home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!gatech!bloom-beacon!eru.mt.luth.se!lunic!sunic!dkuug!diku!terra
- From: terra@diku.dk (Morten Welinder)
- Newsgroups: comp.sys.ibm.pc.hardware
- Subject: Re: What speed rep/movs on a 486?
- Message-ID: <1992Aug23.115654.3988@odin.diku.dk>
- Date: 23 Aug 92 11:56:54 GMT
- References: <eyal.714307145@ise>
- Sender: terra@ran.diku.dk
- Organization: Department of Computer Science, U of Copenhagen
- Lines: 28
-
- eyal@echo.canberra.edu.au (Eyal Lebedinsky) writes:
-
- >Hello Everyone,
-
- >I am eveluating a 486/50 for myself. I have a program that measures the
- >time to rep/mov a 20KB block as bytes, words or Dwords. On my 386 the
- >measured time is accurate. I get 2.000ms which translates to 4 cycles
- >per item (this is a 386/40). Moving words/Dwords kepps the item rate so
- >the total time drops to 1.000 then 0.500.
-
- >The same program on a 486/50 produces strange results. First, the best
- >that I get for rep/movsb is 2.000ms for 20KB which is 5 cycles/item. The
- >spec says it should take 3 cycles/item. Moving words/dwords gets a worse
- >per-item result of up to 7 cycles/item. The result is not regulat enough
- >to suggest a fixed wait state. With 8 megs installed there should be
- >enough interleave to allow burst move, but it does not. Is the cache
- >getting in the way? the 256KB cache should latch the whole 20KB block
- >and work at full speed?
-
- 1) Just a reminder: When you move Dwords they must be aligned 0 mod 4.
- This is the same for a 386 but not a 386sx where aligment can be
- 2 mod 4 also.
-
- 2) Rep Movs? were not optimized in the design of the 486. An aligned loop
- (0 mod 16) might perform better.
-
- Morten Welinder
- terra@diku.dk
-