NOTE: Linear blending is currently disabled in this CL as the
feature is still a work in progress
Android currently performs all blending (any kind of linear math
on colors really) on gamma-encoded colors. Since Android assumes
that the default color space is sRGB, all bitmaps and colors
are encoded with the sRGB Opto-Electronic Conversion Function
(OECF, which can be approximated with a power function). Since
the power curve is not linear, our linear math is incorrect.
The result is that we generate colors that tend to be too dark;
this affects blending but also anti-aliasing, gradients, blurs,
etc.
The solution is to convert gamma-encoded colors back to linear
space before doing any math on them, using the sRGB Electo-Optical
Conversion Function (EOCF). This is achieved in different
ways in different parts of the pipeline:
- Using hardware conversions when sampling from OpenGL textures
or writing into OpenGL frame buffers
- Using software conversion functions, to translate app-supplied
colors to and from sRGB
- Using Skia's color spaces
Any type of processing on colors must roughly ollow these steps:
[sRGB input]->EOCF->[linear data]->[processing]->OECF->[sRGB output]
For the sRGB color space, the conversion functions are defined as
follows:
OECF(linear) :=
linear <= 0.0031308 ? linear * 12.92 : (pow(linear, 1/2.4) * 1.055) - 0.055
EOCF(srgb) :=
srgb <= 0.04045 ? srgb / 12.92 : pow((srgb + 0.055) / 1.055, 2.4)
The EOCF is simply the reciprocal of the OECF.
While it is highly recommended to use the exact sRGB conversion
functions everywhere possible, it is sometimes useful or beneficial
to rely on approximations:
- pow(x,2.2) and pow(x,1/2.2)
- x^2 and sqrt(x)
The latter is particularly useful in fragment shaders (for instance
to apply dithering in sRGB space), especially if the sqrt() can be
replaced with an inversesqrt().
Here is a fairly exhaustive list of modifications implemented
in this CL:
- Set TARGET_ENABLE_LINEAR_BLENDING := false in BoardConfig.mk
to disable linear blending. This is only for GLES 2.0 GPUs
with no hardware sRGB support. This flag is currently assumed
to be false (see note above)
- sRGB writes are disabled when entering a functor (WebView).
This will need to be fixed at some point
- Skia bitmaps are created with the sRGB color space
- Bitmaps using a 565 config are expanded to 888
- Linear blending is disabled when entering a functor
- External textures are not properly sampled (see below)
- Gradients are interpolated in linear space
- Texture-based dithering was replaced with analytical dithering
- Dithering is done in the quantization color space, which is
why we must do EOCF(OECF(color)+dither)
- Text is now gamma corrected differently depending on the luminance
of the source pixel. The asumption is that a bright pixel will be
blended on a dark background and the other way around. The source
alpha is gamma corrected to thicken dark on bright and thin
bright on dark to match the intended design of fonts. This also
matches the behavior of popular design/drawing applications
- Removed the asset atlas. It did not contain anything useful and
could not be sampled in sRGB without a yet-to-be-defined GL
extension
- The last column of color matrices is converted to linear space
because its value are added to linear colors
Missing features:
- Resource qualifier?
- Regeneration of goldeng images for automated tests
- Handle alpha8/grey8 properly
- Disable sRGB write for layers with external textures
Test: Manual testing while work in progress
Bug: 29940137
Change-Id: I6a07b15ab49b554377cd33a36b6d9971a15e9a0b
bug:17478770
This removes a lot of redundant property query code, and puts the
queries all in one place, so defining them automatically will be simpler
in the future.
Change-Id: I0428550e6081f07bc6554ffdf73b22284325abb8
Adds remaining missing overrides and nullptr usages, missed due to
an extreme failure in tool usage.
Change-Id: I56abd72975a3999ad13330003c348db40f59aebf
Narrow the use of #include directives in hwui, replacing with forward
declarations where straightforward. Speeds compiles; doesn't do any
restructuring of code.
Change-Id: Icac2baffb5896f55d8c6718e9bd9d4bfa02d3ca0
A Patch can be fairly large, holding bitmap data, but
is also frequently leaked which adds to the severity.
The feature is used in many important processes such
as Home, SystemUI and Chrome.
The following leaks are solved:
1. The Patch itself was not always freed.
PatchCache::removeDeferred() can mark patches to be
cared for by PatchCache::clearGarbage(). But
mCache.remove() would only destroy the container
and the pointer, not the Patch object itself.
2. The vertices stored in the Patch at Patch::createMesh()
would always leak. The empty/default destructor in Patch
would not properly destroy "vertices" since it's just a
pointer.
3. A BufferBlock that's added to the mFreeBlocks
in PatchCache could leak. The leak happened when a
patch later needed the entire free block, because the
object was removed from the list but never deleted
in PatchCache::setupMesh().
Change-Id: I41e60824479230b67426fc546d3dbff294c8891f
Some caches(PatchCache, TextureCache, PathCache) for HWUI
uses deferred removal for their cache entries even though
actual resource objects are immediately freed by
ResourceCache.
