There's little benefit to deduping matrices and regions: they're infrequently
used, and doubly infrequently reused. Their use-weighted byte cost is tiny.
There is some downside to deduping matrices and regions. Even when they're not
used, we prepare dictionaries for deduping them for every picture. Each of
these dictionaries costs 160 bytes, so two unused dictionaries make a big chunk
of the ~1100 bytes it takes to allocate an SkPictureRecord. (~330 come from
parent class SkCanvas, 768 from SkPictureRecord itself, here reduced to 448).
One side benefit of not deduping these guys is that the change weighs -140 lines of code.
It may go without saying, but this breaks the picture format.
Testing: out/Debug/tests && out/Debug/dm (which runs all picture modes by default)
BUG=skia:1850
R=reed@google.com, bensong@google.com, robertphillips@google.com
Author: mtklein@google.com
Review URL: https://codereview.chromium.org/143883006
git-svn-id: http://skia.googlecode.com/svn/trunk@13149 2bbb7eff-a529-9590-31e7-b0007b416f81
Make leak counters implemented with SK_DECLARE_INST_COUNT thread-safe.
Enable the leak counting for Debug builds when Skia is built as a
static library. Having SK_DECLARE_INST_COUNT without
SK_DEFINE_INST_COUNT relies on static variables in member functions
declared in the header files. These might be duplicated in the clients
of the library when Skia is built as a dynamic library, producing
incorrect operation.
Protect the instance counter initialization step (initStep) by
using SkOnce.
Makes SkOnce.h part of the public API, since SkInstCnt is public.
Protect the per-class child list shared variable with a per-class mutex.
Changes the behavior in the way that if the child list has been
"cleaned up", it will still try to create subsequent child lists.
BUG=skia:1219
R=robertphillips@google.com, mtklein@google.com, bsalomon@google.com, bungeman@google.com, djsollen@google.com
Author: kkinnunen@nvidia.com
Review URL: https://codereview.chromium.org/99483003
git-svn-id: http://skia.googlecode.com/svn/trunk@13120 2bbb7eff-a529-9590-31e7-b0007b416f81
This reduces the allocation overhead of a null picture (create, beginRecording(), endRecording) from about 18K to about 1.9K. (There's still lots more to prune.)
SkPictureFlat can exploit the fact that Writer32 is contiguous simplify its memory management. The Writer32 itself becomes the scratch buffer.
Remove lots and lots of arbitrary magic numbers that were size guesses and minimum allocation sizes. Keep your eyes open for the big obvious DUH why we save 16K per picture! (Spoiler alert. It's because that first save we issue in beginRecording() forces the old SkWriter32 to allocate 16K.)
Tests passing, DM passing.
bench --match writer: ~20% faster
null bench_record: ~30% faster
bench_record on buildbot .skps: ~3-6% slower, ranging 25% faster to 20% slower
bench_pictures on buildbot .skps: ~1-2% faster, ranging 13% faster to 28% slower
BUG=skia:1850
R=reed@google.com
Author: mtklein@google.com
Review URL: https://codereview.chromium.org/137433003
git-svn-id: http://skia.googlecode.com/svn/trunk@13073 2bbb7eff-a529-9590-31e7-b0007b416f81
SkOnceFlag is now statically initializable on all platforms.
Also adds sk_atomic_cas, used to implement new SkSpinlock.
Going to punt on making SkOnceFlag any smaller (for now, it's 8 bytes). We could conceivably get it down to two bits, one for done and one for a one-bit spinlock (we'd need atomic-& and atomic-| to make that work, but they appear to be available everywhere).
