Adds a saturated_add function that was on SkNx and used in
SkXfermode_opts, but hadn't been ported to skvx yet.
Removes the Sk4px_opts variants and simplifies some of its functions;
many were already defined skvx.
The largest change is that Sk4px does not extend skvx::byte16, since it
used to extend Sk16b. Now it just has a vector as a data type. This
was necessary so that we could define operators that were typed for
Sk4px and Wide w/o conflicting with the free operators that were
defined for the base skvx types.
Change-Id: I8c667ba86f662ccf07ad85aa32e78abfc0a8c7ae
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/542645
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Change-Id: I1a5490f546a3cb046c64b114a30be991d2d9f2cc
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/541064
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Removes specializations for all() on AVX2 and SSE 4.1, which give the
wrong results if the ints didn't have all bits set (inconsistent with
other platforms and non-SIMD). Added a unit test that checks this case.
The mirror specializations for any() on AVX2 and SSE 4.1 are actually
valid, so added those, and added a 2 instruction specialization for
SSE for any() and all(). This is what clang-trunk produces on -O3, but
ToT clang struggles to vectorize it.
Also adds specializations for NEON for any() and all(), since even
clang-trunk was struggling to vectorize it automatically. In
particular, this will help skgpu::graphite::Rect's implementations of
intersect and contains, which use any/all to get a final boolean value.
In the Instruments app, I had see Rect's intersection as a hotspot
on the Mac M1, and this vectorization helps a bit.
Also takes the opportunity to remove fake C++14 constexpr for a
real constexpr.
Change-Id: Ib142e305ae5615056a777424e379b6da82d44f0c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/542296
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Change-Id: I511f6105537b24953de1533ad7b73d1186afd4fc
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/541060
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This adds aliases like skvx::float2, float4, etc. to SkVx.h and goes
through existing usages of SkVx to standardize on those aliases, or
refer to the full name directly.
In particular, this lets us clean up the equivalent aliases in
src/gpu/tessellate, src/gpu/graphite/VectorTypes and src/gpu/ganesh/GrVx
Where possible, I switched to using skvx::Foo directly and leveraged
auto to make it less redundant. Headers always used the full type
except for PatchWriter.h and Rect.h because of the number of their
usages. In this case, the alias is scoped to private so it can't leak.
This is prep to migrate older code that is still using SkNx and its
aliases like Sk4f to SkVx as well.
Change-Id: I9dd104e83cf17c2b88995a047cfd2e2b0fe6fac2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/541058
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
SkVx.h attempted to not use vector instructions if SKNX_NO_SIMD was set.
However, this was incomplete and could lead to a wasm build with
SKNX_NO_SIMD still trying to use some vector instructions which are
not present if __wasm_simd128__ is defined.
This change requires some additional "include what you use" includes
since some other files were depending on SkVx.h including the vector
instruction headers.
Bug: cl/421848579
Change-Id: I6a878d64b76677a925b94724926c62f3e42ddd4c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/496313
Reviewed-by: Herb Derby <herb@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Ben Wagner <bungeman@google.com>
This is a reland of 35a74eab5d
Added guard for SKNX_NO_SIMD. I guess they don't want speedy goodness.
Original change's description:
> add a scaled uint32x4_t divided by uint32_t to SkVx
>
> This extracts the divide used in SkImageBlurFilter.cpp, and
> encapsulates it into ScaledDividerU32. It generates results that
> are with in +/- 1 of the rounded answer generated by doubles.
>
> I have added hand coded implementations for sse and for neon to
> hopefully to avoid code generation problems.
>
> Bug: skia:12522
>
> Change-Id: Ia7372d45895c799f69f8c0fd9fdea5efac321139
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/458216
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: Herb Derby <herb@google.com>
Bug: skia:12522
Change-Id: I9833a98f159827f483147c8155f1b92b7a7130ed
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/458716
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Herb Derby <herb@google.com>
This reverts commit 35a74eab5d.
