I measured relative runtimes on my laptop:
pack_int_uint16_t_ss…
1036 …e41 1x …se3 1.01x …e2_b 3.01x …e2_a 3.02x
I've run into Clang problems with the actual _mm_packus_epi32 instruction, I think,
so I'm going to exercise a little cowardice and leave that option disabled for now.
The ssse3 version probably looks a little faster than it will be in practice.
We'll usually need to load its mask, which here is hoisted out of the bench loop.
The two sse2 variants are close enough in speed that I'm tie breaking them on other
concerns: the <<16, >>16 version doesn't need any scratch registers or to load any
constants, so it wins.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2150343002
CQ_INCLUDE_TRYBOTS=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
Review-Url: https://codereview.chromium.org/2150343002
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
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/+/3296bee70d074bb8094b3229dbe12fa016657e90TBR=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
Looks like this code is using sk_calloc(), NULL on failure, accidentally
instead of sk_calloc_throw(). We're using sk_malloc_throw() in the parallel
code path, so it really seems like we're not checking the result pointer.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2152753002
Review-Url: https://codereview.chromium.org/2152753002
I've also changed it so all attachment views (texture, color, and resolve) are created separately and not shared with each other. This just added a lot more complexity than we were probably even saving in time.
A quick fix to make sure we don't reuse keys in resource tracking also
got merged into this change.
BUG=skia:5223
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2146103002
Review-Url: https://codereview.chromium.org/2146103002
These two new types are in support of Vulkan and the ability to send
separate texture and sampler uniforms to the shader. They don't really fit
well in the current system, since the current system ties together to idea
of intended use and how to emit shader code into the same GrSLType enum.
In vulkan, I want the GrGLSLSampler object to be used as a Sampler2D, but
when appending its declaration it will emit a Texture2D and sampler object.
Our query for GrSLTypeIsSamplerType refers more to the combination of texture
and sampler and not just the sampler part. The GrSLTypeIs2DTextureType query
is for is a a SamplerType that uses Texture2Ds. My new types don't really fit
into either these categories as they are just half of the whole.
In some refactoring down the road (possibly connected with SkSL), I suggest we
split apart the concept of how we intend to use a GrGLSLSampler (Sampler2D, SamplerBuffer,
etc.), from how we actually add it to the code (sampler, texture2D, sampler2D, etc.).
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2143143002
Review-Url: https://codereview.chromium.org/2143143002
When we start a new MonotonePoly due to a handedness change, we don't need to
increase the vertex count, since that edge (and vertex) has already been
accounted for in the previous MonotonePoly.
This was not a correctness issue, but was causing us to allocate
extra vertices which would go unused.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2146063002
Review-Url: https://codereview.chromium.org/2146063002
On some platforms, a newly-created buffer was liable to be CPU backed.
This would break code that expected a VBO (aka instanced rendering).
This change adds an optional flag to GrResourceProvider that requires
a buffer to be created in GPU memory.
It also moves the CPU backing logic into Gr land in order to properly
cache real VBOs on platforms that prefer client-side buffers.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2143333002
Review-Url: https://codereview.chromium.org/2143333002
