| Commit message (Collapse) | Author | Age |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a reland of 78cb579f33943421afc8423a39867fcfd69fed44
This time, lowp stages are controlled by !defined(JUMPER_IS_SCALAR), not
by defined(__clang__). The two are usually the same, except when we opt
Clang builds into JUMPER_IS_SCALAR artificially.
Some Google3 builds use compilers old enough that they barf when
compiling our NEON code. It's conceivably also possible to define
JUMPER_IS_SCALAR yourself, but I don't think anyone does that.
Original change's description:
> Reland "make SkJumper stages normal Skia code"
>
> This is a reland of 22e536e3a1a09405d1c0e6f071717a726d86e8d4
>
> Now with fixed #include paths in SkRasterPipeline_opts.h,
> and -ffp-contract=fast for the :hsw target to minimize
> diffs on non-Windows Clang AVX2/AVX-512 bots.
>
> Original change's description:
> > make SkJumper stages normal Skia code
> >
> > Enough clients are using Clang now that we can say, use Clang to build
> > if you want these software pipeline stages to go fast.
> >
> > This lets us drop the offline build aspect of SkJumper stages, instead
> > building as part of Skia using the SkOpts framework.
> >
> > I think everything should work, except I've (temporarily) removed
> > AVX-512 support. I will put this back in a follow up.
> >
> > I have had to drop Windows down to __vectorcall and our narrower
> > stage calling convention that keeps the d-registers on the stack.
> > I tried forcing sysv_abi, but that crashed Clang. :/
> >
> > Added a TODO to up the same narrower stage calling convention
> > for lowp stages... we just *don't* today, for no good reason.
> >
> > Change-Id: Iaaa792ffe4deab3508d2dc5d0008c163c24b3383
> > Reviewed-on: https://skia-review.googlesource.com/110641
> > Commit-Queue: Mike Klein <mtklein@chromium.org>
> > Reviewed-by: Herb Derby <herb@google.com>
> > Reviewed-by: Florin Malita <fmalita@chromium.org>
>
> Change-Id: I44f2c03d33958e3807747e40904b6351957dd448
> Reviewed-on: https://skia-review.googlesource.com/112742
> Reviewed-by: Mike Klein <mtklein@chromium.org>
Change-Id: I3d71197d4bbb19ca4a94961a97fa2e54d5cbfb0d
Reviewed-on: https://skia-review.googlesource.com/112744
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This reverts commit 78cb579f33943421afc8423a39867fcfd69fed44.
Reason for revert: lowp should be controlled by defined(JUMPER_IS_SCALAR), not defined(__clang__). So close.
Original change's description:
> Reland "make SkJumper stages normal Skia code"
>
> This is a reland of 22e536e3a1a09405d1c0e6f071717a726d86e8d4
>
> Now with fixed #include paths in SkRasterPipeline_opts.h,
> and -ffp-contract=fast for the :hsw target to minimize
> diffs on non-Windows Clang AVX2/AVX-512 bots.
>
> Original change's description:
> > make SkJumper stages normal Skia code
> >
> > Enough clients are using Clang now that we can say, use Clang to build
> > if you want these software pipeline stages to go fast.
> >
> > This lets us drop the offline build aspect of SkJumper stages, instead
> > building as part of Skia using the SkOpts framework.
> >
> > I think everything should work, except I've (temporarily) removed
> > AVX-512 support. I will put this back in a follow up.
> >
> > I have had to drop Windows down to __vectorcall and our narrower
> > stage calling convention that keeps the d-registers on the stack.
> > I tried forcing sysv_abi, but that crashed Clang. :/
> >
> > Added a TODO to up the same narrower stage calling convention
> > for lowp stages... we just *don't* today, for no good reason.
