image/draw: increase performances by applying special case if mask is *image.Alpha #46395
Labels
FrozenDueToAge
NeedsInvestigation
Someone must examine and confirm this is a valid issue and not a duplicate of an existing one.
Performance
Comments
seankhliao
added
NeedsInvestigation
|
I'm open in principle to the change, with two comments.
Line 148 in 6ff0ae2 |
|
Thanks for your reply. I thought about adding a private function to replace all the calls to func rgbaAt(x, y int, img image.Image) (r, g, b, a uint32) {
switch img0 := img.(type) {
case *image.RGBA:
off := img0.PixOffset(x, y)
r = uint32(img0.Pix[off])
g = uint32(img0.Pix[off+1])
b = uint32(img0.Pix[off+2])
a = uint32(img0.Pix[off+3])
r |= r << 8
g |= g << 8
b |= b << 8
a |= a << 8
return
case *image.Gray:
off := img0.PixOffset(x, y)
y := uint32(img0.Pix[off])
y |= y << 8
return y, y, y, uint32(0xffff)
case *image.Alpha:
off := img0.PixOffset(x, y)
a := uint32(img0.Pix[off])
a |= a << 8
return a, a, a, a
default:
return img0.At(x, y).RGBA()
}
}I tested it, and the results look very promising. Besides the computation time, it is allocating less memory.
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
We don't want to end up with a function call and type switch per pixel. Better to hoist it out of the loop, hence writing a new top-level function. See also https://go-review.googlesource.com/c/go/+/311129 "image: add RGBA64Image interface". |
|
Change https://golang.org/cl/323749 mentions this issue: |
|
I'd be interested in new "compare to the status quo" benchmark numbers now that "image/draw: add RGBA64Image fast path" https://go-review.googlesource.com/c/go/+/340049 has been submitted. BTW the "delta" column in the tables from your earlier comment has no numbers. |
|
I made a bench comparing the master branch (commit ab7c904) (that holds the change you mention including the The last bench are: package draw
import (
"image"
"testing"
)
func Benchmark_drawRGBA(b *testing.B) {
r := image.Rect(0, 0, 500, 500)
src := image.NewRGBA(r)
dst := image.NewRGBA(r)
mask := image.NewAlpha(r)
for i := 0; i < b.N; i++ {
DrawMask(dst, r, src, image.Point{}, mask, image.Point{}, Over)
}
}
func Benchmark_drawRGBAGray(b *testing.B) {
r := image.Rect(0, 0, 500, 500)
// Colors are defined by Red, Green, Blue, Alpha uint8 values.
src := image.NewGray(r)
dst := image.NewRGBA(r)
mask := image.NewAlpha(r)
for i := 0; i < b.N; i++ {
DrawMask(dst, r, src, image.Point{}, mask, image.Point{}, Over)
}
}I am not sure that is what you want. |
|
Change https://golang.org/cl/351852 mentions this issue: |
|
Thanks for the new numbers. I overlooked a code path in https://golang.org/cl/340049 which should be addressed by https://golang.org/cl/351852 Once 351852 lands, I'd be curious if you could re-run your benchmarks. I'll repeat that the "delta" column in your tables have no numbers, so it's hard to tell which rows I should focus on. |
|
https://golang.org/cl/351852 has landed |
gopherbot
pushed a commit
that referenced
this issue
Sep 27, 2021
