Fixes the cases where clip stack reduction would cause clip to be
re-rendered to stencil for each draw call. This causes unneeded
slowdown.
Stencil cache would not be used because the clip stack generation id communicated
by the clip stack element list would be invalid. This happended due to
a) clip stack reduction creating new elements in the element list.
b) purging logic removing the generation id, but reduction logic
selecting already purged element, and thus the generation id, as
the representative state of the clip.
Cases of a) where reduction would flatten the stack to a single new
element were fixed by assigning the generation id of the top-most
element of the clip stack as the generation id of the new
element. This is not strictly minimal, but enables more caching than
using invalid id.
Cases of a) where reduction would substitute a stack element with a
new element the generation id of the substituted element is used.
The b) part was fixed by removing the purging logic. It was not
exactly correct, as the previously purged states were actually
used. The purging was not used for anything.
Changes SkClipStack API to highlight that invalid generation id is
never returned by SkClipStack. Empty stacks are wide open. Changes the
clients to reflect this.
Fixes a crash when not passing anti-alias out parameter to
GrReducedClip::ReduceClipStack. The crash is not exercised in the
current code.
R=bsalomon@google.com, robertphillips@google.com
Author: kkinnunen@nvidia.com
Review URL: https://codereview.chromium.org/48593003
git-svn-id: http://skia.googlecode.com/svn/trunk@12084 2bbb7eff-a529-9590-31e7-b0007b416f81
In some cases, the allocated array into which the data will be read is using getArrayCount() to allocate itself, which should be safe, but some cases use fixed length arrays or compute the array size before reading, which could overflow if the stream is compromised.
To prevent that from happening, I added a check that will verify that the number of bytes to read will not exceed the capacity of the input buffer argument passed to all the read...Array() functions.
I chose to use the byte array for this initial version, so that "size" represents the same value across all read...Array() functions, but I could also use the element count, if it is preferred.
Note : readPointArray and writePointArray are unused, so I could also remove them
BUG=
R=reed@google.com, mtklein@google.com, senorblanco@chromium.org
Author: sugoi@chromium.org
Review URL: https://codereview.chromium.org/37803002
git-svn-id: http://skia.googlecode.com/svn/trunk@12058 2bbb7eff-a529-9590-31e7-b0007b416f81
- SkScaledImageCache:
- Add new FindAndLock/AddAndLock variants that work well with
SkLazyPixelRefs (take width, height, generation_id).
- Add static versions of these new variants.
- SkLazyPixelRef:
- If NULL passed in as SkImageCache* in the constructor, it will
now default to using the static SkScaledImageCache methods to
cache decoded images.
- If (fImageCache==NULL), the default allocator can be changed
with the setAllocator method. If (fImageCache!=NULL), the
SkImageCache handles allocation.
- CachedDecodingPixelRefTest to test the new functionality.
BUG=
R=scroggo@google.com, mtklein@google.com, reed@google.com
Author: halcanary@google.com
Review URL: https://codereview.chromium.org/37343002
git-svn-id: http://skia.googlecode.com/svn/trunk@12006 2bbb7eff-a529-9590-31e7-b0007b416f81
This adds an invalidation listener mechanism to SkPixelRef to let it send this message while still staying ignorant of who's listening.
These messages are tricky to deliver. The SkPixelRefs they originates from and the GrResourceCaches they ultimately end up at may be on different threads; neither class is threadsafe; their object lifetimes are totally independent; it's a many-senders-to-many-receivers relation; and neither codebase should really know about the other.
So I've added a per-message-type global message bus to broadcast messages to threadsafe inboxes. Anyone can post() a message, which will show up in all the inboxes of that type, read whenever the inbox's owner calls poll(). The implementation is _dumb_; it can be improved in several dimensions (inbox size limits, lock-free message delivery) if we find the need.
I took some care to make sure not to send the invalidation message for any SkPixelRef that's sharing a generation ID with another SkPixelRef.
