diff options
| author | Torbjorn Granlund <tg@gmplib.org> | 2017-07-21 18:32:31 +0200 |
|---|---|---|
| committer | Torbjorn Granlund <tg@gmplib.org> | 2017-07-21 18:32:31 +0200 |
| commit | 2ca58a400fbee53ccae13fad241a346e4e43e21a (patch) | |
| tree | 5173ddce0a84608912a5c53e7f4311dcebe28b45 /tune | |
| parent | 9b13801dce562250c91304c3dd2d8574a0a73449 (diff) | |
| download | gmp-2ca58a400fbee53ccae13fad241a346e4e43e21a.tar.gz | |
Measure mpn_sbpi1_bdiv_r.
Diffstat (limited to 'tune')
| -rw-r--r-- | tune/common.c | 10 | ||||
| -rw-r--r-- | tune/speed.c | 1 | ||||
| -rw-r--r-- | tune/speed.h | 41 |
3 files changed, 50 insertions, 2 deletions
diff --git a/tune/common.c b/tune/common.c index 60b29648b..3c1f8df44 100644 --- a/tune/common.c +++ b/tune/common.c @@ -225,8 +225,9 @@ speed_measure (double (*fun) (struct speed_params *s), struct speed_params *s) fprintf (stderr, "speed_measure() could not get %d results within %.1f%%\n", e, (TOLERANCE-1.0)*100.0); fprintf (stderr, " unsorted sorted\n"); - fprintf (stderr, " %.12f %.12f is about 0.5%%\n", - t_unsorted[0]*(TOLERANCE-1.0), t[0]*(TOLERANCE-1.0)); + fprintf (stderr, " %.12f %.12f is about %.1f%%\n", + t_unsorted[0]*(TOLERANCE-1.0), t[0]*(TOLERANCE-1.0), + 100*(TOLERANCE-1.0)); for (i = 0; i < numberof (t); i++) fprintf (stderr, " %.09f %.09f\n", t_unsorted[i], t[i]); @@ -895,6 +896,11 @@ speed_mpn_dcpi1_bdiv_q (struct speed_params *s) SPEED_ROUTINE_MPN_PI1_BDIV_Q (mpn_dcpi1_bdiv_q); } double +speed_mpn_sbpi1_bdiv_r (struct speed_params *s) +{ + SPEED_ROUTINE_MPN_PI1_BDIV_R (mpn_sbpi1_bdiv_r); +} +double speed_mpn_mu_bdiv_q (struct speed_params *s) { SPEED_ROUTINE_MPN_MU_BDIV_Q (mpn_mu_bdiv_q, mpn_mu_bdiv_q_itch); diff --git a/tune/speed.c b/tune/speed.c index d7bec7afc..b1c2cd44e 100644 --- a/tune/speed.c +++ b/tune/speed.c @@ -382,6 +382,7 @@ const struct routine_t { { "mpn_dcpi1_bdiv_qr", speed_mpn_dcpi1_bdiv_qr }, { "mpn_sbpi1_bdiv_q", speed_mpn_sbpi1_bdiv_q }, { "mpn_dcpi1_bdiv_q", speed_mpn_dcpi1_bdiv_q }, + { "mpn_sbpi1_bdiv_r", speed_mpn_sbpi1_bdiv_r }, { "mpn_broot", speed_mpn_broot, FLAG_R }, { "mpn_broot_invm1", speed_mpn_broot_invm1, FLAG_R }, diff --git a/tune/speed.h b/tune/speed.h index c8c01093c..6ea000b4c 100644 --- a/tune/speed.h +++ b/tune/speed.h @@ -299,6 +299,7 @@ double speed_mpn_sbpi1_bdiv_qr (struct speed_params *); double speed_mpn_dcpi1_bdiv_qr (struct speed_params *); double speed_mpn_sbpi1_bdiv_q (struct speed_params *); double speed_mpn_dcpi1_bdiv_q (struct speed_params *); +double speed_mpn_sbpi1_bdiv_r (struct speed_params *); double speed_mpn_mu_bdiv_q (struct speed_params *); double speed_mpn_mu_bdiv_qr (struct speed_params *); double speed_mpn_broot (struct speed_params *); @@ -2033,6 +2034,46 @@ int speed_routine_count_zeros_setup (struct speed_params *, mp_ptr, int, int); TMP_FREE; \ return t; \ } +#define SPEED_ROUTINE_MPN_PI1_BDIV_R(function) \ + { \ + unsigned i; \ + mp_ptr dp, tp, ap; \ + mp_limb_t inv; \ + double t; \ + TMP_DECL; \ + \ + SPEED_RESTRICT_COND (s->size >= 1); \ + \ + TMP_MARK; \ + SPEED_TMP_ALLOC_LIMBS (ap, 2*s->size, s->align_xp); \ + SPEED_TMP_ALLOC_LIMBS (dp, s->size, s->align_yp); \ + SPEED_TMP_ALLOC_LIMBS (tp, 2*s->size, s->align_wp2); \ + \ + MPN_COPY (ap, s->xp, s->size); \ + MPN_COPY (ap+s->size, s->xp, s->size); \ + \ + /* divisor must be odd */ \ + MPN_COPY (dp, s->yp, s->size); \ + dp[0] |= 1; \ + binvert_limb (inv, dp[0]); \ + inv = -inv; \ + \ + speed_operand_src (s, ap, 2*s->size); \ + speed_operand_dst (s, tp, 2*s->size); \ + speed_operand_src (s, dp, s->size); \ + speed_cache_fill (s); \ + \ + speed_starttime (); \ + i = s->reps; \ + do { \ + MPN_COPY (tp, ap, 2*s->size); \ + function (tp, 2*s->size, dp, s->size, inv); \ + } while (--i != 0); \ + t = speed_endtime (); \ + \ + TMP_FREE; \ + return t; \ + } #define SPEED_ROUTINE_MPN_MU_BDIV_Q(function,itchfn) \ { \ unsigned i; \ |