golang-image/draw/scale.go

526 lines
15 KiB
Go
Raw Permalink Normal View History

// Copyright 2015 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
//go:generate go run gen.go
package draw
import (
"image"
"image/color"
"math"
draw: use a sync.Pool for kernel scaling's temporary buffers. 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>
2015-03-27 01:21:16 +01:00
"sync"
2018-12-06 17:03:52 +01:00
"git.gutmet.org/golang-image.git/math/f64"
)
// Copy copies the part of the source image defined by src and sr and writes
// the result of a Porter-Duff composition to the part of the destination image
// defined by dst and the translation of sr so that sr.Min translates to dp.
func Copy(dst Image, dp image.Point, src image.Image, sr image.Rectangle, op Op, opts *Options) {
var o Options
if opts != nil {
o = *opts
}
dr := sr.Add(dp.Sub(sr.Min))
if o.DstMask == nil {
DrawMask(dst, dr, src, sr.Min, o.SrcMask, o.SrcMaskP.Add(sr.Min), op)
} else {
NearestNeighbor.Scale(dst, dr, src, sr, op, opts)
}
}
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
// Scaler scales the part of the source image defined by src and sr and writes
// the result of a Porter-Duff composition to the part of the destination image
// defined by dst and dr.
//
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
// A Scaler is safe to use concurrently.
type Scaler interface {
Scale(dst Image, dr image.Rectangle, src image.Image, sr image.Rectangle, op Op, opts *Options)
}
// Transformer transforms the part of the source image defined by src and sr
// and writes the result of a Porter-Duff composition to the part of the
// destination image defined by dst and the affine transform m applied to sr.
//
// For example, if m is the matrix
//
// m00 m01 m02
// m10 m11 m12
//
// then the src-space point (sx, sy) maps to the dst-space point
// (m00*sx + m01*sy + m02, m10*sx + m11*sy + m12).
//
// A Transformer is safe to use concurrently.
type Transformer interface {
Transform(dst Image, m f64.Aff3, src image.Image, sr image.Rectangle, op Op, opts *Options)
}
// Options are optional parameters to Copy, Scale and Transform.
//
// A nil *Options means to use the default (zero) values of each field.
type Options struct {
// Masks limit what parts of the dst image are drawn to and what parts of
// the src image are drawn from.
//
// A dst or src mask image having a zero alpha (transparent) pixel value in
// the respective coordinate space means that dst pixel is entirely
// unaffected or that src pixel is considered transparent black. A full
// alpha (opaque) value means that the dst pixel is maximally affected or
// the src pixel contributes maximally. The default values, nil, are
// equivalent to fully opaque, infinitely large mask images.
//
// The DstMask is otherwise known as a clip mask, and its pixels map 1:1 to
// the dst image's pixels. DstMaskP in DstMask space corresponds to
// image.Point{X:0, Y:0} in dst space. For example, when limiting
// repainting to a 'dirty rectangle', use that image.Rectangle and a zero
// image.Point as the DstMask and DstMaskP.
//
// The SrcMask's pixels map 1:1 to the src image's pixels. SrcMaskP in
// SrcMask space corresponds to image.Point{X:0, Y:0} in src space. For
// example, when drawing font glyphs in a uniform color, use an
// *image.Uniform as the src, and use the glyph atlas image and the
// per-glyph offset as SrcMask and SrcMaskP:
// Copy(dst, dp, image.NewUniform(color), image.Rect(0, 0, glyphWidth, glyphHeight), &Options{
// SrcMask: glyphAtlas,
// SrcMaskP: glyphOffset,
// })
DstMask image.Image
DstMaskP image.Point
SrcMask image.Image
SrcMaskP image.Point
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
// TODO: a smooth vs sharp edges option, for arbitrary rotations?
}
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
// Interpolator is an interpolation algorithm, when dst and src pixels don't
// have a 1:1 correspondence.
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
//
// Of the interpolators provided by this package:
// - NearestNeighbor is fast but usually looks worst.
// - CatmullRom is slow but usually looks best.
// - ApproxBiLinear has reasonable speed and quality.
//
// The time taken depends on the size of dr. For kernel interpolators, the
// speed also depends on the size of sr, and so are often slower than
// non-kernel interpolators, especially when scaling down.
type Interpolator interface {
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
Scaler
Transformer
}
// Kernel is an interpolator that blends source pixels weighted by a symmetric
// kernel function.
type Kernel struct {
// Support is the kernel support and must be >= 0. At(t) is assumed to be
// zero when t >= Support.
