This Is What Started AMD's Open-Source Strategy
3D Support
Given that most state-of-the-art graphics chipsets no longer facilitate a separate fixed-function 3D pipeline block, all OpenGL commands have to be translated into vertex/fragment shaders. The Mesa Project already provides some basic functionality to support this in the OpenGL client library (libGL). The shaders required for transformation and lighting are well known.
At NOVELL we are convinced that we are able to implement OpenGL shader support at least to the extent as it is currently available in the free DRI driver for the R300 chipset generation. This will allow for some basic 3D programs like compositing managers, and it would allow to implement 2D and Xv acceleration support using the OpenGL shader language in Glucose. The complexity of adding loops and conditional branches, which are the main conceptual improvements compared to previous chipsets, is difficult to evaluate without additional information about the interface design of the GPU.
AMD has the option to replace this basic shader support with its own proprietary module to provide extra optimizations that may be required in certain enterprise level use cases. Hooks required by a such a proprietary module which provides shader optimization and arbitration could be added to the 3D driver. The basic 3D implementation would remain free software which would help to diffuse a lot of criticism from the community.
Implementing 3D support will touch four components while 2D support will benefit from some of these changes as well:
DRM kernel module
This kernel module watches over memory allocation and is required to provide verification of the command buffers before they are passed down to hardware to avoid malicious commands from being executed. Currently all existing DRM drivers provide this command validation in the kernel module. The R600 chipset generation includes a fully programmable command processor (CP). Taking advantage of this CP instruction validation the required functionality could be implemented in the GPU, freeing the CPU from this work and removing a lot of context switches. To do this however, it would be required to limit the programming of the CP to kernel space. As there is presently no infrastructure in place for the implementation of this functionality, this should probably be introduced as an optimization in a second step.
In any case, to implement such verification steps detailed knowledge of the CP and GPU stream command set is required.
If NOVELL implements separate 2D acceleration in the X.Org driver this will benefit from support for DMA command buffers, in contrast to the old IO/MMIO-port programming model.
One big unknown in estimating the complexity of this work is the influence of virtualized buffers that are used everywhere in the new design of the latest ATI chipset generation. An initial implementation might use fixed size buffers only.libdrm
This is the user-space part of DRM, mostly coupled to the kernel module. Thus library interfaces to system calls in the kernel need to be added here.DRI driver
The main part of the 3D driver resides in the Mesa OpenGL client side library. This library abstracts many details of OpenGL already. Drivers are free to provide acceleration for certain functionalities and short-circuits for special code paths.X.Org driver
The communication infrastructure doesn't have to be extended significantly, as almost no driver-dependent code is required there. The 2D driver needs to accommodate for the 3D driver using graphics memory, though, and as the agent in charge of overall control over the graphics hardware needs to setup and enable the 3D components by calling the appropriate hooks.
Development should take place following roughly the steps below (extension lists not complete but exemplary):
Implementation of memory management for z-, shader, and command buffers.
Implementation of command buffer functionality.
Implementation of engine setup and buffer management.
Basic shader support: ARB_fragment_program, ARB_vertex_program.
Primitive support, basic framebuffer blending operations, z- and scissor tests. Clipping.
Emulation of a fixed function pipeline with shaders as far as not already supported in Mesa.
implementation of basic texturing (RGBA8 only), including extended memory management.
Given the complexity of 3D graphics drivers \u2013 and lack of knowledge of the new ATI chipset generation \u2013 a time evaluation is difficult to make. NOVELL is confident, however, to have basic 3D support available (development step 7, shaders without optimization) within six to nine man months.
This comprises only very basic 3D support, additional functionalities will be added as necessary to accommodate for composited desktops like Compiz, first running Xgl (which is simpler due to less required functionality), than AIGLX:
Add EXT_texture_from_pixmap, ARB_texture_non_power_of_two, EXT_subtexture.
Provide AIGLX support.
Subsequently, functionality may be added in order of relevance to get full OpenGL 1.4 support implemented. Mesa's backward compatibility features may help to attain OpenGL 2.1 support automatically, though shaders will have to include loops and conditional branches.
The order given in the list below is just a rough estimate (this list is not exhaustive):
Enhanced shader support: full swizzling, implementation of less common instructions.
Implementation of less common blending operators, stencil buffers, alpha and stencil tests.
More enhanced shader support: loops and conditional jumps, complex instructions.
Enhanced shader support: virtualization, as far as this is not already implemented in 1. and 3.
Implementation of additional texture formats (for environment + normal mapping, float textures, luminance)
Support for antialiasing.
Enhanced shader support: geometry shaders.
Support for pBuffer + frame buffer objects (needs general support in DRI).
Floating point framebuffer (general support in Mesa/DRI).
Improving usage of the CP for command verification and engine setup.
Improving arbitration in the dispatch processor, as far as this is programmable / configurable.
Support for Tessellator (needs new extension).
These functionalities will be verified against simple test programs throughout development. Verification against complex OpenGL programs like games will be performed with the help of the community and AMD.