| Commit message (Collapse) | Author | Age |
|
|
|
|
|
|
|
|
| |
#floats
BUG=skia:
BUG=skia:3592
Review URL: https://codereview.chromium.org/1047823002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The primary feature this delivers is SkNf and SkNd for arbitrary power-of-two N. Non-specialized types or types larger than 128 bits should now Just Work (and we can drop in a specialization to make them faster). Sk4s is now just a typedef for SkNf<4, SkScalar>; Sk4d is SkNf<4, double>, Sk2f SkNf<2, float>, etc.
This also makes implementing new specializations easier and more encapsulated. We're now using template specialization, which means the specialized versions don't have to leak out so much from SkNx_sse.h and SkNx_neon.h.
This design leaves us room to grow up, e.g to SkNf<8, SkScalar> == Sk8s, and to grown down too, to things like SkNi<8, uint16_t> == Sk8h.
To simplify things, I've stripped away most APIs (swizzles, casts, reinterpret_casts) that no one's using yet. I will happily add them back if they seem useful.
You shouldn't feel bad about using any of the typedef Sk4s, Sk4f, Sk4d, Sk2s, Sk2f, Sk2d, Sk4i, etc. Here's how you should feel:
- Sk4f, Sk4s, Sk2d: feel awesome
- Sk2f, Sk2s, Sk4d: feel pretty good
No public API changes.
TBR=reed@google.com
BUG=skia:3592
Review URL: https://codereview.chromium.org/1048593002
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add and test trunc(), which is what get() used to be before rounding.
Using trunc() is a ~40% speedup on our linear gradient bench.
#neon #floats
BUG=skia:3592
#n5
#n9
CQ_INCLUDE_TRYBOTS=client.skia.android:Test-Android-Nexus5-Adreno330-Arm7-Debug-Trybot;client.skia.android:Test-Android-Nexus9-TegraK1-Arm64-Release-Trybot
Review URL: https://codereview.chromium.org/1032243002
|
|
|
|
|
|
|
|
|
| |
NOPRESUBMIT=true
BUG=skia:
DOCS_PREVIEW= https://skia.org/?cl=1037793002
Review URL: https://codereview.chromium.org/1037793002
|
|
|
|
|
|
|
|
|
|
|
| |
There is no reason to require the 4 SkPMFloats (registers) to be adjacent.
The only potential win in loads and stores comes from the SkPMColors being adjacent.
Makes no difference to existing bench.
BUG=skia:
Review URL: https://codereview.chromium.org/1035583002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
SkMatrix::mapPts() using aacc/bbdd was always worse than using badc():
- On Intel, it was faster than exisiting swizzle, but badc() is 10% faster still (one pshufd instead of two).
- On ARM, existing swizzle < badc() < aacc()+bbdd(), even though aacc() then bbdd() is really a single vtrn instruction.
I will revert SkMatrix.cpp before submitting. Just thought you might like to look.
Will think more and try to gear up Instruments on ARM.
BUG=skia:
Review URL: https://codereview.chromium.org/1012573003
|
|
|
|
|
|
|
|
|
| |
This removes all the existing Sk4x swizzles and adds badc(), which is
both fast on all implementations and currently useful.
BUG=skia:
Review URL: https://codereview.chromium.org/997353005
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We don't have control over which way _mm_cvtps_epi32 rounds.
- This makes the SSE SkPMFloat rounding consistent with _neon and _none.
- Sk4f::cast<Sk4i>() is closer to (int)float's behavior. (Correct when >=0).
Add tests that would fail at head.
BUG=skia:
Review URL: https://codereview.chromium.org/1029163002
|
|
|
|
|
|
| |
BUG=skia:
Review URL: https://codereview.chromium.org/1024993002
|
|
|
|
|
|
|
|
|
|
| |
Tests pass on N7 + N9.
BUG=skia:
CQ_EXTRA_TRYBOTS=client.skia.compile:Build-Mac10.7-Clang-Arm7-Debug-iOS-Trybot,Build-Ubuntu-GCC-Arm64-Release-Android-Trybot
Review URL: https://codereview.chromium.org/1027753003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The implementation is nearly identical to Sk2f, with these changes:
- float32x2_t -> float64x2_t
- vfoo -> vfooq
- one extra Newton's method step in sqrt().
Also, generally fix NEON detection to be defined(SK_ARM_HAS_NEON).
SK_ARM_HAS_NEON is not being set on ARM64 bots right now (nor does the compiler
seem to set __ARM_NEON__), so this CL fixes everything up.
BUG=skia:
Committed: https://skia.googlesource.com/skia/+/e57b5cab261a243dcbefa74c91c896c28959bf09
CQ_EXTRA_TRYBOTS=client.skia.compile:Build-Mac10.7-Clang-Arm7-Debug-iOS-Trybot,Build-Ubuntu-GCC-Arm64-Release-Android-Trybot
Review URL: https://codereview.chromium.org/1020963002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Adds an SSE2 version of the Color32A_D565 function, to replace
the existing SSE4 version. Also does some minor cleanup.
Performance improvement in the following Skia benchmarks.
Measured on Atom Silvermont:
Xfermode_SrcOver - x3
luma_colorfilter_large - x4.6
luma_colorfilter_small - x2
tablebench - ~15%
chart_bw - ~10%
Measured on Corei7 Haswell:
luma_colorfilter_large running SSE2 - x2
luma_colorfilter_large running SSE4 - x2.3
Also improves performance in WPS Office application and 2D subtest of 0xbenchmark on Android.
