status-im
/
resize
mirror of https://github.com/status-im/resize.git
2
0
Fork 0
Code Issues Projects Releases Wiki Activity
resize/converter.go

276 lines
7.7 KiB
Go
Raw Normal View History

Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
/*
Copyright (c) 2012, Jan Schlicht <jan.schlicht@gmail.com>
Permission to use, copy, modify, and/or distribute this software for any purpose
with or without fee is hereby granted, provided that the above copyright notice
and this permission notice appear in all copies.
THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND
FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER
TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF
THIS SOFTWARE.
*/
package resize
import (
"image"
"image/color"
)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
// Keep value in [0,255] range.
func clampUint8(in int32) uint8 {
if in < 0 {
return 0
Replicate border instead of using a constant Constant border could lead to wrong values alpha values with certain configurations.
2012-12-10 18:27:21 +00:00
}
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
if in > 255 {
return 255
}
return uint8(in)
Replicate border instead of using a constant Constant border could lead to wrong values alpha values with certain configurations.
2012-12-10 18:27:21 +00:00
}
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
// Keep value in [0,65535] range.
func clampUint16(in int64) uint16 {
if in < 0 {
return 0
}
if in > 65535 {
return 65535
}
return uint16(in)
Replicate border instead of using a constant Constant border could lead to wrong values alpha values with certain configurations.
2012-12-10 18:27:21 +00:00
}
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
func resizeGeneric(in image.Image, out *image.RGBA64, scale float64, coeffs []int32, offset []int, filterLength int) {
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
oldBounds := in.Bounds()
newBounds := out.Bounds()
Change ordering
2012-12-11 18:57:34 +00:00
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for x := newBounds.Min.X; x < newBounds.Max.X; x++ {
for y := newBounds.Min.Y; y < newBounds.Max.Y; y++ {
var rgba [4]int64
var sum int64
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
start := offset[y]
ci := (y - newBounds.Min.Y) * filterLength
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for i := 0; i < filterLength; i++ {
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
coeff := coeffs[ci+i]
if coeff != 0 {
xi := start + i
switch {
case uint(xi) < uint(oldBounds.Max.X):
break
case xi >= oldBounds.Max.X:
xi = oldBounds.Min.X
default:
xi = oldBounds.Max.X - 1
}
r, g, b, a := in.At(xi, x).RGBA()
rgba[0] += int64(coeff) * int64(r)
rgba[1] += int64(coeff) * int64(g)
rgba[2] += int64(coeff) * int64(b)
rgba[3] += int64(coeff) * int64(a)
sum += int64(coeff)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
}
}
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
offset := (y-newBounds.Min.Y)*out.Stride + (x-newBounds.Min.X)*8
value := clampUint16(rgba[0] / sum)
out.Pix[offset+0] = uint8(value >> 8)
out.Pix[offset+1] = uint8(value)
value = clampUint16(rgba[1] / sum)
out.Pix[offset+2] = uint8(value >> 8)
out.Pix[offset+3] = uint8(value)
value = clampUint16(rgba[2] / sum)
out.Pix[offset+4] = uint8(value >> 8)
out.Pix[offset+5] = uint8(value)
value = clampUint16(rgba[3] / sum)
out.Pix[offset+6] = uint8(value >> 8)
out.Pix[offset+7] = uint8(value)
}
}
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
}
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
func resizeRGBA(in *image.RGBA, out *image.RGBA, scale float64, coeffs []int16, offset []int, filterLength int) {
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
oldBounds := in.Bounds()
newBounds := out.Bounds()
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
minX := oldBounds.Min.X * 4
maxX := (oldBounds.Max.X - oldBounds.Min.X - 1) * 4
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for x := newBounds.Min.X; x < newBounds.Max.X; x++ {
row := in.Pix[(x-oldBounds.Min.Y)*in.Stride:]
for y := newBounds.Min.Y; y < newBounds.Max.Y; y++ {
var rgba [4]int32
var sum int32
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
start := offset[y]
ci := (y - newBounds.Min.Y) * filterLength
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for i := 0; i < filterLength; i++ {
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
coeff := coeffs[ci+i]
if coeff != 0 {
xi := start + i
switch {
case uint(xi) < uint(oldBounds.Max.X):
xi *= 4
case xi >= oldBounds.Max.X:
xi = maxX
default:
xi = minX
}
rgba[0] += int32(coeff) * int32(row[xi+0])
rgba[1] += int32(coeff) * int32(row[xi+1])
rgba[2] += int32(coeff) * int32(row[xi+2])
rgba[3] += int32(coeff) * int32(row[xi+3])
sum += int32(coeff)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
}
}
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
xo := (y-newBounds.Min.Y)*out.Stride + (x-newBounds.Min.X)*4
out.Pix[xo+0] = clampUint8(rgba[0] / sum)
out.Pix[xo+1] = clampUint8(rgba[1] / sum)
out.Pix[xo+2] = clampUint8(rgba[2] / sum)
out.Pix[xo+3] = clampUint8(rgba[3] / sum)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
}
}
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
}
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
func resizeRGBA64(in *image.RGBA64, out *image.RGBA64, scale float64, coeffs []int32, offset []int, filterLength int) {
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
oldBounds := in.Bounds()
newBounds := out.Bounds()
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
minX := oldBounds.Min.X * 8
maxX := (oldBounds.Max.X - oldBounds.Min.X - 1) * 8
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for x := newBounds.Min.X; x < newBounds.Max.X; x++ {
row := in.Pix[(x-oldBounds.Min.Y)*in.Stride:]
for y := newBounds.Min.Y; y < newBounds.Max.Y; y++ {
var rgba [4]int64
var sum int64
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