Reason for revert: Breaks Google3
Original change's description:
> add a scaled uint32x4_t divided by uint32_t to SkVx
>
> This extracts the divide used in SkImageBlurFilter.cpp, and
> encapsulates it into ScaledDividerU32. It generates results that
> are with in +/- 1 of the rounded answer generated by doubles.
>
> I have added hand coded implementations for sse and for neon to
> hopefully to avoid code generation problems.
>
> Bug: skia:12522
>
> Change-Id: Ia7372d45895c799f69f8c0fd9fdea5efac321139
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/458216
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: Herb Derby <herb@google.com>
Bug: skia:12522
Change-Id: Id5d6968c813322dfc68e549e2f3afea7da9a0e18
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/458258
Auto-Submit: Herb Derby <herb@google.com>
Commit-Queue: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
This extracts the divide used in SkImageBlurFilter.cpp, and
encapsulates it into ScaledDividerU32. It generates results that
are with in +/- 1 of the rounded answer generated by doubles.
I have added hand coded implementations for sse and for neon to
hopefully to avoid code generation problems.
Bug: skia:12522
Change-Id: Ia7372d45895c799f69f8c0fd9fdea5efac321139
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/458216
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Bug: skia:12515
Change-Id: I8db3501c129d93fc1eb822c90840119a7a7f2b4b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/457478
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: Chris Dalton <csmartdalton@google.com>
This is step 1 of a less surgical method of specializing vectors with
constructors and swizzles.
Bug: skia:12515
Change-Id: I4d65ad595387b35fa74df8564c73952e0a8b681c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/457477
Commit-Queue: Chris Dalton <csmartdalton@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
This reverts commit 01b02956c7.
Reason for revert: Codegen regressions
Original change's description:
> Add convenient "xyzw" accessors and swizzles to skvx
>
> Change-Id: Ic300285d10679a4e34190ab7b6b08bd1f6d80330
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/454309
> Reviewed-by: Michael Ludwig <michaelludwig@google.com>
> Commit-Queue: Chris Dalton <csmartdalton@google.com>
Bug: skia:12515
Change-Id: Id853e4d9e25c6d2ae622668ef064e1b2b078b824
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/457476
Auto-Submit: Chris Dalton <csmartdalton@google.com>
Reviewed-by: Chris Dalton <csmartdalton@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Change-Id: Ic300285d10679a4e34190ab7b6b08bd1f6d80330
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/454309
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: Chris Dalton <csmartdalton@google.com>
These macros are not meant to leak out of the file.
Change-Id: I7e24f65a3053785410c7fac760fd3af46c5c1f1c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337739
Auto-Submit: Mike Klein <mtklein@google.com>
Commit-Queue: Chris Dalton <csmartdalton@google.com>
Reviewed-by: Chris Dalton <csmartdalton@google.com>
Adds an SkChopCubicAt overload that performs two chops at once in
SIMD. Also updates SkChopCubicAt to accept T values of 0 and 1. This
has been the source of bugs in the past.
Bug: skia:10419
Change-Id: Ic8a482a69192fb1685f3766411cbdceed830f9b7
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/327436
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Chris Dalton <csmartdalton@google.com>
Now that we have skvx::map(), anyone can write this sort of
scalar-to-vector code. There are no vector instructions for these, so
they'll never going to be particularly interesting for SkVx to provide.
We did work out _approximate_ versions of each of these for SkVM, and
that's what we use to evaluate these programs there. So if this stuff
really matters we could port that logic back over to SkVx.h.
But in terms of pure refactoring, I think this is where we want to sit
until we decide to use those approximations. I don't really want to
invest much time in the SkSLByteCode interpreter any more.