For this reason, the uniqueness of a resource address in
the caches is not guaranteed in specific cases.
(Because malloc() can return the same address when malloc()
and free() called very frequently.)
So it can be possible the cache have two cache entries for
two different resources but the same memory address.
(Of course one of the resources is already freed.)
It also can be possible mGarbage vector in PatchCache has
duplicated addresses and this can lead to duplicated free
blocks in the free block list and graphics corruption.
(Deferred removal was implmeneted based on an assumption of
unique resource addresses.)
So this patch makes sure resource objects are freed after
the resources are removed from the caches to guarantee
the uniqueness of a resource address and prevent graphics
corruption.
Change-Id: I040f033a4fc783d2c4bc04b113589657c36fb15b
Signed-off-by: Sangkyu Lee <sk82.lee@lge.com>
This change adds refcounting of Res_png_9patch instances, the native
data structure used to represent 9-patches. The Dalvik NinePatch class
now holds a native pointer instead of a Dalvik byte[]. This pointer
is used whenever we need to draw the 9-patch (software or hardware.)
Since we are now tracking garbage collection of NinePatch objects
libhwui's PatchCache must keep a list of free blocks in the VBO
used to store the meshes.
This change also removes unnecessary instances tracking from
GLES20DisplayList. Bitmaps and 9-patches are refcounted at the
native level and do not need to be tracked by the Dalvik layer.
Change-Id: Ib8682d573a538aaf1945f8ec5a9bd5da5d16f74b
This change also fixes the way batched bitmaps were handled
inside a layer. The layer is now correctly dirtied to minimize
the amount of pixels to blend.
Fix alpha, mode and opaque computations for DrawPatchOp.
Change-Id: I1b6cd581c0f0db66c1002bb4fb1a9811e55bfa78
This optimization saves up to 0.3ms per frame on the Play Store's
front page, on a Nexus 4 device.
Change-Id: Iaa4ef33c6e3b37e175efd5b9eea9ef59b43f14f3
The 9aptch cache was reinitialized after destroying/recreating
the EGL context but not after clearing it during a normal
memory trim.
Change-Id: If6155bfc8a62439e9878bc742a4766b3bd6c6aec
When the Android runtime starts, the system preloads a series of assets
in the Zygote process. These assets are shared across all processes.
Unfortunately, each one of these assets is later uploaded in its own
OpenGL texture, once per process. This wastes memory and generates
unnecessary OpenGL state changes.
This CL introduces an asset server that provides an atlas to all processes.
Note: bitmaps used by skia shaders are *not* sampled from the atlas.
It's an uncommon use case and would require extra texture transforms
in the GL shaders.
WHAT IS THE ASSETS ATLAS
The "assets atlas" is a single, shareable graphic buffer that contains
all the system's preloaded bitmap drawables (this includes 9-patches.)
The atlas is made of two distinct objects: the graphic buffer that
contains the actual pixels and the map which indicates where each
preloaded bitmap can be found in the atlas (essentially a pair of
x and y coordinates.)
HOW IS THE ASSETS ATLAS GENERATED
Because we need to support a wide variety of devices and because it
is easy to change the list of preloaded drawables, the atlas is
generated at runtime, during the startup phase of the system process.
There are several steps that lead to the atlas generation:
1. If the device is booting for the first time, or if the device was
updated, we need to find the best atlas configuration. To do so,
the atlas service tries a number of width, height and algorithm
variations that allows us to pack as many assets as possible while
using as little memory as possible. Once a best configuration is found,
it gets written to disk in /data/system/framework_atlas
2. Given a best configuration (algorithm variant, dimensions and
number of bitmaps that can be packed in the atlas), the atlas service
packs all the preloaded bitmaps into a single graphic buffer object.
3. The packing is done using Skia in a temporary native bitmap. The
Skia bitmap is then copied into the graphic buffer using OpenGL ES
to benefit from texture swizzling.
HOW PROCESSES USE THE ATLAS
Whenever a process' hardware renderer initializes its EGL context,
it queries the atlas service for the graphic buffer and the map.
It is important to remember that both the context and the map will
be valid for the lifetime of the hardware renderer (if the system
process goes down, all apps get killed as well.)
Every time the hardware renderer needs to render a bitmap, it first
checks whether the bitmap can be found in the assets atlas. When
the bitmap is part of the atlas, texture coordinates are remapped
appropriately before rendering.
Change-Id: I8eaecf53e7f6a33d90da3d0047c5ceec89ea3af0
Bug #7970966
The bug described in #7970966 should normally never happen but just in
case, change the detection code to be more robust.
Change-Id: I7040a6087590e34abe8803cb8f83f051d77f3944
This change detects empty quads in 9patches and removes them from
the mesh to avoid unnecessary blending.
Change-Id: I4500566fb4cb6845d64dcb59b522c0be7a0ec704
