| Commit message (Collapse) | Author | Age |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
It's become clear we need to sometimes deal with values <0 or >1.
I'm not yet convinced we care about NaN or +-inf.
We had some fairly clever tricks and optimizations here for NEON
and SSE. I've thrown them out in favor of a single implementation.
If we find the specializations mattered, we can certainly figure out
how to extend them to this new range/domain.
This happens to add a vectorized float -> half for ARMv7, which was
missing from the _01 version. (The SSE strategy was not portable to
platforms that flush denorm floats to zero.)
I've tested the full float range for FloatToHalf on my desktop and a 5x.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2145663003
CQ_INCLUDE_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot,Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-Fast-Trybot
Committed: https://skia.googlesource.com/skia/+/3296bee70d074bb8094b3229dbe12fa016657e90
Review-Url: https://codereview.chromium.org/2145663003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
(patchset #7 id:120001 of https://codereview.chromium.org/2145663003/ )
Reason for revert:
Unit tests fail on Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-Fast
Original issue's description:
> Expand _01 half<->float limitation to _finite. Simplify.
>
> It's become clear we need to sometimes deal with values <0 or >1.
> I'm not yet convinced we care about NaN or +-inf.
>
> We had some fairly clever tricks and optimizations here for NEON
> and SSE. I've thrown them out in favor of a single implementation.
> If we find the specializations mattered, we can certainly figure out
> how to extend them to this new range/domain.
>
> This happens to add a vectorized float -> half for ARMv7, which was
> missing from the _01 version. (The SSE strategy was not portable to
> platforms that flush denorm floats to zero.)
>
> I've tested the full float range for FloatToHalf on my desktop and a 5x.
>
> BUG=skia:
> GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2145663003
> CQ_INCLUDE_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot,Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-Fast-Trybot
>
> Committed: https://skia.googlesource.com/skia/+/3296bee70d074bb8094b3229dbe12fa016657e90
TBR=msarett@google.com,mtklein@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Review-Url: https://codereview.chromium.org/2151023003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
It's become clear we need to sometimes deal with values <0 or >1.
I'm not yet convinced we care about NaN or +-inf.
We had some fairly clever tricks and optimizations here for NEON
and SSE. I've thrown them out in favor of a single implementation.
If we find the specializations mattered, we can certainly figure out
how to extend them to this new range/domain.
This happens to add a vectorized float -> half for ARMv7, which was
missing from the _01 version. (The SSE strategy was not portable to
platforms that flush denorm floats to zero.)
I've tested the full float range for FloatToHalf on my desktop and a 5x.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2145663003
CQ_INCLUDE_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Review-Url: https://codereview.chromium.org/2145663003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This removes the code paths where we make SkCpu::Supports() calls
from within a tight loop. It keeps code paths using SkCpu::Supports()
to choose entire routines from src/opts/.
We can't rely on these hyper-local checks to be hoisted up reliably enough.
It worked pretty well with the first couple platforms we tried (e.g. Clang
on Linux/Mac) but we can't gaurantee it works everywhere.
Further, I'm not able to actually do anything fancy with those tests
outside of x86... I've not found a way to get, say, NEON+F16 conversion
code embedded into ordinary NEON code outside writing then entire function
in external assembly.
This whole idea becomes less important now that we've got a way to chain
separate function calls together efficiently. We can now, e.g., use an
AVX+F16C method to load some pixels, then chain that into an ordinary AVX
method to color filter them.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2138073002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Review-Url: https://codereview.chromium.org/2138073002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- floor with roundps is about 4.5x faster when available
- f16 srcover_n is similar to but a little faster than the version in https://codereview.chromium.org/1884683002. This new one fuses the dst load/stores into the f16<->f32 conversions:
+0x180 movups (%r15), %xmm1
+0x184 vcvtph2ps (%rbx), %xmm2
+0x189 movaps %xmm1, %xmm3
+0x18c shufps $255, %xmm3, %xmm3
+0x190 movaps %xmm0, %xmm4
+0x193 subps %xmm3, %xmm4
+0x196 mulps %xmm2, %xmm4
+0x199 addps %xmm1, %xmm4
+0x19c vcvtps2ph $0, %xmm4, (%rbx)
+0x1a2 addq $16, %r15
+0x1a6 addq $8, %rbx
+0x1aa decl %r14d
+0x1ad jne +0x180
If we decide to land this it'd be a good idea to convert most or all users of SkFloatToHalf_01 and SkHalfToFloat_01 over to the pointer-based versions.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1891513002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Committed: https://skia.googlesource.com/skia/+/cbe3c1af987d622ea67ef560d855b41bb14a0ce9
Committed: https://skia.googlesource.com/skia/+/3faf74b8364491ca806f523fbb1d8a97be592663
Review URL: https://codereview.chromium.org/1891513002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
https://codereview.chromium.org/1891513002/ )
Reason for revert:
this depends on a CL I want to revert
Original issue's description:
> skcpu: sse4.1 floor, f16c f16<->f32
>
> - floor with roundps is about 4.5x faster when available
> - f16 srcover_n is similar to but a little faster than the version in https://codereview.chromium.org/1884683002. This new one fuses the dst load/stores into the f16<->f32 conversions:
>
> +0x180 movups (%r15), %xmm1
> +0x184 vcvtph2ps (%rbx), %xmm2
> +0x189 movaps %xmm1, %xmm3
> +0x18c shufps $255, %xmm3, %xmm3
> +0x190 movaps %xmm0, %xmm4
> +0x193 subps %xmm3, %xmm4
> +0x196 mulps %xmm2, %xmm4
> +0x199 addps %xmm1, %xmm4
> +0x19c vcvtps2ph $0, %xmm4, (%rbx)
> +0x1a2 addq $16, %r15
> +0x1a6 addq $8, %rbx
> +0x1aa decl %r14d
> +0x1ad jne +0x180
>
> If we decide to land this it'd be a good idea to convert most or all users of SkFloatToHalf_01 and SkHalfToFloat_01 over to the pointer-based versions.