BUG=
R=bsalomon@google.com, scroggo@google.com, reed@google.com
Author: mtklein@google.com
Review URL: https://codereview.chromium.org/26734003
git-svn-id: http://skia.googlecode.com/svn/trunk@11949 2bbb7eff-a529-9590-31e7-b0007b416f81
This returns true if (1) the picture has finished recording and
(2) this picture or any picture drawn into it refers to any bitmaps.
It allows clients doing complicated manipulations of the picture to
early-out when there are no bitmaps present.
BUG=303281
R=reed@google.com
git-svn-id: http://skia.googlecode.com/svn/trunk@11935 2bbb7eff-a529-9590-31e7-b0007b416f81
Adds GrEffect::willUseInputColor() which indicates whether or not the
input color affects the output of the effect. This is needed for
certain Xfermodes, such as kSrc_Mode. For these modes the color filter
will not use the input color.
An effect with GrEffect::willUseInputColor() true will cause all color
or coverage effects before it to be discarded, as their computations
cannot affect the output. In these cases program is marked as having
white input color.
This fixes an assert when Skia is compiled in a mode that prefers
using uniforms instead of attributes for constants. (Flags
GR_GL_USE_NV_PATH_RENDERING or GR_GL_NO_CONSTANT_ATTRIBUTES). Using
attributes hides the problem where the fragment shader does not need
input color for color filters that ignore DST part of the filter. The
assert would be hit when uniform manager tries to bind an uniform which
has been optimized away by the shader compiler.
Adds specific GrGLSLExpr4 and GrGLSLExpr1 classes. This way the GLSL
expressions like "(v - src.a)" can remain somewhat readable in form of
"(v - src.a())". The GrGLSLExpr<typename> template implements the
generic functionality, GrGLSLExprX is the specialization that exposes
the type-safe interface to this functionality.
Also adds operators so that GLSL binary operators of the form
"(float * vecX)" can be expressed in C++. Before only the equivalent
"(vecX * float)" was possible. This reverts the common blending
calculations to more conventional order, such as "(1-a) * c" instead of
"c * (1-a)".
Changes GrGLSLExpr1::OnesStr from 1 to 1.0 in order to preserve the
color filter blending formula string the same (with the exception of
variable name change).
Shaders change in case of input color being needed:
- vec4 filteredColor;
- filteredColor = (((1.0 - uFilterColor.a) * output_Stage0) + uFilterColor);
- fsColorOut = filteredColor;
+ vec4 output_Stage1;
+ { // Stage 1: ModeColorFilterEffect
+ output_Stage1 = (((1.0 - uFilterColor_Stage1.a) * output_Stage0) + uFilterColor_Stage1);
+ }
+ fsColorOut = output_Stage1;
Shaders change in case of input color being not needed:
-uniform vec4 uFilterColor;
-in vec4 vColor;
+uniform vec4 uFilterColor_Stage0;
out vec4 fsColorOut;
void main() {
- vec4 filteredColor;
- filteredColor = uFilterColor;
- fsColorOut = filteredColor;
+ vec4 output_Stage0;
+ { // Stage 0: ModeColorFilterEffect
+ output_Stage0 = uFilterColor_Stage0;
+ }
+ fsColorOut = output_Stage0;
}
R=bsalomon@google.com, robertphillips@google.com, jvanverth@google.com
Author: kkinnunen@nvidia.com
Review URL: https://codereview.chromium.org/25023003
git-svn-id: http://skia.googlecode.com/svn/trunk@11912 2bbb7eff-a529-9590-31e7-b0007b416f81
Make GrProgramsTest check how many texture coordinate sets are
available and select random effects up until the amount runs out.
Otherwise, following effect sequence would fail the shader compilation
when Skia is compiled with nv_path_rendering on (eg. when fixed
function codepath is used):
* Stage 0: TextureDomain (1 texcoord)
* Stage 1: Convolution (1 texcoord)
* Stage 2: Bitmap Alpha Threshold (2 texcoords)
* Stage 3: DisplacementMap (2 texcoords)
* Stage 4: Config Conversion (1 texcoords)
* Stage 5: Two-Point Conical Gradient (2 texcoords)
This would use more texture coordinate sets than 8, which is fairly
common amount currently.