start := offset[y]
ci := (y - newBounds.Min.Y) * filterLength
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for i := 0; i < filterLength; i++ {
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
coeff := coeffs[ci+i]
if coeff != 0 {
xi := start + i
switch {
case uint(xi) < uint(oldBounds.Max.X):
xi *= 8
case xi >= oldBounds.Max.X:
xi = maxX
default:
xi = minX
}
rgba[0] += int64(coeff) * int64(uint16(row[xi+0])<<8|uint16(row[xi+1]))
rgba[1] += int64(coeff) * int64(uint16(row[xi+2])<<8|uint16(row[xi+3]))
rgba[2] += int64(coeff) * int64(uint16(row[xi+4])<<8|uint16(row[xi+5]))
rgba[3] += int64(coeff) * int64(uint16(row[xi+6])<<8|uint16(row[xi+7]))
sum += int64(coeff)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
}
}
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
xo := (y-newBounds.Min.Y)*out.Stride + (x-newBounds.Min.X)*8
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
value := clampUint16(rgba[0] / sum)
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
out.Pix[xo+0] = uint8(value >> 8)
out.Pix[xo+1] = uint8(value)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
value = clampUint16(rgba[1] / sum)
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
out.Pix[xo+2] = uint8(value >> 8)
out.Pix[xo+3] = uint8(value)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
value = clampUint16(rgba[2] / sum)
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
out.Pix[xo+4] = uint8(value >> 8)
out.Pix[xo+5] = uint8(value)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
value = clampUint16(rgba[3] / sum)
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
out.Pix[xo+6] = uint8(value >> 8)
out.Pix[xo+7] = uint8(value)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
}
}
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
}
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
func resizeGray(in *image.Gray, out *image.Gray, scale float64, coeffs []int16, offset []int, filterLength int) {
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
oldBounds := in.Bounds()
newBounds := out.Bounds()
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
minX := oldBounds.Min.X
maxX := (oldBounds.Max.X - oldBounds.Min.X - 1)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for x := newBounds.Min.X; x < newBounds.Max.X; x++ {
row := in.Pix[(x-oldBounds.Min.Y)*in.Stride:]
for y := newBounds.Min.Y; y < newBounds.Max.Y; y++ {
var gray int32
var sum int32
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
start := offset[y]
ci := (y - newBounds.Min.Y) * filterLength
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for i := 0; i < filterLength; i++ {
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
coeff := coeffs[ci+i]
if coeff != 0 {
xi := start + i
switch {
case uint(xi) < uint(oldBounds.Max.X):
break
case xi >= oldBounds.Max.X:
xi = maxX
default:
xi = minX
}
gray += int32(coeff) * int32(row[xi])
sum += int32(coeff)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
}
}
offset := (y-newBounds.Min.Y)*out.Stride + (x - newBounds.Min.X)
out.Pix[offset] = clampUint8(gray / sum)
}
}
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
}
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
func resizeGray16(in *image.Gray16, out *image.Gray16, scale float64, coeffs []int32, offset []int, filterLength int) {
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
oldBounds := in.Bounds()
newBounds := out.Bounds()
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
minX := oldBounds.Min.X * 2
maxX := (oldBounds.Max.X - oldBounds.Min.X - 1) * 2
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for x := newBounds.Min.X; x < newBounds.Max.X; x++ {
row := in.Pix[(x-oldBounds.Min.Y)*in.Stride:]
for y := newBounds.Min.Y; y < newBounds.Max.Y; y++ {
var gray int64
var sum int64
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
start := offset[y]
ci := (y - newBounds.Min.Y) * filterLength
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
for i := 0; i < filterLength; i++ {
Pre-calculate start offset, optimize bound check Increased performance by 30% for RGBA and 45% for Gray images, minor performance increase for 16-bit images. The start offset calculated by createWeights are stored in a slice and passed to the resize functions to prevent duplication of effort.
2014-07-29 20:53:35 +00:00
coeff := coeffs[ci+i]
if coeff != 0 {
xi := start + i
switch {
case uint(xi) < uint(oldBounds.Max.X):
xi *= 2
case xi >= oldBounds.Max.X:
xi = maxX
default:
xi = minX
}
gray += int64(coeff) * int64(uint16(row[xi+0])<<8|uint16(row[xi+1]))
sum += int64(coeff)
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
}
}
offset := (y-newBounds.Min.Y)*out.Stride + (x-newBounds.Min.X)*2
value := clampUint16(gray / sum)
out.Pix[offset+0] = uint8(value >> 8)
out.Pix[offset+1] = uint8(value)
}
}
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
}
Optimize data-locality for a huge increase in processing speed. This is a complete rewrite! The tight scaling loop needs data locality for optimal performance. The old version used lots of pointer redirections to access image data which was bad for data locality. By providing the complete loop for each image type, this problem is solved. Unfortunately this increases code duplication but the result should be worth it: I could measure a ~6x speed-up for certain test cases!
2014-07-19 11:19:31 +00:00
func convertYCbCrToRGBA(in *image.YCbCr) *image.RGBA {
out := image.NewRGBA(in.Bounds())
for y := 0; y < out.Bounds().Dy(); y++ {
for x := 0; x < out.Bounds().Dx(); x++ {
p := out.Pix[y*out.Stride+4*x:]
yi := in.YOffset(x, y)
ci := in.COffset(x, y)
r, g, b := color.YCbCrToRGB(in.Y[yi], in.Cb[ci], in.Cr[ci])
p[0] = r
p[1] = g
p[2] = b
p[3] = 0xff
}
}
return out
Speed up computation: Try to avoid Image.At() as much as possible -> specialized color access for some image types
2012-09-21 18:02:25 +00:00
}