Free Fortran, C, C++ Compilers & Tools for CPUs and GPUs

Use the PGI Profiler (a/k/a PGPROF®) to help you analyze your application's performance.

There are two profiling modes: Command-line profiling and Visual profiling. You can enter the command-line mode by opening a shell and entering the commands that follow below. You can enter the graphical mode by typing pgprof in a command shell. You can follow this guide in either mode.

Using the PGI Profiler consists of two basic steps: Profiling your application and then analyzing the profile. Both steps can be accomplished in either mode. In command-line mode, two distinct commands are used to profile and analyze. In graphical mode, both profiling and analysis can be accomplished in the same session. Additionally, it's possible to save a profile either at the command-line or within graphical mode. The saved profile can be analyzed within the graphical mode by using the File | Import option.

How long does it take my application to run?

Profile a.out and save the performance results in the file a.prof.

$ pgprof -o a.prof a.out

The pgprof --cpu-profiling-mode top-down option orients the call tree to show main at the top and the functions it called below.

$ pgprof --cpu-profiling-mode top-down -i a.prof

======== CPU profiling result (top down):
Time(%)      Time  Name
 97.36%  35.2596s  main
 97.36%  35.2596s  | MAIN_
 32.02%  11.5976s  |   swim_mod_calc3_
 29.98%  10.8578s  |   swim_mod_calc2_
 25.93%  9.38965s  |   swim_mod_calc1_
  6.82%  2.46976s  |   swim_mod_inital_
  1.76%  637.36ms  |   | __fvd_sin_vex_256
  1.73%  625.98ms  |   | | __fvd_sin_vex
  1.51%  546.31ms  |   | __fvd_cos_vex_256
  1.48%  534.93ms  |   | | __fvd_cos_vex
  0.06%  22.763ms  |   | __fvd_cos_vex
  0.03%  11.381ms  |   | __fvd_sin_vex
  0.03%  11.381ms  |   | _mp_penter
  0.03%  11.381ms  |   |   _mp_cpenter
  0.03%  11.381ms  |   |     _mp_create_team
  0.03%  11.381ms  |   |       _mp_barrierw
  1.79%  648.74ms  |   swim_mod_calc3z_
  0.44%  159.34ms  |   _mp_pexit
  0.44%  159.34ms  |     _mp_cpexit
  0.44%  159.34ms  |       _mp_barrierw
  2.64%  956.04ms  _mp_slave
  2.64%  956.04ms    _mp_cslave
  2.64%  956.04ms      _mp_barrier_tw

To start a new profiling session, after launching the PGI Profiler, open the File menu and select New Session.
In the dialog box, browse to the executable file you want to profile. Then add any command line arguments with which to launch it.
Click Next then Finish.
In the CPU Details tab, click on the "Show the top-down (callers first) call tree view" button as shown below.

What part of my application takes the longest time to run?

Profile a.out and save the performance results in the file a.prof.

$ pgprof -o a.prof a.out

The pgprof --cpu-profiling-mode bottom-up option orients the call tree to show each function followed by functions that called it working backwards to main.

$ pgprof --cpu-profiling-mode bottom-up -i a.prof

======== CPU profiling result (bottom up):
Time(%)      Time  Name
 32.02%  11.5976s  swim_mod_calc3_
 32.02%  11.5976s  | MAIN_
 32.02%  11.5976s  |   main
 29.98%  10.8578s  swim_mod_calc2_
 29.98%  10.8578s  | MAIN_
 29.98%  10.8578s  |   main
 25.93%  9.38965s  swim_mod_calc1_
 25.93%  9.38965s  | MAIN_
 25.93%  9.38965s  |   main
  3.43%  1.24057s  swim_mod_inital_
  3.43%  1.24057s  | MAIN_
  3.43%  1.24057s  |   main
  2.64%  956.04ms  _mp_barrier_tw
  2.64%  956.04ms  | _mp_cslave
  2.64%  956.04ms  |   _mp_slave
  1.79%  648.74ms  swim_mod_calc3z_
  1.79%  648.74ms  | MAIN_
  1.79%  648.74ms  |   main
  1.76%  637.36ms  __fvd_sin_vex
  1.73%  625.98ms  | __fvd_sin_vex_256
  1.73%  625.98ms  | | swim_mod_inital_
  1.73%  625.98ms  | |   MAIN_
  1.73%  625.98ms  | |     main
  0.03%  11.381ms  | swim_mod_inital_
  0.03%  11.381ms  |   MAIN_
  0.03%  11.381ms  |     main
  1.54%  557.69ms  __fvd_cos_vex
  1.48%  534.93ms  | __fvd_cos_vex_256
  1.48%  534.93ms  | | swim_mod_inital_
  1.48%  534.93ms  | |   MAIN_
  1.48%  534.93ms  | |     main
  0.06%  22.763ms  | swim_mod_inital_
  0.06%  22.763ms  |   MAIN_
  0.06%  22.763ms  |     main
  0.47%  170.72ms  _mp_barrierw
  0.44%  159.34ms  | _mp_cpexit
  0.44%  159.34ms  | | _mp_pexit
  0.44%  159.34ms  | |   MAIN_
  0.44%  159.34ms  | |     main
  0.03%  11.381ms  | _mp_create_team
  0.03%  11.381ms  |   _mp_cpenter
  0.03%  11.381ms  |     _mp_penter
  0.03%  11.381ms  |       swim_mod_inital_
  0.03%  11.381ms  |         MAIN_
  0.03%  11.381ms  |           main
  0.38%  136.58ms  MAIN_
  0.38%  136.58ms  | main
  0.03%  11.381ms  __fvd_cos_vex_256
  0.03%  11.381ms  | swim_mod_inital_
  0.03%  11.381ms  |   MAIN_
  0.03%  11.381ms  |     main
  0.03%  11.381ms  __fvd_sin_vex_256
  0.03%  11.381ms    swim_mod_inital_
  0.03%  11.381ms      MAIN_
  0.03%  11.381ms        main

