Software Tuning, Performance Optimization & Platform Monitoring
Discussion regarding monitoring and software tuning methodologies, Performance Monitoring Unit (PMU) of Intel microprocessors, and platform updating.

Counting UOPS dispatched to execution ports on Haswell

BGoel
Beginner
811 Views

I am trying to count the UOPS dispatched to execution ports on Haswell platform using the hardware performance counters. The relevant performance event for this is UOPS_EXECUTED_PORT with individual umasks for counting executions on PORT_0 to PORT_7. This event works fine for counting UOPS dispatched to individual ports.

But I will like to count UOPS dispatched to non-memory ports (Port 0, 1, 5, 6) using one counter but I am not sure how to do it. On Nehalem, there was a specific umask for this: UOPS_EXECUTED:PORT015 but this is not listed for Haswell. I tried to combine umasks for Port_0(0x01), Port_1(0x02), Port_5(0x20) and Port_6(0x40) into one event as UOPS_EXECUTED_PORT:0x63 but the value provided by this event doesn't match the summation of executions on individual ports. 

Does anyone know if it is possible?

0 Kudos
3 Replies
McCalpinJohn
Honored Contributor III
811 Views

The UOPS_EXECUTED_PORT.PORT_* events for Haswell carry the description:

Cycles which a uop is dispatched on port * in this thread.

Combining the umasks for different ports should count cycles in which a uop is dispatched on any one of the requested ports -- not the sum of the uops requested on those ports.  

Based on this description as a "cycle count" event, it does not appear that you can get the sum with a single event -- you will need to count separately on the four ports and add the results.

0 Kudos
BGoel
Beginner
811 Views

Oh dang! You are right. I missed that piece of information. Thanks.

0 Kudos
BGoel
Beginner
811 Views

I just noticed that the event description changed after Nehalem. On Nehalem, the relevant event was UOPS_EXECUTED:PORT* and description was "Counts the number of Uops executed on port *". On both Sandy Bridge (UOPS_DISPATCHED_PORT.PORT_*) and Haswell (UOPS_EXECUTED_PORT.PORT_*), the description changed to cycles in which Uop is dispatched to port. That was partly why I was confused.

I guess I have to stick to using separate counters for counting these events.

0 Kudos
Reply