Although focus was and still is on graphical APIs, apitrace has a generic infrastructure to trace any kind of API:
-
the APIs types and calls are specified in Python files in specs sub-directory;
- there is a type hierarchy in
specs/stdapi.py
, capable of representing most types in C language, and additional semantic metadata
- there is a type hierarchy in
-
Python scripts generate C++ code to trace and serialize calls parameters to a file, and vice-versa.
-
The Visitor design pattern is used to navigate over the types.
-
The Template design pattern is used to enable any step of code generation to be overriden by derived classes, allowing to handle cases that need special treatment without sacrifycing code reuse.
-
apitrace's architecture is composed of several layers. Too many to show in a single graph, so only those relevant for OpenGL API are shown below:
specs
^
|
dispatch <-------------- glws
^ ^
| /
helpers <--- glstate /
^ ^ ^ /
/ \ | /
/ \ | /
trace retrace | /
^ ^ | /
/ \ | /
gltrace glretrace
^ ^ ^ ^
/ | \ \
/ | \ \
/ | \ \
/ | \ \
/ | \ \
glxtrace wgltrace cgltrace qapitrace
Here is a quick synopsis of what the layers do:
-
specs
-- specification of the API, expressed in a Python class hierarchy -
dispatch
-- runtime dispatch of calls to DLLs (open the DLL, get the symbol address, and call it passing all arguments as-is) -
helpers
-- helper functions to determine sizes of arrays, blobs, etc. It often needs to dispatch calls to give the answers. -
trace
-- generate C++ code for tracing an API based on its spec-
gltrace
-- specialization of the tracing generation for GL API, with extra code to generate -
glxtrace
,wgltrace
,cgltrace
-- specialization of the tracing code for the GLX, WGL, and CGL APIs.
-
-
retrace
-- generate C++ code to interpret calls from trace, based on the API's specglretrace
-- specialization of the retrace code for the GL API
-
glstate
-- code to dump OpenGL state to a JSON file -
glws
-- abstraction of the window system specific APIs (GXL, WGL, CGL), to enable cross-platform portability ofglretrace
-
qapitrace
-- the GUI; it reads traces directly, and gets JSON state by invokingglretrace
These are guidelines for new code. Admittedly some of the existing code hasn't been updated to follow these conventions yet.
Whitespace (all languages):
-
indentation is 4 spaces
-
never use tabs as indents, except on Makefiles
-
otherwise tab equals to 8 spaces
-
separate classes with two empty lines
Naming convention:
-
camelCase
for functions/methods -
UpperCase
for structures/classes -
lowercase
for namespaces/modules -
UPPER_CASE
for #defines -
single underscore prefix for variables/functions in automatically generated code
C++:
-
enclose single statement
if
clauses in{
}
, specially for automatically generated code -
} else {
-
use
inline
keyword for functions/methods which are called with high-frequency
CMake:
-
lower_case
commands -
space between
(
and precedent name
And when in doubt, be consistent with the existing code.
Feature development:
-
Existing features in master branch should not degrade at any time, for any platform. Unless they are seldom used or redundant, and there is agreement.
-
In particular, new features / changes must not introduce any sort of instability when tracing.
While application developers and driver developers may be able to workaround quirks in apitrace, we want to be able to obtain traces from non-technical end-users with minimal intervention.
This implies that tracing should not make any non-standard assumptions, and care must be taken to ensure the tracing code is robust against invalid parameters, multiple threads, etc.
-
-
It's fine to add new features for only some platforms or APIs.
-
Non-trivial changes should be staged in a branch, to allow review and regression testing. Feature branches should be deleted once they have been merged.
-
Releases are tagged commits from master. There are no stable branches.
Backwards compatibility:
-
Backwards binary compatibility with old traces must be always maintained: all tools, including
glretrace
, must handle old traces without regressions. -
No backwards compatibility guarantees for derived data (ASCII dumps, state, images, etc).
-
There should be no gratuitous changes to command line tool interfaces, but no guarantees are given.
There is a regression test suite under development in https://github.com/apitrace/apitrace-tests .
All OpenGL (and OpenGL ES) function prototypes live in specs/glapi.py
. This
file is semi-automatically derived from Khronos XML description of OpenGL /
OpenGL ES. To refresh do
$ make -C specs/scripts/ glapi.py
$ meld specs/glapi.py specs/scripts/glapi.py
and then port over new prototypes. See also specs/scripts/README.markdown
.
The upstream XML description is not rich enough to describe all semantic details that apitrace needs, therefore one needs to manually tweak the specifications:
-
Fix-up the types of array, blob, and pointer arguments.
-
For
glGet*
you can use"_gl_param_size(pname)"
for automatically determining the number of parameters written, e.g.GlFunction(Void, "glGetIntegerv", [(GLenum, "pname"), Out(Array(GLint, "_gl_param_size(pname)"), "params")], sideeffects=False),
-
-
Add the
sideeffects=False
keyword argument where appropriate, so that those calls can be merely ignored byglretrace
. -
Replace generically type
GLuint
object IDs with typed ones, (e.g., replace(GLuint, "texture")
into(GLtexture, "texture")
, so thatglretrace
can swizzle the objects IDs, when replaying on a different OpenGL implementation.