OProfile
This page describes how to use oprofile on Sigma 8654 board. Currently, oprofile on mips supports 24K core only.
There is a required patch to the Sigma kernel. Until this is included in Sigma’s kernel, please apply the patch.
This page assumes you know oprofile, and only need MIPS specific details. For general info, manuals, etc., please see the project page here: http://oprofile.sourceforge.net/docs
Getting started
This section of commands is typed at the Android shell prompt, using the serial console, or an adb shell:
Load the kernel module by typing:
insmod /lib/modules/oprofile.ko
Do a simple setup (profiling by CYCLES only):
opcontrol --setup --event=CYCLES:40000 --vmlinux=/vmlinux --kernel-range=0x84000000,0x84400000
Start profiling:
opcontrol --start
Do something that you want to profile ….
Stop profiling:
opcontrol --stop
Pulling samples to your host workstation
At this point, you have collected sample data and you need to move it from your target board to your host and analyze it there.
I use the python script provided at external/oprofile/opimport_pull. I have copied it to my working directory, and modified the first line from this:
#!/usr/bin/python2.4 -E
to this:
#!/usr/bin/python -E
Before you can run this script, you must have set up your paths by doing this at the top of your build tree:
source build/envsetup.sh
setpaths
Then you must make sure that you have a live ADB connection, by:
export ADBHOST=<your_target_board_IP_addr>
adb kill-server
adb shell
(log out of adb shell, once you know it works)
Then you can run the script like this:
opimport_pull -r my_samples
That tells the script to make a subdirectory called “my_samples” (and first remove it if it already exists), and pull the samples over from your target board.
A little digression
On my system, adb pull (used inside the opimport_pull script) runs very slowly, and it takes 5 minutes to pull over the samples. I wrote a workaround for this, by reading the samples directly off the NFS server that my target board mounts its root filesystem.
See the script nfs_opimport_pull which is attached to this page. You will need to modify line 9 of this script to point to the location of your root filesystem, which must also be mounted on your host workstation. The script copies the files as root using “sudo”, because Android makes its /data directory private, and therefore only readable by root.
Sample Analysis
Finally, if all the above steps are OK, the opimport_pull script will fire off opreport to give you a summary analysis of your trace. The relevant part will look something like this:
CPU: MIPS 24K, speed 0 MHz (estimated) Counted CYCLES events (Cycles) with a unit mask of 0x00 (No unit mask) count 40000 samples % app name symbol name 57230 37.4630 libcutils.so /system/lib/libcutils.so 40837 26.7321 libskia.so /system/lib/libskia.so 18923 12.3871 libwebcore.so /system/lib/libwebcore.so 14348 9.3923 vmlinux /vmlinux 7862 5.1465 libc.so /system/lib/libc.so 7463 4.8853 libdvm.so /system/lib/libdvm.so 1774 1.1613 oprofiled /system/xbin/oprofiled 996 0.6520 oprofile /oprofile 616 0.4032 libffi.so /system/lib/libffi.so 586 0.3836 libbinder.so /system/lib/libbinder.so 533 0.3489 libm.so /system/lib/libm.so 424 0.2776 libutils.so /system/lib/libutils.so 279 0.1826 libandroid_runtime.so /system/lib/libandroid_runtime.so 245 0.1604 libui.so /system/lib/libui.so 115 0.0753 libsurfaceflinger.so /system/lib/libsurfaceflinger.so (truncated)
You can then call opreport with the --symbols directive to let it break down the profile by function name. You have to tell it where your symbols are located. I cd into the sample directory, and give it a command like this (note that the -p option seems to want an absolute path, so substitute your own path to top of android tree):
opreport -p /u/android/mips2-eclair/out/target/product/smp86xx/symbols/system/lib --symbols --session-dir=. | head -100
That will produce this type of detail:
samples % app name symbol name 29443 19.2822 libcutils.so mips_memset 26507 17.3594 libcutils.so android_memset16 14348 9.3965 vmlinux /vmlinux 8777 5.7481 libskia.so S32A_D565_Opaque_Dither(unsigned short*, unsigned int const*, int, unsigned int, int, int) 8445 5.5306 libskia.so SkRGB16_Black_Blitter::blitMask(SkMask const&, SkIRect const&) 2247 1.4716 libwebcore.so WebCore::WidthIterator::advance(int, WebCore::GlyphBuffer*) 2079 1.3615 libskia.so SkRGB16_Opaque_Blitter::blitMask(SkMask const&, SkIRect const&) 1774 1.1618 oprofiled /system/xbin/oprofiled 1440 0.9431 libwebcore.so WebCore::Font::glyphDataForCharacter(int, bool, bool) const 1165 0.7630 libc.so memcmp 1023 0.6700 libc.so free 1018 0.6667 libwebcore.so WebCore::InlineFlowBox::paint(WebCore::RenderObject::PaintInfo&, int, int) 1007 0.6595 libwebcore.so WebCore::InlineTextBox::paint(WebCore::RenderObject::PaintInfo&, int, int) 996 0.6523 oprofile /oprofile 921 0.6032 libskia.so SkPaint::descriptorProc(SkMatrix const*, void (*)(SkDescriptor const*, void*), void*) const 861 0.5639 libwebcore.so WebCore::GlyphBuffer::add(unsigned short, WebCore::SimpleFontData const*, float, WebCore::FloatSize const*) 778 0.5095 libdvm.so dvmMterp_invokeMethod 717 0.4696 libskia.so SkPicturePlayback::draw(SkCanvas&)
Finally, you can also give it a path to your kernel vmlinux file, and it will get those symbols as well. I again give it a a full path, substitute your own prefix:
opreport -p /u/android/mips2-eclair/out/target/product/smp86xx/symbols/system/lib,/u/android/kernel-sigma --symbols --session-dir=. | head -100
samples % app name symbol name 29443 19.4091 libcutils.so mips_memset 26507 17.4736 libcutils.so android_memset16 8777 5.7859 libskia.so S32A_D565_Opaque_Dither(unsigned short*, unsigned int const*, int, unsigned int, int, int) 8445 5.5670 libskia.so SkRGB16_Black_Blitter::blitMask(SkMask const&, SkIRect const&) 2247 1.4812 libwebcore.so WebCore::WidthIterator::advance(int, WebCore::GlyphBuffer*) 2079 1.3705 libskia.so SkRGB16_Opaque_Blitter::blitMask(SkMask const&, SkIRect const&) 1799 1.1859 vmlinux __do_softirq 1774 1.1694 oprofiled /system/xbin/oprofiled 1440 0.9493 libwebcore.so WebCore::Font::glyphDataForCharacter(int, bool, bool) const 1205 0.7943 vmlinux ring_buffer_consume 1188 0.7831 vmlinux finish_task_switch 1165 0.7680 libc.so memcmp 1023 0.6744 libc.so free
A few other notes:
Once you have taken a sample and uploaded it, if you want to take another using the same settings, do a --reset first, to clear out the old samples:
opcontrol --reset
opcontrol --start
<do something interesting>
opcontrol --stop
If you want to change the setup, must must first shutdown the oprofile daemon, and clean up:
opcontrol --shutdown
opcontrol --setup --event=CYCLES:50000 --event=ICACHE_MISSES:3000 --vmlinux=/vmlinux --kernel-range=0x84000000,0x84400000
opcontrol --start
<do something interesting>
opcontrol --stop
The 24K has two performance counters, and they differ in capability. Please read the 24K core manual to see what you can profile, and use the “opcontrol –list” command to see the names. You can specify two “–event” specs to opcontrol, as above.
It makes the most sense to specify things that you think can be correlated to each other. The opreport function will print them side-by-side.