R=bsalomon@google.com, cdalton@nvidia.com
Author: kkinnunen@nvidia.com
Review URL: https://codereview.chromium.org/32403002
git-svn-id: http://skia.googlecode.com/svn/trunk@11881 2bbb7eff-a529-9590-31e7-b0007b416f81
This change removes sk_stdint.h since it is only needed for vs2008 and earlier.
This change removes SK_MMAP_SUPPORT define since it is no longer used.
This change removes the stdio.h include from SkTypes.h since on many systems
this is a very large header, few Skia files actually use it, it is
available everywhere standard, and SkDebugf should be used instead.
After this change there is no need for external users to put Skia's
include/config into their own list of includes, saving the headache
of having two header files of the same name and sometimes getting the
wrong one depending on include order.
R=bsalomon@google.com, djsollen@google.com
Review URL: https://codereview.chromium.org/27044002
git-svn-id: http://skia.googlecode.com/svn/trunk@11738 2bbb7eff-a529-9590-31e7-b0007b416f81
Add new runtime configuration variable,
images.gif.suppressDecoderWarnings, which suppresses warning and
errors from the GIF library. It defaults to "true", which is current
behavior.
(This setting can be changed by setting the environment variable
skia_images_gif_suppressDecoderWarnings="false".)
Some conditions which were errors before are now warnings:
- If the image width or height is greater than the GIF screen width or
height (respectively) we expand the screen to hold the image.
- If the offset of the image inside the screen would place the
image outside of the screen, we shift the image to fix this.
- If the image lacks a color table, we create a default color table.
- If the image is truncated, then the rest of the image is filled with
the fill color.
In all four cases, if images.gif.suppressDecoderWarnings is set to
false, then a warning message is printed via SkDebugf.
In the event of another kind of error, SkGIFImageDecoder::onDecode()
will still return false. But with this change, if
images.gif.suppressDecoderWarnings is set to false, a description of
the error is printed via SkDebugf.
Also, added a new unit test GifTest, which tests the deconing of both
good GIf files and corrupted files that should now work with this
change. This unit test is disabled on Win32, iOS, and Mac.
BUG=skia:1689
R=scroggo@google.com
Review URL: https://codereview.chromium.org/26743002
git-svn-id: http://skia.googlecode.com/svn/trunk@11734 2bbb7eff-a529-9590-31e7-b0007b416f81
I noticed SkMatrix <-> SkMatrix44 conversions were dropping the
perspective values on the floor. As we use SkMatrix44 heavily in
Chromium, I'm concerned this missing code will cause a bug eventually.
It should be correct to simply use the bottom row of the 4x4 matrix
excluding the third column.
Previously committed and reverted, second attempt with fix for
incorrect use of SkMScalar/SkScalar.
BUG=
R=reed@google.com, caryclark@google.com
Author: aelias@chromium.org
Review URL: https://codereview.chromium.org/25484006
git-svn-id: http://skia.googlecode.com/svn/trunk@11624 2bbb7eff-a529-9590-31e7-b0007b416f81
I noticed SkMatrix <-> SkMatrix44 conversions were dropping the
perspective values on the floor. As we use SkMatrix44 heavily in
Chromium, I'm concerned this missing code will cause a bug eventually.
It should be correct to simply use the bottom row of the 4x4 matrix
excluding the third column.
BUG=
R=reed@google.com
Author: aelias@chromium.org
Review URL: https://codereview.chromium.org/25484006
git-svn-id: http://skia.googlecode.com/svn/trunk@11622 2bbb7eff-a529-9590-31e7-b0007b416f81
This change address what happens when a jpeg is partially downloaded
before failing. Many browsers will render it anyway: we want Skia to
do the same. The JpegTest takes a perfectly cromulent jpeg file and
only passes into the ImageDecoder the first half of the image. We
then verify that the image decoder returns a valid bitmap of the
correct dimensions.