Support float64
// At is the kernel function. It will only be called with t in the
// range [0, Support).
At func(t float64) float64
}
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
// Scale implements the Scaler interface.
func (q *Kernel) Scale(dst Image, dr image.Rectangle, src image.Image, sr image.Rectangle, op Op, opts *Options) {
q.newScaler(dr.Dx(), dr.Dy(), sr.Dx(), sr.Dy(), false).Scale(dst, dr, src, sr, op, opts)
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
}
// NewScaler returns a Scaler that is optimized for scaling multiple times with
// the same fixed destination and source width and height.
func (q *Kernel) NewScaler(dw, dh, sw, sh int) Scaler {
draw: use a sync.Pool for kernel scaling's temporary buffers. 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>
2015-03-27 01:21:16 +01:00
return q.newScaler(dw, dh, sw, sh, true)
}
func (q *Kernel) newScaler(dw, dh, sw, sh int, usePool bool) Scaler {
z := &kernelScaler{
kernel: q,
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
dw: int32(dw),
dh: int32(dh),
sw: int32(sw),
sh: int32(sh),
horizontal: newDistrib(q, int32(dw), int32(sw)),
vertical: newDistrib(q, int32(dh), int32(sh)),
}
draw: use a sync.Pool for kernel scaling's temporary buffers. 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>
2015-03-27 01:21:16 +01:00
if usePool {
z.pool.New = func() interface{} {
tmp := z.makeTmpBuf()
return &tmp
}
}
return z
}
var (
// NearestNeighbor is the nearest neighbor interpolator. It is very fast,
// but usually gives very low quality results. When scaling up, the result
// will look 'blocky'.
NearestNeighbor = Interpolator(nnInterpolator{})
// ApproxBiLinear is a mixture of the nearest neighbor and bi-linear
// interpolators. It is fast, but usually gives medium quality results.
//
// It implements bi-linear interpolation when upscaling and a bi-linear
// blend of the 4 nearest neighbor pixels when downscaling. This yields
// nicer quality than nearest neighbor interpolation when upscaling, but
// the time taken is independent of the number of source pixels, unlike the
// bi-linear interpolator. When downscaling a large image, the performance
// difference can be significant.
ApproxBiLinear = Interpolator(ablInterpolator{})
// BiLinear is the tent kernel. It is slow, but usually gives high quality
// results.
BiLinear = &Kernel{1, func(t float64) float64 {
return 1 - t
}}
// CatmullRom is the Catmull-Rom kernel. It is very slow, but usually gives
// very high quality results.
//
// It is an instance of the more general cubic BC-spline kernel with parameters
// B=0 and C=0.5. See Mitchell and Netravali, "Reconstruction Filters in
// Computer Graphics", Computer Graphics, Vol. 22, No. 4, pp. 221-228.
CatmullRom = &Kernel{2, func(t float64) float64 {
if t < 1 {
return (1.5*t-2.5)*t*t + 1
}
return ((-0.5*t+2.5)*t-4)*t + 2
}}
// TODO: a Kaiser-Bessel kernel?
)
type nnInterpolator struct{}
type ablInterpolator struct{}
type kernelScaler struct {
draw: make Scale an Interpolator method instead of a function. 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>
2015-03-03 06:54:53 +01:00
kernel *Kernel
dw, dh, sw, sh int32
horizontal, vertical distrib
draw: use a sync.Pool for kernel scaling's temporary buffers. 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>
2015-03-27 01:21:16 +01:00
pool sync.Pool
}
func (z *kernelScaler) makeTmpBuf() [][4]float64 {
return make([][4]float64, z.dw*z.sh)
}
// source is a range of contribs, their inverse total weight, and that ITW
// divided by 0xffff.
type source struct {
i, j int32
invTotalWeight float64
invTotalWeightFFFF float64
}
// contrib is the weight of a column or row.
type contrib struct {
coord int32
weight float64
}
// distrib measures how source pixels are distributed over destination pixels.
type distrib struct {
// sources are what contribs each column or row in the source image owns,
// and the total weight of those contribs.
sources []source
// contribs are the contributions indexed by sources[s].i and sources[s].j.
contribs []contrib
}
// newDistrib returns a distrib that distributes sw source columns (or rows)
// over dw destination columns (or rows).
