Intel® Inspector User Guide for Linux* OS

User Guide

  • 2021.4
  • 09/27/2021
  • Public Content
Contents

APIs for Collection Control

Intel Inspector
 provides a set of collection control APIs to help you specify which parts of your application should be included in, or excluded from, analysis.
To support independence of implementation between library implementers and library users, the collection control APIs allow nesting by pushing and popping state information.
There are two kinds of collection control APIs:
  • Time-oriented APIs to control which time periods of thread execution to analyze or not analyze
  • Class-oriented APIs to control which data objects to analyze or not analyze

Using Time-oriented Collection Control APIs in Your Code

Use This in C/C++ Code
Use This in Fortran Code
To Do This 
void __itt_suppress_push( 
   unsigned int etype) 

subroutine itt_suppress_push(mask)
   integer, intent(in), value :: mask
end subroutine itt_suppress_push
Tell the
Intel Inspector
 to stop analyzing for errors on the current thread.
C/C++
etype
 is:
  • __itt_suppress_memory_errors
     to stop analyzing for memory errors
  • __itt_suppress_threading_errors
     to stop analyzing for threading errors
  • __itt_suppress_memory_errors| _itt_suppress_threading_errors
     to stop analyzing for memory or threading errors
Fortran
mask
 is:
  • itt_suppress_memory_errors
     to stop analyzing for memory errors
  • itt_suppress_threading_errors
     to stop analyzing for threading errors
  • itt_suppress_all_errors
     to stop analyzing for memory or threading errors 
void __itt_suppress_pop (
   void) 

subroutine itt_suppress_pop()
end subroutine itt_suppress_pop
Tell the
Intel Inspector
 to undo the action corresponding to the most recent matching nested
push
 call.
Push
 calls nest and are additive, so the
Intel Inspector
 does not resume analyzing for:
  • Memory errors until there are an equal number of
    pop
     calls corresponding to
    push
     calls for memory errors
  • Threading errors until there are an equal number of
    pop
     calls corresponding to
    push
     calls for threading errors
For errors that might involve two threads, the state the other thread was in at the time the memory operation happened is more important than the current thread state. For example:
  1. Thread A writes to variable X, then makes a
    push
     call.
  2. Thread B reads variable X.
Even though thread A is currently suppressed, neither thread was suppressed at the time the conflicting accesses occurred, so the race is reported.
Conversely:
  1. Thread A suppresses, writes to variable Y, then unsuppresses.
  2. Thread B reads variable Y.
Even though both threads are currently unsuppressed, the error is not reported because thread A was suppressed at the time the conflicting accesses occurred.

Using Object-oriented Collection Control APIs in Your Code

Use This in C/C++ Code
Use This in Fortran Code
To Do This 
void __itt_suppress_mark_range(
   __itt_suppress_mode_t mode,
   unsigned int etype,
   void * address,
   size_t size); 
subroutine itt_suppress_mark_range(
   action, mask, addr, size)
   integer, intent(in), value :: action
   integer, intent(in), value :: mask
   integer(kind=itt_ptr), intent(in), value :: addr
   integer(kind=itt_ptr), intent(in), value :: size
end subroutine itt_suppress_mark_range
C/C++
Tell the
Intel Inspector
 to mark the memory range defined by
address
 and
size
 with the flags
mode
 and
etype
.
mode
 is
__itt_suppress_range
 or
__itt_unsuppress_range
.
etype
 is:
  • __itt_suppress_memory_errors
     to stop analyzing for memory errors
  • __itt_suppress_threading_errors
     to stop analyzing for threading errors
  • __itt_suppress_memory_errors | __itt_suppress_threading_errors
     to stop analyzing for memory or threading errors
addr
 is the address of the first byte to suppress.
size
 is the number of bytes to suppress.
Fortran
Tell the
Intel Inspector
 to mark the memory range defined by
addr
 and
size
 with the flags
action
 and
mask
.
action
 is
itt_suppress_range
 or
itt_unsuppress_range
.
mask
 is:
  • itt_suppress_memory_errors
     to stop analyzing for memory errors
  • itt_suppress_threading_errors
     to stop analyzing for threading errors
  • itt_suppress_all_errors
     to stop analyzing for memory or threading errors
addr
 is the address of the first byte to suppress.
size
 is the number of bytes to suppress. 
void __itt_suppress_clear_range(
   __itt_suppress_mode_t mode, 
   unsigned int etype, 
   void * address, 
   size_t size); 

