That said, feel free to file an issue and/or send an email to the mailing list
if:
-* send an email to the mailing list if you want discuss the rationale, propose
- your ideas on how to address it, or volunteer to work on it;
+* you want discuss the rationale, propose your ideas on how to address it, or
+ volunteer to work on it;
-* file the issue in the issue tracker (or comment to it if it already exists)
- if it is important for you and you would like to see it addressed sooner
- rather than later.
+* a known issue is important for you and you would like to see it addressed
+ sooner rather than later.
Tracing
There is no way to distinguish the fake calls from those actually
made by the application yet.
+* Huge traces will be generated for applications that make extensive use of
+ immediate mode drawing (ie., glBegin/glEnd), use vertex/element arrays in
+ user memory, or generate a lot of dynamic vertex data per frame. This is
+ because large quantities of (often redundant) data will be recorded over and
+ over; and although the traces are compressed, the compression algorithms used
+ aim a good performance/compression balance, hence fail to identify and remove
+ the redundancy at this scale.
+
+ However, this should not be a problem for modern OpenGL applications that
+ make an efficient use of VBOs and vertex shaders.
+
* On MacOSX, the internal OpenGL calls done by GLU are not traced yet.
Retracing
---------
+* Replaying can take substantially more CPU due to the overhead of reading the
+ trace file from disk.
+
+ However, the overhead should not be significant for modern 3D applications
+ that do efficient use of VBOs and vertex shaders. The communication between
+ CPU to GPU can easily be a bottleneck on the powerful discrete GPUs of
+ nowadays, and trace overhead tend to be propotional so it is not a
+ coincidence that applications that are optimized for modern GPUs will also be
+ traced and replayed efficiently.
+
* glretrace needs to infer window sizes from glViewport calls, because calls
that create/resize windows happen outside of OpenGL and are not traced.
Therefore window will be missing if the application relies on the default
a standard 32bit RGBA, 24bit depth, 8bit alpha, double buffer visual. Unless
overridden on command line.
-* Multi-threaded OpenGL is not yet supported.
-
-* OpenGL context sharing is not fully respected -- all contexts are expected to
- share state, and most likely there
+* OpenGL context sharing is not fully respected -- all contexts are assumed to
+ share state. This is by far the most common case though.
GUI
Please attach long logs to https://gist.github.com/ and paste the URL into the
issue description.
-For big attachments, such as traces, please upload it temporarily to a web
-server you control, or use a file upload service such as
-http://www.megaupload.com/ or http://dropbox.com/ and paste the URL into the
-issue description.
+For big attachments, such as traces, please upload them temporarily to a web
+server you control, or use a file upload service such as http://dropbox.com/
+and paste the URL into the issue description.
Trace files are only slightly compressed (for performance reasons). You can
further reduce their size when attaching/uploading by compressing with
------------
Please rebuild apitrace with debugging information, by passing
-`-DCMAKE_BUILD_TYPE=Debug` to cmake.
+`-DCMAKE_BUILD_TYPE=Debug` to cmake, or editing its value in `CMakeCache.txt`
+and rebuilding.
To obtain a stack back-trace, run the application with gdb from a terminal:
...
(gdb) bt
+On Linux, to trace an application inside gdb the `LD_PRELOAD` environment
+variable should be set from within gdb like:
+
+ gdb --ex 'set exec-wrapper env LD_PRELOAD=/path/to/glxtrace.so' --args application arg1 ...
See also more detailed and Distro specific instructions: