Analyzers
Talk to fellow users of Intel Analyzer tools (Intel VTune™ Profiler, Intel Advisor)

OMPT support

Intel_C_9
Beginner
569 Views

Hello;

Does VTune support OMPT APIs for OpenMP performance analysis?

Intel OpenMP Runtime Lib. supports OMPT APIs.

 

Thanks,

 

 

0 Kudos
2 Replies
Peter_W_Intel
Employee
569 Views

Does Intel compiler OMP libarary (libiomp5.so) support OMPT APIs? I saw other thread "...Since support for OMPT is not built into the libiomp5.so which ships with the Intel compiler, you must be building your own version..."

VTune(TM) Amplifier XE just collect performance data on hot functions, overall serial time and parallel time for application, overhead time & spin time for multithreaded application (including OpenMP*), what data VTune can give are OpenMP* region in timeline report, and other OpenMP metrics like Parallel Region time and Potential Performance Gains.

VTune Amplifier XE only monitor standard system APIs, thread APIs (including OpenMP to know thread creation/terminate/suspend/resume, wait objectives, etc) , I think there is no specific monitoring on OMPT APIs - treat them as common functions!

0 Kudos
Dmitry_P_Intel1
Employee
569 Views

Hello,

Currently for OpenMP analysis in VTune we use VTune User API instrumentation inside Intel OpenMP runtime library. Particularly we get from there region/barrier markup, barrier imbalance, parallel loop parameters (scheduling, chunking, iteration number), thread naming. We combine tracing info from OpenMP with sampling info from VTune collectors and building metrics like serial time vs parallel time, wall-time based potential gain for regions/worksharing constructs with break-down by reasons (imbalance, scheduling, lock contention, parallel work arrangement overhead etc).

VTune OpenMP analysis does not support OMPT. OMPT is not yet a part of standard and we are looking at opportunity to use it when it is finalized. BTW - more precisely Intel OpenMP runtime implementation that is a part of Intel Compiler does not support OMPT currently (it is done in Intel OpenMP RTL open source tree so far).

Is the root of the question refer to a fact that you see something that is missed in the current analysis or lack of support for openmp implementations that already adopted OMPT?

Thanks & Regards, Dmitry 

0 Kudos
Reply