Enable AVX support. Up to 2x faster is some cases.

RazrFalcon · Oct 20, 2020 · a5737da · a5737da
1 parent 37d21ee
commit a5737da
Show file tree

Hide file tree

Showing 8 changed files with 319 additions and 236 deletions.
diff --git a/Cargo.toml b/Cargo.toml
@@ -19,7 +19,7 @@ float-cmp = { version = "0.8", default-features = false, features = ["std"] }
 memmap2 = { version = "0.1", optional = true }
 num-ext = { git = "https://github.com/RazrFalcon/num-ext", rev = "109f01a" }
 png = { version = "0.16.6", optional = true }
-wide = { version = "0.6.1", features = ["std"] }
+wide = { git = "https://github.com/Lokathor/wide", rev = "cee7b6f", features = ["std"] }
 
 [features]
 default = ["png-format"]

diff --git a/README.md b/README.md
@@ -15,7 +15,7 @@ with a focus on a rendering quality, speed and binary size.
 
 The main motivation behind this library is to have a small, high-quality 2D rendering
 library that can be used by [resvg]. And the choice is rather limited.
-You basically have to choose between cairo, Qt and Skia. And all of them are
+You basically have to choose between [cairo], Qt and Skia. And all of them are
 relatively bloated, hard to compile and distribute. Not to mention that none of them
 is written in Rust.
 
@@ -32,42 +32,38 @@ uses an obscure build system (`gn`) which still uses Python2
 and doesn't really support 32bit targets.
 
 `tiny-skia` tries to be small, simple and easy to build.
+Currently, it has around 12 KLOC and compiles in less than 5s on a modern CPU.
 
 ## Performance
 
-Does `tiny-skia` as fast as [Skia]? The short answer is no. The longer one is: it depends.
+Currently, `tiny-skia` is 20-100% slower than Skia.
+Which is still faster than [cairo] and [raqote].
 
 The heart of Skia's CPU rendering is
 [SkRasterPipeline](https://github.com/google/skia/blob/master/src/opts/SkRasterPipeline_opts.h).
 And this is an extremely optimized piece of code.
 But to be a bit pedantic, it's not really a C++ code. It relies on clang's
-non-standard vector extensions, which means that you must build it with clang.
+non-standard vector extensions, which means that it works only with clang.
 You can actually build it with gcc/msvc, but it will simply ignore all the optimizations
-and become 15-30 *times* slower! Which makes it kinda useless. And `tiny-skia`
-is way closer to a clang version.
-
-Also, `SkRasterPipeline` supports AVX2 instructions, which provide 256-bits wide types.
-This makes common operations almost 2x faster, compared to a generic SSE2/128-bits one.
-Which is no surprise.<br>
-The problem is that Skia doesn't support dynamic CPU detection.
-So by enabling AVX2 you're making the resulting binary non-portable,
-since you need a Haswell processor or newer.<br>
-Right now, `tiny-skia` directly supports only SSE2 instructions
-and relies on autovectorization for newer one.
+and become 15-30 *times* slower! Which makes it kinda useless.
 
 Skia also supports ARM NEON instructions, which are unavailable in a stable Rust at the moment.
-Therefore a default scalar implementation will be used instead on ARM and other non-x86 targets.
+Therefore a fallback scalar implementation will be used instead on ARM and other non-x86 targets.
+So if you're targeting ARM, you better stick with Skia.
 
-Accounting all above, `tiny-skia` is 20-100% slower than "a Skia built for a generic x86_64 CPU".
-Which still makes if faster than `cairo` in many cases.
+Also note, that neither Skia or `tiny-skia` are supporting dynamic CPU detection,
+so by enabling newer instructions you're making the resulting binary non-portable.
 
-We can technically use the `SkRasterPipeline` directly, to achive the same performance as Skia has.
-But it means that we have to complicate our build process quite a lot.
-Mainly because we have to use only clang.
-So having a pure Rust library, even a bit slower one, is still a good trade off.
+Essentially, you will get a decent performance on x86 targets by default.
+But if you are looking for an even better performance, you should compile your application
+with `RUSTFLAGS="-Ctarget-cpu=haswell"` env variables to enable AVX instructions.
 
 You can find more information in [benches/README.md](./benches/README.md).
 
+## Rendering quality
+
+Unless there is a bug, `tiny-skia` must produce exactly the same results as Skia.
+
 ## API overview
 
 The API is a bit unconventional. It doesn't look like cairo, QPainter (Qt), HTML Canvas or even Skia.
@@ -163,7 +159,7 @@ Therefore we have to compromise or even rewrite some parts from scratch.
 
 ## Alternatives
 
-Right now, the only pure Rust alternative is [raqote](https://github.com/jrmuizel/raqote).
+Right now, the only pure Rust alternative is [raqote].
 