Signed-off-by: Henrik Smiding <henrik.smiding@intel.com>
Review URL: https://codereview.chromium.org/923523002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
https://codereview.chromium.org/1020963002/)
Reason for revert:
https://uberchromegw.corp.google.com/i/client.skia.compile/builders/Build-Mac10.7-Clang-Arm7-Debug-iOS/builds/2441/steps/build%20most/logs/stdio
https://uberchromegw.corp.google.com/i/client.skia.compile/builders/Build-Mac10.7-Clang-Arm7-Release-iOS/builds/2424/steps/build%20most/logs/stdio
https://uberchromegw.corp.google.com/i/client.skia.compile/builders/Build-Ubuntu-GCC-Arm64-Release-Android/builds/8/steps/build%20most/logs/stdio
Original issue's description:
> Specialize Sk2d for ARM64
>
> The implementation is nearly identical to Sk2f, with these changes:
> - float32x2_t -> float64x2_t
> - vfoo -> vfooq
> - one extra Newton's method step in sqrt().
>
> Also, generally fix NEON detection to be defined(SK_ARM_HAS_NEON).
> SK_ARM_HAS_NEON is not being set on ARM64 bots right now (nor does the compiler
> seem to set __ARM_NEON__), so this CL fixes everything up.
>
> BUG=skia:
>
> Committed: https://skia.googlesource.com/skia/+/e57b5cab261a243dcbefa74c91c896c28959bf09
TBR=msarett@google.com,reed@google.com,mtklein@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Review URL: https://codereview.chromium.org/1028523003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The implementation is nearly identical to Sk2f, with these changes:
- float32x2_t -> float64x2_t
- vfoo -> vfooq
- one extra Newton's method step in sqrt().
Also, generally fix NEON detection to be defined(SK_ARM_HAS_NEON).
SK_ARM_HAS_NEON is not being set on ARM64 bots right now (nor does the compiler
seem to set __ARM_NEON__), so this CL fixes everything up.
BUG=skia:
Review URL: https://codereview.chromium.org/1020963002
|
|
|
|
|
|
|
|
|
|
|
| |
Also decreases the precision of Sk4f::rsqrt() for speed, keeping Sk4f::sqrt() the same:
instead of doing two estimation steps in rsqrt(), do one there and one more in sqrt().
Tests pass on my Nexus 7. float64x2_t is still a TODO for when I get a hold of a Nexus 9.
BUG=skia:
Review URL: https://codereview.chromium.org/1018423003
|
|
|
|
|
|
|
|
| |
This adds an API, an SSE impl, a portable impl, and some tests for Sk2f/Sk2d/Sk2s.
BUG=skia:
Review URL: https://codereview.chromium.org/1025463002
|
|
|
|
|
|
|
|
|
|
| |
No real changes here, just moving files around:
- move impl files into src/opts
- rename _portable _none
BUG=skia:
Review URL: https://codereview.chromium.org/1021713004
|
|
|
|
|
|
| |
BUG=skia:
Review URL: https://codereview.chromium.org/1021583002
|
|
|
|
| |
Review URL: https://codereview.chromium.org/1020563002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A store/load pair like this is a redundant no-op:
store simd_register_a, memory_address
load memory_address, simd_register_a
Everyone seems to be good at removing those when using SSE, but GCC and Clang
are pretty terrible at this for NEON. We end up issuing both redundant
commands, usually to and from the stack. That's slow. Let's not do that.
This CL unions in the native SIMD register type into SkPMFloat, so that we can
assign to and from it directly, which is generating a lot better NEON code. On
my Nexus 5, the benchmarks improve from 36ns to 23ns.
SSE is just as fast either way, but I paralleled the NEON code for consistency.
It's a little terser. And because it needed the platform headers anyway, I
moved all includes into SkPMFloat.h, again only for consistency.
I'd union in Sk4f too to make its conversion methods a little clearer,
but MSVC won't let me (it has a copy constructor... they're apparently not up
to speed with C++11 unrestricted unions).
BUG=skia:
Review URL: https://codereview.chromium.org/1015083004
|
|
|
|
|
|
|
|
|
| |
clone (+rebase) of https://codereview.chromium.org/1009183002/
BUG=skia:
TBR=scroggo@google.com
Review URL: https://codereview.chromium.org/1014533004
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Clamping 4 at a time is now about 15% faster than 1 at a time with SSSE3.
Clamping 4 at a time is now about 20% faster with SSE2,
and this applies to non-clamping too (we still just clamp there).
In all cases, 4 at a time is never worse than 1 at a time,
and not clamping is never slower than clamping.
