEEE 754 Converter