Time to select a projet, profiling ffmpeg2theora

Time has come to choose a project for the DPS915 CUDA programming course.

After looking online, a project that caught my eyes was the ffmpeg2theora

ffmpeg2theora is built on top of the ffmpeg project and its goal is to provide a command line interface to convert videos to the theora format wrapped in an Ogg container

My idea for the project is to add GPU optimization support to the converter, specifically using the CUDA api for Nvidia graphics cards.
At the moment it is not clear how or if it’s even possible to do that since the converter itself has a lot of dependencies and talking with some developers in the #vorbis channel I was told that the optimizations would have to be done in the libtheora and a big chunk of the library it’s already written in assembly for performance reasons.

So for now I’m trying to gather as much information as possible.

Here is a list of some resources relevant to the research


To get an idea of the project I decided to build it and play around with the converter

As I would expect to build ffmpeg2theora from source requires a bunch of dependencies.
The developers created two scripts that make the process easy.
One script clones the latest stable release of ffmpeg from their git repository and builds it and the other does the same thing but for libkate

Besides installing ffmpeg and libkate I also needed to install

  • libvorbis
  • libogg

On Ubuntu I also had to install

  • yasm
  • gawk

The Build system

For the build system they use SCons
SCons is a software construction tool implemented in Python, it is a replacement for the famous make.

I have to say that at first I was kind of skeptic, but after reading their user docs and hacking around some scripts I fell in love immediately.
SCons doesn’t try to solve all the problems in the world, but they take a very pragmatic approach towards build tools and have some info to back it up.

Here is the SCons script used in the ffmpeg2theora projet:

# SCons build specification
# vi:si:et:sw=2:sts=2:ts=2
from glob import glob
import os

import SCons

def version():
    f = os.popen("./version.sh")
    version = f.read().strip()
    return version




except TypeError:
    print 'SCons %d.%d.%d or greater is required, but you have an older version' % scons_version

opts = Variables()
  BoolVariable('static', 'Set to 1 for static linking', 0),
  BoolVariable('debug', 'Set to 1 to enable debugging', 0),
  BoolVariable('build_ffmpeg', 'Set to 1 to build local copy of ffmpeg', 0),
  ('prefix', 'install files in', '/usr/local'),
  ('bindir', 'user executables', 'PREFIX/bin'),
  ('mandir', 'man documentation', 'PREFIX/man'),
  ('destdir', 'extra install time prefix', ''),
  ('APPEND_CCFLAGS', 'Additional C/C++ compiler flags'),
  ('APPEND_LINKFLAGS', 'Additional linker flags'),
  BoolVariable('libkate', 'enable libkate support', 1),
  BoolVariable('crossmingw', 'Set to 1 for crosscompile with mingw', 0)
env = Environment(options = opts)

pkg_flags="--cflags --libs"
if env['static']:
  pkg_flags+=" --static"

if env['crossmingw']:
    env.Tool('crossmingw', toolpath = ['scons-tools'])

prefix = env['prefix']
if env['destdir']:
  if prefix.startswith('/'): prefix = prefix[1:]
  prefix = os.path.join(env['destdir'], prefix)
man_dir = env['mandir'].replace('PREFIX', prefix)
bin_dir = env['bindir'].replace('PREFIX', prefix)

  '-DPACKAGE_VERSION=\\"%s\\"' % pkg_version,
  '-DPACKAGE_STRING=\\"%s-%s\\"' % (pkg_name, pkg_version),
  '-DPACKAGE=\\"%s\\"' % pkg_name,

env.Append(CCFLAGS = Split('$APPEND_CCFLAGS'))

if env['debug'] and env['CC'] == 'gcc':
  env.Append(CCFLAGS=["-g", "-O2", "-Wall"])

if GetOption("help"):

def ParsePKGConfig(env, name):
  if os.environ.get('PKG_CONFIG_PATH', ''):
    action = 'PKG_CONFIG_PATH=%s pkg-config %s "%s"' % (os.environ['PKG_CONFIG_PATH'], pkg_flags, name)
    action = 'pkg-config %s "%s"' % (pkg_flags, name)
  return env.ParseConfig(action)