Change-Id: I4e595dee5fd9e608905305e46b2aebcab986c561
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/326277
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Cq-Include-Trybots: luci.chromium.try:linux_chromium_cfi_rel_ng
Bug: chromium:1137652, chromium:1137958
Change-Id: I8575b588f9a1ba89740b95382b2462338e34bec5
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/326478
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
These don't return reliable portable results, so I don't want to promote
them as good ideas to use. You can get at least 5 different results
from these across the four main architectures we support, and they've
been the root cause of bugs uncovered only in production on undertested
platforms.
Luckily, unused outside of tests.
Change-Id: I532731fe4cddf127253341e5ace8d9c5c9ebb0f1
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/326108
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Just a little refactor no-op.
Change-Id: I1842a0190cd96c60da2fe3c7f88fa56c9f73af81
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/325681
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Rewrite map() to allow any number of arguments,
now also used for 2-argument (pow) and 3-argument (fma) operations.
I left a note about fma()... I can't understand why, but calling as
map(fmaf, x,y,z) ends up with scalar calls to fmaf(), but with the
lambda indirection we see perfect vector codegen.
I had to break map() back into two parts. I don't see any way to pass
both a variadic number of arguments and play our trick with the default
std::index_sequence parameter. The lane lambda similarly exists only to
split up the expansion of the Rest... type pack from the I... index
pack; you can't use two pack expansions in the same expression.
Change-Id: Ia156a7fd846237f687d6018a7f95550c9fd4a56d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/325736
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
This is a reland of e24f7f3de7
... with fix for ~0 constants for the pedantic MSVC.
Original change's description:
> refactor any()/all(), ptest for all()
>
> Part of this is a simple refactor, adapting any() and all() to the new
> style of specialization.
>
> And with that refactor in place, add AVX2/SSE4.1 for all() using ptest.
> This isn't terribly important, but it does help make Op::asserts run
> faster in the SkVM interpreter. I like to run with asserts enabled, and
> this makes passing asserts much cheaper---failing asserts are expensive
> still of course, printing to SkDebugf(), etc.
>
> Change-Id: Iebdeee701fab7c50cce8e457674b565f7dd2ec21
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317422
> Reviewed-by: Herb Derby <herb@google.com>
> Commit-Queue: Mike Klein <mtklein@google.com>
Cq-Include-Trybots: luci.skia.skia.primary:Build-Win-MSVC-x86_64-Debug
Change-Id: I93f08177ef3439e65e4383cc517dba60c0c4ef3e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317638
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
This is a reland of 4985db413d
...with a better implementation of map(). I don't understand
why we had to revert, but it had something with calling the
function pointer in map_(), so maybe this will help.
I've flattened the map_() / map() merge CL into this one,
and marked the resulting map() as no_sanitize("cfi"). I
don't see anything wrong, so I think it's a false positive.
Original change's description:
> update skvx scalar-fallback strategy
>
> Turns out Clang's a lot better at auto-vectorizing "obvious" scalar code
> into obvious vector code when it's written out the long way, e.g.
>
> F32x4 x = ...;
> x = { sqrtf(x[0]), sqrtf(x[1]), sqrtf(x[2]), sqrtf(x[3]) };
>
> vectorizes into sqrtps a lot more reliably than our recurse-onto-scalars
> strategy, and also better than the other naive approach,
>
> F32x4 x = ...;
> for (int i = 0; i < 4; i++) { x[i] = sqrtf(x[i]); }
>
> So here I've added a map(V, fn) -> V' using C++14 tricks to let the
> compiler handle the expansion of x = { fn(x[0]), fn(x[1]), ...
> fn(x[N-1]) } for any N, and implemented most skvx scalar fallback code
> using that.
>
> With these now vectorizing well at any N, we can remove any
> specializations we'd written for particular N, really tidying up.
>
> Over in the SkVM interpreter, this is a big improvement for ceil and
> floor, which were being done 2 floats at a time instead of 8. They're
> now slimmed way down to
>
> shlq $6, %r13
> vroundps $K, (%r12,%r13), %ymm0
> vroundps $K, 32(%r12,%r13), %ymm1
> jmp ...
>
> where K is 9 or 10 depending on the op.
>
> I haven't found a scalar function that Clang will vectorize to vcvtps2pd
> (the rounding one, not truncating vcvttps2pd), so I've kept lrint()
> written the long way, updated to the style I've been using lately with
> specializations inline.
>
> Change-Id: Ia97abe3c876008228bf62b1daacd6f6140408fc4
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317375
> Reviewed-by: Herb Derby <herb@google.com>
> Commit-Queue: Mike Klein <mtklein@google.com>
Cq-Include-Trybots: luci.chromium.try:linux_chromium_cfi_rel_ng
Bug: chromium:1129408
Change-Id: Ia9c14074b9a14a67dd221f4925894d35a551f9d7
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317551
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
This reverts commit 4985db413d.
Reason for revert:
../../third_party/skia/include/private/SkVx.h:491:14: runtime error: control flow integrity check for type 'float (float)' failed during indirect function call
(/lib/x86_64-linux-gnu/libm.so.6+0x36460): note: (unknown) defined here
../../third_party/skia/include/private/SkVx.h:491:14: note: check failed in /b/s/w/ir/out/Release/viz_unittests, destination function located in /lib/x86_64-linux-gnu/libm.so.6
#0 0x55e964d3c1f9 in skvx::Vec<4, float> skvx::map_<4, float, float, 0ul, 1ul, 2ul, 3ul>(skvx::Vec<4, float> const&, float (*)(float), std::__1::integer_sequence<unsigned long, 0ul, 1ul, 2ul, 3ul>) ./../../third_party/skia/include/private/SkVx.h:491
I don't understand what's wrong here, but I have a better map() coming up anyway.
Original change's description:
> update skvx scalar-fallback strategy
>
> Turns out Clang's a lot better at auto-vectorizing "obvious" scalar code
> into obvious vector code when it's written out the long way, e.g.
>
> F32x4 x = ...;
> x = { sqrtf(x[0]), sqrtf(x[1]), sqrtf(x[2]), sqrtf(x[3]) };
>
> vectorizes into sqrtps a lot more reliably than our recurse-onto-scalars
> strategy, and also better than the other naive approach,
>
> F32x4 x = ...;
> for (int i = 0; i < 4; i++) { x[i] = sqrtf(x[i]); }
>
> So here I've added a map(V, fn) -> V' using C++14 tricks to let the
> compiler handle the expansion of x = { fn(x[0]), fn(x[1]), ...
> fn(x[N-1]) } for any N, and implemented most skvx scalar fallback code
> using that.
>
> With these now vectorizing well at any N, we can remove any
> specializations we'd written for particular N, really tidying up.
>
> Over in the SkVM interpreter, this is a big improvement for ceil and
> floor, which were being done 2 floats at a time instead of 8. They're
> now slimmed way down to
>
> shlq $6, %r13
> vroundps $K, (%r12,%r13), %ymm0
> vroundps $K, 32(%r12,%r13), %ymm1
> jmp ...
>
> where K is 9 or 10 depending on the op.
>
> I haven't found a scalar function that Clang will vectorize to vcvtps2pd
> (the rounding one, not truncating vcvttps2pd), so I've kept lrint()
> written the long way, updated to the style I've been using lately with
> specializations inline.
>
> Change-Id: Ia97abe3c876008228bf62b1daacd6f6140408fc4
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317375
> Reviewed-by: Herb Derby <herb@google.com>
> Commit-Queue: Mike Klein <mtklein@google.com>
TBR=mtklein@google.com,herb@google.com
Change-Id: I27b5eff3328bf2ddf7063ee0dee14a378ff23b89
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317546
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
This reverts commit e24f7f3de7.
