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>
This reverts commit e3b110dc6e.
Reason for revert: bot failures
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>
TBR=mtklein@google.com,michaelludwig@google.com
Change-Id: I72357b9775685efcc2cd75db220711c8145b8ac4
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: chromium:952502
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/208680
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
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>
These replicate the base logic of Sk4px::Wide::div255() and
Sk4px::approxMulDiv255(), and will come in handy replacing them.
No platform specializations yet... want to remind myself what
codegen they get from these vanilla versions first, and then
I'll fill in the platform specific stuff as needed. The tests
should cover everything pretty exhaustively.
Change-Id: I5854d1bc0902a85cbb2351f669c4da7cc31a8775
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/207683
Commit-Queue: Mike Klein <mtklein@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
For some reason, Clang can infer <N,T> but GCC can't.
No big deal... we know exactly the ones we want anyway.
Change-Id: I15ba4d4edbd3bc0f37ebe3c2b6e411726cd9fb69
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/207341
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Was starting to use this and ran into a few problems with clashing
symbols, namely SI and cast(). Seemed simple enough to not use SI,
and to move all the free-standing types into skvx: skvx::cast,
skvx::shuffle, etc.
Change-Id: Ia5d8ef6d0ae5375bf80d76be88d16f0c9cde56e7
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/207340
Commit-Queue: Mike Klein <mtklein@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Guarding the implict constructors and scalar/vector
operations with std::is_convertible ought to make SkVx
types feel more like normal C types, allowing implicit
conversions exactly when the scalar equivalents would.
This shouldn't change the behavior of any code, or make
anything new possible... just nicer to read and write.
Change-Id: Iff4b89012c5b8c7f7933e6841c925b81186bc614
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/201402
Commit-Queue: Mike Klein <mtklein@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
Since these are all already static, it doesn't have any real functional
impact in terms of linking or codegen. But it does supress unused
function warnings in compilation units that don't use everything.
Add a new SI boilerplate macro to go along with SINT and SIT.
Change-Id: If2c09951b7453338dd20a3a88e3abbee5eefcd27
Reviewed-on: https://skia-review.googlesource.com/c/195921
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Add SSE, SSE4.1, and NEON specializations.
The if_then_else() unit tests in SkVxTest.cpp should cover this.
I had to give up on my dream of not using Skia headers for now. There's
really no good way of knowing whether we've got SSE4.1 support in MSVC
except when we explicitly define SK_CPU_SSE_LEVEL=SK_CPU_SSE_LEVEL_SSE41.
This refactor to use SK_CPU_SSE_LEVEL let MSVC point out a slight
ordering problem that would cause an infinite loop calling any of
the specializions like sqrt(float2). I believe moving them after
the float4 specializations will fix that.
Change-Id: I83639f378a182716d1b37e92b6d725472698f874
Reviewed-on: https://skia-review.googlesource.com/c/195920
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Just trying to get things mostly under 100 cols.
Change-Id: Ifc8f4f0b78a89dfc5ba6ca2e310e969f1880e194
Reviewed-on: https://skia-review.googlesource.com/c/191001
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
- remove ALWAYS_INLINE until we find we need it
- make bit_puns explicit
- implement everything recursively so, e.g.
sqrt(float8) picks up sqrt(float4) when
not otherwise specialized.
- implement SSE specializations:
of the operations I tested, only sqrt, rcp, and rsqrt
needed any help. The others look good as-is.
Change-Id: I1b679c7bd9a99f952272b118d7ade2469b55d604
Reviewed-on: https://skia-review.googlesource.com/c/190222
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Change-Id: Ie3e5b353f84e74d398a5350dc0baff5541789119
Reviewed-on: https://skia-review.googlesource.com/c/189982
Commit-Queue: Mike Klein <mtklein@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Obviously lots of these new operations like sqrt() will want platform
specialization. That'll come later.
Change-Id: Ia0758425d4ec5911968a3d0ad63fa387b9b4cb39
Reviewed-on: https://skia-review.googlesource.com/c/189848
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
Change-Id: I1cb8113af243ed6327179d295835295834a752aa
Reviewed-on: https://skia-review.googlesource.com/c/189581
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
