The Android graphics path, in depth

The Android graphics path
in depth
The Android graphics path 1 Copyright ©2011-2014, 2net Limited

License
These slides are available under a Creative Commons Attribution-ShareAlike 3.0
license. You can read the full text of the license here
http://creativecommons.org/licenses/by-sa/3.0/legalcode
You are free to
• copy, distribute, display, and perform the work
• make derivative works
• make commercial use of the work
Under the following conditions
• Attribution: you must give the original author credit
• Share Alike: if you alter, transform, or build upon this work, you may distribute
the resulting work only under a license identical to this one (i.e. include this
page exactly as it is)
• For any reuse or distribution, you must make clear to others the license terms of
this work
The orginals are at http://2net.co.uk/slides/android-graphics-abs-2014.pdf

About Chris Simmonds
• Consultant and trainer
• Working with embedded Linux since 1999
• Android since 2009
• Speaker at many conferences and
workshops
"Looking after the Inner Penguin" blog at http://2net.co.uk/
https://uk.linkedin.com/in/chrisdsimmonds/
https://google.com/+chrissimmonds

Overview
• The Android graphics stack changed a lot in Jelly
Bean as a result of project Butter
• This presentation describes the current (JB) graphics
stack from top to bottom
• Main topics covered
• The application layer
• SurfaceFlinger, interfaces and buffer queues
• The hardware modules HWComposer and Gralloc
• OpenGL ES and EGL

The big picture
SurfaceFlinger
Activity
OpenGL vendor
libraries
OpenGL ES
GPU driver
HWComposer
ION
Gralloc alloc
FB driver
Gralloc FB
Android
framework
Vendor HAL
library
Kernel
driver
Activity
Activity
Manager
Window
Manager
BuﬀerQueue

Inception of a pixel
• Everything begins when an activity draws to a surface
• 2D applications can use
• drawing functions in Canvas to write to a Bitmap:
android.graphics.Canvas.drawRect(), drawText(), etc
• descendants of the View class to draw objects such
as buttons and lists
• a custom View class to implement your own
appearance and behaviour
• In all cases the drawing is rendered to a Surface
which contains a GraphicBuffer

2D rendering path
Activity
Canvas
Surface
SkiaOpenGL ES
Vendor Open GL
GPU driver
Activity
HWUI

Skia and hwui
• For 2D drawing there are two rendering paths
• hwui: (libwhui.so) hardware accelerated using
OpenGL ES 2.0
• skia: (libskia.so) software render engine
• hwui is the default
• Hardware rendering can be disabled per view,
window, activity, application or for the whole device
• Maybe for comparability reasons: hwui produces
results different to skia in some (rare) cases

3D rendering path
• An activity can instead create a GLSurfaceView and
use OpenGL ES bindings for Java (the
android.opengl.* classes)
• Using either the vendor GPU driver (which must
support OpenGL ES 2.0 and optinally 3.0)
• Or as a fall-back, using PixelFlinger, a software GPU
that implements OpenGL ES 1.0 only
• Once again, the drawing is rendered to a Surface

3D rendering path
Activity
Surface
OpenGL ES
Vendor Open GL
GPU driver
Activity
Android GL
PixelFlinger

Composition
SurfaceFlinger
Wallpaper
Launcher
Navigation
bar
Status
bar

SurfaceFlinger
frameworks/native/services/surfaceflinger
• A high-priority native (C++) daemon, started by init
with UID=system
• Services connections from activities via Binder
interface ISurfaceComposer
• Receives activity status from Activity Manager
• Receives window status (visibility, Z-order) from
Window Manager
• Composits multiple Surfaces into a single image
• Passes image to one or more displays
• Manages buffer allocation, synchronisation

SurfaceFlinger binder interfaces
SurfaceFlinger
Client
ISurfaceComposer ISurfaceComposerClient IDisplayEventConnection
e.g. activity
createConnection()
createDisplayEventConnection()
createSurface()
destroySurface()
getDataChannel()

ISurfaceComposer
• ISurfaceComposer
• Clients use this interface to set up a connection with
SurfaceFlinger
• Client begins by calling createConnection() which
spawns an ISurfaceComposerClient
• Client calls createGraphicBufferAlloc() to create an
instance of IGraphicBufferAlloc (discussed later)
• Client calls createDisplayEventConnection() to create
an instance of IDisplayEventConnection
• Other methods include captureScreen() and
setTransactionState()

ISurfaceComposerClient
• ISurfaceComposerClient
• This interface has two methods:
• createSurface() asks SufraceFlinger to create a new
Surface
• destroySurface() destroys a Surface

IDisplayEventConnection
• IDisplayEventConnection
• This interface passes vsync event information from
SurfaceFlinger to the client
• setVsyncRate() sets the vsync event delivery rate:
value of 1 returns all events, 0 returns none
• requestNextVsync() schedules the next vsync event:
has no effect if the vsync rate is non zero
• getDataChannel() returns a BitTube which can be
used to receive events