def TryAction(action):
    import os
    ret = os.system(action)
    if ret == 0:
        return (1, '')
    return (0, '')

def CheckPKGConfig(context, version):
  context.Message( 'Checking for pkg-config... ' )
  ret = TryAction('pkg-config --atleast-pkgconfig-version=%s' % version)[0]
  context.Result( ret )
  return ret

def CheckPKG(context, name):
  context.Message( 'Checking for %s... ' % name )
  if os.environ.get('PKG_CONFIG_PATH', ''):
    action = 'PKG_CONFIG_PATH=%s pkg-config --exists "%s"' % (os.environ['PKG_CONFIG_PATH'], name)
    action = 'pkg-config --exists "%s"' % name
  ret = TryAction(action)[0]
  context.Result( ret )
  return ret

env.PrependENVPath ('PATH', os.environ['PATH'])

conf = Configure(env, custom_tests = {
  'CheckPKGConfig' : CheckPKGConfig,
  'CheckPKG' : CheckPKG,

if env["build_ffmpeg"]:
  if env.GetOption('clean'):
    TryAction("cd ffmpeg;make distclean")

if not env.GetOption('clean'):
  if not conf.CheckPKGConfig(pkgconfig_version):
     print 'pkg-config >= %s not found.' % pkgconfig_version

  if not conf.CheckPKG("ogg >= 1.1"):
    print 'ogg >= 1.1 missing'

  if not conf.CheckPKG("vorbis"):
    print 'vorbis missing'

  if not conf.CheckPKG("vorbisenc"):
    print 'vorbisenc missing'

  if not conf.CheckPKG("theoraenc >= 1.1.0"):
    print 'theoraenc >= 1.1.0 missing'

  XIPH_LIBS="ogg >= 1.1 vorbis vorbisenc theoraenc >= 1.1.0"

  if not conf.CheckPKG(XIPH_LIBS):
    print 'some xiph libs are missing, ffmpeg2theora depends on %s' % XIPH_LIBS
  ParsePKGConfig(env, XIPH_LIBS)

      "libavcodec >= 52.30.0",
  if os.path.exists("./ffmpeg"):
    pkg_path = list(set(map(os.path.dirname, glob('./ffmpeg/*/*.pc'))))
    pkg_path.append(os.environ.get('PKG_CONFIG_PATH', ''))
    os.environ['PKG_CONFIG_PATH'] = ':'.join(pkg_path)

  if not conf.CheckPKG(' '.join(FFMPEG_LIBS)):
    print """
        Could not find %s.
        You can install it via
         sudo apt-get install %s
        or update PKG_CONFIG_PATH to point to ffmpeg's source folder
        or run ./get_ffmpeg.sh (for more information see INSTALL)
    """ %(" ".join(FFMPEG_LIBS), " ".join(["%s-dev"%l.split()[0] for l in FFMPEG_LIBS]))

  for lib in FFMPEG_LIBS:
      ParsePKGConfig(env, lib)

  if conf.CheckCHeader('libavformat/framehook.h'):

  if env['libkate']:
    if os.path.exists("./libkate/misc/pkgconfig"):
      os.environ['PKG_CONFIG_PATH'] = "./libkate/misc/pkgconfig:" + os.environ.get('PKG_CONFIG_PATH', '')
    if os.path.exists("./libkate/pkg/pkgconfig"):
      os.environ['PKG_CONFIG_PATH'] = "./libkate/pkg/pkgconfig:" + os.environ.get('PKG_CONFIG_PATH', '')
    if conf.CheckPKG(KATE_LIBS):
      ParsePKGConfig(env, KATE_LIBS)
      env.Append(CCFLAGS=['-DHAVE_KATE', '-DHAVE_OGGKATE'])
      print """
          Could not find libkate. Subtitles support will be disabled.
          You can also run ./get_libkate.sh (for more information see INSTALL)
          or update PKG_CONFIG_PATH to point to libkate's source folder

  if conf.CheckCHeader('iconv.h'):
      if conf.CheckLib('iconv'):

  if env['crossmingw']:
  elif env['static']:
      env.Append(LIBS=['m', 'dl'])

# Flags for profiling

env = conf.Finish()