>
> BUG=skia:
> GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1891513002
> CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
>
> Committed: https://skia.googlesource.com/skia/+/cbe3c1af987d622ea67ef560d855b41bb14a0ce9
>
> Committed: https://skia.googlesource.com/skia/+/3faf74b8364491ca806f523fbb1d8a97be592663
TBR=fmalita@chromium.org,herb@google.com,reed@google.com,mtklein@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Review URL: https://codereview.chromium.org/1897433002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- floor with roundps is about 4.5x faster when available
- f16 srcover_n is similar to but a little faster than the version in https://codereview.chromium.org/1884683002. This new one fuses the dst load/stores into the f16<->f32 conversions:
+0x180 movups (%r15), %xmm1
+0x184 vcvtph2ps (%rbx), %xmm2
+0x189 movaps %xmm1, %xmm3
+0x18c shufps $255, %xmm3, %xmm3
+0x190 movaps %xmm0, %xmm4
+0x193 subps %xmm3, %xmm4
+0x196 mulps %xmm2, %xmm4
+0x199 addps %xmm1, %xmm4
+0x19c vcvtps2ph $0, %xmm4, (%rbx)
+0x1a2 addq $16, %r15
+0x1a6 addq $8, %rbx
+0x1aa decl %r14d
+0x1ad jne +0x180
If we decide to land this it'd be a good idea to convert most or all users of SkFloatToHalf_01 and SkHalfToFloat_01 over to the pointer-based versions.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1891513002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Committed: https://skia.googlesource.com/skia/+/cbe3c1af987d622ea67ef560d855b41bb14a0ce9
Review URL: https://codereview.chromium.org/1891513002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
https://codereview.chromium.org/1891513002/ )
Reason for revert:
Need to change around my #if guards so that clang-cl is treated like GCC and Clang, rather than MSVC.
Original issue's description:
> skcpu: sse4.1 floor, f16c f16<->f32
>
> - floor with roundps is about 4.5x faster when available
> - f16 srcover_n is similar to but a little faster than the version in https://codereview.chromium.org/1884683002. This new one fuses the dst load/stores into the f16<->f32 conversions:
>
> +0x180 movups (%r15), %xmm1
> +0x184 vcvtph2ps (%rbx), %xmm2
> +0x189 movaps %xmm1, %xmm3
> +0x18c shufps $255, %xmm3, %xmm3
> +0x190 movaps %xmm0, %xmm4
> +0x193 subps %xmm3, %xmm4
> +0x196 mulps %xmm2, %xmm4
> +0x199 addps %xmm1, %xmm4
> +0x19c vcvtps2ph $0, %xmm4, (%rbx)
> +0x1a2 addq $16, %r15
> +0x1a6 addq $8, %rbx
> +0x1aa decl %r14d
> +0x1ad jne +0x180
>
> If we decide to land this it'd be a good idea to convert most or all users of SkFloatToHalf_01 and SkHalfToFloat_01 over to the pointer-based versions.
>
> BUG=skia:
> GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1891513002
> CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
>
> Committed: https://skia.googlesource.com/skia/+/cbe3c1af987d622ea67ef560d855b41bb14a0ce9
TBR=fmalita@chromium.org,herb@google.com,reed@google.com,mtklein@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Review URL: https://codereview.chromium.org/1891993002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- floor with roundps is about 4.5x faster when available
- f16 srcover_n is similar to but a little faster than the version in https://codereview.chromium.org/1884683002. This new one fuses the dst load/stores into the f16<->f32 conversions:
+0x180 movups (%r15), %xmm1
+0x184 vcvtph2ps (%rbx), %xmm2
+0x189 movaps %xmm1, %xmm3
+0x18c shufps $255, %xmm3, %xmm3
+0x190 movaps %xmm0, %xmm4
+0x193 subps %xmm3, %xmm4
+0x196 mulps %xmm2, %xmm4
+0x199 addps %xmm1, %xmm4
+0x19c vcvtps2ph $0, %xmm4, (%rbx)
+0x1a2 addq $16, %r15
+0x1a6 addq $8, %rbx
+0x1aa decl %r14d
+0x1ad jne +0x180
If we decide to land this it'd be a good idea to convert most or all users of SkFloatToHalf_01 and SkHalfToFloat_01 over to the pointer-based versions.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1891513002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Review URL: https://codereview.chromium.org/1891513002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Adds f32 <-> f16 ARMv7 and ARMv8 NEON code.
