Buffer Objects is the general term for unformatted linear memory allocated by the OpenGL context. These can be used to store vertex data, pixel data retrieved from images or the framebuffer, and a variety of other things.
Buffer objects were initially introduced in the ARB_vertex_buffer_object extension. They have been part of core OpenGL since version 1.5. This discussion will pertain specifically to the core version of buffer objects, though it is essentially identical to the extension version.
While buffer objects themselves are relatively old, some of the uses of them are much more recent.
Buffer Objects are OpenGL Objects; they therefore follow all of the rules of regular OpenGL objects. To create a buffer object, you call
glGenBuffers. Deleting them uses
glDeleteBuffers. These use the standard Gen/Delete paradigm as most OpenGL objects.
As with the standard OpenGL object paradighm, this only creates the object's name, the reference to the object. To actually create the object itself, you must bind it to the context. You do this using the following API:
void glBindBuffer(enum target, uint bufferName)
target defines how you intend to use this binding of the buffer object. When you're just creating and/or filling the buffer object with data, the target you use doesn't technically matter. It matters more when you intend to tell OpenGL to use the data in the buffer in some way.
- Performance Note: While it technically does not matter how you bind the buffer for creations vs. how you use it (create it with one target, unbind and bind it for use as another target) this is not recommended if you can at all help it. OpenGL implementations can and will make assumptions based on what target you used to create a buffer object. Try to be consistent where reasonable; don't use a buffer that normally stores vertex data to store pixel data or uniforms.
- Implementations will handle some cross-target usages, but only those that make sense semantically (vertex buffers being used for transform feedback makes sense. Vertex buffers being used for pixel transfer does not).
To create the actual memory storage for a buffer object, you use this API:
void glBufferData(enum target, sizeiptr size, const void *data, enum usage)
target parameter is just like the one for
glBindBuffer; it says which bound buffer to modify.
size represents how many bytes you want to allocate in this buffer object.
data parameter is a pointer to user memory that will be copied into the buffer object's data store. If this value is NULL, then no copying will occur, and the buffer object's data will be undefined.
usage parameter can be very confusing.
Buffer Object Usage
Buffer objects are general purpose memory storage blocks allocated by OpenGL. They are intended to be used in a great many ways. To give the implementation great flexibility in exactly what a particular buffer object's data store will be, so as to better optimize performance, the user is required to give usage hints. These provide a general description as to how exactly the user will be using the buffer object.
There are two independent parts to the usage pattern: how the user will be reading/writing from/to the buffer, and how often the user will be changing it relative to the use of the data.
There are two ways for data to cause a change in the data contents of a buffer object. One way is for the user to explicitly upload some binary data. The other way is for the user to issue GL commands that cause the buffer to be filled in. For example, if you want to have a buffer store the results of a vertex shader computation through the use of transform feedback, the user is not directly changing the buffer information. So this is the latter kind of change.
Similarly, the user can read a buffer's data, using a variety of commands. Or, the user can execute an OpenGL command that causes the GL to read the contents of the buffer and do something based on it. Buffers storing vertex data are read by the GL when rendering.
There are three hints that the user can specify the data. They are all based on what the user will be doing with the buffer. That is, whether the user will be directly reading or writing the buffer's data.
- DRAW: The user will be writing data to the buffer, but the user will not read it.
- READ: The user will not be writing data, but the user will be reading it back.
- COPY: The user will be neither writing nor reading the data.
DRAW is useful for, as the name suggests, drawing. The user is uploading data, but only the GL is reading it. Buffer objects holding vertex data are generally specified as DRAW, though there can be exceptions.
READ is used when a buffer object is an intermediate for asynchronous delivery of image data. You can use
glGetTexSubImage to read data into a buffer object.
COPY is used when a buffer object is used to pass data from one place in OpenGL to another. For example, you can read image data into a buffer, then use that image data as vertex data in a draw call. You can also use transform feedback to achieve the same thing in a more direct way. You have the feedback data go to a buffer object, then use that buffer object as vertex data.
There are three hints for how frequently the user will be changing the buffer's data.
- STATIC: The user will set the data once.
