About 9x faster than Murmur3 for long inputs.
Most of this is a mechanical change from SkChecksum::Murmur3(...) to SkOpts::hash(...).
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2208903002
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;master.client.skia.compile:Build-Ubuntu-GCC-x86_64-Release-CMake-Trybot,Build-Mac-Clang-x86_64-Release-CMake-Trybot
Review-Url: https://codereview.chromium.org/2208903002
Reason for revert:
Bot failures
Original issue's description:
> AVX 2 SrcOver blits: color32, blitmask.
>
> As a follow up to the SSE 4.1 CL, this should look pretty familiar.
>
> I've made some organizational changes around how we load, store, pack, and unpack data that I think makes things clearer and more orthogonal, and it'll make it easier to try out a pmaddubsw lerp. I have backported these changes to the SSE 4.1 code, and I hope that I can actually get a lot of this code templated for sharing between the two later.
>
> Perf changes (relative to SSE 4.1):
> Xfermode_SrcOver: 1650 -> 1180 (0.71x) // large opaque blit
> Xfermode_SrcOver_aa: 1794 -> 1653 (0.92x) // large opaque + small transparent
> text_16_AA_{FF,BK,WT}: 1.72 -> 1.59 (0.92x) // small opaque blit
> text_16_AA_88: 1.83 -> 1.77 (0.97x) // small transparent blit
>
> This should be a big throughout win, and a small latency win.
> This should all be pixel-exact to the previous SSE 4.1 code.
>
>
> GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1532613002
> CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;client.skia.compile:Build-Ubuntu-GCC-x86_64-Release-CMake-Trybot,Build-Mac10.9-Clang-x86_64-Release-CMake-Trybot
>
> Committed: https://skia.googlesource.com/skia/+/5d2117015eb271e09faf4a7ddd89093c9d618a36TBR=herb@google.com,mtklein@google.com,mtklein@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
Review URL: https://codereview.chromium.org/1632713002
As a follow up to the SSE 4.1 CL, this should look pretty familiar.
I've made some organizational changes around how we load, store, pack, and unpack data that I think makes things clearer and more orthogonal, and it'll make it easier to try out a pmaddubsw lerp. I have backported these changes to the SSE 4.1 code, and I hope that I can actually get a lot of this code templated for sharing between the two later.
Perf changes (relative to SSE 4.1):
Xfermode_SrcOver: 1650 -> 1180 (0.71x) // large opaque blit
Xfermode_SrcOver_aa: 1794 -> 1653 (0.92x) // large opaque + small transparent
text_16_AA_{FF,BK,WT}: 1.72 -> 1.59 (0.92x) // small opaque blit
text_16_AA_88: 1.83 -> 1.77 (0.97x) // small transparent blit
This should be a big throughout win, and a small latency win.
This should all be pixel-exact to the previous SSE 4.1 code.
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1532613002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;client.skia.compile:Build-Ubuntu-GCC-x86_64-Release-CMake-Trybot,Build-Mac10.9-Clang-x86_64-Release-CMake-Trybot
Review URL: https://codereview.chromium.org/1532613002
Xfermode_ColorDodge_aa 10.3ms -> 7.85ms 0.76x
Xfermode_SoftLight_aa 13.8ms -> 10.2ms 0.74x
Xfermode_ColorBurn_aa 10.7ms -> 7.82ms 0.73x
Xfermode_SoftLight 33.6ms -> 23.2ms 0.69x
Xfermode_ColorDodge 25ms -> 16.5ms 0.66x
Xfermode_ColorBurn 26.1ms -> 16.6ms 0.63x
Ought to be no pixel diffs:
https://gold.skia.org/search2?issue=1432903002&unt=true&query=source_type%3Dgm&master=false
Incidental stuff:
I made the SkNx(T) constructors implicit to make writing math expressions simpler.
This allows us to write expressions like
Sk4f v;
...
v = v*4;
rather than
Sk4f v;
...
v = v * Sk4f(4);
As written it only works when the constant is on the right-hand side,
so expressions like `(Sk4f(1) - da)` have to stay for now. I plan on
following up with a CL that lets those become `(1 - da)` too.
BUG=skia:4117
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Review URL: https://codereview.chromium.org/1432903002
CL (1 of 3) adds empty lists in our .gypi,
and builds the files in those empty lists with the appropriate flags.
CL (2 of 3) will have Chrome's GYP and GN files read these lists,
and build them with the appropriate flags.
CL (3 of 3) will add runtime detection and stub files to the lists
with empty Init_sse42(), Init_avx(), Init_avx2() methods.
After that, we should be able to use SSE 4.2, AVX, and AVX2 if desired.
Some motivation:
- SSE 4.2 adds some sweet string-oriented methods that
can help us write fast high quality 32-bit hashes.
- AVX is SSE doubled, e.g. 8 floats or two SkPMFloat at a time.
- AVX2 is SSE2 doubled, e.g. 8 pixels at a time.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1290423007
It's sort of pointless: all our clients that will have SSE2 at runtime have it
unconditionally at compile time, so the functions in namespace portable will
pick up the SSE2 code. The procs in SkOpts_sse2.o were just duplicate code.
A couple of the procs we had in _sse2.cpp can benefit slightly when compiled
with SSSE3. I've moved those to _ssse3.cpp. This should lead to small speedups
on platforms like Linux and Windows that have a baseline of SSE2.