func newDistrib(q *Kernel, dw, sw int32) distrib {
scale := float64(sw) / float64(dw)
halfWidth, kernelArgScale := q.Support, 1.0
// When shrinking, broaden the effective kernel support so that we still
// visit every source pixel.
if scale > 1 {
halfWidth *= scale
kernelArgScale = 1 / scale
}
// Make the sources slice, one source for each column or row, and temporarily
// appropriate its elements' fields so that invTotalWeight is the scaled
// coordinate of the source column or row, and i and j are the lower and
// upper bounds of the range of destination columns or rows affected by the
// source column or row.
n, sources := int32(0), make([]source, dw)
for x := range sources {
center := (float64(x)+0.5)*scale - 0.5
i := int32(math.Floor(center - halfWidth))
if i < 0 {
i = 0
}
j := int32(math.Ceil(center + halfWidth))
if j > sw {
j = sw
if j < i {
j = i
}
}
sources[x] = source{i: i, j: j, invTotalWeight: center}
n += j - i
}
contribs := make([]contrib, 0, n)
for k, b := range sources {
totalWeight := 0.0
l := int32(len(contribs))
for coord := b.i; coord < b.j; coord++ {
t := abs((b.invTotalWeight - float64(coord)) * kernelArgScale)
if t >= q.Support {
continue
}
weight := q.At(t)
if weight == 0 {
continue
}
totalWeight += weight
contribs = append(contribs, contrib{coord, weight})
}
totalWeight = 1 / totalWeight
sources[k] = source{
i: l,
j: int32(len(contribs)),
invTotalWeight: totalWeight,
invTotalWeightFFFF: totalWeight / 0xffff,
}
}
return distrib{sources, contribs}
}
// abs is like math.Abs, but it doesn't care about negative zero, infinities or
// NaNs.
func abs(f float64) float64 {
if f < 0 {
f = -f
}
return f
}
// ftou converts the range [0.0, 1.0] to [0, 0xffff].
func ftou(f float64) uint16 {
i := int32(0xffff*f + 0.5)
if i > 0xffff {
return 0xffff
}
if i > 0 {
return uint16(i)
}
return 0
}
// fffftou converts the range [0.0, 65535.0] to [0, 0xffff].
func fffftou(f float64) uint16 {
i := int32(f + 0.5)
if i > 0xffff {
return 0xffff
}
if i > 0 {
return uint16(i)
}
return 0
}
// invert returns the inverse of m.
//
// TODO: move this into the f64 package, once we work out the convention for
// matrix methods in that package: do they modify the receiver, take a dst
// pointer argument, or return a new value?
func invert(m *f64.Aff3) f64.Aff3 {
m00 := +m[3*1+1]
m01 := -m[3*0+1]
m02 := +m[3*1+2]*m[3*0+1] - m[3*1+1]*m[3*0+2]
m10 := -m[3*1+0]
m11 := +m[3*0+0]
m12 := +m[3*1+0]*m[3*0+2] - m[3*1+2]*m[3*0+0]
det := m00*m11 - m10*m01
return f64.Aff3{
m00 / det,
m01 / det,
m02 / det,
m10 / det,
m11 / det,
m12 / det,
}
}
func matMul(p, q *f64.Aff3) f64.Aff3 {
return f64.Aff3{
p[3*0+0]*q[3*0+0] + p[3*0+1]*q[3*1+0],
p[3*0+0]*q[3*0+1] + p[3*0+1]*q[3*1+1],
p[3*0+0]*q[3*0+2] + p[3*0+1]*q[3*1+2] + p[3*0+2],
p[3*1+0]*q[3*0+0] + p[3*1+1]*q[3*1+0],
p[3*1+0]*q[3*0+1] + p[3*1+1]*q[3*1+1],
p[3*1+0]*q[3*0+2] + p[3*1+1]*q[3*1+2] + p[3*1+2],
}
}
// transformRect returns a rectangle dr that contains sr transformed by s2d.
func transformRect(s2d *f64.Aff3, sr *image.Rectangle) (dr image.Rectangle) {
ps := [...]image.Point{
{sr.Min.X, sr.Min.Y},
{sr.Max.X, sr.Min.Y},
{sr.Min.X, sr.Max.Y},
{sr.Max.X, sr.Max.Y},
}
for i, p := range ps {
sxf := float64(p.X)
syf := float64(p.Y)
dx := int(math.Floor(s2d[0]*sxf + s2d[1]*syf + s2d[2]))
dy := int(math.Floor(s2d[3]*sxf + s2d[4]*syf + s2d[5]))
// The +1 adjustments below are because an image.Rectangle is inclusive
// on the low end but exclusive on the high end.
