When DstMask is not nil, this shortcut causes stack overflow because
Copy function in turn will call Scale with same dr and sr.
Fixesgolang/go#23107
Change-Id: I8ccadbd9b7f16363ac17b6114308527d6fa9456e
Reviewed-on: https://go-review.googlesource.com/83537
Reviewed-by: Nigel Tao <nigeltao@golang.org>
Run-TryBot: Nigel Tao <nigeltao@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This is the golang.org/x equivalent of the golang.org/cl/36732 change to
the standard library.
Change-Id: I71c09a72e24c8e1c013769a6fed8f9a031724d9d
Reviewed-on: https://go-review.googlesource.com/36733
Run-TryBot: Nigel Tao <nigeltao@golang.org>
Reviewed-by: Rob Pike <r@golang.org>
This is only for the NearestNeighbor and ApproxBiLinear Interpolators. A
Kernel interpolator will add blur even when the dst and src rectangles
are the same size.
We do not bother recognizing Transforms that are Scales. The performance
difference is minimal, as you still need to do a per-dst-pixel inverse
mapping either way.
benchmark old ns/op new ns/op delta
BenchmarkSimpleScaleCopy-8 4866297 29586 -99.39%
BenchmarkSimpleTransformCopy-8 4875991 29531 -99.39%
BenchmarkSimpleTransformScale-8 1208147 1223206 +1.25%
Change-Id: If649ad27a4e81bcbb24b18315745c02c9186a5b7
Reviewed-on: https://go-review.googlesource.com/13004
Reviewed-by: Rob Pike <r@golang.org>
It's more args (in terms of bytes), but fewer allocations (Transformer
is an interface). Either way, it's not really that big of a deal, but
the value instead of the pointer seems conceptually more correct.
Change-Id: Ibea76da17cbda0d9633110fd56044b4e2c690e81
Reviewed-on: https://go-review.googlesource.com/12669
Reviewed-by: Rob Pike <r@golang.org>
This is the golang.org/x equivalent of the golang.org/cl/12220 change to
the standard library.
Change-Id: I220d1398902dec374472ef5b08db8501607cc971
Reviewed-on: https://go-review.googlesource.com/12222
Reviewed-by: Rob Pike <r@golang.org>
This will make the upcoming "implement masks" change easier.
Change-Id: I2b316a0b25f98779d9cd076e64266c617464a08e
Reviewed-on: https://go-review.googlesource.com/8999
Reviewed-by: Rob Pike <r@golang.org>
It'll only be used by the fallback code paths, but it's easiest if all
code paths take the extra argument.
Change-Id: I09c4067df6af7c4cd270e8ffcad96d1d8af29057
Reviewed-on: https://go-review.googlesource.com/8909
Reviewed-by: Rob Pike <r@golang.org>
This change only adds the fields, more or less. Follow-up changes will
actually honor the masks.
Change-Id: I81411dc1aac4b3c846dcdf13e2cb0b5cd60fb2b4
Reviewed-on: https://go-review.googlesource.com/8902
Reviewed-by: Rob Pike <r@golang.org>
The raw computation can produce red > alpha when some weights are
negative.
Change-Id: Ic6701354770f012d3ef21a390a8400e14e9d1e25
Reviewed-on: https://go-review.googlesource.com/8740
Reviewed-by: Rob Pike <r@golang.org>
The two Ops are equivalent, but Src is faster.
Change-Id: I2c73a13755047c224c71fb5af786875f02681de9
Reviewed-on: https://go-review.googlesource.com/8640
Reviewed-by: Rob Pike <r@golang.org>
This change just does the mechanical codegen for the op == Over cases.
The actual compositing operator is still effectively Src. Fixing that,
which is less mechanical, will be a follow-up change.
Change-Id: I87805114d49e7ce7087066187a2f4c722a883c01
Reviewed-on: https://go-review.googlesource.com/8524
Reviewed-by: Rob Pike <r@golang.org>
This change only *prepares* the codegen to handle multiple Ops. The
actual generated code still only supports one Op (Src) and not the other
(Over). A follow-up change will add Over.
This Op switch (an eventual x2 multiplier in the amount of code
generated) should be the last of the codegen LoC multipliers. The dst
and src mask options will be implemented in the slow path fallback.