Similarly, I've removed the call to Init_neon() when NEON is available globally... it's a no-op.
Renaming namespace portable to something clearer is TBD.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1294213002
Pretty vanilla translation. I cleaned up who calls whom a little.
Used to be utils -> opts -> utils, now it's just utils -> opts.
I may follow up with a pass over the NEON code for readability
and to clean up dead code.
This turns on NEON A8->R11EAC conversion for ARMv8.
Unit tests which now hit the NEON code still pass.
I can't find any related bench.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1273103002
Nothing too fancy.
Direction enums become enum classes so they don't get all confused. An
alternative is to create one single Direction enum that both blur and
morphology opts use.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1267343004
+268 -535 lines
I also rearranged the code a little bit to encapsulate itself better,
mostly replacing static helper functions with lambdas. This also
let me merge the SSE2 and SSE4.1 code paths.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1264103004
Renames Sk4pxXfermode.h to SkXfermode_opts.h,
and refactors it a tiny bit internally.
This moves xfermode optimization from being "compile-time everywhere but NEON"
to simply "runtime everywhere". I don't anticipate any effect on perf or
correctness.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1264543006
With this new arrangement, the benefits of inlining sk_memset16/32 have changed.
On x86, they're not significantly different, except for small N<=10 where the inlined code is significantly slower.
On ARMv7 with NEON, our custom code is still significantly faster for N>10 (up to 2x faster). For small N<=10 inlining is still significantly faster.
On ARMv7 without NEON, our custom code is still ridiculously faster (up to 10x) than inlining for N>10, though for small N<=10 inlining is still a little faster.
We were not using the NEON memset16 and memset32 procs on ARMv8. At first blush, that seems to be an oversight, but if so it's an extremely lucky one. The ARMv8 code generation for our memset16/32 procs is total garbage, leaving those methods ~8x slower than just inlining the memset, using the compiler's autovectorization.
So, no need to inline any more on x86, and still inline for N<=10 on ARMv7. Always inline for ARMv8.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1270573002
This doesn't really do anything yet. It's just the CPU detection code, skeleton new .cpp files, and a few little .gyp tweaks.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1255193002
Now that SK_SUPPORT_LEGACY_XFERMODES is unused, tons of code becomes dead.
Nothing is needed in opts/ anymore for x86.
We still do runtime NEON detection, which just duplicates Sk4pxXfermode.
TBR=reed@google.com
BUG=skia:
Review URL: https://codereview.chromium.org/1230023011
According to bench/MemsetBench.cpp, I've got them somewhere between 10% slower
and a percent or two faster than the old assembly.
BUG=skia:
CQ_EXTRA_TRYBOTS=client.skia.android:Test-Android-GCC-Nexus5-CPU-NEON-Arm7-Debug-Trybot
Review URL: https://codereview.chromium.org/1075003002
Step 1 of a zillion in the quest for NEON on iOS,
and step 1 of a different zillion in the Great Assembly Purge.
ios, arm, arm64, arm_v7, arm_v7_neon all build.
BUG=skia:
Review URL: https://codereview.chromium.org/1072063002
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
Reason for revert:
Going to punt on 16-bit float support for now. Can't figure out ARM 64.
Original issue's description:
> GYP groudwork for half-float opts support.
>
> This sets us up two new opts targets with the immediate goal of adding half-float (SkHalf.h) opts:
> - opts_neon_fp16: uses hardware support on most ARM chips with NEON to do 4 conversions at a time;
> - opts_avx: uses hardware support on Intel chips with AVX to do 8 conversions at a time.
>
> opts_avx will be a handy thing to have around later too, especially if we want to work with floats.
>
> This doesn't actually add any new source files to these libraries yet, so they're no-ops for now.
> I'll need to write a parallel change to Chrome's GN and GYPs before we can start adding sources.
>
> This also rolls GYP up to head, to get suppport for EnableEnhancedInstructionSet: '3' on Windows,
> which is how we turn on AVX there. There's no Mac-specific flag, so we use OTHER_CPLUSPLUSFLAGS.
>
> BUG=skia:
>
> TBR=reed@google.com
>
> Committed: https://skia.googlesource.com/skia/+/46b80833394d7919cadf2abf2b93802141dd21c5TBR=reed@google.com,mtklein@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Review URL: https://codereview.chromium.org/912223002
This sets us up two new opts targets with the immediate goal of adding half-float (SkHalf.h) opts:
- opts_neon_fp16: uses hardware support on most ARM chips with NEON to do 4 conversions at a time;
- opts_avx: uses hardware support on Intel chips with AVX to do 8 conversions at a time.
opts_avx will be a handy thing to have around later too, especially if we want to work with floats.
This doesn't actually add any new source files to these libraries yet, so they're no-ops for now.
I'll need to write a parallel change to Chrome's GN and GYPs before we can start adding sources.
This also rolls GYP up to head, to get suppport for EnableEnhancedInstructionSet: '3' on Windows,
which is how we turn on AVX there. There's no Mac-specific flag, so we use OTHER_CPLUSPLUSFLAGS.
BUG=skia:
TBR=reed@google.com
Review URL: https://codereview.chromium.org/915693002
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/+/4988891a1173cd405bf1c1dd3a3668c451f45e4cTBR=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
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/+/6dbfb21a6c88af6d94e8c823c3ad559f1a41b493TBR=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
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
