GMP bignum results using double cells.

[Rob Browning <rlb@defaultvalue.org>]

OK.  I'd say that's a bit better -- guile-gmp-old is the smob based
implementation with the extra indirection.  guile-gmp has bignums that
store the mpz_t inside the 3 words of a double cell.

  $ ./run-benchmark guile-1.6 ./guile-1.7 ./guile-gmp-old ./guile-gmp foo.scm
  starting trials
  ("bench-random-op +" "guile-1.6" ((user . 206) (sys . 0) (gc . 91)))
  ("bench-random-op -" "guile-1.6" ((user . 186) (sys . 1) (gc . 59)))
  ("bench-random-op *" "guile-1.6" ((user . 226) (sys . 0) (gc . 101)))
  ("bench-random-op /" "guile-1.6" ((user . 267) (sys . 0) (gc . 116)))
  ("bench-random-op gcd" "guile-1.6" ((user . 126) (sys . 0) (gc . 46)))
  ("bench-random-op lcm" "guile-1.6" ((user . 288) (sys . 0) (gc . 121)))
  starting trials
  ("bench-random-op +" "./guile-1.7" ((user . 353) (sys . 1) (gc . 62)))
  ("bench-random-op -" "./guile-1.7" ((user . 327) (sys . 0) (gc . 62)))
  ("bench-random-op *" "./guile-1.7" ((user . 364) (sys . 0) (gc . 56)))
  ("bench-random-op /" "./guile-1.7" ((user . 392) (sys . 0) (gc . 56)))
  ("bench-random-op gcd" "./guile-1.7" ((user . 276) (sys . 0) (gc . 59)))
  ("bench-random-op lcm" "./guile-1.7" ((user . 392) (sys . 0) (gc . 57)))
  starting trials
  ("bench-random-op +" "./guile-gmp-old" ((user . 414) (sys . 0) (gc . 66)))
  ("bench-random-op -" "./guile-gmp-old" ((user . 399) (sys . 0) (gc . 65)))
  ("bench-random-op *" "./guile-gmp-old" ((user . 438) (sys . 0) (gc . 68)))
  ("bench-random-op /" "./guile-gmp-old" ((user . 434) (sys . 0) (gc . 65)))
  ("bench-random-op gcd" "./guile-gmp-old" ((user . 407) (sys . 0) (gc . 66)))
  ("bench-random-op lcm" "./guile-gmp-old" ((user . 466) (sys . 0) (gc . 67)))
  starting trials
  ("bench-random-op +" "./guile-gmp" ((user . 349) (sys . 1) (gc . 67)))
  ("bench-random-op -" "./guile-gmp" ((user . 302) (sys . 0) (gc . 32)))
  ("bench-random-op *" "./guile-gmp" ((user . 321) (sys . 0) (gc . 33)))
  ("bench-random-op /" "./guile-gmp" ((user . 334) (sys . 0) (gc . 32)))
  ("bench-random-op gcd" "./guile-gmp" ((user . 312) (sys . 0) (gc . 33)))
  ("bench-random-op lcm" "./guile-gmp" ((user . 366) (sys . 0) (gc . 32)))

It looks like the new code is substantially better than 1.7, though if
I understand correctly, and gc time is included in user time, then it
appears the improvement was on the gc side, not the computation side.

Marius, for now, rather than use SCM_CELL_ADDR_1, I mirrored the
handling of real values like this:

  #define SCM_I_BIG_MPZ(x) (((scm_t_big_mpz *) SCM2PTR (x))->mpz)
  #define SCM_BIGP(x) (!SCM_IMP (x) && SCM_TYP16 (x) == scm_tc16_big)

  typedef struct scm_t_big_mpz
  {
    SCM type;
    mpz_t mpz;
  } scm_t_big_mpz;

However, this means that numbers.h now has to include gmp.h, and with
the SCM_CELL_ADDR_1 approach, it wouldn't.  I'd be happy to arrange
things whichever way you prefer.

Also, I was trying to figure out how to add the test for 

  sizeof (mpz_t) <= 3 * (sizeof (scm_t_bits))

In order to put it in configure.in, we have to have access to the
definition of scm_t_bits (or at least SIZEOF_SCM_T_BITS) at configure
time.  Any thoughts?

-- 
Rob Browning
rlb @defaultvalue.org, @linuxdevel.com, and @debian.org
Previously @cs.utexas.edu
GPG starting 2002-11-03 = 14DD 432F AE39 534D B592  F9A0 25C8 D377 8C7E 73A4


_______________________________________________
Guile-devel mailing list
Guile-devel@gnu.org
http://mail.gnu.org/mailman/listinfo/guile-devel