Robert,
On Wed, Apr 2, 2008 at 3:30 PM, Robert Richter <[EMAIL PROTECTED]> wrote:
> Stephane,
>
> thanks for your comments.
>
>
> On 27.03.08 22:59:29, stephane eranian wrote:
> > > -/*
> > > - * AMD64 specific parameters for the library
> > > - */
> > > typedef struct {
> > > - pfmlib_amd64_counter_t
> pfp_amd64_counters[PMU_AMD64_NUM_COUNTERS]; /* extended counter features
> */
> > > - uint64_t reserved[4]; /* for
> future use */
> > > + pfmlib_amd64_counter_t
> pfp_amd64_counters[PMU_AMD64_MAX_COUNTERS]; /* extended counter features */
> > > + uint32_t flags; /* use flags */
> > > + uint32_t reserved1; /* for future use */
> > > + ibsfetchctl_t ibsfetchctl; /* IBS fetch control */
> > > + ibsopctl_t ibsopctl; /* IBS execution control */
> > > + uint64_t reserved2; /* for future use */
> > > } pfmlib_amd64_input_param_t;
> > >
> >
> > I don't understand why you need to pass the full ibsfetchctl and
> > ibsopctl structures.
> > There is actually only one field that is configurable: sampling period
> > (cnt). I do not
> > account for the enable bit which has to be set anyway. So why not just
> > pass 2 uint32_t
> > values for the periods of each component of IBS and let the library
> > internals deal with
> > the actual layout. This was done this way on Itanium, for instance.
>
> You are right, for IbsOpCtl and IbsFetchCtl there are only the
> IbsFetchMaxCnt and IbsOpMaxCnt values to set. Additionally there is an
> option flag in IbsFetchCtl that enables randomization (IbsRandEn). For
> parameter passing we have to use a struct. Instead of having to
> different structures (the input arguments and the PMD registers as in
> pfmlib_ita2_ear_t and pfm_ita2_pmc_reg_t), I decided to have one
> structure for both and use the IBS register layout to pass
> parameters. So, I found it easier to simply pass a control register
> value as input parameter. Sure, the user has to know that to setup
> here, but this should not be hard to do for whom being familiar with
> IBS. Do you think it would be easier to have only a subset of those
> parameters needed for configuration (currently this are
> ibsfetchmaxcnt, ibsopmaxcnt and ibsranden)?
>
Yes, I do. No need to expose fields the user cannot influence on input.
>
> > > +/* A bit mask, meaning multiple usage types may be defined */
> > > +#define PFMLIB_AMD64_USE_IBSFETCH 1
> > > +#define PFMLIB_AMD64_USE_IBSOP 2
> > > +
> > > typedef struct {
> > > - uint64_t reserved[8]; /* for future use */
> > > + uint32_t ibsfetch_base; /* Perfmon2 base register index
> */
> > > + uint32_t ibsop_base; /* Perfmon2 base register index
> */
> > > + uint64_t reserved[7]; /* for future use */
> > > } pfmlib_amd64_output_param_t;
> >
> > Could you explain the goal of the two bases? I was envisoning a scheme
> > where in dispatch_ibs()
> > you would simply add the list of PMDs that are used. It is okay to
> > have more PMDs that PMCs.
> > The key is that the PMDs corresponding to the events must be first AND
> > in the order the events
> > where given. After that PMDS can be passed in any order.
>
> That exactly is the problem, the dispatch function sets up new PMCs
> and PMDs, but it is not clear which register of pfp_pmds[] is used for
> IBS fetch and op. There is a reordering required because of the
> counter 1:1 mapping between PMCs and PMDs. Thus, the index of the IBS
> PMDs may vary and this information can be found in the base values.
>
The rule of libpfm is that PMD related to events (hosted in counting PMD)
are ALWAYS put first in pfp_pmds[] AND in the order the events where
specified. Here is an example:
I want to measure events A, B. If event A is measured by PMD3 and B
by PMD1, then
pfp_pmd[0].reg_num = 3
pfp_pmd[1].reg_num = 1
After those PMDs, there can be more PMDs but there is no order
guarantee. Those PMDs
are related to extended features which use PMDs. The values return by
libpfm for those
are mostly informational. It is expected that programs using extended
features KNOW
which PMD to use and know how to decode their content.
-------------------------------------------------------------------------
Check out the new SourceForge.net Marketplace.
It's the best place to buy or sell services for
just about anything Open Source.
http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace
_______________________________________________
perfmon2-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