Reason for revert: Build-Win-MSVC-x86_64-Debug
Original change's description:
> refactor any()/all(), ptest for all()
>
> Part of this is a simple refactor, adapting any() and all() to the new
> style of specialization.
>
> And with that refactor in place, add AVX2/SSE4.1 for all() using ptest.
> This isn't terribly important, but it does help make Op::asserts run
> faster in the SkVM interpreter. I like to run with asserts enabled, and
> this makes passing asserts much cheaper---failing asserts are expensive
> still of course, printing to SkDebugf(), etc.
>
> Change-Id: Iebdeee701fab7c50cce8e457674b565f7dd2ec21
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317422
> Reviewed-by: Herb Derby <herb@google.com>
> Commit-Queue: Mike Klein <mtklein@google.com>
TBR=mtklein@google.com,herb@google.com
Change-Id: Ib3ecbe93aa9d14b10dd87e8aa247f275c2c3eb67
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317545
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Part of this is a simple refactor, adapting any() and all() to the new
style of specialization.
And with that refactor in place, add AVX2/SSE4.1 for all() using ptest.
This isn't terribly important, but it does help make Op::asserts run
faster in the SkVM interpreter. I like to run with asserts enabled, and
this makes passing asserts much cheaper---failing asserts are expensive
still of course, printing to SkDebugf(), etc.
Change-Id: Iebdeee701fab7c50cce8e457674b565f7dd2ec21
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317422
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Turns out Clang's a lot better at auto-vectorizing "obvious" scalar code
into obvious vector code when it's written out the long way, e.g.
F32x4 x = ...;
x = { sqrtf(x[0]), sqrtf(x[1]), sqrtf(x[2]), sqrtf(x[3]) };
vectorizes into sqrtps a lot more reliably than our recurse-onto-scalars
strategy, and also better than the other naive approach,
F32x4 x = ...;
for (int i = 0; i < 4; i++) { x[i] = sqrtf(x[i]); }
So here I've added a map(V, fn) -> V' using C++14 tricks to let the
compiler handle the expansion of x = { fn(x[0]), fn(x[1]), ...
fn(x[N-1]) } for any N, and implemented most skvx scalar fallback code
using that.
With these now vectorizing well at any N, we can remove any
specializations we'd written for particular N, really tidying up.
Over in the SkVM interpreter, this is a big improvement for ceil and
floor, which were being done 2 floats at a time instead of 8. They're
now slimmed way down to
shlq $6, %r13
vroundps $K, (%r12,%r13), %ymm0
vroundps $K, 32(%r12,%r13), %ymm1
jmp ...
where K is 9 or 10 depending on the op.
I haven't found a scalar function that Clang will vectorize to vcvtps2pd
(the rounding one, not truncating vcvttps2pd), so I've kept lrint()
written the long way, updated to the style I've been using lately with
specializations inline.
Change-Id: Ia97abe3c876008228bf62b1daacd6f6140408fc4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317375
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Implement min and max using if_then_else(y<x,...) on vectors
rather than recursing to std::min/std::max applied to scalars.
But actually, factor out and use naive_if_then_else(), which Clang can
reason through better than it can our specialized if_then_else(). This
lets every min() or max() I've looked at compile down to ideal codegen,
vmaxps, vpminsw, etc, where if you use if_then_else() you'd see the
literal comparison and blend as written.
I've been looking at q14x2 codegen in the interpreter, and most things
were already good, unexpectedly even uavg_q14x2. The biggest surprise
was how bad the min/max codegen was, and looking back, even the min_f32
and max_f32 codegen is super bad. This CL fixes all that, leaving us
with the ideal codegen using the specific instruction you'd want,
replacing a giant mess of code that recursed down to scalars.
mul_q14x2 is still bad, but an easy follow up.
