RCL-57

[Werner]

(05-20-2022 01:06 PM)pauln Wrote:  Interesting. Did you use a different algorithm or did you optimize the implementation?

Both ;-)
- all routines use a single dot product routine at their core. the LU-decomp of the Master Lib uses a very badly written and slow axpy (y:=y+a*x) routine.
- pivotation vector is stored in a single register
- inversion is done solving U*X*L = I (so X= inv(A) ), making max. use of the innermost dot product routine. The Master Lib calculates the inverse as inv(U)*inv(L), in three separate steps, spending quite a bit more time in loop setup and control.
- only 640 steps (I simplified the interface a little bit as well). I do use some HIR registers (2 or 3), and of course the Opcodes to add or subtract 1 from a reg, which are not used in Pgm 02.

(05-21-2022 02:13 PM)pauln Wrote:  Paradoxically, in a real device or emulator it could be a tie between your program and the one in the Master Library.

Indeed the same program will run faster if run from a module than from regular memory. I think it is because module programs are guaranteed, if coded correctly, to never run in an infinite loop, so there is no need to check for user input ("R/S") while they are running.

Seeing that there's a factor 3 difference in LU factorisation, I'm hoping to be left with some speed improvement ;-)
But, on an emulator, I would not have both 100 registers and 960 program steps, I'm afraid, and so I would only be able to test small matrices (up to 5x5 I think), where the speed difference would not be so pronounced. Unless you can 'tweak' it a little.

Cheers, Werner