if i == 0 {
dr = image.Rectangle{
Min: image.Point{dx + 0, dy + 0},
Max: image.Point{dx + 1, dy + 1},
}
continue
}
if dr.Min.X > dx {
dr.Min.X = dx
}
dx++
if dr.Max.X < dx {
dr.Max.X = dx
}
if dr.Min.Y > dy {
dr.Min.Y = dy
}
dy++
if dr.Max.Y < dy {
dr.Max.Y = dy
}
}
return dr
}
func clipAffectedDestRect(adr image.Rectangle, dstMask image.Image, dstMaskP image.Point) (image.Rectangle, image.Image) {
if dstMask == nil {
return adr, nil
}
draw: re-enable the image.Rectangle DstMask fast path This change reverts the remaining part of CL 9463, the image.Rectangle DstMask fast path, since Go 1.5 has been released. $ go test -bench=. -count=5 -timeout=1h > before.txt $ # apply change $ go test -bench=. -count=5 -timeout=1h > after.txt $ benchstat before.txt after.txt name old time/op new time/op delta SimpleScaleCopy-8 23.4µs ± 3% 21.8µs ± 4% -7.14% (p=0.016 n=5+5) SimpleTransformCopy-8 22.6µs ± 6% 21.4µs ± 1% -5.07% (p=0.008 n=5+5) SimpleTransformScale-8 904µs ± 0% 900µs ± 0% ~ (p=0.056 n=5+5) ScaleNNLargeDown-8 641µs ± 1% 643µs ± 1% ~ (p=0.548 n=5+5) ScaleABLargeDown-8 1.50ms ± 1% 1.50ms ± 1% ~ (p=0.413 n=5+4) ScaleBLLargeDown-8 192ms ± 5% 196ms ± 4% ~ (p=0.095 n=5+5) ScaleCRLargeDown-8 355ms ± 4% 350ms ± 2% ~ (p=0.690 n=5+5) ScaleNNDown-8 134µs ± 2% 126µs ± 1% -5.68% (p=0.008 n=5+5) ScaleABDown-8 261µs ± 3% 256µs ± 1% -1.92% (p=0.016 n=5+5) ScaleBLDown-8 2.42ms ± 3% 2.44ms ± 8% ~ (p=0.841 n=5+5) ScaleCRDown-8 4.36ms ± 1% 4.85ms ± 4% +11.22% (p=0.008 n=5+5) ScaleNNUp-8 6.27ms ± 2% 6.72ms ± 3% +7.20% (p=0.008 n=5+5) ScaleABUp-8 12.8ms ± 1% 13.8ms ± 2% +7.26% (p=0.008 n=5+5) ScaleBLUp-8 17.5ms ± 7% 17.4ms ± 2% ~ (p=1.000 n=5+5) ScaleCRUp-8 23.2ms ± 2% 23.1ms ± 4% ~ (p=0.690 n=5+5) ScaleNNSrcRGBA-8 419µs ± 1% 421µs ± 3% ~ (p=0.690 n=5+5) ScaleNNSrcUnif-8 3.35µs ± 3% 3.35µs ± 2% ~ (p=1.000 n=5+5) ScaleNNOverRGBA-8 470µs ± 1% 479µs ± 5% ~ (p=0.548 n=5+5) ScaleNNOverUnif-8 102µs ± 2% 102µs ± 1% ~ (p=0.222 n=5+5) TformNNSrcRGBA-8 215µs ± 1% 213µs ± 1% ~ (p=0.222 n=5+5) TformNNSrcUnif-8 84.2µs ± 1% 86.6µs ± 2% +2.82% (p=0.016 n=5+5) TformNNOverRGBA-8 351µs ± 5% 359µs ± 0% ~ (p=0.151 n=5+5) TformNNOverUnif-8 101µs ± 1% 104µs ± 1% +3.47% (p=0.008 n=5+5) ScaleABSrcGray-8 399µs ± 2% 400µs ± 2% ~ (p=0.690 