BUG=skia:1929
R=bungeman@google.com, reed@google.com
Author: mtklein@google.com
Review URL: https://codereview.chromium.org/123093002
git-svn-id: http://skia.googlecode.com/svn/trunk@12968 2bbb7eff-a529-9590-31e7-b0007b416f81
To do this, this patch changes the "offset/loc" parameter in filterImage() / onFilterImage() from an inout-param to an out-param only, so that the calling filter can know how much the input filter wants its result offset (and doesn't include the original primitive position). This offset can then be applied to the current filter's crop rect. (I've renamed the parameter "offset" in all cases to make this clear.) This makes the call sites in SkCanvas/SkGpuDevice responsible for applying the resulting offset to the primitive's position, which is actually a fairly small change.
This change also fixes SkTileImageFilter and SkOffsetImageFilter to correctly handle an input offset, which they weren't before. This required modifying the GM's, since they assumed the broken behaviour.
NOTE: this will require rebaselining the imagefiltersgraph test, since it has a new test case.
NOTE: this will "break" the Blink layout tests css3/filters/effect-reference-subregion-chained-hw.html and css3/filters/effect-reference-subregion-hw.html, but it actually makes them give correct results. It should be suppressed on the skia roll, and I'll rebaseline it.
R=reed@google.com
Review URL: https://codereview.chromium.org/112803004
git-svn-id: http://skia.googlecode.com/svn/trunk@12895 2bbb7eff-a529-9590-31e7-b0007b416f81
This works in a way that is similar to SkData.
SkMallocPixelRef::NewWithProc
Motivation: Chrome has a ETC1PixelRef which calls delete[] on the
pixles on destruction. There is no reason for them to almost
duplicate our class, when we can provide them a more flexible
class. Example use:
static void delete_uint8_proc(void* ptr, void*) {
delete[] static_cast<uint8_t>(ptr);
}
SkPixelRef* new_delete_pixref(const SkImageInfo& info,
SkColorTable* ctable) {
size_t rb = info.minRowBytes();
return SkMallocPixelRef::NewWithProc(
info, rb, ctable,
new uint8_t[info.getSafeSize(rb)],
delete_uint8_proc, NULL);
}
SkMallocPixelRef::NewWithData
Motivation: This allows up to eliminate SkDataPixelRef. We
modified SkImage_Raster to use MallocPixelRef rather than
SkDataPixlRef.
Also: Unit tests in tests/MallocPixelRefTest.
BUG=
R=reed@google.com
Review URL: https://codereview.chromium.org/106883006
git-svn-id: http://skia.googlecode.com/svn/trunk@12861 2bbb7eff-a529-9590-31e7-b0007b416f81
Motivation: We want to remove redundant classes from Skia. To
that end we want to remove SkImageRef and its subclasses and
replace their uses with SkDiscardablePixelRef +
SkDecodingImageGenerator. Since Android uses SkImageRef, we need
to make sure that SkDecodingImageGenerator allows all of the
settings that Android exposes in BitmapFactory.Options.
To that end, we have created an Options struct for the
SkDecodingImageGenerator which lets the client of the generator set
sample size, dithering, and bitmap config.
We have made the SkDecodingImageGenerator constructor private
and replaced the SkDecodingImageGenerator::Install functions
with a SkDecodingImageGenerator::Create functions (one for
SkData and one for SkStream) which now take a
SkDecodingImageGenerator::Options struct.
Also added a ImageDecoderOptions test which loops through a list
of sets of options and tries them on a set of 5 small encoded
images.
Also updated several users of SkDecodingImageGenerator::Install to
follow new call signature - gm/factory.cpp, LazyDecodeBitmap.cpp,
and PictureTest.cpp, CachedDecodingPixelRefTest.cpp.
We also added a new ImprovedBitmapFactory Test which simulates the
exact function that Android will need to modify to use this,
installPixelRef() in BitmapFactory.
R=reed@google.com, scroggo@google.com
Committed: https://code.google.com/p/skia/source/detail?r=12744
Review URL: https://codereview.chromium.org/93703004
git-svn-id: http://skia.googlecode.com/svn/trunk@12855 2bbb7eff-a529-9590-31e7-b0007b416f81
