golang-image/vector/raster_fixed.go
Nigel Tao fa54d6fa1c vector: simplify fixedLineTo computation.
name                              old time/op  new time/op   delta
GlyphAlpha16Over-8                3.38µs ± 0%   3.36µs ± 0%   -0.54%        (p=0.000 n=10+10)
GlyphAlpha16Src-8                 3.28µs ± 0%   3.26µs ± 0%   -0.69%         (p=0.000 n=10+9)
GlyphAlpha32Over-8                5.23µs ± 1%   5.20µs ± 0%   -0.58%         (p=0.000 n=10+8)
GlyphAlpha32Src-8                 4.83µs ± 0%   4.81µs ± 1%   -0.46%         (p=0.001 n=10+9)
GlyphAlpha64Over-8                10.2µs ± 0%   10.2µs ± 0%   -0.21%         (p=0.003 n=10+9)
GlyphAlpha64Src-8                 8.68µs ± 1%   8.62µs ± 0%   -0.70%         (p=0.000 n=10+9)
GlyphAlpha128Over-8               24.2µs ± 1%   24.1µs ± 0%   -0.58%         (p=0.001 n=10+9)
GlyphAlpha128Src-8                18.0µs ± 1%   17.9µs ± 0%   -0.61%         (p=0.001 n=10+9)
GlyphAlpha256Over-8               70.3µs ± 0%   70.1µs ± 0%   -0.37%        (p=0.019 n=10+10)
GlyphAlpha256Src-8                45.4µs ± 0%   45.2µs ± 1%   -0.38%         (p=0.041 n=8+10)
GlyphRGBA16Over-8                 5.14µs ± 0%   5.12µs ± 0%   -0.43%         (p=0.000 n=9+10)
GlyphRGBA16Src-8                  4.59µs ± 0%   4.57µs ± 1%   -0.43%          (p=0.005 n=9+9)
GlyphRGBA32Over-8                 12.2µs ± 1%   12.1µs ± 0%   -0.70%        (p=0.000 n=10+10)
GlyphRGBA32Src-8                  10.0µs ± 0%   10.0µs ± 0%     ~            (p=0.092 n=7+10)
GlyphRGBA64Over-8                 37.5µs ± 1%   37.2µs ± 0%   -0.75%         (p=0.000 n=10+9)
GlyphRGBA64Src-8                  29.1µs ± 0%   29.0µs ± 1%     ~            (p=0.243 n=10+9)
GlyphRGBA128Over-8                 135µs ± 0%    134µs ± 1%   -0.72%          (p=0.000 n=9+9)
GlyphRGBA128Src-8                  101µs ± 1%    100µs ± 1%     ~           (p=0.197 n=10+10)
GlyphRGBA256Over-8                 511µs ± 0%    506µs ± 0%   -0.97%        (p=0.000 n=10+10)
GlyphRGBA256Src-8                  374µs ± 0%    373µs ± 0%   -0.29%        (p=0.002 n=10+10)

Change-Id: Ic05a900935cb59e55711374db1e62b055d75c8e3
Reviewed-on: https://go-review.googlesource.com/31116
Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-16 03:06:30 +00:00

345 lines
11 KiB
Go

// Copyright 2016 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.
package vector
// This file contains a fixed point math implementation of the vector
// graphics rasterizer.
import (
"golang.org/x/image/math/f32"
)
const (
// ϕ is the number of binary digits after the fixed point.
//
// For example, if ϕ == 10 (and int1ϕ is based on the int32 type) then we
// are using 22.10 fixed point math.
//
// When changing this number, also change the assembly code (search for ϕ
// in the .s files).
//
// TODO: drop ϕ from 10 to 9, so that ±1<<(3*ϕ+3) doesn't overflow an int32
// and we can therefore use int32 math instead of the slower int64 math in
// Rasterizer.fixedLineTo below.
ϕ = 10
fxOne int1ϕ = 1 << ϕ
fxOneAndAHalf int1ϕ = 1<<ϕ + 1<<(ϕ-1)
fxOneMinusIota int1ϕ = 1<<ϕ - 1 // Used for rounding up.
)
// int1ϕ is a signed fixed-point number with 1*ϕ binary digits after the fixed
// point.
type int1ϕ int32
// int2ϕ is a signed fixed-point number with 2*ϕ binary digits after the fixed
// point.