# ffmpeg2theora
ffmpeg2theora = env.Clone()
ffmpeg2theora_sources = glob('src/*.c')
ffmpeg2theora.Program('ffmpeg2theora', ffmpeg2theora_sources)

ffmpeg2theora.Install(bin_dir, 'ffmpeg2theora')
ffmpeg2theora.Install(man_dir + "/man1", 'ffmpeg2theora.1')
ffmpeg2theora.Alias('install', prefix)

The script just set some configurations for the build and checks for some dependencies.
I added some extra flags because I wanted to generate a profile of the application


Summarizing the steps to build ffmpeg2theora:

Download the source code


sudo ./get_ffmpeg.sh
sudo ./get_libkate.sh
sudo scons
sudo scons install

**If you need to install any other dependencies the configure scripts will output to the terminal
**On the mac I had some problems in running “sudo scons”, the pkg-config path would get corrupted and the build would fail, by loggin in the shell as root and sourcing the environment variables of my profile solved the problem (I didn’t have this issue on Ubuntu)
**If you don’t run the get_ffmpeg script as root the libraries won’t be installed in the system and the build will fail during the linking stage


Next step was to generate a profile of the program and see which area of the application was consuming most of the CPU time.
I used the Instruments Timer Profiler to create a profile of the application.
I have previously blogged about how to use the Instruments Timer Profiler on the mac

instruments -t "/Applications/Xcode.app/Contents/Applications/Instruments.app/Contents/Resources/templates/Time Profiler.tracetemplate" ./ffmpeg2theora myvideo.mp4

and the profile information was generated:

Looking at the profile gave me a better idea of how the converter works but I still need to run the converter with a larger video to see where the heavy processing takes place.

What’s next?

This is a very intimidating projet considering that I’m not very familiar in video encoding and CUDA programming, but what better way to learn something than by doing it? 🙂
I still remember taking the Topics in OpenSource Development last year with David Humphrey here at Seneca college and how we started hacking Firefox. At the beginning it was very hard and overwelming, but after a while the beast didn’t look as scary as before. That just proved to me that as long you put the time into something you will get the results no matter what. In the end hard works does pay off, indeed.

With that being said, I’m a little scare about diving into an area that I don’t much about and trying to implement something new, but at the same time I welcome the challenge and I will try to learn as much as I can during the process.
Video Processing and GPU Programming are two topics that interest me so I’m sure it will be a lot of fun 🙂


Building GNU Scientific Library(GSL) from source

In the DPS915 workshop 2 we started working with the GNU Scientific Library(GSL) to perform some matrix operations.

GSL is an implementation of Basic Linear Algebra Subprograms(BLAS)

By default, the GSL libraries don’t come pre install on OSX, so if you want to use them you gotta install it yourself


You can download the latest stable release from here.
At this time, the latest stable one is version 1.15


Unzip the files and navigate to the folder.
The steps to install:

Check if all the dev dependencies are installed


Build the libraries.

time make -j12 > /dev/null


  • time will output how much the build process took
  • -j lets you specify how many cores to use during the build
  • > /dev/null redircts the output to make the build faster

Finally you need to install the libraries

sudo make install

The header files are installed in:

And some libraries in:

Test it

To check if the gsl libs are working you can try to run a sample program:

 // matMult.cpp

 extern "C" {
 #include <gsl/gsl_cblas.h>
 using namespace std;
 #define WIDTH 5

 int main(int argc, char** argv) {
     if (argc != 2)
         cerr << "** invalid number of arguments**" << endl;
     else {
         int n = atof(argv[1]);
         float* a = new float[n * n];
         float* b = new float[n * n];
         float* c = new float[n * n];
         int k = 0;
         double f = 1.0 / RAND_MAX;
         for (int i = 0; i < n; i++)
             for (int j = 0; j < n; j++)
                 a[k++] = rand() * f;
         k = 0;
         for (int i = 0; i < n; i++)
             for (int j = 0; j < n; j++)
                 b[k++] = rand() * f;

         // compute matrix product
         cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
          n, n, n, 1.0, a, n, b, n, 0.0, c, n);