n=5+5) ScaleABSrcNRGBA-8 795µs ± 1% 809µs ± 4% ~ (p=0.095 n=5+5) ScaleABSrcRGBA-8 807µs ± 1% 807µs ± 2% ~ (p=1.000 n=5+5) ScaleABSrcYCbCr-8 1.50ms ± 1% 1.48ms ± 0% -0.76% (p=0.032 n=5+5) ScaleABOverGray-8 396µs ± 0% 395µs ± 1% ~ (p=0.421 n=5+5) ScaleABOverNRGBA-8 886µs ± 1% 885µs ± 1% ~ (p=0.556 n=4+5) ScaleABOverRGBA-8 893µs ± 1% 895µs ± 1% ~ (p=0.310 n=5+5) ScaleABOverYCbCr-8 1.49ms ± 1% 1.49ms ± 1% ~ (p=0.548 n=5+5) TformABSrcGray-8 359µs ± 1% 364µs ± 3% ~ (p=0.421 n=5+5) TformABSrcNRGBA-8 618µs ± 1% 660µs ± 3% +6.81% (p=0.008 n=5+5) TformABSrcRGBA-8 614µs ± 1% 638µs ± 5% +3.94% (p=0.008 n=5+5) TformABSrcYCbCr-8 903µs ± 1% 913µs ± 2% ~ (p=0.056 n=5+5) TformABOverGray-8 362µs ± 3% 358µs ± 1% ~ (p=0.421 n=5+5) TformABOverNRGBA-8 691µs ± 1% 690µs ± 1% ~ (p=0.841 n=5+5) TformABOverRGBA-8 671µs ± 1% 673µs ± 1% ~ (p=0.690 n=5+5) TformABOverYCbCr-8 904µs ± 1% 906µs ± 0% ~ (p=0.190 n=5+4) ScaleCRSrcGray-8 9.24ms ± 1% 9.16ms ± 1% ~ (p=0.222 n=5+5) ScaleCRSrcNRGBA-8 21.9ms ± 2% 21.7ms ± 1% ~ (p=0.222 n=5+5) ScaleCRSrcRGBA-8 22.5ms ± 8% 21.8ms ± 1% ~ (p=1.000 n=5+5) ScaleCRSrcYCbCr-8 43.6ms ± 1% 43.9ms ± 1% ~ (p=0.222 n=5+5) ScaleCROverGray-8 9.24ms ± 0% 9.35ms ± 6% ~ (p=0.690 n=5+5) ScaleCROverNRGBA-8 21.9ms ± 1% 21.8ms ± 1% ~ (p=0.548 n=5+5) ScaleCROverRGBA-8 21.8ms ± 0% 22.0ms ± 1% +0.94% (p=0.008 n=5+5) ScaleCROverYCbCr-8 43.5ms ± 1% 44.1ms ± 1% +1.53% (p=0.008 n=5+5) TformCRSrcGray-8 3.16ms ± 4% 3.06ms ± 1% ~ (p=0.056 n=5+5) TformCRSrcNRGBA-8 4.21ms ± 1% 4.20ms ± 1% ~ (p=0.841 n=5+5) TformCRSrcRGBA-8 4.29ms ± 1% 4.29ms ± 1% ~ (p=0.841 n=5+5) TformCRSrcYCbCr-8 5.55ms ± 1% 5.58ms ± 0% +0.67% (p=0.032 n=5+4) TformCROverGray-8 3.08ms ± 1% 3.07ms ± 2% ~ (p=0.421 n=5+5) TformCROverNRGBA-8 4.40ms ± 1% 4.39ms ± 0% ~ (p=0.841 n=5+5) TformCROverRGBA-8 4.50ms ± 5% 4.44ms ± 0% ~ (p=0.730 n=5+4) TformCROverYCbCr-8 5.57ms ± 0% 5.61ms ± 1% ~ (p=0.095 n=5+5) Change-Id: I981861c28e103b68275a82e051d2999a8e714502 Reviewed-on: https://go-review.googlesource.com/c/148575 Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
2018-11-09 00:34:23 +01:00
if r, ok := dstMask.(image.Rectangle); ok {
return adr.Intersect(r.Sub(dstMaskP)), nil
}
// TODO: clip to dstMask.Bounds() if the color model implies that out-of-bounds means 0 alpha?