> >
> > Change-Id: Iaaa792ffe4deab3508d2dc5d0008c163c24b3383
> > Reviewed-on: https://skia-review.googlesource.com/110641
> > Commit-Queue: Mike Klein <mtklein@chromium.org>
> > Reviewed-by: Herb Derby <herb@google.com>
> > Reviewed-by: Florin Malita <fmalita@chromium.org>
>
> Change-Id: I44f2c03d33958e3807747e40904b6351957dd448
> Reviewed-on: https://skia-review.googlesource.com/112742
> Reviewed-by: Mike Klein <mtklein@chromium.org>
TBR=mtklein@chromium.org,herb@google.com,fmalita@chromium.org
Change-Id: Ie64da98f5187d44e03c0ce05d7cb189d4a6e6663
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/112743
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a reland of 22e536e3a1a09405d1c0e6f071717a726d86e8d4
Now with fixed #include paths in SkRasterPipeline_opts.h,
and -ffp-contract=fast for the :hsw target to minimize
diffs on non-Windows Clang AVX2/AVX-512 bots.
Original change's description:
> make SkJumper stages normal Skia code
>
> Enough clients are using Clang now that we can say, use Clang to build
> if you want these software pipeline stages to go fast.
>
> This lets us drop the offline build aspect of SkJumper stages, instead
> building as part of Skia using the SkOpts framework.
>
> I think everything should work, except I've (temporarily) removed
> AVX-512 support. I will put this back in a follow up.
>
> I have had to drop Windows down to __vectorcall and our narrower
> stage calling convention that keeps the d-registers on the stack.
> I tried forcing sysv_abi, but that crashed Clang. :/
>
> Added a TODO to up the same narrower stage calling convention
> for lowp stages... we just *don't* today, for no good reason.
>
> Change-Id: Iaaa792ffe4deab3508d2dc5d0008c163c24b3383
> Reviewed-on: https://skia-review.googlesource.com/110641
> Commit-Queue: Mike Klein <mtklein@chromium.org>
> Reviewed-by: Herb Derby <herb@google.com>
> Reviewed-by: Florin Malita <fmalita@chromium.org>
Change-Id: I44f2c03d33958e3807747e40904b6351957dd448
Reviewed-on: https://skia-review.googlesource.com/112742
Reviewed-by: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This reverts commit 22e536e3a1a09405d1c0e6f071717a726d86e8d4.
Reason for revert: wrong include path :/
Original change's description:
> make SkJumper stages normal Skia code
>
> Enough clients are using Clang now that we can say, use Clang to build
> if you want these software pipeline stages to go fast.
>
> This lets us drop the offline build aspect of SkJumper stages, instead
> building as part of Skia using the SkOpts framework.
>
> I think everything should work, except I've (temporarily) removed
> AVX-512 support. I will put this back in a follow up.
>
> I have had to drop Windows down to __vectorcall and our narrower
> stage calling convention that keeps the d-registers on the stack.
> I tried forcing sysv_abi, but that crashed Clang. :/
>
> Added a TODO to up the same narrower stage calling convention
> for lowp stages... we just *don't* today, for no good reason.
>
> Change-Id: Iaaa792ffe4deab3508d2dc5d0008c163c24b3383
> Reviewed-on: https://skia-review.googlesource.com/110641
> Commit-Queue: Mike Klein <mtklein@chromium.org>
> Reviewed-by: Herb Derby <herb@google.com>
> Reviewed-by: Florin Malita <fmalita@chromium.org>
TBR=mtklein@chromium.org,herb@google.com,fmalita@chromium.org
Change-Id: I2bdc709c80cdfa6b13ff24e024b3721bef887f46
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/112741
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Enough clients are using Clang now that we can say, use Clang to build
if you want these software pipeline stages to go fast.
This lets us drop the offline build aspect of SkJumper stages, instead
building as part of Skia using the SkOpts framework.
I think everything should work, except I've (temporarily) removed
AVX-512 support. I will put this back in a follow up.
I have had to drop Windows down to __vectorcall and our narrower
stage calling convention that keeps the d-registers on the stack.