         // output result
         k = 0;
         cout << setprecision(6) << fixed;
         for (int i = 0; i < n; i++) {
             cout << setw(5) << i+1 << ':';
             for (int j = 0; j < n; j++) {
                 cout << setw(10) << c[k++] << ' ';
                 if (j % WIDTH == WIDTH - 1)
                     cout << endl << setw(5) << i+1 << ':';
             cout << endl;
         cout << endl;
         delete [] a;
         delete [] b;
         delete [] c;

To compile:

g++ -std=c++0x -Wall matMult.cpp -o matMult -lgslcblas

-std=c++0x enables c++0x support

To run the program:

./matMult 2

The first argument is the size of the matrix you want to generate

If you see a 2 by 2 matrix then it means it worked 🙂

**The code sample was taken from https://scs.senecac.on.ca/~gpu610/pages/content/lines.html

Firefox Bug 784402, Pointer Lock must respect iframe sandbox flag

Recently I’ve worked on the Firefox Bug 784402 – Pointer Lock must respect iframe sandbox flag.

This is a quick overview of what had to be done on the bug.

Sandbox flags

First lets check what the sandbox attribute does:
A quote from the w3c spec

The sandbox attribute, when specified, enables a set of extra restrictions on any content hosted by the iframe. Its value must be an unordered set of unique space-separated tokens that are ASCII case-insensitive. The allowed values are allow-forms, allow-popups, allow-same-origin, allow-scripts, and allow-top-navigation. When the attribute is set, the content is treated as being from a unique origin, forms and scripts are disabled, links are prevented from targeting other browsing contexts, and plugins are secured. The allow-same-origin keyword allows the content to be treated as being from the same origin instead of forcing it into a unique origin, the allow-top-navigation keyword allows the content to navigate its top-level browsing context, and the allow-forms, allow-popups and allow-scripts keywords re-enable forms, popups, and scripts respectively.

With pointerlock landing on Firefox 15, it was decided that a new sandbox flag should be created to restrict the pointerlock usage on embedded scripts in a page, so for example: if you add an advertisement script on your page, you don’t want to give the permissions to the advertisement to lock the pointer to itself.
To manage that, the allow-pointer-lock sandbox was created.

An overview of how the sandbox flags work:
List of flags:

 * This flag prevents content from navigating browsing contexts other than
 * the sandboxed browsing context itself (or browsing contexts further
 * nested inside it), and the top-level browsing context.
const unsigned long SANDBOXED_NAVIGATION  = 0x1;

 * This flag prevents content from navigating their top-level browsing
 * context.
const unsigned long SANDBOXED_TOPLEVEL_NAVIGATION = 0x2;

 * This flag prevents content from instantiating plugins, whether using the
 * embed element, the object element, the applet element, or through
 * navigation of a nested browsing context, unless those plugins can be
 * secured.
const unsigned long SANDBOXED_PLUGINS = 0x4;

 * This flag forces content into a unique origin, thus preventing it from
 * accessing other content from the same origin.
 * This flag also prevents script from reading from or writing to the
 * document.cookie IDL attribute, and blocks access to localStorage.
const unsigned long SANDBOXED_ORIGIN = 0x8;

 * This flag blocks form submission.
const unsigned long SANDBOXED_FORMS = 0x10;

 * This flag blocks script execution.
const unsigned long SANDBOXED_SCRIPTS = 0x20;

 * This flag blocks features that trigger automatically, such as
 * automatically playing a video or automatically focusing a form control.
const unsigned long SANDBOXED_AUTOMATIC_FEATURES = 0x40;

 * This flag blocks the document from acquiring pointerlock.
const unsigned long SANDBOXED_POINTER_LOCK = 0x80;

Parsing the flags

So we have a 32 bit integer to store the sandbox flags.

Breaking down the integer we have 8 bytes
We can represent each byte in hexadecimal format:

So the number 0xFFFFFFFF has all the bits turned ON

Knowing that, we could use each bit of the integer to represent a flag.
We don’t care about the decimal value of that integer, since we are using it to store flags and not values.
So by saying 0x1, we are telling to turn the first bit of the first byte on, 0x2 turns the second bit of the first byte on
0x10 on the other hand tells to turn the first bit of the second byte on.
Remember that we are using hexadecimal notation.

So in the end, what’s happening is that each flag is turning a different bit on the integer

Later we’ll be able to check if that specific bit is ON or OFF and determine the status of the flag.

