* Magnum 4000 caches @ 1997-10-23 17:05 Ralf Baechle 1997-10-23 17:21 ` William J. Earl 0 siblings, 1 reply; 8+ messages in thread From: Ralf Baechle @ 1997-10-23 17:05 UTC (permalink / raw) To: linux Hi, maybe one of the old Mips people can answer me two questions: - why do the Magnum 4000PC instruction and data caches have a different linesize for the L1 caches, but the Magnum 4000SC not? - does the Magnum 4000SC use a split instruction/data L2 cache? TIA, Ralf ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: Magnum 4000 caches 1997-10-23 17:05 Magnum 4000 caches Ralf Baechle @ 1997-10-23 17:21 ` William J. Earl 1997-10-23 17:56 ` Ralf Baechle 0 siblings, 1 reply; 8+ messages in thread From: William J. Earl @ 1997-10-23 17:21 UTC (permalink / raw) To: Ralf Baechle; +Cc: linux Ralf Baechle writes: > Hi, > > maybe one of the old Mips people can answer me two questions: > > - why do the Magnum 4000PC instruction and data caches have a different > linesize for the L1 caches, but the Magnum 4000SC not? It turned out that, at least for some workloads, the 4000PC worked better that way. (If I remember correctly, a 32-byte linesize is better for the icache, since a 4-instruction basic block is unusual.) However, not all boxes use the same linesize values, because there were hardware bugs with at least some of the memory controllers which were affected by the choice of linesize. I don't remember the details anymore, although I might find them in my mail archivies. I have a 4000PC system (32-byte I, 16-byte D, MCT version 3) and two 4000SC systems (16-byte I, D, and S, MCT version 2) still online. I think that MCT version 2 would not support a 32-byte line. > - does the Magnum 4000SC use a split instruction/data L2 cache? No, it has a unified writeback L2 cache. ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: Magnum 4000 caches 1997-10-23 17:21 ` William J. Earl @ 1997-10-23 17:56 ` Ralf Baechle 1997-10-23 17:56 ` Ralf Baechle 1997-10-23 18:08 ` William J. Earl 0 siblings, 2 replies; 8+ messages in thread From: Ralf Baechle @ 1997-10-23 17:56 UTC (permalink / raw) To: William J. Earl; +Cc: ralf, linux > It turned out that, at least for some workloads, the 4000PC > worked better that way. (If I remember correctly, a 32-byte linesize > is better for the icache, since a 4-instruction basic block is > unusual.) However, not all boxes use the same linesize values, > because there were hardware bugs with at least some of the memory > controllers which were affected by the choice of linesize. I don't > remember the details anymore, although I might find them in my mail > archivies. I have a 4000PC system (32-byte I, 16-byte D, MCT version 3) > and two 4000SC systems (16-byte I, D, and S, MCT version 2) still online. > I think that MCT version 2 would not support a 32-byte line. So as the easy solution, as I understand things we should be safe by just leaving the linesize as the firmware chooses them for us Magnums? I actually intended to experiment with the linesize on R4k but as things look now this isn't a good thing. Btw, seems we'll get an interesting new target for Linux. It's currently ftp.uni-erlangen.de and one of the admins who might do the job, a Linux/68k hacker told me it's a 3 x R6000 box with 256mb of RAM currently running ES/IX something like that ... Ralf ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: Magnum 4000 caches 1997-10-23 17:56 ` Ralf Baechle @ 1997-10-23 17:56 ` Ralf Baechle 1997-10-23 18:08 ` William J. Earl 1 sibling, 0 replies; 8+ messages in thread From: Ralf Baechle @ 1997-10-23 17:56 UTC (permalink / raw) To: William J. Earl; +Cc: ralf, linux > It turned out that, at least for some workloads, the 4000PC > worked better that way. (If I remember correctly, a 32-byte linesize > is better for the icache, since a 4-instruction basic block is > unusual.) However, not all boxes use the same linesize values, > because there were hardware bugs with at least some of the memory > controllers which were affected by the choice of linesize. I don't > remember the details anymore, although I might find them in my mail > archivies. I have a 4000PC system (32-byte I, 16-byte D, MCT version 3) > and two 4000SC systems (16-byte I, D, and S, MCT version 2) still online. > I think that MCT version 2 would not support a 32-byte line. So as the easy solution, as I understand things we should be safe by just leaving