kirchner@informatik.uni-kl.de (Reinhard Kirchner) (05/29/91)
Hello, the trouble with FP discussed under IEEE Floating Point Standard subject is exactly the reason why the people from the Kulisch group at Karlsruhe University developed their theory, algorithms, language extensions etc. The problem when going with these things to market is that people want to compute fast, not accurate. The Kulisch people deliver accurate, numerically verified results, or their programs tell you the problem is not solvable, which is far better than random numbers. There are products based on these things in the market ( I tell some names when asked, don't want to be called a commercial poster ), but these are selling poor simply because people want to compute FAST. The verfiying algorithms need 2 to 3 times as long ( I remember, we should ask Karlsruhe for exact numbers ) and this is too long for many. ( Nevertheless I would prefer all nuklear power plants being computed with accurate arithmetic ). These things are, obviously, based on interval computations. Using IEEE machines for this helps, but it would be nice to have the rounding mode in the opcode, since operations with different rounding follow immediately. So changing the mode needs at least one additional instruction between the FP instructions. What is also needed and is not in IEEE is a exact dot product, a dot product with NO intermediate rounding. This is ( or was ) on some machines in microcode, but not in hardware. We should have this on supercomputers! Sun could be convinced to put at least 64*64 bit => 128 bit FP multiplication into their architecture, this helps a lot in building such a dot product in software. So far for this time, my students are waiting. Send mail or post to this group for more info. Reinhard Kirchner Univ. of Kaiserslautern kirchner@uklirb.informatik.uni-kl.de ( I worked many years for Kulisch and still have connections, so I know a lot about their work )