One thing to keep in mind is that if the iframe doesn’t have the sandbox attribute, then all the flags are turned OFF by default.

<i frame></i frame>

If the iframe has an empty sandbox attribute, then all the flags are ON by default

<i frame sandbox=""></i frame>

To turn the flags off, you can specify the feature you want to enable in the sandbox attribute:

<i frame sandbox="allow-pointer-lock allow-same-origin></i frame>

In the snippet above both the allow-pointer-lock and allow-same-origin flag would be turned OFF, all the other flags would be ON

This is the code that parses the sandbox flags:

 * A helper function that parses a sandbox attribute (of an <iframe> or
 * a CSP directive) and converts it to the set of flags used internally.
 * @param aAttribute    the value of the sandbox attribute
 * @return              the set of flags
nsContentUtils::ParseSandboxAttributeToFlags(const nsAString& aSandboxAttrValue)
  // If there's a sandbox attribute at all (and there is if this is being
  // called), start off by setting all the restriction flags.
  uint32_t out = SANDBOXED_NAVIGATION |
                 SANDBOXED_PLUGINS |
                 SANDBOXED_ORIGIN |
                 SANDBOXED_FORMS |
                 SANDBOXED_SCRIPTS |

  if (!aSandboxAttrValue.IsEmpty()) {
    // The separator optional flag is used because the HTML5 spec says any
    // whitespace is ok as a separator, which is what this does.
    HTMLSplitOnSpacesTokenizer tokenizer(aSandboxAttrValue, ' ',

    while (tokenizer.hasMoreTokens()) {
      nsDependentSubstring token = tokenizer.nextToken();
      if (token.LowerCaseEqualsLiteral("allow-same-origin")) {
        out &= ~SANDBOXED_ORIGIN;
      } else if (token.LowerCaseEqualsLiteral("allow-forms")) {
        out &= ~SANDBOXED_FORMS;
      } else if (token.LowerCaseEqualsLiteral("allow-scripts")) {
        // allow-scripts removes both SANDBOXED_SCRIPTS and
        out &= ~SANDBOXED_SCRIPTS;
      } else if (token.LowerCaseEqualsLiteral("allow-top-navigation")) {
      } else if (token.LowerCaseEqualsLiteral("allow-pointer-lock")) {

  return out;

First all the flags are turned ON.
Then it checks if the sandbox attribute has any values, if it does it splits them and compares against the possible flags.
Once it finds a match, it does a BIT NEGATION on the flag and a BIT AND with the integer that has all the other flags.
What happens is that the flag being parsed is turned OFF.

In the end the integer with the status of all the flags is returned.

Locking the pointer

Now lets take a look at the code that checks for the allow-pointer-lock flag when an element requests pointerlock

nsDocument::ShouldLockPointer(Element* aElement)
  // Check if pointer lock pref is enabled
  if (!Preferences::GetBool("full-screen-api.pointer-lock.enabled")) {
    NS_WARNING("ShouldLockPointer(): Pointer Lock pref not enabled");
    return false;

  if (aElement != GetFullScreenElement()) {
    NS_WARNING("ShouldLockPointer(): Element not in fullscreen");
    return false;

  if (!aElement->IsInDoc()) {
    NS_WARNING("ShouldLockPointer(): Element without Document");
    return false;

  if (mSandboxFlags & SANDBOXED_POINTER_LOCK) {
    NS_WARNING("ShouldLockPointer(): Document is sandboxed and doesn't allow pointer-lock");
    return false;

  // Check if the element is in a document with a docshell.
  nsCOMPtr ownerDoc = aElement->OwnerDoc();
  if (!ownerDoc) {
    return false;
  if (!nsCOMPtr(ownerDoc->GetContainer())) {
    return false;
  nsCOMPtr ownerWindow = ownerDoc->GetWindow();
  if (!ownerWindow) {
    return false;
  nsCOMPtr ownerInnerWindow = ownerDoc->GetInnerWindow();
  if (!ownerInnerWindow) {
    return false;
  if (ownerWindow->GetCurrentInnerWindow() != ownerInnerWindow) {
    return false;

  return true;

The ShouldLockPointer method is called every time an element requests pointerlock, the method does some sanity checks and makes sure everything is correct.
To check for the allow-pointer-lock sandbox flag, a BIT AND with the mSandBoxFlags and the SANDBOX_POINTER_LOCK const is performed, we’ve looked at the SANDBOX_POINTER_LOCK flag before, it has the value of 0x80
So if pointerlock is allowed, the mSandboxFlags would have the SANDBOX_POINTER_LOCK flag OFF and the BIT AND would be false.