- DYNAMIC: The user will set the data occasionally.
- STREAM: The user will be changing the data after every use. Or almost every use.
STREAM is pretty easy to understand: the buffer object's contents will be updated after almost every use. STATIC is pretty easy to understand too. The buffer object's contents will be updated once and never changed.
What is unclear is when DYNAMIC becomes STREAM or STATIC. These are only hints, after all. It is perfectly legal OpenGL code to modify a STATIC buffer after it has been created, or to never modify a STREAM buffer.
Is it better to use STATIC for buffers that are updated very infrequently? Is it better to use DYNAMIC for buffers that get updated frequently, but not at STREAM speed? Is it better to use DYNAMIC for buffers that get partially updated? These are questions that can only be answered with careful profiling. And even then, the answer will only be accurate for that particular driver version from that particular hardware vendor.
In any case, STREAM, STATIC, and DYNAMIC can be matched with READ, DRAW, and COPY in any combination. STREAM_COPY means that you will be doing transform feedback writes (or other kinds of GL-based writes) into the buffer after almost every use; it will not be updated with
BufferSubData or similar functions. STATIC_READ means that you will fill the buffer up from the GL, but you will only do this once.
We have seen that
BufferData can be used to update the data in a buffer object. However, this also recreates the buffer object, causing it to be reallocated. This is not usually what one wants, as recreating the buffer can often be a heavyweight operation.
Instead, one can use the following API:
void glBufferSubData(enum target, intptr offset, sizeiptr size, const void *data)
offset parameter is an integer offset into the buffer object where we should begin updating. The
size parameter is the number of bytes we should copy out of
data. For obvious reasons,
data cannot be NULL.
glBufferSubData is a nice way to present data to a buffer object. But it can be wasteful in performance, depending on your use patterns.
For example, if you have an algorithm that generates data that you want to store in the buffer object, you must first allocate some temporary memory to store that data in. Then you can use
glBufferSubData to transfer it to OpenGL's memory. Similarly, if you want to read data back,
glGetBufferSubData is perhaps not what you need, though this is less likely. It would be really nice if you could just get a pointer to the buffer object's storage and write directly to it.
You can. To do this, you must map the buffer. This gives you a pointer to memory that you can write to or read from, theoretically, just like any other. When you unmap the buffer, this invalidates the pointer (don't use it again), and the buffer object will be updated with the changes you made to it.
While a buffer is mapped, you cannot do anything to it. Do not call any OpenGL function that would cause OpenGL to write to or read from that buffer. Doing so can have unfortunate consequences. Most important of all, do not unbind the buffer while it is mapped. It's OK to do these things before you map the buffer, but not while it is mapped.
To map a buffer, you call
glMapBufferRange. The signature of this function is this:
void *glMapBufferRange( GLenum target, GLintptr offset, GLsizeiptr length, GLbitfield access );
The return value is the pointer to the buffer object's data. The parameters
length allow you to specify a particular range within the buffer to map; you do not have to map the entire buffer. The
target parameter refers to the particular target that you have the buffer you want to map bound to.
access parameter is a bit complicated.
You can tell OpenGL what you intend to do with the pointer. If you're just adding new data to the buffer, then the pointer that gets returned isn't something you need to read from. Similarly, it could be your intention to read from this pointer and not change the buffer; OpenGL could do a simple copy from the buffer's memory to scratch memory in the client address space to make reading the buffer faster (buffer memory may not be optimized for reading).
You specify this by setting bitflags in the
access parameter. This parameter must have either the GL_MAP_READ_BIT or GL_MAP_WRITE_BIT set; it can have both (the memory should be fit for reading and writing) but it can't have neither. There are a number of other flags that can be set; the description of these is best left to a dedicated article.
- Legacy Note:
glMapBufferRangeis fairly new. It is a core extensions and is core in GL 3.0 and above. Implementations that don't support this function will support its predecessor,
glMapBuffer. The range version above is a strict superset of the original glMapBuffer function. The old function cannot map a particular range of bytes; it can only map all of the buffer. And the old function does not support some of the advanced access features of glMapBufferRange. If your implementation supports glMapBufferRange, you should use it; only use glMapBuffer as a fallback.