Here's all the bench results, with the numbers for portable code as a fun point
of reference:
SSSE3:
maxrss loops min median mean max stddev samples config bench
10M 2291 4.66ns 4.66ns 4.66ns 4.68ns 0% ▆█▁▁▁▇▁▇▁▃ nonrendering SkPMFloat_get_1x
10M 2040 5.29ns 5.3ns 5.3ns 5.32ns 0% ▃▆▃▃▁▁▆▃▃█ nonrendering SkPMFloat_clamp_1x
10M 7175 4.62ns 4.62ns 4.62ns 4.63ns 0% ▁▄▃████▃▄▇ nonrendering SkPMFloat_get_4x
10M 5801 4.89ns 4.89ns 4.89ns 4.91ns 0% █▂▄▃▁▃▄█▁▁ nonrendering SkPMFloat_clamp_4x
SSE2:
maxrss loops min median mean max stddev samples config bench
10M 1601 6.02ns 6.05ns 6.04ns 6.08ns 0% █▅▄▅▄▂▁▂▂▂ nonrendering SkPMFloat_get_1x
10M 2918 6.05ns 6.06ns 6.05ns 6.06ns 0% ▂▇▁▇▇▁▇█▇▂ nonrendering SkPMFloat_clamp_1x
10M 3569 5.43ns 5.45ns 5.44ns 5.45ns 0% ▄█▂██▇▁▁▇▇ nonrendering SkPMFloat_get_4x
10M 4168 5.43ns 5.43ns 5.43ns 5.44ns 0% █▄▇▁▇▄▁▁▁▁ nonrendering SkPMFloat_clamp_4x
Portable:
maxrss loops min median mean max stddev samples config bench
10M 500 27.8ns 28.1ns 28ns 28.2ns 0% ▃█▆▃▇▃▆▁▇▂ nonrendering SkPMFloat_get_1x
10M 770 40.1ns 40.2ns 40.2ns 40.3ns 0% ▅▁▃▂▆▄█▂▅▂ nonrendering SkPMFloat_clamp_1x
10M 1269 28.4ns 28.8ns 29.1ns 32.7ns 4% ▂▂▂█▂▁▁▂▁▁ nonrendering SkPMFloat_get_4x
10M 1439 40.2ns 40.4ns 40.4ns 40.5ns 0% ▆▆▆█▁▆▅█▅▆ nonrendering SkPMFloat_clamp_4x
SkPMFloat_neon.h is still one big TODO as far as 4-at-a-time APIs go.
BUG=skia:
Review URL: https://codereview.chromium.org/982123002
|
|
|
|
|
|
|
|
|
|
|
| |
Instead of set(SkPMColor), add a constructor SkPMFloat(SkPMColor).
Replace setA(), setR(), etc. with a 4 float constructor.
And, promise to stick to SkPMColor order.
BUG=skia:
Review URL: https://codereview.chromium.org/977773002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
With SSSE3, we can use the Swiss Army Knife byte shuffler pshufb,
a.k.a. _mm_shuffle_epi8(), to jump directly between 32 and 128 bits.
In microbench isolation, this looks like an additional 10-15% speedup:
SkPMFloat_get: 2.35ns -> 1.98ns
SkPMFloat_clamp: 2.35ns -> 2.18ns
Before this CL, get() and clamp() were identical code. The _get benchmark improves because both set() and get() become faster; the _clamp benchmark shows the improvement from set() getting faster with clamp() staying the same.
BUG=skia:
Review URL: https://codereview.chromium.org/976493002
|
|
|
|
|
|
|
|
|
|
|
| |
SSE rounds for free (that was a happy accident: they also have a truncating version).
NEON does not, nor obviously the portable code, so they add 0.5 before truncating.
NOPRESUBMIT=true
BUG=skia:
Review URL: https://codereview.chromium.org/974643002
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This pushes the cost of the *255 and *1/255 conversions onto only those code
paths that need it. We're not doing it any more efficiently than can be done
with Sk4f.
In microbenchmark isolation, this is about a 15% speedup.
BUG=skia:
NOPRESUBMIT=true
Review URL: https://codereview.chromium.org/973603002
|
|
|
|
|
|
| |
BUG=skia:
Review URL: https://codereview.chromium.org/960023002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We've always been using the portable ColorRect32, so we don't need the
ColorRectProc plumbing.
Furthermore, ColorRect32 doesn't seem to be very important (we're only using
it in the opaque case, which our row-by-row procs already specialize for).
Remove that too.
If we find we want specialization for really narrow rects again, let's put it in
blitRect() directly. It's pretty unlikely we're going to get platform-specific
speedup for blits to non-contiguous memory.
My local SKP comparison is +- 3%... most neutral I've ever seen.
BUG=skia:
Review URL: https://codereview.chromium.org/959873002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Removes the disabled SSE2 optimization of ColorRect32 and deletes
the two files containing the code.
Measured on both Core Haswell and Atom Silvermont, and only got
some miniscule improvement compared to the default implementation.
Also tried to write a new, ultimate, version of this optimization,
but only got ~5% improvement on ColorRect32-heavy tests.
