OGRE
2.2.4
Object-Oriented Graphics Rendering Engine
|
The Command Buffer was the response to a very practical problem: Ogre couldn't assume persistent mapping is supported as this would break compatibility with DX10-level hardware and DX11 API.
This means that we need to map a buffer to update its contents. But we can't issue a draw call while a buffer is mapped. Unmapping and mapping before and after each draw call would be extremely costly; and therefore a command buffer was needed that would record all the API calls that have to be made; in order for buffers (i.e. constant and texture buffers to send data to the shaders) to be mapped only once.
Once we're done updating all buffers, the command buffer is executed, playing back all the API calls in the right order.
Note that the command buffer is written for high performance, low footprint; thus its inner workings and interface are very C-like.
A C++ template is used to simplify the process; though you are not forced to use it.
The function "addCommand" creates an uninitialized command, and it's your job to initialize it to valid values.
The following snippet will record the command to set a new Vao, which will eventually calling RenderSystem::_setVertexArrayObject
when it gets executed:
The following snippet will issue a v1-style indexed draw call:
All the CommandBuffer-related code is under the CommandBuffer folder. By convention, all commands start with the Cb
prefix.
The returned pointer should not be kept for long as further calls to addCommand
may invalidate the pointer.
Each command is, at the time of writting, of a fixed size of 32 bytes. This value is controlled by CommandBuffer::COMMAND_FIXED_SIZE
. The first 2 bytes are reserved for the command type; which leaves you with 30 bytes of actual space.
Beware of storing pointers in the command: In a 32-bit environment each pointer is 4 bytes. But in a 64-bit environment, each pointer is 8 bytes. Make sure they still fit within the size of the command.
If 32 bytes is not enough, you can try researching into implementing multi-commands (an action that spans multiple commands and would read back and forth through the command buffer to obtain all the necessary data, beware of respecting the header in each command!); or raise the COMMAND_FIXED_SIZE
constant.
Keeping the constant size low is vital to achieve good bandwidth and cache utilization, and hence keep the overhead to a minimum.
The command buffer is implemented via a manual virtual table. The variable CommandBuffer::CommandBufferExecuteFunc* CbExecutionTable[MAX_COMMAND_BUFFER+1]is populated with all the entries to this vtable. See enum
CbType` for all the possible entries.
Each function accepts a pointer to this
state and the command itself.
Commands are stored in structs and its constructors are used to ensure each command populates the command type value properly.
Example:
To implement a command that will call renderSystem->_setHlmsMacroblock( myMacroblock )
; a structure is defined:
Block
is the myMacroblock
we want to set. We derive from CbBase
which conveniently has already the 16-bit value defined. Since CbBase
is only 2 bytes, take in mind compilers may add padding between the end of CbBase and the beginning of CbMacroblock unless compiler packing extensions[^16] are used; which leaves us with an usable space of 28 bytes in 32-bit environments, or 24 bytes in 64-bit. Note that you're not required to derive from CbBase
(in case you're pressed with the command's size). This is just for convenience and consistency.
The constructor will populate all the proper values:
The constructor sets the command type to CB_SET_MACROBLOCK
which is already part of CbType
.
Therefore running *commandBuffer->addCommand<CbMacroblock>() = CbMacroblock( myMacroblock )
; will just work.
But there's one more thing: We need to implement the actual command!
For that, we declare and define the following function:
Last but not least, we have to add the function to the table:
The comment's remark is very important: You can't add execute_setMacroblock
anywhere, you must store it at the same index as the value of CB_SET_MACROBLOCK
, so that CbExecutionTable[CB_SET_MACROBLOCK]
points to to execute_setMacroblock
.
Maintaining a manual vtable may seem like a lot of trouble, but the amount of commands is manageable (don't feel tempted to add one command for every little thing!) and it has very powerful benefits: You can manipulate the command type with simple integer arithmetic, thus saving expensive branch evaluations per object.
For example, Ogre supports indirect buffers when available. But when indirect buffers aren't supported, not only we have to store the draw command data in CPU memory rather than GPU memory, but we also have to use very different draw calls. This is shown in RenderSystem::_render
vs RenderSystem::_renderEmulated
.
The former uses indirect buffer draw calls, the latter doesn't. Both do the exact same thing.
Usually, during command generation, the RenderQueue could look like this:
This involves a branch per renderable. Since there can be literally hundred of thousands of them, this is costly.
Instead, you can just store the indirectBuffer version and the emulated together in the vtable, and use integer arithmetic to modify the index to the table:
Then during the command generation in the RenderQueue just write:
It's not just cleaner, it's also more efficient. A cmp, a jmp and a mov instruction (plus a potential miss-prediction) was just transformed into a single add.
Another advantage of a command buffer, is that you may have more information about the rendering sequence only after you have processed all objects. You could also try to detect redundant state changes (Ogre doesn't do this though).
For example, both texture buffers require binding them to a shader by specifying an offset and a size.
The offset is known at the time the texture buffer is needed, but the size is not. The same texture buffer may be used to store data of many subsequent Renderables, and thus required size may not be known until we're done processing all of them.
For example, the Hlms implementations shipped with Ogre populate this size at the end. For that, it saves the command location with CommandBuffer::getCommandOffset
called after adding the command:
The size at this point is set to 0, which is wrong. This will be corrected once we're done:
The function getCommandFromOffset
will only return a null pointer if the value from texBufferOffset
is invalid.
We need to use these functions, as saving the command's pointer is not safe as it is not guaranteed to be valid after the next addCommand
call.
[^14]: For example DX11 doesn't allow this level of flexibility. However we can still batch all meshes into the same vertex buffer, and in the same index buffer; texture buffers are also very flexible.
[^15]: On DX11, we can emulate fences using dummy occlusion queries.
[^16]: Like #pragma pack