Unmapping the buffer is done when you are finished with the pointer and want to let OpenGL know that the buffer is free to be used. This is done with the function
glUnmapBuffer. This function takes only the target that the buffer in question is bound to. After calling this function, you should not use the pointer returned in the map call again.
There is one major gotcha when mapping a buffer.
During normal OpenGL operations, the OpenGL specification requires that all data stored in OpenGL objects be preserved. Thus, if something in the operating system or other such things causes video memory to be trashed, the OpenGL implementation must ensure that this data is restored properly.
Mapping is not a normal operation. Because of its low-level nature, these protections have to be relaxed. Therefore, it is possible that, during the time a buffer is mapped, some kind of corruption happens. If this occurs, calling
glUnmapBuffer will return GL_FALSE. At that point, the contents of the buffer in question are considered undefined. It may have your data, or it may have random garbage.
How often does this happen? On Microsoft Windows 5.1 (XP) and below, video memory could get trashed anytime an application didn't have input focus. This is why alt-tabbing away from games takes a long time to recover from; the application/OpenGL has to reload all of this data back to video memory. Fortunately, on Windows 6.0 (Vista) and above, this is fixed; Windows itself manages video memory and will ensure that all video memory is retained. Thus, at least theoretically, this should never be a problem on Vista or above machines.
One thing to remember about buffer mapping is this: the implementation is not obligated in any way to give you an actual pointer to the buffer object's memory. It is perfectly capable of giving you a pointer to some memory that OpenGL allocated just for the purpose of mapping, then it will do the copy on its own time.
The other thing to remember is that you should not care. OpenGL likes to give implementations flexibility to make performance optimizations. If mapping a certain buffer gives you some other pointer, and the implementation will do the copy on its own time, you should assume that this is probably the fastest way to work. It's still potentially faster than glBufferSubData, since the copy in glBufferSubData must happen before the function call returns, whereas the copy in the mapping case can happen in a thread the GL spawns. The worst case is that it's no slower than glBufferSubData.
However, you should not use the pointer you are given like any other pointer you might have. If this pointer is a pointer to non-standard memory (uncached or video memory), then writing to it haphazardly can be problematic. If you are attempting to stream data to the buffer, you should always map the buffer only for writing and you should write sequentially. Even if you
The purpose in writing sequentially is to be able to use write-combine memory, a feature of some processors (most x86's). It allows sequential writes to uncached memory to flow fairly quickly, compared to random writes to uncached memory. Always write sequentially, even if you have to write padding bytes manually.
Buffer objects provide a number of possible usage patterns for streaming. Exactly which will work best depends on the particulars of the hardware.
Tf you're streaming data, STREAM is going to need to be in your usage. And since we're talking about updating from the user side, you should be using STREAM_DRAW.
There is a parallelism problem that can occur when streaming data. The OpenGL specification permits an implementation to delay the execution of drawing commands. This allows you to draw a lot of stuff, and then let OpenGL handle things on its own time. Because of this, it is entirely possible that, well after you called the rendering function with a buffer object, you might start trying to stream vertex data into that buffer. If this happens, the OpenGL specification requires that the thread halt until all drawing commands that could be affected by your update of the buffer object complete. This obviously misses the whole point of streaming.
This is going to be your main source of woe.
There is one tried-and-true method of avoiding this: manual double-buffering. That is, allocate two buffer objects of the same size. Fill one up and render with it, then switch to the other one when you need to stream some new vertices in.
This is nice, and it gets around the above issue. But it has problems. Namely, that it takes up 2x the memory. Also, the STREAM hint is designed to deal with precisely this issue, so it is entirely possible that the implementation may double-buffer for you.
Instead, you can try a variety of techniques to force the implementation to do what you need.
glMapBufferRange with the GL_MAP_INVALIDATE_BUFFER_BIT set is one way to do it. Invalidating the buffer tells OpenGL that the entire buffer's contents will not be needed. This gives OpenGL the opertunity to orphan the buffer and allocate a new one. It also conveniently maps the buffer, so if you need to map the buffer to upload your data, there you are.
