This should be a drop-in replacement for most for-loops to make them run in parallel:
for (int i = 0; i < N; i++) { code... }
~~~>
sk_parallel_for(N, [&](int i) { code... });
This is just syntax sugar over SkTaskGroup to make this use case really easy to write.
There's no more overhead that we weren't already forced to add using an interface like batch(),
and no extra heap allocations.
I've replaced 3 uses of SkTaskGroup with sk_parallel_for:
1) My unit tests for SkOnce.
2) Cary's path fuzzer.
3) SkMultiPictureDraw.
Performance should be the same. Please compare left and right for readability. :)
BUG=skia:
No public API changes.
TBR=reed@google.com
Review URL: https://codereview.chromium.org/1184373003
Use the form SkDebugf("%s", arbitraryString) instead of
SkDebugf(arbitraryString).
Fixes the case where SkString::appendf-ing a string with "%%" and then
printing the string with SkDebugf would cause uninitialized read and
corrupted debug print.
ninja -C out/Debug tools && valgrind --leak-check=full
./out/Debug/render_pictures --config gpu -w q -r ...
...
==7307== Conditional jump or move depends on uninitialised value(s)
==7307== at 0x6908475: __printf_fp (printf_fp.c:1180)
==7307== by 0x6904267: vfprintf (vfprintf.c:1629)
==7307== by 0x6906E53: buffered_vfprintf (vfprintf.c:2313)
==7307== by 0x690188D: vfprintf (vfprintf.c:1316)
==7307== by 0x67E8F5: SkDebugf(char const*, ...) (SkDebug_stdio.cpp:18)
==7307== by 0x7983F1: GrContext::printCacheStats() const (GrTest.cpp:54)
==7307== by 0x408ECF: tool_main(int, char**) (render_pictures_main.cpp:480)
==7307== by 0x40913E: main (render_pictures_main.cpp:511)
==7307==
Budget: 2048 items 100663296 bytes
Entry Count: current 652 (651 budgeted, 0 wrapped, 297 locked, 638 scratch 32 0.000000ull), high 652
Entry Bytes: current 51087658 (budgeted 49826658, 49 0.000000ull, 1261000 unbudgeted) high 51087658
(observe "ull" instead of "% full")
(from mtklein)
This CL is not editing public API.
TBR=reed@google.com
Review URL: https://codereview.chromium.org/943453002
This fixes every case where virtual and SK_OVERRIDE were on the same line,
which should be the bulk of cases. We'll have to manually clean up the rest
over time unless I level up in regexes.
for f in (find . -type f); perl -p -i -e 's/virtual (.*)SK_OVERRIDE/\1SK_OVERRIDE/g' $f; end
BUG=skia:
Review URL: https://codereview.chromium.org/806653007
SkTaskGroup is like SkThreadPool except the threads stay in
one global pool. Each SkTaskGroup itself is tiny (4 bytes)
and its wait() method applies only to tasks add()ed to that
instance, not the whole thread pool.
This means we don't need to bring up new thread pools when
tests themselves want to use multithreading (e.g. pathops,
quilt). We just create a new SkTaskGroup and wait for that
to complete. This should be more efficient, and allow us
to expand where we use threads to really latency sensitive
places. E.g. we can probably now use these in nanobench
for CPU .skp rendering.
Now that all threads are sharing the same pool, I think we
can remove most of the custom mechanism pathops tests use
to control threading. They'll just ride on the global pool
with all other tests now.
This (temporarily?) removes the GPU multithreading feature
from DM, which we don't use.
On my desktop, DM runs a little faster (57s -> 55s) in
Debug, and a lot faster in Release (36s -> 24s). The bots
show speedups of similar proportions, cutting more than a
minute off the N4/Release and Win7/Debug runtimes.
BUG=skia:
Committed: https://skia.googlesource.com/skia/+/9c7207b5dc71dc5a96a2eb107d401133333d5b6fR=caryclark@google.com, bsalomon@google.com, bungeman@google.com, mtklein@google.com, reed@google.com
Author: mtklein@chromium.org
Review URL: https://codereview.chromium.org/531653002
Reason for revert:
Leaks, leaks, leaks.
Original issue's description:
> SkThreadPool ~~> SkTaskGroup
>
> SkTaskGroup is like SkThreadPool except the threads stay in
> one global pool. Each SkTaskGroup itself is tiny (4 bytes)
> and its wait() method applies only to tasks add()ed to that
> instance, not the whole thread pool.
>
> This means we don't need to bring up new thread pools when
> tests themselves want to use multithreading (e.g. pathops,
> quilt). We just create a new SkTaskGroup and wait for that
> to complete. This should be more efficient, and allow us
> to expand where we use threads to really latency sensitive
> places. E.g. we can probably now use these in nanobench
> for CPU .skp rendering.
>
> Now that all threads are sharing the same pool, I think we
> can remove most of the custom mechanism pathops tests use
> to control threading. They'll just ride on the global pool
> with all other tests now.