A big thanks to Ian Melven.
Ian is the one who implemented the sandbox attribute on Firefox and gave me some guidance on the PointerLock sandbox attribute bug.

DPS915 Workshop 1 – Initial Profile

Int the first workshop for the DPS915 course(Parallel Programming Fundamentals) we had to profile a simple application.
I wrote a previous blog post listing the steps to profile an application on osx.

The application we had to profile was:

// Profile a Serial Application - Workshop 1
 // w1.cpp

 #include <iostream>
 #include <iomanip>
 #include <cstdlib>
 #include <ctime>
 using namespace std;

 void init(float** a, int n) {
     float f = 1.0f / RAND_MAX;
     for (int i = 0; i < n; i++)
         for (int j = 0; j < n; j++)
             a[i][j] = rand() * f;

 void add(float** a, float** b, float** c, int n) {
     for (int i = 0; i < n; i++)
         for (int j = 0; j < n; j++)
             c[i][j] = a[i][j] + 3.0f * b[i][j];

 void multiply(float** a, float** b, float** c, int n) {
     for (int i = 0; i < n; i++)
         for (int j = 0; j < n; j++) {
             float sum = 0.0f;
             for (int k = 0; k < n; k++)
                 sum += a[i][k] * b[k][j];
             c[i][j] = sum;

 int main(int argc, char* argv[]) {
     // start timing
     time_t ts, te;
     ts = time(nullptr);

     // interpret command-line arguments
     if (argc != 3) {
         cerr << "**invalid number of arguments**" << endl;
         return 1;
     int n  = atoi(argv[1]);   // size of matrices
     int nr = atoi(argv[2]);   // number of runs

     float** a = new float*[n];
     for (int i = 0; i < n; i++)
        a[i] = new float[n];
     float** b = new float*[n];
     for (int i = 0; i < n; i++)
        b[i] = new float[n];
     float** c = new float*[n];
     for (int i = 0; i < n; i++)
        c[i] = new float[n];
     init(a, n);
     init(b, n);

     for (int i = 0; i < nr; i++) {
         add(a, b, c, n);
         multiply(a, b, c, n);

     for (int i = 0; i < n; i++)
        delete [] a[i];
     delete [] a;
     for (int i = 0; i < n; i++)
        delete [] b[i];
     delete [] b;
     for (int i = 0; i < n; i++)
        delete [] c[i];
     delete [] c;

     // elapsed time
     te = time(nullptr);
     cout << setprecision(0);
     cout << "Elapsed time : " << difftime(te, ts) << endl;

We had to run the application with 12 different combinations to see how much time the program spent executing the “add” and “multiply” functions.

Here is the profile results:

To easy the process of generating the profile data, I create a bash script to automate the runs:


# First Set



# Second Set



# Third Set



# Fourth Set



if [ $(uname) = "Darwin" ]

echo "OS $OS"

OPTIONS="-std=c++0x -O2 -g -pg"

INSTRUMENT_TEMPLATE="/Applications/Xcode.app/Contents/Applications/Instruments.app/Contents/Resources/templates/Time Profiler.tracetemplate"
#compile workshop

#generate profile info
for i in {0..11}
echo "Running ${i}th set"
  if [ $OS = "mac" ]
echo "Running on MacOS"
    instruments -t "$INSTRUMENT_TEMPLATE" -D results/mac/"${N[$i]}x${NR[$i]}.log" $OBJ ${N[$i]} ${NR[$i]}
echo "Running some linux distro."
    ./$OBJ ${N[$i]} ${NR[$i]}
    gprof -p $OBJ > "results/linux/${N[$i]}x${NR[$i]}.log"

The script works both on mac and linux.
If it’s running on a mac, it uses the Instruments Time Profiler, on a linux distro it uses gprof.