I tried forcing sysv_abi, but that crashed Clang. :/
Added a TODO to up the same narrower stage calling convention
for lowp stages... we just *don't* today, for no good reason.
Change-Id: Iaaa792ffe4deab3508d2dc5d0008c163c24b3383
Reviewed-on: https://skia-review.googlesource.com/110641
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Herb Derby <herb@google.com>
Reviewed-by: Florin Malita <fmalita@chromium.org>
|
|
|
|
|
|
|
|
|
| |
CQ_INCLUDE_TRYBOTS=skia.primary:Test-Debian9-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD
Bug: skia:
Change-Id: If6f0d0a57463bf99a66d674e65a62ce3931d0116
Reviewed-on: https://skia-review.googlesource.com/24644
Commit-Queue: Mike Reed <reed@google.com>
Reviewed-by: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
No pixel diffs.
Performance on 8888 looks like an overall win.
Before:
micros bench
222.41 bitmap_scale_filter_64_256
40.06 bitmap_scale_filter_256_64
8.17 bitmap_scale_filter_90_10
10.32 bitmap_scale_filter_90_30
22.50 bitmap_scale_filter_90_80
1.80 bitmap_scale_filter_90_90
57.51 bitmap_scale_filter_80_90
41.99 bitmap_scale_filter_30_90
31.51 bitmap_scale_filter_10_90
After:
micros bench
193.60 bitmap_scale_filter_64_256
46.26 bitmap_scale_filter_256_64
7.81 bitmap_scale_filter_90_10
9.99 bitmap_scale_filter_90_30
22.05 bitmap_scale_filter_90_80
1.96 bitmap_scale_filter_90_90
52.07 bitmap_scale_filter_80_90
37.73 bitmap_scale_filter_30_90
27.63 bitmap_scale_filter_10_90
Change-Id: I2f29366b0fd503176c5af4d825fa524e632da21b
Reviewed-on: https://skia-review.googlesource.com/7630
Reviewed-by: Matt Sarett <msarett@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
__AVX2__ will not be defined if you omit -mavx2. Android does this
intentionally for x86 builds. (No mobile CPU supports AVX2 AFAIK.)
This should fix the Android roll.
Change-Id: Ib94c862641abc11fbb46863afc53bcc049f362ad
Reviewed-on: https://skia-review.googlesource.com/7633
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Mike Klein <mtklein@chromium.org>
Reviewed-by: Matt Sarett <msarett@google.com>
|
|
|
|
|
|
|
|
|
| |
These new names reflect the actual pixels stored in each register.
Change-Id: I8e626196cd8bcbef622e4fb87ac3566a79d3573a
Reviewed-on: https://skia-review.googlesource.com/7624
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Matt Sarett <msarett@google.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I'm warming back up to the idea of very careful use of SkOpts_hsw.
But if we're going to do that, we need a strict header discipline.
No header can be assumed to be safe without vetting, and most aren't.
Today there's only one function defined in SkOpts_hsw, so this CL mostly
rewrites that convolve_vertically() to use no headers beyond immintrin.h
and stdint.h, both safe. It shared very little code with the others
anyway, so we're not losing anything by putting it directly into
SkOpts_hsw.cpp. I have also streamlined the implementation considerably
to improve maintainability and readability.
Change-Id: Ia03daae660e54125a0d2e2988464cfc930349e80
Reviewed-on: https://skia-review.googlesource.com/7611
Reviewed-by: Matt Sarett <msarett@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
It's proving too difficult to keep on top of all the ways we might cause ODR violations that crash Chrome. I'd rather focus on other ways of running the pipelines that won't have that particular problem. Our -Fast bots will keep testing and benchmarking AVX2 raster pipelines.
BUG=chromium:679147,chromium:654213,chromium:664864,chromium:666707,etc.
CQ_INCLUDE_TRYBOTS=skia.primary:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD
Change-Id: I35ba8f5f4303107237fd78a6ce442d7c26e5fbef
Reviewed-on: https://skia-review.googlesource.com/6827
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This time, with manual program memory management instead of std::vector<void*>.
