Free drivers for ARM graphics

At the 2013 Linux Plumbers Conference in New Orleans, Rob Clark presented an update of the progress toward free software graphic drivers on ARM systems. Free drivers for desktop display hardware are more actively developed, with many users simply resigning themselves to the use of binary-only drivers in the ARM system-on-chip (SoC) world. But as Clark explained, the effort to provide free drivers has gained significant momentum in the past year alone.

A few years ago, Clark said, the situation looked very bleak, and there were no projects attempting to write free drivers for ARM SoCs. Then Luc Verhaegen started his Lima project, an attempt to reverse engineer the Mali GPU. That project "motivated the rest of us," Clark said, and now there are four active projects, each targeting a different GPU.

The first that Clark described is the Etnaviv project for GPUs made by Vivante. The Vivante GPUs are on the low-end, usually found in mobile phones and other small devices like the nettop-class CuBox PC. These GPUs have a fairly straightforward design that is similar to many desktop GPUs. They use a unified shader architecture, meaning that there are not separate shader units designated for separate purposes (e.g., vertex shaders, pixel shaders, and geometry shaders). Most Vivante models have supported OpenGL ES 2.0, while the latest revision has moved to OpenGL ES 3.0 and OpenCL. The instruction set is all vector-based, and historically offered only floating-point data—although here, too, there is a recent change: the newest models now offer integer support, too.

Etnaviv is making very rapid progress, Clark said: work only started in late 2012. As of now, the project has produced a working Gallium3D driver, which is capable of playing some games. But it only supports the Linux framebuffer (fbdev) backend. The X backend needs a lot of help, he said: Direct Rendering Manager (DRM), Direct Rendering Infrastructure (DRI), and 2D X (DDX) support are all missing.

For NVIDIA's Tegra SoC series, there is the Grate project. Tegra is a widely-adopted SoC that can be found in a number of well-known products, such as Samsung Galaxy tablets, the (first-generation) Google Nexus 7, and Trim-Slice nettops. The GPU architecture is more sophisticated than Vivante's, with separate vertex and fragment shaders. The instruction set is "minimalist," Clark said, not even providing loops, but the GPUs tend to offer good performance by incorporating a massive number of cores. The devices support OpenGL ES 2.0.

Grate is still in the very early stages, he said, and is not yet usable. Grate can capture and replay a command stream, the basic GL state is well understood, and the project has reverse engineered the vertex shader instruction set architecture. But the vertex shader is straightforward; the challenge is the fragment shader, which "is more weird," he said. It uses three separate instruction streams: one for texture lookup, one for varying variable interpolation and the "Special Function Unit" or SFU (which implements out-of-the-ordinary functions like reciprocal square roots), and one for the standard arithmetic logic unit (ALU). But even the ALU is unusual; it accepts packets of three to four instructions and offers just four opcodes.

Clark next discussed the Lima project, which was the first such ARM GPU driver project to get started. Most of the Lima effort is directed at the Mali 200 and 400 series, although work has recently started on the Mali 600 series. The 200 and 400 are similar: both support OpenGL ES 2.0 and offer separate vertex and fragment shaders, and both support 2D and cubemap textures. The 600 series diverges considerably: it supports OpenGL ES 3.0 and OpenCL 1.1, uses a unified shader architecture (with different models using varying numbers of cores and ALUs of varying register widths), and supports 3D textures in addition to 2D and cubemap textures. Mali 200/400 chips are found in high-end phones, Allwinner devices, and some Samsung tablets, while the Mali 600 series is found in Google's Chromebooks and Nexus 10 tablet.

The 200/400 series vertex shader is a bit unusual, he said. It is single-threaded but deeply pipelined. It uses very long instruction words (VLIW), each of which can include two additions, two multiplications, one complex ALU operation, one pass-through ALU operation, one attribute load, one register load, one uniform load, and one varying store. In addition, there are no explicit output registers: the outputs from previous instructions are routed into the current instruction, a job which the shader compiler is responsible for sequencing correctly. The fragment shader is not quite as strange, he said, although it uses variable-length VLIW rather than the fixed-length VLIW design of the vertex shader.

As of now, the Lima driver is starting to reach usable status on Mali 200 and 400 series GPUs. There is a Mesa-based DRI driver that runs es2gears (the OpenGL ES version of the famous "glxgears" demo) as well as some other 3D demos. Connor Abbott has been developing a shader compiler, although it has not been hooked up to Lima yet. For the Mali 600 series, work has only recently begun, and so there is little progress to report.

Clark himself is the developer behind Freedreno, the driver project for Adreno GPUs. Adreno GPUs are found in Qualcomm Snapdragon SoCs, HP TouchPads, and several high-end Android phones. The GPU is available in two generations. The 200 series chips support OpenGL ES 2.0, while the 300 series chips support OpenGL ES 3.0 and Open CL 1.1. Both use a unified shader architecture, although there are differences between them. The 200 series uses VLIW instructions on vectors with the ability to co-dispatch scalars, while the 300 series uses explicitly pipelined scalar instructions.

As of now, the Freedreno driver is the furthest along of the free driver projects. There is an initial DRM/KMS driver in for kernel 3.12, and there is a working Gallium3D driver that supports both 200 and 300 series GPUs. The Gallium3D driver also works on the Kernel Graphics Support Layer (KGSL)/fbdev backend for Android. There is an X driver, too, which can utilize Adreno Z180 2D vector graphics cores. Freedreno currently implements support for OpenGL ES 1.0 and 2.0 and offers OpenGL 1.4 support on what Clark called "a best effort basis." That best effort is good enough to run GNOME Shell, XBMC, and several well-known 3D games like Xonotic and OpenArena. The DRM/KMS backend also supports the Weston compositor for Wayland.

Binary graphics drivers are a thorny issue: many free software supporters decry them, but they are a common sight—particularly on mobile Linux devices. Considering the rapid progress that has been made on free drivers for desktop GPUs in recent years, it can be easy to forget just how recently binary drivers were considered a necessary evil for that hardware as well. As Clark's talk illustrated, free drivers for ARM SoC systems still have a ways to go, but they are also making rapid progress, which should give hope to those who feel stuck with proprietary GPUs in their pockets.