>
> This (temporarily?) removes the GPU multithreading feature
> from DM, which we don't use.
>
> On my desktop, DM runs a little faster (57s -> 55s) in
> Debug, and a lot faster in Release (36s -> 24s). The bots
> show speedups of similar proportions, cutting more than a
> minute off the N4/Release and Win7/Debug runtimes.
>
> BUG=skia:
>
> Committed: https://skia.googlesource.com/skia/+/9c7207b5dc71dc5a96a2eb107d401133333d5b6fR=caryclark@google.com, bsalomon@google.com, bungeman@google.com, reed@google.com, mtklein@chromium.orgTBR=bsalomon@google.com, bungeman@google.com, caryclark@google.com, mtklein@chromium.org, reed@google.com
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Author: mtklein@google.com
Review URL: https://codereview.chromium.org/533393002
SkTaskGroup is like SkThreadPool except the threads stay in
one global pool. Each SkTaskGroup itself is tiny (4 bytes)
and its wait() method applies only to tasks add()ed to that
instance, not the whole thread pool.
This means we don't need to bring up new thread pools when
tests themselves want to use multithreading (e.g. pathops,
quilt). We just create a new SkTaskGroup and wait for that
to complete. This should be more efficient, and allow us
to expand where we use threads to really latency sensitive
places. E.g. we can probably now use these in nanobench
for CPU .skp rendering.
Now that all threads are sharing the same pool, I think we
can remove most of the custom mechanism pathops tests use
to control threading. They'll just ride on the global pool
with all other tests now.
This (temporarily?) removes the GPU multithreading feature
from DM, which we don't use.
On my desktop, DM runs a little faster (57s -> 55s) in
Debug, and a lot faster in Release (36s -> 24s). The bots
show speedups of similar proportions, cutting more than a
minute off the N4/Release and Win7/Debug runtimes.
BUG=skia:
R=caryclark@google.com, bsalomon@google.com, bungeman@google.com, mtklein@google.com, reed@google.com
Author: mtklein@chromium.org
Review URL: https://codereview.chromium.org/531653002
This fixes all but one of those failures.
Major changes include:
- Replace angle indices with angle pointers. This was motivated by the need to add angles later but not renumber existing angles.
- Aggressive segment chase. When the winding is known on a segment, more aggressively passing that winding to adjacent segments allows fragmented data sets to succeed.
- Line segments with ends nearly the same are treated as coincident first.
- Transfer partial coincidence by observing that if segment A is partially coincident to B and C then B and C may be partially coincident.
TBR=reed
Author: caryclark@google.com
Review URL: https://codereview.chromium.org/272153002
Mike K: please sanity check Test.cpp and skia_test.cpp
Feel free to look at the rest, but I don't expect any in depth review of path ops innards.
Path Ops first iteration used QuickSort to order segments radiating from an intersection to compute the winding rule.
This revision uses a circular sort instead. Breaking out the circular sort into its own long-lived structure (SkOpAngle) allows doing less work and provides a home for caching additional sorting data.
The circle sort is more stable than the former sort, has a robust ordering and fewer exceptions. It finds unsortable ordering less often. It is less reliant on the initial curve tangent, using convex hulls instead whenever it can.
Additional debug validation makes sure that the computed structures are self-consistent. A new visualization tool helps verify that the angle ordering is correct.
The 70+M tests pass with this change on Windows, Mac, Linux 32 and Linux 64 in debug and release.
R=mtklein@google.com, reed@google.com
Author: caryclark@google.com
Review URL: https://codereview.chromium.org/131103009
git-svn-id: http://skia.googlecode.com/svn/trunk@14183 2bbb7eff-a529-9590-31e7-b0007b416f81
Modify line intersections to first
- match exact ends
- compute intersections
- match near ends
where the exact ends are preferred, then near matches, then
computed matches. This pulls matches towards existing end points
when possible, and keeps intersection distances consistent with
different line/line line/quad and line/cubic computations.
BUG=
Review URL: https://codereview.chromium.org/19183003
git-svn-id: http://skia.googlecode.com/svn/trunk@10073 2bbb7eff-a529-9590-31e7-b0007b416f81
This is a major change resulting from a minor
tweak. In the old code, the intersection point
of two curves was shared between them, but the
intersection points and end points of sorted edges was
computed directly from the intersection T value.
In this CL, both intersection points and sorted points
are the same, and intermediate control points are computed
to preserve their slope.
The sort itself has been completely rewritten to be more
robust and remove 'magic' checks, conditions that empirically
worked but couldn't be rationalized.
This CL was triggered by errors generated computing the clips
of SKP files. At this point, all 73M standard tests work and
at least the first troublesome SKPs work.
Review URL: https://codereview.chromium.org/15338003
git-svn-id: http://skia.googlecode.com/svn/trunk@9432 2bbb7eff-a529-9590-31e7-b0007b416f81
