Subversion Repositories openrisc

The profile_gprof package provides a mechanism to measure the

runtime performance of an application.  This is done by gathering

an execution histogram.

When profiling is started on the target device, a

TFTP

server will be started

which exports the single file

PROFILE.DAT

This analysis data can then be fetched

by connecting to the target with a TFTP

client program

and then be processed by the

gprof

utility program.

Be sure and specify binary mode transfers for this data file,

which may not be the default with on some TFTP

client programs.

The port used for this TFTP server is

configurable.  The default will be the IETF standard port

of 69/UDP, but it may be changed to any UDP port via the

CYGNUM_PROFILE_TFTP_PORT CDL option.

In order for profile data to be gathered for an application, the

program has to initiate the process.

Once started, execution histogram data will be collected in a

dynamic memory buffer.

This data can be uploaded to a host using TFTP.

A side effect of the upload of the data is that the histogram

is reset.

This is useful, especially for high resolution histograms, since

the histogram data are collected as 16-bit counters which can be quickly

saturated.

For example, if the histogram is being collected at a rate of 10,000

samples per second, a hot spot in the program could saturate after

only 6.5 seconds.

 The API for the application profiling functions can be

found in the file <cyg/profile/profile.h>.

This function is used to initiate the gathering of the

runtime execution histogram data.

void profile_on(void *start, void *end, int bucket_size, int resolution);

Calling this function will initiate execution profiling.

An execution histogram is collected at the rate of

resolution times per second.

The area between start and end

will be divided up into a number of buckets, each representing

bucket_size

program bytes in length.  Using statistical sampling (via a high speed timer), when

the program counter is found to be within the range

start..end, the appropriate

bucket (histogram entry) will be incremented.

The choice of resolution and bucket_size

control how large the data gathered will be, as well as how much overhead is

encumbered for gathering the histogram.

Smaller values for bucket_size will garner better

results (gprof can more closely align the data with

actual function names) at the expense of a larger data buffer.

The value of bucket_size will be rounded up to a power of two.