//
// The Rasterizer's bufU32 field, nominally of type []uint32 (since that slice
// is also used by other code), can be thought of as a []int2ϕ during the
// fixedLineTo method. Lines of code that are actually like:
// buf[i] += uint32(etc) // buf has type []uint32.
// can be thought of as
// buf[i] += int2ϕ(etc) // buf has type []int2ϕ.
type int2ϕ int32
func fixedMax(x, y int1ϕ) int1ϕ {
if x > y {
return x
}
return y
}
func fixedMin(x, y int1ϕ) int1ϕ {
if x < y {
return x
}
return y
}
func fixedFloor(x int1ϕ) int32 { return int32(x >> ϕ) }
func fixedCeil(x int1ϕ) int32 { return int32((x + fxOneMinusIota) >> ϕ) }
func (z *Rasterizer) fixedLineTo(b f32.Vec2) {
a := z.pen
z.pen = b
dir := int1ϕ(1)
if a[1] > b[1] {
dir, a, b = -1, b, a
}
// Horizontal line segments yield no change in coverage. Almost horizontal
// segments would yield some change, in ideal math, but the computation
// further below, involving 1 / (b[1] - a[1]), is unstable in fixed point
// math, so we treat the segment as if it was perfectly horizontal.
if b[1]-a[1] <= 0.000001 {
return
}
dxdy := (b[0] - a[0]) / (b[1] - a[1])
ay := int1ϕ(a[1] * float32(fxOne))
by := int1ϕ(b[1] * float32(fxOne))
x := int1ϕ(a[0] * float32(fxOne))
y := fixedFloor(ay)
yMax := fixedCeil(by)
if yMax > int32(z.size.Y) {
yMax = int32(z.size.Y)
}
width := int32(z.size.X)
for ; y < yMax; y++ {
dy := fixedMin(int1ϕ(y+1)<<ϕ, by) - fixedMax(int1ϕ(y)<<ϕ, ay)
xNext := x + int1ϕ(float32(dy)*dxdy)
if y < 0 {
x = xNext
continue
}
buf := z.bufU32[y*width:]
d := dy * dir // d ranges up to ±1<<(1*ϕ).
x0, x1 := x, xNext
if x > xNext {
x0, x1 = x1, x0
}
x0i := fixedFloor(x0)
x0Floor := int1ϕ(x0i) << ϕ
x1i := fixedCeil(x1)
x1Ceil := int1ϕ(x1i) << ϕ
if x1i <= x0i+1 {
xmf := (x+xNext)>>1 - x0Floor
if i := clamp(x0i+0, width); i < uint(len(buf)) {
buf[i] += uint32(d * (fxOne - xmf))
}
if i := clamp(x0i+1, width); i < uint(len(buf)) {
buf[i] += uint32(d * xmf)
}
} else {
oneOverS := x1 - x0
twoOverS := 2 * oneOverS
x0f := x0 - x0Floor
oneMinusX0f := fxOne - x0f
oneMinusX0fSquared := oneMinusX0f * oneMinusX0f
x1f := x1 - x1Ceil + fxOne
x1fSquared := x1f * x1f
// These next two variables are unused, as rounding errors are
// minimized when we delay the division by oneOverS for as long as
// possible. These lines of code (and the "In ideal math" comments
// below) are commented out instead of deleted in order to aid the
// comparison with the floating point version of the rasterizer.
//
// a0 := ((oneMinusX0f * oneMinusX0f) >> 1) / oneOverS
// am := ((x1f * x1f) >> 1) / oneOverS
if i := clamp(x0i, width); i < uint(len(buf)) {
// In ideal math: buf[i] += uint32(d * a0)
D := oneMinusX0fSquared // D ranges up to ±1<<(1*ϕ).
D *= d // D ranges up to ±1<<(2*ϕ).
D /= twoOverS
buf[i] += uint32(D)
}
if x1i == x0i+2 {
if i := clamp(x0i+1, width); i < uint(len(buf)) {
// In ideal math: buf[i] += uint32(d * (fxOne - a0 - am))
//
// (x1i == x0i+2) and (twoOverS == 2 * (x1 - x0)) implies
// that int64(twoOverS) ranges up to +1<<(1*ϕ+2).
//
// Convert to int64 to avoid overflow. Without that,
// TestRasterize30Degrees fails.
D := int64(twoOverS<<ϕ - oneMinusX0fSquared - x1fSquared) // D ranges up to ±1<<(2*ϕ+2).