BufferQueue
frameworks/native/include/gui/BufferQueue.h
• Mechanism for passing GraphicBuffers to
SurfaceFlinger
• Contains an array of between 2 and 32
GraphicBuffers
• Uses interface IGraphicBufferAlloc to allocate buffers
(see later)
• Provides two Binder interfaces
• IGraphicBufferProducer for the client (Activity)
• IGraphicBufferConsumer for the consumer
(SurfaceFlinger)
• Buffers cycle between producer and consumer

BufferQueue state diagram
FREE
QUEUED
DEQUEUEDACQUIRED
IGraphicBufferProducer::
dequeueBuffer()
IGraphicBufferConsumer::
releaseBuffer()
queueBuffer()
IGraphicBufferConsumer::
acquireBuffer()
cancelBuffer()

BufferQueue
• Default number of buffer slots since JB is 3
(previously 2)
• In JB you can compile Layer.cpp with
TARGET_DISABLE_TRIPLE_BUFFERING to return to 2 slots
• Call setBufferCount() to change the number of slots
• BufferQueue operates in two modes:
• Synchronous: client blocks until there is a free slot
• Asynchronous: queueBuffer() discards any existing
buffers in QUEUED state so the queue only holds the
most recent frame

GraphicBuffer
frameworks/native/include/ui/GraphicBuffer.h
• Represents a buffer, wraps ANativeWindowBuffer
• Attributes including width, height, format, usage
inherited from ANativeWindowBuffer

Composition
• On a vsync event, SurfaceFlinger calls
handleMessageRefresh() which goes through a
composition cycle:
• preComposition(): sort layers by Z order and call
onPreComposition() for each
• doComposition(): loop through displays: if there is a
dirty region, mark it to be drawn then call
postFameBuffer() to do the drawing
• postComposition(): loop through layers in Z order and
call onPostComposition()

Layer
frameworks/native/services/surfaceflinger/Layer.h
• Each Layer has
• Z order
• Alpha value from 0 to 255
• visibleRegion
• crop region
• transformation: rotate 0, 90, 180, 270: ﬂip H, V: scale
• SurfaceFlinger composits the layers using
• HWComposer, if it supports the operation
• Fall back to the GPU, via OpenGL ES (version 1.0
only, for historical reasons)

HWComposer
hardware/libhardware/include/hardware/hwcomposer.h
• HWComposer is a vendor-supplied library, at run-time
in /system/lib/hw/hwcomposer.[product name].so
• Optional: in all cases there are fall-backs if HWC is
absent
• HWC does several different things
• sync framework (vsync callback)
• modesetting, display hotplug (e.g. hdmi)
• compositing layers together using features of the
display controller
• displaying frames on the screen

prepare() and set()
• SurfaceFlinger calls HWComposer in two stages
• prepare()
• Passes a list of layers
• For each layer, HWComposer returns
• HWC_FRAMEBUFFER: SurfaceFlinger should write this
layer (using OpenGL)
• HWC_OVERLAY: will be composed by HWComposer
• set()
• Passes the list of layers for HWComposer to handle
• set() is used in place of eglSwapBuffers()

vsync
• Since JB 4.1 SurfaceFlinger is synchronised to a
60Hz (16.7ms period) vsync event
• If HWComposer present, it is responsible for vsync
• Usually using an interrupt from the display: if no h/w
trigger, fake in software
• vsync() is a callback registered with HWComposer
• Each callback includes a display identiﬁer and a
timestamp (in ns)
• If no HWComposer, SurfaceFlinger uses 16ms
timeout in s/w

Displays
• HWComposer deﬁnes three display types
HWC_DISPLAY_PRIMARY e.g. built-in LCD screen
HWC_DISPLAY_EXTERNAL e.g. HDMI, WiDi
HWC_DISPLAY_VIRTUAL not a real display
• For each display there is an instance of
DisplayDevice in SurfaceFlinger

IGraphicBufferAlloc and friends
frameworks/native/include/gui/IGraphicBufferAlloc.h
• Binder interface used by SurfaceFlinger to allocate
buffers
• Has one function createGraphicBuffer
• Implemented by class GraphicBufferAllocator, which
wraps the ANativeWindowBuffer class
• Uses Gralloc.alloc to the the actual allocation
• Underlying buffer is referenced by a buffer_handle_t
which is a ﬁle descriptor (returned by gralloc alloc)
• Binder can pass open ﬁle descriptors from process to
process
• Access buffer data using mmap