 - It doesn't support high-quality antialiasing (hairline stroking in particular).
 - It's very slow (see [benchmarks](./benches/README.md)).
@@ -179,4 +175,6 @@ The project relies on some unsafe code.
 The same as used by [Skia]: [New BSD License](./LICENSE)
 
 [Skia]: https://skia.org/
+[cairo]: https://www.cairographics.org/
+[raqote]: https://github.com/jrmuizel/raqote
 [resvg]: https://github.com/RazrFalcon/resvg
diff --git a/benches/Cargo.lock b/benches/Cargo.lock
diff --git a/benches/README.md b/benches/README.md
@@ -44,17 +44,17 @@ Filling a shape with a solid color.
 | overlay              |        826,240 |        514,436 |    733,158 |    385,684 |  3,412,192 |  7,016,443 |
 | darken               |        631,307 |        434,548 |    610,493 |    317,215 |  3,579,931 |  5,230,384 |
 | lighten              |        633,910 |        434,123 |    621,340 |    319,867 |  3,584,409 |  5,255,202 |
-| color_dodge          |      1,794,356 |      1,386,241 |    780,556 |    679,996 |  5,151,488 |  9,634,447 |
-| color_burn           |      1,878,874 |      1,437,832 |    861,359 |    708,962 |  5,007,181 |  9,617,102 |
+| color_dodge          |      1,794,356 |      1,002,832 |    780,556 |    679,996 |  5,151,488 |  9,634,447 |
+| color_burn           |      1,878,874 |      1,055,831 |    861,359 |    708,962 |  5,007,181 |  9,617,102 |
 | hard_light           |        804,040 |        508,038 |    734,878 |    370,307 |  3,442,584 |  7,034,417 |
-| soft_light           |      2,730,721 |      2,219,540 |  1,225,756 |    986,403 |  5,900,415 | 11,941,630 |
+| soft_light           |      2,730,721 |      1,412,907 |  1,225,756 |    986,403 |  5,900,415 | 11,941,630 |
 | difference           |        650,915 |        440,554 |    632,797 |    326,285 |  3,936,284 |  5,718,776 |
 | exclusion            |        530,865 |        399,089 |    618,645 |    316,235 |  3,842,131 |  6,082,393 |
 | multiply             |        586,406 |        434,776 |    627,202 |    316,841 |  3,608,817 |  5,986,364 |
-| hue                  |      3,662,119 |      3,099,846 |  1,705,745 |  1,413,791 |  7,517,902 | 13,716,827 |
-| saturation           |      3,675,423 |      3,026,723 |  1,676,187 |  1,411,117 |  7,443,261 | 13,752,382 |
-| color                |      3,125,045 |      2,549,478 |  1,431,804 |  1,115,800 |  6,070,058 | 10,537,391 |
-| luminosity           |      3,117,391 |      2,419,344 |  1,351,584 |  1,090,402 |  6,124,294 | 10,488,916 |
+| hue                  |      3,662,119 |      2,008,062 |  1,705,745 |  1,413,791 |  7,517,902 | 13,716,827 |
+| saturation           |      3,675,423 |      2,006,886 |  1,676,187 |  1,411,117 |  7,443,261 | 13,752,382 |
+| color                |      3,125,045 |      1,655,401 |  1,431,804 |  1,115,800 |  6,070,058 | 10,537,391 |
+| luminosity           |      3,117,391 |      1,599,196 |  1,351,584 |  1,090,402 |  6,124,294 | 10,488,916 |
 
 *Destination* is faster in `tiny-skia`, because we're exiting immediately,
 while Skia uses null blitter, so edges processing is still in place.
@@ -144,17 +144,17 @@ Draws a large spiral using a subpixel stroke width.
 | linear, three stops, evenly spread   |      2,068,385 |      1,940,036 |  1,806,614 |    781,344 |  2,413,454 |  4,021,338 |
 | linear, three stops, unevenly spread |      2,068,394 |      1,939,826 |  1,805,379 |    687,479 |  2,423,142 |  3,412,176 |
 | simple radial                        |      2,293,883 |      2,205,554 |  2,050,437 |    805,376 |  4,704,141 |  5,531,178 |
-| two point radial                     |      3,627,946 |      3,716,636 |  1,943,448 |  1,083,230 |  4,709,760 | 13,454,676 |
+| two point radial                     |      3,627,946 |      2,516,614 |  1,943,448 |  1,083,230 |  4,709,760 | 13,454,676 |
 
 ### pattern
 
 `pattern.rs`
 
 | Test/Library                | tiny-skia SSE2 | tiny-skia AVX2 | Skia SSE2  | Skia AVX2  | cairo       | raqote     |
 | --------------------------- | -------------: | -------------: | ---------: | ---------: | ----------: | ---------: |
-| plain (nearest, no ts)      |      3,438,646 |      3,326,305 |  1,315,079 |  1,122,982 |     785,550 |  1,865,327 |
-| lq (bilinear, with ts)      |      9,982,212 |      8,247,785 |  4,484,023 |  2,646,523 |  17,612,685 | 24,906,379 |
-| hq (bicubic/gauss, with ts) |     34,931,424 |     27,438,092 | 12,386,848 |  9,364,356 | 162,771,632 |          - |
+| plain (nearest, no ts)      |      3,438,646 |      2,263,839 |  1,315,079 |  1,122,982 |     785,550 |  1,865,327 |
+| lq (bilinear, with ts)      |      9,982,212 |      4,865,226 |  4,484,023 |  2,646,523 |  17,612,685 | 24,906,379 |
+| hq (bicubic/gauss, with ts) |     34,931,424 |     14,760,398 | 12,386,848 |  9,364,356 | 162,771,632 |          - |
 
 Note that `raqote` doesn't support high quality filtering.