Signed-off-by: Henrik Smiding <henrik.smiding@intel.com>
Review URL: https://codereview.chromium.org/957433002
|
|
|
|
|
|
|
|
|
|
|
|
| |
BUG=skia:
Committed: https://skia.googlesource.com/skia/+/50d2b3114b3e59dc84811881591bf25b2c1ecb9f
CQ_EXTRA_TRYBOTS=client.skia.compile:Build-Ubuntu13.10-GCC4.8-Arm7-Release-Android_Neon-Trybot
http://build.chromium.org/p/client.skia.compile/builders/Build-Ubuntu13.10-GCC4.8-Arm7-Release-Android_Neon/builds/2120/steps/build%20most/logs/stdio
Review URL: https://codereview.chromium.org/936633002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
https://codereview.chromium.org/936633002/)
Reason for revert:
http://build.chromium.org/p/client.skia.compile/builders/Build-Ubuntu13.10-GCC4.8-Arm7-Release-Android_Neon/builds/2120/steps/build%20most/logs/stdio
Original issue's description:
> Sketch SkPMFloat
>
> BUG=skia:
>
> Committed: https://skia.googlesource.com/skia/+/50d2b3114b3e59dc84811881591bf25b2c1ecb9f
TBR=reed@google.com,msarrett@google.com,mtklein@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Review URL: https://codereview.chromium.org/952453004
|
|
|
|
|
|
| |
BUG=skia:
Review URL: https://codereview.chromium.org/936633002
|
|
|
|
|
|
|
|
|
|
| |
Shouldn't call Fast Blur path(DoubleRowBoxBlur_NEON)
when kernelsize is 1. Or, uint16x8_t resultPixels will be overflow.
BUG=skia:2845
R=senorblanco@chromium.org
Review URL: https://codereview.chromium.org/587543003
|
|
|
|
|
|
| |
BUG=skia:
Review URL: https://codereview.chromium.org/916713003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
https://codereview.chromium.org/916563002/)
Reason for revert:
Going to punt on 16-bit float support for now. Can't figure out ARM 64.
Original issue's description:
> add dummy avx file so xcode will build
>
> BUG=skia:
>
> Committed: https://skia.googlesource.com/skia/+/a0921f2563701d54e4e022de99f2705f4ada8a6e
TBR=reed@google.com
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Review URL: https://codereview.chromium.org/912213002
|
|
|
|
|
|
| |
BUG=skia:
Review URL: https://codereview.chromium.org/916563002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Adds an SSE4.1 version of the Color32A_D565 function.
Performance improvement in the following benchmarks:
Xfermode_SrcOver - ~100%
luma_colorfilter_large - ~150%
luma_colorfilter_small - ~60%
tablebench - ~10%
chart_bw - ~10%
(Measured on a Atom Silvermont core)
Signed-off-by: Henrik Smiding <henrik.smiding@intel.com>
Review URL: https://codereview.chromium.org/892623002
|
|
|
|
|
|
|
|
|
|
| |
Removes the opaque-only versions of this function from the factory since
they will never be used. Opaque source colors are handled in
SkRGB16_Opaque_Blitter instead, which doesn't use the factory function.
Signed-off-by: Henrik Smiding <henrik.smiding@intel.com>
Review URL: https://codereview.chromium.org/901593002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
(patchset #1 id:1 of https://codereview.chromium.org/873553003/)
Reason for revert:
Reverted the wrong CL.
Original issue's description:
> Revert of SSE4 opaque blend using intrinsics instead of assembly. (patchset #16 id:300001 of https://codereview.chromium.org/874863002/)
>
> Reason for revert:
> This causes a bug on the 'hittestpath' GM on MacMini 4,1
>
> See:
>
> https://gold.skia.org/#/triage/hittestpath?head=0
>
> for details.
>
> Original issue's description:
> > SSE4 opaque blend using intrinsics instead of assembly.
> >
> > Since we had such a hard time with the assembly versions of this blit (to the
> > point that we have them completely disabled everywhere), I thought I'd take
> > a shot at writing a version of the blit using intrinsics.
> >
> > The key feature of SSE4 we're exploiting is that we can use ptest (_mm_test*)
> > to skip the blend when the 16 src pixels we consider each loop are all opaque
> > or all transparent. _mm_shuffle_epi8 from SSSE3 also lends a hand to extract
> > all those alphas.
> >
> > It's worth looking to see if we can backport this type of logic to SSE2 using
> > _mm_movemask_epi8, or up to 32 pixels at a time using AVX.
> >
> > My local performance testing doesn't show this to be an unambiguous win
> > (there are probably microbenchmarks and SKPs where we'd be better off just
> > powering through the blend rather than looking at alphas), but the potential
> > does seem tantalizing enough to let skiaperf vet it on the bots. (< 1.0x is a win.)
> >
> > DM says it draws pixel perfect compare to the old code.
> >
> > Microbenchmarks:
> > bitmap_RGBA_8888_A_source_stripes_two 14us -> 14.4us 1.03x
> > bitmap_RGBA_8888_A_source_stripes_three 14.3us -> 14.5us 1.01x
> > bitmap_RGBA_8888_scale_bilerp 61.9us -> 62.2us 1.01x
> > bitmap_RGBA_8888_update_volatile_scale_rotate_bilerp 102us -> 101us 0.99x
> > bitmap_RGBA_8888_scale_rotate_bilerp 103us -> 101us 0.99x
> > bitmap_RGBA_8888_scale 18.4us -> 18.2us 0.99x
> > bitmap_RGBA_8888_A_scale_rotate_bicubic 71us -> 70us 0.99x
> > bitmap_RGBA_8888_update_scale_rotate_bilerp 103us -> 101us 0.99x
> > bitmap_RGBA_8888_A_scale_rotate_bilerp 112us -> 109us 0.98x
> > bitmap_RGBA_8888_update_volatile 5.72us -> 5.58us 0.98x
> > bitmap_RGBA_8888 5.73us -> 5.58us 0.97x
> > bitmap_RGBA_8888_update 5.78us -> 5.6us 0.97x
> > bitmap_RGBA_8888_A_scale_bilerp 70.7us -> 68us 0.96x
> > bitmap_RGBA_8888_A_scale_bicubic 23.7us -> 21.8us 0.92x
> > bitmap_RGBA_8888_A 13.9us -> 10.9us 0.78x
> > bitmap_RGBA_8888_A_source_opaque 14us -> 6.29us 0.45x
> > bitmap_RGBA_8888_A_source_transparent 14us -> 3.65us 0.26x
> >
> > Running over our ~70 SKP web page captures, this looks like we spend 0.7x
> > the time in S32A_Opaque_BlitRow compared to the SSE2 version, which should
> > be a decent predictor of real-world impact.