Change-Id: Iecbcc6fad063e2aac36d78d5380c0a0947c709df
Reviewed-on: https://go-review.googlesource.com/8488
Reviewed-by: Rob Pike <r@golang.org>
benchmark old ns/op new ns/op delta
BenchmarkScaleBLLargeDown 257715146 260286012 +1.00%
BenchmarkScaleCRLargeDown 426797448 430078734 +0.77%
BenchmarkScaleBLDown 4449939 4222542 -5.11%
BenchmarkScaleCRDown 8160446 8010056 -1.84%
BenchmarkScaleBLUp 22290312 21044122 -5.59%
BenchmarkScaleCRUp 33010722 32021468 -3.00%
BenchmarkScaleCRSrcGray 13307961 13020192 -2.16%
BenchmarkScaleCRSrcNRGBA 40567431 40801939 +0.58%
BenchmarkScaleCRSrcRGBA 39892971 40240558 +0.87%
BenchmarkScaleCRSrcYCbCr 59020222 59686699 +1.13%
benchmark old allocs new allocs delta
BenchmarkScaleBLLargeDown 1 1 +0.00%
BenchmarkScaleCRLargeDown 1 2 +100.00%
BenchmarkScaleBLDown 1 0 -100.00%
BenchmarkScaleCRDown 1 0 -100.00%
BenchmarkScaleBLUp 1 0 -100.00%
BenchmarkScaleCRUp 1 0 -100.00%
BenchmarkScaleCRSrcGray 1 0 -100.00%
BenchmarkScaleCRSrcNRGBA 1 0 -100.00%
BenchmarkScaleCRSrcRGBA 1 0 -100.00%
BenchmarkScaleCRSrcYCbCr 1 0 -100.00%
benchmark old bytes new bytes delta
BenchmarkScaleBLLargeDown 14745600 2949200 -80.00%
BenchmarkScaleCRLargeDown 14745600 4915333 -66.67%
BenchmarkScaleBLDown 1523712 5079 -99.67%
BenchmarkScaleCRDown 1523712 7619 -99.50%
BenchmarkScaleBLUp 10117120 101175 -99.00%
BenchmarkScaleCRUp 10117120 202350 -98.00%
BenchmarkScaleCRSrcGray 4915200 49156 -99.00%
BenchmarkScaleCRSrcNRGBA 4915200 163853 -96.67%
BenchmarkScaleCRSrcRGBA 4915200 163853 -96.67%
BenchmarkScaleCRSrcYCbCr 4915200 245780 -95.00%
The increase in BenchmarkScale??LargeDown number of allocs I think is an
accounting error due to the low number of iterations: a low denominator.
I suspect that there are one or two extra allocs up front for using the
sync.Pool, but one fewer alloc per iteration. The number of iterations
is only 5 for BL and 3 for CR, for the default timeout. If I increase
the -test.benchtime value to 5s, then the reported average (allocs/op)
drop from 2 to 0, so the delta should actually be -100% instead of +0 or
+100%.
Change-Id: I21d9bb0086bdb25517b6a430e8a21bdf3db026f6
Reviewed-on: https://go-review.googlesource.com/8150
Reviewed-by: Rob Pike <r@golang.org>
Only the YCbCr benchmarks show significant changes. The other benchmark
changes look noisy. The PixOffset and YOffset calls were previously
already inlined by the gc compiler. COffset was different because it's
more complicated than YOffset, and the switch inside the COffset body is
redundant when you already know the src image sratio.
http://golang.org/src/image/ycbcr.go?s=2377:2414#L77
benchmark old ns/op new ns/op delta
BenchmarkScaleLargeDownNN 1037504 908236 -12.46%
BenchmarkScaleLargeDownAB 3196568 2735776 -14.42%
BenchmarkScaleLargeDownBL 357165552 311463393 -12.80%
BenchmarkScaleLargeDownCR 649403305 544985134 -16.08%
BenchmarkScaleSrcYCbCr 3204063 2699147 -15.76%
BenchmarkTformABSrcYCbCr 2155142 1968540 -8.66%
BenchmarkTformCRSrcYCbCr 11672625 9865358 -15.48%
Change-Id: Ifa109363a1282ab114b2fdb0b577dcafef927333
Reviewed-on: https://go-review.googlesource.com/7880
Reviewed-by: Rob Pike <r@golang.org>
benchmark old ns/op new ns/op delta
BenchmarkTformCRSrcGray 6111610 5344117 -12.56%
BenchmarkTformCRSrcNRGBA 62070281 59295178 -4.47%
BenchmarkTformCRSrcRGBA 13840290 10612547 -23.32%
BenchmarkTformCRSrcUniform 591637 587621 -0.68%
BenchmarkTformCRSrcYCbCr 72219184 69404747 -3.90%
As of current origin/master, Gray and RGBA have fast paths but the other src
image types do not. They have more fat, so the relative improvement is smaller.
Change-Id: Ibbae91cd3cb3c139efb1dcc8fda1cb6432505189
Reviewed-on: https://go-review.googlesource.com/7794
Reviewed-by: Rob Pike <r@golang.org>
These code paths aren't actually fast yet. That will be a follow-up
change.
