[Libre-soc-dev] SVP64 Vectorised add-carry => big int add

[lkcl]

On Mon, Apr 18, 2022 at 4:32 PM lkcl <luke.leighton at gmail.com> wrote:

> any clues as to why?

this is about the oddest multiply operation i could imagine... but it works.
whhyyyy? how could it be turned into a proper double-width
mul-and-subtract? what information needs to be transferred
from the first loop to the second?  is it as simple as checking
if un[i+j] < (p&0xffffffff)? that would cause a carry-borrow into
phi, which could be compensated for, couldn't it?

      // double-width multiply with strange subtract only on bottom half
      for (i = 0; i < n; i++) {
        unsigned long long p = qhat*vn[i];
        plo[i] = un[i+j] - (p&0xffffffffLL);
        phi[i] = p >> 32;
      }
      for (i = 0; i < n; i++) {
         t = plo[i] - k; // subtract previous carry
         un[i+j] = (t & 0xffffffffLL);
         k = phi[i] - (t >> 32); // take top-halves for new carry
      }