Change-Id: I77b5d7c9aa20a9a2f5ceb3e40f1e18ace2a1b5c1
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317310
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Change-Id: Ib10215e1e5a86bf78cc34f9dca670417bb217b73
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317271
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
The default implementation of if_then_else is logically bitwise,
(cond & true_val) | (~cond & false_val)
The existing skvx specializations work only for 32-bit lanes, but we can
easily make them work for any type where the whole vector is the right
size by reducing the granularity down to byte level.
Existing code using 32-bit values and 0xffff'ffff or 0x0000'0000 masks
will continue to work the same. But this now lets us use, e.g. 16-bit
values with 0xffff and 0x0000 masks, or even things like 32-bit values
and a mask like 0xff00ff00, selecting byte by byte.
We can't go any lower without falling back on the generic bitwise
implementation, so we'll have to settle for not getting to use a mask
like 0x0f0f0f0f.
Change-Id: I8518cb3cafc7f6e1480b4ae8af50daad2d28c5df
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317170
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
- remove some workarounds
- more SI/SIN/SIT/SINT use
- rewrap a lot of code to 100 cols
- etc. misc.
Change-Id: I78b7ff272afcbb8658cf147aad8af85d0e2acf42
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/314676
Auto-Submit: Mike Klein <mtklein@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Reviewed-by: Herb Derby <herb@google.com>
GCC intrisics type validation is stricter than the one in Clang, so
passing a uint16x4_t to a function expected to accept float16x4_t
is not valid.
Bug: chromium:819294
Change-Id: I6d68e5458345e78bdb05dd028481fe9cae36c5ff
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/307276
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
- add f32<->f16 functions to skvx
- add f32<->f16 x86 instructions to skvm::Assembler
- add f32<->f16 ops to skvm,
using the skvx functions in the interpreter
Still TODO:
use the new x86 instructions in the JIT
(For now like in many other ways, the aarch64 JIT
continues to languish. Will pick that back up one day.)
Change-Id: Ib8dc1ccdc75ecb23769ea4947d66d3ab22520f23
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/302942
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Although it appeared that the experimental_simd CanvasKit build was
working, the build was not producing actual wasm SIMD operations. This
CL fixes that issue by changing the build arguments.
This issue also fixes an incorrect type issue with the SkVx wasm SIMD
implementation.
Bug: skia:10453
Change-Id: If26f84b09e4d84df36be589245878c821972dffc
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/302669
Reviewed-by: Kevin Lubick <kjlubick@google.com>
Reviewed-by: Mike Klein <mtklein@google.com>
First move if_then_else() specializations inline using a
quasi-constexpr-if approach, letting them operate on any types of the
right vector and lane size. We can't use constexpr-if per se because
this header is sometimes used in C++14 contexts.
Then, add AVX specialization for 8x32-bit types.
SkVM's interpreter uses if_then_else() on three i32x16, and these
changes allow that to vectorize ideally, as two vblendvps instructions.
Change-Id: I8355c47975c736c1fbc32b1f8ceddb772978d271
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/302080
Auto-Submit: Mike Klein <mtklein@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
The `experimental_simd` build target builds
CanvasKit using the Emscripten `-msimd128` flag, to build CanvasKit
with SIMD instructions in the compiled WASM. This build of
CanvasKit works in Chrome Canary 86.0.4186.0
with chrome://flags#enable-webassembly-simd enabled.
Also add WebAssembly-specific intrinsics to SkVx.h to enable
support for almost all native SIMD operations in CanvasKit WebAssmebly.
Also add a Skia/modules/canvaskit/wasm_tools/SIMD folder which contains
build_simd_test.sh for testing whether WASM SIMD intrinsics operations
are actually being used by skvx, and for testing correctness of
WASM SIMD operations. Also contains simd_float_test.cpp and
simd_int_test.cpp which serve as documentation for which operations are
correctly turned into WASM SIMD operations by emscripten.