I’m committing all my course work to github

Any suggestions are more than welcome 🙂

Using Instruments Time Profiler

Gprof problem

On OSX 10.8.1 (Mountain Lion) the gnu profiling tool wasn’t working.
I’ve looked it up online and there was very little documentation about the problem.
I read in a couple of places saying that gprof in fact didn’t work but I couldn’t find any final answers.
Basically what happened is that when the program was compiled with the “pg” option, the gmon.out file was not created, thus not being able to run gprof to gather profile information for a specific program.

At first I thought the problem could be related to the fact that I was running gcc 4.2.1(the one that comes by default with XCode) so I tried to compile the latest version of gcc from source to check if it solved the problem.
I compiled gcc version 4.7.1. However it didn’t fix the problem.

I even try linking the profiling lib manually, but the gmon.out file was not being created.

**I’m still trying to find why the gmon.out file wasn’t being created, if anybody knows the reason or have any suggestions please leave a comment below.
My next step will be to compile the libc from source to add some profile symbols.
I’m following these references:

A couple of resources that are not related to gprof but nevertheless very useful:


Time Profiler

With all that being said, I needed to profile a c++ program on the mac, so I went looking for alternatives.

Luckly, I found that XCode comes with some extra tools called Instruments
A few tools included in the Instruments toolset are:

  • Allocations
  • Leaks
  • Activity Monitor
  • Time Profiler
  • System Trace
  • Automation
  • Energy Diagnotics

To get started with the Time Profiler is very simple, you first need to create a Xcode project.

Select the Profile option under Product (Command + I)

Select the Time Profiler template

Finally it will display the profile of your application

So far so good, I managed to generate profile information for my application. However, what if I wanted to get the information via the command line?
In my case I had to run the same application several times with different arguments to inspect how some functions behaved in certain situations and if they needed some optimizations.
With that in mind, running the time profiler via XCode was out of the question since I would need to manually modify the arguments and run the profiler each single time.
Instead I created a bash script to automate the runs.

Now I needed to find how to run the Instruments Time Profiler via the command line.
It wasn’t easy, there is very few documentation online and the manual has some outdated information.
Instead of [-d document] the correct is [-D document]
Anyway, to run Instruments from the command line:

instruments -t PathToTemplate -D ProfileResults YourApplication [list of arguments]

To see a list with all the available templates:

instruments -s

The result is a trace file that will contain the information regarding the profiling of the application.

Building Firefox on Mountain Lion 10.8

All the work that I’ve done on Firefox so far has been on a linux box.
I bought a mac recently so I’m in the process of switching all my dev tools.
To build Firefox on a mac is almost as straight forward as building on a linux distro.

Here are the steps:


First you’ll need to install macports.
Download the pkg installer for Mountain Lion or whatever version you are running and install macports

After the installation you’ll need to restart your shell so the $PATH gets updated.
You can find more details here

Once macports is installed:

$ sudo port selfupdate
$ sudo port sync
$ sudo port install libidl autoconf213 yasm mercurial ccache

The commands above will install all the dependencies you need to build firefox.

**More info on how to configure ccache here


Next it’s time to checkout the source code.

hg clone http://hg.mozilla.org/mozilla-central

It might take a while to clone the whole repo.


Now that you have both the dev dependencies and the source code the last thing missing is a .mozconfig file.
Below is a default configuration:

ac_add_options --enable-debug
ac_add_options --enable-trace-malloc
ac_add_options --enable-accessibility
ac_add_options --enable-signmar

# Enable parallel compiling
mk_add_options MOZ_MAKE_FLAGS="-j12"

# Treat warnings as errors in directories with FAIL_ON_WARNINGS.
ac_add_options --enable-warnings-as-errors
ac_add_options --with-ccache

# Package js shell.

You can find more info about .mozconfig here


Now it is time to start building.

First run:

make -f client.mk configure

That will make sure everything is setup properly, if you don’t see any error messages then you can start the build:

make -f client.mk build > build.out

A trick is to redirect the output of make to a file, it not only makes it easier to spot errors but it also decreases the build time.