the linesize as the firmware chooses them for us Magnums? I actually intended to experiment with the linesize on R4k but as things look now this isn't a good thing. Btw, seems we'll get an interesting new target for Linux. It's currently ftp.uni-erlangen.de and one of the admins who might do the job, a Linux/68k hacker told me it's a 3 x R6000 box with 256mb of RAM currently running ES/IX something like that ... Ralf ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: Magnum 4000 caches 1997-10-23 17:56 ` Ralf Baechle 1997-10-23 17:56 ` Ralf Baechle @ 1997-10-23 18:08 ` William J. Earl 1997-10-23 18:23 ` Ralf Baechle 1 sibling, 1 reply; 8+ messages in thread From: William J. Earl @ 1997-10-23 18:08 UTC (permalink / raw) To: Ralf Baechle; +Cc: linux Ralf Baechle writes: > > It turned out that, at least for some workloads, the 4000PC > > worked better that way. (If I remember correctly, a 32-byte linesize > > is better for the icache, since a 4-instruction basic block is > > unusual.) However, not all boxes use the same linesize values, > > because there were hardware bugs with at least some of the memory > > controllers which were affected by the choice of linesize. I don't > > remember the details anymore, although I might find them in my mail > > archivies. I have a 4000PC system (32-byte I, 16-byte D, MCT version 3) > > and two 4000SC systems (16-byte I, D, and S, MCT version 2) still online. > > I think that MCT version 2 would not support a 32-byte line. > > So as the easy solution, as I understand things we should be safe by > just leaving the linesize as the firmware chooses them for us Magnums? > I actually intended to experiment with the linesize on R4k but as > things look now this isn't a good thing. Yes, I think you should leave the linesize alone, at least on the MCT version 2. You could try alternate linesize values on the MCT version 3. The cache control code should use the linesize values from $config. By the way, I have found that it is generally faster on the R4000 (and especially on later processors) to compute such values from $config each time than to load them from variables in memory. The amortized cost of the extra cache misses can easily outweigh the extra instructions needed to recompute the value (since the latter are really very few, if you code the sequence cleverly in assembly language). Similarly, if you need to test for processor type, it is usually faster to fetch and decode $prid than to use a variable in memory. > Btw, seems we'll get an interesting new target for Linux. It's currently > ftp.uni-erlangen.de and one of the admins who might do the job, a Linux/68k > hacker told me it's a 3 x R6000 box with 256mb of RAM currently running > ES/IX something like that ... The R6000 is a real challenge. It is substantially different from the R3000 and R4000. In particular, it uses part of the cache as a sort of second level TLB. ES/IX is the CDC version of MP RISC/os. (MIPS worked on MP RISC/os, but never shipped it to customers; CDC took it over as a product.) The I/O is fairly straightforward, being mostly VME devices, but it does include I/O address mapping hardware which must be configured. ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: Magnum 4000 caches 1997-10-23 18:08 ` William J. Earl @ 1997-10-23 18:23 ` Ralf Baechle 1997-10-23 18:23 ` Ralf Baechle 1997-10-23 18:47 ` William J. Earl 0 siblings, 2 replies; 8+ messages in thread From: Ralf Baechle @ 1997-10-23 18:23 UTC (permalink / raw) To: William J. Earl; +Cc: ralf, linux > Yes, I think you should leave the linesize alone, at least on > the MCT version 2. You could try alternate linesize values on the MCT > version 3. The cache control code should use the linesize values from > $config. By the way, I have found that it is generally faster on the R4000 (and > especially on later processors) to compute such values from $config each > time than to load them from variables in memory. The amortized cost > of the extra cache misses can easily outweigh the extra instructions needed > to recompute the value (since the latter are really very few, if you > code the sequence cleverly in assembly language). Similarly, if you need > to test for processor type, it is usually faster to fetch and decode > $prid than to use a variable in memory. Indeed, Linux could need a little polish in that respect ... > The R6000 is a real challenge. It is