D *= int64(d) // D ranges up to ±1<<(3*ϕ+2).
D /= int64(twoOverS)
buf[i] += uint32(D)
}
} else {
// This is commented out for the same reason as a0 and am.
//
// a1 := ((fxOneAndAHalf - x0f) << ϕ) / oneOverS
if i := clamp(x0i+1, width); i < uint(len(buf)) {
// In ideal math:
// buf[i] += uint32(d * (a1 - a0))
// or equivalently (but better in non-ideal, integer math,
// with respect to rounding errors),
// buf[i] += uint32(A * d / twoOverS)
// where
// A = (a1 - a0) * twoOverS
// = a1*twoOverS - a0*twoOverS
// Noting that twoOverS/oneOverS equals 2, substituting for
// a0 and then a1, given above, yields:
// A = a1*twoOverS - oneMinusX0fSquared
// = (fxOneAndAHalf-x0f)<<(ϕ+1) - oneMinusX0fSquared
// = fxOneAndAHalf<<(ϕ+1) - x0f<<(ϕ+1) - oneMinusX0fSquared
//
// This is a positive number minus two non-negative
// numbers. For an upper bound on A, the positive number is
// P = fxOneAndAHalf<<(ϕ+1)
// < (2*fxOne)<<(ϕ+1)
// = fxOne<<(ϕ+2)
// = 1<<(2*ϕ+2)
//
// For a lower bound on A, the two non-negative numbers are
// N = x0f<<(ϕ+1) + oneMinusX0fSquared
// ≤ x0f<<(ϕ+1) + fxOne*fxOne
// = x0f<<(ϕ+1) + 1<<(2*ϕ)
// < x0f<<(ϕ+1) + 1<<(2*ϕ+1)
// ≤ fxOne<<(ϕ+1) + 1<<(2*ϕ+1)
// = 1<<(2*ϕ+1) + 1<<(2*ϕ+1)
// = 1<<(2*ϕ+2)
//
// Thus, A ranges up to ±1<<(2*ϕ+2). It is possible to
// derive a tighter bound, but this bound is sufficient to
// reason about overflow.
//
// Convert to int64 to avoid overflow. Without that,
// TestRasterizePolygon fails.
D := int64((fxOneAndAHalf-x0f)<<(ϕ+1) - oneMinusX0fSquared) // D ranges up to ±1<<(2*ϕ+2).
D *= int64(d) // D ranges up to ±1<<(3*ϕ+2).
D /= int64(twoOverS)
buf[i] += uint32(D)
}
dTimesS := uint32((d << (2 * ϕ)) / oneOverS)
for xi := x0i + 2; xi < x1i-1; xi++ {
if i := clamp(xi, width); i < uint(len(buf)) {
buf[i] += dTimesS
}
}
// This is commented out for the same reason as a0 and am.
//
// a2 := a1 + (int1ϕ(x1i-x0i-3)<<(2*ϕ))/oneOverS
if i := clamp(x1i-1, width); i < uint(len(buf)) {
// In ideal math:
// buf[i] += uint32(d * (fxOne - a2 - am))
// or equivalently (but better in non-ideal, integer math,
// with respect to rounding errors),
// buf[i] += uint32(A * d / twoOverS)
// where
// A = (fxOne - a2 - am) * twoOverS
// = twoOverS<<ϕ - a2*twoOverS - am*twoOverS
// Noting that twoOverS/oneOverS equals 2, substituting for
// am and then a2, given above, yields:
// A = twoOverS<<ϕ - a2*twoOverS - x1f*x1f
// = twoOverS<<ϕ - a1*twoOverS - (int1ϕ(x1i-x0i-3)<<(2*ϕ))*2 - x1f*x1f
// = twoOverS<<ϕ - a1*twoOverS - int1ϕ(x1i-x0i-3)<<(2*ϕ+1) - x1f*x1f
// Substituting for a1, given above, yields:
// A = twoOverS<<ϕ - ((fxOneAndAHalf-x0f)<<ϕ)*2 - int1ϕ(x1i-x0i-3)<<(2*ϕ+1) - x1f*x1f