This should have been part of https://golang.org/cl/340049 but I overlooked it. That commit added fast path code when the destination image was *not* an *image.RGBA. This commit edits func drawRGBA. name old time/op new time/op delta RGBA1-4 5.11ms ± 1% 1.12ms ± 1% -78.01% (p=0.008 n=5+5) RGBA2-4 8.69ms ± 1% 2.98ms ± 1% -65.77% (p=0.008 n=5+5) Updates #44808. Updates #46395. Change-Id: I899d46d985634fc81ea47ff4f0d436630e8a961c Reviewed-on: https://go-review.googlesource.com/c/go/+/351852 Trust: Nigel Tao <nigeltao@golang.org> Reviewed-by: Dmitri Shuralyov <dmitshur@golang.org>
|
Here are the results of my tests: ❯ cat ~/go.516d75ccf1.bench
goos: darwin
goarch: amd64
pkg: image/draw
cpu: Intel(R) Core(TM) i7-8850H CPU @ 2.60GHz
BenchmarkFillOver-12 1240 919849 ns/op 0 B/op 0 allocs/op
BenchmarkFillSrc-12 50834 24552 ns/op 0 B/op 0 allocs/op
BenchmarkCopyOver-12 1705 699634 ns/op 0 B/op 0 allocs/op
BenchmarkCopySrc-12 55306 20768 ns/op 0 B/op 0 allocs/op
BenchmarkNRGBAOver-12 1422 841424 ns/op 0 B/op 0 allocs/op
BenchmarkNRGBASrc-12 2649 464700 ns/op 0 B/op 0 allocs/op
BenchmarkYCbCr-12 2835 436423 ns/op 0 B/op 0 allocs/op
BenchmarkGray-12 8022 152508 ns/op 0 B/op 0 allocs/op
BenchmarkCMYK-12 2582 461211 ns/op 0 B/op 0 allocs/op
BenchmarkGlyphOver-12 5244 228985 ns/op 0 B/op 0 allocs/op
BenchmarkRGBA1-12 1944 611994 ns/op 0 B/op 0 allocs/op
BenchmarkRGBA2-12 746 1626558 ns/op 0 B/op 0 allocs/op
BenchmarkPalettedFill-12 185510 6629 ns/op 0 B/op 0 allocs/op
BenchmarkPalettedRGBA-12 771 1607542 ns/op 40 B/op 2 allocs/op
BenchmarkGenericOver-12 745 1711477 ns/op 0 B/op 0 allocs/op
BenchmarkGenericMaskOver-12 1329 1115717 ns/op 0 B/op 0 allocs/op
BenchmarkGenericSrc-12 1329 899930 ns/op 0 B/op 0 allocs/op
BenchmarkGenericMaskSrc-12 982 1181788 ns/op 0 B/op 0 allocs/op
Benchmark_drawRGBA-12 355 3524321 ns/op 6392 B/op 0 allocs/op
Benchmark_drawRGBAGray-12 432 2776462 ns/op 3508 B/op 0 allocs/op
PASS
ok image/draw 28.130s❯ cat ~/go.25a0774.bench
goos: darwin
goarch: amd64
pkg: image/draw
cpu: Intel(R) Core(TM) i7-8850H CPU @ 2.60GHz
BenchmarkFillOver-12 1226 937482 ns/op 0 B/op 0 allocs/op
BenchmarkFillSrc-12 50300 24057 ns/op 0 B/op 0 allocs/op
BenchmarkCopyOver-12 1728 688587 ns/op 0 B/op 0 allocs/op
BenchmarkCopySrc-12 56661 21351 ns/op 0 B/op 0 allocs/op
BenchmarkNRGBAOver-12 1413 844938 ns/op 0 B/op 0 allocs/op
BenchmarkNRGBASrc-12 2599 459764 ns/op 0 B/op 0 allocs/op
BenchmarkYCbCr-12 2786 425523 ns/op 0 B/op 0 allocs/op
BenchmarkGray-12 8143 147877 ns/op 0 B/op 0 allocs/op
BenchmarkCMYK-12 2365 469891 ns/op 0 B/op 0 allocs/op
BenchmarkGlyphOver-12 4761 233635 ns/op 0 B/op 0 allocs/op
BenchmarkRGBA1-12 1956 604964 ns/op 0 B/op 0 allocs/op
BenchmarkRGBA2-12 1064 1131638 ns/op 0 B/op 0 allocs/op
BenchmarkPalettedFill-12 182772 6491 ns/op 0 B/op 0 allocs/op
BenchmarkPalettedRGBA-12 786 1536837 ns/op 40 B/op 2 allocs/op
BenchmarkGenericOver-12 775 1532559 ns/op 0 B/op 0 allocs/op
BenchmarkGenericMaskOver-12 1360 884728 ns/op 0 B/op 0 allocs/op
BenchmarkGenericSrc-12 1386 868615 ns/op 0 B/op 0 allocs/op
BenchmarkGenericMaskSrc-12 1164 1060872 ns/op 0 B/op 0 allocs/op
Benchmark_drawRGBA-12 500 2335115 ns/op 4538 B/op 0 allocs/op
Benchmark_drawRGBAGray-12 1375 882834 ns/op 1102 B/op 0 allocs/op
PASS
ok image/draw 26.094s❯ benchstat -delta-test none ~/go.516d75ccf1.bench ~/go.25a0774.bench
name old time/op new time/op delta
FillOver-12 920µs ± 0% 937µs ± 0% +1.92%
FillSrc-12 24.6µs ± 0% 24.1µs ± 0% -2.02%
CopyOver-12 700µs ± 0% 689µs ± 0% -1.58%
CopySrc-12 20.8µs ± 0% 21.4µs ± 0% +2.81%