> >
> > BUG=chromium:399842
> >
> > Committed: https://skia.googlesource.com/skia/+/04bc91b972417038fecfa87c484771eac2b9b785
> >
> > CQ_EXTRA_TRYBOTS=client.skia:Test-Mac10.6-MacMini4.1-GeForce320M-x86_64-Release-Trybot
> >
> > Committed: https://skia.googlesource.com/skia/+/6dbfb21a6c88af6d94e8c823c3ad559f1a41b493
>
> TBR=henrik.smiding@intel.com,mtklein@google.com,herb@google.com,reed@google.com,thakis@chromium.org,mtklein@chromium.org
> NOPRESUBMIT=true
> NOTREECHECKS=true
> NOTRY=true
> BUG=chromium:399842
>
> Committed: https://skia.googlesource.com/skia/+/4988891a1173cd405bf1c1dd3a3668c451f45e4c
TBR=henrik.smiding@intel.com,mtklein@google.com,herb@google.com,reed@google.com,thakis@chromium.org,mtklein@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=chromium:399842
Review URL: https://codereview.chromium.org/894083002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
#16 id:300001 of https://codereview.chromium.org/874863002/)
Reason for revert:
This causes a bug on the 'hittestpath' GM on MacMini 4,1
See:
https://gold.skia.org/#/triage/hittestpath?head=0
for details.
Original issue's description:
> SSE4 opaque blend using intrinsics instead of assembly.
>
> Since we had such a hard time with the assembly versions of this blit (to the
> point that we have them completely disabled everywhere), I thought I'd take
> a shot at writing a version of the blit using intrinsics.
>
> The key feature of SSE4 we're exploiting is that we can use ptest (_mm_test*)
> to skip the blend when the 16 src pixels we consider each loop are all opaque
> or all transparent. _mm_shuffle_epi8 from SSSE3 also lends a hand to extract
> all those alphas.
>
> It's worth looking to see if we can backport this type of logic to SSE2 using
> _mm_movemask_epi8, or up to 32 pixels at a time using AVX.
>
> My local performance testing doesn't show this to be an unambiguous win
> (there are probably microbenchmarks and SKPs where we'd be better off just
> powering through the blend rather than looking at alphas), but the potential
> does seem tantalizing enough to let skiaperf vet it on the bots. (< 1.0x is a win.)
>
> DM says it draws pixel perfect compare to the old code.
>
> Microbenchmarks:
> bitmap_RGBA_8888_A_source_stripes_two 14us -> 14.4us 1.03x
> bitmap_RGBA_8888_A_source_stripes_three 14.3us -> 14.5us 1.01x
> bitmap_RGBA_8888_scale_bilerp 61.9us -> 62.2us 1.01x
> bitmap_RGBA_8888_update_volatile_scale_rotate_bilerp 102us -> 101us 0.99x
> bitmap_RGBA_8888_scale_rotate_bilerp 103us -> 101us 0.99x
> bitmap_RGBA_8888_scale 18.4us -> 18.2us 0.99x
> bitmap_RGBA_8888_A_scale_rotate_bicubic 71us -> 70us 0.99x
> bitmap_RGBA_8888_update_scale_rotate_bilerp 103us -> 101us 0.99x
> bitmap_RGBA_8888_A_scale_rotate_bilerp 112us -> 109us 0.98x
> bitmap_RGBA_8888_update_volatile 5.72us -> 5.58us 0.98x
> bitmap_RGBA_8888 5.73us -> 5.58us 0.97x
> bitmap_RGBA_8888_update 5.78us -> 5.6us 0.97x
> bitmap_RGBA_8888_A_scale_bilerp 70.7us -> 68us 0.96x
> bitmap_RGBA_8888_A_scale_bicubic 23.7us -> 21.8us 0.92x
> bitmap_RGBA_8888_A 13.9us -> 10.9us 0.78x
> bitmap_RGBA_8888_A_source_opaque 14us -> 6.29us 0.45x
> bitmap_RGBA_8888_A_source_transparent 14us -> 3.65us 0.26x
>
> Running over our ~70 SKP web page captures, this looks like we spend 0.7x
> the time in S32A_Opaque_BlitRow compared to the SSE2 version, which should
> be a decent predictor of real-world impact.
>
> BUG=chromium:399842
>
> Committed: https://skia.googlesource.com/skia/+/04bc91b972417038fecfa87c484771eac2b9b785
>
> CQ_EXTRA_TRYBOTS=client.skia:Test-Mac10.6-MacMini4.1-GeForce320M-x86_64-Release-Trybot
>
> Committed: https://skia.googlesource.com/skia/+/6dbfb21a6c88af6d94e8c823c3ad559f1a41b493
TBR=henrik.smiding@intel.com,mtklein@google.com,herb@google.com,reed@google.com,thakis@chromium.org,mtklein@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=chromium:399842
Review URL: https://codereview.chromium.org/873553003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Backport optimization from https://codereview.chromium.org/874863002/.
