NAME

• PAPI_profil - generate a histogram of hardware counter overflows vs. PC addresses

Synopsis

C Interface

#include <papi.h>
int PAPI_profil(void *  buf,  unsigned  bufsiz,  unsigned long  offset,
unsigned  scale,  int  EventSet,  int  EventCode,   int   threshold,
int  flags );

Fortran Interface

The profiling routines have no Fortran interface.

Description

PAPI_profil() provides hardware event statistics by profiling the occurence of specified hardware counter events. It is designed to mimic the UNIX SVR4 profil call. The statistics are generated by creating a histogram of hardware counter event overflows vs. program counter addresses for the current process. The histogram is defined for a specific region of program code to be profiled, and the identified region is logically broken up into a set of equal size subdivisions, each of which corresponds to a count in the histogram. With each hardware event overflow, the current subdivision is identified and its corresponding histogram count is incremented. These counts establish a relative measure of how many hardware counter events are occuring in each code subdivision. The resulting histogram counts for a profiled region can be used to identify those program addresses that generate a disproportionately high percentage of the event of interest.

Events to be profiled are specified with the EventSet and EventCode parameters. More than one event can be simultaneously profiled by calling PAPI_profil() several times with different EventCode values. Profiling can be turned off for a given event by calling PAPI_profil() with a threshold value of 0.

Arguments

*buf -- pointer to a buffer of bufsiz bytes in which the histogram counts are stored in an array of unsigned short, unsigned int, or unsigned long long values, or 'buckets'. The size of the buckets is determined by values in the flags argument.

bufsiz -- the size of the histogram buffer in bytes. It is computed from the length of the code region to be profiled, the size of the buckets, and the scale factor as discussed below.

offset -- the start address of the region to be profiled.

scale -- broadly and historically speaking, a contraction factor that indicates how much smaller the histogram buffer is than the region to be profiled. More precisely, scale is interpreted as an unsigned 16-bit fixed-point fraction with the decimal point implied on the left. Its value is the reciprocal of the number of addresses in a subdivision, per counter of histogram buffer. Below is a table of representative values for scale:

Representative values for the scale variable

Historically, the scale factor was introduced to allow the allocation of buffers smaller than the code size to be profiled. Data and instruction sizes were assumed to be multiples of 16-bits. These assumptions are no longer necessarily true. PAPI_profil has preserved the traditional definition of scale where appropriate, but deprecated the definitions for 0 and 1 (disable scaling) and extended the range of scale to include 65536 and 131072 to allow for exactly two addresses and exactly one address per profiling bucket.

The value of bufsiz is computed as follows:

bufsiz = (end - start)*(bucket_size/2)*(scale/65536) where

bufsiz - the size of the buffer in bytes

end, start - the ending and starting addresses of the profiled region

bucket_size - the size of each bucket in bytes; 2, 4, or 8 as defined in flags

scale - as defined above

EventSet -- The PAPI EventSet to profile. This EventSet is marked as profiling-ready, but profiling doesn't actually start until a PAPI_start() call is issued.

EventCode -- Code of the Event in the EventSet to profile. This event must already be a member of the EventSet.

threshold -- minimum number of events that must occur before the PC is sampled. If hardware overflow is supported for your substrate, this threshold will trigger an interrupt when reached. Otherwise, the counters will be sampled periodically and the PC will be recorded for the first sample that exceeds the threshold. If the value of threshold is 0, profiling will be disabled for this event.

flags -- bit pattern to control profiling behavior. Defined values are shown in the table below:

Defined bits for the flags variable

Errors

PAPI_EINVAL One or more of the arguments is invalid.

PAPI_ENOMEM Insufficient memory to complete the operation.

PAPI_ENOEVST The EventSet specified does not exist.

PAPI_EISRUN The EventSet is currently counting events.

PAPI_ECNFLCT The underlying counter hardware can not count this event and other events in the EventSet simultaneously.

PAPI_ENOEVNT The PAPI preset is not available on the underlying hardware.

Examples

int retval;
unsigned long length;
PAPI_exe_info_t *prginfo;
unsigned short *profbuf;
if ((prginfo = PAPI_get_executable_info()) == NULL)
handle_error(1);
length = (unsigned long)(prginfo->text_end - prginfo->text_start);
profbuf = (unsigned short *)malloc(length);
if (profbuf == NULL)
handle_error(1);
memset(profbuf,0x00,length);
if ((retval = PAPI_profil(profbuf, length, start, 65536, EventSet,
PAPI_FP_INS, 1000000, PAPI_PROFIL_POSIX | PAPI_PROFIL_BUCKET_16)) != PAPI_OK)
handle_error(retval);

Bugs

If you call PAPI_profil, PAPI allocates buffer space that will not be freed if you call PAPI_shutdown or PAPI_cleanup_eventset. To clean all memory, you must call PAPI_profil on the Events with a 0 threshold.

See Also

PAPI_get_executable_info(3), PAPI_overflow(3), PAPI_sprofil(3)