[Gems-users] simulating large instruction window/rob sizes

[Miray Kas <miraykas@xxxxxxxxx>]

Hi everyone,

 

I have a question on simulating large rob/iwin sizes.

In opal's config folder, there are a bunch of configurations but none of them model extremely large parameters.

 

For instance, in std-512-L1_64K_1-L2_4M_4 following settings are given:

#include ./config/width-standard.txt
#include ./config/register-512.txt
#include ./config/alu-standard.txt
#include ./config/l1-64B-64K-1.txt
#include ./config/l2-64B-4M-4.txt
#include ./config/memconf-standard.txt
#include ./config/yags-standard.txt
#include ./config/ras-standard.txt
#include ./config/indirect-standard.txt

IWINDOW_WIN_SIZE: 512
PSEQ_SLIDING_WINSIZE: 64

To model large window effect, I assume IWINDOW_ROB_SIZE and registers should also be scaled accordingly.

 

So, if we set the IWINDOW_ROB_SIZE to be equal to IWINDOW_WIN_SIZE and change the included register-512 file to be something like this:

 

// physical = 160 (logical) + 512 (rename)
CONFIG_IREG_PHYSICAL: 672

// physical = 64 (logical) + 1024
// note: must be < 256
CONFIG_FPREG_PHYSICAL: 1086

// physical = 5 (logical) + 512
CONFIG_CCREG_PHYSICAL: 517
 

the simulation fails some assertions. I tried to do this kind of scaling taking register-64 and register-128 configuration files as example. But if I dont mess with the register configurations, the simulation runs normally...

 

This is the output I got:

 

Where it fails assertion-->Used slots: 511
Where it fails assertion-->Rob size:512
ASSERT_FAILURE_CYCLE[ 3785 ]
simics-common: system/pseq.h:733: bool pseq_t::isSlotAvailable(int): Assertion `(m_iwin[k].getNumSlotsTaken() == IWINDOW_ROB_SIZE)' failed.

I did not make any changes in the code, except setting robsize equal to iwindow size and adding two lines for assertion print out and everything else is MOSI_SMP_bcast with default parameters. What would be the proper way of scaling registers and any other parameters (that I potentially might be missing) to be able to model large rob effect ?

 

Many thanks,

Miray