* Hard ware long dooble operations are slower.
@ 1999-12-30 2:14 M.V.S. Phani Narasimham
1999-12-31 22:24 ` M.V.S. Phani Narasimham
0 siblings, 1 reply; 2+ messages in thread
From: M.V.S. Phani Narasimham @ 1999-12-30 2:14 UTC (permalink / raw)
To: gcc-help
hi,
Why quad floatine point instructions of sparcv9 are very slower,
compared to the instructions which make call to the library and
perform the operations.
The following c code takes more time when compiled with cc1 of gcc
with option -mhard-quad-float than when compiled with sun cc.
-----------------------------------------------------------
c code:-
1 main(){
2 long double a=2e0;
3 int i;
4 for(i=0;i<10000000;i++)
5 a=a+2;
6 printf("%Lfg\n",a);
7 }
The c code generates an faddq instruction.
----------------------------------------------------------------------
Time taken details:-
real 4m36.96s
user 0m23.33s
sys 1m8.59s
-------------------------------------------------------------------
If Sun cc is used it makes a call to the library (_Qp_add_)
and the details of the time are.
real 0m14.77s
user 0m5.01s
sys 0m0.01s
------------------------------------------------------------------
I could not see any trap handler that get's invoked when the hard quad
instructions are generated. Is it that they are implimented in the system
software or firmware. If so why is it slow.
phani.
-----------------------------------------------------------------------
M.V.S. Phani Narasimham
Global R & D Solutions
1 st Block,7 th main road,No:8
Kormangla
BANGLORE-36 , India
Ph- 5530053 ext.1053
E-MAIL- phani.sesha@wipro.com
www.wipro.com
The World's First SEI CMM LEVEL 5 Software Services Company
-----------------------------------------------------------------------
^ permalink raw reply [flat|nested] 2+ messages in thread
* Hard ware long dooble operations are slower.
1999-12-30 2:14 Hard ware long dooble operations are slower M.V.S. Phani Narasimham
@ 1999-12-31 22:24 ` M.V.S. Phani Narasimham
0 siblings, 0 replies; 2+ messages in thread
From: M.V.S. Phani Narasimham @ 1999-12-31 22:24 UTC (permalink / raw)
To: gcc-help
hi,
Why quad floatine point instructions of sparcv9 are very slower,
compared to the instructions which make call to the library and
perform the operations.
The following c code takes more time when compiled with cc1 of gcc
with option -mhard-quad-float than when compiled with sun cc.
-----------------------------------------------------------
c code:-
1 main(){
2 long double a=2e0;
3 int i;
4 for(i=0;i<10000000;i++)
5 a=a+2;
6 printf("%Lfg\n",a);
7 }
The c code generates an faddq instruction.
----------------------------------------------------------------------
Time taken details:-
real 4m36.96s
user 0m23.33s
sys 1m8.59s
-------------------------------------------------------------------
If Sun cc is used it makes a call to the library (_Qp_add_)
and the details of the time are.
real 0m14.77s
user 0m5.01s
sys 0m0.01s
------------------------------------------------------------------
I could not see any trap handler that get's invoked when the hard quad
instructions are generated. Is it that they are implimented in the system
software or firmware. If so why is it slow.
phani.
-----------------------------------------------------------------------
M.V.S. Phani Narasimham
Global R & D Solutions
1 st Block,7 th main road,No:8
Kormangla
BANGLORE-36 , India
Ph- 5530053 ext.1053
E-MAIL- phani.sesha@wipro.com
www.wipro.com
The World's First SEI CMM LEVEL 5 Software Services Company
-----------------------------------------------------------------------
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~1999-12-31 22:24 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
1999-12-30 2:14 Hard ware long dooble operations are slower M.V.S. Phani Narasimham
1999-12-31 22:24 ` M.V.S. Phani Narasimham
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).