Using STL types from SkOpts_hsw.cpp is not safe. Things like std::vector<void*>
are inlined but not anonymous, so they're deduped by the linker arbitrarily. This
is bad when we pick the version compiled with AVX instructions on a machine that
doesn't support AVX...
std::vector<Stage> was safe before because Stage itself was anonymous. While not anonymous, std::vector<Stage> is unique to the compilation unit, because you can only refer to the anonymous Stage in the compilation unit.
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.linux:linux_chromium_asan_rel_ng;skia.primary:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD
Change-Id: I015e27583b6b6ff06b5e9f63e3f40ee6b27d6dbd
Reviewed-on: https://skia-review.googlesource.com/6550
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
ConvolveVertically time is reduced about 60% using haswell cpu.
Nanobench results:
before after
bitmap_scale_filter_64_256 611us 302us
bitmap_scale_filter_80_90 101us 64.9us
bitmap_scale_filter_30_90 82.3us 51.4us
bitmap_scale_filter_10_90 73.6us 42.4us
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2526733002
CQ_INCLUDE_TRYBOTS=skia.primary:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD
Review-Url: https://codereview.chromium.org/2526733002
|
|
|
|
|
|
|
|
|
| |
CQ_INCLUDE_TRYBOTS=skia.primary:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD
Change-Id: I308b6d75f2987a667eead9a55760a2ff6aec2984
Reviewed-on: https://skia-review.googlesource.com/5353
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
BUG=666707
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=5089
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Change-Id: I3bebfdf635d541d92fb84236f0f6fae2da39d691
Reviewed-on: https://skia-review.googlesource.com/5089
Reviewed-by: Bruce Dawson <brucedawson@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
| |
I'm not yet caching these in the blitter, and speed is essentially unchanged in the bench where I am now building and compiling the pipeline only once. This may not be able to stay a simple std::function after I figure out caching, but for now it's a nice fit.
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=3911
Change-Id: I9545af589f73baf9f17cb4e6ace9a814c2478fe9
Reviewed-on: https://skia-review.googlesource.com/3911
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The portable code now becomes entirely focused on enum+ptr descriptions, leaving the concrete implementation of the pipeline to SkOpts::run_pipeline().
As implemented, the concrete implementation is basically the same, with a little more type safety.
Speed is essentially unchanged on my laptop, and that's having run_pipeline() rebuild its concrete state every call. There's room for improvement there if we split this into a compile_pipeline() / run_pipeline() sort of thing, which is my next planned CL.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=3920
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Change-Id: Ie4c554f51040426de7c5c144afa5d9d9d8938012
Reviewed-on: https://skia-review.googlesource.com/3920
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Give body and tail functions separate types. This frees a register in body functions, especially important for Windows.
- Fill out default, SSE4.1, and HSW versions of all functions. This means we don't have to mess around with SkNf_abi... all functions come from the same compilation unit where SkNf is a single consistent type.
- Move Stage::next() into SkRasterPipeline_opts.h as a static inline function.
- Remove Stage::ctx() entirely... fCtx is literally the same thing.
This is a step along the way toward building the entire pipeline in src/opts, removing the need for all the stages to be functions living in SkOpts.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=3680
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot,Test-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Debug-ASAN-Trybot
Change-Id: I7de78ffebc15b9bad4eda187c9f50369cd7e5e42
Reviewed-on: https://skia-review.googlesource.com/3680
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Originally reviewed: https://skia-review.googlesource.com/3667
This time around, don't forget swap_src_dst.
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Debug-ASAN-Trybot
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=3676
Change-Id: I127e7fb2bf9d3bfee61c3749fc1c334c9476cb4e
Reviewed-on: https://skia-review.googlesource.com/3676
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This reverts commit 4f02ce7995554d899cdde2b7d82f600fc8017fcc.