subroutine itt_suppress_clear_range(
   action, mask, addr, size)
   integer, intent(in), value :: action
   integer, intent(in), value :: mask
   integer(kind=itt_ptr), intent(in), value :: addr
   integer(kind=itt_ptr), intent(in), value :: size
end subroutine itt_suppress_clear_range
C/C++
Tell the
Intel Inspector
 to clear the previously marked range defined by
address
 and
size
 with the flags
mode
 and
etype
.
mode
 is
__itt_suppress_range
 or
__itt_unsuppress_range
.
etype
 is:
  • __itt_suppress_memory_errors
     to stop analyzing for memory errors
  • __itt_suppress_threading_errors
     to stop analyzing for threading errors
  • __itt_suppress_memory_errors | __itt_suppress_threading_errors
     to stop analyzing for memory or threading errors
addr
 is the address of the first byte to not suppress.
size
 is the number of bytes to not suppress.
Fortran
Tell the
Intel Inspector
 to clear the previously marked range defined by
addr
 and
size
 with the flags
action
 and
mask
.
action
 is
itt_suppress_range
 or
itt_unsuppress_range
.
mask
 is:
  • itt_suppress_memory_errors
     to stop analyzing for memory errors
  • itt_suppress_threading_errors
     to stop analyzing for threading errors
  • itt_suppress_all_errors
     to stop analyzing for memory or threading errors
addr
 is the address of the first byte to not suppress.
size
 is the number of bytes to not suppress.
When the
Intel Inspector
 detects an error, it checks the set of marked ranges and, using the smallest enclosing range, filters out the error if the
mode/action
 is
suppress_range
 and the
etype/mask
 matches the detected error.
The largest supported range is 2 GB.
The default
mode/action
 for all of memory is
unsuppress_range
. You can change the default by creating a range with address NULL and size 0.
You can nest ranges only if the new range is a subset of existing ranges.
Two calls to
mark_range
 that specify the same range and intersecting
etype/mask
 may not specify both
suppress_range
 and
unsuppress_range
 as the
mode/action
.

Usage Example: Time-oriented Collection Control

C/C++ Example
Fortran Example 
#include <ittnotify.h>
 
...
#pragma omp parallel
    __itt_suppress_push(
       __itt_suppress_threading_errors);
    /* Any threading errors here will be
       ignored by the calling thread.
       In this case, each thread in the region */
 
    __itt_suppress_pop();
    /* Any threading errors here will be
       seen by Inspector*/
}
use ittnotify

...!$omp parallel
    call itt_suppress_push(
       itt_suppress_threading_errors)
    !Any threading errors here will be
       !ignored by the calling thread.
       !In this case, each thread in the region 
 
    call itt_suppress_pop()
    ! Any threading errors here will be
       !seen by Inspector

!$omp end parallel

Usage Example: Object-oriented Collection Control

C/C++ Example
Fortran Example 
#include <ittnotify.h>

int variable_to_watch;
int other_variable;
…
     //change the default mode by using NULL and 0
       as address and size
    __itt_suppress_mark_range(
       __itt_suppress_range,
       __itt_suppress_threading_errors,
       NULL, 
       0);
     //ensure we see errors on variable_to_watch
    __itt_suppress_mark_range(
      __itt_unsuppress_range,
      __itt_suppress_threading_errors,
      &variable_to_watch,
      sizeof(variable_to_watch));
#pragma omp parallel
     …
     variable_to_watch++; //race will be reported
     other_variable++; //race will not be reported
}
     //clear the record for all of memory
    __itt_suppress_clear_range(
      __itt_suppress_range,
      __itt_suppress_threading_errors,
      NULL,
      0);
    //clear the record for variable_to_watch
    __itt_suppress_clear_range(
      __itt_unsuppress_range,
      __itt_suppress_threading_errors,
      &variable_to_watch, 
      sizeof(variable_to_watch));
    //mark the range for other_variable 
      so we don’t see errors
    __itt_suppress_mark_range(
      __itt_suppress_range,
      __itt_suppress_threading_errors,
      &other_variable,
      sizeof(other_variable));
#pragma omp parallel
     …
     variable_to_watch++; //race will be reported
     other_variable++; //race not be reported
}
use ittnotify

common /x/ variable_to_watch, other_variable
integer variable_to_watch
integer other_variable
…
     !change the default mode by using 0 and 0
       !as address and size
    call itt_suppress_mark_range(
      itt_suppress_range,
      itt_suppress_threading_errors,
      0,
      0)
     !ensure we see errors on variable_to_watch
    call itt_suppress_mark_range(
      itt_unsuppress_range,
      itt_suppress_threading_errors, 
      LOC(variable_to_watch),
      4)
!$omp parallel
     …
     variable_to_watch = variable_to_watch + 1
     !race will be reported
     other_variable = other_variable + 1
     !race will not be reported
!$omp end parallel
     !clear the record for all of memory
    call itt_suppress_clear_range(
      itt_suppress_range,
      itt_suppress_threading_errors,
      0,
      0);
    !clear the record for variable_to_watch
    call itt_suppress_clear_range(
      itt_unsuppress_range,
      itt_suppress_threading_errors,
      LOC(variable_to_watch),
      4)
    !mark the range for other_variable 
      !so we don’t see errors
    call itt_suppress_mark_range(
      itt_suppress_range,
      itt_suppress_threading_errors,
      LOC(other_variable),
      4)
!$omp parallel
     …
     variable_to_watch = variable_to_watch + 1
     !race will be reported
     other_variable = other_variable + 1
     !race not be reported
!$omp end parallel

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.