Welcome to the Etna_viv project.

Project Etnaviv is an attempt to make an open source user-space driver for the Vivante GCxxx series of embedded GPUs.

The current state of the project is experimental. It is currently only of use to developers interested in helping develop open source drivers for the hardware, reverse engineering, or in interfacing with GPU hardware directly. It is NOT usable as a driver for end users.

Once the understanding of the hardware is good enough, we'd likely want to fork Mesa/Gallium and create a GL driver. All help is appreciated.

ARM-based:

• Google TV (Marvell Armada 1500, contains a GC1000)
• OLPC (also Marvell Armada something with GC1000)
• CuBox, including pro variant (Marvell Armada 510, GC600)
• Many older tablets and such based on Rockchip 2918 SoC (GC800)
• Devices based on Freescale i.MX6 Series (GC2000 + GC320 + GC355)

MIPS-based:

• Devices based on Ingenic JZ4770 MIPS SoC (GC860), such as the GCW zero, and JZ4760 (GC200, 2D only).

See also wikipedia.

For the Vivante GPUs on many platforms the detailed features and specs are known, these can be found in doc/gpus_comparison.html.

The repository contains various tools and documentation related to figuring out how to program Vivante GCxxx GPU chips.

To exercise the initial-stage driver there are a few framebuffer tests in:

native/fb/

These demos do double-buffered animated rendering of 1000 frames to the framebuffer using the proof-of-concept etna rendering and command stream building API. The goal of this API is to provide a Gallium-like low-level interface to the Vivante hardware while abstracting away kernel interface details.

• companion_cube: Rotating "weighted companion cube", using array or indexed rendering. Exercised in this demo:

  • Array and indexed rendering of arbitrary mesh
  • Video memory allocation
  • Setting up render state
  • Depth buffer
  • Vertex / fragment shader
  • Texturing
  • Double-buffered rendering to framebuffer
  • MSAA (off / 2X / 4X)

• mip_cube_state: Rotating cube with a mipmapped texture loaded from a dds file provided on the command line. One of the example textures have a different color and number on each mipmap level, to explicitly show interpolation between mipmap levels as the surface goes nearer or farther from the camera.

  • Mipmapping
  • DXT1 / DXT3 / DXT5 / ETC1 compressed textures

• alpha_blend: Alpha blending quads

• cubemap_sphere: Cubemap textures

• stencil_test: Test stencil buffer handling

• particle_system: Simple particle system using vertex shader and point sprites

• displacement: Displacement mapping using vertex texture fetch

If you are executing these demos on an Android device, make sure that you are root, otherwise the framebuffer is not accessible.

Running these tests while Android is still writing to the framebuffer will result in stroboscopic effects. To get surfaceflinger out of the way type:

adb shell stop surfaceflinger
(run test)
adb shell start surfaceflinger

Map of documentation for known render state and registers. Mapped in rules-ng-ng (envytools) format:

rnndb/state.xml     Top-level database, global state
rnndb/state_hi.xml  Host interface registers
rnndb/state_2d.xml  2D engine state
rnndb/state_3d.xml  3D engine state

Other scattered bits of documentation about the hardware and ISA can be found in doc/hardware.md.

Shader (both vertex and fragment) instruction set description in rules-ng-ng format can be found here:

rnndb/isa.xml

Some written down notes, and examples of disassembled shaders can be found here:

doc/isa.md

Vivante has a unified, fixed-size, predictable instruction format with explicit inputs and outputs. This does simplify code generation, compared to a weird flow pipe system like the Mali 200/400.

A basic disassembler for the shader instructions (to a custom format) can be found in the tools directory:

tools/disasm.py <shader.bin>

This can be used to disassemble shaders extracted using dump_cmdstream.py --dump-shaders.

There is also an assembler, which accepts the same syntax that is produced by the disassembler:

tools/asm.py <shader.asm> [-o <shader.bin>]

Like other modern GPUs, the primary means of programming the chip is through a command stream interpreted by a DMA engine. This "Front End" takes care of distributing state changes through the individual modules of the GPU, kicking off primitive rendering, synchronization, and also supports basic flow control (branch, call, return).

Most of the relevant bits of this command stream have been deciphered.