We also fixed some png library errors, including issue 1691.
Also, suppressed the majority of warnings from using libpng and
libjpeg. By default, most warnings are *not* suppressed in debug mode.
If you have a debug binary and wish to suppress warnings, set the
following environment variables to true
skia_images_png_suppressDecoderWarnings
skia_images_jpeg_suppressDecoderWarnings
or from within a program that links to Skia:
#if defined(SK_DEBUG)
#include "SkRTConf.h"
SK_CONF_SET("images.jpeg.suppressDecoderWarnings", true);
SK_CONF_SET("images.png.suppressDecoderWarnings", true);
#endif
I tested this, before (control) and after these changes (test), on
364,295 skps from the cluster telemetry.
- number of errors+warnings in control = 2804
- number of errors+warnings fixed = 2283
- number of PNG verbosity fixed = 2152
- number of PNG error fixed = 4
- number of PNG segfault fixed = 3
- number of PNG errors changed to warnings = 62
- number of JPG verbosity fixed = 26
- number of JPG error fixed = 91
Not all errors and warning have been fixed.
These numbers were generated using the find_bad_images_in_skps.py
program. This program may be useful going forward for testing
image-decoding libraries on skp files from the cluster telemetry.
find_bad_images_in_skps.py depends on the test_image_decoder program,
which simply executes the SkImageDecoder::DecodeFile function and uses
its exit status to report success or failure.
BUG=skia:1649
BUG=skia:1691
BUG=skia:1680
R=scroggo@google.com
Review URL: https://codereview.chromium.org/24449003
git-svn-id: http://skia.googlecode.com/svn/trunk@11597 2bbb7eff-a529-9590-31e7-b0007b416f81
These guys are not heavily contended nor speed critical. No need for atomics,
plus this makes tsan stop complaining (correctly) about reading fNextIndex
unsafely in onEnd.
I took a look at failCount/fFailCount, which I think is safely atomic and quite
conveniently so: It's never read until all the threads which could possibly
increment it have terminated (except for the one where it was created,
obviously). We could guard it with a mutex too, but maybe we can let this one
slide.
BUG=
R=bungeman@google.com
Author: mtklein@google.com
Review URL: https://codereview.chromium.org/25357002
git-svn-id: http://skia.googlecode.com/svn/trunk@11561 2bbb7eff-a529-9590-31e7-b0007b416f81
Only implemented for PNG.
Add a getter and setter, and sets the default to false in the
constructor. Also copies the setting in copyFieldsToOther.
Fix an indpendent bug where fDitherImage was not being copied in
copyFieldsToOther.
In SkScaledBitmapSampler::begin, consolidate the settings passed in
by passing a const reference to the decoder. The decoder can be
referenced for its settings of dither, unpremultiplied, and now
skipping writing zeroes. Update callers to use the new API. In png
decoder, rather than passing around a pointer to an initial
read of getDitherImage, and potentially changing it, look at the
field on the decoder itself, and modify it directly. This is a
change in behavior - now if that same decoder is used to decode
a different image, the dither setting has changed. I think this is
okay because A) the typical use case is to use a new decoder for
each decode, B) we do not make any promises that a decode does not
change the decoder and C) it makes the code in SkScaledBitmapSampler
much cleaner.
In SkScaledBitmapScampler, add new row procs for skipping zeroes. Now
that choosing the row proc has five dimensions (src config, dst config,
dither, skip writing zeroes, unpremultiplied), use a new method: each
src/dst combination has a function for choosing the right proc depending
on the decoder.
SkScaledBitmapScampler::RowProc is now public for convenience.
Remove Sample_Gray_D8888_Unpremul, which is effectively no different
from Sample_Gray_D8888.
In cases where unpremultiplied was trivial, such as 565 and when
sampling from gray, decoding may now succeed.