Also adds NEON f32 <-> u16 code to make the comparison fair.
The NDK GCC does not support the ARMv8 NEON intrinsics needed to go fastest, so we use a tiny amount of inline assembly.
The ARMv7 half -> float is different enough from the SSE version that it does not make sense to use SkNx.
Still TODO:
ARMv7 float -> half. Naively translating the SSE version results in 0x0000 where we'd expect a denormal output.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1700473003
CQ_EXTRA_TRYBOTS=client.skia.android:Test-Android-GCC-Nexus5-CPU-NEON-Arm7-Release-Trybot,Test-Android-GCC-Nexus9-CPU-Denver-Arm64-Release-Trybot;client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Review URL: https://codereview.chromium.org/1700473003
|
|
|
|
|
|
|
|
|
| |
This is a little faster than the previous version, and much better explained.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1688233002
Review URL: https://codereview.chromium.org/1688233002
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
These are basically inlined, 4-at-a-time versions of our existing functions,
but cut down to avoid any work that's only necessary outside [0,1].
Both f16 and f32 denorms should work fine modulo the usual ARMv7 NEON denorm==zero caveat.
In exchange for a little speed, f32->f16 does not round properly.
Instead it truncates, so it's never off by more than 1 bit.
Support for finite values >1 or <0 is straightforward to add back.
>1 might already work as-is.
Getting close to _u16 performance:
micros bench
261.13 xferu64_bw_1_opaque_u16
1833.51 xferu64_bw_1_alpha_u16
2762.32 ? xferu64_aa_1_opaque_u16
3334.29 xferu64_aa_1_alpha_u16
249.78 xferu64_bw_1_opaque_f16
3383.18 xferu64_bw_1_alpha_f16
4214.72 xferu64_aa_1_opaque_f16
4701.19 xferu64_aa_1_alpha_f16
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1685133005
Committed: https://skia.googlesource.com/skia/+/9ea11a4235b3e3521cc8bf914a27c2d0dc062db9
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Review URL: https://codereview.chromium.org/1685133005
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
https://codereview.chromium.org/1685133005/ )
Reason for revert:
Gotta fix Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD
Original issue's description:
> SkHalfToFloat_01 / SkFloatToHalf_01
>
> These are basically inlined, 4-at-a-time versions of our existing functions,
> but cut down to avoid any work that's only necessary outside [0,1].
>
> Both f16 and f32 denorms should work fine modulo the usual ARMv7 NEON denorm==zero caveat.
>
> In exchange for a little speed, f32->f16 does not round properly.
> Instead it truncates, so it's never off by more than 1 bit.
>
> Support for finite values >1 or <0 is straightforward to add back.
> >1 might already work as-is.
>
> Getting close to _u16 performance:
> micros bench
> 261.13 xferu64_bw_1_opaque_u16
> 1833.51 xferu64_bw_1_alpha_u16
> 2762.32 ? xferu64_aa_1_opaque_u16
> 3334.29 xferu64_aa_1_alpha_u16
> 249.78 xferu64_bw_1_opaque_f16
> 3383.18 xferu64_bw_1_alpha_f16
> 4214.72 xferu64_aa_1_opaque_f16
> 4701.19 xferu64_aa_1_alpha_f16
>
>
> BUG=skia:
> GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1685133005
>
> Committed: https://skia.googlesource.com/skia/+/9ea11a4235b3e3521cc8bf914a27c2d0dc062db9
TBR=jvanverth@google.com,reed@google.com,mtklein@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:
Review URL: https://codereview.chromium.org/1693443003
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
These are basically inlined, 4-at-a-time versions of our existing functions,
but cut down to avoid any work that's only necessary outside [0,1].
Both f16 and f32 denorms should work fine modulo the usual ARMv7 NEON denorm==zero caveat.
In exchange for a little speed, f32->f16 does not round properly.
Instead it truncates, so it's never off by more than 1 bit.
Support for finite values >1 or <0 is straightforward to add back.
>1 might already work as-is.
Getting close to _u16 performance:
micros bench
261.13 xferu64_bw_1_opaque_u16
1833.51 xferu64_bw_1_alpha_u16
2762.32 ? xferu64_aa_1_opaque_u16
3334.29 xferu64_aa_1_alpha_u16
249.78 xferu64_bw_1_opaque_f16
3383.18 xferu64_bw_1_alpha_f16
4214.72 xferu64_aa_1_opaque_f16
4701.19 xferu64_aa_1_alpha_f16
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1685133005
Review URL: https://codereview.chromium.org/1685133005
|
|
|
|
|
|
|
|
|
| |
This allows us to create distance field textures with better precision,
which may help text quality.
BUG=skia:3103
Review URL: https://codereview.chromium.org/762923003
|
|
Based on code by Fabian Giesen at
https://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/.
These will be needed for creating binary16 textures from floating point data.
BUG=skia:3103
Review URL: https://codereview.chromium.org/760753003
|