Microbenchmarks data compared to previous SSE2 implementation:
bitmap_BGRA_8888_A_source_stripes_three 7.52us -> 8.67us 1.15x
bitmap_BGRA_8888_A_source_stripes_two 7.48us -> 8.56us 1.15x
bitmap_BGRA_8888_update_scale_rotate_bilerp 63.4us -> 64us 1.01x
bitmap_BGRA_8888_update_volatile 3.31us -> 3.33us 1.01x
bitmap_BGRA_8888_scale 11.1us -> 11.2us 1x
bitmap_BGRA_8888_scale_bilerp 35.8us -> 35.9us 1x
bitmap_BGRA_8888 3.33us -> 3.33us 1x
bitmap_BGRA_8888_A_scale_rotate_bicubic 66.7us -> 66.5us 1x
bitmap_BGRA_8888_update_volatile_scale_rotate_bilerp 65.1us -> 64us 0.98x
bitmap_BGRA_8888_scale_rotate_bilerp 65.1us -> 64us 0.98x
bitmap_BGRA_8888_A_scale_bicubic 30.6us -> 29.9us 0.98x
bitmap_BGRA_8888_A_scale_bilerp 42.7us -> 41.4us 0.97x
bitmap_BGRA_8888_A_scale_rotate_bilerp 71us -> 67.7us 0.95x
bitmap_BGRA_8888_A 7.44us -> 5.7us 0.77x
bitmap_BGRA_8888_A_source_opaque 7.46us -> 3.72us 0.5x
bitmap_BGRA_8888_A_source_transparent 7.46us -> 1.96us 0.26x
BUG=skia:
Review URL: https://codereview.chromium.org/886403002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This includes blend32_16_row neon implementation
for aarch32 and aarch64.
For performance,
blend32_16_row is called in following tests in nanobench.
- Xfermode_SrcOver
- tablebench
- rotated_rects_bw_alternating_transparent_and_opaque_srcover
- rotated_rects_bw_changing_transparent_srcover
- rotated_rects_bw_same_transparent_srcover
- luma_colorfilter_large
- luma_colorfilter_small
- chart_bw
I can see perf increase in following two tests, especially. For others, looks
similar.
For each, I tried to run two times.
1) Xfermode_SrcOver
<org>
- D/skia ( 2000): 3M 57 17.3µs 17.4µs 17.4µs 17.7µs 1%
█▃▂▃▂▂▂▁▃▂ 565 Xfermode_SrcOver
- D/skia ( 1915): 3M 70 13.5µs 16.9µs 16.7µs 18.8µs 9%
▆█▄▅█▁▅▅▆▄ 565 Xfermode_SrcOver
<new>
- D/skia ( 2000): 3M 8 11.6µs 11.8µs 12.1µs 14.4µs 7%
▃█▁▁▂▁▁▁▂▂ 565 Xfermode_SrcOver
- D/skia ( 2004): 3M 62 10.3µs 12.9µs 13µs 15.2µs 11%
█▅▅▆▁▅▅▅▇▃ 565 Xfermode_SrcOver
2)
luma_colorfilter_large
<org>
- D/skia ( 2000): 159M 8 136µs 136µs 136µs 139µs 1%
█▃▁▂▁▁▁▁▁▁ 565 luma_colorfilter_large
- D/skia ( 1915): 158M 2 135µs 177µs 182µs 269µs 22%
▆▃█▁▁▃▃▃▃▃ 565 luma_colorfilter_large
<new>
- D/skia ( 2000): 157M 5 84.2µs 85.3µs 87.5µs 110µs 9%
█▁▂▁▁▁▁▁▁▁ 565 luma_colorfilter_large
- D/skia ( 2004): 159M 6 84.7µs 110µs 112µs 144µs 18%
█▄▇▁▁▄▃▄▄▆ 565 luma_colorfilter_large
Review URL: https://codereview.chromium.org/847363002
|
|
|
|
|
|
|
|
| |
NOTREECHECKS=true
BUG=chromium:399842
Review URL: https://codereview.chromium.org/881253003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Since we had such a hard time with the assembly versions of this blit (to the
point that we have them completely disabled everywhere), I thought I'd take
a shot at writing a version of the blit using intrinsics.
The key feature of SSE4 we're exploiting is that we can use ptest (_mm_test*)
to skip the blend when the 16 src pixels we consider each loop are all opaque
or all transparent. _mm_shuffle_epi8 from SSSE3 also lends a hand to extract
all those alphas.
It's worth looking to see if we can backport this type of logic to SSE2 using
_mm_movemask_epi8, or up to 32 pixels at a time using AVX.
My local performance testing doesn't show this to be an unambiguous win
(there are probably microbenchmarks and SKPs where we'd be better off just
powering through the blend rather than looking at alphas), but the potential
does seem tantalizing enough to let skiaperf vet it on the bots. (< 1.0x is a win.)
DM says it draws pixel perfect compare to the old code.
