mersenneforum.org AMD64 opcode timings
 User Name Remember Me? Password
 Register FAQ Search Today's Posts Mark Forums Read

 2005-01-16, 04:04 #1 Prime95 P90 years forever!     Aug 2002 Yeehaw, FL 748310 Posts AMD64 opcode timings Does anyone know of AMD documents that give the instruction latencies and throughput for different opcodes? Right now, I'm mostly interested in div, idiv, mul, imul. I'm working on the trial factoring rewrite for x86-64.
 2005-01-17, 00:26 #2 gdf   Aug 2003 10102 Posts This one ? http://www.amd.com/us-en/assets/cont...docs/25112.PDF Chapter C has latency tables. gdf
 2005-01-17, 14:48 #3 Prime95 P90 years forever!     Aug 2002 Yeehaw, FL 7·1,069 Posts That's the one. Thanks, I missed it browsing AMD's web site. 71 clocks for division, 4 for a multiply. Looks like I need to go to great lengths to avoid divsions!
 2005-01-18, 14:55 #4 Jeff Gilchrist     Jun 2003 Ottawa, Canada 3×17×23 Posts Wow. So you cannot divide and conquer this problem but instead must go forth and multiply.
 2005-01-18, 16:30 #5 ThomRuley     May 2003 23×31 Posts If it would be so inefficient to divide, would it be more efficient to simply subtract when performing the modulo step?
2005-01-18, 21:00   #6
ewmayer
2ω=0

Sep 2002
República de California

7×11×151 Posts

Quote:
 Originally Posted by Jeff Gilchrist Wow. So you cannot divide and conquer this problem but instead must go forth and multiply.
...at which point the 2 snakes turned around and replied to Noah: "We can't - we're adders."

(One bad pun deserves another.)

The best solution for the extremely-long-latency divides is probably the standard Newtonian iterative inverse.

 2005-01-19, 02:35 #7 cheesehead     "Richard B. Woods" Aug 2002 Wisconsin USA 22·3·641 Posts Reminds me of the ole' IBM 1620, which could divide a 66,000-digit number by a 33,000-digit number, but it would take ... hours ( on the order of 20 microseconds times number of digits in dividend times number of digits in divisor )
2005-01-19, 14:25   #8
xilman
Bamboozled!

"𒉺𒌌𒇷𒆷𒀭"
May 2003
Down not across

13·821 Posts

Quote:
 Originally Posted by ewmayer ...at which point the 2 snakes turned around and replied to Noah: "We can't - we're adders." (One bad pun deserves another.)
Noah then stomped off, muttering under his breath.

A year later, he did another census. Everything else around had multiplied manyfold, but he couldn't see the vipers anywhere. "Damn things have probably gone extinct", he muttered and went off to collect some fuel for his stove. He pulled some branches off the woodpile and was astounded to see in there dozens of tiny snakes squirming around the two adults. "How did you do it?" he exclaimed. "It was easy once we'd thought about it", the male viper replied, "we used logs."

Paul

 2005-01-26, 13:22 #9 CraigB   Jan 2005 Cape Town, ZA 2 Posts Please mail me on craigb@mailme.co.za so i can put you in contact with AMD's main software optimiser. He is interested in assisting you with writing optimised code for AMD64. He is a REALLY rare resourse to have access to so please dont miss this opertunity as he has next week available to speak to you.
 2005-01-26, 23:40 #10 Prime95 P90 years forever!     Aug 2002 Yeehaw, FL 7×1,069 Posts I just sent you a private message
 2005-01-27, 15:53 #11 CraigB   Jan 2005 Cape Town, ZA 28 Posts Thanks & ditto :-)

 Similar Threads Thread Thread Starter Forum Replies Last Post nuggetprime Software 5 2011-02-21 08:28 axn Operation Billion Digits 1 2009-02-06 16:14 ewmayer Factoring 7 2008-12-11 22:12 paulunderwood 3*2^n-1 Search 14 2008-04-17 22:27 MooooMoo Riesel Prime Search 2 2007-10-11 08:56

All times are UTC. The time now is 14:02.

Thu May 6 14:02:26 UTC 2021 up 28 days, 8:43, 1 user, load averages: 1.57, 2.01, 2.07