Bug: skia:10453
Change-Id: Icd312b4d189e8d8667d3ffe12a72bfa6febaab2f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/299705
Reviewed-by: Mike Klein <mtklein@google.com>
- add atan, fract, dividef, subtractf
Also wants mix(), but I'm still learning how to handle 2 args
functions (e.g. how to support atan(y,x) as well)
Change-Id: Ib9f233cd1c4266110cfea68a7d444f834f875f1f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/286276
Reviewed-by: Florin Malita <fmalita@chromium.org>
Reviewed-by: Mike Klein <mtklein@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Mike Reed <reed@google.com>
This lets us get rid of VECTOR_UNARY_FN_VEC.
I don't know exactly what was wrong with VECTOR_UNARY_FN_VEC,
but `color.rgb = color.rgb + a*(sin(6.28*color.rgb)*0.159)` looks
ok to me now when run through the interpreter.
Change-Id: I700398cd55eca1b8e1b3b46858415ecae5585a32
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/286065
Reviewed-by: Florin Malita <fmalita@chromium.org>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
While I think trunc(mad(x, scale, 0.5)) is fine for doing our float
to fixed point conversions, round(mul(x, scale)) was kind of better
all around:
- better rounding than +0.5 and trunc
- faster when mad() is not an fma
- often now no need to use the constant 0.5f or have it in a register
- allows the mul() in to_unorm to use mul_f32_imm
Those last two points are key... this actually frees up 2 registers in
the x86 JIT when using to_unorm().
So I think maybe we can resurrect round and still guarantee our desired
intra-machine stability by committing to using instructions that follow
the current rounding mode, which is what [v]cvtps2dq inextricably uses.
Left some notes on the ARM impl... we're rounding to nearest even there,
which is probably the current mode anyway, but to be more correct we
need a slightly longer impl that rounds float->float then "truncates".
Unsure whether it matters in practice. Same deal in the unit test that
I added back, now testing negative and 0.5 cases too. The expectations
assume the current mode is nearest even.
I had the idea to resurrect this when I was looking at adding _imm Ops
for fma_f32. I noticed that the y and z arguments to an fma_f32 were by
far most likely to be constants, and when they are, they're by far likely
to both be constants, e.g. 255.0f & 0.5f from to_unorm(8,...).
llvm disassembly for SkVM_round unit test looks good:
~ $ llc -mcpu=haswell /tmp/skvm-jit-1231521224.bc -o -
.section __TEXT,__text,regular,pure_instructions
.macosx_version_min 10, 15
.globl "_skvm-jit-1231521224" ## -- Begin function skvm-jit-1231521224
.p2align 4, 0x90
"_skvm-jit-1231521224": ## @skvm-jit-1231521224
.cfi_startproc
cmpl $8, %edi
jl LBB0_3
.p2align 4, 0x90
LBB0_2: ## %loopK
## =>This Inner Loop Header: Depth=1
vcvtps2dq (%rsi), %ymm0
vmovupd %ymm0, (%rdx)
addl $-8, %edi
addq $32, %rsi
addq $32, %rdx
cmpl $8, %edi
jge LBB0_2
LBB0_3: ## %hoist1
xorl %eax, %eax
testl %edi, %edi
jle LBB0_6
.p2align 4, 0x90
LBB0_5: ## %loop1
## =>This Inner Loop Header: Depth=1
vcvtss2si (%rsi,%rax), %ecx
movl %ecx, (%rdx,%rax)
decl %edi
addq $4, %rax
testl %edi, %edi
jg LBB0_5
LBB0_6: ## %leave
vzeroupper
retq
.cfi_endproc
## -- End function
Change-Id: Ib59eb3fd8a6805397850d93226c6c6d37cc3ab84
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/276738
Auto-Submit: Mike Klein <mtklein@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Reviewed-by: Herb Derby <herb@google.com>
This is the easiest way to guarantee Op::fma_f32
actually fuses, by using platform intrinsics.
While implementing this we noticed that quad-pumping
was actually slower than double-pumping by about 25%,
and single-pumping was between the two. Switch from
quad to double pumping.
