Author: Robert Hyatt
Date: 13:17:37 02/08/06
Go up one level in this thread
On February 07, 2006 at 16:52:27, h.g.muller wrote: >OK, I already discovered my mistake, see other message. > >But about the Cray vs PC: sure I can't get that memory bandwidth out of my PC, >but the point is that I don't NEED such memory bandwidth on a PC. The code on >the Cray was vectorized by breaking it into loops that only did a single >operation and you had to stream the results to and from memory between >operations. The PC works best if you do as much as possible in a loop: combine >all loops that run in the same direction through your data, so that, once data >is brought in the cache, you chew it as much as you can before it is flushed. >Most algorithms allow you to do some 10 operations on the data while it resides >in L1, and that reduces the required bandwidth by a factor of 10 compared to >what the Cray would need for the same computation. And when you need large >strides you simply work on a number of colums simulaneously to make sure you >process the entire cache line. There are a couple of high-performance computing books that spend a lot of time talking about "blocked" algorithms. Which do just what you are describing. Access a cache-sized block of data and beat on it until done, then go to next block, rather than thrashing cache by running the loops from start to end in their entirety... Makes the code look ugly and awkward as hell, but it runs on "cachey" systems...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.