return adr, dstMask
}
func transform_Uniform(dst Image, dr, adr image.Rectangle, d2s *f64.Aff3, src *image.Uniform, sr image.Rectangle, bias image.Point, op Op) {
switch op {
case Over:
switch dst := dst.(type) {
case *image.RGBA:
pr, pg, pb, pa := src.C.RGBA()
pa1 := (0xffff - pa) * 0x101
for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
dyf := float64(dr.Min.Y+int(dy)) + 0.5
d := dst.PixOffset(dr.Min.X+adr.Min.X, dr.Min.Y+int(dy))
for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx, d = dx+1, d+4 {
dxf := float64(dr.Min.X+int(dx)) + 0.5
sx0 := int(d2s[0]*dxf+d2s[1]*dyf+d2s[2]) + bias.X
sy0 := int(d2s[3]*dxf+d2s[4]*dyf+d2s[5]) + bias.Y
if !(image.Point{sx0, sy0}).In(sr) {
continue
}
dst.Pix[d+0] = uint8((uint32(dst.Pix[d+0])*pa1/0xffff + pr) >> 8)
dst.Pix[d+1] = uint8((uint32(dst.Pix[d+1])*pa1/0xffff + pg) >> 8)
dst.Pix[d+2] = uint8((uint32(dst.Pix[d+2])*pa1/0xffff + pb) >> 8)
dst.Pix[d+3] = uint8((uint32(dst.Pix[d+3])*pa1/0xffff + pa) >> 8)
}
}
default:
pr, pg, pb, pa := src.C.RGBA()
pa1 := 0xffff - pa
dstColorRGBA64 := &color.RGBA64{}
dstColor := color.Color(dstColorRGBA64)
for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
dyf := float64(dr.Min.Y+int(dy)) + 0.5
for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx++ {
dxf := float64(dr.Min.X+int(dx)) + 0.5
sx0 := int(d2s[0]*dxf+d2s[1]*dyf+d2s[2]) + bias.X
sy0 := int(d2s[3]*dxf+d2s[4]*dyf+d2s[5]) + bias.Y
if !(image.Point{sx0, sy0}).In(sr) {
continue
}
qr, qg, qb, qa := dst.At(dr.Min.X+int(dx), dr.Min.Y+int(dy)).RGBA()
dstColorRGBA64.R = uint16(qr*pa1/0xffff + pr)
dstColorRGBA64.G = uint16(qg*pa1/0xffff + pg)
dstColorRGBA64.B = uint16(qb*pa1/0xffff + pb)
dstColorRGBA64.A = uint16(qa*pa1/0xffff + pa)
dst.Set(dr.Min.X+int(dx), dr.Min.Y+int(dy), dstColor)
}
}
}
case Src:
switch dst := dst.(type) {
case *image.RGBA:
pr, pg, pb, pa := src.C.RGBA()
pr8 := uint8(pr >> 8)
pg8 := uint8(pg >> 8)
pb8 := uint8(pb >> 8)
pa8 := uint8(pa >> 8)
for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
dyf := float64(dr.Min.Y+int(dy)) + 0.5
d := dst.PixOffset(dr.Min.X+adr.Min.X, dr.Min.Y+int(dy))
for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx, d = dx+1, d+4 {
dxf := float64(dr.Min.X+int(dx)) + 0.5
sx0 := int(d2s[0]*dxf+d2s[1]*dyf+d2s[2]) + bias.X
sy0 := int(d2s[3]*dxf+d2s[4]*dyf+d2s[5]) + bias.Y
if !(image.Point{sx0, sy0}).In(sr) {
continue
}
dst.Pix[d+0] = pr8
dst.Pix[d+1] = pg8
dst.Pix[d+2] = pb8
dst.Pix[d+3] = pa8
}
}
default:
pr, pg, pb, pa := src.C.RGBA()
dstColorRGBA64 := &color.RGBA64{
uint16(pr),
uint16(pg),
uint16(pb),
uint16(pa),
}
dstColor := color.Color(dstColorRGBA64)
for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
dyf := float64(dr.Min.Y+int(dy)) + 0.5
for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx++ {
dxf := float64(dr.Min.X+int(dx)) + 0.5
sx0 := int(d2s[0]*dxf+d2s[1]*dyf+d2s[2]) + bias.X
sy0 := int(d2s[3]*dxf+d2s[4]*dyf+d2s[5]) + bias.Y
if !(image.Point{sx0, sy0}).In(sr) {
continue
}
dst.Set(dr.Min.X+int(dx), dr.Min.Y+int(dy), dstColor)
}
}
}
}
}
func opaque(m image.Image) bool {
o, ok := m.(interface {
Opaque() bool
})
return ok && o.Opaque()
}