Add a benchmark (currently disabled) for comparing the speed of skipping
writing zeroes versus not skipping. For this particular image, which is
mostly transparent pixels, normal decoding took about 3.6 milliseconds,
while skipping zeroes in the decode took only about 2.5 milliseconds
(this is on a Nexus 4). Presumably it would be slower on an image
with a small amount of transparency, but there will be no slowdown
for an image which reports that it has no transparency.
In SkImageRef_ashmem, always skip writing zeroes, since ashmem
memory is guaranteed to be initialized to 0.
Add a flag to skip writing zeroes in skimage.
Add a regression test for choosing the rowproc to ensure I did not
change any behavior accidentally.
BUG=skia:1661
R=reed@google.com
Review URL: https://codereview.chromium.org/24269006
git-svn-id: http://skia.googlecode.com/svn/trunk@11558 2bbb7eff-a529-9590-31e7-b0007b416f81
Only affects factories, static functions that will use the factories,
and subset decoding, which all require rewinding. The decoders
themselves continue to take an SkStream. This is merely documentation
stating which functions will possibly rewind the passed in SkStream.
This is part of the general change to coordinate SkStreams with
Android's streams, which don't necessarily support rewinding in all
cases.
Update callers to use SkStreamRewindable.
BUG=skia:1572
R=bungeman@google.com, reed@google.com
Review URL: https://codereview.chromium.org/23477009
git-svn-id: http://skia.googlecode.com/svn/trunk@11460 2bbb7eff-a529-9590-31e7-b0007b416f81
I'm seeing basically no difference between malloc + bzero and calloc on my desktop, but on a Galaxy Nexus calloc is never slower, and significantly faster once the allocation size becomes large, both for allocation and for _reading_.
BUG=skia:1662
R=reed@google.com
Review URL: https://codereview.chromium.org/24251008
git-svn-id: http://skia.googlecode.com/svn/trunk@11414 2bbb7eff-a529-9590-31e7-b0007b416f81
Last week, the fuzzer found a few numerical issue with filters and I had written some fixes for them. Here are the fixes with some unit tests.
For senorblanco : So I figured out what was asserting when we'd get a 0 width "result" in SkBicubicImageFilter::onFilterImage(). Basically, if the "result" SkBitmap object calls SkBitmap::setConfig() with "width" and/or "height" set to 0, then the SkBitmap object will call SkBitmap::reset(), making the SkBitmap object's config invalid. At this point, calling SkBitmap::getAddr32() will assert, even without attempting to dereference the data pointer, because the SkBitmap's config is invalid. If height is valid, but width is 0, then this call to SkBitmap::getAddr32() happens directly in SkBicubicImageFilter::onFilterImage() a few lines lower and asserts right away.
BUG=
R=senorblanco@google.com, senorblanco@chromium.org, bsalomon@google.com
Author: sugoi@chromium.org
Review URL: https://chromiumcodereview.appspot.com/23533042
git-svn-id: http://skia.googlecode.com/svn/trunk@11249 2bbb7eff-a529-9590-31e7-b0007b416f81
Obviously these are all currently function pointers of type T(*)(P) for various
T and P. In bench refactoring, I'm trying to register a function pointer of
type T(*)(), which can't be done as is (passing P=void doesn't work). This
also lets us register things like primitives, which is conceivable useful.
BUG=
R=reed@google.com, scroggo@google.com
Review URL: https://codereview.chromium.org/23453031
git-svn-id: http://skia.googlecode.com/svn/trunk@11082 2bbb7eff-a529-9590-31e7-b0007b416f81
cleaned up SkPictureFlat.h quite a bit while working on this.
