If we want to support VEX-encoded instructions (AVX, F16C, etc.) without a ridiculous slowdown, we need to make sure we're running either all VEX-encoded instructions or all non-VEX-encoded instructions. That means we cannot mix arbitrary user-defined SkRasterPipeline::Fn (never VEX) with those living in SkOpts (maybe VEX)... it's SkOpts or bust.
This ports the existing user-defined SkRasterPipeline::Fn use cases over to use stock stages from SkOpts. I rewrote the unit test to use stock stages, and moved the SkXfermode implementations to SkOpts. The code deleted for SkArithmeticMode_scalar should already be dead.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2940
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Change-Id: I94dbe766b2d65bfec6e544d260f71d721f0f5cb0
Reviewed-on: https://skia-review.googlesource.com/2940
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Mike Reed <reed@google.com>
This lets them pick up runtime CPU specializations. Here I've plugged in SSE4.1. This is still one of the N prelude CLs to full 8-at-a-time AVX.
I've moved the union of the stages used by SkRasterPipelineBench and SkRasterPipelineBlitter to SkOpts... they'll all be used by the blitter eventually. Picking up SSE4.1 specialization here (even still just 4 pixels at a time) is a significant speedup, especially to store_srgb(), so much that it's no longer really interesting to compare against the fused-but-default-instruction-set version in the bench. So that's gone now.
That left the SkRasterPipeline unit test as the only other user of the EasyFn simplified interface to SkRasterPipeline. So I converted that back down to the bare-metal interface, and EasyFn and its friends became SkRasterPipeline_opts.h exclusive abbreviations (now called Kernel_Sk4f). This isn't really unexpected: SkXfermode also wanted to build up its own little abstractions, and once you build your own abstraction, the value of an additional EasyFn-like layer plummets to negative.
For simplicity I've left the SkXfermode stages alone, except srcover() which was always part of the blitter. No particular reason except keeping the churn down while I hack. These _can_ be in SkOpts, but don't have to be until we go 8-at-a-time.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2752
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Change-Id: I3b476b18232a1598d8977e425be2150059ab71dc
Reviewed-on: https://skia-review.googlesource.com/2752
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>
We used to step at a 4-pixel stride as long as possible, then run up to 3 times, one pixel at a time. Now replace those 1-at-a-time runs with a single tail stamp if there are 1-3 remaining pixels.
This style is simply more efficient: e.g. we'll blend and lerp once for 3 pixels instead of 3 times. This should make short blits significantly more efficient. It's also more future-oriented... AVX+ on Intel and SVE on ARM support masked loads and stores, so we can do the entire tail in one direct step.
This also makes it possible to re-arrange the code a bit to encapsulate each stage better. I think generally this code reads more clearly than the old code, but YMMV. I've arranged things so you write one function, but it's compiled into two specializations, one for tail=0 (Body) and one for tail>0 (Tail). It's pretty tidy.
For now I've just burned a register to pass around tail. It's 2 bits now, maybe soon 3 with AVX, and capped at 4 for even the craziest new toys, so there are plenty of places we can pack it if we want to get clever.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2717
Change-Id: I45852a3e5d4c5b5e9315302c46601aee0d32265f
Reviewed-on: https://skia-review.googlesource.com/2717
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
Most visibly this adds a macro SK_RASTER_STAGE that cuts down on the boilerplate of defining a raster pipeline stage function.
Most interestingly, SK_RASTER_STAGE doesn't define a SkRasterPipeline::Fn, but rather a new type EasyFn. This function is always static and inlined, and the details of interacting with the SkRasterPipeline::Stage are taken care of for you: ctx is just passed as a void*, and st->next() is always called. All EasyFns have to do is take care of the meat of the work: update r,g,b, etc. and read and write from their context.
The really neat new feature here is that you can either add EasyFns to a pipeline with the new append() functions, _or_ call them directly yourself. This lets you use the same set of pieces to build either a pipelined version of the function or a custom, fused version. The bench shows this off.
On my desktop, the pipeline version of the bench takes about 25% more time to run than the fused one.
The old approach to creating stages still works fine. I haven't updated SkXfermode.cpp or SkArithmeticMode.cpp because they seemed just as clear using Fn directly as they would have using EasyFn.
If this looks okay to you I will rework the comments in SkRasterPipeline to explain SK_RASTER_STAGE and EasyFn a bit as I've done here in the CL description.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2195853002
Review-Url: https://codereview.chromium.org/2195853002
