Improvement LD_LIBRARY_PATH explanation.

[apitrace] / README.markdown

About **apitrace**
==================

**apitrace** consists of a set of tools to:

* trace OpenGL, OpenGL ES, Direct3D, and DirectDraw APIs calls to a file;

* retrace OpenGL and OpenGL ES calls from a file;

* inspect OpenGL state at any call while retracing;

* visualize and edit trace files.


Obtaining **apitrace**
======================

To obtain apitrace either [download the latest
binaries](https://github.com/apitrace/apitrace/downloads) for your platform if
available, or follow the instructions in INSTALL.markdown to build it yourself.
On 64bits Linux and Windows platforms you'll need apitrace binaries that match
the architecture (32bits or 64bits) of the application being traced.


Basic usage
===========

Run the application you want to trace as

    apitrace trace --api API /path/to/application [args...]

and it will generate a trace named `application.trace` in the current
directory.  You can specify the written trace filename by passing the
`--output` command line option.

Problems while tracing (e.g, if the application uses calls/parameters
unsupported by apitrace) will be reported via stderr output on Unices.  On
Windows you'll need to run
[DebugView](http://technet.microsoft.com/en-us/sysinternals/bb896647) to view
these messages.

Follow the "Tracing manually" instructions below if you cannot obtain a trace.

View the trace with

    apitrace dump application.trace

Replay an OpenGL trace with

    glretrace application.trace

Pass the `-sb` option to use a single buffered visual.  Pass `--help` to
glretrace for more options.


Basic GUI usage
===============

Start the GUI as

    qapitrace application.trace

You can also tell the GUI to go directly to a specific call

    qapitrace application.trace 12345


Advanced command line usage
===========================


Call sets
---------

Several tools take `CALLSET` arguments, e.g:

    apitrace dump --calls CALLSET foo.trace
    glretrace -S CALLSET foo.trace

The call syntax is very flexible. Here are a few examples:

 * `4`             one call

 * `1,2,4,5`       set of calls

 * `"1 2 4 5"`     set of calls (commas are optional and can be replaced with whitespace)

 * `1-100/2`       calls 1, 3, 5, ...,  99

 * `1-1000/draw`   all draw calls between 1 and 1000

 * `1-1000/fbo`    all fbo changes between calls 1 and 1000

 * `frame`         all calls at end of frames

 * `@foo.txt`      read call numbers from `foo.txt`, using the same syntax as above



Tracing manually
----------------

### Linux ###

On 64 bits systems, you'll need to determine ether the application is 64 bits
or 32 bits.  This can be done by doing

    file /path/to/application

But beware of wrapper shell scripts -- what matters is the architecture of the
main process.

Run the application you want to trace as

     LD_PRELOAD=/path/to/apitrace/wrappers/glxtrace.so /path/to/application

and it will generate a trace named `application.trace` in the current
directory.  You can specify the written trace filename by setting the
`TRACE_FILE` environment variable before running.

The `LD_PRELOAD` mechanism should work with the majority applications.  There
are some applications (e.g., Unigine Heaven, Android GPU emulator, etc.), that
have global function pointers with the same name as GL entrypoints, living in a
shared object that wasn't linked with `-Bsymbolic` flag, so relocations to
those globals function pointers get overwritten with the address to our wrapper
library, and the application will segfault when trying to write to them.  For
these applications it is possible to trace by using `glxtrace.so` as an
ordinary `libGL.so` and injecting it via `LD_LIBRARY_PATH`:

    ln -s glxtrace.so wrappers/libGL.so
    ln -s glxtrace.so wrappers/libGL.so.1
    ln -s glxtrace.so wrappers/libGL.so.1.2
    export LD_LIBRARY_PATH=/path/to/apitrace/wrappers:$LD_LIBRARY_PATH
    export TRACE_LIBGL=/path/to/real/libGL.so.1
    /path/to/application

If you are an application developer, you can avoid this either by linking with
`-Bsymbolic` flag, or by using some unique prefix for your function pointers.

See the `ld.so` man page for more information about `LD_PRELOAD` and
`LD_LIBRARY_PATH` environment flags.