NRGBAOver-12 841µs ± 0% 845µs ± 0% +0.42%
NRGBASrc-12 465µs ± 0% 460µs ± 0% -1.06%
YCbCr-12 436µs ± 0% 426µs ± 0% -2.50%
Gray-12 153µs ± 0% 148µs ± 0% -3.04%
CMYK-12 461µs ± 0% 470µs ± 0% +1.88%
GlyphOver-12 229µs ± 0% 234µs ± 0% +2.03%
RGBA1-12 612µs ± 0% 605µs ± 0% -1.15%
RGBA2-12 1.63ms ± 0% 1.13ms ± 0% -30.43%
PalettedFill-12 6.63µs ± 0% 6.49µs ± 0% -2.08%
PalettedRGBA-12 1.61ms ± 0% 1.54ms ± 0% -4.40%
GenericOver-12 1.71ms ± 0% 1.53ms ± 0% -10.45%
GenericMaskOver-12 1.12ms ± 0% 0.88ms ± 0% -20.70%
GenericSrc-12 900µs ± 0% 869µs ± 0% -3.48%
GenericMaskSrc-12 1.18ms ± 0% 1.06ms ± 0% -10.23%
_drawRGBA-12 3.52ms ± 0% 2.34ms ± 0% -33.74%
_drawRGBAGray-12 2.78ms ± 0% 0.88ms ± 0% -68.20%
name old alloc/op new alloc/op delta
FillOver-12 0.00B 0.00B 0.00%
FillSrc-12 0.00B 0.00B 0.00%
CopyOver-12 0.00B 0.00B 0.00%
CopySrc-12 0.00B 0.00B 0.00%
NRGBAOver-12 0.00B 0.00B 0.00%
NRGBASrc-12 0.00B 0.00B 0.00%
YCbCr-12 0.00B 0.00B 0.00%
Gray-12 0.00B 0.00B 0.00%
CMYK-12 0.00B 0.00B 0.00%
GlyphOver-12 0.00B 0.00B 0.00%
RGBA1-12 0.00B 0.00B 0.00%
RGBA2-12 0.00B 0.00B 0.00%
PalettedFill-12 0.00B 0.00B 0.00%
PalettedRGBA-12 40.0B ± 0% 40.0B ± 0% 0.00%
GenericOver-12 0.00B 0.00B 0.00%
GenericMaskOver-12 0.00B 0.00B 0.00%
GenericSrc-12 0.00B 0.00B 0.00%
GenericMaskSrc-12 0.00B 0.00B 0.00%
_drawRGBA-12 6.39kB ± 0% 4.54kB ± 0% -29.01%
_drawRGBAGray-12 3.51kB ± 0% 1.10kB ± 0% -68.59%
name old allocs/op new allocs/op delta
FillOver-12 0.00 0.00 0.00%
FillSrc-12 0.00 0.00 0.00%
CopyOver-12 0.00 0.00 0.00%
CopySrc-12 0.00 0.00 0.00%
NRGBAOver-12 0.00 0.00 0.00%
NRGBASrc-12 0.00 0.00 0.00%
YCbCr-12 0.00 0.00 0.00%
Gray-12 0.00 0.00 0.00%
CMYK-12 0.00 0.00 0.00%
GlyphOver-12 0.00 0.00 0.00%
RGBA1-12 0.00 0.00 0.00%
RGBA2-12 0.00 0.00 0.00%
PalettedFill-12 0.00 0.00 0.00%
PalettedRGBA-12 2.00 ± 0% 2.00 ± 0% 0.00%
GenericOver-12 0.00 0.00 0.00%
GenericMaskOver-12 0.00 0.00 0.00%
GenericSrc-12 0.00 0.00 0.00%
GenericMaskSrc-12 0.00 0.00 0.00%
_drawRGBA-12 0.00 0.00 0.00%
_drawRGBAGray-12 0.00 0.00 0.00% |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
What version of Go are you using (
go version)?Does this issue reproduce with the latest release?
yes
What operating system and processor architecture are you using (
go env)?go envOutputWhat did you do?
I am building a tool that is manipulating a lot of images. The process is a bit slow.
I made some performance analysis, and it appears that the
drawRGBAmethod is really time-consuming.Specially the call to
mask.At(mx,my).RGBA()is causing performance penalty.I think that most of the use cases of
DrawMaskfunction are using an*image.Alphastructure as a mask. Therefore doing a special treatment if the mask is*image.Alphacould be valuable for most of the usage.I made a simple test and bench for the
drawRGBAmethod (it should be done for other methods as well).Then I made a patch as an experiment:
The performance enhancement is not negligible:
This is a test for opening the discussion, and I guess that other controls should be added as it makes some tests of the draw package fail. Specifically, the
test.maskcauses panic, and I cannot figure out whattest.maskis nor where it is defined.What did you expect to see?
N/A
What did you see instead?
N/A
The text was updated successfully, but these errors were encountered: