OpenGL.GL.ARB.framebuffer_object

OpenGL extension ARB.framebuffer_object

This module customises the behaviour of the OpenGL.raw.GL.ARB.framebuffer_object to provide a more Python-friendly API

Overview (from the spec)

ARB_framebuffer_object is an extension intended to address the following goals:

EXT_framebuffer_object EXT_framebuffer_blit EXT_framebuffer_multisample EXT_packed_depth_stencil

ARB_framebuffer_object defines an interface for drawing to rendering destinations other than the buffers provided to the GL by the window-system.

In this extension, these newly defined rendering destinations are known collectively as "framebuffer-attachable images". This extension provides a mechanism for attaching framebuffer-attachable images to the GL framebuffer as one of the standard GL logical buffers: color, depth, and stencil. (Attaching a framebuffer-attachable image to the accum logical buffer is left for a future extension to define). When a framebuffer-attachable image is attached to the framebuffer, it is used as the source and destination of fragment operations as described in Chapter 4.

By allowing the use of a framebuffer-attachable image as a rendering destination, this extension enables a form of "offscreen" rendering. Furthermore, "render to texture" is supported by allowing the images of a texture to be used as framebuffer-attachable images. A particular image of a texture object is selected for use as a framebuffer-attachable image by specifying the mipmap level, cube map face (for a cube map texture), and layer (for a 3D texture) that identifies the image. The "render to texture" semantics of this extension are similar to performing traditional rendering to the framebuffer, followed immediately by a call to CopyTexSubImage. However, by using this extension instead, an application can achieve the same effect, but with the advantage that the GL can usually eliminate the data copy that would have been incurred by calling CopyTexSubImage.

This extension also defines a new GL object type, called a "renderbuffer", which encapsulates a single 2D pixel image. The image of renderbuffer can be used as a framebuffer-attachable image for generalized offscreen rendering and it also provides a means to support rendering to GL logical buffer types which have no corresponding texture format (stencil, accum, etc). A renderbuffer is similar to a texture in that both renderbuffers and textures can be independently allocated and shared among multiple contexts. The framework defined by this extension is general enough that support for attaching images from GL objects other than textures and renderbuffers could be added by layered extensions.

To facilitate efficient switching between collections of framebuffer-attachable images, this extension introduces another new GL object, called a framebuffer object. A framebuffer object contains the state that defines the traditional GL framebuffer, including its set of images. Prior to this extension, it was the window-system which defined and managed this collection of images, traditionally by grouping them into a "drawable". The window-system API's would also provide a function (i.e., wglMakeCurrent, glXMakeCurrent, aglSetDrawable, etc.) to bind a drawable with a GL context (as is done in the WGL_ARB_pbuffer extension). In this extension however, this functionality is subsumed by the GL and the GL provides the function BindFramebufferARB to bind a framebuffer object to the current context. Later, the context can bind back to the window-system-provided framebuffer in order to display rendered content.

Previous extensions that enabled rendering to a texture have been much more complicated. One example is the combination of ARB_pbuffer and ARB_render_texture, both of which are window-system extensions. This combination requires calling MakeCurrent, an operation that may be expensive, to switch between the window and the pbuffer drawables. An application must create one pbuffer per renderable texture in order to portably use ARB_render_texture. An application must maintain at least one GL context per texture format, because each context can only operate on a single pixelformat or FBConfig. All of these characteristics make ARB_render_texture both inefficient and cumbersome to use.

ARB_framebuffer_object, on the other hand, is both simpler to use and more efficient than ARB_render_texture. The ARB_framebuffer_object API is contained wholly within the GL API and has no (non-portable) window-system components. Under ARB_framebuffer_object, it is not necessary to create a second GL context when rendering to a texture image whose format differs from that of the window. Finally, unlike the pbuffers of ARB_render_texture, a single framebuffer object can facilitate rendering to an unlimited number of texture objects.

This extension differs from EXT_framebuffer_object by splitting the framebuffer object binding point into separate DRAW and READ bindings (incorporating functionality introduced by EXT_framebuffer_blit). This allows copying directly from one framebuffer to another. In addition, a new high performance blit function is added to facilitate these blits and perform some data conversion where allowed.

This extension also enables usage of multisampling in conjunction with renderbuffers (incorporating functionality from EXT_packed_depth_stencil), as follows:

The new operation RenderbufferStorageMultisample() allocates storage for a renderbuffer object that can be used as a multisample buffer. A multisample render buffer image differs from a single-sample render buffer image in that a multisample image has a number of SAMPLES that is greater than zero. No method is provided for creating multisample texture images.

All of the framebuffer-attachable images attached to a framebuffer object must have the same number of SAMPLES or else the framebuffer object is not "framebuffer complete". If a framebuffer object with multisample attachments is "framebuffer complete", then the framebuffer object behaves as if SAMPLE_BUFFERS is one.

In traditional multisample rendering, where DRAW_FRAMEBUFFER_BINDING is zero and SAMPLE_BUFFERS is one, the GL spec states that "the color sample values are resolved to a single, displayable color each time a pixel is updated." There are, however, several modern hardware implementations that do not actually resolve for each sample update, but instead postpones the resolve operation to a later time and resolve a batch of sample updates at a time. This is OK as long as the implementation behaves "as if" it had resolved a sample-at-a-time. Unfortunately, however, honoring the "as if" rule can sometimes degrade performance.