Change-Id: Ib93fd175fb8f6aaf49f769a95edfa9fd6b2674f6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/275299
Commit-Queue: Mike Klein <mtklein@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Reviewed-by: Herb Derby <herb@google.com>
These are inline, but still subject to the ODR,
and in Debug builds they might not be inlined.
This fixes one unit test failure on the x86 Debug GCC Test bot.
Bug: skia:9664
Change-Id: Id3837fdfbf69bd7012339d89d16e8dedaf113de2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/260520
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Remove the need to include SkTypes.h in SkVx.h,
making SkVx entirely independent of Skia.
As an experiment, switch to checking Clang/GCC-style __SSE__ /
__ARM_NEON defines directly instead of the slightly more abstract
SK_CPU_SSE_LEVEL / SK_ARM_HAS_NEON.
Those SK_ defines only exist to help SSE detection on MSVC, which SkVx
generates serial code for anyway.
If this sticks I may do this same sort of change all through Skia.
Change-Id: I1c51fd6ba1fa48f199ce623824d5ef20ff6be995
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/219541
Reviewed-by: Brian Osman <brianosman@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Current strategy: everything from the top
Things to look at first are the manual changes:
- added tools/rewrite_includes.py
- removed -Idirectives from BUILD.gn
- various compile.sh simplifications
- tweak tools/embed_resources.py
- update gn/find_headers.py to write paths from the top
- update gn/gn_to_bp.py SkUserConfig.h layout
so that #include "include/config/SkUserConfig.h" always
gets the header we want.
No-Presubmit: true
Change-Id: I73a4b181654e0e38d229bc456c0d0854bae3363e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/209706
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Hal Canary <halcanary@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Reviewed-by: Florin Malita <fmalita@chromium.org>
This is a reland of e3b110dc6e
PS1 is the original, so best to diff against that.
This is the original with compiler workarounds.
Original change's description:
> align skvx::Vec<N,T> to N*sizeof(T)
>
> This increases the alignment of these vector types. I would have liked
> to keep the alignment minimal, but it's probably no big deal either way.
>
> In terms of code generation, it doesn't make much difference for x86 or
> ARMv8, but it seems hugely important for good ARMv7 NEON code. It's a
> ~10x difference for the bench I've been playing around with that spends
> most of its time in that SkOpts::blit_row_color32 routine.
>
> Bug: chromium:952502
> Change-Id: Ib12caad6b9b3f3f6e821ed70bfb57099db37b15f
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/208581
> Commit-Queue: Michael Ludwig <michaelludwig@google.com>
> Reviewed-by: Michael Ludwig <michaelludwig@google.com>
> Auto-Submit: Mike Klein <mtklein@google.com>
Bug: chromium:952502
Cq-Include-Trybots: skia.primary:Test-Win2016-MSVC-GCE-CPU-AVX2-x86-Release-All,Build-Debian9-GCC-mips64el-Debug
Change-Id: Ief99e14ab4de5a56840ed6bb326cf7669c51dc97
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/208681
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Yet another surprising finding when looking at ARM code generation is
that passing these values to functions by const& does make a difference,
even when fully inlined. I can only guess that the compiler's somehow
more sure that way that the values won't change? Anyway, convert all
skvx functions that take Vec arguments to take const Vec& instead.
This tweak is enough to let the natural implementation of mull()
actually produce good code generation, so I've promoted that to SkVx.h
and added a unit test. Notice in the NEON case we've got a base case at
N=8 and two recursive cases, one down to 8 as usual when N > 8, but also
one up to 8 when N < 8.
This also is another big speedup for ARMv7 NEON, bringing it to nearly
the same speed as ARMv8 NEON on the same device.
Bug: chromium:952502
Change-Id: I0f19bab45cf02222ccc8090053ea2a4a380f1dfe
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/208582
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