bench --match picture_record_ shows some improvement:
compare.sh bench --match picture_record_ --repeat 100
master -> usehash
N=3 p=0.001000 (corrected to 0.000333)
sig? rel. speed bench
y 1.0x picture_record_dictionaries
y 1.5x picture_record_recurring_paint_dictionary
y 3.8x picture_record_unique_paint_dictionary
Overall relative speed: 1.9x
bench_pictures --record is pretty much neutral:
compare.sh bench_pictures -r ../skp --mode record --repeat 30
master -> usehash
N=63 p=0.001000 (corrected to 0.000016)
sig? rel. speed bench
n 0.9x desk_pokemonwiki.skp
y 0.9x desk_googlespreadsheet.skp
y 0.9x tabl_pravda.skp
y 1.0x desk_googlespreadsheetdashed.skp
n 1.0x tabl_onlinewsj.skp
n 1.0x tabl_nytimes.skp
n 1.0x desk_googlehome.skp
y 1.0x desk_techcrunch.skp
n 1.0x tabl_slashdot.skp
n 1.0x tabl_techmeme.skp
n 1.0x desk_googleplus.skp
n 1.0x desk_sfgate.skp
n 1.0x tabl_transformice.skp
n 1.0x desk_espn.skp
n 1.0x desk_baidu.skp
n 1.0x tabl_worldjournal.skp
n 1.0x desk_chalkboard.skp
n 1.0x tabl_frantzen.skp
n 1.0x desk_gws.skp
n 1.0x tabl_androidpolice.skp
n 1.0x desk_linkedin.skp
n 1.0x mobi_wikipedia.skp
n 1.0x desk_wowwiki.skp
n 1.0x desk_css3gradients.skp
n 1.0x desk_gmailthread.skp
n 1.0x desk_yahoogames.skp
n 1.0x desk_facebook.skp
n 1.0x desk_wordpress.skp
n 1.0x tabl_vnexpress.skp
n 1.0x desk_br337.skp
n 1.0x tabl_engadget.skp
n 1.0x tabl_theverge.skp
n 1.0x desk_amazon.skp
n 1.0x desk_ebay.skp
n 1.0x tabl_hsfi.skp
n 1.0x tabl_sahadan.skp
n 1.0x desk_weather.skp
n 1.0x tabl_digg.skp
n 1.0x desk_youtubetvbrowse.skp
n 1.0x tabl_culturalsolutions.skp
n 1.0x tabl_ukwsj.skp
n 1.0x desk_youtube.skp
n 1.0x tabl_googlecalendar.skp
y 1.0x desk_yahooanswers.skp
n 1.0x desk_blogger.skp
n 1.0x desk_yahoonews.skp
y 1.0x desk_yahoosports.skp
y 1.0x tabl_mercurynews.skp
n 1.0x desk_youtubetvvideo.skp
y 1.0x tabl_gspro.skp
y 1.1x tabl_googleblog.skp
y 1.1x tabl_cnet.skp
y 1.1x tabl_mlb.skp
y 1.1x tabl_cuteoverload.skp
y 1.1x desk_booking.skp
y 1.1x tabl_deviantart.skp
y 1.1x desk_twitter.skp
y 1.1x tabl_cnn.skp
y 1.1x tabl_gamedeksiam.skp
y 1.1x tabl_gmail.skp
y 1.1x tabl_nofolo.skp
y 1.1x tabl_mozilla.skp
y 1.1x desk_pinterest.skp
Overall relative speed: 1.0x
(I'd take this to mean that the microbenches are probably drifting away from relevance.)
BUG=
R=reed@google.com
Review URL: https://codereview.chromium.org/21564008
git-svn-id: http://skia.googlecode.com/svn/trunk@10825 2bbb7eff-a529-9590-31e7-b0007b416f81
- add validate()
- make add() and remove() strict
- fill in maybeShrink()
- make grow and shrink thresholds configurable.
- fix issue where we were getting filled with deleted items - we need to count them as occupied when determining if we should grow
BUG=
R=reed@google.com
Review URL: https://codereview.chromium.org/22571010
git-svn-id: http://skia.googlecode.com/svn/trunk@10677 2bbb7eff-a529-9590-31e7-b0007b416f81
Use SkIntToScalar.
Reflect that 8888 can be copied to 4444 in BitmapCopyTest.
Unreviewed.
git-svn-id: http://skia.googlecode.com/svn/trunk@10623 2bbb7eff-a529-9590-31e7-b0007b416f81