Microbenchmarks:
bitmap_RGBA_8888_A_source_stripes_two 14us -> 14.4us 1.03x
bitmap_RGBA_8888_A_source_stripes_three 14.3us -> 14.5us 1.01x
bitmap_RGBA_8888_scale_bilerp 61.9us -> 62.2us 1.01x
bitmap_RGBA_8888_update_volatile_scale_rotate_bilerp 102us -> 101us 0.99x
bitmap_RGBA_8888_scale_rotate_bilerp 103us -> 101us 0.99x
bitmap_RGBA_8888_scale 18.4us -> 18.2us 0.99x
bitmap_RGBA_8888_A_scale_rotate_bicubic 71us -> 70us 0.99x
bitmap_RGBA_8888_update_scale_rotate_bilerp 103us -> 101us 0.99x
bitmap_RGBA_8888_A_scale_rotate_bilerp 112us -> 109us 0.98x
bitmap_RGBA_8888_update_volatile 5.72us -> 5.58us 0.98x
bitmap_RGBA_8888 5.73us -> 5.58us 0.97x
bitmap_RGBA_8888_update 5.78us -> 5.6us 0.97x
bitmap_RGBA_8888_A_scale_bilerp 70.7us -> 68us 0.96x
bitmap_RGBA_8888_A_scale_bicubic 23.7us -> 21.8us 0.92x
bitmap_RGBA_8888_A 13.9us -> 10.9us 0.78x
bitmap_RGBA_8888_A_source_opaque 14us -> 6.29us 0.45x
bitmap_RGBA_8888_A_source_transparent 14us -> 3.65us 0.26x
Running over our ~70 SKP web page captures, this looks like we spend 0.7x
the time in S32A_Opaque_BlitRow compared to the SSE2 version, which should
be a decent predictor of real-world impact.
BUG=chromium:399842
Committed: https://skia.googlesource.com/skia/+/04bc91b972417038fecfa87c484771eac2b9b785
CQ_EXTRA_TRYBOTS=client.skia:Test-Mac10.6-MacMini4.1-GeForce320M-x86_64-Release-Trybot
Review URL: https://codereview.chromium.org/874863002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
#14 id:260001 of https://codereview.chromium.org/874863002/)
Reason for revert:
This kills Mac 10.6 bots.
FAILED: c++ -MMD -MF obj/src/opts/opts_sse4.SkBlitRow_opts_SSE4.o.d -DSK_INTERNAL -DSK_GAMMA_SRGB -DSK_GAMMA_APPLY_TO_A8 -DSK_SCALAR_TO_FLOAT_EXCLUDED -DSK_ALLOW_STATIC_GLOBAL_INITIALIZERS=1 -DSK_SUPPORT_GPU=1 -DSK_SUPPORT_OPENCL=0 -DSK_FORCE_DISTANCE_FIELD_TEXT=0 -DSK_BUILD_FOR_MAC -DSK_CRASH_HANDLER -DSK_DEVELOPER=1 -I../../src/core -I../../src/utils -I../../include/c -I../../include/config -I../../include/core -I../../include/pathops -I../../include/pipe -I../../include/utils/mac -I../../include/effects -O0 -gdwarf-2 -mmacosx-version-min=10.6 -arch x86_64 -mssse3 -Wall -Wextra -Winit-self -Wpointer-arith -Wsign-compare -Wno-unused-parameter -Wno-invalid-offsetof -msse4.1 -c ../../src/opts/SkBlitRow_opts_SSE4.cpp -o obj/src/opts/opts_sse4.SkBlitRow_opts_SSE4.o
../../src/opts/SkBlitRow_opts_SSE4.cpp:15:27: warning: x86intrin.h: No such file or directory
../../src/opts/SkBlitRow_opts_SSE4.cpp: In function 'void S32A_Opaque_BlitRow32_SSE4(SkPMColor*, const SkPMColor*, int, U8CPU)':
../../src/opts/SkBlitRow_opts_SSE4.cpp:40: error: '_mm_testz_si128' was not declared in this scope
../../src/opts/SkBlitRow_opts_SSE4.cpp:45: error: '_mm_testc_si128' was not declared in this scope
Original issue's description:
> SSE4 opaque blend using intrinsics instead of assembly.
>
> Since we had such a hard time with the assembly versions of this blit (to the
> point that we have them completely disabled everywhere), I thought I'd take
> a shot at writing a version of the blit using intrinsics.
>
> The key feature of SSE4 we're exploiting is that we can use ptest (_mm_test*)
> to skip the blend when the 16 src pixels we consider each loop are all opaque
> or all transparent. _mm_shuffle_epi8 from SSSE3 also lends a hand to extract
> all those alphas.
>
> It's worth looking to see if we can backport this type of logic to SSE2 using
> _mm_movemask_epi8, or up to 32 pixels at a time using AVX.
>
> My local performance testing doesn't show this to be an unambiguous win
> (there are probably microbenchmarks and SKPs where we'd be better off just
> powering through the blend rather than looking at alphas), but the potential
> does seem tantalizing enough to let skiaperf vet it on the bots. (< 1.0x is a win.)
>
> DM says it draws pixel perfect compare to the old code.
>
> Microbenchmarks:
> bitmap_RGBA_8888_A_source_stripes_two 14us -> 14.4us 1.03x
> bitmap_RGBA_8888_A_source_stripes_three 14.3us -> 14.5us 1.01x
> bitmap_RGBA_8888_scale_bilerp 61.9us -> 62.2us 1.01x
> bitmap_RGBA_8888_update_volatile_scale_rotate_bilerp 102us -> 101us 0.99x
> bitmap_RGBA_8888_scale_rotate_bilerp 103us -> 101us 0.99x
> bitmap_RGBA_8888_scale 18.4us -> 18.2us 0.99x
> bitmap_RGBA_8888_A_scale_rotate_bicubic 71us -> 70us 0.99x
> bitmap_RGBA_8888_update_scale_rotate_bilerp 103us -> 101us 0.99x
> bitmap_RGBA_8888_A_scale_rotate_bilerp 112us -> 109us 0.98x
> bitmap_RGBA_8888_update_volatile 5.72us -> 5.58us 0.98x
> bitmap_RGBA_8888 5.73us -> 5.58us 0.97x
> bitmap_RGBA_8888_update 5.78us -> 5.6us 0.97x
> bitmap_RGBA_8888_A_scale_bilerp 70.7us -> 68us 0.96x
> bitmap_RGBA_8888_A_scale_bicubic 23.7us -> 21.8us 0.92x
> bitmap_RGBA_8888_A 13.9us -> 10.9us 0.78x
> bitmap_RGBA_8888_A_source_opaque 14us -> 6.29us 0.45x
> bitmap_RGBA_8888_A_source_transparent 14us -> 3.65us 0.26x
>
> Running over our ~70 SKP web page captures, this looks like we spend 0.7x
> the time in S32A_Opaque_BlitRow compared to the SSE2 version, which should
> be a decent predictor of real-world impact.
>
> BUG=chromium:399842
>
> Committed: https://skia.googlesource.com/skia/+/04bc91b972417038fecfa87c484771eac2b9b785
TBR=henrik.smiding@intel.com,mtklein@google.com,herb@google.com,reed@google.com,thakis@chromium.org,mtklein@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=chromium:399842
Review URL: https://codereview.chromium.org/874033004
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Since we had such a hard time with the assembly versions of this blit (to the
point that we have them completely disabled everywhere), I thought I'd take
a shot at writing a version of the blit using intrinsics.
The key feature of SSE4 we're exploiting is that we can use ptest (_mm_test*)
to skip the blend when the 16 src pixels we consider each loop are all opaque
or all transparent. _mm_shuffle_epi8 from SSSE3 also lends a hand to extract
all those alphas.
It's worth looking to see if we can backport this type of logic to SSE2 using
_mm_movemask_epi8, or up to 32 pixels at a time using AVX.
My local performance testing doesn't show this to be an unambiguous win
(there are probably microbenchmarks and SKPs where we'd be better off just
powering through the blend rather than looking at alphas), but the potential
does seem tantalizing enough to let skiaperf vet it on the bots. (< 1.0x is a win.)
DM says it draws pixel perfect compare to the old code.
Microbenchmarks:
bitmap_RGBA_8888_A_source_stripes_two 14us -> 14.4us 1.03x
bitmap_RGBA_8888_A_source_stripes_three 14.3us -> 14.5us 1.01x
bitmap_RGBA_8888_scale_bilerp 61.9us -> 62.2us 1.01x
bitmap_RGBA_8888_update_volatile_scale_rotate_bilerp 102us -> 101us 0.99x
bitmap_RGBA_8888_scale_rotate_bilerp 103us -> 101us 0.99x
bitmap_RGBA_8888_scale 18.4us -> 18.2us 0.99x
bitmap_RGBA_8888_A_scale_rotate_bicubic 71us -> 70us 0.99x
bitmap_RGBA_8888_update_scale_rotate_bilerp 103us -> 101us 0.99x
bitmap_RGBA_8888_A_scale_rotate_bilerp 112us -> 109us 0.98x
bitmap_RGBA_8888_update_volatile 5.72us -> 5.58us 0.98x
bitmap_RGBA_8888 5.73us -> 5.58us 0.97x
bitmap_RGBA_8888_update 5.78us -> 5.6us 0.97x
bitmap_RGBA_8888_A_scale_bilerp 70.7us -> 68us 0.96x
bitmap_RGBA_8888_A_scale_bicubic 23.7us -> 21.8us 0.92x
bitmap_RGBA_8888_A 13.9us -> 10.9us 0.78x
bitmap_RGBA_8888_A_source_opaque 14us -> 6.29us 0.45x
bitmap_RGBA_8888_A_source_transparent 14us -> 3.65us 0.26x
Running over our ~70 SKP web page captures, this looks like we spend 0.7x
the time in S32A_Opaque_BlitRow compared to the SSE2 version, which should
be a decent predictor of real-world impact.
BUG=chromium:399842
Review URL: https://codereview.chromium.org/874863002
|
|
|
|
|
|
|
|
|
|
| |
It's very common (universal?) that alpha is the top byte.
You'd hope the compiler would remove the left shift then,
but I've seen Clang just do a dumb left shift of zero. :(
BUG=skia:
Review URL: https://codereview.chromium.org/872243003
|
|
|
|
|
|
| |
BUG=skia:
Review URL: https://codereview.chromium.org/845303005
|
|
|
|
|
|
| |
BUG=skia:3302
Review URL: https://codereview.chromium.org/847443003
|
|
|
|
|
|
|
|
|
|
| |
Prevent these symbols from being exported from any library which
includes skia by marking them .hidden, as they are implementation
details of skia.
BUG=skia:3303
Review URL: https://codereview.chromium.org/831663004
|