文本文件  |  80行  |  3.88 KB

Design of the GLES Tracing Library

Code Runtime Behavior:

    Initialization:
    
    egl_display_t::initialize() calls initEglTraceLevel() to figure out whether tracing should be
    enabled. Currently, the shell properties "debug.egl.trace" and "debug.egl.debug_proc" together
    control whether tracing should be enabled for a certain process. If tracing is enabled, this
    calls GLTrace_start() to start the trace server.
    
    egl_display_t::initialize() then calls setGLHooksThreadSpecific() where we set the thread
    specific gl_hooks structure to point to the trace implementation. From this point on, every
    GLES call is redirected to the trace implementation.
    
    Application runtime:

    While the application is running, all its GLES calls are directly routed to their corresponding
    trace implementation.

    For EGL calls, the trace library provides a bunch of functions that must be explicitly called
    from the EGL library. These functions are declared in glestrace.h

    Application shutdown:

    Currently, the application is killed when the user stops tracing from the frontend GUI. We need
    to explore if a more graceful method of stopping the application, or detaching tracing from the
    application is required.


Enabling tracing while the application is running:

    In order to allow tracing of an already running application, we allow DdmServer to enable
    OpenGL tracing. In such a case, the application already has its GL hooks set up to point to the
    real GL implementation, and we need to switch them to point to the trace implementation.

    This is achieved by checking whether tracing should be enabled at every eglSwap call.
    (Note: We were already checking for tracing at every eglSwap, the only change now is that
    the tracing could actually be ON/OFF at runtime - earlier it was set once and never changed).

    If eglSwap detects that tracing should be enabled now, then it performs the following steps:
        - switch the gl hooks to point to the trace implementation.
        - call trace eglMakeCurrent to indicate that there is now a new context that is current.
        - continue on with tracing the eglSwap call.
    This switches the hooks to point to the trace implementation only for the current context.
    But the other contexts have their gl hooks updated when they perform eglMakeCurrent.

    The GLTrace version of eglMakeCurrent now has to be updated to allow switching to a context
    it may not know of. In such a case, it creates a context matching the version that it is now
    switching to.

Disabling tracing:

    We disable tracing under two conditions:
        - stop tracing request from DdmServer
        - gltrace transport gets disconnected from the host.
    In either case, both actions simply disable the tracing flag. The current context gets its
    gl hooks restored in the next eglSwap, and the other traced contexts get their gl hooks
    restored when they perform a eglMakeCurrent.

Code Structure:

    glestrace.h declares all the hooks exposed by libglestrace. These are used by EGL/egl.cpp and
    EGL/eglApi.cpp to initialize the trace library, and to inform the library of EGL calls.

    All GL calls are present in GLES_Trace/src/gltrace_api.cpp. This file is generated by the
    GLES_Trace/src/genapi.py script. The structure of all the functions looks like this:

            void GLTrace_glFunction(args) {
                // declare a protobuf
                // copy arguments into the protobuf
                // call the original GLES function
                // if there is a return value, save it into the protobuf
                // fixup the protobuf if necessary
                // transport the protobuf to the host
            }

    The fixupGLMessage() call does any custom processing of the protobuf based on the GLES call.
    This typically amounts to copying the data corresponding to input or output pointers.