Buffer usage and pixel format
frameworks/native/include/ui/GraphicBuffer.h
USAGE_HW_TEXTURE OpenGL ES texture
USAGE_HW_RENDER OpenGL ES render target
USAGE_HW_2D 2D hardware blitter
USAGE_HW_COMPOSER used by the HWComposer HAL
USAGE_HW_VIDEO_ENCODER HW video encoder
frameworks/native/include/ui/PixelFormat.h
PIXEL_FORMAT_RGBA_8888 4x8-bit RGBA
PIXEL_FORMAT_RGBX_8888 4x8-bit RGB0
PIXEL_FORMAT_RGB_888 3x8-bit RGB
PIXEL_FORMAT_RGB_565 16-bit RGB
PIXEL_FORMAT_BGRA_8888 4x8-bit BGRA

Gralloc
hardware/libhardware/include/hardware/gralloc.h
• Gralloc is a vendor-supplied library, at run-time in
/system/lib/hw/gralloc.[product name].so
• Does two things
• gralloc alloc: allocates graphic buffers
• gralloc framebuffer: interface to Linux framebuffer
device, e.g. /dev/graphics/fb0
• gralloc alloc allocates all graphic buffers using a
kernel memory manager, typically ION
• Selects appropriate ION heap based on the buffer
usage ﬂags

OpenGL ES
• The Khronos OpenGL ES and EGL APIs are
implemented in these libraries
• /system/lib/libEGL.so
• /system/lib/libGLESv1_CM.so
• /system/lib/libGLESv2.so
• /system/lib/libGLESv3.so (optional from JB 4.3
onwards: actually a symlink to libGLESv2.so)
• In most cases they simply call down to the
vendor-supplied libraries in /system/lib/egl

EGL
• EGL is the Khronos Native Platform Graphics
Interface
• Rendering operations are executed in an EGLContext
• In most cases the EGLContext is based on the
default display
• The mapping from the EGL generic display type is
done in
frameworks/native/opengl/include/EGL/eglplatform.h
typedef struct ANativeWindow* EGLNativeWindowType;
• EGLNativeWindowType is deﬁned in
system/core/include/system/window.h

OpenGL vendor implementation
• The vendor OpenGL libraries form the interface to the
GPU
• Responsible for
• creating display lists
• scheduling work for the GPU
• managing buffer synchronisation (typically using
fences, see background at the end)
• Usually there is a kernel driver which handles low
level memory management, DMA and interrupts
• The kernel interface is usually a group of ioctl
functions

• Questions?

Background: fences

Buffer synchronisation
• There are many producers and consumers of
graphics buffers
• Pre JB sync was implicit: buffer not released until
operation complete
• Did not encourage parallel processing
• JB introduced explicit sync: each buffer has a sync
object called a fence
• Means a buffer can be passed to the next user before
operations complete
• The next user waits on the fence before accessing
the buffer contents

Synchronisation using fences
• Represented by ﬁle handles: can be passed between
applications in binder messages
• Can also be passed from applications to drivers
• Each device driver (display, camera, video codec...)
has its own timeline
• A fence may have synchronisation points on multiple
timelines
• Allows buffers to be passed between multiple devices

Timeline and sync point
• Timeline
• Per-device (display, GPU, camera, ...)
• Monotonically increasing 32-bit value
• Incremented after each event (essentially it is a count
of the jobs processed by the device)
• Sync point
• A point on a timeline
• Becomes signalled when the timeline passes it

Fence
• Fence
• A collection of one or more sync points, possibly from
different timelines
• Represented by a ﬁle descriptor so an application can
wait using poll()
• Two fences can be merged to create a new fence that
depends on all the sync points of the original pair

Fence: example
FB
timeline
GPU
timeline
Sync pt
Sync pt
Sync
Fence

Background: ION

Memory constraints
• Often necessary for a buffer to be accessed by
hardware
• Example: graphics buffer and display controller or
GPU
• Hardware may constrain memory access
• Example: hardware without IOMMU usually needs
physically contiguous memory
• To avoid copying, the memory must be allocated for
the most constrained device

ION
• Previous memory allocators include pmem
(Qualcomm), cmem (TI), and nvmap (NVIDA)
• ION provides a uniﬁed interface for these needs
• Different allocation constraints
• Different caching requirements
• But the programmer still has to make the right choices

Types of heap
• ION_HEAP_TYPE_SYSTEM
• memory allocated via vmalloc
• ION_HEAP_TYPE_SYSTEM_CONTIG
• memory allocated via kmalloc
• ION_HEAP_TYPE_CARVEOUT
• memory allocated from a pre reserved carveout heap
• allocations are physically contiguous

Heap ﬂags
• ION_FLAG_CACHED
• mappings of this buffer should be cached, ION will do
cache maintenance when the buffer is mapped for
DMA
• ION_FLAG_CACHED_NEEDS_SYNC
• Cache must be managed manually, e.g. using
ION_IOC_SYNC

The Android graphics path, in depth

More Related Content

What's hot (20)

Similar to The Android graphics path, in depth (20)

More from Chris Simmonds (19)

Recently uploaded (20)

The Android graphics path, in depth