[comp.sys.hp] Sun f77 -fswitch overhead. Comparison with HP9000/350.

irf@kuling.UUCP (Bo Thide) (08/22/88)

>From sun-spots Vol 6: Issue 172:

  >>From:    Jim Barby <jabarby@watvlsi.waterloo.edu>
  >Subject: Re: Revenge of the 68881
  >Reference: v6n152

  >In v6152 Dave Bernholdt was surprised about the high overhead -fswitch
  >has.  I did some bench marking a few weeks back and sent off a request to

I ran the ratfor program on my HP9000/350 (where -fswitch is +bfpa) which
has the same hardware as the Sun3/260.  As is clear from the results below,
the HP +bfpa also has a high overhead.  But I was surprised to see how
much faster code the HP f77 produces -- almost twice as fast as Sun's!
Note that I am not using the new, faster (and cheaper, sic) HP FPA.
(I did'n run the test without the write statement since currently HP f77 only
has an assembly code optimizer.)

Comments?

-Bo

>>> Bo Thide', Swedish Institute of Space Physics, S-755 91 Uppsala, Sweden <<<
>>> Phone: (+46) 18-403000.  Telex: 76036 (IRFUPP S).  Fax: (+46) 18-403100 <<<
>>>    INTERNET: bt@irfu.uu.se     UUCP: ...enea!kuling!irfu!bt <<<

==== The posted ratfor program (I changed 0 to 6 in write statemnt - BT) ====

define N 10000000
double precision data(2), add, sub, mul, div
integer i

data(2) = 1

do i=2,N {
	data(1) = data(2)
	data(2) = i
	add = data(2) + data(1)
	sub = data(2) - data(1)
	mul = data(2) * data(1)
	div = data(2) / data(1)
}

write(6,*) add, sub, mul, div

stop
end

------------------------------------------------------------------------

========hardware configuration=========================================

	node name	short name	equipment
	---------------+---------------+------------------------------
	watvlach	vlach		older 3/50 with new MC68881
	vlsisun1	sun1		older 3/160 with old MC68881
	vlsisun2	sun2		3/160 with new MC68881 and FPA
	vlsisun3	sun3		3/260 with MC68881

	node name	short name	equipment
	---------------+---------------+---------------------------------
	heating.irfu    hp350		HP9000/350 with 68881 and old FPA


========test results with and without -O compile flag==================

	cpu	cpu times in minutes without using -O
	name	-fsoft	-f68881	-ffpa	-fswitch
	-------+-------+-------+-------+--------
	vlach	56.0	14.0	----	20.1
	sun1	52.6	10.1	----	20.8
	sun2	52.8	12.2	5.1	10.2
	sun3	29.4	10.3	----	13.5

	name	+M    	(68881)	+ffpa	+bfpa
	-------+-------+-------+-------+--------
	hp350   15:28    9:23   3:57     5:25


	cpu	cpu times in minutes using -O
	name	-fsoft	-f68881	-ffpa	-fswitch
	-------+-------+-------+-------+--------
	vlach	56.7	14.2	----	21.3
	sun1	53.3	15.4	----	22.3
	sun2	53.4	12.5	2.0	10.8
	sun3	29.8	10.4	----	14.0

	name	+M    	(68881)	+ffpa	+bfpa
	-------+-------+-------+-------+--------
	hp350   15:30    4:39   1:44     4:30

-- 
>>> Bo Thide', Swedish Institute of Space Physics, S-755 90 Uppsala, Sweden <<<  Phone (+46) 18-300020.  Telex: 76036 (IRFUPP S).  UUCP: ..enea!kuling!irfu!bt