substantially different > from the R3000 and R4000. In particular, it uses part of the cache > as a sort of second level TLB. ES/IX is the CDC version of MP RISC/os. > (MIPS worked on MP RISC/os, but never shipped it to customers; CDC took > it over as a product.) The I/O is fairly straightforward, being mostly > VME devices, but it does include I/O address mapping hardware which must > be configured. Is the information in newer releases of Kane's MIPS bible sufficient to do the R6000 part of the job? I guess the company (BIT ???) that made the R6000 is no longer around, so it would make sense if Roman'd ask them for a manual ... Ralf ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: Magnum 4000 caches 1997-10-23 18:23 ` Ralf Baechle @ 1997-10-23 18:23 ` Ralf Baechle 1997-10-23 18:47 ` William J. Earl 1 sibling, 0 replies; 8+ messages in thread From: Ralf Baechle @ 1997-10-23 18:23 UTC (permalink / raw) To: William J. Earl; +Cc: ralf, linux > Yes, I think you should leave the linesize alone, at least on > the MCT version 2. You could try alternate linesize values on the MCT > version 3. The cache control code should use the linesize values from > $config. By the way, I have found that it is generally faster on the R4000 (and > especially on later processors) to compute such values from $config each > time than to load them from variables in memory. The amortized cost > of the extra cache misses can easily outweigh the extra instructions needed > to recompute the value (since the latter are really very few, if you > code the sequence cleverly in assembly language). Similarly, if you need > to test for processor type, it is usually faster to fetch and decode > $prid than to use a variable in memory. Indeed, Linux could need a little polish in that respect ... > The R6000 is a real challenge. It is substantially different > from the R3000 and R4000. In particular, it uses part of the cache > as a sort of second level TLB. ES/IX is the CDC version of MP RISC/os. > (MIPS worked on MP RISC/os, but never shipped it to customers; CDC took > it over as a product.) The I/O is fairly straightforward, being mostly > VME devices, but it does include I/O address mapping hardware which must > be configured. Is the information in newer releases of Kane's MIPS bible sufficient to do the R6000 part of the job? I guess the company (BIT ???) that made the R6000 is no longer around, so it would make sense if Roman'd ask them for a manual ... Ralf ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: Magnum 4000 caches 1997-10-23 18:23 ` Ralf Baechle 1997-10-23 18:23 ` Ralf Baechle @ 1997-10-23 18:47 ` William J. Earl 1 sibling, 0 replies; 8+ messages in thread From: William J. Earl @ 1997-10-23 18:47 UTC (permalink / raw) To: Ralf Baechle; +Cc: linux Ralf Baechle writes: ... > > The R6000 is a real challenge. It is substantially different > > from the R3000 and R4000. In particular, it uses part of the cache > > as a sort of second level TLB. ES/IX is the CDC version of MP RISC/os. > > (MIPS worked on MP RISC/os, but never shipped it to customers; CDC took > > it over as a product.) The I/O is fairly straightforward, being mostly > > VME devices, but it does include I/O address mapping hardware which must > > be configured. > > Is the information in newer releases of Kane's MIPS bible sufficient to > do the R6000 part of the job? I guess the company (BIT ???) that made the > R6000 is no longer around, so it would make sense if Roman'd ask them > for a manual ... BIT made the parts originally, but we later got the parts from NEC. The latter were faster, more reliable, and much cheaper. CDC might still provide the manuals. As far as I know, the R6000 is not well documented in the regular MIPS documentation. I might, however, be able to get you some information about processor issues. ^ permalink raw reply [flat|nested] 8+ messages in thread
end of thread, other threads:[~1997-10-23 18:49 UTC | newest] Thread overview: 8+ messages (download: mbox.gz follow: Atom feed -- links below jump to the message on this page -- 1997-10-23 17:05 Magnum 4000 caches Ralf Baechle 1997-10-23 17:21 ` William J. Earl 1997-10-23 17:56 ` Ralf Baechle 1997-10-23 17:56 ` Ralf Baechle 1997-10-23 18:08 ` William J. Earl 1997-10-23 18:23 ` Ralf Baechle 1997-10-23 18:23 ` Ralf Baechle 1997-10-23 18:47 ` William J. Earl
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox