source/libs/gmp/gmp-src/mpn/cray/ieee/mul_basecase.c

   1 /* Cray PVP/IEEE mpn_mul_basecase.
   2
   3 Copyright 2000, 2001 Free Software Foundation, Inc.
   4
   5 This file is part of the GNU MP Library.
   6
   7 The GNU MP Library is free software; you can redistribute it and/or modify
   8 it under the terms of either:
   9
  10   * the GNU Lesser General Public License as published by the Free
  11     Software Foundation; either version 3 of the License, or (at your
  12     option) any later version.
  13
  14 or
  15
  16   * the GNU General Public License as published by the Free Software
  17     Foundation; either version 2 of the License, or (at your option) any
  18     later version.
  19
  20 or both in parallel, as here.
  21
  22 The GNU MP Library is distributed in the hope that it will be useful, but
  23 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  24 or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  25 for more details.
  26
  27 You should have received copies of the GNU General Public License and the
  28 GNU Lesser General Public License along with the GNU MP Library.  If not,
  29 see https://www.gnu.org/licenses/.  */
  30
  31 /* The most critical loop of this code runs at about 5 cycles/limb on a T90.
  32    That is not perfect, mainly due to vector register shortage.  */
  33
  34 #include <intrinsics.h>
  35 #include "gmp.h"
  36 #include "gmp-impl.h"
  37
  38 void
  39 mpn_mul_basecase (mp_ptr rp,
  40                   mp_srcptr up, mp_size_t un,
  41                   mp_srcptr vp, mp_size_t vn)
  42 {
  43   mp_limb_t cy[un + vn];
  44   mp_limb_t vl;
  45   mp_limb_t a, b, r, s0, s1, c0, c1;
  46   mp_size_t i, j;
  47   int more_carries;
  48
  49   for (i = 0; i < un + vn; i++)
  50     {
  51       rp[i] = 0;
  52       cy[i] = 0;
  53     }
  54
  55 #pragma _CRI novector
  56   for (j = 0; j < vn; j++)
  57     {
  58       vl = vp[j];
  59
  60       a = up[0] * vl;
  61       r = rp[j];
  62       s0 = a + r;
  63       rp[j] = s0;
  64       c0 = ((a & r) | ((a | r) & ~s0)) >> 63;
  65       cy[j] += c0;
  66
  67 #pragma _CRI ivdep
  68       for (i = 1; i < un; i++)
  69         {
  70           a = up[i] * vl;
  71           b = _int_mult_upper (up[i - 1], vl);
  72           s0 = a + b;
  73           c0 = ((a & b) | ((a | b) & ~s0)) >> 63;
  74           r = rp[j + i];
  75           s1 = s0 + r;
  76           rp[j + i] = s1;
  77           c1 = ((s0 & r) | ((s0 | r) & ~s1)) >> 63;
  78           cy[j + i] += c0 + c1;
  79         }
  80       rp[j + un] = _int_mult_upper (up[un - 1], vl);
  81     }
  82
  83   more_carries = 0;
  84 #pragma _CRI ivdep
  85   for (i = 1; i < un + vn; i++)
  86     {
  87       r = rp[i];
  88       c0 = cy[i - 1];
  89       s0 = r + c0;
  90       rp[i] = s0;
  91       c0 = (r & ~s0) >> 63;
  92       more_carries += c0;
  93     }
  94   /* If that second loop generated carry, handle that in scalar loop.  */
  95   if (more_carries)
  96     {
  97       mp_limb_t cyrec = 0;
  98       for (i = 1; i < un + vn; i++)
  99         {
 100           r = rp[i];
 101           c0 = (r < cy[i - 1]);
 102           s0 = r + cyrec;
 103           rp[i] = s0;
 104           c1 = (r & ~s0) >> 63;
 105           cyrec = c0 | c1;
 106         }
 107     }
 108 }