The command stream format represented in rules-ng-ng XML format can be found here:

rnndb/cmdstream.xml

A significant part of reverse engineering was done by intercepting command streams while running GL simple demos. viv_hook is a library to intercept and log the traffic between libGAL (the Vivante user space blob) and the kernel driver / hardware.

This library uses ELF hooks to intercept only system calls such as ioctl and mmap coming from the driver, not from other parts of the application, unlike more crude hacks using LD_PRELOAD.

At the beginning of the program call the_hook, at the end of the program call end_hook to finalize and flush buffers. This should even work for native android applications that fork from the zygote.

The raw binary structures interchanged with the kernel are written to disk in a .fdr file, along with updates to video memory, to be parsed by the accompanying command stream dumper and other tools.

Other tools live in:

tools/

The most useful ones, aside from the assembler and disassembler mentioned before are:

• show_egl2_log.sh (uses dump_cmdstream.py, you may have to adapt this script to use another structure definition json depending on your kernel interface)

Decodes and dumps the intercepted command stream in human readable format, making use of rnndb state maps.

• fdr_dump_mem.py

Extract areas of video memory, images, and command buffers at certain points of execution.

The replay tests replay the command stream and ioctl commands of the EGL demos, to get the same output.

They can be found in:

native/replay/

Currently this is available for the cube example that renders a smoothed cube, and the cube_companion example that renders a textured cube.

A beginning has been made of a simple driver that builds the command stream from scratch and submits it to the kernel driver:

native/lib/viv.(c|h)
native/replay/etna.(c|h)
native/replay/etna_test.c (to experiment with shaders)
native/replay/cube_etna.c (renders the GLES2 smoothed cube)

The headers and implementation files for the Vivante GPL kernel drivers are also included:

kernel_drivers/

Four GPL kernel driver versions, gc600_driver_dove, v2 and v4 and imx6, are provided. They are useful in understanding the kernel interface, and the hardware at a basic level.

As open source drivers for the kernel are available, there are currently no plans to write a DRM/DRI kernel driver for Vivante. (There may be other reasons to do this anyway, such as allowing the driver to work without losing a fixed 128MB amount of memory to the GPU)

Envytools (https://github.com/pathscale/envytools) is a set of tools aimed at developers of the open source NVidia driver Nouveau, however some parts such as rnndb can be applied more generally. The repository contains a slightly modified subset of envytools for header generation from the state / cmdstream / isa rnndb files, so they can be used from the C code (etna), build with

cd envytools
mkdir build
cd build
cmake ..
make
cd ../..

Then generate the headers with

rnndb/gen_headers.sh

The build process is complicated by the existence of many different kernel drivers, with their subtly different interface (different headers, different offsets for fields, different management of context, and so on). These values for environment variable GCABI are supported out of the box:

• dove: Marvell Dove, newer drivers (0.8.0.3184)
• dove_old: Marvell Dove, older drivers (0.8.0.1998, 0.8.0.1123)
• arnova: Android, Arnova 10B G3 tablet (RK2918)
• v2: Various Android, for older chips (RK2918 etc)
• imx6: Various Android, for newer chips (i.MX6 specific)
• v4: Various Android, for newer chips (i.MX6 etc)

If possible get the gc_*.h headers for your specific kernel version. If that's not possible, try to find which of the above sets of headers is most similar, and adapt that.

gc_abi.h is an extra header that defines the following flags describing the kernel interface to etna, for a certain setting of the environment variable GCABI:

• GCABI_CONTEXT_HAS_PHYSICAL: struct _gcoCONTEXT has physical and bytes fields
• GCABI_HAS_MINOR_FEATURES_2: struct _gcsHAL_QUERY_CHIP_IDENTITY has chipMinorFeatures2 field
• GCABI_HAS_MINOR_FEATURES_3: struct _gcsHAL_QUERY_CHIP_IDENTITY has chipMinorFeatures3 field
• GCABI_USER_SIGNAL_HAS_TYPE: struct _gcsHAL_USER_SIGNAL has signalType field
• GCABI_HAS_CONTEXT: struct _gcsHAL_COMMIT has contextBuffer field
• GCABI_HAS_STATE_DELTAS: struct _gcsHAL_COMMIT has delta field

It would be very useful to have an auto-detection of the Vivante kernel version, to prevent crashes and such from wrong interfaces. However, I don't currently know any way to do this. The kernel does check the size of the passed ioctl structure, however this guarantees nothing about the field offsets. There is /proc/driver/gc that in some cases contains a version number. In very new drivers there is an ioctl call gcvHAL_VERSION that returns the major, minor and build version.

