Re: broadwell cycle_activity unit mask

On 06/11/2015 02:46 PM, Andi Kleen wrote:
>> Shouldn't that also be using a name default rather than a numerical value?
> 
> Yes it should be using a named default.
> 
> BTW I wish we would just get rid of the defaults. There's really no
> such concept in the original event lists and often it makes very
> little sense. Most of the times my script just picked the first.
> 
> -Andi
> 

Hi Andi,

The unit_mask use on the processor hardware has morphed.  On Intel hardware they are now used more in conjunction with the event number to select a particular event.  The unit mask as a bit mask to select whether to count reads or writes in memory references is pretty much non-existent in newer Intel processors.  Maybe in a future round the Intel events should be written using mandatory unit masks.

In the mean time which named unit mask makes the most sense as a default?

Looking at Table 19-3.  Non-Architectural Performance Events In the Processor Core of 4th Generation Intel® Core™ Processors of 
http://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32-architectures-software-developer-vol-3b-part-2-manual.pdf for A3H event there seems to be a lot more entries in events/i386/broadwell/unit_masks than in the table and some of those seem to have the same bit patterns.  Are those name aliases intentional (for example cycles_l2_pending and cycles_l2_miss)?  Also I don't see entries in the manual table matching stalls_ldm_pending or cycles_no_execute of the unit_masks.

-Will

------------------------------------------------------------------------------