If you call
glBufferData with a NULL data pointer and the same usage hints and size, the OpenGL implementation can take this as a sign that you no longer care about the current contents of the buffer. Again, this allows OpenGL to orphan the buffer and allocate a new one.
Both of these can give the effect of double-buffering.
The deepest of the deep magic comes in
glMapBufferRange with GL_MAP_UNSYNCHRONIZED_BIT. This guarantees that you will never halt due to the buffer being in use. Unfortunately, it also means that you can get a race condition, where you are updating a buffer object while it is being read from. The unsyncrhonized flag will prevent OpenGL from trying to stop this, but it won't prevent OpenGL from rendering wrong stuff when it does happen.
To prevent it on your end, you can use the ARB_sync core extension (core in version 3.2). This allows you to ask whether a particular rendering command has finished by putting a fence after that command. Thus, if it has finished, you can do the streaming. If it hasn't, you can choose to do something else. That way, your thread isn't stopped.
Most of the uses of buffer objects involve binding them to a certain target, then calling a function that behaves differently based on having a buffer object in that target. Usually, these functions take a pointer as one of their parameters. When a buffer object is bound to certain targets, it causes some functions that take a pointer parameter to treat that parameter as an offset into the buffer object. Thus, rather than pulling data from client memory pointers, they pull it from the bound buffer object.
Indeed, in more recent versions of OpenGL, some of these functions are no longer allowed to take a regular pointer at all. They can only be used when a buffer object is bound to the proper target.
Easily, the most common usage for buffer objects is as vertex array data. This was the original use behind them, though other uses were expected and designed into the feature. OpenGL 3.1 and above no longer allow the use of vertex arrays without buffer objects.
When a buffer object is bound to the GL_ARRAY_BINDING target, all
glVertexAttribPointer commands, as well as the equivalent fixed function commands (
glTexCoordPointer, etc) will use the pointer value as an absolute offset from the beginning of the buffer object.
Similarly, when a buffer object is bound to GL_ELEMENT_ARRAY_BINDING, all functions of the form
glDraw*Elements will use the pointer value as an absolute offset from the beginning of the bound buffer object. The indices will be taken from the buffer object.
In GL 3.1 and above (without compatibility), none of these functions will work in the absence of a buffer object. You simply are not allowed to use client memory for drawing in these versions of OpenGL.
The GL_COPY_READ_BUFFER and GL_COPY_WRITE_BUFFER (core extension GL_ARB_copy_buffer, core in version 3.1 and above) are used for copying buffer objects. The function
glCopyBufferSubData is used to copy data between buffers.
The GL_PIXEL_PACK_BUFFER and GL_PIXEL_UNPACK_BUFFER (extension GL_ARB_pixel_buffer_object, core in versions 2.1 and above) are used to store or read pixel data. If a buffer is bound to GL_PIXEL_PACK_BUFFER,
glDrawPixels are all affected. These functions will read their data from the bound buffer object instead of a client pointer. Similarly, if a buffer is bound to GL_PIXEL_UNPACK_BUFFER,
The pack and unpack bindings are best used for asynchronous transfers of data to/from the GPU. When these functions read/write from/to client memory, they must block for a period of time. By having their source/destinations be buffer objects, which have a very controlled access scheme, it allows these commands to work asynchronously.
The GL_TEXTURE_BUFFER (core extension GL_ARB_texture_buffer_object, core in versions 3.0 and above) binding allows a special one-dimensional texture to have its storage be a buffer object. This binding point is how the buffer gets attached to the texture object.
The GL_TRANSFORM_FEEDBACK_BUFFER (core in versions 3.0 and above) binding is used to implement storing the results of the vertex (and geometry, where available) shader in a buffer object. This binding point is the buffer that gets written into.
The GL_UNIFORM_BUFFER_BINDING (core extension GL_ARB_uniform_buffer_object, core in versions 3.1 and above) binding is used to allow buffer objects to store uniforms for shader programs. This binding point is used to attach buffers containing uniforms to program objects.