To start a new profiling session, after launching the PGI Profiler, open the File menu and select New Session.
In the dialog box, browse to the executable file you want to profile. Then add any command line arguments with which to launch it.
Click Next then Finish.
In the CPU Details tab, click on the "Show the bottom-up (callees first) call tree view" button as shown below.

How can I visualize both CPU and GPU performance data?

Profile a.out and save the performance results in the file a.prof. $ pgprof -o a.prof a.out Then display the contents of the output file. $ pgprof -i a.prof The results are broken into four sections: GPU kernel execution profile. CUDA API execution profile. OpenACC execution profile. CPU execution profile.
====== Profiling result: Time(%) Time Calls Avg Min Max Name 38.14% 1.41393s 20 70.696ms 70.666ms 70.731ms calc2_198_gpu 31.11% 1.15312s 18 64.062ms 64.039ms 64.083ms calc3_273_gpu 23.35% 865.68ms 20 43.284ms 43.244ms 43.325ms calc1_142_gpu 5.17% 191.78ms 141 1.3602ms 1.3120us 1.6409ms [CUDA memcpy HtoD] ... ======== API calls: Time(%) Time Calls Avg Min Max Name 92.65% 3.49314s 62 56.341ms 1.8850us 70.771ms cuStreamSynchronize 3.78% 142.36ms 1 142.36ms 142.36ms 142.36ms cuDevicePrimaryCtxRetain ... ======== OpenACC (excl): Time(%) Time Calls Avg Min Max Name 36.27% 1.41470s 20 70.735ms 70.704ms 70.773ms acc_wait@swim-acc-data.f:223 29.60% 1.15449s 18 64.138ms 64.114ms 64.159ms acc_wait@swim-acc-data.f:302 22.22% 866.66ms 20 43.333ms 43.294ms 43.376ms acc_wait@swim-acc-data.f:169 9.06% 353.49ms 1 353.49ms 353.49ms 353.49ms acc_update@swim-acc-data.f:402 ... ======== CPU profiling result (bottom up): Time(%) Time Name 59.09% 8.55785s cudbgGetAPIVersion 59.09% 8.55785s \| start_thread 59.09% 8.55785s \| clone 25.75% 3.73007s cuStreamSynchronize 25.75% 3.73007s \| __pgi_uacc_cuda_wait 25.75% 3.73007s \| __pgi_uacc_computedone 10.38% 1.50269s \| swim_mod_calc2_ 10.38% 1.50269s \| \| MAIN_ 10.38% 1.50269s \| \| main 8.54% 1.23625s \| swim_mod_calc3_ 8.54% 1.23625s \| \| MAIN_ 8.54% 1.23625s \| \| main 6.48% 937.85ms \| swim_mod_calc1_ 6.48% 937.85ms \| \| MAIN_ 6.48% 937.85ms \| \| main 0.37% 53.287ms \| swim_mod_calc3z_ 0.37% 53.287ms \| MAIN_ 0.37% 53.287ms \| main 6.03% 873.9ms swim_mod_inital_ ...
To start a new profiling session, after launching the PGI Profiler, open the File menu and select New Session. In the dialog box, browse to the executable file you want to profile. Then add any command line arguments with which to launch it. Click Next then Finish. The Timeline view will show the profiled events ordered by the time they occurred. The GPU Details tab lists the performance details for each GPU kernel. The CPU Details tab shows the CPU call tree. The Properties tab shows the details of the events selected on the timeline.

What information can the compiler tell me about how my application is structured for performance?

Line level profiling is not currently available on OpenPOWER. Use the Visual mode to view compiler messages.

The PGI Profiler can show you information about how your program was compiled. First add the following option when compiling and linking:

-Minfo=ccff

Profile a.out and save the performance results in the file a.prof.