// = twoOverS<<ϕ - (fxOneAndAHalf-x0f)<<(ϕ+1) - int1ϕ(x1i-x0i-3)<<(2*ϕ+1) - x1f*x1f
// = B<<ϕ - x1f*x1f
// where
// B = twoOverS - (fxOneAndAHalf-x0f)<<1 - int1ϕ(x1i-x0i-3)<<(ϕ+1)
// = (x1-x0)<<1 - (fxOneAndAHalf-x0f)<<1 - int1ϕ(x1i-x0i-3)<<(ϕ+1)
//
// Re-arranging the defintions given above:
// x0Floor := int1ϕ(x0i) << ϕ
// x0f := x0 - x0Floor
// x1Ceil := int1ϕ(x1i) << ϕ
// x1f := x1 - x1Ceil + fxOne
// combined with fxOne = 1<<ϕ yields:
// x0 = x0f + int1ϕ(x0i)<<ϕ
// x1 = x1f + int1ϕ(x1i-1)<<ϕ
// so that expanding (x1-x0) yields:
// B = (x1f-x0f + int1ϕ(x1i-x0i-1)<<ϕ)<<1 - (fxOneAndAHalf-x0f)<<1 - int1ϕ(x1i-x0i-3)<<(ϕ+1)
// = (x1f-x0f)<<1 + int1ϕ(x1i-x0i-1)<<(ϕ+1) - (fxOneAndAHalf-x0f)<<1 - int1ϕ(x1i-x0i-3)<<(ϕ+1)
// A large part of the second and fourth terms cancel:
// B = (x1f-x0f)<<1 - (fxOneAndAHalf-x0f)<<1 - int1ϕ(-2)<<(ϕ+1)
// = (x1f-x0f)<<1 - (fxOneAndAHalf-x0f)<<1 + 1<<(ϕ+2)
// = (x1f - fxOneAndAHalf)<<1 + 1<<(ϕ+2)
// The first term, (x1f - fxOneAndAHalf)<<1, is a negative
// number, bounded below by -fxOneAndAHalf<<1, which is
// greater than -fxOne<<2, or -1<<(ϕ+2). Thus, B ranges up
// to ±1<<(ϕ+2). One final simplification:
// B = x1f<<1 + (1<<(ϕ+2) - fxOneAndAHalf<<1)
//
// Convert to int64 to avoid overflow. Without that,
// TestRasterizePolygon fails.
const C = 1<<(ϕ+2) - fxOneAndAHalf<<1
D := int64(x1f<<1 + C) // D ranges up to ±1<<(1*ϕ+2).
D <<= ϕ // D ranges up to ±1<<(2*ϕ+2).
D -= int64(x1fSquared) // D ranges up to ±1<<(2*ϕ+3).
D *= int64(d) // D ranges up to ±1<<(3*ϕ+3).
D /= int64(twoOverS)
buf[i] += uint32(D)
}
}
if i := clamp(x1i, width); i < uint(len(buf)) {
// In ideal math: buf[i] += uint32(d * am)
D := x1fSquared // D ranges up to ±1<<(2*ϕ).
D *= d // D ranges up to ±1<<(3*ϕ).
D /= twoOverS
buf[i] += uint32(D)
}
}
x = xNext
}
}
func fixedAccumulateOpOver(dst []uint8, src []uint32) {
// Sanity check that len(dst) >= len(src).
if len(dst) < len(src) {
return
}
acc := int2ϕ(0)
for i, v := range src {
acc += int2ϕ(v)
a := acc
if a < 0 {
a = -a
}
a >>= 2*ϕ - 16
if a > 0xffff {
a = 0xffff
}
// This algorithm comes from the standard library's image/draw package.
dstA := uint32(dst[i]) * 0x101
maskA := uint32(a)
outA := dstA*(0xffff-maskA)/0xffff + maskA
dst[i] = uint8(outA >> 8)
}
}
func fixedAccumulateOpSrc(dst []uint8, src []uint32) {
// Sanity check that len(dst) >= len(src).
if len(dst) < len(src) {
return
}
acc := int2ϕ(0)
for i, v := range src {
acc += int2ϕ(v)
a := acc
if a < 0 {
a = -a
}
a >>= 2*ϕ - 8
if a > 0xff {
a = 0xff
}
dst[i] = uint8(a)
}
}
func fixedAccumulateMask(buf []uint32) {
acc := int2ϕ(0)
for i, v := range buf {
acc += int2ϕ(v)
a := acc
if a < 0 {
a = -a
}
a >>= 2*ϕ - 16
if a > 0xffff {
a = 0xffff
}
buf[i] = uint32(a)
}
}