3 ms·
This is really cool and I think it might be a re-discovery of https://dl.acm.org/doi/abs/10.1145/3434310 https://dl.acm.org/doi/abs/10.1145/3434310.
by adgjlsfhk1 3y ago
This is really cool and I think it might be a re-discovery of https://dl.acm.org/doi/abs/10.1145/3434310 https://dl.acm.org/doi/abs/10.1145/3434310.
- nsajko 3y agoYeah, some ideas are the same. Some notes: 1. Their LP constraint matrix is just a Vandermonde matrix according to the linked paper, while my code formulates the LP problem in a smarter way that should make the job much easier for the LP solver. I guess this improvement would be easy to adopt on their end. 2. The linked paper focuses on tiny types like BFloat16, but they mention scalability with large data types as a future goal. My focus was on larger types like IEEE-754 binary64 (double in C) from the start. 3. They account for range reduction and output compensation! This seems really nice, I hope that whatever they did can be applied to my approach to improve it.