$ pgprof -o a.prof a.out

Now show the profile with the CCFF information and break down the performance results by line.

$ pgprof --cpu-profiling-show-ccff on -i a.prof

======== CPU profiling result (bottom up):
Time(%)      Time  Name
  3.81%  1.42602s  swim_mod_calc3_ (src/swim-omp.f:259 0xe27)
                 [CCFF] (MSGINTENSITY) Intensity = 1.00
                 [CCFF] (MSGVECT) Generated 3 alternate versions of the loop
                 [CCFF] (MSGVECT) Generated vector sse code for the loop
                 [CCFF] (MSGPREFETCH) Generated 9 prefetch instructions for the loop
  3.81%  1.42602s  | MAIN_ (src/swim-omp.f:423 0x3d1)
  3.81%  1.42602s  |   main (0x44)
  3.74%  1.40205s  swim_mod_calc1_ (src/swim-omp.f:140 0xe0e)
                 [CCFF] (MSGINTENSITY) Intensity = 1.93
                 [CCFF] (MSGLRE) 2 loop-carried redundant expressions removed with 2
                                 operations and 4 arrays
                 [CCFF] (MSGVECT) Generated 5 alternate versions of the loop
                 [CCFF] (MSGVECT) Generated vector sse code for the loop
                 [CCFF] (MSGPREFETCH) Generated 6 prefetch instructions for the loop
  3.74%  1.40205s  | MAIN_ (src/swim-omp.f:384 0x398)
                |  [CCFF] (MSGINTENSITY) Intensity = 1.00
                |  [CCFF] (MSGINTENSITY) Intensity = 1.00
                |  [CCFF] (MSGNEGVECT) Loop not vectorized/parallelized: contains call
  3.74%  1.40205s  |   main (0x44)
  3.58%  1.34214s  swim_mod_calc3_ (src/swim-omp.f:259 0xe01)
                 [CCFF] (MSGINTENSITY) Intensity = 1.00
                 [CCFF] (MSGVECT) Generated 3 alternate versions of the loop
                 [CCFF] (MSGVECT) Generated vector sse code for the loop
                 [CCFF] (MSGPREFETCH) Generated 9 prefetch instructions for the loop
  3.58%  1.34214s  | MAIN_ (src/swim-omp.f:423 0x3d1)
  3.58%  1.34214s  |   main (0x44)
  3.46%   1.2942s  swim_mod_calc3_ (src/swim-omp.f:259 0xde6)
                 [CCFF] (MSGINTENSITY) Intensity = 1.00
                 [CCFF] (MSGVECT) Generated 3 alternate versions of the loop
                 [CCFF] (MSGVECT) Generated vector sse code for the loop
                 [CCFF] (MSGPREFETCH) Generated 9 prefetch instructions for the loop
  3.46%   1.2942s  | MAIN_ (src/swim-omp.f:423 0x3d1)
  3.46%   1.2942s  |   main (0x44)
  3.07%   1.1504s  swim_mod_calc3_ (src/swim-omp.f:259 0xdd7)
                 [CCFF] (MSGINTENSITY) Intensity = 1.00
                 [CCFF] (MSGVECT) Generated 3 alternate versions of the loop
                 [CCFF] (MSGVECT) Generated vector sse code for the loop
                 [CCFF] (MSGPREFETCH) Generated 9 prefetch instructions for the loop
  3.07%   1.1504s  | MAIN_ (src/swim-omp.f:423 0x3d1)
  3.07%   1.1504s  |   main (0x44)
  2.98%  1.11445s  swim_mod_calc3_ (src/swim-omp.f:259 0xdc4)
                 [CCFF] (MSGINTENSITY) Intensity = 1.00
                 [CCFF] (MSGVECT) Generated 3 alternate versions of the loop
                 [CCFF] (MSGVECT) Generated vector sse code for the loop
                 [CCFF] (MSGPREFETCH) Generated 9 prefetch instructions for the loop
  2.98%  1.11445s  | MAIN_ (src/swim-omp.f:423 0x3d1)
  2.98%  1.11445s  |   main (0x44)

The PGI Profiler can provide you with information about how your program was compiled. First add the following option when compiling and linking:

-Minfo=ccff

Then profile.

To start a new profiling session, after launching the PGI Profiler, open the File menu and select New Session.
In the dialog box, browse to the executable file you want to profile. Then add any command line arguments with which to launch it.
Click Next then Finish.
On the CPU Details tab, click the "Show the code structure view" button.
Double-click on a file entry, function name or line number to bring up that file or specific line in the source code view.
To the left of the line numbers in the source code view you’ll see a series of compiler notes; hover the mouse over a note to bring up the feedback provided by the compiler.

For more information about the PGI Profiler, use the '--help' option in command-line mode or in graphical mode, use the Help | Contents menu item.

Profiler Quick Start Guide