Depending on your computer the build might take some time, don’t expect the build to finish before 15min, it will probably take something between 30min to 2h


Once the build is done, you can run Firefox by going to dir obj-dir/dist/NightlyDebug.app/Contents/MacOS and launch the firefox executable.

Simple Firefox build
Mac OS X Build Prerequisites

Getting started with CUDA on OSX 10.8 – Driver Problems

To install all the dev dependencies for CUDA enabled GPUs is not that bad, I faced a few issues but overall the documentation is pretty good.

You can find more information about how to get started here, it has all the links for the download of the driver + toolkit + SDK for windows, linux and mac

They also posted a PDF giving detail instructions about how to install everything.


Road Blocks

I’m running a MacBook Pro 2012 that comes with a GeForce GTM 650M.
On their website, they have the driver version 4.2 for download. However, I can update the CUDA driver to version 5.0.24 through the CUDA Preferences window under the System Preferences tab.

So after following the instructions they have posted on the Get Started pdf, I would get the message “Driver not supported” when running the deviceQuery test script.
I looked up online and found that this problem usually happened when the driver had a lower version than the SDK, I thought it was weird since I had downloaded all files they had instructed on the website.

I started browsing on the System Preferences when I saw the CUDA preferences tab.
On the tab it had the option to update the driver.
After the update, my driver was on version 5.0.24, and the deviceQuery test would work. \o/

After running the deviceQuery test, they suggested to run the bandwithTest to make sure the communication with the GPU was working properly.
To my surprise, when I ran the bandwithTest the computer crashed, some weird noises came from the case and a kernel panic messaged appeared.

Interval Since Last Panic Report:  75 sec
Panics Since Last Report:          2
Anonymous UUID:                    CD3F065C-4392-433E-8B7B-9D466743EE14

Tue Sep 11 23:16:23 2012
panic(cpu 4 caller 0xffffff802e8b7b95): Kernel trap at 0xffffff7faef9d18e, type 14=page fault, registers:
CR0: 0x0000000080010033, CR2: 0xffffff8191902000, CR3: 0x000000006b34b06c, CR4: 0x00000000001606e0
RAX: 0xffffff815123d000, RBX: 0x00000000406c5000, RCX: 0x00000000101b1400, RDX: 0xffffff8043302374
RSP: 0xffffff815117b650, RBP: 0xffffff815117b650, RSI: 0xffffff8043302004, RDI: 0xffffff80432ff804
R8:  0x00000000003f6a01, R9:  0xffffff815117b664, R10: 0x0000000000ffffff, R11: 0xffffff8100d10004
R12: 0xffffff80432ff804, R13: 0xffffff8043302374, R14: 0x0000000000000000, R15: 0xffffff8043302004
RFL: 0x0000000000010206, RIP: 0xffffff7faef9d18e, CS:  0x0000000000000008, SS:  0x0000000000000010
Fault CR2: 0xffffff8191902000, Error code: 0x0000000000000002, Fault CPU: 0x4

I wasn’t sure if the kernel panic was connected with the driver update, so I went back and ran some other scripts that come with the CUDA SDK, I ran the particles, simpleGL, volumeRender and a few others, then to my surprise again, when I ran the mergeSort another kernel panic was generated.
By now I was starting to get worried, I went back to the scripts dir and run a few others to make sure my GPU was still functioning properly, I ran the particles, simpleGL, volumeRender and the clock script, and again, after starting the clock script another kernel panic.

Now I knew for sure something was wrong, that shouldn’t be happening.

It was almost 12pm and I was getting tired and frustrated.
I did the only logical thing left to do… google it.

I entered the search: “mac 2012 crash with cuda driver 5”



To my relief it appeared that the kernel panics were in fact a known problem with the CUDA driver version 5 for the MacBook pro 2012.
I found this post on Adobe’s blog explaining the issue.
Apparently having the “Automatic Graphics Switching” option enable causes some CUDA applications to crash.
Turning the option off solves the problem.

Without the automatic graphics switching ON I ran the bandwithTest, mergeSort and clock apps and they worked just fine.

That Adobe’s blog post was created on August 29, so I believe that a fix for this problem should be coming out very soon.
Only Mountain Lion (Mac OSX v10.8) and Lion (Mac OSX v10.7) are affected by this bug.