To build for an Android device, install the Android NDK and define the cross-build environment by setting environment variables, for example like this:

export NDK="/opt/ndk"
export TOOLCHAIN="/opt/ndk/toolchains/arm-linux-androideabi-4.6/prebuilt/linux-x86"
export SYSROOT="/opt/ndk/platforms/android-14/arch-arm"
export PATH="$PATH:$TOOLCHAIN/bin"

export GCCPREFIX="arm-linux-androideabi-"
export CXXABI="armeabi-v7a"
export PLATFORM_CFLAGS="--sysroot=${SYSROOT} -DANDROID -DPIPE_ARCH_LITTLE_ENDIAN"
export PLATFORM_CXXFLAGS="--sysroot=${SYSROOT} -DANDROID -DPIPE_ARCH_LITTLE_ENDIAN -I${NDK}/sources/cxx-stl/gnu-libstdc++/4.6/include -I${NDK}/sources/cxx-stl/gnu-libstdc++/4.6/libs/${CXXABI}/include"
export PLATFORM_LDFLAGS="--sysroot=${SYSROOT} -L${NDK}/sources/cxx-stl/gnu-libstdc++/4.6/libs/${CXXABI} -lgnustl_static"
# Set GC kernel ABI (important!)
#export GCABI="v2"
#export GCABI="v4"
export GCABI="arnova"

To build the egl samples (for command stream interception), you need to copy libEGL_VIVANTE.so libGLESv2_VIVANTE.so from the device /system/lib/egl to native/lib/egl. This is not needed if you just want to build the replay, etna or fb tests, which do not rely in any way on the userspace blob.

For non-Android Linux ARM cross compile, create a script like this (example for CuBox) to set up the build environment. Don't forget to also copy the EGL/GLES2/KDR headers from some place and put them in a directory include under the location where the script is installed, and get the libEGL.so and libGLESv2.so from the device into lib:

DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
export GCCPREFIX="arm-linux-gnueabi-"
export PLATFORM_CFLAGS="-I${DIR}/include -D_POSIX_C_SOURCE=200809 -D_GNU_SOURCE"
export PLATFORM_CXXFLAGS="-I${DIR}/include -D_POSIX_C_SOURCE=200809 -D_GNU_SOURCE"
export PLATFORM_LDFLAGS="-ldl -L${DIR}/lib"
export PLATFORM_GL_LIBS="-lEGL -lGLESv2 -L${TOP}/lib/egl -Xlinker --allow-shlib-undefined"
# Set GC kernel ABI to dove (important!)
#export GCABI="dove"      # 0.8.0.3184
export GCABI="dove_old"  # 0.8.0.1998, 0.8.0.1123

If you haven't got the arm-linux-gnueabi- bintools, on an Debian/Ubuntu host they can be installed with

apt-get install gcc-arm-linux-gnueabi g++-arm-linux-gnueabi

On hardfloat targets you should use gcc-arm-linux-gnueabihf- instead.

Run make in native, or alternatively in native/gallium, native/fb, native/replay and/or native/egl separately.

Etna_pipe is currently compatible with the following GC chips at least:

• GC600
• GC800

GC2000 support is underway.

The command stream on different device GCxxx variants will also likely be slightly different; the features bit system allows for a ton of slightly different chips. When porting it, look for:

• Number of bits per tile (2 on my hw), depends on 2BIT_PER_TILE feature flag

• depth buffer (hierarchical or normal)

• location of shader memory in state space (will be at 0x0C000/0x08000 or 0x20000 for recent models with more than 256 shader instructions support)

There is currently no mailing list for this project, and looking at other GPU reverse engineering projects the mailing lists usually see very little traffic, so I won't bother.

We usually hang out in #etnaviv and #lima on irc.freenode.net.

• Wladimir J. van der Laan
• Steven J. Hill

• Luc Verhaegen (libv) of Lima project (basic framework, general idea)
• Nouveau developers (rnndb, envytools)