Native bindings to the EXT_direct_state_access extension.

This extension introduces a set of new "direct state access" commands (meaning no selector is involved) to access (update and query) OpenGL state that previously depended on the OpenGL state selectors for access. These new commands supplement the existing selector-based OpenGL commands to access the same state.

The intent of this extension is to make it more efficient for libraries to avoid disturbing selector and latched state. The extension also allows more efficient command usage by eliminating the need for selector update commands.

Two derivative advantages of this extension are 1) display lists can be executed using these commands that avoid disturbing selectors that subsequent commands may depend on, and 2) drivers implemented with a dual-thread partitioning with OpenGL command buffering from an application thread and then OpenGL command dispatching in a concurrent driver thread can avoid thread synchronization created by selector saving, setting, command execution, and selector restoration.

This extension does not itself add any new OpenGL state.

We call a state variable in OpenGL an "OpenGL state selector" or simply a "selector" if OpenGL commands depend on the state variable to determine what state to query or update. The matrix mode and active texture are both selectors. Object bindings for buffers, programs, textures, and framebuffer objects are also selectors.

We call OpenGL state "latched" if the state is set by one OpenGL command but then that state is saved by a subsequent command or the state determines how client memory or buffer object memory is accessed by a subsequent command. The array and element array buffer bindings are latched by vertex array specification commands to determine which buffer a given vertex array uses. Vertex array state and pixel pack/unpack state decides how client memory or buffer object memory is accessed by subsequent vertex pulling or image specification commands.

The existence of selectors and latched state in the OpenGL API reduces the number of parameters to various sets of OpenGL commands but complicates the access to state for layered libraries which seek to access state without disturbing other state, namely the state of state selectors and latched state. In many cases, selectors and latched state were introduced by extensions as OpenGL evolved to minimize the disruption to the OpenGL API when new functionality, particularly the pluralization of existing functionality as when texture objects and later multiple texture units, was introduced.

The OpenGL API involves several selectors (listed in historical order of introduction):

• The matrix mode.
• The current bound texture for each supported texture target.
• The active texture.
• The active client texture.
• The current bound program for each supported program target.
• The current bound buffer for each supported buffer target.
• The current GLSL program.
• The current framebuffer object.

The new selector-free update commands can be compiled into display lists.

The OpenGL API has latched state for vertex array buffer objects and pixel store state. When an application issues a GL command to unpack or pack pixels (for example, glTexImage2D or glReadPixels respectively), the current unpack and pack pixel store state determines how the pixels are unpacked from/packed to client memory or pixel buffer objects. For example, consider:

glPixelStorei(GL_UNPACK_SWAP_BYTES, GL_TRUE);
glPixelStorei(GL_UNPACK_ROW_LENGTH, 640);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 47);
glDrawPixels(100, 100, GL_RGB, GL_FLOAT, pixels);

The unpack swap bytes and row length state set by the preceding glPixelStorei commands (as well as the 6 other unpack pixel store state variables) control how data is read (unpacked) from buffer of data pointed to by pixels. The glBindBuffer command also specifies an unpack buffer object (47) so the pixel pointer is actually treated as a byte offset into buffer object 47.

When an application issues a command to configure a vertex array, the current array buffer state is latched as the binding for the particular vertex array being specified. For example, consider:

glBindBuffer(GL_ARRAY_BUFFER, 23);
glVertexPointer(3, GL_FLOAT, 12, pointer);

The glBindBuffer command updates the array buffering binding (GL_ARRAY_BUFFER_BINDING) to the buffer object named 23. The subsequent glVertexPointer command specifies explicit parameters for the size, type, stride, and pointer to access the position vertex array BUT ALSO latches the current array buffer binding for the vertex array buffer binding (GL_VERTEX_ARRAY_BUFFER_BINDING). Effectively the current array buffer binding buffer object becomes an implicit fifth parameter to glVertexPointer and this applies to all the gl*Pointer vertex array specification commands.

Selectors and latched state create problems for layered libraries using OpenGL because selectors require the selector state to be modified to update some other state and latched state means implicit state can affect the operation of commands specifying, packing, or unpacking data through pointers/offsets. For layered libraries, a state update performed by the library may attempt to save the selector state, set the selector, update/query some state the selector controls, and then restore the selector to its saved state. Layered libraries can skip the selector save/restore but this risks introducing uncertainty about the state of a selector after calling layered library routines. Such selector side-effects are difficult to document and lead to compatibility issues as the layered library evolves or its usage varies. For latched state, layered libraries may find commands such as glDrawPixels do not work as expected because latched pixel store state is not what the library expects. Querying or pushing the latched state, setting the latched state explicitly, performing the operation involving latched state, and then restoring or popping the latched state avoids entanglements with latched state but at considerable cost.

EXAMPLE USAGE OF THIS EXTENSION'S FUNCTIONALITY

Consider the following routine to set the modelview matrix involving the matrix mode selector:

void setModelviewMatrix(const GLfloat matrix[16])
{
    GLenum savedMatrixMode;

    glGetIntegerv(GL_MATRIX_MODE, &savedMatrixMode);
    glMatrixMode(GL_MODELVIEW);
    glLoadMatrixf(matrix);
    glMatrixMode(savedMatrixMode);
}

Notice that four OpenGL commands are required to update the current modelview matrix without disturbing the matrix mode selector.

OpenGL query commands can also substantially reduce the performance of modern OpenGL implementations which may off-load OpenGL state processing to another CPU core/thread or to the GPU itself.

An alternative to querying the selector is to use the glPushAttrib/glPopAttrib commands. However this approach typically involves pushing far more state than simply the one or two selectors that need to be saved and restored. Because so much state is associated with a given push/pop attribute bit, the glPushAttrib and glPopAttrib commands are considerably more costly than the save/restore approach. Additionally glPushAttrib risks overflowing the attribute stack.

The reliability and performance of layered libraries and applications can be improved by adding to the OpenGL API a new set of commands to access directly OpenGL state that otherwise involves selectors to access.

The above example can be reimplemented more efficiently and without selector side-effects:

void setModelviewMatrix(const GLfloat matrix[16])
{
    glMatrixLoadfEXT(GL_MODELVIEW, matrix);
}

Consider a layered library seeking to load a texture:

void loadTexture(GLint texobj, GLint width, GLint height, void *data)
{
    glBindTexture(GL_TEXTURE_2D, texobj);
    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB8, width, height, GL_RGB, GL_FLOAT, data);
}

The library expects the data to be packed into the buffer pointed to by data. But what if the current pixel unpack buffer binding is not zero so the current pixel unpack buffer, rather than client memory, will be read? Or what if the application has modified the GL_UNPACK_ROW_LENGTH pixel store state before loadTexture is called?

We can fix the routine by calling glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0) and setting all the pixel store unpack state to the initial state the loadTexture routine expects, but this is expensive. It also risks disturbing the state so when loadTexture returns to the application, the application doesn't realize the current texture object (for whatever texture unit the current active texture happens to be) and pixel store state has changed.

We can more efficiently implement this routine without disturbing selector or latched state as follows:

void loadTexture(GLint texobj, GLint width, GLint height, void *data)
{
    glPushClientAttribDefaultEXT(GL_CLIENT_PIXEL_STORE_BIT);
    glTextureImage2D(texobj, GL_TEXTURE_2D, 0, GL_RGB8, width, height, GL_RGB, GL_FLOAT, data);
    glPopClientAttrib();
}

Now loadTexture does not have to worry about inappropriately configured pixel store state or a non-zero pixel unpack buffer binding. And loadTexture has no unintended side-effects for selector or latched state (assuming the client attrib state does not overflow).