Reason for revert: missed a stage
TBR=mtklein@chromium.org,msarett@google.com,herb@google.com,reviews@skia.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
Change-Id: I1dc1229183d67fe72977e492977a97b19dc630d2
Reviewed-on: https://skia-review.googlesource.com/3675
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
All SkNx are now in anonymous namespaces and all their methods are force-inlined. We should not have any ODR problems.
This is still a near 2x speedup, more so for f16.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=3667
Change-Id: I6db9a46f7164f49827ab4d7983e80bf8cea99995
Reviewed-on: https://skia-review.googlesource.com/3667
Reviewed-by: Matt Sarett <msarett@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
| |
This reverts commit Id0ba250037e271a9475fe2f0989d64f0aa909bae.
crbug.com/654213
Looks like Chrome Canary's picking up Haswell code on non-Haswell machines.
Change-Id: I16f976da24db86d5c99636c472ffad56db213a2a
Reviewed-on: https://skia-review.googlesource.com/3108
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Original review here: https://skia-review.googlesource.com/c/2990/
Changes since:
- simpler implementations of load_tail() / store_tail(): slower, but more obviously correct to all compilers
- fleshed out math ops on Sk8i and Sk8u to make unit tests happy on -Fast bot (where we always have AVX2)
- now storing stage functions as void(*)() to avoid undefined behavior and/or linker problems. This restores 32-bit Windows.
- all AVX2 Sk8x methods are marked always-inline, to avoid linking the "wrong" version on Debug builds.
CQ_INCLUDE_TRYBOTS=master.client.skia:Perf-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Debug-ASAN-Trybot,Perf-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Debug-GN,Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot,Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-Fast-Trybot;master.client.skia.compile:Build-Win-MSVC-x86_64-Debug-Trybot
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=3064
Change-Id: Id0ba250037e271a9475fe2f0989d64f0aa909bae
Reviewed-on: https://skia-review.googlesource.com/3064
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This reverts commit I1c82e5755d8e44cc0b9c6673d04b117f85d71a3a.
Reason for revert: lots of failing bots.
TBR=mtklein@chromium.org,msarett@google.com,reviews@skia.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
Change-Id: I653bed3905187f43196504f19424985fa2a765b5
Reviewed-on: https://skia-review.googlesource.com/3063
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Mike Klein <mtklein@chromium.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Bench runtime changes:
sRGB: 7194 -> 3735 = 1.93x faster
F16: 6531 -> 2559 = 2.55x faster
Instead of building 4x and 1-3x pipelines and then maybe 8x and 1-7x, instead build either the short ones or the long ones, but not both. If we just take care to use a compatible run_pipeline(), there's some cross-module type disagreement but everything works out in the end.
Oddly, a few places that looked like they'd be faster using SkNx_fma() or Sk4f_round()/Sk8f_round() are actually faster the long way, e.g. multiply, add 0.5, truncate. Curious! In all the other places you see here that I've used SkNx_fma(), it's been a significant speedup.
This folds in a couple refactors and cleanups that I've been meaning to do. Hope you don't mind... if find the new code considerably easier to read than the old code.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2990
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Change-Id: I1c82e5755d8e44cc0b9c6673d04b117f85d71a3a
Reviewed-on: https://skia-review.googlesource.com/2990
Reviewed-by: Matt Sarett <msarett@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
|
|
Haswell brought a whole slew of handy new instructions for us (AVX2, FMA, BMI1+BMI2) and also feature F16C, which came one generation earlier on Ivybridge. We work with integers often enough that we really want to target AVX2 instead of AVX, and this means it's pretty practical to ask for all those other goodies along with it.
Chrome's GN files and Google3's BUILD file will need an update, before or after this CL.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2840
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Change-Id: I826daf77b5104664c5d31ddaabee347e287b87a2
Reviewed-on: https://skia-review.googlesource.com/2840
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Herb Derby <herb@google.com>
|