Change-Id: I814992573cc6af422e49d0ddf336003e662897a5
Reviewed-on: https://go-review.googlesource.com/7791
Reviewed-by: Rob Pike <r@golang.org>
Also fix the NN and ABL fast paths to only apply if we can access the
Pix elements without src-bounds checking.
Change-Id: Ie9fc96b28e0665df49d00c4c53cb81385faee4db
Reviewed-on: https://go-review.googlesource.com/7675
Reviewed-by: Rob Pike <r@golang.org>
I simply forgot to do this in the previous change.
Change-Id: I0ea5d6bb0a09154cb995a665476ffceb155bd49e
Reviewed-on: https://go-review.googlesource.com/7651
Reviewed-by: Rob Pike <r@golang.org>
Just stub implementations for now. Actual implementations will be
follow-up changes.
Change-Id: Id21d9042a2073c2dc0f78c9977c4940f000a41df
Reviewed-on: https://go-review.googlesource.com/6805
Reviewed-by: Rob Pike <r@golang.org>
This means that only Kernel values have a NewScaler method, which
re-uses computation when scaling multiple images of the same dst and src
dimensions. The NearestNeighbor and ApproxBiLinear scalers don't get any
pre-computation to re-use, so don't need a NewScaler method just to
satisfy the previous Interpolator interface. As a small bonus, NN.Scale
and ABL.Scale should no longer allocate on the fast paths.
This change is consistent the upcoming Transformer method, so that the
Interpolator interface will be
type Interpolator interface {
Scale(etc)
Transform(etc)
}
instead of
type Interpolator interface {
NewScaler(etc) Scaler
Transform(etc)
}
I don't have a good theory for why the "func (ablInterpolator)
scale_RGBA_RGBA" benchmark is such a dramatic improvement, but at least
it's in the right direction. I'm calling the other benchmark changes as
noise.
benchmark old ns/op new ns/op delta
BenchmarkScaleLargeDownNN 3233406 3169060 -1.99%
BenchmarkScaleLargeDownAB 12018178 12011348 -0.06%
BenchmarkScaleLargeDownBL 1420827834 1409335695 -0.81%
BenchmarkScaleLargeDownCR 2820669690 2795534035 -0.89%
BenchmarkScaleDownNN 866628 869241 +0.30%
BenchmarkScaleDownAB 3175963 3216041 +1.26%
BenchmarkScaleDownBL 26639767 26677003 +0.14%
BenchmarkScaleDownCR 51720996 51621628 -0.19%
BenchmarkScaleUpNN 42758485 43258611 +1.17%
BenchmarkScaleUpAB 156693813 156943367 +0.16%
BenchmarkScaleUpBL 69511444 69621698 +0.16%
BenchmarkScaleUpCR 124530191 124885601 +0.29%
BenchmarkScaleSrcGray 8992205 9129321 +1.52%
BenchmarkScaleSrcNRGBA 9807837 9894466 +0.88%
BenchmarkScaleSrcRGBA 1333188 1104282 -17.17%
BenchmarkScaleSrcUniform 1147788 1162488 +1.28%
BenchmarkScaleSrcYCbCr 12164542 12305373 +1.16%
Change-Id: I2aee6c392eb7437e843260775aed97ce145b4d47
Reviewed-on: https://go-review.googlesource.com/6556
Reviewed-by: Rob Pike <r@golang.org>
This will lessen the diff in a follow-up change.
Change-Id: I67257b935e4b5fc6fcc462c83910a3f09f26702f
Reviewed-on: https://go-review.googlesource.com/6554
Reviewed-by: David Symonds <dsymonds@golang.org>
Optimizing this code path will be follow-up changes.
Change-Id: Ifece6c7d3685bb8f2d20cf8828a121c9ff346434
Reviewed-on: https://go-review.googlesource.com/6238
Reviewed-by: Rob Pike <r@golang.org>
Adding fast paths for concrete dst and src image types will be follow-up
changes.
Change-Id: Ib23e3cc46d9c7a52ebd5eddfbad09177b0b421af
Reviewed-on: https://go-review.googlesource.com/5417
Reviewed-by: Rob Pike <r@golang.org>
This is just a copy/paste/rename in anticipation of automatically
generating fast-path code for specific dst-image types and src-image
types, a la the standard library's image/draw package.
Change-Id: If6385b347cb5e1d134b0ce841ef38ed4450a6fef
Reviewed-on: https://go-review.googlesource.com/5252
Reviewed-by: David Symonds <dsymonds@golang.org>
gutmet
Jiulong Wang
Nigel Tao