To trace the application inside gdb, invoke gdb as:

    gdb --ex 'set exec-wrapper env LD_PRELOAD=/path/to/glxtrace.so' --args /path/to/application

### Android ###

The following instructions should work at least for Android Ice Scream
Sandwitch:

For standalone applications the instructions above for Linux should
work. To trace applications started from within the Android VM process
(`app_process` aka zygote) you'll have to wrap this process and enable
tracing dynamically for the application to be traced.

- Wrapping the android main VM process:

  In the Android root /init.rc add the `LD_PRELOAD` setting to zygote's
  environment in the 'service zygote' section:

        service zygote ...
           setenv LD_PRELOAD /data/egltrace.so
           ...

  Note that ICS will overwrite the /init.rc during each boot with the
  version in the recovery image. So you'll have to change the file in
  your ICS source tree, rebuild and reflash the device.
  Rebuilding/reflashing only the recovery image should be sufficient.

- Copy egltrace.so to /data

  On the host:

        adb push /path/to/apitrace/build/wrappers/egltrace.so /data

- Adjust file permissions to store the trace file:

  By default egltrace.so will store the trace in
  `/data/app_process.trace`. For this to work for applications running
  with a uid other than 0, you have to allow writes to the `/data`
  directory on the device:

        chmod 0777 /data

- Enable tracing for a specific process name:

  To trace for example the Settings application:

        setprop debug.apitrace.procname com.android.settings

  In general this name will match what `ps` reports.

- Start the application:

  If the application was already running, for example due to ICS's way
  of pre-starting the apps, you might have to kill the application
  first:

        kill <pid of app>

  Launch the application for example from the application menu.

### Mac OS X ###

Run the application you want to trace as

    DYLD_LIBRARY_PATH=/path/to/apitrace/wrappers /path/to/application

Note that although Mac OS X has an `LD_PRELOAD` equivalent,
`DYLD_INSERT_LIBRARIES`, it is mostly useless because it only works with
`DYLD_FORCE_FLAT_NAMESPACE=1` which breaks most applications.  See the `dyld` man
page for more details about these environment flags.

### Windows ###

When tracing third-party applications, you can identify the target
application's main executable, either by:

* right clicking on the application's icon in the _Start Menu_, choose
  _Properties_, and see the _Target_ field;

* or by starting the application, run Windows Task Manager (taskmgr.exe), right
  click on the application name in the _Applications_ tab, choose _Go To Process_,
  note the highlighted _Image Name_, and search it on `C:\Program Files` or
  `C:\Program Files (x86)`.

On 64 bits Windows, you'll need to determine ether the application is a 64 bits
or 32 bits. 32 bits applications will have a `*32` suffix in the _Image Name_
column of the _Processes_ tab of _Windows Task Manager_ window.

Copy the appropriate `opengl32.dll`, `d3d8.dll`, or `d3d9.dll` from the
wrappers directory to the directory with the application you want to trace.
Then run the application as usual.

You can specify the written trace filename by setting the `TRACE_FILE`
environment variable before running.


Emitting annotations to the trace
---------------------------------

From OpenGL applications you can embed annotations in the trace file through the
[`GL_GREMEDY_string_marker`](http://www.opengl.org/registry/specs/GREMEDY/string_marker.txt)
and
[`GL_GREMEDY_frame_terminator`](http://www.opengl.org/registry/specs/GREMEDY/frame_terminator.txt)
GL extensions.

**apitrace** will advertise and intercept these GL extensions independently of
the GL implementation.  So all you have to do is to use these extensions when
available.

For example, if you use [GLEW](http://glew.sourceforge.net/) to dynamically
detect and use GL extensions, you could easily accomplish this by doing:

    void foo() {
    
      if (GLEW_GREMEDY_string_marker) {
        glStringMarkerGREMEDY(0, __FUNCTION__ ": enter");
      }
      
      ...
      
      if (GLEW_GREMEDY_string_marker) {
        glStringMarkerGREMEDY(0, __FUNCTION__ ": leave");
      }
      
    }

This has the added advantage of working equally well with gDEBugger.


From OpenGL ES applications you can embed annotations in the trace file through the
[`GL_EXT_debug_marker`](http://www.khronos.org/registry/gles/extensions/EXT/EXT_debug_marker.txt)
extension.


For Direct3D applications you can follow the same procedure used for 
[instrumenting an application for PIX](http://technet.microsoft.com/en-us/query/ee417250)


Dump GL state at a particular call
----------------------------------

You can get a dump of the bound GL state at call 12345 by doing:

    glretrace -D 12345 application.trace > 12345.json

This is precisely the mechanism the GUI obtains its own state.

You can compare two state dumps by doing:

    apitrace diff-state 12345.json 67890.json


Comparing two traces side by side
---------------------------------

    apitrace diff trace1.trace trace2.trace

This works only on Unices, and it will truncate the traces due to performance
limitations.


Recording a video with FFmpeg
-----------------------------

You can make a video of the output by doing

    glretrace -s - application.trace \
    | ffmpeg -r 30 -f image2pipe -vcodec ppm -i pipe: -vcodec mpeg4 -y output.mp4


Trimming a trace
----------------

You can make a smaller trace by doing:

    apitrace trim --callset 100-1000 -o trimed.trace applicated.trace

If you need precise control over which calls to trim you can specify the
individual call numbers a plaintext file, as described in the 'Call sets'
section above.


Profiling a trace
-----------------

You can perform gpu and cpu profiling with the command line options:

 * `-pgpu` record gpu times for frames and draw calls.

 * `-pcpu` record cpu times for frames and draw calls.

 * `-ppd` record pixels drawn for each draw call.

The results from this can then be read by hand or analysed with a script.

`scripts/profileshader.py` will read the profile results and format them into a
table which displays profiling results per shader.

For example, to record all profiling data and utilise the per shader script:

    ./glretrace -pgpu -pcpu -ppd foo.trace | ./scripts/profileshader.py


Advanced usage for OpenGL implementors
======================================

There are several advanced usage examples meant for OpenGL implementors.


Regression testing
------------------

These are the steps to create a regression test-suite around **apitrace**:

* obtain a trace

* obtain reference snapshots, by doing on a reference system:

        mkdir /path/to/reference/snapshots/
        glretrace -s /path/to/reference/snapshots/ application.trace

* prune the snapshots which are not interesting

* to do a regression test, do:

        glretrace -c /path/to/reference/snapshots/ application.trace

  Alternatively, for a HTML summary, use `apitrace diff-images`:

        glretrace -s /path/to/test/snapshots/ application.trace
        apitrace diff-images --output summary.html /path/to/reference/snapshots/ /path/to/test/snapshots/


Automated git-bisection
-----------------------

With tracecheck.py it is possible to automate git bisect and pinpoint the
commit responsible for a regression.

Below is an example of using tracecheck.py to bisect a regression in the
Mesa-based Intel 965 driver.  But the procedure could be applied to any GL
driver hosted on a git repository.

First, create a build script, named build-script.sh, containing:

    #!/bin/sh
    set -e
    export PATH=/usr/lib/ccache:$PATH
    export CFLAGS='-g'
    export CXXFLAGS='-g'
    ./autogen.sh --disable-egl --disable-gallium --disable-glut --disable-glu --disable-glw --with-dri-drivers=i965
    make clean
    make "$@"

It is important that builds are both robust, and efficient.  Due to broken
dependency discovery in Mesa's makefile system, it was necessary invoke `make
clean` in every iteration step.  `ccache` should be installed to avoid
recompiling unchanged source files.

Then do:

    cd /path/to/mesa
    export LIBGL_DEBUG=verbose
    export LD_LIBRARY_PATH=$PWD/lib
    export LIBGL_DRIVERS_DIR=$PWD/lib
    git bisect start \
        6491e9593d5cbc5644eb02593a2f562447efdcbb 71acbb54f49089b03d3498b6f88c1681d3f649ac \
        -- src/mesa/drivers/dri/intel src/mesa/drivers/dri/i965/
    git bisect run /path/to/tracecheck.py \
        --precision-threshold 8.0 \
        --build /path/to/build-script.sh \
        --gl-renderer '.*Mesa.*Intel.*' \
        --retrace=/path/to/glretrace \
        -c /path/to/reference/snapshots/ \
        topogun-1.06-orc-84k.trace

The trace-check.py script will skip automatically when there are build
failures.

The `--gl-renderer` option will also cause a commit to be skipped if the
`GL_RENDERER` is unexpected (e.g., when a software renderer or another GL
driver is unintentionally loaded due to missing symbol in the DRI driver, or
another runtime fault).


Side by side retracing
----------------------

In order to determine which draw call a regression first manifests one could
generate snapshots for every draw call, using the `-S` option.  That is, however,
very inefficient for big traces with many draw calls.

A faster approach is to run both the bad and a good GL driver side-by-side.
The latter can be either a previously known good build of the GL driver, or a
reference software renderer.

This can be achieved with retracediff.py script, which invokes glretrace with
different environments, allowing to choose the desired GL driver by
manipulating variables such as `LD_LIBRARY_PATH`, `LIBGL_DRIVERS_DIR`, or
`TRACE_LIBGL`.

For example, on Linux:

    ./scripts/retracediff.py \
        --ref-env LD_LIBRARY_PATH=/path/to/reference/GL/implementation \
        --retrace /path/to/glretrace \
        --diff-prefix=/path/to/output/diffs \
        application.trace

Or on Windows:

    python scripts\retracediff.py --retrace \path\to\glretrace.exe --ref-env TRACE_LIBGL=\path\to\reference\opengl32.dll application.trace


Links
=====

About **apitrace**:

* [Official mailing list](http://lists.freedesktop.org/mailman/listinfo/apitrace)

* [Zack Rusin's blog introducing the GUI](http://zrusin.blogspot.com/2011/04/apitrace.html)

* [Jose's Fonseca blog introducing the tool](http://jrfonseca.blogspot.com/2008/07/tracing-d3d-applications.html)


Direct3D
--------

Open-source:

* [Proxy DLL](http://www.mikoweb.eu/index.php?node=21)

  * [Intercept Calls to DirectX with a Proxy DLL](http://www.codeguru.com/cpp/g-m/directx/directx8/article.php/c11453/)

* [Direct3D 9 API Interceptor](http://graphics.stanford.edu/~mdfisher/D3D9Interceptor.html)

Closed-source:

* [Microsoft PIX](http://msdn.microsoft.com/en-us/library/ee417062.aspx)

  * [D3DSpy](http://doc.51windows.net/Directx9_SDK/?url=/directx9_sdk/graphics/programmingguide/TutorialsAndSamplesAndToolsAndTips/Tools/D3DSpy.htm): the predecessor of PIX

* [NVIDIA PerfKit](http://developer.nvidia.com/nvidia-perfkit)

* [AMD GPU PerfStudio](http://developer.amd.com/gpu/PerfStudio/pages/APITraceWindow.aspx)

* [Intel Graphics Performance Analyzers](http://www.intel.com/software/gpa/)


OpenGL
------

Open-source:

* [BuGLe](http://www.opengl.org/sdk/tools/BuGLe/)

* [GLIntercept](http://code.google.com/p/glintercept/)

* [tracy](https://gitorious.org/tracy): OpenGL ES and OpenVG trace, retrace, and state inspection

* [WebGL-Inspector](http://benvanik.github.com/WebGL-Inspector/)

Closed-source:

* [AMD CodeXL](http://developer.amd.com/tools/hc/CodeXL/Pages/default.aspx)

* [AMD GPU PerfStudio](http://developer.amd.com/gpu/PerfStudio/pages/APITraceWindow.aspx)

* [gDEBugger](http://www.gremedy.com/products.php) and [AMD gDEBugger](http://developer.amd.com/tools/gDEBugger/Pages/default.aspx)

* [glslDevil](http://cumbia.informatik.uni-stuttgart.de/glsldevil/index.html)