To clarify, I set opLat=24 and issueLat=12 for FloatSqrt in
FuncUnitConfig.py in order to fix the problem.
-Rick
Rick Strong wrote:
I think you may be right about the latency. I went into
src/cpu/o3/FuncUnitConfig.py and set the latency of FloatSqrt to 12,
and it appears to work. I hope that points someone in the right
direction and towards a solution.
-Rick
Kevin Lim wrote:
Hi Rick,
There may be an issue with the CPU thinking there's no activity left
due to the long latency of the instruction, and then not waking up
properly when the instruction completes. I'm not sure why it'd be on
that instruction in particular though. I'm going to try your example
and see if I can get some more information about why it might be
doing this. This may take me some time as I don't have a clean up to
date tree at the moment. Could you send me your exact command line,
as well as any diffs on any of the configuration files you made?
In the meantime, try this: in your configuration file where you
select the O3CPU (most likely Simulation.py if you're just using our
default configs), try setting the parameter "activity" to 1. This
should (hopefully) prevent the O3CPU from going to sleep if it's
inactive. If the program runs properly as a result, then this is
most likely the culprit. You may get very minor statistics
differences if you set this to 1.
Kevin
Rick Strong wrote:
Hi all,
I have been tracking down a problem with resuming a few benchmarks
which include (fma3d, ammp, apsi, mgrid, and others) at certain
checkpoints. The culprit appears to be a sqrtt (FloatSqrt) operation
that gets stuck at the head of the queue and stalls the detailed
model (O3CPU). If I run the benchmark in AtomicSimpleCPU there is no
problem. I have confirmed that sqrtt is at the head of the queue in
four samples (two in ammp and two in apsi). The simulation ends with
("Exiting @ cycle 9223372036854775807 because simulate() limit
reached.") . The same symptoms are seen in other benchmarks
suffering a similar phenomena and I guess that sqrtt may also be the
culprit. If anyone has some insight on what might be happening, it
would be appreciated.
-Rick
_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