In contrast, when DRAW_FRAMEBUFFER_BINDING is an application-created framebuffer object, MULTISAMPLE is enabled, and SAMPLE_BUFFERS is one, there is no implicit per-sample-update resolve. Instead, the application explicitly controls when the resolve operation is performed. The resolve operation is affected by calling BlitFramebuffer where the source is a multisample application-created framebuffer object and the destination is a single-sample framebuffer object (either application-created or window-system provided).

This design for multisample resolve more closely matches current hardware, but still permits implementations which choose to resolve a single sample at a time. If hardware that implements the multisample resolution "one sample at a time" exposes ARB_framebuffer_object, it could perform the implicit resolve to a driver-managed hidden surface, then read from that surface when the application calls BlitFramebuffer.

Another motivation for granting the application explicit control over the multisample resolve operation has to do with the flexibility afforded by ARB_framebuffer_object. Previously, a drawable (window or pbuffer) had exclusive access to all of its buffers. There was no mechanism for sharing a buffer across multiple drawables. Under ARB_framebuffer_object, however, a mechanism exists for sharing a framebuffer-attachable image across several framebuffer objects, as well as sharing an image between a framebuffer object and a texture. If we had retained the "implicit" resolve from traditional multisampled rendering, and allowed the creation of "multisample" format renderbuffers, then this type of sharing would have lead to two problematic situations:

Using BlitFramebuffer as an explicit resolve to serialize access to the multisampled contents and eliminate the implicit per-sample resolve operation, we avoid both of these problems.

This extension also enables usage of packed depth-stencil formats in renderbuffers (incorporating functionality from EXT_packed_depth_stencil), as follows:

Many OpenGL implementations have chosen to interleave the depth and stencil buffers into one buffer, often with 24 bits of depth precision and 8 bits of stencil data. 32 bits is more than is needed for the depth buffer much of the time; a 24-bit depth buffer, on the other hand, requires that reads and writes of depth data be unaligned with respect to power-of-two boundaries. On the other hand, 8 bits of stencil data is more than sufficient for most applications, so it is only natural to pack the two buffers into a single buffer with both depth and stencil data. OpenGL never provides direct access to the buffers, so the OpenGL implementation can provide an interface to applications where it appears the one merged buffer is composed of two logical buffers.

One disadvantage of this scheme is that OpenGL lacks any means by which this packed data can be handled efficiently. For example, when an application reads from the 24-bit depth buffer, using the type GL_UNSIGNED_SHORT will lose 8 bits of data, while GL_UNSIGNED_INT has 8 too many. Both require expensive format conversion operations. A 24-bit format would be no more suitable, because it would also suffer from the unaligned memory accesses that made the standalone 24-bit depth buffer an unattractive proposition in the first place.

Many applications, such as parallel rendering applications, may also wish to draw to or read back from both the depth and stencil buffers at the same time. Currently this requires two separate operations, reducing performance. Since the buffers are interleaved, drawing to or reading from both should be no more expensive than using just one; in some cases, it may even be cheaper.

This extension provides a new data format, GL_DEPTH_STENCIL, that can be used with the glDrawPixels, glReadPixels, and glCopyPixels commands, as well as a packed data type, GL_UNSIGNED_INT_24_8, that is meant to be used with GL_DEPTH_STENCIL. No other data types are supported with GL_DEPTH_STENCIL. If ARB_depth_texture or SGIX_depth_texture is supported, GL_DEPTH_STENCIL/GL_UNSIGNED_INT_24_8 data can also be used for textures; this provides a more efficient way to supply data for a 24-bit depth texture.

GL_DEPTH_STENCIL data, when passed through the pixel path, undergoes both depth and stencil operations. The depth data is scaled and biased by the current GL_DEPTH_SCALE and GL_DEPTH_BIAS, while the stencil data is shifted and offset by the current GL_INDEX_SHIFT and GL_INDEX_OFFSET. The stencil data is also put through the stencil-to-stencil pixel map.

glDrawPixels of GL_DEPTH_STENCIL data operates similarly to that of GL_STENCIL_INDEX data, bypassing the OpenGL fragment pipeline entirely, unlike the treatment of GL_DEPTH_COMPONENT data. The stencil and depth masks are applied, as are the pixel ownership and scissor tests, but all other operations are skipped.

glReadPixels of GL_DEPTH_STENCIL data reads back a rectangle from both the depth and stencil buffers.

glCopyPixels of GL_DEPTH_STENCIL data copies a rectangle from both the depth and stencil buffers. Like glDrawPixels, it applies both the stencil and depth masks but skips the remainder of the OpenGL fragment pipeline.

glTex[Sub]Image[1,2,3]D of GL_DEPTH_STENCIL data loads depth and stencil data into a depth_stencil texture. glGetTexImage of GL_DEPTH_STENCIL data can be used to retrieve depth and stencil data from a depth/stencil texture.

In addition, a new base internal format, GL_DEPTH_STENCIL, can be used by both texture images and renderbuffer storage. When an image with a DEPTH_STENCIL internal format is attached to both the depth and stencil attachment points of a framebuffer object, then it becomes both the depth and stencil buffers of the framebuffer. This fits nicely with hardware that interleaves both depth and stencil data into a single buffer. When a texture with DEPTH_STENCIL data is bound for texturing, only the depth component is accessible through the texture fetcher. The stencil data can be written with TexImage or CopyTexImage, and can be read with GetTexImage. When a DEPTH_STENCIL image is attached to the stencil attachment of the bound framebuffer object, the stencil data can be accessed through any operation that reads from or writes to the framebuffer's stencil buffer.

The official definition of this extension is available here: http://www.opengl.org/registry/specs/ARB/framebuffer_object.txt