1 Feb
2022
1 Feb
'22
12:06 a.m.
On 2022-01-31 23:57, Dave McGuire wrote:
On 1/31/22 5:18 PM, Paul Koning wrote:
Johnny says that RSX said so, for the fast machines like 11/70 and
11/44. And that makes sense, with basic instruction times under a
microsecond and Unibus accesses rather expensive due to mapping. The
software CRC takes 7 instructions per byte, all register instructions
except for one table reference. The KG11 does a bit at a time, with a
10 MHz clock if I read it right, so that fits the statement that the
operation is complete in a microsecond so software may not need to
test the Done bit depending on what machine you have. At least two
unibus cycles plus a microsecond processing time is not obviously
faster than 7 fast CPU instructions.
The KG11 dates back to the 11/20; as I said earlier, it very obviously
makes sense on that machine where software CRC would probably take
around 30 microseconds per byte.
Ok, executing mostly out of cache I can see now how that might be
faster, especially on something like an 11/70 with a screaming fast
processor but a slow Unibus interface. Interesting!
Well, the DECnet/RSX manuals explicitly state that you don't want to
enable the KG11 even if you have one, if you run on an 11/44 or 11/70.
So yeah, slow...
Johnny
--
Johnny Billquist || "I'm on a bus
|| on a psychedelic trip
email: bqt(a)softjar.se || Reading murder books
pdp is alive! || tryin' to stay hip" - B. Idol
I did an unsupported software DDCMP, but that does CRC
in software
(8 bits at a time, the classic table driven fast software
implementation). That probably outruns a KG11 and in any case I
never had access to that hardware.
The KG11 looks like all TTL logic; are you really sure it'd be
faster to do it in software running on similar-generation hardware?